linux/drivers/md/raid5.c

8614 lines
243 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/*
* raid5.c : Multiple Devices driver for Linux
* Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
* Copyright (C) 1999, 2000 Ingo Molnar
* Copyright (C) 2002, 2003 H. Peter Anvin
*
* RAID-4/5/6 management functions.
* Thanks to Penguin Computing for making the RAID-6 development possible
* by donating a test server!
*/
/*
* BITMAP UNPLUGGING:
*
* The sequencing for updating the bitmap reliably is a little
* subtle (and I got it wrong the first time) so it deserves some
* explanation.
*
* We group bitmap updates into batches. Each batch has a number.
* We may write out several batches at once, but that isn't very important.
* conf->seq_write is the number of the last batch successfully written.
* conf->seq_flush is the number of the last batch that was closed to
* new additions.
* When we discover that we will need to write to any block in a stripe
* (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
* the number of the batch it will be in. This is seq_flush+1.
* When we are ready to do a write, if that batch hasn't been written yet,
* we plug the array and queue the stripe for later.
* When an unplug happens, we increment bm_flush, thus closing the current
* batch.
* When we notice that bm_flush > bm_write, we write out all pending updates
* to the bitmap, and advance bm_write to where bm_flush was.
* This may occasionally write a bit out twice, but is sure never to
* miss any bits.
*/
#include <linux/blkdev.h>
#include <linux/kthread.h>
#include <linux/raid/pq.h>
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
#include <linux/async_tx.h>
#include <linux/module.h>
#include <linux/async.h>
#include <linux/seq_file.h>
#include <linux/cpu.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
#include <linux/slab.h>
#include <linux/ratelimit.h>
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
#include <linux/nodemask.h>
#include <trace/events/block.h>
#include <linux/list_sort.h>
#include "md.h"
#include "raid5.h"
#include "raid0.h"
#include "md-bitmap.h"
#include "raid5-log.h"
#define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
#define cpu_to_group(cpu) cpu_to_node(cpu)
#define ANY_GROUP NUMA_NO_NODE
md/raid5: disable 'DISCARD' by default due to safety concerns. It has come to my attention (thanks Martin) that 'discard_zeroes_data' is only a hint. Some devices in some cases don't do what it says on the label. The use of DISCARD in RAID5 depends on reads from discarded regions being predictably zero. If a write to a previously discarded region performs a read-modify-write cycle it assumes that the parity block was consistent with the data blocks. If all were zero, this would be the case. If some are and some aren't this would not be the case. This could lead to data corruption after a device failure when data needs to be reconstructed from the parity. As we cannot trust 'discard_zeroes_data', ignore it by default and so disallow DISCARD on all raid4/5/6 arrays. As many devices are trustworthy, and as there are benefits to using DISCARD, add a module parameter to over-ride this caution and cause DISCARD to work if discard_zeroes_data is set. If a site want to enable DISCARD on some arrays but not on others they should select DISCARD support at the filesystem level, and set the raid456 module parameter. raid456.devices_handle_discard_safely=Y As this is a data-safety issue, I believe this patch is suitable for -stable. DISCARD support for RAID456 was added in 3.7 Cc: Shaohua Li <shli@kernel.org> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: Heinz Mauelshagen <heinzm@redhat.com> Cc: stable@vger.kernel.org (3.7+) Acked-by: Martin K. Petersen <martin.petersen@oracle.com> Acked-by: Mike Snitzer <snitzer@redhat.com> Fixes: 620125f2bf8ff0c4969b79653b54d7bcc9d40637 Signed-off-by: NeilBrown <neilb@suse.de>
2014-10-02 03:45:00 +00:00
static bool devices_handle_discard_safely = false;
module_param(devices_handle_discard_safely, bool, 0644);
MODULE_PARM_DESC(devices_handle_discard_safely,
"Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
static struct workqueue_struct *raid5_wq;
static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
{
int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
return &conf->stripe_hashtbl[hash];
}
static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
{
return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
}
static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
{
spin_lock_irq(conf->hash_locks + hash);
spin_lock(&conf->device_lock);
}
static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
{
spin_unlock(&conf->device_lock);
spin_unlock_irq(conf->hash_locks + hash);
}
static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
{
int i;
spin_lock_irq(conf->hash_locks);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
spin_lock(&conf->device_lock);
}
static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
{
int i;
spin_unlock(&conf->device_lock);
for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
spin_unlock(conf->hash_locks + i);
spin_unlock_irq(conf->hash_locks);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
}
/* Find first data disk in a raid6 stripe */
static inline int raid6_d0(struct stripe_head *sh)
{
if (sh->ddf_layout)
/* ddf always start from first device */
return 0;
/* md starts just after Q block */
if (sh->qd_idx == sh->disks - 1)
return 0;
else
return sh->qd_idx + 1;
}
static inline int raid6_next_disk(int disk, int raid_disks)
{
disk++;
return (disk < raid_disks) ? disk : 0;
}
/* When walking through the disks in a raid5, starting at raid6_d0,
* We need to map each disk to a 'slot', where the data disks are slot
* 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
* is raid_disks-1. This help does that mapping.
*/
static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
int *count, int syndrome_disks)
{
int slot = *count;
if (sh->ddf_layout)
(*count)++;
if (idx == sh->pd_idx)
return syndrome_disks;
if (idx == sh->qd_idx)
return syndrome_disks + 1;
if (!sh->ddf_layout)
(*count)++;
return slot;
}
static void print_raid5_conf (struct r5conf *conf);
static int stripe_operations_active(struct stripe_head *sh)
{
return sh->check_state || sh->reconstruct_state ||
test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
test_bit(STRIPE_COMPUTE_RUN, &sh->state);
}
static bool stripe_is_lowprio(struct stripe_head *sh)
{
return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
!test_bit(STRIPE_R5C_CACHING, &sh->state);
}
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
{
struct r5conf *conf = sh->raid_conf;
struct r5worker_group *group;
int thread_cnt;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
int i, cpu = sh->cpu;
if (!cpu_online(cpu)) {
cpu = cpumask_any(cpu_online_mask);
sh->cpu = cpu;
}
if (list_empty(&sh->lru)) {
struct r5worker_group *group;
group = conf->worker_groups + cpu_to_group(cpu);
if (stripe_is_lowprio(sh))
list_add_tail(&sh->lru, &group->loprio_list);
else
list_add_tail(&sh->lru, &group->handle_list);
group->stripes_cnt++;
sh->group = group;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
}
if (conf->worker_cnt_per_group == 0) {
md_wakeup_thread(conf->mddev->thread);
return;
}
group = conf->worker_groups + cpu_to_group(sh->cpu);
group->workers[0].working = true;
/* at least one worker should run to avoid race */
queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
/* wakeup more workers */
for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
if (group->workers[i].working == false) {
group->workers[i].working = true;
queue_work_on(sh->cpu, raid5_wq,
&group->workers[i].work);
thread_cnt--;
}
}
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
struct list_head *temp_inactive_list)
{
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
int i;
int injournal = 0; /* number of date pages with R5_InJournal */
BUG_ON(!list_empty(&sh->lru));
BUG_ON(atomic_read(&conf->active_stripes)==0);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (r5c_is_writeback(conf->log))
for (i = sh->disks; i--; )
if (test_bit(R5_InJournal, &sh->dev[i].flags))
injournal++;
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
/*
* In the following cases, the stripe cannot be released to cached
* lists. Therefore, we make the stripe write out and set
* STRIPE_HANDLE:
* 1. when quiesce in r5c write back;
* 2. when resync is requested fot the stripe.
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
*/
if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
(conf->quiesce && r5c_is_writeback(conf->log) &&
!test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
if (test_bit(STRIPE_R5C_CACHING, &sh->state))
r5c_make_stripe_write_out(sh);
set_bit(STRIPE_HANDLE, &sh->state);
}
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (test_bit(STRIPE_HANDLE, &sh->state)) {
if (test_bit(STRIPE_DELAYED, &sh->state) &&
!test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
list_add_tail(&sh->lru, &conf->delayed_list);
else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
sh->bm_seq - conf->seq_write > 0)
list_add_tail(&sh->lru, &conf->bitmap_list);
else {
clear_bit(STRIPE_DELAYED, &sh->state);
clear_bit(STRIPE_BIT_DELAY, &sh->state);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
if (conf->worker_cnt_per_group == 0) {
if (stripe_is_lowprio(sh))
list_add_tail(&sh->lru,
&conf->loprio_list);
else
list_add_tail(&sh->lru,
&conf->handle_list);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
} else {
raid5_wakeup_stripe_thread(sh);
return;
}
}
md_wakeup_thread(conf->mddev->thread);
} else {
BUG_ON(stripe_operations_active(sh));
if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
if (atomic_dec_return(&conf->preread_active_stripes)
< IO_THRESHOLD)
md_wakeup_thread(conf->mddev->thread);
atomic_dec(&conf->active_stripes);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
if (!r5c_is_writeback(conf->log))
list_add_tail(&sh->lru, temp_inactive_list);
else {
WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
if (injournal == 0)
list_add_tail(&sh->lru, temp_inactive_list);
else if (injournal == conf->raid_disks - conf->max_degraded) {
/* full stripe */
if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
atomic_inc(&conf->r5c_cached_full_stripes);
if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
atomic_dec(&conf->r5c_cached_partial_stripes);
list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
r5c_check_cached_full_stripe(conf);
md/r5cache: enable chunk_aligned_read with write back cache Chunk aligned read significantly reduces CPU usage of raid456. However, it is not safe to fully bypass the write back cache. This patch enables chunk aligned read with write back cache. For chunk aligned read, we track stripes in write back cache at a bigger granularity, "big_stripe". Each chunk may contain more than one stripe (for example, a 256kB chunk contains 64 4kB-page, so this chunk contain 64 stripes). For chunk_aligned_read, these stripes are grouped into one big_stripe, so we only need one lookup for the whole chunk. For each big_stripe, struct big_stripe_info tracks how many stripes of this big_stripe are in the write back cache. We count how many stripes of this big_stripe are in the write back cache. These counters are tracked in a radix tree (big_stripe_tree). r5c_tree_index() is used to calculate keys for the radix tree. chunk_aligned_read() calls r5c_big_stripe_cached() to look up big_stripe of each chunk in the tree. If this big_stripe is in the tree, chunk_aligned_read() aborts. This look up is protected by rcu_read_lock(). It is necessary to remember whether a stripe is counted in big_stripe_tree. Instead of adding new flag, we reuses existing flags: STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these two flags are set, the stripe is counted in big_stripe_tree. This requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to r5c_try_caching_write(); and moving clear_bit of STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to r5c_finish_stripe_write_out(). Signed-off-by: Song Liu <songliubraving@fb.com> Reviewed-by: NeilBrown <neilb@suse.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-01-11 21:39:14 +00:00
} else
/*
* STRIPE_R5C_PARTIAL_STRIPE is set in
* r5c_try_caching_write(). No need to
* set it again.
*/
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
}
}
}
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
struct list_head *temp_inactive_list)
{
if (atomic_dec_and_test(&sh->count))
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
do_release_stripe(conf, sh, temp_inactive_list);
}
/*
* @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
*
* Be careful: Only one task can add/delete stripes from temp_inactive_list at
* given time. Adding stripes only takes device lock, while deleting stripes
* only takes hash lock.
*/
static void release_inactive_stripe_list(struct r5conf *conf,
struct list_head *temp_inactive_list,
int hash)
{
int size;
bool do_wakeup = false;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
unsigned long flags;
if (hash == NR_STRIPE_HASH_LOCKS) {
size = NR_STRIPE_HASH_LOCKS;
hash = NR_STRIPE_HASH_LOCKS - 1;
} else
size = 1;
while (size) {
struct list_head *list = &temp_inactive_list[size - 1];
/*
* We don't hold any lock here yet, raid5_get_active_stripe() might
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
* remove stripes from the list
*/
if (!list_empty_careful(list)) {
spin_lock_irqsave(conf->hash_locks + hash, flags);
if (list_empty(conf->inactive_list + hash) &&
!list_empty(list))
atomic_dec(&conf->empty_inactive_list_nr);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
list_splice_tail_init(list, conf->inactive_list + hash);
do_wakeup = true;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
spin_unlock_irqrestore(conf->hash_locks + hash, flags);
}
size--;
hash--;
}
if (do_wakeup) {
wake_up(&conf->wait_for_stripe);
if (atomic_read(&conf->active_stripes) == 0)
wake_up(&conf->wait_for_quiescent);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (conf->retry_read_aligned)
md_wakeup_thread(conf->mddev->thread);
}
}
/* should hold conf->device_lock already */
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
static int release_stripe_list(struct r5conf *conf,
struct list_head *temp_inactive_list)
{
struct stripe_head *sh, *t;
int count = 0;
struct llist_node *head;
head = llist_del_all(&conf->released_stripes);
head = llist_reverse_order(head);
llist_for_each_entry_safe(sh, t, head, release_list) {
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int hash;
/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
smp_mb();
clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
/*
* Don't worry the bit is set here, because if the bit is set
* again, the count is always > 1. This is true for
* STRIPE_ON_UNPLUG_LIST bit too.
*/
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
hash = sh->hash_lock_index;
__release_stripe(conf, sh, &temp_inactive_list[hash]);
count++;
}
return count;
}
void raid5_release_stripe(struct stripe_head *sh)
{
struct r5conf *conf = sh->raid_conf;
unsigned long flags;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
struct list_head list;
int hash;
bool wakeup;
raid5: avoid release list until last reference of the stripe The (lockless) release_list reduces lock contention, but there is excessive queueing and dequeuing of stripes on this list. A stripe will currently be queued on the release_list with a stripe reference count > 1. This can cause the raid5 kernel thread(s) to dequeue the stripe and decrement the refcount without doing any other useful processing of the stripe. The are two cases when the stripe can be put on the release_list multiple times before it is actually handled by the kernel thread(s). 1) make_request() activates the stripe processing in 4k increments. When a write request is large enough to span multiple chunks of a stripe_head, the first 4k chunk adds the stripe to the plug list. The next 4k chunk that is processed for the same stripe puts the stripe on the release_list with a refcount=2. This can cause the kernel thread to process and decrement the stripe before the stripe us unplugged, which again will put it back on the release_list. 2) Whenever IO is scheduled on a stripe (pre-read and/or write), the stripe refcount is set to the number of active IO (for each chunk). The stripe is released as each IO complete, and can be queued and dequeued multiple times on the release_list, until its refcount finally reached zero. This simple patch will ensure a stripe is only queued on the release_list when its refcount=1 and is ready to be handled by the kernel thread(s). I added some instrumentation to raid5 and counted the number of times striped were queued on the release_list for a variety of write IO sizes. Without this patch the number of times stripes got queued on the release_list was 100-500% higher than with the patch. The excess queuing will increase with the IO size. The patch also improved throughput by 5-10%. Signed-off-by: Eivind Sarto <esarto@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-05-28 03:39:23 +00:00
/* Avoid release_list until the last reference.
*/
if (atomic_add_unless(&sh->count, -1, 1))
return;
if (unlikely(!conf->mddev->thread) ||
test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
goto slow_path;
wakeup = llist_add(&sh->release_list, &conf->released_stripes);
if (wakeup)
md_wakeup_thread(conf->mddev->thread);
return;
slow_path:
/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
INIT_LIST_HEAD(&list);
hash = sh->hash_lock_index;
do_release_stripe(conf, sh, &list);
spin_unlock_irqrestore(&conf->device_lock, flags);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
release_inactive_stripe_list(conf, &list, hash);
}
}
static inline void remove_hash(struct stripe_head *sh)
{
pr_debug("remove_hash(), stripe %llu\n",
(unsigned long long)sh->sector);
hlist_del_init(&sh->hash);
}
static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
{
struct hlist_head *hp = stripe_hash(conf, sh->sector);
pr_debug("insert_hash(), stripe %llu\n",
(unsigned long long)sh->sector);
hlist_add_head(&sh->hash, hp);
}
/* find an idle stripe, make sure it is unhashed, and return it. */
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
{
struct stripe_head *sh = NULL;
struct list_head *first;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (list_empty(conf->inactive_list + hash))
goto out;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
first = (conf->inactive_list + hash)->next;
sh = list_entry(first, struct stripe_head, lru);
list_del_init(first);
remove_hash(sh);
atomic_inc(&conf->active_stripes);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
BUG_ON(hash != sh->hash_lock_index);
if (list_empty(conf->inactive_list + hash))
atomic_inc(&conf->empty_inactive_list_nr);
out:
return sh;
}
static void shrink_buffers(struct stripe_head *sh)
{
struct page *p;
int i;
int num = sh->raid_conf->pool_size;
for (i = 0; i < num ; i++) {
WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
p = sh->dev[i].page;
if (!p)
continue;
sh->dev[i].page = NULL;
put_page(p);
}
}
static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
{
int i;
int num = sh->raid_conf->pool_size;
for (i = 0; i < num; i++) {
struct page *page;
if (!(page = alloc_page(gfp))) {
return 1;
}
sh->dev[i].page = page;
sh->dev[i].orig_page = page;
}
return 0;
}
static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
struct stripe_head *sh);
static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
{
struct r5conf *conf = sh->raid_conf;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int i, seq;
BUG_ON(atomic_read(&sh->count) != 0);
BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
BUG_ON(stripe_operations_active(sh));
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
pr_debug("init_stripe called, stripe %llu\n",
(unsigned long long)sector);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
retry:
seq = read_seqcount_begin(&conf->gen_lock);
sh->generation = conf->generation - previous;
sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
sh->sector = sector;
stripe_set_idx(sector, conf, previous, sh);
sh->state = 0;
for (i = sh->disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (dev->toread || dev->read || dev->towrite || dev->written ||
test_bit(R5_LOCKED, &dev->flags)) {
pr_err("sector=%llx i=%d %p %p %p %p %d\n",
(unsigned long long)sh->sector, i, dev->toread,
dev->read, dev->towrite, dev->written,
test_bit(R5_LOCKED, &dev->flags));
WARN_ON(1);
}
dev->flags = 0;
dev->sector = raid5_compute_blocknr(sh, i, previous);
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (read_seqcount_retry(&conf->gen_lock, seq))
goto retry;
sh->overwrite_disks = 0;
insert_hash(conf, sh);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
sh->cpu = smp_processor_id();
set_bit(STRIPE_BATCH_READY, &sh->state);
}
static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
short generation)
{
struct stripe_head *sh;
pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
hlist: drop the node parameter from iterators I'm not sure why, but the hlist for each entry iterators were conceived list_for_each_entry(pos, head, member) The hlist ones were greedy and wanted an extra parameter: hlist_for_each_entry(tpos, pos, head, member) Why did they need an extra pos parameter? I'm not quite sure. Not only they don't really need it, it also prevents the iterator from looking exactly like the list iterator, which is unfortunate. Besides the semantic patch, there was some manual work required: - Fix up the actual hlist iterators in linux/list.h - Fix up the declaration of other iterators based on the hlist ones. - A very small amount of places were using the 'node' parameter, this was modified to use 'obj->member' instead. - Coccinelle didn't handle the hlist_for_each_entry_safe iterator properly, so those had to be fixed up manually. The semantic patch which is mostly the work of Peter Senna Tschudin is here: @@ iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host; type T; expression a,c,d,e; identifier b; statement S; @@ -T b; <+... when != b ( hlist_for_each_entry(a, - b, c, d) S | hlist_for_each_entry_continue(a, - b, c) S | hlist_for_each_entry_from(a, - b, c) S | hlist_for_each_entry_rcu(a, - b, c, d) S | hlist_for_each_entry_rcu_bh(a, - b, c, d) S | hlist_for_each_entry_continue_rcu_bh(a, - b, c) S | for_each_busy_worker(a, c, - b, d) S | ax25_uid_for_each(a, - b, c) S | ax25_for_each(a, - b, c) S | inet_bind_bucket_for_each(a, - b, c) S | sctp_for_each_hentry(a, - b, c) S | sk_for_each(a, - b, c) S | sk_for_each_rcu(a, - b, c) S | sk_for_each_from -(a, b) +(a) S + sk_for_each_from(a) S | sk_for_each_safe(a, - b, c, d) S | sk_for_each_bound(a, - b, c) S | hlist_for_each_entry_safe(a, - b, c, d, e) S | hlist_for_each_entry_continue_rcu(a, - b, c) S | nr_neigh_for_each(a, - b, c) S | nr_neigh_for_each_safe(a, - b, c, d) S | nr_node_for_each(a, - b, c) S | nr_node_for_each_safe(a, - b, c, d) S | - for_each_gfn_sp(a, c, d, b) S + for_each_gfn_sp(a, c, d) S | - for_each_gfn_indirect_valid_sp(a, c, d, b) S + for_each_gfn_indirect_valid_sp(a, c, d) S | for_each_host(a, - b, c) S | for_each_host_safe(a, - b, c, d) S | for_each_mesh_entry(a, - b, c, d) S ) ...+> [akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c] [akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c] [akpm@linux-foundation.org: checkpatch fixes] [akpm@linux-foundation.org: fix warnings] [akpm@linux-foudnation.org: redo intrusive kvm changes] Tested-by: Peter Senna Tschudin <peter.senna@gmail.com> Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Gleb Natapov <gleb@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 01:06:00 +00:00
hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
if (sh->sector == sector && sh->generation == generation)
return sh;
pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
return NULL;
}
/*
* Need to check if array has failed when deciding whether to:
* - start an array
* - remove non-faulty devices
* - add a spare
* - allow a reshape
* This determination is simple when no reshape is happening.
* However if there is a reshape, we need to carefully check
* both the before and after sections.
* This is because some failed devices may only affect one
* of the two sections, and some non-in_sync devices may
* be insync in the section most affected by failed devices.
*/
int raid5_calc_degraded(struct r5conf *conf)
{
int degraded, degraded2;
int i;
rcu_read_lock();
degraded = 0;
for (i = 0; i < conf->previous_raid_disks; i++) {
struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
if (rdev && test_bit(Faulty, &rdev->flags))
rdev = rcu_dereference(conf->disks[i].replacement);
if (!rdev || test_bit(Faulty, &rdev->flags))
degraded++;
else if (test_bit(In_sync, &rdev->flags))
;
else
/* not in-sync or faulty.
* If the reshape increases the number of devices,
* this is being recovered by the reshape, so
* this 'previous' section is not in_sync.
* If the number of devices is being reduced however,
* the device can only be part of the array if
* we are reverting a reshape, so this section will
* be in-sync.
*/
if (conf->raid_disks >= conf->previous_raid_disks)
degraded++;
}
rcu_read_unlock();
if (conf->raid_disks == conf->previous_raid_disks)
return degraded;
rcu_read_lock();
degraded2 = 0;
for (i = 0; i < conf->raid_disks; i++) {
struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
if (rdev && test_bit(Faulty, &rdev->flags))
rdev = rcu_dereference(conf->disks[i].replacement);
if (!rdev || test_bit(Faulty, &rdev->flags))
degraded2++;
else if (test_bit(In_sync, &rdev->flags))
;
else
/* not in-sync or faulty.
* If reshape increases the number of devices, this
* section has already been recovered, else it
* almost certainly hasn't.
*/
if (conf->raid_disks <= conf->previous_raid_disks)
degraded2++;
}
rcu_read_unlock();
if (degraded2 > degraded)
return degraded2;
return degraded;
}
static int has_failed(struct r5conf *conf)
{
int degraded;
if (conf->mddev->reshape_position == MaxSector)
return conf->mddev->degraded > conf->max_degraded;
degraded = raid5_calc_degraded(conf);
if (degraded > conf->max_degraded)
return 1;
return 0;
}
struct stripe_head *
raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
int previous, int noblock, int noquiesce)
{
struct stripe_head *sh;
int hash = stripe_hash_locks_hash(conf, sector);
int inc_empty_inactive_list_flag;
pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
spin_lock_irq(conf->hash_locks + hash);
do {
wait_event_lock_irq(conf->wait_for_quiescent,
conf->quiesce == 0 || noquiesce,
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
*(conf->hash_locks + hash));
sh = __find_stripe(conf, sector, conf->generation - previous);
if (!sh) {
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
sh = get_free_stripe(conf, hash);
if (!sh && !test_bit(R5_DID_ALLOC,
&conf->cache_state))
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
set_bit(R5_ALLOC_MORE,
&conf->cache_state);
}
if (noblock && sh == NULL)
break;
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
r5c_check_stripe_cache_usage(conf);
if (!sh) {
set_bit(R5_INACTIVE_BLOCKED,
&conf->cache_state);
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
r5l_wake_reclaim(conf->log, 0);
wait_event_lock_irq(
conf->wait_for_stripe,
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
!list_empty(conf->inactive_list + hash) &&
(atomic_read(&conf->active_stripes)
< (conf->max_nr_stripes * 3 / 4)
|| !test_bit(R5_INACTIVE_BLOCKED,
&conf->cache_state)),
*(conf->hash_locks + hash));
clear_bit(R5_INACTIVE_BLOCKED,
&conf->cache_state);
} else {
init_stripe(sh, sector, previous);
atomic_inc(&sh->count);
}
} else if (!atomic_inc_not_zero(&sh->count)) {
spin_lock(&conf->device_lock);
if (!atomic_read(&sh->count)) {
if (!test_bit(STRIPE_HANDLE, &sh->state))
atomic_inc(&conf->active_stripes);
BUG_ON(list_empty(&sh->lru) &&
!test_bit(STRIPE_EXPANDING, &sh->state));
inc_empty_inactive_list_flag = 0;
if (!list_empty(conf->inactive_list + hash))
inc_empty_inactive_list_flag = 1;
list_del_init(&sh->lru);
if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
atomic_inc(&conf->empty_inactive_list_nr);
if (sh->group) {
sh->group->stripes_cnt--;
sh->group = NULL;
}
}
atomic_inc(&sh->count);
spin_unlock(&conf->device_lock);
}
} while (sh == NULL);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
spin_unlock_irq(conf->hash_locks + hash);
return sh;
}
static bool is_full_stripe_write(struct stripe_head *sh)
{
BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
__acquires(&sh1->stripe_lock)
__acquires(&sh2->stripe_lock)
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
{
if (sh1 > sh2) {
spin_lock_irq(&sh2->stripe_lock);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
spin_lock_nested(&sh1->stripe_lock, 1);
} else {
spin_lock_irq(&sh1->stripe_lock);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
spin_lock_nested(&sh2->stripe_lock, 1);
}
}
static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
__releases(&sh1->stripe_lock)
__releases(&sh2->stripe_lock)
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
{
spin_unlock(&sh1->stripe_lock);
spin_unlock_irq(&sh2->stripe_lock);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
}
/* Only freshly new full stripe normal write stripe can be added to a batch list */
static bool stripe_can_batch(struct stripe_head *sh)
{
struct r5conf *conf = sh->raid_conf;
if (raid5_has_log(conf) || raid5_has_ppl(conf))
return false;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
return test_bit(STRIPE_BATCH_READY, &sh->state) &&
!test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
is_full_stripe_write(sh);
}
/* we only do back search */
static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
{
struct stripe_head *head;
sector_t head_sector, tmp_sec;
int hash;
int dd_idx;
int inc_empty_inactive_list_flag;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
/* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
tmp_sec = sh->sector;
if (!sector_div(tmp_sec, conf->chunk_sectors))
return;
head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
hash = stripe_hash_locks_hash(conf, head_sector);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
spin_lock_irq(conf->hash_locks + hash);
head = __find_stripe(conf, head_sector, conf->generation);
if (head && !atomic_inc_not_zero(&head->count)) {
spin_lock(&conf->device_lock);
if (!atomic_read(&head->count)) {
if (!test_bit(STRIPE_HANDLE, &head->state))
atomic_inc(&conf->active_stripes);
BUG_ON(list_empty(&head->lru) &&
!test_bit(STRIPE_EXPANDING, &head->state));
inc_empty_inactive_list_flag = 0;
if (!list_empty(conf->inactive_list + hash))
inc_empty_inactive_list_flag = 1;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
list_del_init(&head->lru);
if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
atomic_inc(&conf->empty_inactive_list_nr);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head->group) {
head->group->stripes_cnt--;
head->group = NULL;
}
}
atomic_inc(&head->count);
spin_unlock(&conf->device_lock);
}
spin_unlock_irq(conf->hash_locks + hash);
if (!head)
return;
if (!stripe_can_batch(head))
goto out;
lock_two_stripes(head, sh);
/* clear_batch_ready clear the flag */
if (!stripe_can_batch(head) || !stripe_can_batch(sh))
goto unlock_out;
if (sh->batch_head)
goto unlock_out;
dd_idx = 0;
while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
dd_idx++;
if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
goto unlock_out;
if (head->batch_head) {
spin_lock(&head->batch_head->batch_lock);
/* This batch list is already running */
if (!stripe_can_batch(head)) {
spin_unlock(&head->batch_head->batch_lock);
goto unlock_out;
}
/*
* We must assign batch_head of this stripe within the
* batch_lock, otherwise clear_batch_ready of batch head
* stripe could clear BATCH_READY bit of this stripe and
* this stripe->batch_head doesn't get assigned, which
* could confuse clear_batch_ready for this stripe
*/
sh->batch_head = head->batch_head;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
/*
* at this point, head's BATCH_READY could be cleared, but we
* can still add the stripe to batch list
*/
list_add(&sh->batch_list, &head->batch_list);
spin_unlock(&head->batch_head->batch_lock);
} else {
head->batch_head = head;
sh->batch_head = head->batch_head;
spin_lock(&head->batch_lock);
list_add_tail(&sh->batch_list, &head->batch_list);
spin_unlock(&head->batch_lock);
}
if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
if (atomic_dec_return(&conf->preread_active_stripes)
< IO_THRESHOLD)
md_wakeup_thread(conf->mddev->thread);
if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
int seq = sh->bm_seq;
if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
sh->batch_head->bm_seq > seq)
seq = sh->batch_head->bm_seq;
set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
sh->batch_head->bm_seq = seq;
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
atomic_inc(&sh->count);
unlock_out:
unlock_two_stripes(head, sh);
out:
raid5_release_stripe(head);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
}
/* Determine if 'data_offset' or 'new_data_offset' should be used
* in this stripe_head.
*/
static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
{
sector_t progress = conf->reshape_progress;
/* Need a memory barrier to make sure we see the value
* of conf->generation, or ->data_offset that was set before
* reshape_progress was updated.
*/
smp_rmb();
if (progress == MaxSector)
return 0;
if (sh->generation == conf->generation - 1)
return 0;
/* We are in a reshape, and this is a new-generation stripe,
* so use new_data_offset.
*/
return 1;
}
static void dispatch_bio_list(struct bio_list *tmp)
{
struct bio *bio;
while ((bio = bio_list_pop(tmp)))
submit_bio_noacct(bio);
}
static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
{
const struct r5pending_data *da = list_entry(a,
struct r5pending_data, sibling);
const struct r5pending_data *db = list_entry(b,
struct r5pending_data, sibling);
if (da->sector > db->sector)
return 1;
if (da->sector < db->sector)
return -1;
return 0;
}
static void dispatch_defer_bios(struct r5conf *conf, int target,
struct bio_list *list)
{
struct r5pending_data *data;
struct list_head *first, *next = NULL;
int cnt = 0;
if (conf->pending_data_cnt == 0)
return;
list_sort(NULL, &conf->pending_list, cmp_stripe);
first = conf->pending_list.next;
/* temporarily move the head */
if (conf->next_pending_data)
list_move_tail(&conf->pending_list,
&conf->next_pending_data->sibling);
while (!list_empty(&conf->pending_list)) {
data = list_first_entry(&conf->pending_list,
struct r5pending_data, sibling);
if (&data->sibling == first)
first = data->sibling.next;
next = data->sibling.next;
bio_list_merge(list, &data->bios);
list_move(&data->sibling, &conf->free_list);
cnt++;
if (cnt >= target)
break;
}
conf->pending_data_cnt -= cnt;
BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
if (next != &conf->pending_list)
conf->next_pending_data = list_entry(next,
struct r5pending_data, sibling);
else
conf->next_pending_data = NULL;
/* list isn't empty */
if (first != &conf->pending_list)
list_move_tail(&conf->pending_list, first);
}
static void flush_deferred_bios(struct r5conf *conf)
{
struct bio_list tmp = BIO_EMPTY_LIST;
if (conf->pending_data_cnt == 0)
return;
spin_lock(&conf->pending_bios_lock);
dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
BUG_ON(conf->pending_data_cnt != 0);
spin_unlock(&conf->pending_bios_lock);
dispatch_bio_list(&tmp);
}
static void defer_issue_bios(struct r5conf *conf, sector_t sector,
struct bio_list *bios)
{
struct bio_list tmp = BIO_EMPTY_LIST;
struct r5pending_data *ent;
spin_lock(&conf->pending_bios_lock);
ent = list_first_entry(&conf->free_list, struct r5pending_data,
sibling);
list_move_tail(&ent->sibling, &conf->pending_list);
ent->sector = sector;
bio_list_init(&ent->bios);
bio_list_merge(&ent->bios, bios);
conf->pending_data_cnt++;
if (conf->pending_data_cnt >= PENDING_IO_MAX)
dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
spin_unlock(&conf->pending_bios_lock);
dispatch_bio_list(&tmp);
}
static void
raid5_end_read_request(struct bio *bi);
static void
raid5_end_write_request(struct bio *bi);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
struct r5conf *conf = sh->raid_conf;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int i, disks = sh->disks;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct stripe_head *head_sh = sh;
struct bio_list pending_bios = BIO_EMPTY_LIST;
bool should_defer;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
might_sleep();
if (log_stripe(sh, s) == 0)
return;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
should_defer = conf->batch_bio_dispatch && conf->group_cnt;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
for (i = disks; i--; ) {
int op, op_flags = 0;
int replace_only = 0;
struct bio *bi, *rbi;
struct md_rdev *rdev, *rrdev = NULL;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
sh = head_sh;
2010-09-03 09:56:18 +00:00
if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
op = REQ_OP_WRITE;
2010-09-03 09:56:18 +00:00
if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
op_flags = REQ_FUA;
if (test_bit(R5_Discard, &sh->dev[i].flags))
op = REQ_OP_DISCARD;
2010-09-03 09:56:18 +00:00
} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
op = REQ_OP_READ;
else if (test_and_clear_bit(R5_WantReplace,
&sh->dev[i].flags)) {
op = REQ_OP_WRITE;
replace_only = 1;
} else
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
continue;
if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
op_flags |= REQ_SYNC;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
again:
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
bi = &sh->dev[i].req;
rbi = &sh->dev[i].rreq; /* For writing to replacement */
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
rcu_read_lock();
rrdev = rcu_dereference(conf->disks[i].replacement);
smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
rdev = rcu_dereference(conf->disks[i].rdev);
if (!rdev) {
rdev = rrdev;
rrdev = NULL;
}
if (op_is_write(op)) {
if (replace_only)
rdev = NULL;
if (rdev == rrdev)
/* We raced and saw duplicates */
rrdev = NULL;
} else {
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
rdev = rrdev;
rrdev = NULL;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (rdev && test_bit(Faulty, &rdev->flags))
rdev = NULL;
if (rdev)
atomic_inc(&rdev->nr_pending);
if (rrdev && test_bit(Faulty, &rrdev->flags))
rrdev = NULL;
if (rrdev)
atomic_inc(&rrdev->nr_pending);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
rcu_read_unlock();
/* We have already checked bad blocks for reads. Now
* need to check for writes. We never accept write errors
* on the replacement, so we don't to check rrdev.
*/
while (op_is_write(op) && rdev &&
test_bit(WriteErrorSeen, &rdev->flags)) {
sector_t first_bad;
int bad_sectors;
int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
&first_bad, &bad_sectors);
if (!bad)
break;
if (bad < 0) {
set_bit(BlockedBadBlocks, &rdev->flags);
if (!conf->mddev->external &&
conf->mddev->sb_flags) {
/* It is very unlikely, but we might
* still need to write out the
* bad block log - better give it
* a chance*/
md_check_recovery(conf->mddev);
}
/*
* Because md_wait_for_blocked_rdev
* will dec nr_pending, we must
* increment it first.
*/
atomic_inc(&rdev->nr_pending);
md_wait_for_blocked_rdev(rdev, conf->mddev);
} else {
/* Acknowledged bad block - skip the write */
rdev_dec_pending(rdev, conf->mddev);
rdev = NULL;
}
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (rdev) {
if (s->syncing || s->expanding || s->expanded
|| s->replacing)
md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_IO_STARTED, &sh->state);
bio_set_dev(bi, rdev->bdev);
bio_set_op_attrs(bi, op, op_flags);
bi->bi_end_io = op_is_write(op)
? raid5_end_write_request
: raid5_end_read_request;
bi->bi_private = sh;
pr_debug("%s: for %llu schedule op %d on disc %d\n",
__func__, (unsigned long long)sh->sector,
bi->bi_opf, i);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
atomic_inc(&sh->count);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (sh != head_sh)
atomic_inc(&head_sh->count);
if (use_new_offset(conf, sh))
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
bi->bi_iter.bi_sector = (sh->sector
+ rdev->new_data_offset);
else
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
bi->bi_iter.bi_sector = (sh->sector
+ rdev->data_offset);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
bi->bi_opf |= REQ_NOMERGE;
if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
if (!op_is_write(op) &&
test_bit(R5_InJournal, &sh->dev[i].flags))
/*
* issuing read for a page in journal, this
* must be preparing for prexor in rmw; read
* the data into orig_page
*/
sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
else
sh->dev[i].vec.bv_page = sh->dev[i].page;
bi->bi_vcnt = 1;
bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
bi->bi_io_vec[0].bv_offset = 0;
bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
bi->bi_write_hint = sh->dev[i].write_hint;
if (!rrdev)
sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
/*
* If this is discard request, set bi_vcnt 0. We don't
* want to confuse SCSI because SCSI will replace payload
*/
if (op == REQ_OP_DISCARD)
bi->bi_vcnt = 0;
if (rrdev)
set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
if (conf->mddev->gendisk)
trace_block_bio_remap(bi->bi_disk->queue,
bi, disk_devt(conf->mddev->gendisk),
sh->dev[i].sector);
if (should_defer && op_is_write(op))
bio_list_add(&pending_bios, bi);
else
submit_bio_noacct(bi);
}
if (rrdev) {
if (s->syncing || s->expanding || s->expanded
|| s->replacing)
md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
set_bit(STRIPE_IO_STARTED, &sh->state);
bio_set_dev(rbi, rrdev->bdev);
bio_set_op_attrs(rbi, op, op_flags);
BUG_ON(!op_is_write(op));
rbi->bi_end_io = raid5_end_write_request;
rbi->bi_private = sh;
pr_debug("%s: for %llu schedule op %d on "
"replacement disc %d\n",
__func__, (unsigned long long)sh->sector,
rbi->bi_opf, i);
atomic_inc(&sh->count);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (sh != head_sh)
atomic_inc(&head_sh->count);
if (use_new_offset(conf, sh))
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
rbi->bi_iter.bi_sector = (sh->sector
+ rrdev->new_data_offset);
else
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
rbi->bi_iter.bi_sector = (sh->sector
+ rrdev->data_offset);
if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
sh->dev[i].rvec.bv_page = sh->dev[i].page;
rbi->bi_vcnt = 1;
rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
rbi->bi_io_vec[0].bv_offset = 0;
rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
rbi->bi_write_hint = sh->dev[i].write_hint;
sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
/*
* If this is discard request, set bi_vcnt 0. We don't
* want to confuse SCSI because SCSI will replace payload
*/
if (op == REQ_OP_DISCARD)
rbi->bi_vcnt = 0;
if (conf->mddev->gendisk)
trace_block_bio_remap(rbi->bi_disk->queue,
rbi, disk_devt(conf->mddev->gendisk),
sh->dev[i].sector);
if (should_defer && op_is_write(op))
bio_list_add(&pending_bios, rbi);
else
submit_bio_noacct(rbi);
}
if (!rdev && !rrdev) {
if (op_is_write(op))
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_DEGRADED, &sh->state);
pr_debug("skip op %d on disc %d for sector %llu\n",
bi->bi_opf, i, (unsigned long long)sh->sector);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
clear_bit(R5_LOCKED, &sh->dev[i].flags);
set_bit(STRIPE_HANDLE, &sh->state);
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (!head_sh->batch_head)
continue;
sh = list_first_entry(&sh->batch_list, struct stripe_head,
batch_list);
if (sh != head_sh)
goto again;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
if (should_defer && !bio_list_empty(&pending_bios))
defer_issue_bios(conf, head_sh->sector, &pending_bios);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static struct dma_async_tx_descriptor *
async_copy_data(int frombio, struct bio *bio, struct page **page,
sector_t sector, struct dma_async_tx_descriptor *tx,
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
struct stripe_head *sh, int no_skipcopy)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
block: Convert bio_for_each_segment() to bvec_iter More prep work for immutable biovecs - with immutable bvecs drivers won't be able to use the biovec directly, they'll need to use helpers that take into account bio->bi_iter.bi_bvec_done. This updates callers for the new usage without changing the implementation yet. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Paul Clements <Paul.Clements@steeleye.com> Cc: Jim Paris <jim@jtan.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Nagalakshmi Nandigama <Nagalakshmi.Nandigama@lsi.com> Cc: Sreekanth Reddy <Sreekanth.Reddy@lsi.com> Cc: support@lsi.com Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Matthew Wilcox <matthew.r.wilcox@intel.com> Cc: Keith Busch <keith.busch@intel.com> Cc: Stephen Hemminger <shemminger@vyatta.com> Cc: Quoc-Son Anh <quoc-sonx.anh@intel.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: Vivek Goyal <vgoyal@redhat.com> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Jan Kara <jack@suse.cz> Cc: linux-m68k@lists.linux-m68k.org Cc: linuxppc-dev@lists.ozlabs.org Cc: drbd-user@lists.linbit.com Cc: nbd-general@lists.sourceforge.net Cc: cbe-oss-dev@lists.ozlabs.org Cc: xen-devel@lists.xensource.com Cc: virtualization@lists.linux-foundation.org Cc: linux-raid@vger.kernel.org Cc: linux-s390@vger.kernel.org Cc: DL-MPTFusionLinux@lsi.com Cc: linux-scsi@vger.kernel.org Cc: devel@driverdev.osuosl.org Cc: linux-fsdevel@vger.kernel.org Cc: cluster-devel@redhat.com Cc: linux-mm@kvack.org Acked-by: Geoff Levand <geoff@infradead.org>
2013-11-24 01:19:00 +00:00
struct bio_vec bvl;
struct bvec_iter iter;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct page *bio_page;
int page_offset;
struct async_submit_ctl submit;
enum async_tx_flags flags = 0;
struct r5conf *conf = sh->raid_conf;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
if (bio->bi_iter.bi_sector >= sector)
page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
else
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
if (frombio)
flags |= ASYNC_TX_FENCE;
init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
block: Convert bio_for_each_segment() to bvec_iter More prep work for immutable biovecs - with immutable bvecs drivers won't be able to use the biovec directly, they'll need to use helpers that take into account bio->bi_iter.bi_bvec_done. This updates callers for the new usage without changing the implementation yet. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Paul Clements <Paul.Clements@steeleye.com> Cc: Jim Paris <jim@jtan.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Nagalakshmi Nandigama <Nagalakshmi.Nandigama@lsi.com> Cc: Sreekanth Reddy <Sreekanth.Reddy@lsi.com> Cc: support@lsi.com Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Matthew Wilcox <matthew.r.wilcox@intel.com> Cc: Keith Busch <keith.busch@intel.com> Cc: Stephen Hemminger <shemminger@vyatta.com> Cc: Quoc-Son Anh <quoc-sonx.anh@intel.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: Vivek Goyal <vgoyal@redhat.com> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Jan Kara <jack@suse.cz> Cc: linux-m68k@lists.linux-m68k.org Cc: linuxppc-dev@lists.ozlabs.org Cc: drbd-user@lists.linbit.com Cc: nbd-general@lists.sourceforge.net Cc: cbe-oss-dev@lists.ozlabs.org Cc: xen-devel@lists.xensource.com Cc: virtualization@lists.linux-foundation.org Cc: linux-raid@vger.kernel.org Cc: linux-s390@vger.kernel.org Cc: DL-MPTFusionLinux@lsi.com Cc: linux-scsi@vger.kernel.org Cc: devel@driverdev.osuosl.org Cc: linux-fsdevel@vger.kernel.org Cc: cluster-devel@redhat.com Cc: linux-mm@kvack.org Acked-by: Geoff Levand <geoff@infradead.org>
2013-11-24 01:19:00 +00:00
bio_for_each_segment(bvl, bio, iter) {
int len = bvl.bv_len;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int clen;
int b_offset = 0;
if (page_offset < 0) {
b_offset = -page_offset;
page_offset += b_offset;
len -= b_offset;
}
if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
clen = RAID5_STRIPE_SIZE(conf) - page_offset;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
else
clen = len;
if (clen > 0) {
block: Convert bio_for_each_segment() to bvec_iter More prep work for immutable biovecs - with immutable bvecs drivers won't be able to use the biovec directly, they'll need to use helpers that take into account bio->bi_iter.bi_bvec_done. This updates callers for the new usage without changing the implementation yet. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Paul Clements <Paul.Clements@steeleye.com> Cc: Jim Paris <jim@jtan.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Nagalakshmi Nandigama <Nagalakshmi.Nandigama@lsi.com> Cc: Sreekanth Reddy <Sreekanth.Reddy@lsi.com> Cc: support@lsi.com Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Matthew Wilcox <matthew.r.wilcox@intel.com> Cc: Keith Busch <keith.busch@intel.com> Cc: Stephen Hemminger <shemminger@vyatta.com> Cc: Quoc-Son Anh <quoc-sonx.anh@intel.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: Vivek Goyal <vgoyal@redhat.com> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Jan Kara <jack@suse.cz> Cc: linux-m68k@lists.linux-m68k.org Cc: linuxppc-dev@lists.ozlabs.org Cc: drbd-user@lists.linbit.com Cc: nbd-general@lists.sourceforge.net Cc: cbe-oss-dev@lists.ozlabs.org Cc: xen-devel@lists.xensource.com Cc: virtualization@lists.linux-foundation.org Cc: linux-raid@vger.kernel.org Cc: linux-s390@vger.kernel.org Cc: DL-MPTFusionLinux@lsi.com Cc: linux-scsi@vger.kernel.org Cc: devel@driverdev.osuosl.org Cc: linux-fsdevel@vger.kernel.org Cc: cluster-devel@redhat.com Cc: linux-mm@kvack.org Acked-by: Geoff Levand <geoff@infradead.org>
2013-11-24 01:19:00 +00:00
b_offset += bvl.bv_offset;
bio_page = bvl.bv_page;
if (frombio) {
if (conf->skip_copy &&
b_offset == 0 && page_offset == 0 &&
clen == RAID5_STRIPE_SIZE(conf) &&
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
!no_skipcopy)
*page = bio_page;
else
tx = async_memcpy(*page, bio_page, page_offset,
b_offset, clen, &submit);
} else
tx = async_memcpy(bio_page, *page, b_offset,
page_offset, clen, &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
/* chain the operations */
submit.depend_tx = tx;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (clen < len) /* hit end of page */
break;
page_offset += len;
}
return tx;
}
static void ops_complete_biofill(void *stripe_head_ref)
{
struct stripe_head *sh = stripe_head_ref;
int i;
struct r5conf *conf = sh->raid_conf;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
/* clear completed biofills */
for (i = sh->disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
/* acknowledge completion of a biofill operation */
/* and check if we need to reply to a read request,
* new R5_Wantfill requests are held off until
* !STRIPE_BIOFILL_RUN
*/
if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct bio *rbi, *rbi2;
BUG_ON(!dev->read);
rbi = dev->read;
dev->read = NULL;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (rbi && rbi->bi_iter.bi_sector <
dev->sector + RAID5_STRIPE_SECTORS(conf)) {
rbi2 = r5_next_bio(conf, rbi, dev->sector);
bio_endio(rbi);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
rbi = rbi2;
}
}
}
clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void ops_run_biofill(struct stripe_head *sh)
{
struct dma_async_tx_descriptor *tx = NULL;
struct async_submit_ctl submit;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int i;
struct r5conf *conf = sh->raid_conf;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
for (i = sh->disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (test_bit(R5_Wantfill, &dev->flags)) {
struct bio *rbi;
spin_lock_irq(&sh->stripe_lock);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
dev->read = rbi = dev->toread;
dev->toread = NULL;
spin_unlock_irq(&sh->stripe_lock);
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (rbi && rbi->bi_iter.bi_sector <
dev->sector + RAID5_STRIPE_SECTORS(conf)) {
tx = async_copy_data(0, rbi, &dev->page,
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
dev->sector, tx, sh, 0);
rbi = r5_next_bio(conf, rbi, dev->sector);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
}
}
atomic_inc(&sh->count);
init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
async_trigger_callback(&submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void mark_target_uptodate(struct stripe_head *sh, int target)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
struct r5dev *tgt;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (target < 0)
return;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
tgt = &sh->dev[target];
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(R5_UPTODATE, &tgt->flags);
BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
clear_bit(R5_Wantcompute, &tgt->flags);
}
static void ops_complete_compute(void *stripe_head_ref)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
struct stripe_head *sh = stripe_head_ref;
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
/* mark the computed target(s) as uptodate */
mark_target_uptodate(sh, sh->ops.target);
mark_target_uptodate(sh, sh->ops.target2);
clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
if (sh->check_state == check_state_compute_run)
sh->check_state = check_state_compute_result;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
/* return a pointer to the address conversion region of the scribble buffer */
static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
{
return percpu->scribble + i * percpu->scribble_obj_size;
}
/* return a pointer to the address conversion region of the scribble buffer */
static addr_conv_t *to_addr_conv(struct stripe_head *sh,
struct raid5_percpu *percpu, int i)
{
return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
}
static struct dma_async_tx_descriptor *
ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
int disks = sh->disks;
struct page **xor_srcs = to_addr_page(percpu, 0);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int target = sh->ops.target;
struct r5dev *tgt = &sh->dev[target];
struct page *xor_dest = tgt->page;
int count = 0;
struct dma_async_tx_descriptor *tx;
struct async_submit_ctl submit;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int i;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu block: %d\n",
__func__, (unsigned long long)sh->sector, target);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
for (i = disks; i--; )
if (i != target)
xor_srcs[count++] = sh->dev[i].page;
atomic_inc(&sh->count);
init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (unlikely(count == 1))
tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
else
tx = async_xor(xor_dest, xor_srcs, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
return tx;
}
/* set_syndrome_sources - populate source buffers for gen_syndrome
* @srcs - (struct page *) array of size sh->disks
* @sh - stripe_head to parse
*
* Populates srcs in proper layout order for the stripe and returns the
* 'count' of sources to be used in a call to async_gen_syndrome. The P
* destination buffer is recorded in srcs[count] and the Q destination
* is recorded in srcs[count+1]].
*/
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
static int set_syndrome_sources(struct page **srcs,
struct stripe_head *sh,
int srctype)
{
int disks = sh->disks;
int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
int d0_idx = raid6_d0(sh);
int count;
int i;
for (i = 0; i < disks; i++)
srcs[i] = NULL;
count = 0;
i = d0_idx;
do {
int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
struct r5dev *dev = &sh->dev[i];
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (i == sh->qd_idx || i == sh->pd_idx ||
(srctype == SYNDROME_SRC_ALL) ||
(srctype == SYNDROME_SRC_WANT_DRAIN &&
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
(test_bit(R5_Wantdrain, &dev->flags) ||
test_bit(R5_InJournal, &dev->flags))) ||
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
(srctype == SYNDROME_SRC_WRITTEN &&
(dev->written ||
test_bit(R5_InJournal, &dev->flags)))) {
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (test_bit(R5_InJournal, &dev->flags))
srcs[slot] = sh->dev[i].orig_page;
else
srcs[slot] = sh->dev[i].page;
}
i = raid6_next_disk(i, disks);
} while (i != d0_idx);
return syndrome_disks;
}
static struct dma_async_tx_descriptor *
ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
{
int disks = sh->disks;
struct page **blocks = to_addr_page(percpu, 0);
int target;
int qd_idx = sh->qd_idx;
struct dma_async_tx_descriptor *tx;
struct async_submit_ctl submit;
struct r5dev *tgt;
struct page *dest;
int i;
int count;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
if (sh->ops.target < 0)
target = sh->ops.target2;
else if (sh->ops.target2 < 0)
target = sh->ops.target;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
else
/* we should only have one valid target */
BUG();
BUG_ON(target < 0);
pr_debug("%s: stripe %llu block: %d\n",
__func__, (unsigned long long)sh->sector, target);
tgt = &sh->dev[target];
BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
dest = tgt->page;
atomic_inc(&sh->count);
if (target == qd_idx) {
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
blocks[count] = NULL; /* regenerating p is not necessary */
BUG_ON(blocks[count+1] != dest); /* q should already be set */
init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
ops_complete_compute, sh,
to_addr_conv(sh, percpu, 0));
tx = async_gen_syndrome(blocks, 0, count+2,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
} else {
/* Compute any data- or p-drive using XOR */
count = 0;
for (i = disks; i-- ; ) {
if (i == target || i == qd_idx)
continue;
blocks[count++] = sh->dev[i].page;
}
init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
NULL, ops_complete_compute, sh,
to_addr_conv(sh, percpu, 0));
tx = async_xor(dest, blocks, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
return tx;
}
static struct dma_async_tx_descriptor *
ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
{
int i, count, disks = sh->disks;
int syndrome_disks = sh->ddf_layout ? disks : disks-2;
int d0_idx = raid6_d0(sh);
int faila = -1, failb = -1;
int target = sh->ops.target;
int target2 = sh->ops.target2;
struct r5dev *tgt = &sh->dev[target];
struct r5dev *tgt2 = &sh->dev[target2];
struct dma_async_tx_descriptor *tx;
struct page **blocks = to_addr_page(percpu, 0);
struct async_submit_ctl submit;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
pr_debug("%s: stripe %llu block1: %d block2: %d\n",
__func__, (unsigned long long)sh->sector, target, target2);
BUG_ON(target < 0 || target2 < 0);
BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
/* we need to open-code set_syndrome_sources to handle the
* slot number conversion for 'faila' and 'failb'
*/
for (i = 0; i < disks ; i++)
blocks[i] = NULL;
count = 0;
i = d0_idx;
do {
int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
blocks[slot] = sh->dev[i].page;
if (i == target)
faila = slot;
if (i == target2)
failb = slot;
i = raid6_next_disk(i, disks);
} while (i != d0_idx);
BUG_ON(faila == failb);
if (failb < faila)
swap(faila, failb);
pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
__func__, (unsigned long long)sh->sector, faila, failb);
atomic_inc(&sh->count);
if (failb == syndrome_disks+1) {
/* Q disk is one of the missing disks */
if (faila == syndrome_disks) {
/* Missing P+Q, just recompute */
init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
ops_complete_compute, sh,
to_addr_conv(sh, percpu, 0));
return async_gen_syndrome(blocks, 0, syndrome_disks+2,
RAID5_STRIPE_SIZE(sh->raid_conf),
&submit);
} else {
struct page *dest;
int data_target;
int qd_idx = sh->qd_idx;
/* Missing D+Q: recompute D from P, then recompute Q */
if (target == qd_idx)
data_target = target2;
else
data_target = target;
count = 0;
for (i = disks; i-- ; ) {
if (i == data_target || i == qd_idx)
continue;
blocks[count++] = sh->dev[i].page;
}
dest = sh->dev[data_target].page;
init_async_submit(&submit,
ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
NULL, NULL, NULL,
to_addr_conv(sh, percpu, 0));
tx = async_xor(dest, blocks, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf),
&submit);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
init_async_submit(&submit, ASYNC_TX_FENCE, tx,
ops_complete_compute, sh,
to_addr_conv(sh, percpu, 0));
return async_gen_syndrome(blocks, 0, count+2,
RAID5_STRIPE_SIZE(sh->raid_conf),
&submit);
}
} else {
init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
ops_complete_compute, sh,
to_addr_conv(sh, percpu, 0));
if (failb == syndrome_disks) {
/* We're missing D+P. */
return async_raid6_datap_recov(syndrome_disks+2,
RAID5_STRIPE_SIZE(sh->raid_conf),
faila,
blocks, &submit);
} else {
/* We're missing D+D. */
return async_raid6_2data_recov(syndrome_disks+2,
RAID5_STRIPE_SIZE(sh->raid_conf),
faila, failb,
blocks, &submit);
}
}
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
static void ops_complete_prexor(void *stripe_head_ref)
{
struct stripe_head *sh = stripe_head_ref;
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (r5c_is_writeback(sh->raid_conf->log))
/*
* raid5-cache write back uses orig_page during prexor.
* After prexor, it is time to free orig_page
*/
r5c_release_extra_page(sh);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static struct dma_async_tx_descriptor *
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
struct dma_async_tx_descriptor *tx)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
int disks = sh->disks;
struct page **xor_srcs = to_addr_page(percpu, 0);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int count = 0, pd_idx = sh->pd_idx, i;
struct async_submit_ctl submit;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
/* existing parity data subtracted */
struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
/* Only process blocks that are known to be uptodate */
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (test_bit(R5_InJournal, &dev->flags))
xor_srcs[count++] = dev->orig_page;
else if (test_bit(R5_Wantdrain, &dev->flags))
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
xor_srcs[count++] = dev->page;
}
init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
tx = async_xor(xor_dest, xor_srcs, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
return tx;
}
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
static struct dma_async_tx_descriptor *
ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
struct dma_async_tx_descriptor *tx)
{
struct page **blocks = to_addr_page(percpu, 0);
int count;
struct async_submit_ctl submit;
pr_debug("%s: stripe %llu\n", __func__,
(unsigned long long)sh->sector);
count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
tx = async_gen_syndrome(blocks, 0, count+2,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
return tx;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
static struct dma_async_tx_descriptor *
ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
struct r5conf *conf = sh->raid_conf;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
int disks = sh->disks;
int i;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct stripe_head *head_sh = sh;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
for (i = disks; i--; ) {
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct r5dev *dev;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct bio *chosen;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
sh = head_sh;
if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct bio *wbi;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
again:
dev = &sh->dev[i];
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
/*
* clear R5_InJournal, so when rewriting a page in
* journal, it is not skipped by r5l_log_stripe()
*/
clear_bit(R5_InJournal, &dev->flags);
spin_lock_irq(&sh->stripe_lock);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
chosen = dev->towrite;
dev->towrite = NULL;
sh->overwrite_disks = 0;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
BUG_ON(dev->written);
wbi = dev->written = chosen;
spin_unlock_irq(&sh->stripe_lock);
WARN_ON(dev->page != dev->orig_page);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (wbi && wbi->bi_iter.bi_sector <
dev->sector + RAID5_STRIPE_SECTORS(conf)) {
if (wbi->bi_opf & REQ_FUA)
2010-09-03 09:56:18 +00:00
set_bit(R5_WantFUA, &dev->flags);
if (wbi->bi_opf & REQ_SYNC)
set_bit(R5_SyncIO, &dev->flags);
if (bio_op(wbi) == REQ_OP_DISCARD)
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
set_bit(R5_Discard, &dev->flags);
else {
tx = async_copy_data(1, wbi, &dev->page,
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
dev->sector, tx, sh,
r5c_is_writeback(conf->log));
if (dev->page != dev->orig_page &&
!r5c_is_writeback(conf->log)) {
set_bit(R5_SkipCopy, &dev->flags);
clear_bit(R5_UPTODATE, &dev->flags);
clear_bit(R5_OVERWRITE, &dev->flags);
}
}
wbi = r5_next_bio(conf, wbi, dev->sector);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head_sh->batch_head) {
sh = list_first_entry(&sh->batch_list,
struct stripe_head,
batch_list);
if (sh == head_sh)
continue;
goto again;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
}
return tx;
}
static void ops_complete_reconstruct(void *stripe_head_ref)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
struct stripe_head *sh = stripe_head_ref;
int disks = sh->disks;
int pd_idx = sh->pd_idx;
int qd_idx = sh->qd_idx;
int i;
bool fua = false, sync = false, discard = false;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
for (i = disks; i--; ) {
2010-09-03 09:56:18 +00:00
fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
discard |= test_bit(R5_Discard, &sh->dev[i].flags);
}
2010-09-03 09:56:18 +00:00
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
2010-09-03 09:56:18 +00:00
if (dev->written || i == pd_idx || i == qd_idx) {
if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
set_bit(R5_UPTODATE, &dev->flags);
if (test_bit(STRIPE_EXPAND_READY, &sh->state))
set_bit(R5_Expanded, &dev->flags);
}
2010-09-03 09:56:18 +00:00
if (fua)
set_bit(R5_WantFUA, &dev->flags);
if (sync)
set_bit(R5_SyncIO, &dev->flags);
2010-09-03 09:56:18 +00:00
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
if (sh->reconstruct_state == reconstruct_state_drain_run)
sh->reconstruct_state = reconstruct_state_drain_result;
else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
sh->reconstruct_state = reconstruct_state_prexor_drain_result;
else {
BUG_ON(sh->reconstruct_state != reconstruct_state_run);
sh->reconstruct_state = reconstruct_state_result;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void
ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
struct dma_async_tx_descriptor *tx)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
int disks = sh->disks;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct page **xor_srcs;
struct async_submit_ctl submit;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
int count, pd_idx = sh->pd_idx, i;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct page *xor_dest;
int prexor = 0;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
unsigned long flags;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
int j = 0;
struct stripe_head *head_sh = sh;
int last_stripe;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
for (i = 0; i < sh->disks; i++) {
if (pd_idx == i)
continue;
if (!test_bit(R5_Discard, &sh->dev[i].flags))
break;
}
if (i >= sh->disks) {
atomic_inc(&sh->count);
set_bit(R5_Discard, &sh->dev[pd_idx].flags);
ops_complete_reconstruct(sh);
return;
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
again:
count = 0;
xor_srcs = to_addr_page(percpu, j);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
/* check if prexor is active which means only process blocks
* that are part of a read-modify-write (written)
*/
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
prexor = 1;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (head_sh->dev[i].written ||
test_bit(R5_InJournal, &head_sh->dev[i].flags))
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
xor_srcs[count++] = dev->page;
}
} else {
xor_dest = sh->dev[pd_idx].page;
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (i != pd_idx)
xor_srcs[count++] = dev->page;
}
}
/* 1/ if we prexor'd then the dest is reused as a source
* 2/ if we did not prexor then we are redoing the parity
* set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
* for the synchronous xor case
*/
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
last_stripe = !head_sh->batch_head ||
list_first_entry(&sh->batch_list,
struct stripe_head, batch_list) == head_sh;
if (last_stripe) {
flags = ASYNC_TX_ACK |
(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
atomic_inc(&head_sh->count);
init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
to_addr_conv(sh, percpu, j));
} else {
flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
init_async_submit(&submit, flags, tx, NULL, NULL,
to_addr_conv(sh, percpu, j));
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (unlikely(count == 1))
tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
else
tx = async_xor(xor_dest, xor_srcs, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (!last_stripe) {
j++;
sh = list_first_entry(&sh->batch_list, struct stripe_head,
batch_list);
goto again;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void
ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
struct dma_async_tx_descriptor *tx)
{
struct async_submit_ctl submit;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct page **blocks;
int count, i, j = 0;
struct stripe_head *head_sh = sh;
int last_stripe;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
int synflags;
unsigned long txflags;
pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
for (i = 0; i < sh->disks; i++) {
if (sh->pd_idx == i || sh->qd_idx == i)
continue;
if (!test_bit(R5_Discard, &sh->dev[i].flags))
break;
}
if (i >= sh->disks) {
atomic_inc(&sh->count);
set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
ops_complete_reconstruct(sh);
return;
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
again:
blocks = to_addr_page(percpu, j);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
synflags = SYNDROME_SRC_WRITTEN;
txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
} else {
synflags = SYNDROME_SRC_ALL;
txflags = ASYNC_TX_ACK;
}
count = set_syndrome_sources(blocks, sh, synflags);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
last_stripe = !head_sh->batch_head ||
list_first_entry(&sh->batch_list,
struct stripe_head, batch_list) == head_sh;
if (last_stripe) {
atomic_inc(&head_sh->count);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
head_sh, to_addr_conv(sh, percpu, j));
} else
init_async_submit(&submit, 0, tx, NULL, NULL,
to_addr_conv(sh, percpu, j));
tx = async_gen_syndrome(blocks, 0, count+2,
RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (!last_stripe) {
j++;
sh = list_first_entry(&sh->batch_list, struct stripe_head,
batch_list);
goto again;
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void ops_complete_check(void *stripe_head_ref)
{
struct stripe_head *sh = stripe_head_ref;
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
sh->check_state = check_state_check_result;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
int disks = sh->disks;
int pd_idx = sh->pd_idx;
int qd_idx = sh->qd_idx;
struct page *xor_dest;
struct page **xor_srcs = to_addr_page(percpu, 0);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
struct dma_async_tx_descriptor *tx;
struct async_submit_ctl submit;
int count;
int i;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
pr_debug("%s: stripe %llu\n", __func__,
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
(unsigned long long)sh->sector);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
count = 0;
xor_dest = sh->dev[pd_idx].page;
xor_srcs[count++] = xor_dest;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
for (i = disks; i--; ) {
if (i == pd_idx || i == qd_idx)
continue;
xor_srcs[count++] = sh->dev[i].page;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
init_async_submit(&submit, 0, NULL, NULL, NULL,
to_addr_conv(sh, percpu, 0));
tx = async_xor_val(xor_dest, xor_srcs, 0, count,
RAID5_STRIPE_SIZE(sh->raid_conf),
&sh->ops.zero_sum_result, &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
atomic_inc(&sh->count);
init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
tx = async_trigger_callback(&submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
{
struct page **srcs = to_addr_page(percpu, 0);
struct async_submit_ctl submit;
int count;
pr_debug("%s: stripe %llu checkp: %d\n", __func__,
(unsigned long long)sh->sector, checkp);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
if (!checkp)
srcs[count] = NULL;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
atomic_inc(&sh->count);
init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
sh, to_addr_conv(sh, percpu, 0));
async_syndrome_val(srcs, 0, count+2,
RAID5_STRIPE_SIZE(sh->raid_conf),
&sh->ops.zero_sum_result, percpu->spare_page, &submit);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
{
int overlap_clear = 0, i, disks = sh->disks;
struct dma_async_tx_descriptor *tx = NULL;
struct r5conf *conf = sh->raid_conf;
int level = conf->level;
struct raid5_percpu *percpu;
unsigned long cpu;
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
cpu = get_cpu();
percpu = per_cpu_ptr(conf->percpu, cpu);
if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
ops_run_biofill(sh);
overlap_clear++;
}
if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
if (level < 6)
tx = ops_run_compute5(sh, percpu);
else {
if (sh->ops.target2 < 0 || sh->ops.target < 0)
tx = ops_run_compute6_1(sh, percpu);
else
tx = ops_run_compute6_2(sh, percpu);
}
/* terminate the chain if reconstruct is not set to be run */
if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
async_tx_ack(tx);
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
if (level < 6)
tx = ops_run_prexor5(sh, percpu, tx);
else
tx = ops_run_prexor6(sh, percpu, tx);
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
tx = ops_run_partial_parity(sh, percpu, tx);
if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
tx = ops_run_biodrain(sh, tx);
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
overlap_clear++;
}
if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
if (level < 6)
ops_run_reconstruct5(sh, percpu, tx);
else
ops_run_reconstruct6(sh, percpu, tx);
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
if (sh->check_state == check_state_run)
ops_run_check_p(sh, percpu);
else if (sh->check_state == check_state_run_q)
ops_run_check_pq(sh, percpu, 0);
else if (sh->check_state == check_state_run_pq)
ops_run_check_pq(sh, percpu, 1);
else
BUG();
}
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (overlap_clear && !sh->batch_head)
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (test_and_clear_bit(R5_Overlap, &dev->flags))
wake_up(&sh->raid_conf->wait_for_overlap);
}
put_cpu();
md: raid5_run_ops - run stripe operations outside sh->lock When the raid acceleration work was proposed, Neil laid out the following attack plan: 1/ move the xor and copy operations outside spin_lock(&sh->lock) 2/ find/implement an asynchronous offload api The raid5_run_ops routine uses the asynchronous offload api (async_tx) and the stripe_operations member of a stripe_head to carry out xor+copy operations asynchronously, outside the lock. To perform operations outside the lock a new set of state flags is needed to track new requests, in-flight requests, and completed requests. In this new model handle_stripe is tasked with scanning the stripe_head for work, updating the stripe_operations structure, and finally dropping the lock and calling raid5_run_ops for processing. The following flags outline the requests that handle_stripe can make of raid5_run_ops: STRIPE_OP_BIOFILL - copy data into request buffers to satisfy a read request STRIPE_OP_COMPUTE_BLK - generate a missing block in the cache from the other blocks STRIPE_OP_PREXOR - subtract existing data as part of the read-modify-write process STRIPE_OP_BIODRAIN - copy data out of request buffers to satisfy a write request STRIPE_OP_POSTXOR - recalculate parity for new data that has entered the cache STRIPE_OP_CHECK - verify that the parity is correct STRIPE_OP_IO - submit i/o to the member disks (note this was already performed outside the stripe lock, but it made sense to add it as an operation type The flow is: 1/ handle_stripe sets STRIPE_OP_* in sh->ops.pending 2/ raid5_run_ops reads sh->ops.pending, sets sh->ops.ack, and submits the operation to the async_tx api 3/ async_tx triggers the completion callback routine to set sh->ops.complete and release the stripe 4/ handle_stripe runs again to finish the operation and optionally submit new operations that were previously blocked Note this patch just defines raid5_run_ops, subsequent commits (one per major operation type) modify handle_stripe to take advantage of this routine. Changelog: * removed ops_complete_biodrain in favor of ops_complete_postxor and ops_complete_write. * removed the raid5_run_ops workqueue * call bi_end_io for reads in ops_complete_biofill, saves a call to handle_stripe * explicitly handle the 2-disk raid5 case (xor becomes memcpy), Neil Brown * fix race between async engines and bi_end_io call for reads, Neil Brown * remove unnecessary spin_lock from ops_complete_biofill * remove test_and_set/test_and_clear BUG_ONs, Neil Brown * remove explicit interrupt handling for channel switching, this feature was absorbed (i.e. it is now implicit) by the async_tx api * use return_io in ops_complete_biofill Signed-off-by: Dan Williams <dan.j.williams@intel.com> Acked-By: NeilBrown <neilb@suse.de>
2007-01-02 20:52:30 +00:00
}
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
{
if (sh->ppl_page)
__free_page(sh->ppl_page);
kmem_cache_free(sc, sh);
}
static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
int disks, struct r5conf *conf)
{
struct stripe_head *sh;
int i;
sh = kmem_cache_zalloc(sc, gfp);
if (sh) {
spin_lock_init(&sh->stripe_lock);
spin_lock_init(&sh->batch_lock);
INIT_LIST_HEAD(&sh->batch_list);
INIT_LIST_HEAD(&sh->lru);
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
INIT_LIST_HEAD(&sh->r5c);
INIT_LIST_HEAD(&sh->log_list);
atomic_set(&sh->count, 1);
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
sh->raid_conf = conf;
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
sh->log_start = MaxSector;
for (i = 0; i < disks; i++) {
struct r5dev *dev = &sh->dev[i];
bio_init(&dev->req, &dev->vec, 1);
bio_init(&dev->rreq, &dev->rvec, 1);
}
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (raid5_has_ppl(conf)) {
sh->ppl_page = alloc_page(gfp);
if (!sh->ppl_page) {
free_stripe(sc, sh);
sh = NULL;
}
}
}
return sh;
}
static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
{
struct stripe_head *sh;
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
if (!sh)
return 0;
if (grow_buffers(sh, gfp)) {
shrink_buffers(sh);
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
free_stripe(conf->slab_cache, sh);
return 0;
}
sh->hash_lock_index =
conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
/* we just created an active stripe so... */
atomic_inc(&conf->active_stripes);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
raid5_release_stripe(sh);
conf->max_nr_stripes++;
return 1;
}
static int grow_stripes(struct r5conf *conf, int num)
{
struct kmem_cache *sc;
size_t namelen = sizeof(conf->cache_name[0]);
int devs = max(conf->raid_disks, conf->previous_raid_disks);
if (conf->mddev->gendisk)
snprintf(conf->cache_name[0], namelen,
"raid%d-%s", conf->level, mdname(conf->mddev));
else
snprintf(conf->cache_name[0], namelen,
"raid%d-%p", conf->level, conf->mddev);
snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
conf->active_name = 0;
sc = kmem_cache_create(conf->cache_name[conf->active_name],
sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
0, 0, NULL);
if (!sc)
return 1;
conf->slab_cache = sc;
conf->pool_size = devs;
while (num--)
if (!grow_one_stripe(conf, GFP_KERNEL))
return 1;
return 0;
}
/**
* scribble_alloc - allocate percpu scribble buffer for required size
* of the scribble region
* @percpu: from for_each_present_cpu() of the caller
* @num: total number of disks in the array
* @cnt: scribble objs count for required size of the scribble region
*
* The scribble buffer size must be enough to contain:
* 1/ a struct page pointer for each device in the array +2
* 2/ room to convert each entry in (1) to its corresponding dma
* (dma_map_page()) or page (page_address()) address.
*
* Note: the +2 is for the destination buffers of the ddf/raid6 case where we
* calculate over all devices (not just the data blocks), using zeros in place
* of the P and Q blocks.
*/
static int scribble_alloc(struct raid5_percpu *percpu,
int num, int cnt)
{
size_t obj_size =
sizeof(struct page *) * (num+2) +
sizeof(addr_conv_t) * (num+2);
void *scribble;
/*
* If here is in raid array suspend context, it is in memalloc noio
* context as well, there is no potential recursive memory reclaim
* I/Os with the GFP_KERNEL flag.
*/
scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
if (!scribble)
return -ENOMEM;
kvfree(percpu->scribble);
percpu->scribble = scribble;
percpu->scribble_obj_size = obj_size;
return 0;
}
static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
{
unsigned long cpu;
int err = 0;
/*
* Never shrink. And mddev_suspend() could deadlock if this is called
* from raid5d. In that case, scribble_disks and scribble_sectors
* should equal to new_disks and new_sectors
*/
if (conf->scribble_disks >= new_disks &&
conf->scribble_sectors >= new_sectors)
return 0;
mddev_suspend(conf->mddev);
get_online_cpus();
for_each_present_cpu(cpu) {
struct raid5_percpu *percpu;
percpu = per_cpu_ptr(conf->percpu, cpu);
err = scribble_alloc(percpu, new_disks,
new_sectors / RAID5_STRIPE_SECTORS(conf));
if (err)
break;
}
put_online_cpus();
mddev_resume(conf->mddev);
if (!err) {
conf->scribble_disks = new_disks;
conf->scribble_sectors = new_sectors;
}
return err;
}
static int resize_stripes(struct r5conf *conf, int newsize)
{
/* Make all the stripes able to hold 'newsize' devices.
* New slots in each stripe get 'page' set to a new page.
*
* This happens in stages:
* 1/ create a new kmem_cache and allocate the required number of
* stripe_heads.
* 2/ gather all the old stripe_heads and transfer the pages across
* to the new stripe_heads. This will have the side effect of
* freezing the array as once all stripe_heads have been collected,
* no IO will be possible. Old stripe heads are freed once their
* pages have been transferred over, and the old kmem_cache is
* freed when all stripes are done.
* 3/ reallocate conf->disks to be suitable bigger. If this fails,
* we simple return a failure status - no need to clean anything up.
* 4/ allocate new pages for the new slots in the new stripe_heads.
* If this fails, we don't bother trying the shrink the
* stripe_heads down again, we just leave them as they are.
* As each stripe_head is processed the new one is released into
* active service.
*
* Once step2 is started, we cannot afford to wait for a write,
* so we use GFP_NOIO allocations.
*/
struct stripe_head *osh, *nsh;
LIST_HEAD(newstripes);
struct disk_info *ndisks;
int err = 0;
struct kmem_cache *sc;
int i;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int hash, cnt;
md_allow_write(conf->mddev);
/* Step 1 */
sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
0, 0, NULL);
if (!sc)
return -ENOMEM;
/* Need to ensure auto-resizing doesn't interfere */
mutex_lock(&conf->cache_size_mutex);
for (i = conf->max_nr_stripes; i; i--) {
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
if (!nsh)
break;
list_add(&nsh->lru, &newstripes);
}
if (i) {
/* didn't get enough, give up */
while (!list_empty(&newstripes)) {
nsh = list_entry(newstripes.next, struct stripe_head, lru);
list_del(&nsh->lru);
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
free_stripe(sc, nsh);
}
kmem_cache_destroy(sc);
mutex_unlock(&conf->cache_size_mutex);
return -ENOMEM;
}
/* Step 2 - Must use GFP_NOIO now.
* OK, we have enough stripes, start collecting inactive
* stripes and copying them over
*/
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
hash = 0;
cnt = 0;
list_for_each_entry(nsh, &newstripes, lru) {
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
lock_device_hash_lock(conf, hash);
wait_event_cmd(conf->wait_for_stripe,
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
!list_empty(conf->inactive_list + hash),
unlock_device_hash_lock(conf, hash),
lock_device_hash_lock(conf, hash));
osh = get_free_stripe(conf, hash);
unlock_device_hash_lock(conf, hash);
for(i=0; i<conf->pool_size; i++) {
nsh->dev[i].page = osh->dev[i].page;
nsh->dev[i].orig_page = osh->dev[i].page;
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
nsh->hash_lock_index = hash;
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
free_stripe(conf->slab_cache, osh);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
cnt++;
if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
!!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
hash++;
cnt = 0;
}
}
kmem_cache_destroy(conf->slab_cache);
/* Step 3.
* At this point, we are holding all the stripes so the array
* is completely stalled, so now is a good time to resize
* conf->disks and the scribble region
*/
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
if (ndisks) {
for (i = 0; i < conf->pool_size; i++)
ndisks[i] = conf->disks[i];
for (i = conf->pool_size; i < newsize; i++) {
ndisks[i].extra_page = alloc_page(GFP_NOIO);
if (!ndisks[i].extra_page)
err = -ENOMEM;
}
if (err) {
for (i = conf->pool_size; i < newsize; i++)
if (ndisks[i].extra_page)
put_page(ndisks[i].extra_page);
kfree(ndisks);
} else {
kfree(conf->disks);
conf->disks = ndisks;
}
} else
err = -ENOMEM;
mutex_unlock(&conf->cache_size_mutex);
md: update slab_cache before releasing new stripes when stripes resizing When growing raid5 device on machine with small memory, there is chance that mdadm will be killed and the following bug report can be observed. The same bug could also be reproduced in linux-4.10.6. [57600.075774] BUG: unable to handle kernel NULL pointer dereference at (null) [57600.083796] IP: [<ffffffff81a6aa87>] _raw_spin_lock+0x7/0x20 [57600.110378] PGD 421cf067 PUD 4442d067 PMD 0 [57600.114678] Oops: 0002 [#1] SMP [57600.180799] CPU: 1 PID: 25990 Comm: mdadm Tainted: P O 4.2.8 #1 [57600.187849] Hardware name: To be filled by O.E.M. To be filled by O.E.M./MAHOBAY, BIOS QV05AR66 03/06/2013 [57600.197490] task: ffff880044e47240 ti: ffff880043070000 task.ti: ffff880043070000 [57600.204963] RIP: 0010:[<ffffffff81a6aa87>] [<ffffffff81a6aa87>] _raw_spin_lock+0x7/0x20 [57600.213057] RSP: 0018:ffff880043073810 EFLAGS: 00010046 [57600.218359] RAX: 0000000000000000 RBX: 000000000000000c RCX: ffff88011e296dd0 [57600.225486] RDX: 0000000000000001 RSI: ffffe8ffffcb46c0 RDI: 0000000000000000 [57600.232613] RBP: ffff880043073878 R08: ffff88011e5f8170 R09: 0000000000000282 [57600.239739] R10: 0000000000000005 R11: 28f5c28f5c28f5c3 R12: ffff880043073838 [57600.246872] R13: ffffe8ffffcb46c0 R14: 0000000000000000 R15: ffff8800b9706a00 [57600.253999] FS: 00007f576106c700(0000) GS:ffff88011e280000(0000) knlGS:0000000000000000 [57600.262078] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [57600.267817] CR2: 0000000000000000 CR3: 00000000428fe000 CR4: 00000000001406e0 [57600.274942] Stack: [57600.276949] ffffffff8114ee35 ffff880043073868 0000000000000282 000000000000eb3f [57600.284383] ffffffff81119043 ffff880043073838 ffff880043073838 ffff88003e197b98 [57600.291820] ffffe8ffffcb46c0 ffff88003e197360 0000000000000286 ffff880043073968 [57600.299254] Call Trace: [57600.301698] [<ffffffff8114ee35>] ? cache_flusharray+0x35/0xe0 [57600.307523] [<ffffffff81119043>] ? __page_cache_release+0x23/0x110 [57600.313779] [<ffffffff8114eb53>] kmem_cache_free+0x63/0xc0 [57600.319344] [<ffffffff81579942>] drop_one_stripe+0x62/0x90 [57600.324915] [<ffffffff81579b5b>] raid5_cache_scan+0x8b/0xb0 [57600.330563] [<ffffffff8111b98a>] shrink_slab.part.36+0x19a/0x250 [57600.336650] [<ffffffff8111e38c>] shrink_zone+0x23c/0x250 [57600.342039] [<ffffffff8111e4f3>] do_try_to_free_pages+0x153/0x420 [57600.348210] [<ffffffff8111e851>] try_to_free_pages+0x91/0xa0 [57600.353959] [<ffffffff811145b1>] __alloc_pages_nodemask+0x4d1/0x8b0 [57600.360303] [<ffffffff8157a30b>] check_reshape+0x62b/0x770 [57600.365866] [<ffffffff8157a4a5>] raid5_check_reshape+0x55/0xa0 [57600.371778] [<ffffffff81583df7>] update_raid_disks+0xc7/0x110 [57600.377604] [<ffffffff81592b73>] md_ioctl+0xd83/0x1b10 [57600.382827] [<ffffffff81385380>] blkdev_ioctl+0x170/0x690 [57600.388307] [<ffffffff81195238>] block_ioctl+0x38/0x40 [57600.393525] [<ffffffff811731c5>] do_vfs_ioctl+0x2b5/0x480 [57600.399010] [<ffffffff8115e07b>] ? vfs_write+0x14b/0x1f0 [57600.404400] [<ffffffff811733cc>] SyS_ioctl+0x3c/0x70 [57600.409447] [<ffffffff81a6ad97>] entry_SYSCALL_64_fastpath+0x12/0x6a [57600.415875] Code: 00 00 00 00 55 48 89 e5 8b 07 85 c0 74 04 31 c0 5d c3 ba 01 00 00 00 f0 0f b1 17 85 c0 75 ef b0 01 5d c3 90 31 c0 ba 01 00 00 00 <f0> 0f b1 17 85 c0 75 01 c3 55 89 c6 48 89 e5 e8 85 d1 63 ff 5d [57600.435460] RIP [<ffffffff81a6aa87>] _raw_spin_lock+0x7/0x20 [57600.441208] RSP <ffff880043073810> [57600.444690] CR2: 0000000000000000 [57600.448000] ---[ end trace cbc6b5cc4bf9831d ]--- The problem is that resize_stripes() releases new stripe_heads before assigning new slab cache to conf->slab_cache. If the shrinker function raid5_cache_scan() gets called after resize_stripes() starting releasing new stripes but right before new slab cache being assigned, it is possible that these new stripe_heads will be freed with the old slab_cache which was already been destoryed and that triggers this bug. Signed-off-by: Dennis Yang <dennisyang@qnap.com> Fixes: edbe83ab4c27 ("md/raid5: allow the stripe_cache to grow and shrink.") Cc: stable@vger.kernel.org (4.1+) Reviewed-by: NeilBrown <neilb@suse.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-03-29 07:46:13 +00:00
conf->slab_cache = sc;
conf->active_name = 1-conf->active_name;
/* Step 4, return new stripes to service */
while(!list_empty(&newstripes)) {
nsh = list_entry(newstripes.next, struct stripe_head, lru);
list_del_init(&nsh->lru);
for (i=conf->raid_disks; i < newsize; i++)
if (nsh->dev[i].page == NULL) {
struct page *p = alloc_page(GFP_NOIO);
nsh->dev[i].page = p;
nsh->dev[i].orig_page = p;
if (!p)
err = -ENOMEM;
}
raid5_release_stripe(nsh);
}
/* critical section pass, GFP_NOIO no longer needed */
if (!err)
conf->pool_size = newsize;
return err;
}
static int drop_one_stripe(struct r5conf *conf)
{
struct stripe_head *sh;
int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
spin_lock_irq(conf->hash_locks + hash);
sh = get_free_stripe(conf, hash);
spin_unlock_irq(conf->hash_locks + hash);
if (!sh)
return 0;
BUG_ON(atomic_read(&sh->count));
shrink_buffers(sh);
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
free_stripe(conf->slab_cache, sh);
atomic_dec(&conf->active_stripes);
conf->max_nr_stripes--;
return 1;
}
static void shrink_stripes(struct r5conf *conf)
{
while (conf->max_nr_stripes &&
drop_one_stripe(conf))
;
kmem_cache_destroy(conf->slab_cache);
conf->slab_cache = NULL;
}
static void raid5_end_read_request(struct bio * bi)
{
struct stripe_head *sh = bi->bi_private;
struct r5conf *conf = sh->raid_conf;
int disks = sh->disks, i;
char b[BDEVNAME_SIZE];
struct md_rdev *rdev = NULL;
sector_t s;
for (i=0 ; i<disks; i++)
if (bi == &sh->dev[i].req)
break;
pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
(unsigned long long)sh->sector, i, atomic_read(&sh->count),
bi->bi_status);
if (i == disks) {
bio_reset(bi);
BUG();
return;
}
if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
/* If replacement finished while this request was outstanding,
* 'replacement' might be NULL already.
* In that case it moved down to 'rdev'.
* rdev is not removed until all requests are finished.
*/
rdev = conf->disks[i].replacement;
if (!rdev)
rdev = conf->disks[i].rdev;
if (use_new_offset(conf, sh))
s = sh->sector + rdev->new_data_offset;
else
s = sh->sector + rdev->data_offset;
if (!bi->bi_status) {
set_bit(R5_UPTODATE, &sh->dev[i].flags);
if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
/* Note that this cannot happen on a
* replacement device. We just fail those on
* any error
*/
pr_info_ratelimited(
"md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
(unsigned long long)s,
bdevname(rdev->bdev, b));
atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
clear_bit(R5_ReadError, &sh->dev[i].flags);
clear_bit(R5_ReWrite, &sh->dev[i].flags);
} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
if (test_bit(R5_InJournal, &sh->dev[i].flags))
/*
* end read for a page in journal, this
* must be preparing for prexor in rmw
*/
set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
if (atomic_read(&rdev->read_errors))
atomic_set(&rdev->read_errors, 0);
} else {
const char *bdn = bdevname(rdev->bdev, b);
int retry = 0;
int set_bad = 0;
clear_bit(R5_UPTODATE, &sh->dev[i].flags);
if (!(bi->bi_status == BLK_STS_PROTECTION))
atomic_inc(&rdev->read_errors);
if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
pr_warn_ratelimited(
"md/raid:%s: read error on replacement device (sector %llu on %s).\n",
mdname(conf->mddev),
(unsigned long long)s,
bdn);
else if (conf->mddev->degraded >= conf->max_degraded) {
set_bad = 1;
pr_warn_ratelimited(
"md/raid:%s: read error not correctable (sector %llu on %s).\n",
mdname(conf->mddev),
(unsigned long long)s,
bdn);
} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
/* Oh, no!!! */
set_bad = 1;
pr_warn_ratelimited(
"md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
mdname(conf->mddev),
(unsigned long long)s,
bdn);
} else if (atomic_read(&rdev->read_errors)
> conf->max_nr_stripes) {
if (!test_bit(Faulty, &rdev->flags)) {
pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
mdname(conf->mddev),
atomic_read(&rdev->read_errors),
conf->max_nr_stripes);
pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
mdname(conf->mddev), bdn);
}
} else
retry = 1;
if (set_bad && test_bit(In_sync, &rdev->flags)
&& !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
retry = 1;
if (retry)
if (sh->qd_idx >= 0 && sh->pd_idx == i)
set_bit(R5_ReadError, &sh->dev[i].flags);
else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
set_bit(R5_ReadError, &sh->dev[i].flags);
clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
} else
set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
else {
clear_bit(R5_ReadError, &sh->dev[i].flags);
clear_bit(R5_ReWrite, &sh->dev[i].flags);
if (!(set_bad
&& test_bit(In_sync, &rdev->flags)
&& rdev_set_badblocks(
rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
md_error(conf->mddev, rdev);
}
}
rdev_dec_pending(rdev, conf->mddev);
bio_reset(bi);
clear_bit(R5_LOCKED, &sh->dev[i].flags);
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
}
static void raid5_end_write_request(struct bio *bi)
{
struct stripe_head *sh = bi->bi_private;
struct r5conf *conf = sh->raid_conf;
int disks = sh->disks, i;
treewide: Remove uninitialized_var() usage Using uninitialized_var() is dangerous as it papers over real bugs[1] (or can in the future), and suppresses unrelated compiler warnings (e.g. "unused variable"). If the compiler thinks it is uninitialized, either simply initialize the variable or make compiler changes. In preparation for removing[2] the[3] macro[4], remove all remaining needless uses with the following script: git grep '\buninitialized_var\b' | cut -d: -f1 | sort -u | \ xargs perl -pi -e \ 's/\buninitialized_var\(([^\)]+)\)/\1/g; s:\s*/\* (GCC be quiet|to make compiler happy) \*/$::g;' drivers/video/fbdev/riva/riva_hw.c was manually tweaked to avoid pathological white-space. No outstanding warnings were found building allmodconfig with GCC 9.3.0 for x86_64, i386, arm64, arm, powerpc, powerpc64le, s390x, mips, sparc64, alpha, and m68k. [1] https://lore.kernel.org/lkml/20200603174714.192027-1-glider@google.com/ [2] https://lore.kernel.org/lkml/CA+55aFw+Vbj0i=1TGqCR5vQkCzWJ0QxK6CernOU6eedsudAixw@mail.gmail.com/ [3] https://lore.kernel.org/lkml/CA+55aFwgbgqhbp1fkxvRKEpzyR5J8n1vKT1VZdz9knmPuXhOeg@mail.gmail.com/ [4] https://lore.kernel.org/lkml/CA+55aFz2500WfbKXAx8s67wrm9=yVJu65TpLgN_ybYNv0VEOKA@mail.gmail.com/ Reviewed-by: Leon Romanovsky <leonro@mellanox.com> # drivers/infiniband and mlx4/mlx5 Acked-by: Jason Gunthorpe <jgg@mellanox.com> # IB Acked-by: Kalle Valo <kvalo@codeaurora.org> # wireless drivers Reviewed-by: Chao Yu <yuchao0@huawei.com> # erofs Signed-off-by: Kees Cook <keescook@chromium.org>
2020-06-03 20:09:38 +00:00
struct md_rdev *rdev;
sector_t first_bad;
int bad_sectors;
int replacement = 0;
for (i = 0 ; i < disks; i++) {
if (bi == &sh->dev[i].req) {
rdev = conf->disks[i].rdev;
break;
}
if (bi == &sh->dev[i].rreq) {
rdev = conf->disks[i].replacement;
if (rdev)
replacement = 1;
else
/* rdev was removed and 'replacement'
* replaced it. rdev is not removed
* until all requests are finished.
*/
rdev = conf->disks[i].rdev;
break;
}
}
pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
(unsigned long long)sh->sector, i, atomic_read(&sh->count),
bi->bi_status);
if (i == disks) {
bio_reset(bi);
BUG();
return;
}
if (replacement) {
if (bi->bi_status)
md_error(conf->mddev, rdev);
else if (is_badblock(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf),
&first_bad, &bad_sectors))
set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
} else {
if (bi->bi_status) {
set_bit(STRIPE_DEGRADED, &sh->state);
set_bit(WriteErrorSeen, &rdev->flags);
set_bit(R5_WriteError, &sh->dev[i].flags);
if (!test_and_set_bit(WantReplacement, &rdev->flags))
set_bit(MD_RECOVERY_NEEDED,
&rdev->mddev->recovery);
} else if (is_badblock(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf),
&first_bad, &bad_sectors)) {
set_bit(R5_MadeGood, &sh->dev[i].flags);
if (test_bit(R5_ReadError, &sh->dev[i].flags))
/* That was a successful write so make
* sure it looks like we already did
* a re-write.
*/
set_bit(R5_ReWrite, &sh->dev[i].flags);
}
}
rdev_dec_pending(rdev, conf->mddev);
if (sh->batch_head && bi->bi_status && !replacement)
set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
bio_reset(bi);
if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
clear_bit(R5_LOCKED, &sh->dev[i].flags);
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (sh->batch_head && sh != sh->batch_head)
raid5_release_stripe(sh->batch_head);
}
static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
{
char b[BDEVNAME_SIZE];
struct r5conf *conf = mddev->private;
unsigned long flags;
pr_debug("raid456: error called\n");
spin_lock_irqsave(&conf->device_lock, flags);
if (test_bit(In_sync, &rdev->flags) &&
mddev->degraded == conf->max_degraded) {
/*
* Don't allow to achieve failed state
* Don't try to recover this device
*/
conf->recovery_disabled = mddev->recovery_disabled;
spin_unlock_irqrestore(&conf->device_lock, flags);
return;
}
set_bit(Faulty, &rdev->flags);
clear_bit(In_sync, &rdev->flags);
mddev->degraded = raid5_calc_degraded(conf);
spin_unlock_irqrestore(&conf->device_lock, flags);
set_bit(MD_RECOVERY_INTR, &mddev->recovery);
md: make it easier to wait for bad blocks to be acknowledged. It is only safe to choose not to write to a bad block if that bad block is safely recorded in metadata - i.e. if it has been 'acknowledged'. If it hasn't we need to wait for the acknowledgement. We support that using rdev->blocked wait and md_wait_for_blocked_rdev by introducing a new device flag 'BlockedBadBlock'. This flag is only advisory. It is cleared whenever we acknowledge a bad block, so that a waiter can re-check the particular bad blocks that it is interested it. It should be set by a caller when they find they need to wait. This (set after test) is inherently racy, but as md_wait_for_blocked_rdev already has a timeout, losing the race will have minimal impact. When we clear "Blocked" was also clear "BlockedBadBlocks" incase it was set incorrectly (see above race). We also modify the way we manage 'Blocked' to fit better with the new handling of 'BlockedBadBlocks' and to make it consistent between externally managed and internally managed metadata. This requires that each raidXd loop checks if the metadata needs to be written and triggers a write (md_check_recovery) if needed. Otherwise a queued write request might cause raidXd to wait for the metadata to write, and only that thread can write it. Before writing metadata, we set FaultRecorded for all devices that are Faulty, then after writing the metadata we clear Blocked for any device for which the Fault was certainly Recorded. The 'faulty' device flag now appears in sysfs if the device is faulty *or* it has unacknowledged bad blocks. So user-space which does not understand bad blocks can continue to function correctly. User space which does, should not assume a device is faulty until it sees the 'faulty' flag, and then sees the list of unacknowledged bad blocks is empty. Signed-off-by: NeilBrown <neilb@suse.de>
2011-07-28 01:31:48 +00:00
set_bit(Blocked, &rdev->flags);
set_mask_bits(&mddev->sb_flags, 0,
BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
"md/raid:%s: Operation continuing on %d devices.\n",
mdname(mddev),
bdevname(rdev->bdev, b),
mdname(mddev),
conf->raid_disks - mddev->degraded);
r5c_update_on_rdev_error(mddev, rdev);
}
/*
* Input: a 'big' sector number,
* Output: index of the data and parity disk, and the sector # in them.
*/
sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
int previous, int *dd_idx,
struct stripe_head *sh)
{
sector_t stripe, stripe2;
sector_t chunk_number;
unsigned int chunk_offset;
int pd_idx, qd_idx;
int ddf_layout = 0;
sector_t new_sector;
int algorithm = previous ? conf->prev_algo
: conf->algorithm;
int sectors_per_chunk = previous ? conf->prev_chunk_sectors
: conf->chunk_sectors;
int raid_disks = previous ? conf->previous_raid_disks
: conf->raid_disks;
int data_disks = raid_disks - conf->max_degraded;
/* First compute the information on this sector */
/*
* Compute the chunk number and the sector offset inside the chunk
*/
chunk_offset = sector_div(r_sector, sectors_per_chunk);
chunk_number = r_sector;
/*
* Compute the stripe number
*/
stripe = chunk_number;
*dd_idx = sector_div(stripe, data_disks);
stripe2 = stripe;
/*
* Select the parity disk based on the user selected algorithm.
*/
pd_idx = qd_idx = -1;
switch(conf->level) {
case 4:
pd_idx = data_disks;
break;
case 5:
switch (algorithm) {
case ALGORITHM_LEFT_ASYMMETRIC:
pd_idx = data_disks - sector_div(stripe2, raid_disks);
if (*dd_idx >= pd_idx)
(*dd_idx)++;
break;
case ALGORITHM_RIGHT_ASYMMETRIC:
pd_idx = sector_div(stripe2, raid_disks);
if (*dd_idx >= pd_idx)
(*dd_idx)++;
break;
case ALGORITHM_LEFT_SYMMETRIC:
pd_idx = data_disks - sector_div(stripe2, raid_disks);
*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
break;
case ALGORITHM_RIGHT_SYMMETRIC:
pd_idx = sector_div(stripe2, raid_disks);
*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
break;
case ALGORITHM_PARITY_0:
pd_idx = 0;
(*dd_idx)++;
break;
case ALGORITHM_PARITY_N:
pd_idx = data_disks;
break;
default:
BUG();
}
break;
case 6:
switch (algorithm) {
case ALGORITHM_LEFT_ASYMMETRIC:
pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
qd_idx = pd_idx + 1;
if (pd_idx == raid_disks-1) {
(*dd_idx)++; /* Q D D D P */
qd_idx = 0;
} else if (*dd_idx >= pd_idx)
(*dd_idx) += 2; /* D D P Q D */
break;
case ALGORITHM_RIGHT_ASYMMETRIC:
pd_idx = sector_div(stripe2, raid_disks);
qd_idx = pd_idx + 1;
if (pd_idx == raid_disks-1) {
(*dd_idx)++; /* Q D D D P */
qd_idx = 0;
} else if (*dd_idx >= pd_idx)
(*dd_idx) += 2; /* D D P Q D */
break;
case ALGORITHM_LEFT_SYMMETRIC:
pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
qd_idx = (pd_idx + 1) % raid_disks;
*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
break;
case ALGORITHM_RIGHT_SYMMETRIC:
pd_idx = sector_div(stripe2, raid_disks);
qd_idx = (pd_idx + 1) % raid_disks;
*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
break;
case ALGORITHM_PARITY_0:
pd_idx = 0;
qd_idx = 1;
(*dd_idx) += 2;
break;
case ALGORITHM_PARITY_N:
pd_idx = data_disks;
qd_idx = data_disks + 1;
break;
case ALGORITHM_ROTATING_ZERO_RESTART:
/* Exactly the same as RIGHT_ASYMMETRIC, but or
* of blocks for computing Q is different.
*/
pd_idx = sector_div(stripe2, raid_disks);
qd_idx = pd_idx + 1;
if (pd_idx == raid_disks-1) {
(*dd_idx)++; /* Q D D D P */
qd_idx = 0;
} else if (*dd_idx >= pd_idx)
(*dd_idx) += 2; /* D D P Q D */
ddf_layout = 1;
break;
case ALGORITHM_ROTATING_N_RESTART:
/* Same a left_asymmetric, by first stripe is
* D D D P Q rather than
* Q D D D P
*/
stripe2 += 1;
pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
qd_idx = pd_idx + 1;
if (pd_idx == raid_disks-1) {
(*dd_idx)++; /* Q D D D P */
qd_idx = 0;
} else if (*dd_idx >= pd_idx)
(*dd_idx) += 2; /* D D P Q D */
ddf_layout = 1;
break;
case ALGORITHM_ROTATING_N_CONTINUE:
/* Same as left_symmetric but Q is before P */
pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
ddf_layout = 1;
break;
case ALGORITHM_LEFT_ASYMMETRIC_6:
/* RAID5 left_asymmetric, with Q on last device */
pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
if (*dd_idx >= pd_idx)
(*dd_idx)++;
qd_idx = raid_disks - 1;
break;
case ALGORITHM_RIGHT_ASYMMETRIC_6:
pd_idx = sector_div(stripe2, raid_disks-1);
if (*dd_idx >= pd_idx)
(*dd_idx)++;
qd_idx = raid_disks - 1;
break;
case ALGORITHM_LEFT_SYMMETRIC_6:
pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
qd_idx = raid_disks - 1;
break;
case ALGORITHM_RIGHT_SYMMETRIC_6:
pd_idx = sector_div(stripe2, raid_disks-1);
*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
qd_idx = raid_disks - 1;
break;
case ALGORITHM_PARITY_0_6:
pd_idx = 0;
(*dd_idx)++;
qd_idx = raid_disks - 1;
break;
default:
BUG();
}
break;
}
if (sh) {
sh->pd_idx = pd_idx;
sh->qd_idx = qd_idx;
sh->ddf_layout = ddf_layout;
}
/*
* Finally, compute the new sector number
*/
new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
return new_sector;
}
sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
{
struct r5conf *conf = sh->raid_conf;
int raid_disks = sh->disks;
int data_disks = raid_disks - conf->max_degraded;
sector_t new_sector = sh->sector, check;
int sectors_per_chunk = previous ? conf->prev_chunk_sectors
: conf->chunk_sectors;
int algorithm = previous ? conf->prev_algo
: conf->algorithm;
sector_t stripe;
int chunk_offset;
sector_t chunk_number;
int dummy1, dd_idx = i;
sector_t r_sector;
struct stripe_head sh2;
chunk_offset = sector_div(new_sector, sectors_per_chunk);
stripe = new_sector;
if (i == sh->pd_idx)
return 0;
switch(conf->level) {
case 4: break;
case 5:
switch (algorithm) {
case ALGORITHM_LEFT_ASYMMETRIC:
case ALGORITHM_RIGHT_ASYMMETRIC:
if (i > sh->pd_idx)
i--;
break;
case ALGORITHM_LEFT_SYMMETRIC:
case ALGORITHM_RIGHT_SYMMETRIC:
if (i < sh->pd_idx)
i += raid_disks;
i -= (sh->pd_idx + 1);
break;
case ALGORITHM_PARITY_0:
i -= 1;
break;
case ALGORITHM_PARITY_N:
break;
default:
BUG();
}
break;
case 6:
if (i == sh->qd_idx)
return 0; /* It is the Q disk */
switch (algorithm) {
case ALGORITHM_LEFT_ASYMMETRIC:
case ALGORITHM_RIGHT_ASYMMETRIC:
case ALGORITHM_ROTATING_ZERO_RESTART:
case ALGORITHM_ROTATING_N_RESTART:
if (sh->pd_idx == raid_disks-1)
i--; /* Q D D D P */
else if (i > sh->pd_idx)
i -= 2; /* D D P Q D */
break;
case ALGORITHM_LEFT_SYMMETRIC:
case ALGORITHM_RIGHT_SYMMETRIC:
if (sh->pd_idx == raid_disks-1)
i--; /* Q D D D P */
else {
/* D D P Q D */
if (i < sh->pd_idx)
i += raid_disks;
i -= (sh->pd_idx + 2);
}
break;
case ALGORITHM_PARITY_0:
i -= 2;
break;
case ALGORITHM_PARITY_N:
break;
case ALGORITHM_ROTATING_N_CONTINUE:
/* Like left_symmetric, but P is before Q */
if (sh->pd_idx == 0)
i--; /* P D D D Q */
else {
/* D D Q P D */
if (i < sh->pd_idx)
i += raid_disks;
i -= (sh->pd_idx + 1);
}
break;
case ALGORITHM_LEFT_ASYMMETRIC_6:
case ALGORITHM_RIGHT_ASYMMETRIC_6:
if (i > sh->pd_idx)
i--;
break;
case ALGORITHM_LEFT_SYMMETRIC_6:
case ALGORITHM_RIGHT_SYMMETRIC_6:
if (i < sh->pd_idx)
i += data_disks + 1;
i -= (sh->pd_idx + 1);
break;
case ALGORITHM_PARITY_0_6:
i -= 1;
break;
default:
BUG();
}
break;
}
chunk_number = stripe * data_disks + i;
r_sector = chunk_number * sectors_per_chunk + chunk_offset;
check = raid5_compute_sector(conf, r_sector,
previous, &dummy1, &sh2);
if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
|| sh2.qd_idx != sh->qd_idx) {
pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
mdname(conf->mddev));
return 0;
}
return r_sector;
}
/*
* There are cases where we want handle_stripe_dirtying() and
* schedule_reconstruction() to delay towrite to some dev of a stripe.
*
* This function checks whether we want to delay the towrite. Specifically,
* we delay the towrite when:
*
* 1. degraded stripe has a non-overwrite to the missing dev, AND this
* stripe has data in journal (for other devices).
*
* In this case, when reading data for the non-overwrite dev, it is
* necessary to handle complex rmw of write back cache (prexor with
* orig_page, and xor with page). To keep read path simple, we would
* like to flush data in journal to RAID disks first, so complex rmw
* is handled in the write patch (handle_stripe_dirtying).
*
* 2. when journal space is critical (R5C_LOG_CRITICAL=1)
*
* It is important to be able to flush all stripes in raid5-cache.
* Therefore, we need reserve some space on the journal device for
* these flushes. If flush operation includes pending writes to the
* stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
* for the flush out. If we exclude these pending writes from flush
* operation, we only need (conf->max_degraded + 1) pages per stripe.
* Therefore, excluding pending writes in these cases enables more
* efficient use of the journal device.
*
* Note: To make sure the stripe makes progress, we only delay
* towrite for stripes with data already in journal (injournal > 0).
* When LOG_CRITICAL, stripes with injournal == 0 will be sent to
* no_space_stripes list.
*
* 3. during journal failure
* In journal failure, we try to flush all cached data to raid disks
* based on data in stripe cache. The array is read-only to upper
* layers, so we would skip all pending writes.
*
*/
static inline bool delay_towrite(struct r5conf *conf,
struct r5dev *dev,
struct stripe_head_state *s)
{
/* case 1 above */
if (!test_bit(R5_OVERWRITE, &dev->flags) &&
!test_bit(R5_Insync, &dev->flags) && s->injournal)
return true;
/* case 2 above */
if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
s->injournal > 0)
return true;
/* case 3 above */
if (s->log_failed && s->injournal)
return true;
return false;
}
static void
schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
int rcw, int expand)
{
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
struct r5conf *conf = sh->raid_conf;
int level = conf->level;
if (rcw) {
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
/*
* In some cases, handle_stripe_dirtying initially decided to
* run rmw and allocates extra page for prexor. However, rcw is
* cheaper later on. We need to free the extra page now,
* because we won't be able to do that in ops_complete_prexor().
*/
r5c_release_extra_page(sh);
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (dev->towrite && !delay_towrite(conf, dev, s)) {
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantdrain, &dev->flags);
if (!expand)
clear_bit(R5_UPTODATE, &dev->flags);
s->locked++;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
} else if (test_bit(R5_InJournal, &dev->flags)) {
set_bit(R5_LOCKED, &dev->flags);
s->locked++;
}
}
/* if we are not expanding this is a proper write request, and
* there will be bios with new data to be drained into the
* stripe cache
*/
if (!expand) {
if (!s->locked)
/* False alarm, nothing to do */
return;
sh->reconstruct_state = reconstruct_state_drain_run;
set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
} else
sh->reconstruct_state = reconstruct_state_run;
set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
if (s->locked + conf->max_degraded == disks)
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
atomic_inc(&conf->pending_full_writes);
} else {
BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
BUG_ON(level == 6 &&
(!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (i == pd_idx || i == qd_idx)
continue;
if (dev->towrite &&
(test_bit(R5_UPTODATE, &dev->flags) ||
test_bit(R5_Wantcompute, &dev->flags))) {
set_bit(R5_Wantdrain, &dev->flags);
set_bit(R5_LOCKED, &dev->flags);
clear_bit(R5_UPTODATE, &dev->flags);
s->locked++;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
} else if (test_bit(R5_InJournal, &dev->flags)) {
set_bit(R5_LOCKED, &dev->flags);
s->locked++;
}
}
if (!s->locked)
/* False alarm - nothing to do */
return;
sh->reconstruct_state = reconstruct_state_prexor_drain_run;
set_bit(STRIPE_OP_PREXOR, &s->ops_request);
set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
}
/* keep the parity disk(s) locked while asynchronous operations
* are in flight
*/
set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
s->locked++;
if (level == 6) {
int qd_idx = sh->qd_idx;
struct r5dev *dev = &sh->dev[qd_idx];
set_bit(R5_LOCKED, &dev->flags);
clear_bit(R5_UPTODATE, &dev->flags);
s->locked++;
}
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
!test_bit(STRIPE_FULL_WRITE, &sh->state) &&
test_bit(R5_Insync, &sh->dev[pd_idx].flags))
set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
__func__, (unsigned long long)sh->sector,
s->locked, s->ops_request);
}
/*
* Each stripe/dev can have one or more bion attached.
* toread/towrite point to the first in a chain.
* The bi_next chain must be in order.
*/
static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
int forwrite, int previous)
{
struct bio **bip;
struct r5conf *conf = sh->raid_conf;
int firstwrite=0;
pr_debug("adding bi b#%llu to stripe s#%llu\n",
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
(unsigned long long)bi->bi_iter.bi_sector,
(unsigned long long)sh->sector);
spin_lock_irq(&sh->stripe_lock);
sh->dev[dd_idx].write_hint = bi->bi_write_hint;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
/* Don't allow new IO added to stripes in batch list */
if (sh->batch_head)
goto overlap;
if (forwrite) {
bip = &sh->dev[dd_idx].towrite;
if (*bip == NULL)
firstwrite = 1;
} else
bip = &sh->dev[dd_idx].toread;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
goto overlap;
bip = & (*bip)->bi_next;
}
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
goto overlap;
if (forwrite && raid5_has_ppl(conf)) {
/*
* With PPL only writes to consecutive data chunks within a
* stripe are allowed because for a single stripe_head we can
* only have one PPL entry at a time, which describes one data
* range. Not really an overlap, but wait_for_overlap can be
* used to handle this.
*/
sector_t sector;
sector_t first = 0;
sector_t last = 0;
int count = 0;
int i;
for (i = 0; i < sh->disks; i++) {
if (i != sh->pd_idx &&
(i == dd_idx || sh->dev[i].towrite)) {
sector = sh->dev[i].sector;
if (count == 0 || sector < first)
first = sector;
if (sector > last)
last = sector;
count++;
}
}
if (first + conf->chunk_sectors * (count - 1) != last)
goto overlap;
}
if (!forwrite || previous)
clear_bit(STRIPE_BATCH_READY, &sh->state);
BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
if (*bip)
bi->bi_next = *bip;
*bip = bi;
bio_inc_remaining(bi);
md_write_inc(conf->mddev, bi);
if (forwrite) {
/* check if page is covered */
sector_t sector = sh->dev[dd_idx].sector;
for (bi=sh->dev[dd_idx].towrite;
sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
bi && bi->bi_iter.bi_sector <= sector;
bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
if (bio_end_sector(bi) >= sector)
sector = bio_end_sector(bi);
}
if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
sh->overwrite_disks++;
}
pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
(unsigned long long)(*bip)->bi_iter.bi_sector,
(unsigned long long)sh->sector, dd_idx);
if (conf->mddev->bitmap && firstwrite) {
/* Cannot hold spinlock over bitmap_startwrite,
* but must ensure this isn't added to a batch until
* we have added to the bitmap and set bm_seq.
* So set STRIPE_BITMAP_PENDING to prevent
* batching.
* If multiple add_stripe_bio() calls race here they
* much all set STRIPE_BITMAP_PENDING. So only the first one
* to complete "bitmap_startwrite" gets to set
* STRIPE_BIT_DELAY. This is important as once a stripe
* is added to a batch, STRIPE_BIT_DELAY cannot be changed
* any more.
*/
set_bit(STRIPE_BITMAP_PENDING, &sh->state);
spin_unlock_irq(&sh->stripe_lock);
md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0);
spin_lock_irq(&sh->stripe_lock);
clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
if (!sh->batch_head) {
sh->bm_seq = conf->seq_flush+1;
set_bit(STRIPE_BIT_DELAY, &sh->state);
}
}
spin_unlock_irq(&sh->stripe_lock);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (stripe_can_batch(sh))
stripe_add_to_batch_list(conf, sh);
return 1;
overlap:
set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
spin_unlock_irq(&sh->stripe_lock);
return 0;
}
static void end_reshape(struct r5conf *conf);
static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
struct stripe_head *sh)
{
int sectors_per_chunk =
previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
int dd_idx;
int chunk_offset = sector_div(stripe, sectors_per_chunk);
int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
raid5_compute_sector(conf,
stripe * (disks - conf->max_degraded)
*sectors_per_chunk + chunk_offset,
previous,
&dd_idx, sh);
}
static void
handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
struct stripe_head_state *s, int disks)
{
int i;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
for (i = disks; i--; ) {
struct bio *bi;
int bitmap_end = 0;
if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
struct md_rdev *rdev;
rcu_read_lock();
rdev = rcu_dereference(conf->disks[i].rdev);
if (rdev && test_bit(In_sync, &rdev->flags) &&
!test_bit(Faulty, &rdev->flags))
atomic_inc(&rdev->nr_pending);
else
rdev = NULL;
rcu_read_unlock();
if (rdev) {
if (!rdev_set_badblocks(
rdev,
sh->sector,
RAID5_STRIPE_SECTORS(conf), 0))
md_error(conf->mddev, rdev);
rdev_dec_pending(rdev, conf->mddev);
}
}
spin_lock_irq(&sh->stripe_lock);
/* fail all writes first */
bi = sh->dev[i].towrite;
sh->dev[i].towrite = NULL;
sh->overwrite_disks = 0;
spin_unlock_irq(&sh->stripe_lock);
if (bi)
bitmap_end = 1;
log_stripe_write_finished(sh);
raid5: log reclaim support This is the reclaim support for raid5 log. A stripe write will have following steps: 1. reconstruct the stripe, read data/calculate parity. ops_run_io prepares to write data/parity to raid disks 2. hijack ops_run_io. stripe data/parity is appending to log disk 3. flush log disk cache 4. ops_run_io run again and do normal operation. stripe data/parity is written in raid array disks. raid core can return io to upper layer. 5. flush cache of all raid array disks 6. update super block 7. log disk space used by the stripe can be reused In practice, several stripes consist of an io_unit and we will batch several io_unit in different steps, but the whole process doesn't change. It's possible io return just after data/parity hit log disk, but then read IO will need read from log disk. For simplicity, IO return happens at step 4, where read IO can directly read from raid disks. Currently reclaim run if there is specific reclaimable space (1/4 disk size or 10G) or we are out of space. Reclaim is just to free log disk spaces, it doesn't impact data consistency. The size based force reclaim is to make sure log isn't too big, so recovery doesn't scan log too much. Recovery make sure raid disks and log disk have the same data of a stripe. If crash happens before 4, recovery might/might not recovery stripe's data/parity depending on if data/parity and its checksum matches. In either case, this doesn't change the syntax of an IO write. After step 3, stripe is guaranteed recoverable, because stripe's data/parity is persistent in log disk. In some cases, log disk content and raid disks content of a stripe are the same, but recovery will still copy log disk content to raid disks, this doesn't impact data consistency. space reuse happens after superblock update and cache flush. There is one situation we want to avoid. A broken meta in the middle of a log causes recovery can't find meta at the head of log. If operations require meta at the head persistent in log, we must make sure meta before it persistent in log too. The case is stripe data/parity is in log and we start write stripe to raid disks (before step 4). stripe data/parity must be persistent in log before we do the write to raid disks. The solution is we restrictly maintain io_unit list order. In this case, we only write stripes of an io_unit to raid disks till the io_unit is the first one whose data/parity is in log. The io_unit list order is important for other cases too. For example, some io_unit are reclaimable and others not. They can be mixed in the list, we shouldn't reuse space of an unreclaimable io_unit. Includes fixes to problems which were... Reported-by: kbuild test robot <fengguang.wu@intel.com> Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: NeilBrown <neilb@suse.com>
2015-08-13 21:32:00 +00:00
if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
wake_up(&conf->wait_for_overlap);
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (bi && bi->bi_iter.bi_sector <
sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
md_write_end(conf->mddev);
bio_io_error(bi);
bi = nextbi;
}
if (bitmap_end)
md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0, 0);
bitmap_end = 0;
/* and fail all 'written' */
bi = sh->dev[i].written;
sh->dev[i].written = NULL;
if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
sh->dev[i].page = sh->dev[i].orig_page;
}
if (bi) bitmap_end = 1;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (bi && bi->bi_iter.bi_sector <
sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
md_write_end(conf->mddev);
bio_io_error(bi);
bi = bi2;
}
/* fail any reads if this device is non-operational and
* the data has not reached the cache yet.
*/
if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
s->failed > conf->max_degraded &&
(!test_bit(R5_Insync, &sh->dev[i].flags) ||
test_bit(R5_ReadError, &sh->dev[i].flags))) {
spin_lock_irq(&sh->stripe_lock);
bi = sh->dev[i].toread;
sh->dev[i].toread = NULL;
spin_unlock_irq(&sh->stripe_lock);
if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
wake_up(&conf->wait_for_overlap);
if (bi)
s->to_read--;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (bi && bi->bi_iter.bi_sector <
sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
struct bio *nextbi =
r5_next_bio(conf, bi, sh->dev[i].sector);
bio_io_error(bi);
bi = nextbi;
}
}
if (bitmap_end)
md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0, 0);
/* If we were in the middle of a write the parity block might
* still be locked - so just clear all R5_LOCKED flags
*/
clear_bit(R5_LOCKED, &sh->dev[i].flags);
}
s->to_write = 0;
s->written = 0;
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
if (atomic_dec_and_test(&conf->pending_full_writes))
md_wakeup_thread(conf->mddev->thread);
}
static void
handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
struct stripe_head_state *s)
{
int abort = 0;
int i;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
clear_bit(STRIPE_SYNCING, &sh->state);
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
wake_up(&conf->wait_for_overlap);
s->syncing = 0;
s->replacing = 0;
/* There is nothing more to do for sync/check/repair.
* Don't even need to abort as that is handled elsewhere
* if needed, and not always wanted e.g. if there is a known
* bad block here.
* For recover/replace we need to record a bad block on all
* non-sync devices, or abort the recovery
*/
if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
/* During recovery devices cannot be removed, so
* locking and refcounting of rdevs is not needed
*/
rcu_read_lock();
for (i = 0; i < conf->raid_disks; i++) {
struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
if (rdev
&& !test_bit(Faulty, &rdev->flags)
&& !test_bit(In_sync, &rdev->flags)
&& !rdev_set_badblocks(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0))
abort = 1;
rdev = rcu_dereference(conf->disks[i].replacement);
if (rdev
&& !test_bit(Faulty, &rdev->flags)
&& !test_bit(In_sync, &rdev->flags)
&& !rdev_set_badblocks(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0))
abort = 1;
}
rcu_read_unlock();
if (abort)
conf->recovery_disabled =
conf->mddev->recovery_disabled;
}
md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
}
static int want_replace(struct stripe_head *sh, int disk_idx)
{
struct md_rdev *rdev;
int rv = 0;
rcu_read_lock();
rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
if (rdev
&& !test_bit(Faulty, &rdev->flags)
&& !test_bit(In_sync, &rdev->flags)
&& (rdev->recovery_offset <= sh->sector
|| rdev->mddev->recovery_cp <= sh->sector))
rv = 1;
rcu_read_unlock();
return rv;
}
static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
int disk_idx, int disks)
{
struct r5dev *dev = &sh->dev[disk_idx];
struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
&sh->dev[s->failed_num[1]] };
int i;
bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
if (test_bit(R5_LOCKED, &dev->flags) ||
test_bit(R5_UPTODATE, &dev->flags))
/* No point reading this as we already have it or have
* decided to get it.
*/
return 0;
if (dev->toread ||
(dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
/* We need this block to directly satisfy a request */
return 1;
if (s->syncing || s->expanding ||
(s->replacing && want_replace(sh, disk_idx)))
/* When syncing, or expanding we read everything.
* When replacing, we need the replaced block.
*/
return 1;
if ((s->failed >= 1 && fdev[0]->toread) ||
(s->failed >= 2 && fdev[1]->toread))
/* If we want to read from a failed device, then
* we need to actually read every other device.
*/
return 1;
/* Sometimes neither read-modify-write nor reconstruct-write
* cycles can work. In those cases we read every block we
* can. Then the parity-update is certain to have enough to
* work with.
* This can only be a problem when we need to write something,
* and some device has failed. If either of those tests
* fail we need look no further.
*/
if (!s->failed || !s->to_write)
return 0;
if (test_bit(R5_Insync, &dev->flags) &&
!test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
/* Pre-reads at not permitted until after short delay
* to gather multiple requests. However if this
* device is no Insync, the block could only be computed
* and there is no need to delay that.
*/
return 0;
for (i = 0; i < s->failed && i < 2; i++) {
if (fdev[i]->towrite &&
!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
!test_bit(R5_OVERWRITE, &fdev[i]->flags))
/* If we have a partial write to a failed
* device, then we will need to reconstruct
* the content of that device, so all other
* devices must be read.
*/
return 1;
if (s->failed >= 2 &&
(fdev[i]->towrite ||
s->failed_num[i] == sh->pd_idx ||
s->failed_num[i] == sh->qd_idx) &&
!test_bit(R5_UPTODATE, &fdev[i]->flags))
/* In max degraded raid6, If the failed disk is P, Q,
* or we want to read the failed disk, we need to do
* reconstruct-write.
*/
force_rcw = true;
}
/* If we are forced to do a reconstruct-write, because parity
* cannot be trusted and we are currently recovering it, there
* is extra need to be careful.
* If one of the devices that we would need to read, because
* it is not being overwritten (and maybe not written at all)
* is missing/faulty, then we need to read everything we can.
*/
if (!force_rcw &&
sh->sector < sh->raid_conf->mddev->recovery_cp)
/* reconstruct-write isn't being forced */
return 0;
for (i = 0; i < s->failed && i < 2; i++) {
if (s->failed_num[i] != sh->pd_idx &&
s->failed_num[i] != sh->qd_idx &&
!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
!test_bit(R5_OVERWRITE, &fdev[i]->flags))
return 1;
}
return 0;
}
/* fetch_block - checks the given member device to see if its data needs
* to be read or computed to satisfy a request.
*
* Returns 1 when no more member devices need to be checked, otherwise returns
* 0 to tell the loop in handle_stripe_fill to continue
*/
static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
int disk_idx, int disks)
{
struct r5dev *dev = &sh->dev[disk_idx];
/* is the data in this block needed, and can we get it? */
if (need_this_block(sh, s, disk_idx, disks)) {
/* we would like to get this block, possibly by computing it,
* otherwise read it if the backing disk is insync
*/
BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
BUG_ON(test_bit(R5_Wantread, &dev->flags));
BUG_ON(sh->batch_head);
/*
* In the raid6 case if the only non-uptodate disk is P
* then we already trusted P to compute the other failed
* drives. It is safe to compute rather than re-read P.
* In other cases we only compute blocks from failed
* devices, otherwise check/repair might fail to detect
* a real inconsistency.
*/
if ((s->uptodate == disks - 1) &&
((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
(s->failed && (disk_idx == s->failed_num[0] ||
disk_idx == s->failed_num[1])))) {
/* have disk failed, and we're requested to fetch it;
* do compute it
*/
pr_debug("Computing stripe %llu block %d\n",
(unsigned long long)sh->sector, disk_idx);
set_bit(STRIPE_COMPUTE_RUN, &sh->state);
set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
set_bit(R5_Wantcompute, &dev->flags);
sh->ops.target = disk_idx;
sh->ops.target2 = -1; /* no 2nd target */
s->req_compute = 1;
/* Careful: from this point on 'uptodate' is in the eye
* of raid_run_ops which services 'compute' operations
* before writes. R5_Wantcompute flags a block that will
* be R5_UPTODATE by the time it is needed for a
* subsequent operation.
*/
s->uptodate++;
return 1;
} else if (s->uptodate == disks-2 && s->failed >= 2) {
/* Computing 2-failure is *very* expensive; only
* do it if failed >= 2
*/
int other;
for (other = disks; other--; ) {
if (other == disk_idx)
continue;
if (!test_bit(R5_UPTODATE,
&sh->dev[other].flags))
break;
}
BUG_ON(other < 0);
pr_debug("Computing stripe %llu blocks %d,%d\n",
(unsigned long long)sh->sector,
disk_idx, other);
set_bit(STRIPE_COMPUTE_RUN, &sh->state);
set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
set_bit(R5_Wantcompute, &sh->dev[other].flags);
sh->ops.target = disk_idx;
sh->ops.target2 = other;
s->uptodate += 2;
s->req_compute = 1;
return 1;
} else if (test_bit(R5_Insync, &dev->flags)) {
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantread, &dev->flags);
s->locked++;
pr_debug("Reading block %d (sync=%d)\n",
disk_idx, s->syncing);
}
}
return 0;
}
/*
* handle_stripe_fill - read or compute data to satisfy pending requests.
*/
static void handle_stripe_fill(struct stripe_head *sh,
struct stripe_head_state *s,
int disks)
{
int i;
/* look for blocks to read/compute, skip this if a compute
* is already in flight, or if the stripe contents are in the
* midst of changing due to a write
*/
if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
!sh->reconstruct_state) {
/*
* For degraded stripe with data in journal, do not handle
* read requests yet, instead, flush the stripe to raid
* disks first, this avoids handling complex rmw of write
* back cache (prexor with orig_page, and then xor with
* page) in the read path
*/
if (s->injournal && s->failed) {
if (test_bit(STRIPE_R5C_CACHING, &sh->state))
r5c_make_stripe_write_out(sh);
goto out;
}
for (i = disks; i--; )
if (fetch_block(sh, s, i, disks))
break;
}
out:
set_bit(STRIPE_HANDLE, &sh->state);
}
static void break_stripe_batch_list(struct stripe_head *head_sh,
unsigned long handle_flags);
/* handle_stripe_clean_event
* any written block on an uptodate or failed drive can be returned.
* Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
* never LOCKED, so we don't need to test 'failed' directly.
*/
static void handle_stripe_clean_event(struct r5conf *conf,
struct stripe_head *sh, int disks)
{
int i;
struct r5dev *dev;
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
int discard_pending = 0;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
struct stripe_head *head_sh = sh;
bool do_endio = false;
for (i = disks; i--; )
if (sh->dev[i].written) {
dev = &sh->dev[i];
if (!test_bit(R5_LOCKED, &dev->flags) &&
(test_bit(R5_UPTODATE, &dev->flags) ||
test_bit(R5_Discard, &dev->flags) ||
test_bit(R5_SkipCopy, &dev->flags))) {
/* We can return any write requests */
struct bio *wbi, *wbi2;
pr_debug("Return write for disc %d\n", i);
if (test_and_clear_bit(R5_Discard, &dev->flags))
clear_bit(R5_UPTODATE, &dev->flags);
if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
}
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
do_endio = true;
returnbi:
dev->page = dev->orig_page;
wbi = dev->written;
dev->written = NULL;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
while (wbi && wbi->bi_iter.bi_sector <
dev->sector + RAID5_STRIPE_SECTORS(conf)) {
wbi2 = r5_next_bio(conf, wbi, dev->sector);
md_write_end(conf->mddev);
bio_endio(wbi);
wbi = wbi2;
}
md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
RAID5_STRIPE_SECTORS(conf),
!test_bit(STRIPE_DEGRADED, &sh->state),
0);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head_sh->batch_head) {
sh = list_first_entry(&sh->batch_list,
struct stripe_head,
batch_list);
if (sh != head_sh) {
dev = &sh->dev[i];
goto returnbi;
}
}
sh = head_sh;
dev = &sh->dev[i];
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
} else if (test_bit(R5_Discard, &dev->flags))
discard_pending = 1;
}
raid5: add basic stripe log This introduces a simple log for raid5. Data/parity writing to raid array first writes to the log, then write to raid array disks. If crash happens, we can recovery data from the log. This can speed up raid resync and fix write hole issue. The log structure is pretty simple. Data/meta data is stored in block unit, which is 4k generally. It has only one type of meta data block. The meta data block can track 3 types of data, stripe data, stripe parity and flush block. MD superblock will point to the last valid meta data block. Each meta data block has checksum/seq number, so recovery can scan the log correctly. We store a checksum of stripe data/parity to the metadata block, so meta data and stripe data/parity can be written to log disk together. otherwise, meta data write must wait till stripe data/parity is finished. For stripe data, meta data block will record stripe data sector and size. Currently the size is always 4k. This meta data record can be made simpler if we just fix write hole (eg, we can record data of a stripe's different disks together), but this format can be extended to support caching in the future, which must record data address/size. For stripe parity, meta data block will record stripe sector. It's size should be 4k (for raid5) or 8k (for raid6). We always store p parity first. This format should work for caching too. flush block indicates a stripe is in raid array disks. Fixing write hole doesn't need this type of meta data, it's for caching extension. Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: NeilBrown <neilb@suse.com>
2015-08-13 21:31:59 +00:00
log_stripe_write_finished(sh);
raid5: log reclaim support This is the reclaim support for raid5 log. A stripe write will have following steps: 1. reconstruct the stripe, read data/calculate parity. ops_run_io prepares to write data/parity to raid disks 2. hijack ops_run_io. stripe data/parity is appending to log disk 3. flush log disk cache 4. ops_run_io run again and do normal operation. stripe data/parity is written in raid array disks. raid core can return io to upper layer. 5. flush cache of all raid array disks 6. update super block 7. log disk space used by the stripe can be reused In practice, several stripes consist of an io_unit and we will batch several io_unit in different steps, but the whole process doesn't change. It's possible io return just after data/parity hit log disk, but then read IO will need read from log disk. For simplicity, IO return happens at step 4, where read IO can directly read from raid disks. Currently reclaim run if there is specific reclaimable space (1/4 disk size or 10G) or we are out of space. Reclaim is just to free log disk spaces, it doesn't impact data consistency. The size based force reclaim is to make sure log isn't too big, so recovery doesn't scan log too much. Recovery make sure raid disks and log disk have the same data of a stripe. If crash happens before 4, recovery might/might not recovery stripe's data/parity depending on if data/parity and its checksum matches. In either case, this doesn't change the syntax of an IO write. After step 3, stripe is guaranteed recoverable, because stripe's data/parity is persistent in log disk. In some cases, log disk content and raid disks content of a stripe are the same, but recovery will still copy log disk content to raid disks, this doesn't impact data consistency. space reuse happens after superblock update and cache flush. There is one situation we want to avoid. A broken meta in the middle of a log causes recovery can't find meta at the head of log. If operations require meta at the head persistent in log, we must make sure meta before it persistent in log too. The case is stripe data/parity is in log and we start write stripe to raid disks (before step 4). stripe data/parity must be persistent in log before we do the write to raid disks. The solution is we restrictly maintain io_unit list order. In this case, we only write stripes of an io_unit to raid disks till the io_unit is the first one whose data/parity is in log. The io_unit list order is important for other cases too. For example, some io_unit are reclaimable and others not. They can be mixed in the list, we shouldn't reuse space of an unreclaimable io_unit. Includes fixes to problems which were... Reported-by: kbuild test robot <fengguang.wu@intel.com> Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: NeilBrown <neilb@suse.com>
2015-08-13 21:32:00 +00:00
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
if (!discard_pending &&
test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
int hash;
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
if (sh->qd_idx >= 0) {
clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
}
/* now that discard is done we can proceed with any sync */
clear_bit(STRIPE_DISCARD, &sh->state);
/*
* SCSI discard will change some bio fields and the stripe has
* no updated data, so remove it from hash list and the stripe
* will be reinitialized
*/
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
unhash:
hash = sh->hash_lock_index;
spin_lock_irq(conf->hash_locks + hash);
remove_hash(sh);
spin_unlock_irq(conf->hash_locks + hash);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head_sh->batch_head) {
sh = list_first_entry(&sh->batch_list,
struct stripe_head, batch_list);
if (sh != head_sh)
goto unhash;
}
sh = head_sh;
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
set_bit(STRIPE_HANDLE, &sh->state);
}
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
if (atomic_dec_and_test(&conf->pending_full_writes))
md_wakeup_thread(conf->mddev->thread);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if (head_sh->batch_head && do_endio)
break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
}
/*
* For RMW in write back cache, we need extra page in prexor to store the
* old data. This page is stored in dev->orig_page.
*
* This function checks whether we have data for prexor. The exact logic
* is:
* R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
*/
static inline bool uptodate_for_rmw(struct r5dev *dev)
{
return (test_bit(R5_UPTODATE, &dev->flags)) &&
(!test_bit(R5_InJournal, &dev->flags) ||
test_bit(R5_OrigPageUPTDODATE, &dev->flags));
}
static int handle_stripe_dirtying(struct r5conf *conf,
struct stripe_head *sh,
struct stripe_head_state *s,
int disks)
{
int rmw = 0, rcw = 0, i;
sector_t recovery_cp = conf->mddev->recovery_cp;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
/* Check whether resync is now happening or should start.
* If yes, then the array is dirty (after unclean shutdown or
* initial creation), so parity in some stripes might be inconsistent.
* In this case, we need to always do reconstruct-write, to ensure
* that in case of drive failure or read-error correction, we
* generate correct data from the parity.
*/
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (conf->rmw_level == PARITY_DISABLE_RMW ||
(recovery_cp < MaxSector && sh->sector >= recovery_cp &&
s->failed == 0)) {
/* Calculate the real rcw later - for now make it
* look like rcw is cheaper
*/
rcw = 1; rmw = 2;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
conf->rmw_level, (unsigned long long)recovery_cp,
(unsigned long long)sh->sector);
} else for (i = disks; i--; ) {
/* would I have to read this buffer for read_modify_write */
struct r5dev *dev = &sh->dev[i];
if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
i == sh->pd_idx || i == sh->qd_idx ||
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
test_bit(R5_InJournal, &dev->flags)) &&
!test_bit(R5_LOCKED, &dev->flags) &&
!(uptodate_for_rmw(dev) ||
test_bit(R5_Wantcompute, &dev->flags))) {
if (test_bit(R5_Insync, &dev->flags))
rmw++;
else
rmw += 2*disks; /* cannot read it */
}
/* Would I have to read this buffer for reconstruct_write */
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (!test_bit(R5_OVERWRITE, &dev->flags) &&
i != sh->pd_idx && i != sh->qd_idx &&
!test_bit(R5_LOCKED, &dev->flags) &&
!(test_bit(R5_UPTODATE, &dev->flags) ||
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
test_bit(R5_Wantcompute, &dev->flags))) {
if (test_bit(R5_Insync, &dev->flags))
rcw++;
else
rcw += 2*disks;
}
}
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
(unsigned long long)sh->sector, sh->state, rmw, rcw);
set_bit(STRIPE_HANDLE, &sh->state);
if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
/* prefer read-modify-write, but need to get some data */
if (conf->mddev->queue)
blk_add_trace_msg(conf->mddev->queue,
"raid5 rmw %llu %d",
(unsigned long long)sh->sector, rmw);
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (test_bit(R5_InJournal, &dev->flags) &&
dev->page == dev->orig_page &&
!test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
/* alloc page for prexor */
struct page *p = alloc_page(GFP_NOIO);
if (p) {
dev->orig_page = p;
continue;
}
/*
* alloc_page() failed, try use
* disk_info->extra_page
*/
if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
&conf->cache_state)) {
r5c_use_extra_page(sh);
break;
}
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
/* extra_page in use, add to delayed_list */
set_bit(STRIPE_DELAYED, &sh->state);
s->waiting_extra_page = 1;
return -EAGAIN;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
}
}
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
i == sh->pd_idx || i == sh->qd_idx ||
test_bit(R5_InJournal, &dev->flags)) &&
!test_bit(R5_LOCKED, &dev->flags) &&
!(uptodate_for_rmw(dev) ||
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
test_bit(R5_Wantcompute, &dev->flags)) &&
test_bit(R5_Insync, &dev->flags)) {
if (test_bit(STRIPE_PREREAD_ACTIVE,
&sh->state)) {
pr_debug("Read_old block %d for r-m-w\n",
i);
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantread, &dev->flags);
s->locked++;
} else
set_bit(STRIPE_DELAYED, &sh->state);
}
}
}
if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
/* want reconstruct write, but need to get some data */
int qread =0;
rcw = 0;
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (!test_bit(R5_OVERWRITE, &dev->flags) &&
i != sh->pd_idx && i != sh->qd_idx &&
!test_bit(R5_LOCKED, &dev->flags) &&
!(test_bit(R5_UPTODATE, &dev->flags) ||
test_bit(R5_Wantcompute, &dev->flags))) {
rcw++;
if (test_bit(R5_Insync, &dev->flags) &&
test_bit(STRIPE_PREREAD_ACTIVE,
&sh->state)) {
pr_debug("Read_old block "
"%d for Reconstruct\n", i);
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantread, &dev->flags);
s->locked++;
qread++;
} else
set_bit(STRIPE_DELAYED, &sh->state);
}
}
if (rcw && conf->mddev->queue)
blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
(unsigned long long)sh->sector,
rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
}
if (rcw > disks && rmw > disks &&
!test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
set_bit(STRIPE_DELAYED, &sh->state);
/* now if nothing is locked, and if we have enough data,
* we can start a write request
*/
/* since handle_stripe can be called at any time we need to handle the
* case where a compute block operation has been submitted and then a
* subsequent call wants to start a write request. raid_run_ops only
* handles the case where compute block and reconstruct are requested
* simultaneously. If this is not the case then new writes need to be
* held off until the compute completes.
*/
if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
(s->locked == 0 && (rcw == 0 || rmw == 0) &&
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
!test_bit(STRIPE_BIT_DELAY, &sh->state)))
schedule_reconstruction(sh, s, rcw == 0, 0);
return 0;
}
static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
struct stripe_head_state *s, int disks)
{
struct r5dev *dev = NULL;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
set_bit(STRIPE_HANDLE, &sh->state);
switch (sh->check_state) {
case check_state_idle:
/* start a new check operation if there are no failures */
if (s->failed == 0) {
BUG_ON(s->uptodate != disks);
sh->check_state = check_state_run;
set_bit(STRIPE_OP_CHECK, &s->ops_request);
clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
s->uptodate--;
break;
}
dev = &sh->dev[s->failed_num[0]];
/* fall through */
case check_state_compute_result:
sh->check_state = check_state_idle;
if (!dev)
dev = &sh->dev[sh->pd_idx];
/* check that a write has not made the stripe insync */
if (test_bit(STRIPE_INSYNC, &sh->state))
break;
/* either failed parity check, or recovery is happening */
BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
BUG_ON(s->uptodate != disks);
set_bit(R5_LOCKED, &dev->flags);
s->locked++;
set_bit(R5_Wantwrite, &dev->flags);
clear_bit(STRIPE_DEGRADED, &sh->state);
set_bit(STRIPE_INSYNC, &sh->state);
break;
case check_state_run:
break; /* we will be called again upon completion */
case check_state_check_result:
sh->check_state = check_state_idle;
/* if a failure occurred during the check operation, leave
* STRIPE_INSYNC not set and let the stripe be handled again
*/
if (s->failed)
break;
/* handle a successful check operation, if parity is correct
* we are done. Otherwise update the mismatch count and repair
* parity if !MD_RECOVERY_CHECK
*/
if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
/* parity is correct (on disc,
* not in buffer any more)
*/
set_bit(STRIPE_INSYNC, &sh->state);
else {
atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
/* don't try to repair!! */
set_bit(STRIPE_INSYNC, &sh->state);
pr_warn_ratelimited("%s: mismatch sector in range "
"%llu-%llu\n", mdname(conf->mddev),
(unsigned long long) sh->sector,
(unsigned long long) sh->sector +
RAID5_STRIPE_SECTORS(conf));
} else {
sh->check_state = check_state_compute_run;
set_bit(STRIPE_COMPUTE_RUN, &sh->state);
set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
set_bit(R5_Wantcompute,
&sh->dev[sh->pd_idx].flags);
sh->ops.target = sh->pd_idx;
sh->ops.target2 = -1;
s->uptodate++;
}
}
break;
case check_state_compute_run:
break;
default:
pr_err("%s: unknown check_state: %d sector: %llu\n",
__func__, sh->check_state,
(unsigned long long) sh->sector);
BUG();
}
}
static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
struct stripe_head_state *s,
int disks)
{
int pd_idx = sh->pd_idx;
int qd_idx = sh->qd_idx;
struct r5dev *dev;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
set_bit(STRIPE_HANDLE, &sh->state);
BUG_ON(s->failed > 2);
/* Want to check and possibly repair P and Q.
* However there could be one 'failed' device, in which
* case we can only check one of them, possibly using the
* other to generate missing data
*/
switch (sh->check_state) {
case check_state_idle:
/* start a new check operation if there are < 2 failures */
if (s->failed == s->q_failed) {
/* The only possible failed device holds Q, so it
* makes sense to check P (If anything else were failed,
* we would have used P to recreate it).
*/
sh->check_state = check_state_run;
}
if (!s->q_failed && s->failed < 2) {
/* Q is not failed, and we didn't use it to generate
* anything, so it makes sense to check it
*/
if (sh->check_state == check_state_run)
sh->check_state = check_state_run_pq;
else
sh->check_state = check_state_run_q;
}
/* discard potentially stale zero_sum_result */
sh->ops.zero_sum_result = 0;
if (sh->check_state == check_state_run) {
/* async_xor_zero_sum destroys the contents of P */
clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
s->uptodate--;
}
if (sh->check_state >= check_state_run &&
sh->check_state <= check_state_run_pq) {
/* async_syndrome_zero_sum preserves P and Q, so
* no need to mark them !uptodate here
*/
set_bit(STRIPE_OP_CHECK, &s->ops_request);
break;
}
/* we have 2-disk failure */
BUG_ON(s->failed != 2);
/* fall through */
case check_state_compute_result:
sh->check_state = check_state_idle;
/* check that a write has not made the stripe insync */
if (test_bit(STRIPE_INSYNC, &sh->state))
break;
/* now write out any block on a failed drive,
* or P or Q if they were recomputed
*/
dev = NULL;
if (s->failed == 2) {
dev = &sh->dev[s->failed_num[1]];
s->locked++;
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantwrite, &dev->flags);
}
if (s->failed >= 1) {
dev = &sh->dev[s->failed_num[0]];
s->locked++;
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantwrite, &dev->flags);
}
if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
dev = &sh->dev[pd_idx];
s->locked++;
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantwrite, &dev->flags);
}
if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
dev = &sh->dev[qd_idx];
s->locked++;
set_bit(R5_LOCKED, &dev->flags);
set_bit(R5_Wantwrite, &dev->flags);
}
if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
"%s: disk%td not up to date\n",
mdname(conf->mddev),
dev - (struct r5dev *) &sh->dev)) {
clear_bit(R5_LOCKED, &dev->flags);
clear_bit(R5_Wantwrite, &dev->flags);
s->locked--;
}
clear_bit(STRIPE_DEGRADED, &sh->state);
set_bit(STRIPE_INSYNC, &sh->state);
break;
case check_state_run:
case check_state_run_q:
case check_state_run_pq:
break; /* we will be called again upon completion */
case check_state_check_result:
sh->check_state = check_state_idle;
/* handle a successful check operation, if parity is correct
* we are done. Otherwise update the mismatch count and repair
* parity if !MD_RECOVERY_CHECK
*/
if (sh->ops.zero_sum_result == 0) {
/* both parities are correct */
if (!s->failed)
set_bit(STRIPE_INSYNC, &sh->state);
else {
/* in contrast to the raid5 case we can validate
* parity, but still have a failure to write
* back
*/
sh->check_state = check_state_compute_result;
/* Returning at this point means that we may go
* off and bring p and/or q uptodate again so
* we make sure to check zero_sum_result again
* to verify if p or q need writeback
*/
}
} else {
atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
/* don't try to repair!! */
set_bit(STRIPE_INSYNC, &sh->state);
pr_warn_ratelimited("%s: mismatch sector in range "
"%llu-%llu\n", mdname(conf->mddev),
(unsigned long long) sh->sector,
(unsigned long long) sh->sector +
RAID5_STRIPE_SECTORS(conf));
} else {
int *target = &sh->ops.target;
sh->ops.target = -1;
sh->ops.target2 = -1;
sh->check_state = check_state_compute_run;
set_bit(STRIPE_COMPUTE_RUN, &sh->state);
set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
set_bit(R5_Wantcompute,
&sh->dev[pd_idx].flags);
*target = pd_idx;
target = &sh->ops.target2;
s->uptodate++;
}
if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
set_bit(R5_Wantcompute,
&sh->dev[qd_idx].flags);
*target = qd_idx;
s->uptodate++;
}
}
}
break;
case check_state_compute_run:
break;
default:
pr_warn("%s: unknown check_state: %d sector: %llu\n",
__func__, sh->check_state,
(unsigned long long) sh->sector);
BUG();
}
}
static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
{
int i;
/* We have read all the blocks in this stripe and now we need to
* copy some of them into a target stripe for expand.
*/
struct dma_async_tx_descriptor *tx = NULL;
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
BUG_ON(sh->batch_head);
clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
for (i = 0; i < sh->disks; i++)
if (i != sh->pd_idx && i != sh->qd_idx) {
int dd_idx, j;
struct stripe_head *sh2;
struct async_submit_ctl submit;
sector_t bn = raid5_compute_blocknr(sh, i, 1);
sector_t s = raid5_compute_sector(conf, bn, 0,
&dd_idx, NULL);
sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
if (sh2 == NULL)
/* so far only the early blocks of this stripe
* have been requested. When later blocks
* get requested, we will try again
*/
continue;
if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
/* must have already done this block */
raid5_release_stripe(sh2);
continue;
}
/* place all the copies on one channel */
init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
tx = async_memcpy(sh2->dev[dd_idx].page,
sh->dev[i].page, 0, 0, RAID5_STRIPE_SIZE(conf),
&submit);
set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
for (j = 0; j < conf->raid_disks; j++)
if (j != sh2->pd_idx &&
j != sh2->qd_idx &&
!test_bit(R5_Expanded, &sh2->dev[j].flags))
break;
if (j == conf->raid_disks) {
set_bit(STRIPE_EXPAND_READY, &sh2->state);
set_bit(STRIPE_HANDLE, &sh2->state);
}
raid5_release_stripe(sh2);
}
/* done submitting copies, wait for them to complete */
async_tx_quiesce(&tx);
}
/*
* handle_stripe - do things to a stripe.
*
* We lock the stripe by setting STRIPE_ACTIVE and then examine the
* state of various bits to see what needs to be done.
* Possible results:
* return some read requests which now have data
* return some write requests which are safely on storage
* schedule a read on some buffers
* schedule a write of some buffers
* return confirmation of parity correctness
*
*/
static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
{
struct r5conf *conf = sh->raid_conf;
int disks = sh->disks;
struct r5dev *dev;
int i;
int do_recovery = 0;
memset(s, 0, sizeof(*s));
s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
s->failed_num[0] = -1;
s->failed_num[1] = -1;
s->log_failed = r5l_log_disk_error(conf);
/* Now to look around and see what can be done */
rcu_read_lock();
for (i=disks; i--; ) {
struct md_rdev *rdev;
sector_t first_bad;
int bad_sectors;
int is_bad = 0;
dev = &sh->dev[i];
pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
i, dev->flags,
dev->toread, dev->towrite, dev->written);
/* maybe we can reply to a read
*
* new wantfill requests are only permitted while
* ops_complete_biofill is guaranteed to be inactive
*/
if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
!test_bit(STRIPE_BIOFILL_RUN, &sh->state))
set_bit(R5_Wantfill, &dev->flags);
/* now count some things */
if (test_bit(R5_LOCKED, &dev->flags))
s->locked++;
if (test_bit(R5_UPTODATE, &dev->flags))
s->uptodate++;
if (test_bit(R5_Wantcompute, &dev->flags)) {
s->compute++;
BUG_ON(s->compute > 2);
}
if (test_bit(R5_Wantfill, &dev->flags))
s->to_fill++;
else if (dev->toread)
s->to_read++;
if (dev->towrite) {
s->to_write++;
if (!test_bit(R5_OVERWRITE, &dev->flags))
s->non_overwrite++;
}
if (dev->written)
s->written++;
/* Prefer to use the replacement for reads, but only
* if it is recovered enough and has no bad blocks.
*/
rdev = rcu_dereference(conf->disks[i].replacement);
if (rdev && !test_bit(Faulty, &rdev->flags) &&
rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
!is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
&first_bad, &bad_sectors))
set_bit(R5_ReadRepl, &dev->flags);
else {
if (rdev && !test_bit(Faulty, &rdev->flags))
set_bit(R5_NeedReplace, &dev->flags);
else
clear_bit(R5_NeedReplace, &dev->flags);
rdev = rcu_dereference(conf->disks[i].rdev);
clear_bit(R5_ReadRepl, &dev->flags);
}
if (rdev && test_bit(Faulty, &rdev->flags))
rdev = NULL;
if (rdev) {
is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
&first_bad, &bad_sectors);
if (s->blocked_rdev == NULL
&& (test_bit(Blocked, &rdev->flags)
|| is_bad < 0)) {
if (is_bad < 0)
set_bit(BlockedBadBlocks,
&rdev->flags);
s->blocked_rdev = rdev;
atomic_inc(&rdev->nr_pending);
}
}
clear_bit(R5_Insync, &dev->flags);
if (!rdev)
/* Not in-sync */;
else if (is_bad) {
/* also not in-sync */
if (!test_bit(WriteErrorSeen, &rdev->flags) &&
test_bit(R5_UPTODATE, &dev->flags)) {
/* treat as in-sync, but with a read error
* which we can now try to correct
*/
set_bit(R5_Insync, &dev->flags);
set_bit(R5_ReadError, &dev->flags);
}
} else if (test_bit(In_sync, &rdev->flags))
set_bit(R5_Insync, &dev->flags);
else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
/* in sync if before recovery_offset */
set_bit(R5_Insync, &dev->flags);
else if (test_bit(R5_UPTODATE, &dev->flags) &&
test_bit(R5_Expanded, &dev->flags))
/* If we've reshaped into here, we assume it is Insync.
* We will shortly update recovery_offset to make
* it official.
*/
set_bit(R5_Insync, &dev->flags);
if (test_bit(R5_WriteError, &dev->flags)) {
/* This flag does not apply to '.replacement'
* only to .rdev, so make sure to check that*/
struct md_rdev *rdev2 = rcu_dereference(
conf->disks[i].rdev);
if (rdev2 == rdev)
clear_bit(R5_Insync, &dev->flags);
if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
s->handle_bad_blocks = 1;
atomic_inc(&rdev2->nr_pending);
} else
clear_bit(R5_WriteError, &dev->flags);
}
if (test_bit(R5_MadeGood, &dev->flags)) {
/* This flag does not apply to '.replacement'
* only to .rdev, so make sure to check that*/
struct md_rdev *rdev2 = rcu_dereference(
conf->disks[i].rdev);
if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
s->handle_bad_blocks = 1;
atomic_inc(&rdev2->nr_pending);
} else
clear_bit(R5_MadeGood, &dev->flags);
}
if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
struct md_rdev *rdev2 = rcu_dereference(
conf->disks[i].replacement);
if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
s->handle_bad_blocks = 1;
atomic_inc(&rdev2->nr_pending);
} else
clear_bit(R5_MadeGoodRepl, &dev->flags);
}
if (!test_bit(R5_Insync, &dev->flags)) {
/* The ReadError flag will just be confusing now */
clear_bit(R5_ReadError, &dev->flags);
clear_bit(R5_ReWrite, &dev->flags);
}
if (test_bit(R5_ReadError, &dev->flags))
clear_bit(R5_Insync, &dev->flags);
if (!test_bit(R5_Insync, &dev->flags)) {
if (s->failed < 2)
s->failed_num[s->failed] = i;
s->failed++;
if (rdev && !test_bit(Faulty, &rdev->flags))
do_recovery = 1;
md/raid5: fix data corruption of replacements after originals dropped During raid5 replacement, the stripes can be marked with R5_NeedReplace flag. Data can be read from being-replaced devices and written to replacing spares without reading all other devices. (It's 'replace' mode. s.replacing = 1) If a being-replaced device is dropped, the replacement progress will be interrupted and resumed with pure recovery mode. However, existing stripes before being interrupted cannot read from the dropped device anymore. It prints lots of WARN_ON messages. And it results in data corruption because existing stripes write problematic data into its replacement device and update the progress. \# Erase disks (1MB + 2GB) dd if=/dev/zero of=/dev/sda bs=1MB count=2049 dd if=/dev/zero of=/dev/sdb bs=1MB count=2049 dd if=/dev/zero of=/dev/sdc bs=1MB count=2049 dd if=/dev/zero of=/dev/sdd bs=1MB count=2049 mdadm -C /dev/md0 -amd -R -l5 -n3 -x0 /dev/sd[abc] -z 2097152 \# Ensure array stores non-zero data dd if=/root/data_4GB.iso of=/dev/md0 bs=1MB \# Start replacement mdadm /dev/md0 -a /dev/sdd mdadm /dev/md0 --replace /dev/sda Then, Hot-plug out /dev/sda during recovery, and wait for recovery done. echo check > /sys/block/md0/md/sync_action cat /sys/block/md0/md/mismatch_cnt # it will be greater than 0. Soon after you hot-plug out /dev/sda, you will see many WARN_ON messages. The replacement recovery will be interrupted shortly. After the recovery finishes, it will result in data corruption. Actually, it's just an unhandled case of replacement. In commit <f94c0b6658c7> (md/raid5: fix interaction of 'replace' and 'recovery'.), if a NeedReplace device is not UPTODATE then that is an error, the commit just simply print WARN_ON but also mark these corrupted stripes with R5_WantReplace. (it means it's ready for writes.) To fix this case, we can leverage 'sync and replace' mode mentioned in commit <9a3e1101b827> (md/raid5: detect and handle replacements during recovery.). We can add logics to detect and use 'sync and replace' mode for these stripes. Reported-by: Alex Chen <alexchen@synology.com> Reviewed-by: Alex Wu <alexwu@synology.com> Reviewed-by: Chung-Chiang Cheng <cccheng@synology.com> Signed-off-by: BingJing Chang <bingjingc@synology.com> Signed-off-by: Shaohua Li <shli@fb.com>
2018-08-01 09:08:36 +00:00
else if (!rdev) {
rdev = rcu_dereference(
conf->disks[i].replacement);
if (rdev && !test_bit(Faulty, &rdev->flags))
do_recovery = 1;
}
}
md/r5cache: State machine for raid5-cache write back mode This patch adds state machine for raid5-cache. With log device, the raid456 array could operate in two different modes (r5c_journal_mode): - write-back (R5C_MODE_WRITE_BACK) - write-through (R5C_MODE_WRITE_THROUGH) Existing code of raid5-cache only has write-through mode. For write-back cache, it is necessary to extend the state machine. With write-back cache, every stripe could operate in two different phases: - caching - writing-out In caching phase, the stripe handles writes as: - write to journal - return IO In writing-out phase, the stripe behaviors as a stripe in write through mode R5C_MODE_WRITE_THROUGH. STRIPE_R5C_CACHING is added to sh->state to differentiate caching and writing-out phase. Please note: this is a "no-op" patch for raid5-cache write-through mode. The following detailed explanation is copied from the raid5-cache.c: /* * raid5 cache state machine * * With rhe RAID cache, each stripe works in two phases: * - caching phase * - writing-out phase * * These two phases are controlled by bit STRIPE_R5C_CACHING: * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase * * When there is no journal, or the journal is in write-through mode, * the stripe is always in writing-out phase. * * For write-back journal, the stripe is sent to caching phase on write * (r5c_handle_stripe_dirtying). r5c_make_stripe_write_out() kicks off * the write-out phase by clearing STRIPE_R5C_CACHING. * * Stripes in caching phase do not write the raid disks. Instead, all * writes are committed from the log device. Therefore, a stripe in * caching phase handles writes as: * - write to log device * - return IO * * Stripes in writing-out phase handle writes as: * - calculate parity * - write pending data and parity to journal * - write data and parity to raid disks * - return IO for pending writes */ Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:38 +00:00
if (test_bit(R5_InJournal, &dev->flags))
s->injournal++;
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (test_bit(R5_InJournal, &dev->flags) && dev->written)
s->just_cached++;
}
if (test_bit(STRIPE_SYNCING, &sh->state)) {
/* If there is a failed device being replaced,
* we must be recovering.
* else if we are after recovery_cp, we must be syncing
* else if MD_RECOVERY_REQUESTED is set, we also are syncing.
* else we can only be replacing
* sync and recovery both need to read all devices, and so
* use the same flag.
*/
if (do_recovery ||
sh->sector >= conf->mddev->recovery_cp ||
test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
s->syncing = 1;
else
s->replacing = 1;
}
rcu_read_unlock();
}
/*
* Return '1' if this is a member of batch, or '0' if it is a lone stripe or
* a head which can now be handled.
*/
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
static int clear_batch_ready(struct stripe_head *sh)
{
struct stripe_head *tmp;
if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
return (sh->batch_head && sh->batch_head != sh);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
spin_lock(&sh->stripe_lock);
if (!sh->batch_head) {
spin_unlock(&sh->stripe_lock);
return 0;
}
/*
* this stripe could be added to a batch list before we check
* BATCH_READY, skips it
*/
if (sh->batch_head != sh) {
spin_unlock(&sh->stripe_lock);
return 1;
}
spin_lock(&sh->batch_lock);
list_for_each_entry(tmp, &sh->batch_list, batch_list)
clear_bit(STRIPE_BATCH_READY, &tmp->state);
spin_unlock(&sh->batch_lock);
spin_unlock(&sh->stripe_lock);
/*
* BATCH_READY is cleared, no new stripes can be added.
* batch_list can be accessed without lock
*/
return 0;
}
static void break_stripe_batch_list(struct stripe_head *head_sh,
unsigned long handle_flags)
{
struct stripe_head *sh, *next;
int i;
int do_wakeup = 0;
list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
list_del_init(&sh->batch_list);
WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
(1 << STRIPE_SYNCING) |
(1 << STRIPE_REPLACED) |
(1 << STRIPE_DELAYED) |
(1 << STRIPE_BIT_DELAY) |
(1 << STRIPE_FULL_WRITE) |
(1 << STRIPE_BIOFILL_RUN) |
(1 << STRIPE_COMPUTE_RUN) |
(1 << STRIPE_DISCARD) |
(1 << STRIPE_BATCH_READY) |
(1 << STRIPE_BATCH_ERR) |
(1 << STRIPE_BITMAP_PENDING)),
"stripe state: %lx\n", sh->state);
WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
(1 << STRIPE_REPLACED)),
"head stripe state: %lx\n", head_sh->state);
set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
md/raid5: preserve STRIPE_PREREAD_ACTIVE in break_stripe_batch_list break_stripe_batch_list breaks up a batch and copies some flags from the batch head to the members, preserving others. It doesn't preserve or copy STRIPE_PREREAD_ACTIVE. This is not normally a problem as STRIPE_PREREAD_ACTIVE is cleared when a stripe_head is added to a batch, and is not set on stripe_heads already in a batch. However there is no locking to ensure one thread doesn't set the flag after it has just been cleared in another. This does occasionally happen. md/raid5 maintains a count of the number of stripe_heads with STRIPE_PREREAD_ACTIVE set: conf->preread_active_stripes. When break_stripe_batch_list clears STRIPE_PREREAD_ACTIVE inadvertently this could becomes incorrect and will never again return to zero. md/raid5 delays the handling of some stripe_heads until preread_active_stripes becomes zero. So when the above mention race happens, those stripe_heads become blocked and never progress, resulting is write to the array handing. So: change break_stripe_batch_list to preserve STRIPE_PREREAD_ACTIVE in the members of a batch. URL: https://bugzilla.kernel.org/show_bug.cgi?id=108741 URL: https://bugzilla.redhat.com/show_bug.cgi?id=1258153 URL: http://thread.gmane.org/5649C0E9.2030204@zoner.cz Reported-by: Martin Svec <martin.svec@zoner.cz> (and others) Tested-by: Tom Weber <linux@junkyard.4t2.com> Fixes: 1b956f7a8f9a ("md/raid5: be more selective about distributing flags across batch.") Cc: stable@vger.kernel.org (v4.1 and later) Signed-off-by: NeilBrown <neilb@suse.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-03-09 01:58:25 +00:00
(1 << STRIPE_PREREAD_ACTIVE) |
(1 << STRIPE_DEGRADED) |
(1 << STRIPE_ON_UNPLUG_LIST)),
head_sh->state & (1 << STRIPE_INSYNC));
sh->check_state = head_sh->check_state;
sh->reconstruct_state = head_sh->reconstruct_state;
spin_lock_irq(&sh->stripe_lock);
sh->batch_head = NULL;
spin_unlock_irq(&sh->stripe_lock);
for (i = 0; i < sh->disks; i++) {
if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
do_wakeup = 1;
sh->dev[i].flags = head_sh->dev[i].flags &
(~((1 << R5_WriteError) | (1 << R5_Overlap)));
}
if (handle_flags == 0 ||
sh->state & handle_flags)
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
}
spin_lock_irq(&head_sh->stripe_lock);
head_sh->batch_head = NULL;
spin_unlock_irq(&head_sh->stripe_lock);
for (i = 0; i < head_sh->disks; i++)
if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
do_wakeup = 1;
if (head_sh->state & handle_flags)
set_bit(STRIPE_HANDLE, &head_sh->state);
if (do_wakeup)
wake_up(&head_sh->raid_conf->wait_for_overlap);
}
static void handle_stripe(struct stripe_head *sh)
{
struct stripe_head_state s;
struct r5conf *conf = sh->raid_conf;
int i;
int prexor;
int disks = sh->disks;
struct r5dev *pdev, *qdev;
clear_bit(STRIPE_HANDLE, &sh->state);
raid5: call clear_batch_ready before set STRIPE_ACTIVE We tried to only put the head sh of batch list to handle_list, then the handle_stripe doesn't handle other members in the batch list. However, we still got the calltrace in break_stripe_batch_list. [593764.644269] stripe state: 2003 kernel: [593764.644299] ------------[ cut here ]------------ kernel: [593764.644308] WARNING: CPU: 12 PID: 856 at drivers/md/raid5.c:4625 break_stripe_batch_list+0x203/0x240 [raid456] [...] kernel: [593764.644363] Call Trace: kernel: [593764.644370] handle_stripe+0x907/0x20c0 [raid456] kernel: [593764.644376] ? __wake_up_common_lock+0x89/0xc0 kernel: [593764.644379] handle_active_stripes.isra.57+0x35f/0x570 [raid456] kernel: [593764.644382] ? raid5_wakeup_stripe_thread+0x96/0x1f0 [raid456] kernel: [593764.644385] raid5d+0x480/0x6a0 [raid456] kernel: [593764.644390] ? md_thread+0x11f/0x160 kernel: [593764.644392] md_thread+0x11f/0x160 kernel: [593764.644394] ? wait_woken+0x80/0x80 kernel: [593764.644396] kthread+0xfc/0x130 kernel: [593764.644398] ? find_pers+0x70/0x70 kernel: [593764.644399] ? kthread_create_on_node+0x70/0x70 kernel: [593764.644401] ret_from_fork+0x1f/0x30 As we can see, the stripe was set with STRIPE_ACTIVE and STRIPE_HANDLE, and only handle_stripe could set those flags then return. And since the stipe was already in the batch list, we need to return earlier before set the two flags. And after dig a little about git history especially commit 3664847d95e6 ("md/raid5: fix a race condition in stripe batch"), it seems the batched stipe still could be handled by handle_stipe, then handle_stipe needs to return earlier if clear_batch_ready to return true. Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com> Signed-off-by: Song Liu <songliubraving@fb.com>
2020-06-16 09:25:50 +00:00
/*
* handle_stripe should not continue handle the batched stripe, only
* the head of batch list or lone stripe can continue. Otherwise we
* could see break_stripe_batch_list warns about the STRIPE_ACTIVE
* is set for the batched stripe.
*/
if (clear_batch_ready(sh))
return;
if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
/* already being handled, ensure it gets handled
* again when current action finishes */
set_bit(STRIPE_HANDLE, &sh->state);
return;
}
if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
break_stripe_batch_list(sh, 0);
if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
spin_lock(&sh->stripe_lock);
/*
* Cannot process 'sync' concurrently with 'discard'.
* Flush data in r5cache before 'sync'.
*/
if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
!test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
!test_bit(STRIPE_DISCARD, &sh->state) &&
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
set_bit(STRIPE_SYNCING, &sh->state);
clear_bit(STRIPE_INSYNC, &sh->state);
md/raid5: fix interaction of 'replace' and 'recovery'. If a device in a RAID4/5/6 is being replaced while another is being recovered, then the writes to the replacement device currently don't happen, resulting in corruption when the replacement completes and the new drive takes over. This is because the replacement writes are only triggered when 's.replacing' is set and not when the similar 's.sync' is set (which is the case during resync and recovery - it means all devices need to be read). So schedule those writes when s.replacing is set as well. In this case we cannot use "STRIPE_INSYNC" to record that the replacement has happened as that is needed for recording that any parity calculation is complete. So introduce STRIPE_REPLACED to record if the replacement has happened. For safety we should also check that STRIPE_COMPUTE_RUN is not set. This has a similar effect to the "s.locked == 0" test. The latter ensure that now IO has been flagged but not started. The former checks if any parity calculation has been flagged by not started. We must wait for both of these to complete before triggering the 'replace'. Add a similar test to the subsequent check for "are we finished yet". This possibly isn't needed (is subsumed in the STRIPE_INSYNC test), but it makes it more obvious that the REPLACE will happen before we think we are finished. Finally if a NeedReplace device is not UPTODATE then that is an error. We really must trigger a warning. This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2 (md/raid5: detect and handle replacements during recovery.) which introduced replacement for raid5. That was in 3.3-rc3, so any stable kernel since then would benefit from this fix. Cc: stable@vger.kernel.org (3.3+) Reported-by: qindehua <13691222965@163.com> Tested-by: qindehua <qindehua@163.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-07-22 02:57:21 +00:00
clear_bit(STRIPE_REPLACED, &sh->state);
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
}
spin_unlock(&sh->stripe_lock);
}
clear_bit(STRIPE_DELAYED, &sh->state);
pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
(unsigned long long)sh->sector, sh->state,
atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
sh->check_state, sh->reconstruct_state);
analyse_stripe(sh, &s);
if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
goto finish;
if (s.handle_bad_blocks ||
test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
set_bit(STRIPE_HANDLE, &sh->state);
goto finish;
}
if (unlikely(s.blocked_rdev)) {
if (s.syncing || s.expanding || s.expanded ||
s.replacing || s.to_write || s.written) {
set_bit(STRIPE_HANDLE, &sh->state);
goto finish;
}
/* There is nothing for the blocked_rdev to block */
rdev_dec_pending(s.blocked_rdev, conf->mddev);
s.blocked_rdev = NULL;
}
if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
set_bit(STRIPE_BIOFILL_RUN, &sh->state);
}
pr_debug("locked=%d uptodate=%d to_read=%d"
" to_write=%d failed=%d failed_num=%d,%d\n",
s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
s.failed_num[0], s.failed_num[1]);
/*
* check if the array has lost more than max_degraded devices and,
* if so, some requests might need to be failed.
*
* When journal device failed (log_failed), we will only process
* the stripe if there is data need write to raid disks
*/
if (s.failed > conf->max_degraded ||
(s.log_failed && s.injournal == 0)) {
sh->check_state = 0;
sh->reconstruct_state = 0;
break_stripe_batch_list(sh, 0);
if (s.to_read+s.to_write+s.written)
handle_failed_stripe(conf, sh, &s, disks);
if (s.syncing + s.replacing)
handle_failed_sync(conf, sh, &s);
}
/* Now we check to see if any write operations have recently
* completed
*/
prexor = 0;
if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
prexor = 1;
if (sh->reconstruct_state == reconstruct_state_drain_result ||
sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
sh->reconstruct_state = reconstruct_state_idle;
/* All the 'written' buffers and the parity block are ready to
* be written back to disk
*/
BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
!test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
BUG_ON(sh->qd_idx >= 0 &&
!test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
!test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
for (i = disks; i--; ) {
struct r5dev *dev = &sh->dev[i];
if (test_bit(R5_LOCKED, &dev->flags) &&
(i == sh->pd_idx || i == sh->qd_idx ||
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
dev->written || test_bit(R5_InJournal,
&dev->flags))) {
pr_debug("Writing block %d\n", i);
set_bit(R5_Wantwrite, &dev->flags);
if (prexor)
continue;
if (s.failed > 1)
continue;
if (!test_bit(R5_Insync, &dev->flags) ||
((i == sh->pd_idx || i == sh->qd_idx) &&
s.failed == 0))
set_bit(STRIPE_INSYNC, &sh->state);
}
}
if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
s.dec_preread_active = 1;
}
/*
* might be able to return some write requests if the parity blocks
* are safe, or on a failed drive
*/
pdev = &sh->dev[sh->pd_idx];
s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
qdev = &sh->dev[sh->qd_idx];
s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
|| conf->level < 6;
if (s.written &&
(s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
&& !test_bit(R5_LOCKED, &pdev->flags)
&& (test_bit(R5_UPTODATE, &pdev->flags) ||
test_bit(R5_Discard, &pdev->flags))))) &&
(s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
&& !test_bit(R5_LOCKED, &qdev->flags)
&& (test_bit(R5_UPTODATE, &qdev->flags) ||
test_bit(R5_Discard, &qdev->flags))))))
handle_stripe_clean_event(conf, sh, disks);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
if (s.just_cached)
r5c_handle_cached_data_endio(conf, sh, disks);
log_stripe_write_finished(sh);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
/* Now we might consider reading some blocks, either to check/generate
* parity, or to satisfy requests
* or to load a block that is being partially written.
*/
if (s.to_read || s.non_overwrite
|| (s.to_write && s.failed)
|| (s.syncing && (s.uptodate + s.compute < disks))
|| s.replacing
|| s.expanding)
handle_stripe_fill(sh, &s, disks);
md/r5cache: State machine for raid5-cache write back mode This patch adds state machine for raid5-cache. With log device, the raid456 array could operate in two different modes (r5c_journal_mode): - write-back (R5C_MODE_WRITE_BACK) - write-through (R5C_MODE_WRITE_THROUGH) Existing code of raid5-cache only has write-through mode. For write-back cache, it is necessary to extend the state machine. With write-back cache, every stripe could operate in two different phases: - caching - writing-out In caching phase, the stripe handles writes as: - write to journal - return IO In writing-out phase, the stripe behaviors as a stripe in write through mode R5C_MODE_WRITE_THROUGH. STRIPE_R5C_CACHING is added to sh->state to differentiate caching and writing-out phase. Please note: this is a "no-op" patch for raid5-cache write-through mode. The following detailed explanation is copied from the raid5-cache.c: /* * raid5 cache state machine * * With rhe RAID cache, each stripe works in two phases: * - caching phase * - writing-out phase * * These two phases are controlled by bit STRIPE_R5C_CACHING: * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase * * When there is no journal, or the journal is in write-through mode, * the stripe is always in writing-out phase. * * For write-back journal, the stripe is sent to caching phase on write * (r5c_handle_stripe_dirtying). r5c_make_stripe_write_out() kicks off * the write-out phase by clearing STRIPE_R5C_CACHING. * * Stripes in caching phase do not write the raid disks. Instead, all * writes are committed from the log device. Therefore, a stripe in * caching phase handles writes as: * - write to log device * - return IO * * Stripes in writing-out phase handle writes as: * - calculate parity * - write pending data and parity to journal * - write data and parity to raid disks * - return IO for pending writes */ Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:38 +00:00
/*
* When the stripe finishes full journal write cycle (write to journal
* and raid disk), this is the clean up procedure so it is ready for
* next operation.
*/
r5c_finish_stripe_write_out(conf, sh, &s);
/*
* Now to consider new write requests, cache write back and what else,
* if anything should be read. We do not handle new writes when:
* 1/ A 'write' operation (copy+xor) is already in flight.
* 2/ A 'check' operation is in flight, as it may clobber the parity
* block.
md/r5cache: State machine for raid5-cache write back mode This patch adds state machine for raid5-cache. With log device, the raid456 array could operate in two different modes (r5c_journal_mode): - write-back (R5C_MODE_WRITE_BACK) - write-through (R5C_MODE_WRITE_THROUGH) Existing code of raid5-cache only has write-through mode. For write-back cache, it is necessary to extend the state machine. With write-back cache, every stripe could operate in two different phases: - caching - writing-out In caching phase, the stripe handles writes as: - write to journal - return IO In writing-out phase, the stripe behaviors as a stripe in write through mode R5C_MODE_WRITE_THROUGH. STRIPE_R5C_CACHING is added to sh->state to differentiate caching and writing-out phase. Please note: this is a "no-op" patch for raid5-cache write-through mode. The following detailed explanation is copied from the raid5-cache.c: /* * raid5 cache state machine * * With rhe RAID cache, each stripe works in two phases: * - caching phase * - writing-out phase * * These two phases are controlled by bit STRIPE_R5C_CACHING: * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase * * When there is no journal, or the journal is in write-through mode, * the stripe is always in writing-out phase. * * For write-back journal, the stripe is sent to caching phase on write * (r5c_handle_stripe_dirtying). r5c_make_stripe_write_out() kicks off * the write-out phase by clearing STRIPE_R5C_CACHING. * * Stripes in caching phase do not write the raid disks. Instead, all * writes are committed from the log device. Therefore, a stripe in * caching phase handles writes as: * - write to log device * - return IO * * Stripes in writing-out phase handle writes as: * - calculate parity * - write pending data and parity to journal * - write data and parity to raid disks * - return IO for pending writes */ Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:38 +00:00
* 3/ A r5c cache log write is in flight.
*/
md/r5cache: State machine for raid5-cache write back mode This patch adds state machine for raid5-cache. With log device, the raid456 array could operate in two different modes (r5c_journal_mode): - write-back (R5C_MODE_WRITE_BACK) - write-through (R5C_MODE_WRITE_THROUGH) Existing code of raid5-cache only has write-through mode. For write-back cache, it is necessary to extend the state machine. With write-back cache, every stripe could operate in two different phases: - caching - writing-out In caching phase, the stripe handles writes as: - write to journal - return IO In writing-out phase, the stripe behaviors as a stripe in write through mode R5C_MODE_WRITE_THROUGH. STRIPE_R5C_CACHING is added to sh->state to differentiate caching and writing-out phase. Please note: this is a "no-op" patch for raid5-cache write-through mode. The following detailed explanation is copied from the raid5-cache.c: /* * raid5 cache state machine * * With rhe RAID cache, each stripe works in two phases: * - caching phase * - writing-out phase * * These two phases are controlled by bit STRIPE_R5C_CACHING: * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase * * When there is no journal, or the journal is in write-through mode, * the stripe is always in writing-out phase. * * For write-back journal, the stripe is sent to caching phase on write * (r5c_handle_stripe_dirtying). r5c_make_stripe_write_out() kicks off * the write-out phase by clearing STRIPE_R5C_CACHING. * * Stripes in caching phase do not write the raid disks. Instead, all * writes are committed from the log device. Therefore, a stripe in * caching phase handles writes as: * - write to log device * - return IO * * Stripes in writing-out phase handle writes as: * - calculate parity * - write pending data and parity to journal * - write data and parity to raid disks * - return IO for pending writes */ Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:38 +00:00
if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
if (!r5c_is_writeback(conf->log)) {
if (s.to_write)
handle_stripe_dirtying(conf, sh, &s, disks);
} else { /* write back cache */
int ret = 0;
/* First, try handle writes in caching phase */
if (s.to_write)
ret = r5c_try_caching_write(conf, sh, &s,
disks);
/*
* If caching phase failed: ret == -EAGAIN
* OR
* stripe under reclaim: !caching && injournal
*
* fall back to handle_stripe_dirtying()
*/
if (ret == -EAGAIN ||
/* stripe under reclaim: !caching && injournal */
(!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
s.injournal > 0)) {
ret = handle_stripe_dirtying(conf, sh, &s,
disks);
if (ret == -EAGAIN)
goto finish;
}
md/r5cache: State machine for raid5-cache write back mode This patch adds state machine for raid5-cache. With log device, the raid456 array could operate in two different modes (r5c_journal_mode): - write-back (R5C_MODE_WRITE_BACK) - write-through (R5C_MODE_WRITE_THROUGH) Existing code of raid5-cache only has write-through mode. For write-back cache, it is necessary to extend the state machine. With write-back cache, every stripe could operate in two different phases: - caching - writing-out In caching phase, the stripe handles writes as: - write to journal - return IO In writing-out phase, the stripe behaviors as a stripe in write through mode R5C_MODE_WRITE_THROUGH. STRIPE_R5C_CACHING is added to sh->state to differentiate caching and writing-out phase. Please note: this is a "no-op" patch for raid5-cache write-through mode. The following detailed explanation is copied from the raid5-cache.c: /* * raid5 cache state machine * * With rhe RAID cache, each stripe works in two phases: * - caching phase * - writing-out phase * * These two phases are controlled by bit STRIPE_R5C_CACHING: * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase * * When there is no journal, or the journal is in write-through mode, * the stripe is always in writing-out phase. * * For write-back journal, the stripe is sent to caching phase on write * (r5c_handle_stripe_dirtying). r5c_make_stripe_write_out() kicks off * the write-out phase by clearing STRIPE_R5C_CACHING. * * Stripes in caching phase do not write the raid disks. Instead, all * writes are committed from the log device. Therefore, a stripe in * caching phase handles writes as: * - write to log device * - return IO * * Stripes in writing-out phase handle writes as: * - calculate parity * - write pending data and parity to journal * - write data and parity to raid disks * - return IO for pending writes */ Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:38 +00:00
}
}
/* maybe we need to check and possibly fix the parity for this stripe
* Any reads will already have been scheduled, so we just see if enough
* data is available. The parity check is held off while parity
* dependent operations are in flight.
*/
if (sh->check_state ||
(s.syncing && s.locked == 0 &&
!test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
!test_bit(STRIPE_INSYNC, &sh->state))) {
if (conf->level == 6)
handle_parity_checks6(conf, sh, &s, disks);
else
handle_parity_checks5(conf, sh, &s, disks);
}
md/raid5: fix interaction of 'replace' and 'recovery'. If a device in a RAID4/5/6 is being replaced while another is being recovered, then the writes to the replacement device currently don't happen, resulting in corruption when the replacement completes and the new drive takes over. This is because the replacement writes are only triggered when 's.replacing' is set and not when the similar 's.sync' is set (which is the case during resync and recovery - it means all devices need to be read). So schedule those writes when s.replacing is set as well. In this case we cannot use "STRIPE_INSYNC" to record that the replacement has happened as that is needed for recording that any parity calculation is complete. So introduce STRIPE_REPLACED to record if the replacement has happened. For safety we should also check that STRIPE_COMPUTE_RUN is not set. This has a similar effect to the "s.locked == 0" test. The latter ensure that now IO has been flagged but not started. The former checks if any parity calculation has been flagged by not started. We must wait for both of these to complete before triggering the 'replace'. Add a similar test to the subsequent check for "are we finished yet". This possibly isn't needed (is subsumed in the STRIPE_INSYNC test), but it makes it more obvious that the REPLACE will happen before we think we are finished. Finally if a NeedReplace device is not UPTODATE then that is an error. We really must trigger a warning. This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2 (md/raid5: detect and handle replacements during recovery.) which introduced replacement for raid5. That was in 3.3-rc3, so any stable kernel since then would benefit from this fix. Cc: stable@vger.kernel.org (3.3+) Reported-by: qindehua <13691222965@163.com> Tested-by: qindehua <qindehua@163.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-07-22 02:57:21 +00:00
if ((s.replacing || s.syncing) && s.locked == 0
&& !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
&& !test_bit(STRIPE_REPLACED, &sh->state)) {
/* Write out to replacement devices where possible */
for (i = 0; i < conf->raid_disks; i++)
md/raid5: fix interaction of 'replace' and 'recovery'. If a device in a RAID4/5/6 is being replaced while another is being recovered, then the writes to the replacement device currently don't happen, resulting in corruption when the replacement completes and the new drive takes over. This is because the replacement writes are only triggered when 's.replacing' is set and not when the similar 's.sync' is set (which is the case during resync and recovery - it means all devices need to be read). So schedule those writes when s.replacing is set as well. In this case we cannot use "STRIPE_INSYNC" to record that the replacement has happened as that is needed for recording that any parity calculation is complete. So introduce STRIPE_REPLACED to record if the replacement has happened. For safety we should also check that STRIPE_COMPUTE_RUN is not set. This has a similar effect to the "s.locked == 0" test. The latter ensure that now IO has been flagged but not started. The former checks if any parity calculation has been flagged by not started. We must wait for both of these to complete before triggering the 'replace'. Add a similar test to the subsequent check for "are we finished yet". This possibly isn't needed (is subsumed in the STRIPE_INSYNC test), but it makes it more obvious that the REPLACE will happen before we think we are finished. Finally if a NeedReplace device is not UPTODATE then that is an error. We really must trigger a warning. This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2 (md/raid5: detect and handle replacements during recovery.) which introduced replacement for raid5. That was in 3.3-rc3, so any stable kernel since then would benefit from this fix. Cc: stable@vger.kernel.org (3.3+) Reported-by: qindehua <13691222965@163.com> Tested-by: qindehua <qindehua@163.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-07-22 02:57:21 +00:00
if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
set_bit(R5_WantReplace, &sh->dev[i].flags);
set_bit(R5_LOCKED, &sh->dev[i].flags);
s.locked++;
}
md/raid5: fix interaction of 'replace' and 'recovery'. If a device in a RAID4/5/6 is being replaced while another is being recovered, then the writes to the replacement device currently don't happen, resulting in corruption when the replacement completes and the new drive takes over. This is because the replacement writes are only triggered when 's.replacing' is set and not when the similar 's.sync' is set (which is the case during resync and recovery - it means all devices need to be read). So schedule those writes when s.replacing is set as well. In this case we cannot use "STRIPE_INSYNC" to record that the replacement has happened as that is needed for recording that any parity calculation is complete. So introduce STRIPE_REPLACED to record if the replacement has happened. For safety we should also check that STRIPE_COMPUTE_RUN is not set. This has a similar effect to the "s.locked == 0" test. The latter ensure that now IO has been flagged but not started. The former checks if any parity calculation has been flagged by not started. We must wait for both of these to complete before triggering the 'replace'. Add a similar test to the subsequent check for "are we finished yet". This possibly isn't needed (is subsumed in the STRIPE_INSYNC test), but it makes it more obvious that the REPLACE will happen before we think we are finished. Finally if a NeedReplace device is not UPTODATE then that is an error. We really must trigger a warning. This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2 (md/raid5: detect and handle replacements during recovery.) which introduced replacement for raid5. That was in 3.3-rc3, so any stable kernel since then would benefit from this fix. Cc: stable@vger.kernel.org (3.3+) Reported-by: qindehua <13691222965@163.com> Tested-by: qindehua <qindehua@163.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-07-22 02:57:21 +00:00
if (s.replacing)
set_bit(STRIPE_INSYNC, &sh->state);
set_bit(STRIPE_REPLACED, &sh->state);
}
if ((s.syncing || s.replacing) && s.locked == 0 &&
md/raid5: fix interaction of 'replace' and 'recovery'. If a device in a RAID4/5/6 is being replaced while another is being recovered, then the writes to the replacement device currently don't happen, resulting in corruption when the replacement completes and the new drive takes over. This is because the replacement writes are only triggered when 's.replacing' is set and not when the similar 's.sync' is set (which is the case during resync and recovery - it means all devices need to be read). So schedule those writes when s.replacing is set as well. In this case we cannot use "STRIPE_INSYNC" to record that the replacement has happened as that is needed for recording that any parity calculation is complete. So introduce STRIPE_REPLACED to record if the replacement has happened. For safety we should also check that STRIPE_COMPUTE_RUN is not set. This has a similar effect to the "s.locked == 0" test. The latter ensure that now IO has been flagged but not started. The former checks if any parity calculation has been flagged by not started. We must wait for both of these to complete before triggering the 'replace'. Add a similar test to the subsequent check for "are we finished yet". This possibly isn't needed (is subsumed in the STRIPE_INSYNC test), but it makes it more obvious that the REPLACE will happen before we think we are finished. Finally if a NeedReplace device is not UPTODATE then that is an error. We really must trigger a warning. This bug was introduced in commit 9a3e1101b827a59ac9036a672f5fa8d5279d0fe2 (md/raid5: detect and handle replacements during recovery.) which introduced replacement for raid5. That was in 3.3-rc3, so any stable kernel since then would benefit from this fix. Cc: stable@vger.kernel.org (3.3+) Reported-by: qindehua <13691222965@163.com> Tested-by: qindehua <qindehua@163.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-07-22 02:57:21 +00:00
!test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
test_bit(STRIPE_INSYNC, &sh->state)) {
md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
clear_bit(STRIPE_SYNCING, &sh->state);
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
wake_up(&conf->wait_for_overlap);
}
/* If the failed drives are just a ReadError, then we might need
* to progress the repair/check process
*/
if (s.failed <= conf->max_degraded && !conf->mddev->ro)
for (i = 0; i < s.failed; i++) {
struct r5dev *dev = &sh->dev[s.failed_num[i]];
if (test_bit(R5_ReadError, &dev->flags)
&& !test_bit(R5_LOCKED, &dev->flags)
&& test_bit(R5_UPTODATE, &dev->flags)
) {
if (!test_bit(R5_ReWrite, &dev->flags)) {
set_bit(R5_Wantwrite, &dev->flags);
set_bit(R5_ReWrite, &dev->flags);
} else
/* let's read it back */
set_bit(R5_Wantread, &dev->flags);
set_bit(R5_LOCKED, &dev->flags);
s.locked++;
}
}
/* Finish reconstruct operations initiated by the expansion process */
if (sh->reconstruct_state == reconstruct_state_result) {
struct stripe_head *sh_src
= raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
/* sh cannot be written until sh_src has been read.
* so arrange for sh to be delayed a little
*/
set_bit(STRIPE_DELAYED, &sh->state);
set_bit(STRIPE_HANDLE, &sh->state);
if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
&sh_src->state))
atomic_inc(&conf->preread_active_stripes);
raid5_release_stripe(sh_src);
goto finish;
}
if (sh_src)
raid5_release_stripe(sh_src);
sh->reconstruct_state = reconstruct_state_idle;
clear_bit(STRIPE_EXPANDING, &sh->state);
for (i = conf->raid_disks; i--; ) {
set_bit(R5_Wantwrite, &sh->dev[i].flags);
set_bit(R5_LOCKED, &sh->dev[i].flags);
s.locked++;
}
}
if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
!sh->reconstruct_state) {
/* Need to write out all blocks after computing parity */
sh->disks = conf->raid_disks;
stripe_set_idx(sh->sector, conf, 0, sh);
schedule_reconstruction(sh, &s, 1, 1);
} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
clear_bit(STRIPE_EXPAND_READY, &sh->state);
atomic_dec(&conf->reshape_stripes);
wake_up(&conf->wait_for_overlap);
md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
}
if (s.expanding && s.locked == 0 &&
!test_bit(STRIPE_COMPUTE_RUN, &sh->state))
handle_stripe_expansion(conf, sh);
finish:
/* wait for this device to become unblocked */
if (unlikely(s.blocked_rdev)) {
if (conf->mddev->external)
md_wait_for_blocked_rdev(s.blocked_rdev,
conf->mddev);
else
/* Internal metadata will immediately
* be written by raid5d, so we don't
* need to wait here.
*/
rdev_dec_pending(s.blocked_rdev,
conf->mddev);
}
if (s.handle_bad_blocks)
for (i = disks; i--; ) {
struct md_rdev *rdev;
struct r5dev *dev = &sh->dev[i];
if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
/* We own a safe reference to the rdev */
rdev = conf->disks[i].rdev;
if (!rdev_set_badblocks(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0))
md_error(conf->mddev, rdev);
rdev_dec_pending(rdev, conf->mddev);
}
if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
rdev = conf->disks[i].rdev;
rdev_clear_badblocks(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0);
rdev_dec_pending(rdev, conf->mddev);
}
if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
rdev = conf->disks[i].replacement;
if (!rdev)
/* rdev have been moved down */
rdev = conf->disks[i].rdev;
rdev_clear_badblocks(rdev, sh->sector,
RAID5_STRIPE_SECTORS(conf), 0);
rdev_dec_pending(rdev, conf->mddev);
}
}
if (s.ops_request)
raid_run_ops(sh, s.ops_request);
ops_run_io(sh, &s);
if (s.dec_preread_active) {
/* We delay this until after ops_run_io so that if make_request
2010-09-03 09:56:18 +00:00
* is waiting on a flush, it won't continue until the writes
* have actually been submitted.
*/
atomic_dec(&conf->preread_active_stripes);
if (atomic_read(&conf->preread_active_stripes) <
IO_THRESHOLD)
md_wakeup_thread(conf->mddev->thread);
}
clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
}
static void raid5_activate_delayed(struct r5conf *conf)
{
if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
while (!list_empty(&conf->delayed_list)) {
struct list_head *l = conf->delayed_list.next;
struct stripe_head *sh;
sh = list_entry(l, struct stripe_head, lru);
list_del_init(l);
clear_bit(STRIPE_DELAYED, &sh->state);
if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
atomic_inc(&conf->preread_active_stripes);
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
list_add_tail(&sh->lru, &conf->hold_list);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
raid5_wakeup_stripe_thread(sh);
}
}
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
static void activate_bit_delay(struct r5conf *conf,
struct list_head *temp_inactive_list)
{
/* device_lock is held */
struct list_head head;
list_add(&head, &conf->bitmap_list);
list_del_init(&conf->bitmap_list);
while (!list_empty(&head)) {
struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int hash;
list_del_init(&sh->lru);
atomic_inc(&sh->count);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
hash = sh->hash_lock_index;
__release_stripe(conf, sh, &temp_inactive_list[hash]);
}
}
static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
{
struct r5conf *conf = mddev->private;
sector_t sector = bio->bi_iter.bi_sector;
unsigned int chunk_sectors;
unsigned int bio_sectors = bio_sectors(bio);
WARN_ON_ONCE(bio->bi_partno);
chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
return chunk_sectors >=
((sector & (chunk_sectors - 1)) + bio_sectors);
}
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
/*
* add bio to the retry LIFO ( in O(1) ... we are in interrupt )
* later sampled by raid5d.
*/
static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
{
unsigned long flags;
spin_lock_irqsave(&conf->device_lock, flags);
bi->bi_next = conf->retry_read_aligned_list;
conf->retry_read_aligned_list = bi;
spin_unlock_irqrestore(&conf->device_lock, flags);
md_wakeup_thread(conf->mddev->thread);
}
static struct bio *remove_bio_from_retry(struct r5conf *conf,
unsigned int *offset)
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
{
struct bio *bi;
bi = conf->retry_read_aligned;
if (bi) {
*offset = conf->retry_read_offset;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
conf->retry_read_aligned = NULL;
return bi;
}
bi = conf->retry_read_aligned_list;
if(bi) {
conf->retry_read_aligned_list = bi->bi_next;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
bi->bi_next = NULL;
*offset = 0;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
}
return bi;
}
/*
* The "raid5_align_endio" should check if the read succeeded and if it
* did, call bio_endio on the original bio (having bio_put the new bio
* first).
* If the read failed..
*/
static void raid5_align_endio(struct bio *bi)
{
struct bio* raid_bi = bi->bi_private;
struct mddev *mddev;
struct r5conf *conf;
struct md_rdev *rdev;
blk_status_t error = bi->bi_status;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
bio_put(bi);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
rdev = (void*)raid_bi->bi_next;
raid_bi->bi_next = NULL;
mddev = rdev->mddev;
conf = mddev->private;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
rdev_dec_pending(rdev, conf->mddev);
block: don't access bio->bi_error after bio_put() Commit 4246a0b6 ("block: add a bi_error field to struct bio") has added a few dereferences of 'bio' after a call to bio_put(). This causes use-after-frees such as: [521120.719695] BUG: KASan: use after free in dio_bio_complete+0x2b3/0x320 at addr ffff880f36b38714 [521120.720638] Read of size 4 by task mount.ocfs2/9644 [521120.721212] ============================================================================= [521120.722056] BUG kmalloc-256 (Not tainted): kasan: bad access detected [521120.722968] ----------------------------------------------------------------------------- [521120.722968] [521120.723915] Disabling lock debugging due to kernel taint [521120.724539] INFO: Slab 0xffffea003cdace00 objects=32 used=25 fp=0xffff880f36b38600 flags=0x46fffff80004080 [521120.726037] INFO: Object 0xffff880f36b38700 @offset=1792 fp=0xffff880f36b38800 [521120.726037] [521120.726974] Bytes b4 ffff880f36b386f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.727898] Object ffff880f36b38700: 00 88 b3 36 0f 88 ff ff 00 00 d8 de 0b 88 ff ff ...6............ [521120.728822] Object ffff880f36b38710: 02 00 00 f0 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.729705] Object ffff880f36b38720: 01 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00 ................ [521120.730623] Object ffff880f36b38730: 00 00 00 00 00 00 00 00 01 00 00 00 00 02 00 00 ................ [521120.731621] Object ffff880f36b38740: 00 02 00 00 01 00 00 00 d0 f7 87 ad ff ff ff ff ................ [521120.732776] Object ffff880f36b38750: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.733640] Object ffff880f36b38760: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.734508] Object ffff880f36b38770: 01 00 03 00 01 00 00 00 88 87 b3 36 0f 88 ff ff ...........6.... [521120.735385] Object ffff880f36b38780: 00 73 22 ad 02 88 ff ff 40 13 e0 3c 00 ea ff ff .s".....@..<.... [521120.736667] Object ffff880f36b38790: 00 02 00 00 00 04 00 00 00 00 00 00 00 00 00 00 ................ [521120.737596] Object ffff880f36b387a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.738524] Object ffff880f36b387b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.739388] Object ffff880f36b387c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.740277] Object ffff880f36b387d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.741187] Object ffff880f36b387e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.742233] Object ffff880f36b387f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [521120.743229] CPU: 41 PID: 9644 Comm: mount.ocfs2 Tainted: G B 4.2.0-rc6-next-20150810-sasha-00039-gf909086 #2420 [521120.744274] ffff880f36b38000 ffff880d89c8f638 ffffffffb6e9ba8a ffff880101c0e5c0 [521120.745025] ffff880d89c8f668 ffffffffad76a313 ffff880101c0e5c0 ffffea003cdace00 [521120.745908] ffff880f36b38700 ffff880f36b38798 ffff880d89c8f690 ffffffffad772854 [521120.747063] Call Trace: [521120.747520] dump_stack (lib/dump_stack.c:52) [521120.748053] print_trailer (mm/slub.c:653) [521120.748582] object_err (mm/slub.c:660) [521120.749079] kasan_report_error (include/linux/kasan.h:20 mm/kasan/report.c:152 mm/kasan/report.c:194) [521120.750834] __asan_report_load4_noabort (mm/kasan/report.c:250) [521120.753580] dio_bio_complete (fs/direct-io.c:478) [521120.755752] do_blockdev_direct_IO (fs/direct-io.c:494 fs/direct-io.c:1291) [521120.759765] __blockdev_direct_IO (fs/direct-io.c:1322) [521120.761658] blkdev_direct_IO (fs/block_dev.c:162) [521120.762993] generic_file_read_iter (mm/filemap.c:1738) [521120.767405] blkdev_read_iter (fs/block_dev.c:1649) [521120.768556] __vfs_read (fs/read_write.c:423 fs/read_write.c:434) [521120.772126] vfs_read (fs/read_write.c:454) [521120.773118] SyS_pread64 (fs/read_write.c:607 fs/read_write.c:594) [521120.776062] entry_SYSCALL_64_fastpath (arch/x86/entry/entry_64.S:186) [521120.777375] Memory state around the buggy address: [521120.778118] ffff880f36b38600: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [521120.779211] ffff880f36b38680: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [521120.780315] >ffff880f36b38700: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [521120.781465] ^ [521120.782083] ffff880f36b38780: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [521120.783717] ffff880f36b38800: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [521120.784818] ================================================================== This patch fixes a few of those places that I caught while auditing the patch, but the original patch should be audited further for more occurences of this issue since I'm not too familiar with the code. Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-10 23:05:18 +00:00
if (!error) {
bio_endio(raid_bi);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
if (atomic_dec_and_test(&conf->active_aligned_reads))
wake_up(&conf->wait_for_quiescent);
return;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
}
pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
add_bio_to_retry(raid_bi, conf);
}
static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
{
struct r5conf *conf = mddev->private;
int dd_idx;
struct bio* align_bi;
struct md_rdev *rdev;
sector_t end_sector;
if (!in_chunk_boundary(mddev, raid_bio)) {
pr_debug("%s: non aligned\n", __func__);
return 0;
}
/*
* use bio_clone_fast to make a copy of the bio
*/
align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
if (!align_bi)
return 0;
/*
* set bi_end_io to a new function, and set bi_private to the
* original bio.
*/
align_bi->bi_end_io = raid5_align_endio;
align_bi->bi_private = raid_bio;
/*
* compute position
*/
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
align_bi->bi_iter.bi_sector =
raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
0, &dd_idx, NULL);
end_sector = bio_end_sector(align_bi);
rcu_read_lock();
rdev = rcu_dereference(conf->disks[dd_idx].replacement);
if (!rdev || test_bit(Faulty, &rdev->flags) ||
rdev->recovery_offset < end_sector) {
rdev = rcu_dereference(conf->disks[dd_idx].rdev);
if (rdev &&
(test_bit(Faulty, &rdev->flags) ||
!(test_bit(In_sync, &rdev->flags) ||
rdev->recovery_offset >= end_sector)))
rdev = NULL;
}
md/r5cache: enable chunk_aligned_read with write back cache Chunk aligned read significantly reduces CPU usage of raid456. However, it is not safe to fully bypass the write back cache. This patch enables chunk aligned read with write back cache. For chunk aligned read, we track stripes in write back cache at a bigger granularity, "big_stripe". Each chunk may contain more than one stripe (for example, a 256kB chunk contains 64 4kB-page, so this chunk contain 64 stripes). For chunk_aligned_read, these stripes are grouped into one big_stripe, so we only need one lookup for the whole chunk. For each big_stripe, struct big_stripe_info tracks how many stripes of this big_stripe are in the write back cache. We count how many stripes of this big_stripe are in the write back cache. These counters are tracked in a radix tree (big_stripe_tree). r5c_tree_index() is used to calculate keys for the radix tree. chunk_aligned_read() calls r5c_big_stripe_cached() to look up big_stripe of each chunk in the tree. If this big_stripe is in the tree, chunk_aligned_read() aborts. This look up is protected by rcu_read_lock(). It is necessary to remember whether a stripe is counted in big_stripe_tree. Instead of adding new flag, we reuses existing flags: STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these two flags are set, the stripe is counted in big_stripe_tree. This requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to r5c_try_caching_write(); and moving clear_bit of STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to r5c_finish_stripe_write_out(). Signed-off-by: Song Liu <songliubraving@fb.com> Reviewed-by: NeilBrown <neilb@suse.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-01-11 21:39:14 +00:00
if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
rcu_read_unlock();
bio_put(align_bi);
return 0;
}
if (rdev) {
sector_t first_bad;
int bad_sectors;
atomic_inc(&rdev->nr_pending);
rcu_read_unlock();
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
raid_bio->bi_next = (void*)rdev;
bio_set_dev(align_bi, rdev->bdev);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
bio_sectors(align_bi),
&first_bad, &bad_sectors)) {
bio_put(align_bi);
rdev_dec_pending(rdev, mddev);
return 0;
}
/* No reshape active, so we can trust rdev->data_offset */
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
align_bi->bi_iter.bi_sector += rdev->data_offset;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
spin_lock_irq(&conf->device_lock);
wait_event_lock_irq(conf->wait_for_quiescent,
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
conf->quiesce == 0,
conf->device_lock);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
atomic_inc(&conf->active_aligned_reads);
spin_unlock_irq(&conf->device_lock);
if (mddev->gendisk)
trace_block_bio_remap(align_bi->bi_disk->queue,
align_bi, disk_devt(mddev->gendisk),
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
raid_bio->bi_iter.bi_sector);
submit_bio_noacct(align_bi);
return 1;
} else {
rcu_read_unlock();
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
bio_put(align_bi);
return 0;
}
}
static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
{
struct bio *split;
sector_t sector = raid_bio->bi_iter.bi_sector;
unsigned chunk_sects = mddev->chunk_sectors;
unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
if (sectors < bio_sectors(raid_bio)) {
struct r5conf *conf = mddev->private;
split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
bio_chain(split, raid_bio);
submit_bio_noacct(raid_bio);
raid_bio = split;
}
if (!raid5_read_one_chunk(mddev, raid_bio))
return raid_bio;
return NULL;
}
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
/* __get_priority_stripe - get the next stripe to process
*
* Full stripe writes are allowed to pass preread active stripes up until
* the bypass_threshold is exceeded. In general the bypass_count
* increments when the handle_list is handled before the hold_list; however, it
* will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
* stripe with in flight i/o. The bypass_count will be reset when the
* head of the hold_list has changed, i.e. the head was promoted to the
* handle_list.
*/
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
{
struct stripe_head *sh, *tmp;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
struct list_head *handle_list = NULL;
struct r5worker_group *wg;
bool second_try = !r5c_is_writeback(conf->log) &&
!r5l_log_disk_error(conf);
bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
r5l_log_disk_error(conf);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
again:
wg = NULL;
sh = NULL;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
if (conf->worker_cnt_per_group == 0) {
handle_list = try_loprio ? &conf->loprio_list :
&conf->handle_list;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
} else if (group != ANY_GROUP) {
handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
&conf->worker_groups[group].handle_list;
wg = &conf->worker_groups[group];
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
} else {
int i;
for (i = 0; i < conf->group_cnt; i++) {
handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
&conf->worker_groups[i].handle_list;
wg = &conf->worker_groups[i];
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
if (!list_empty(handle_list))
break;
}
}
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
__func__,
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
list_empty(handle_list) ? "empty" : "busy",
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
list_empty(&conf->hold_list) ? "empty" : "busy",
atomic_read(&conf->pending_full_writes), conf->bypass_count);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
if (!list_empty(handle_list)) {
sh = list_entry(handle_list->next, typeof(*sh), lru);
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (list_empty(&conf->hold_list))
conf->bypass_count = 0;
else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
if (conf->hold_list.next == conf->last_hold)
conf->bypass_count++;
else {
conf->last_hold = conf->hold_list.next;
conf->bypass_count -= conf->bypass_threshold;
if (conf->bypass_count < 0)
conf->bypass_count = 0;
}
}
} else if (!list_empty(&conf->hold_list) &&
((conf->bypass_threshold &&
conf->bypass_count > conf->bypass_threshold) ||
atomic_read(&conf->pending_full_writes) == 0)) {
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
list_for_each_entry(tmp, &conf->hold_list, lru) {
if (conf->worker_cnt_per_group == 0 ||
group == ANY_GROUP ||
!cpu_online(tmp->cpu) ||
cpu_to_group(tmp->cpu) == group) {
sh = tmp;
break;
}
}
if (sh) {
conf->bypass_count -= conf->bypass_threshold;
if (conf->bypass_count < 0)
conf->bypass_count = 0;
}
wg = NULL;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
}
if (!sh) {
if (second_try)
return NULL;
second_try = true;
try_loprio = !try_loprio;
goto again;
}
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (wg) {
wg->stripes_cnt--;
sh->group = NULL;
}
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
list_del_init(&sh->lru);
BUG_ON(atomic_inc_return(&sh->count) != 1);
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
return sh;
}
struct raid5_plug_cb {
struct blk_plug_cb cb;
struct list_head list;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
};
static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
{
struct raid5_plug_cb *cb = container_of(
blk_cb, struct raid5_plug_cb, cb);
struct stripe_head *sh;
struct mddev *mddev = cb->cb.data;
struct r5conf *conf = mddev->private;
int cnt = 0;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int hash;
if (cb->list.next && !list_empty(&cb->list)) {
spin_lock_irq(&conf->device_lock);
while (!list_empty(&cb->list)) {
sh = list_first_entry(&cb->list, struct stripe_head, lru);
list_del_init(&sh->lru);
/*
* avoid race release_stripe_plug() sees
* STRIPE_ON_UNPLUG_LIST clear but the stripe
* is still in our list
*/
smp_mb__before_atomic();
clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
/*
* STRIPE_ON_RELEASE_LIST could be set here. In that
* case, the count is always > 1 here
*/
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
hash = sh->hash_lock_index;
__release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
cnt++;
}
spin_unlock_irq(&conf->device_lock);
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
release_inactive_stripe_list(conf, cb->temp_inactive_list,
NR_STRIPE_HASH_LOCKS);
if (mddev->queue)
trace_block_unplug(mddev->queue, cnt, !from_schedule);
kfree(cb);
}
static void release_stripe_plug(struct mddev *mddev,
struct stripe_head *sh)
{
struct blk_plug_cb *blk_cb = blk_check_plugged(
raid5_unplug, mddev,
sizeof(struct raid5_plug_cb));
struct raid5_plug_cb *cb;
if (!blk_cb) {
raid5_release_stripe(sh);
return;
}
cb = container_of(blk_cb, struct raid5_plug_cb, cb);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (cb->list.next == NULL) {
int i;
INIT_LIST_HEAD(&cb->list);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
INIT_LIST_HEAD(cb->temp_inactive_list + i);
}
if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
list_add_tail(&sh->lru, &cb->list);
else
raid5_release_stripe(sh);
}
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
static void make_discard_request(struct mddev *mddev, struct bio *bi)
{
struct r5conf *conf = mddev->private;
sector_t logical_sector, last_sector;
struct stripe_head *sh;
int stripe_sectors;
if (mddev->reshape_position != MaxSector)
/* Skip discard while reshape is happening */
return;
logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
last_sector = bio_end_sector(bi);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
bi->bi_next = NULL;
stripe_sectors = conf->chunk_sectors *
(conf->raid_disks - conf->max_degraded);
logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
stripe_sectors);
sector_div(last_sector, stripe_sectors);
logical_sector *= conf->chunk_sectors;
last_sector *= conf->chunk_sectors;
for (; logical_sector < last_sector;
logical_sector += RAID5_STRIPE_SECTORS(conf)) {
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
DEFINE_WAIT(w);
int d;
again:
sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
prepare_to_wait(&conf->wait_for_overlap, &w,
TASK_UNINTERRUPTIBLE);
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
if (test_bit(STRIPE_SYNCING, &sh->state)) {
raid5_release_stripe(sh);
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
schedule();
goto again;
}
clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
spin_lock_irq(&sh->stripe_lock);
for (d = 0; d < conf->raid_disks; d++) {
if (d == sh->pd_idx || d == sh->qd_idx)
continue;
if (sh->dev[d].towrite || sh->dev[d].toread) {
set_bit(R5_Overlap, &sh->dev[d].flags);
spin_unlock_irq(&sh->stripe_lock);
raid5_release_stripe(sh);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
schedule();
goto again;
}
}
md/raid5: ensure sync and DISCARD don't happen at the same time. A number of problems can occur due to races between resync/recovery and discard. - if sync_request calls handle_stripe() while a discard is happening on the stripe, it might call handle_stripe_clean_event before all of the individual discard requests have completed (so some devices are still locked, but not all). Since commit ca64cae96037de16e4af92678814f5d4bf0c1c65 md/raid5: Make sure we clear R5_Discard when discard is finished. this will cause R5_Discard to be cleared for the parity device, so handle_stripe_clean_event() will not be called when the other devices do become unlocked, so their ->written will not be cleared. This ultimately leads to a WARN_ON in init_stripe and a lock-up. - If handle_stripe_clean_event() does clear R5_UPTODATE at an awkward time for resync, it can lead to s->uptodate being less than disks in handle_parity_checks5(), which triggers a BUG (because it is one). So: - keep R5_Discard on the parity device until all other devices have completed their discard request - make sure we don't try to have a 'discard' and a 'sync' action at the same time. This involves a new stripe flag to we know when a 'discard' is happening, and the use of R5_Overlap on the parity disk so when a discard is wanted while a sync is active, so we know to wake up the discard at the appropriate time. Discard support for RAID5 was added in 3.7, so this is suitable for any -stable kernel since 3.7. Cc: stable@vger.kernel.org (v3.7+) Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com> Tested-by: Jes Sorensen <Jes.Sorensen@redhat.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-03-12 01:18:06 +00:00
set_bit(STRIPE_DISCARD, &sh->state);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
finish_wait(&conf->wait_for_overlap, &w);
sh->overwrite_disks = 0;
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
for (d = 0; d < conf->raid_disks; d++) {
if (d == sh->pd_idx || d == sh->qd_idx)
continue;
sh->dev[d].towrite = bi;
set_bit(R5_OVERWRITE, &sh->dev[d].flags);
bio_inc_remaining(bi);
md_write_inc(mddev, bi);
sh->overwrite_disks++;
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
}
spin_unlock_irq(&sh->stripe_lock);
if (conf->mddev->bitmap) {
for (d = 0;
d < conf->raid_disks - conf->max_degraded;
d++)
md_bitmap_startwrite(mddev->bitmap,
sh->sector,
RAID5_STRIPE_SECTORS(conf),
0);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
sh->bm_seq = conf->seq_flush + 1;
set_bit(STRIPE_BIT_DELAY, &sh->state);
}
set_bit(STRIPE_HANDLE, &sh->state);
clear_bit(STRIPE_DELAYED, &sh->state);
if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
atomic_inc(&conf->preread_active_stripes);
release_stripe_plug(mddev, sh);
}
bio_endio(bi);
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
}
static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
{
struct r5conf *conf = mddev->private;
int dd_idx;
sector_t new_sector;
sector_t logical_sector, last_sector;
struct stripe_head *sh;
const int rw = bio_data_dir(bi);
DEFINE_WAIT(w);
bool do_prepare;
bool do_flush = false;
if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
int ret = log_handle_flush_request(conf, bi);
if (ret == 0)
return true;
if (ret == -ENODEV) {
if (md_flush_request(mddev, bi))
return true;
}
/* ret == -EAGAIN, fallback */
/*
* if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
* we need to flush journal device
*/
do_flush = bi->bi_opf & REQ_PREFLUSH;
}
if (!md_write_start(mddev, bi))
return false;
md/raid5: don't do chunk aligned read on degraded array. When array is degraded, read data landed on failed drives will result in reading rest of data in a stripe. So a single sequential read would result in same data being read twice. This patch is to avoid chunk aligned read for degraded array. The downside is to involve stripe cache which means associated CPU overhead and extra memory copy. Test Results: Following test are done on a enterprise storage node with Seagate 6T SAS drives and Xeon E5-2648L CPU (10 cores, 1.9Ghz), 10 disks MD RAID6 8+2, chunk size 128 KiB. I use FIO, using direct-io with various bs size, enough queue depth, tested sequential and 100% random read against 3 array config: 1) optimal, as baseline; 2) degraded; 3) degraded with this patch. Kernel version is 4.0-rc3. Each individual test I only did once so there might be some variations, but we just focus on big trend. Sequential Read: bs=(KiB) optimal(MiB/s) degraded(MiB/s) degraded-with-patch (MiB/s) 1024 1608 656 995 512 1624 710 956 256 1635 728 980 128 1636 771 983 64 1612 1119 1000 32 1580 1420 1004 16 1368 688 986 8 768 647 953 4 411 413 850 Random Read: bs=(KiB) optimal(IOPS) degraded(IOPS) degraded-with-patch (IOPS) 1024 163 160 156 512 274 273 272 256 426 428 424 128 576 592 591 64 726 724 726 32 849 848 837 16 900 970 971 8 927 940 929 4 948 940 955 Some notes: * In sequential + optimal, as bs size getting smaller, the FIO thread become CPU bound. * In sequential + degraded, there's big increase when bs is 64K and 32K, I don't have explanation. * In sequential + degraded-with-patch, the MD thread mostly become CPU bound. If you want to we can discuss specific data point in those data. But in general it seems with this patch, we have more predictable and in most cases significant better sequential read performance when array is degraded, and almost no noticeable impact on random read. Performance is a complicated thing, the patch works well for this particular configuration, but may not be universal. For example I imagine testing on all SSD array may have very different result. But I personally think in most cases IO bandwidth is more scarce resource than CPU. Signed-off-by: Eric Mei <eric.mei@seagate.com> Signed-off-by: NeilBrown <neilb@suse.de>
2015-03-19 05:39:11 +00:00
/*
* If array is degraded, better not do chunk aligned read because
* later we might have to read it again in order to reconstruct
* data on failed drives.
*/
if (rw == READ && mddev->degraded == 0 &&
mddev->reshape_position == MaxSector) {
bi = chunk_aligned_read(mddev, bi);
if (!bi)
return true;
}
if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
make_discard_request(mddev, bi);
md_write_end(mddev);
return true;
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
}
logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
last_sector = bio_end_sector(bi);
bi->bi_next = NULL;
prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
int previous;
int seq;
do_prepare = false;
retry:
seq = read_seqcount_begin(&conf->gen_lock);
previous = 0;
if (do_prepare)
prepare_to_wait(&conf->wait_for_overlap, &w,
TASK_UNINTERRUPTIBLE);
if (unlikely(conf->reshape_progress != MaxSector)) {
/* spinlock is needed as reshape_progress may be
* 64bit on a 32bit platform, and so it might be
* possible to see a half-updated value
* Of course reshape_progress could change after
* the lock is dropped, so once we get a reference
* to the stripe that we think it is, we will have
* to check again.
*/
spin_lock_irq(&conf->device_lock);
if (mddev->reshape_backwards
? logical_sector < conf->reshape_progress
: logical_sector >= conf->reshape_progress) {
previous = 1;
} else {
if (mddev->reshape_backwards
? logical_sector < conf->reshape_safe
: logical_sector >= conf->reshape_safe) {
spin_unlock_irq(&conf->device_lock);
schedule();
do_prepare = true;
goto retry;
}
}
spin_unlock_irq(&conf->device_lock);
}
new_sector = raid5_compute_sector(conf, logical_sector,
previous,
&dd_idx, NULL);
pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
(unsigned long long)new_sector,
(unsigned long long)logical_sector);
sh = raid5_get_active_stripe(conf, new_sector, previous,
(bi->bi_opf & REQ_RAHEAD), 0);
if (sh) {
if (unlikely(previous)) {
/* expansion might have moved on while waiting for a
* stripe, so we must do the range check again.
* Expansion could still move past after this
* test, but as we are holding a reference to
* 'sh', we know that if that happens,
* STRIPE_EXPANDING will get set and the expansion
* won't proceed until we finish with the stripe.
*/
int must_retry = 0;
spin_lock_irq(&conf->device_lock);
if (mddev->reshape_backwards
? logical_sector >= conf->reshape_progress
: logical_sector < conf->reshape_progress)
/* mismatch, need to try again */
must_retry = 1;
spin_unlock_irq(&conf->device_lock);
if (must_retry) {
raid5_release_stripe(sh);
schedule();
do_prepare = true;
goto retry;
}
}
if (read_seqcount_retry(&conf->gen_lock, seq)) {
/* Might have got the wrong stripe_head
* by accident
*/
raid5_release_stripe(sh);
goto retry;
}
if (test_bit(STRIPE_EXPANDING, &sh->state) ||
!add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
/* Stripe is busy expanding or
* add failed due to overlap. Flush everything
* and wait a while
*/
md_wakeup_thread(mddev->thread);
raid5_release_stripe(sh);
schedule();
do_prepare = true;
goto retry;
}
if (do_flush) {
set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
/* we only need flush for one stripe */
do_flush = false;
}
set_bit(STRIPE_HANDLE, &sh->state);
clear_bit(STRIPE_DELAYED, &sh->state);
RAID5: batch adjacent full stripe write stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k unit. Idealy we should use big size for adjacent full stripe writes. Bigger stripe cache size means less stripes runing in the state machine so can reduce cpu overhead. And also bigger size can cause bigger IO size dispatched to under layer disks. With below patch, we will automatically batch adjacent full stripe write together. Such stripes will be added to the batch list. Only the first stripe of the list will be put to handle_list and so run handle_stripe(). Some steps of handle_stripe() are extended to cover all stripes of the list, including ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes running in handle_stripe() and we send IO of whole stripe list together to increase IO size. Stripes added to a batch list have some limitations. A batch list can only include full stripe write and can't cross chunk boundary to make sure stripes have the same parity disks. Stripes in a batch list must be in the same state (no written, toread and so on). If a stripe is in a batch list, all new read/write to add_stripe_bio will be blocked to overlap conflict till the batch list is handled. The limitations will make sure stripes in a batch list be in exactly the same state in the life circly. I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6 PCIe SSD. This patch improves around 30% performance and IO size to under layer disk is exactly 32k. I also run a 4k randwrite test in the same array to make sure the performance isn't changed with the patch. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:03 +00:00
if ((!sh->batch_head || sh == sh->batch_head) &&
(bi->bi_opf & REQ_SYNC) &&
!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
atomic_inc(&conf->preread_active_stripes);
release_stripe_plug(mddev, sh);
} else {
/* cannot get stripe for read-ahead, just give-up */
bi->bi_status = BLK_STS_IOERR;
break;
}
}
finish_wait(&conf->wait_for_overlap, &w);
if (rw == WRITE)
md_write_end(mddev);
bio_endio(bi);
return true;
}
static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
{
/* reshaping is quite different to recovery/resync so it is
* handled quite separately ... here.
*
* On each call to sync_request, we gather one chunk worth of
* destination stripes and flag them as expanding.
* Then we find all the source stripes and request reads.
* As the reads complete, handle_stripe will copy the data
* into the destination stripe and release that stripe.
*/
struct r5conf *conf = mddev->private;
struct stripe_head *sh;
struct md_rdev *rdev;
sector_t first_sector, last_sector;
int raid_disks = conf->previous_raid_disks;
int data_disks = raid_disks - conf->max_degraded;
int new_data_disks = conf->raid_disks - conf->max_degraded;
int i;
int dd_idx;
sector_t writepos, readpos, safepos;
sector_t stripe_addr;
int reshape_sectors;
struct list_head stripes;
sector_t retn;
if (sector_nr == 0) {
/* If restarting in the middle, skip the initial sectors */
if (mddev->reshape_backwards &&
conf->reshape_progress < raid5_size(mddev, 0, 0)) {
sector_nr = raid5_size(mddev, 0, 0)
- conf->reshape_progress;
} else if (mddev->reshape_backwards &&
conf->reshape_progress == MaxSector) {
/* shouldn't happen, but just in case, finish up.*/
sector_nr = MaxSector;
} else if (!mddev->reshape_backwards &&
conf->reshape_progress > 0)
sector_nr = conf->reshape_progress;
sector_div(sector_nr, new_data_disks);
if (sector_nr) {
mddev->curr_resync_completed = sector_nr;
md: fix deadlock causing by sysfs_notify The following deadlock was captured. The first process is holding 'kernfs_mutex' and hung by io. The io was staging in 'r1conf.pending_bio_list' of raid1 device, this pending bio list would be flushed by second process 'md127_raid1', but it was hung by 'kernfs_mutex'. Using sysfs_notify_dirent_safe() to replace sysfs_notify() can fix it. There were other sysfs_notify() invoked from io path, removed all of them. PID: 40430 TASK: ffff8ee9c8c65c40 CPU: 29 COMMAND: "probe_file" #0 [ffffb87c4df37260] __schedule at ffffffff9a8678ec #1 [ffffb87c4df372f8] schedule at ffffffff9a867f06 #2 [ffffb87c4df37310] io_schedule at ffffffff9a0c73e6 #3 [ffffb87c4df37328] __dta___xfs_iunpin_wait_3443 at ffffffffc03a4057 [xfs] #4 [ffffb87c4df373a0] xfs_iunpin_wait at ffffffffc03a6c79 [xfs] #5 [ffffb87c4df373b0] __dta_xfs_reclaim_inode_3357 at ffffffffc039a46c [xfs] #6 [ffffb87c4df37400] xfs_reclaim_inodes_ag at ffffffffc039a8b6 [xfs] #7 [ffffb87c4df37590] xfs_reclaim_inodes_nr at ffffffffc039bb33 [xfs] #8 [ffffb87c4df375b0] xfs_fs_free_cached_objects at ffffffffc03af0e9 [xfs] #9 [ffffb87c4df375c0] super_cache_scan at ffffffff9a287ec7 #10 [ffffb87c4df37618] shrink_slab at ffffffff9a1efd93 #11 [ffffb87c4df37700] shrink_node at ffffffff9a1f5968 #12 [ffffb87c4df37788] do_try_to_free_pages at ffffffff9a1f5ea2 #13 [ffffb87c4df377f0] try_to_free_mem_cgroup_pages at ffffffff9a1f6445 #14 [ffffb87c4df37880] try_charge at ffffffff9a26cc5f #15 [ffffb87c4df37920] memcg_kmem_charge_memcg at ffffffff9a270f6a #16 [ffffb87c4df37958] new_slab at ffffffff9a251430 #17 [ffffb87c4df379c0] ___slab_alloc at ffffffff9a251c85 #18 [ffffb87c4df37a80] __slab_alloc at ffffffff9a25635d #19 [ffffb87c4df37ac0] kmem_cache_alloc at ffffffff9a251f89 #20 [ffffb87c4df37b00] alloc_inode at ffffffff9a2a2b10 #21 [ffffb87c4df37b20] iget_locked at ffffffff9a2a4854 #22 [ffffb87c4df37b60] kernfs_get_inode at ffffffff9a311377 #23 [ffffb87c4df37b80] kernfs_iop_lookup at ffffffff9a311e2b #24 [ffffb87c4df37ba8] lookup_slow at ffffffff9a290118 #25 [ffffb87c4df37c10] walk_component at ffffffff9a291e83 #26 [ffffb87c4df37c78] path_lookupat at ffffffff9a293619 #27 [ffffb87c4df37cd8] filename_lookup at ffffffff9a2953af #28 [ffffb87c4df37de8] user_path_at_empty at ffffffff9a295566 #29 [ffffb87c4df37e10] vfs_statx at ffffffff9a289787 #30 [ffffb87c4df37e70] SYSC_newlstat at ffffffff9a289d5d #31 [ffffb87c4df37f18] sys_newlstat at ffffffff9a28a60e #32 [ffffb87c4df37f28] do_syscall_64 at ffffffff9a003949 #33 [ffffb87c4df37f50] entry_SYSCALL_64_after_hwframe at ffffffff9aa001ad RIP: 00007f617a5f2905 RSP: 00007f607334f838 RFLAGS: 00000246 RAX: ffffffffffffffda RBX: 00007f6064044b20 RCX: 00007f617a5f2905 RDX: 00007f6064044b20 RSI: 00007f6064044b20 RDI: 00007f6064005890 RBP: 00007f6064044aa0 R8: 0000000000000030 R9: 000000000000011c R10: 0000000000000013 R11: 0000000000000246 R12: 00007f606417e6d0 R13: 00007f6064044aa0 R14: 00007f6064044b10 R15: 00000000ffffffff ORIG_RAX: 0000000000000006 CS: 0033 SS: 002b PID: 927 TASK: ffff8f15ac5dbd80 CPU: 42 COMMAND: "md127_raid1" #0 [ffffb87c4df07b28] __schedule at ffffffff9a8678ec #1 [ffffb87c4df07bc0] schedule at ffffffff9a867f06 #2 [ffffb87c4df07bd8] schedule_preempt_disabled at ffffffff9a86825e #3 [ffffb87c4df07be8] __mutex_lock at ffffffff9a869bcc #4 [ffffb87c4df07ca0] __mutex_lock_slowpath at ffffffff9a86a013 #5 [ffffb87c4df07cb0] mutex_lock at ffffffff9a86a04f #6 [ffffb87c4df07cc8] kernfs_find_and_get_ns at ffffffff9a311d83 #7 [ffffb87c4df07cf0] sysfs_notify at ffffffff9a314b3a #8 [ffffb87c4df07d18] md_update_sb at ffffffff9a688696 #9 [ffffb87c4df07d98] md_update_sb at ffffffff9a6886d5 #10 [ffffb87c4df07da8] md_check_recovery at ffffffff9a68ad9c #11 [ffffb87c4df07dd0] raid1d at ffffffffc01f0375 [raid1] #12 [ffffb87c4df07ea0] md_thread at ffffffff9a680348 #13 [ffffb87c4df07f08] kthread at ffffffff9a0b8005 #14 [ffffb87c4df07f50] ret_from_fork at ffffffff9aa00344 Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com> Signed-off-by: Song Liu <songliubraving@fb.com>
2020-07-14 23:10:26 +00:00
sysfs_notify_dirent_safe(mddev->sysfs_completed);
*skipped = 1;
retn = sector_nr;
goto finish;
}
}
/* We need to process a full chunk at a time.
* If old and new chunk sizes differ, we need to process the
* largest of these
*/
reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
/* We update the metadata at least every 10 seconds, or when
* the data about to be copied would over-write the source of
* the data at the front of the range. i.e. one new_stripe
* along from reshape_progress new_maps to after where
* reshape_safe old_maps to
*/
writepos = conf->reshape_progress;
sector_div(writepos, new_data_disks);
readpos = conf->reshape_progress;
sector_div(readpos, data_disks);
safepos = conf->reshape_safe;
sector_div(safepos, data_disks);
if (mddev->reshape_backwards) {
BUG_ON(writepos < reshape_sectors);
writepos -= reshape_sectors;
readpos += reshape_sectors;
safepos += reshape_sectors;
} else {
writepos += reshape_sectors;
/* readpos and safepos are worst-case calculations.
* A negative number is overly pessimistic, and causes
* obvious problems for unsigned storage. So clip to 0.
*/
readpos -= min_t(sector_t, reshape_sectors, readpos);
safepos -= min_t(sector_t, reshape_sectors, safepos);
}
/* Having calculated the 'writepos' possibly use it
* to set 'stripe_addr' which is where we will write to.
*/
if (mddev->reshape_backwards) {
BUG_ON(conf->reshape_progress == 0);
stripe_addr = writepos;
BUG_ON((mddev->dev_sectors &
~((sector_t)reshape_sectors - 1))
- reshape_sectors - stripe_addr
!= sector_nr);
} else {
BUG_ON(writepos != sector_nr + reshape_sectors);
stripe_addr = sector_nr;
}
/* 'writepos' is the most advanced device address we might write.
* 'readpos' is the least advanced device address we might read.
* 'safepos' is the least address recorded in the metadata as having
* been reshaped.
* If there is a min_offset_diff, these are adjusted either by
* increasing the safepos/readpos if diff is negative, or
* increasing writepos if diff is positive.
* If 'readpos' is then behind 'writepos', there is no way that we can
* ensure safety in the face of a crash - that must be done by userspace
* making a backup of the data. So in that case there is no particular
* rush to update metadata.
* Otherwise if 'safepos' is behind 'writepos', then we really need to
* update the metadata to advance 'safepos' to match 'readpos' so that
* we can be safe in the event of a crash.
* So we insist on updating metadata if safepos is behind writepos and
* readpos is beyond writepos.
* In any case, update the metadata every 10 seconds.
* Maybe that number should be configurable, but I'm not sure it is
* worth it.... maybe it could be a multiple of safemode_delay???
*/
if (conf->min_offset_diff < 0) {
safepos += -conf->min_offset_diff;
readpos += -conf->min_offset_diff;
} else
writepos += conf->min_offset_diff;
if ((mddev->reshape_backwards
? (safepos > writepos && readpos < writepos)
: (safepos < writepos && readpos > writepos)) ||
time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
/* Cannot proceed until we've updated the superblock... */
wait_event(conf->wait_for_overlap,
atomic_read(&conf->reshape_stripes)==0
|| test_bit(MD_RECOVERY_INTR, &mddev->recovery));
if (atomic_read(&conf->reshape_stripes) != 0)
return 0;
mddev->reshape_position = conf->reshape_progress;
mddev->curr_resync_completed = sector_nr;
if (!mddev->reshape_backwards)
/* Can update recovery_offset */
rdev_for_each(rdev, mddev)
if (rdev->raid_disk >= 0 &&
!test_bit(Journal, &rdev->flags) &&
!test_bit(In_sync, &rdev->flags) &&
rdev->recovery_offset < sector_nr)
rdev->recovery_offset = sector_nr;
conf->reshape_checkpoint = jiffies;
set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
md_wakeup_thread(mddev->thread);
wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
test_bit(MD_RECOVERY_INTR, &mddev->recovery));
if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
return 0;
spin_lock_irq(&conf->device_lock);
conf->reshape_safe = mddev->reshape_position;
spin_unlock_irq(&conf->device_lock);
wake_up(&conf->wait_for_overlap);
md: fix deadlock causing by sysfs_notify The following deadlock was captured. The first process is holding 'kernfs_mutex' and hung by io. The io was staging in 'r1conf.pending_bio_list' of raid1 device, this pending bio list would be flushed by second process 'md127_raid1', but it was hung by 'kernfs_mutex'. Using sysfs_notify_dirent_safe() to replace sysfs_notify() can fix it. There were other sysfs_notify() invoked from io path, removed all of them. PID: 40430 TASK: ffff8ee9c8c65c40 CPU: 29 COMMAND: "probe_file" #0 [ffffb87c4df37260] __schedule at ffffffff9a8678ec #1 [ffffb87c4df372f8] schedule at ffffffff9a867f06 #2 [ffffb87c4df37310] io_schedule at ffffffff9a0c73e6 #3 [ffffb87c4df37328] __dta___xfs_iunpin_wait_3443 at ffffffffc03a4057 [xfs] #4 [ffffb87c4df373a0] xfs_iunpin_wait at ffffffffc03a6c79 [xfs] #5 [ffffb87c4df373b0] __dta_xfs_reclaim_inode_3357 at ffffffffc039a46c [xfs] #6 [ffffb87c4df37400] xfs_reclaim_inodes_ag at ffffffffc039a8b6 [xfs] #7 [ffffb87c4df37590] xfs_reclaim_inodes_nr at ffffffffc039bb33 [xfs] #8 [ffffb87c4df375b0] xfs_fs_free_cached_objects at ffffffffc03af0e9 [xfs] #9 [ffffb87c4df375c0] super_cache_scan at ffffffff9a287ec7 #10 [ffffb87c4df37618] shrink_slab at ffffffff9a1efd93 #11 [ffffb87c4df37700] shrink_node at ffffffff9a1f5968 #12 [ffffb87c4df37788] do_try_to_free_pages at ffffffff9a1f5ea2 #13 [ffffb87c4df377f0] try_to_free_mem_cgroup_pages at ffffffff9a1f6445 #14 [ffffb87c4df37880] try_charge at ffffffff9a26cc5f #15 [ffffb87c4df37920] memcg_kmem_charge_memcg at ffffffff9a270f6a #16 [ffffb87c4df37958] new_slab at ffffffff9a251430 #17 [ffffb87c4df379c0] ___slab_alloc at ffffffff9a251c85 #18 [ffffb87c4df37a80] __slab_alloc at ffffffff9a25635d #19 [ffffb87c4df37ac0] kmem_cache_alloc at ffffffff9a251f89 #20 [ffffb87c4df37b00] alloc_inode at ffffffff9a2a2b10 #21 [ffffb87c4df37b20] iget_locked at ffffffff9a2a4854 #22 [ffffb87c4df37b60] kernfs_get_inode at ffffffff9a311377 #23 [ffffb87c4df37b80] kernfs_iop_lookup at ffffffff9a311e2b #24 [ffffb87c4df37ba8] lookup_slow at ffffffff9a290118 #25 [ffffb87c4df37c10] walk_component at ffffffff9a291e83 #26 [ffffb87c4df37c78] path_lookupat at ffffffff9a293619 #27 [ffffb87c4df37cd8] filename_lookup at ffffffff9a2953af #28 [ffffb87c4df37de8] user_path_at_empty at ffffffff9a295566 #29 [ffffb87c4df37e10] vfs_statx at ffffffff9a289787 #30 [ffffb87c4df37e70] SYSC_newlstat at ffffffff9a289d5d #31 [ffffb87c4df37f18] sys_newlstat at ffffffff9a28a60e #32 [ffffb87c4df37f28] do_syscall_64 at ffffffff9a003949 #33 [ffffb87c4df37f50] entry_SYSCALL_64_after_hwframe at ffffffff9aa001ad RIP: 00007f617a5f2905 RSP: 00007f607334f838 RFLAGS: 00000246 RAX: ffffffffffffffda RBX: 00007f6064044b20 RCX: 00007f617a5f2905 RDX: 00007f6064044b20 RSI: 00007f6064044b20 RDI: 00007f6064005890 RBP: 00007f6064044aa0 R8: 0000000000000030 R9: 000000000000011c R10: 0000000000000013 R11: 0000000000000246 R12: 00007f606417e6d0 R13: 00007f6064044aa0 R14: 00007f6064044b10 R15: 00000000ffffffff ORIG_RAX: 0000000000000006 CS: 0033 SS: 002b PID: 927 TASK: ffff8f15ac5dbd80 CPU: 42 COMMAND: "md127_raid1" #0 [ffffb87c4df07b28] __schedule at ffffffff9a8678ec #1 [ffffb87c4df07bc0] schedule at ffffffff9a867f06 #2 [ffffb87c4df07bd8] schedule_preempt_disabled at ffffffff9a86825e #3 [ffffb87c4df07be8] __mutex_lock at ffffffff9a869bcc #4 [ffffb87c4df07ca0] __mutex_lock_slowpath at ffffffff9a86a013 #5 [ffffb87c4df07cb0] mutex_lock at ffffffff9a86a04f #6 [ffffb87c4df07cc8] kernfs_find_and_get_ns at ffffffff9a311d83 #7 [ffffb87c4df07cf0] sysfs_notify at ffffffff9a314b3a #8 [ffffb87c4df07d18] md_update_sb at ffffffff9a688696 #9 [ffffb87c4df07d98] md_update_sb at ffffffff9a6886d5 #10 [ffffb87c4df07da8] md_check_recovery at ffffffff9a68ad9c #11 [ffffb87c4df07dd0] raid1d at ffffffffc01f0375 [raid1] #12 [ffffb87c4df07ea0] md_thread at ffffffff9a680348 #13 [ffffb87c4df07f08] kthread at ffffffff9a0b8005 #14 [ffffb87c4df07f50] ret_from_fork at ffffffff9aa00344 Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com> Signed-off-by: Song Liu <songliubraving@fb.com>
2020-07-14 23:10:26 +00:00
sysfs_notify_dirent_safe(mddev->sysfs_completed);
}
INIT_LIST_HEAD(&stripes);
for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
int j;
int skipped_disk = 0;
sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
set_bit(STRIPE_EXPANDING, &sh->state);
atomic_inc(&conf->reshape_stripes);
/* If any of this stripe is beyond the end of the old
* array, then we need to zero those blocks
*/
for (j=sh->disks; j--;) {
sector_t s;
if (j == sh->pd_idx)
continue;
if (conf->level == 6 &&
j == sh->qd_idx)
continue;
s = raid5_compute_blocknr(sh, j, 0);
if (s < raid5_size(mddev, 0, 0)) {
skipped_disk = 1;
continue;
}
memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
set_bit(R5_Expanded, &sh->dev[j].flags);
set_bit(R5_UPTODATE, &sh->dev[j].flags);
}
if (!skipped_disk) {
set_bit(STRIPE_EXPAND_READY, &sh->state);
set_bit(STRIPE_HANDLE, &sh->state);
}
list_add(&sh->lru, &stripes);
}
spin_lock_irq(&conf->device_lock);
if (mddev->reshape_backwards)
conf->reshape_progress -= reshape_sectors * new_data_disks;
else
conf->reshape_progress += reshape_sectors * new_data_disks;
spin_unlock_irq(&conf->device_lock);
/* Ok, those stripe are ready. We can start scheduling
* reads on the source stripes.
* The source stripes are determined by mapping the first and last
* block on the destination stripes.
*/
first_sector =
raid5_compute_sector(conf, stripe_addr*(new_data_disks),
1, &dd_idx, NULL);
last_sector =
raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
* new_data_disks - 1),
1, &dd_idx, NULL);
if (last_sector >= mddev->dev_sectors)
last_sector = mddev->dev_sectors - 1;
while (first_sector <= last_sector) {
sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
first_sector += RAID5_STRIPE_SECTORS(conf);
}
/* Now that the sources are clearly marked, we can release
* the destination stripes
*/
while (!list_empty(&stripes)) {
sh = list_entry(stripes.next, struct stripe_head, lru);
list_del_init(&sh->lru);
raid5_release_stripe(sh);
}
/* If this takes us to the resync_max point where we have to pause,
* then we need to write out the superblock.
*/
sector_nr += reshape_sectors;
retn = reshape_sectors;
finish:
if (mddev->curr_resync_completed > mddev->resync_max ||
(sector_nr - mddev->curr_resync_completed) * 2
md: update sync_completed and reshape_position even more often. There are circumstances when a user-space process might need to "oversee" a resync/reshape process. For example when doing an in-place reshape of a raid5, it is prudent to take a backup of each section before reshaping it as this is the only way to provide safety against an unplanned shutdown (i.e. crash/power failure). The sync_max sysfs value can be used to stop the resync from advancing beyond a particular point. So user-space can: suspend IO to the first section and back it up set 'sync_max' to the end of the section wait for 'sync_completed' to reach that point resume IO on the first section and move on to the next section. However this process requires the kernel and user-space to run in lock-step which could introduce unnecessary delays. It would be better if a 'double buffered' approach could be used with userspace and kernel space working on different sections with the 'next' section always ready when the 'current' section is finished. One problem with implementing this is that sync_completed is only guaranteed to be updated when the sync process reaches sync_max. (it is updated on a time basis at other times, but it is hard to rely on that). This defeats some of the double buffering. With this patch, sync_completed (and reshape_position) get updated as the current position approaches sync_max, so there is room for userspace to advance sync_max early without losing updates. To be precise, sync_completed is updated when the current sync position reaches half way between the current value of sync_completed and the value of sync_max. This will usually be a good time for user space to update sync_max. If sync_max does not get updated, the updates to sync_completed (together with associated metadata updates) will occur at an exponentially increasing frequency which will get unreasonably fast (one update every page) immediately before the process hits sync_max and stops. So the update rate will be unreasonably fast only for an insignificant period of time. Signed-off-by: NeilBrown <neilb@suse.de>
2009-04-17 01:06:30 +00:00
>= mddev->resync_max - mddev->curr_resync_completed) {
/* Cannot proceed until we've updated the superblock... */
wait_event(conf->wait_for_overlap,
atomic_read(&conf->reshape_stripes) == 0
|| test_bit(MD_RECOVERY_INTR, &mddev->recovery));
if (atomic_read(&conf->reshape_stripes) != 0)
goto ret;
mddev->reshape_position = conf->reshape_progress;
mddev->curr_resync_completed = sector_nr;
if (!mddev->reshape_backwards)
/* Can update recovery_offset */
rdev_for_each(rdev, mddev)
if (rdev->raid_disk >= 0 &&
!test_bit(Journal, &rdev->flags) &&
!test_bit(In_sync, &rdev->flags) &&
rdev->recovery_offset < sector_nr)
rdev->recovery_offset = sector_nr;
conf->reshape_checkpoint = jiffies;
set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
md_wakeup_thread(mddev->thread);
wait_event(mddev->sb_wait,
!test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
|| test_bit(MD_RECOVERY_INTR, &mddev->recovery));
if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
goto ret;
spin_lock_irq(&conf->device_lock);
conf->reshape_safe = mddev->reshape_position;
spin_unlock_irq(&conf->device_lock);
wake_up(&conf->wait_for_overlap);
md: fix deadlock causing by sysfs_notify The following deadlock was captured. The first process is holding 'kernfs_mutex' and hung by io. The io was staging in 'r1conf.pending_bio_list' of raid1 device, this pending bio list would be flushed by second process 'md127_raid1', but it was hung by 'kernfs_mutex'. Using sysfs_notify_dirent_safe() to replace sysfs_notify() can fix it. There were other sysfs_notify() invoked from io path, removed all of them. PID: 40430 TASK: ffff8ee9c8c65c40 CPU: 29 COMMAND: "probe_file" #0 [ffffb87c4df37260] __schedule at ffffffff9a8678ec #1 [ffffb87c4df372f8] schedule at ffffffff9a867f06 #2 [ffffb87c4df37310] io_schedule at ffffffff9a0c73e6 #3 [ffffb87c4df37328] __dta___xfs_iunpin_wait_3443 at ffffffffc03a4057 [xfs] #4 [ffffb87c4df373a0] xfs_iunpin_wait at ffffffffc03a6c79 [xfs] #5 [ffffb87c4df373b0] __dta_xfs_reclaim_inode_3357 at ffffffffc039a46c [xfs] #6 [ffffb87c4df37400] xfs_reclaim_inodes_ag at ffffffffc039a8b6 [xfs] #7 [ffffb87c4df37590] xfs_reclaim_inodes_nr at ffffffffc039bb33 [xfs] #8 [ffffb87c4df375b0] xfs_fs_free_cached_objects at ffffffffc03af0e9 [xfs] #9 [ffffb87c4df375c0] super_cache_scan at ffffffff9a287ec7 #10 [ffffb87c4df37618] shrink_slab at ffffffff9a1efd93 #11 [ffffb87c4df37700] shrink_node at ffffffff9a1f5968 #12 [ffffb87c4df37788] do_try_to_free_pages at ffffffff9a1f5ea2 #13 [ffffb87c4df377f0] try_to_free_mem_cgroup_pages at ffffffff9a1f6445 #14 [ffffb87c4df37880] try_charge at ffffffff9a26cc5f #15 [ffffb87c4df37920] memcg_kmem_charge_memcg at ffffffff9a270f6a #16 [ffffb87c4df37958] new_slab at ffffffff9a251430 #17 [ffffb87c4df379c0] ___slab_alloc at ffffffff9a251c85 #18 [ffffb87c4df37a80] __slab_alloc at ffffffff9a25635d #19 [ffffb87c4df37ac0] kmem_cache_alloc at ffffffff9a251f89 #20 [ffffb87c4df37b00] alloc_inode at ffffffff9a2a2b10 #21 [ffffb87c4df37b20] iget_locked at ffffffff9a2a4854 #22 [ffffb87c4df37b60] kernfs_get_inode at ffffffff9a311377 #23 [ffffb87c4df37b80] kernfs_iop_lookup at ffffffff9a311e2b #24 [ffffb87c4df37ba8] lookup_slow at ffffffff9a290118 #25 [ffffb87c4df37c10] walk_component at ffffffff9a291e83 #26 [ffffb87c4df37c78] path_lookupat at ffffffff9a293619 #27 [ffffb87c4df37cd8] filename_lookup at ffffffff9a2953af #28 [ffffb87c4df37de8] user_path_at_empty at ffffffff9a295566 #29 [ffffb87c4df37e10] vfs_statx at ffffffff9a289787 #30 [ffffb87c4df37e70] SYSC_newlstat at ffffffff9a289d5d #31 [ffffb87c4df37f18] sys_newlstat at ffffffff9a28a60e #32 [ffffb87c4df37f28] do_syscall_64 at ffffffff9a003949 #33 [ffffb87c4df37f50] entry_SYSCALL_64_after_hwframe at ffffffff9aa001ad RIP: 00007f617a5f2905 RSP: 00007f607334f838 RFLAGS: 00000246 RAX: ffffffffffffffda RBX: 00007f6064044b20 RCX: 00007f617a5f2905 RDX: 00007f6064044b20 RSI: 00007f6064044b20 RDI: 00007f6064005890 RBP: 00007f6064044aa0 R8: 0000000000000030 R9: 000000000000011c R10: 0000000000000013 R11: 0000000000000246 R12: 00007f606417e6d0 R13: 00007f6064044aa0 R14: 00007f6064044b10 R15: 00000000ffffffff ORIG_RAX: 0000000000000006 CS: 0033 SS: 002b PID: 927 TASK: ffff8f15ac5dbd80 CPU: 42 COMMAND: "md127_raid1" #0 [ffffb87c4df07b28] __schedule at ffffffff9a8678ec #1 [ffffb87c4df07bc0] schedule at ffffffff9a867f06 #2 [ffffb87c4df07bd8] schedule_preempt_disabled at ffffffff9a86825e #3 [ffffb87c4df07be8] __mutex_lock at ffffffff9a869bcc #4 [ffffb87c4df07ca0] __mutex_lock_slowpath at ffffffff9a86a013 #5 [ffffb87c4df07cb0] mutex_lock at ffffffff9a86a04f #6 [ffffb87c4df07cc8] kernfs_find_and_get_ns at ffffffff9a311d83 #7 [ffffb87c4df07cf0] sysfs_notify at ffffffff9a314b3a #8 [ffffb87c4df07d18] md_update_sb at ffffffff9a688696 #9 [ffffb87c4df07d98] md_update_sb at ffffffff9a6886d5 #10 [ffffb87c4df07da8] md_check_recovery at ffffffff9a68ad9c #11 [ffffb87c4df07dd0] raid1d at ffffffffc01f0375 [raid1] #12 [ffffb87c4df07ea0] md_thread at ffffffff9a680348 #13 [ffffb87c4df07f08] kthread at ffffffff9a0b8005 #14 [ffffb87c4df07f50] ret_from_fork at ffffffff9aa00344 Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com> Signed-off-by: Song Liu <songliubraving@fb.com>
2020-07-14 23:10:26 +00:00
sysfs_notify_dirent_safe(mddev->sysfs_completed);
}
ret:
return retn;
}
static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
int *skipped)
{
struct r5conf *conf = mddev->private;
struct stripe_head *sh;
sector_t max_sector = mddev->dev_sectors;
sector_t sync_blocks;
int still_degraded = 0;
int i;
if (sector_nr >= max_sector) {
/* just being told to finish up .. nothing much to do */
if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
end_reshape(conf);
return 0;
}
if (mddev->curr_resync < max_sector) /* aborted */
md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
&sync_blocks, 1);
else /* completed sync */
conf->fullsync = 0;
md_bitmap_close_sync(mddev->bitmap);
return 0;
}
/* Allow raid5_quiesce to complete */
wait_event(conf->wait_for_overlap, conf->quiesce != 2);
if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
return reshape_request(mddev, sector_nr, skipped);
/* No need to check resync_max as we never do more than one
* stripe, and as resync_max will always be on a chunk boundary,
* if the check in md_do_sync didn't fire, there is no chance
* of overstepping resync_max here
*/
/* if there is too many failed drives and we are trying
* to resync, then assert that we are finished, because there is
* nothing we can do.
*/
if (mddev->degraded >= conf->max_degraded &&
test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
sector_t rv = mddev->dev_sectors - sector_nr;
*skipped = 1;
return rv;
}
if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
!conf->fullsync &&
!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
/* we can skip this block, and probably more */
do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
*skipped = 1;
/* keep things rounded to whole stripes */
return sync_blocks * RAID5_STRIPE_SECTORS(conf);
}
md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
if (sh == NULL) {
sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
/* make sure we don't swamp the stripe cache if someone else
* is trying to get access
*/
schedule_timeout_uninterruptible(1);
}
/* Need to check if array will still be degraded after recovery/resync
* Note in case of > 1 drive failures it's possible we're rebuilding
* one drive while leaving another faulty drive in array.
*/
rcu_read_lock();
for (i = 0; i < conf->raid_disks; i++) {
locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE() Please do not apply this to mainline directly, instead please re-run the coccinelle script shown below and apply its output. For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't harmful, and changing them results in churn. However, for some features, the read/write distinction is critical to correct operation. To distinguish these cases, separate read/write accessors must be used. This patch migrates (most) remaining ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following coccinelle script: ---- // Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and // WRITE_ONCE() // $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: tj@kernel.org Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 21:07:29 +00:00
struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
if (rdev == NULL || test_bit(Faulty, &rdev->flags))
still_degraded = 1;
}
rcu_read_unlock();
md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
set_bit(STRIPE_HANDLE, &sh->state);
raid5_release_stripe(sh);
return RAID5_STRIPE_SECTORS(conf);
}
static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
unsigned int offset)
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
{
/* We may not be able to submit a whole bio at once as there
* may not be enough stripe_heads available.
* We cannot pre-allocate enough stripe_heads as we may need
* more than exist in the cache (if we allow ever large chunks).
* So we do one stripe head at a time and record in
* ->bi_hw_segments how many have been done.
*
* We *know* that this entire raid_bio is in one chunk, so
* it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
*/
struct stripe_head *sh;
int dd_idx;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
sector_t sector, logical_sector, last_sector;
int scnt = 0;
int handled = 0;
block: Abstract out bvec iterator Immutable biovecs are going to require an explicit iterator. To implement immutable bvecs, a later patch is going to add a bi_bvec_done member to this struct; for now, this patch effectively just renames things. Signed-off-by: Kent Overstreet <kmo@daterainc.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: "Ed L. Cashin" <ecashin@coraid.com> Cc: Nick Piggin <npiggin@kernel.dk> Cc: Lars Ellenberg <drbd-dev@lists.linbit.com> Cc: Jiri Kosina <jkosina@suse.cz> Cc: Matthew Wilcox <willy@linux.intel.com> Cc: Geoff Levand <geoff@infradead.org> Cc: Yehuda Sadeh <yehuda@inktank.com> Cc: Sage Weil <sage@inktank.com> Cc: Alex Elder <elder@inktank.com> Cc: ceph-devel@vger.kernel.org Cc: Joshua Morris <josh.h.morris@us.ibm.com> Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Neil Brown <neilb@suse.de> Cc: Alasdair Kergon <agk@redhat.com> Cc: Mike Snitzer <snitzer@redhat.com> Cc: dm-devel@redhat.com Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: linux390@de.ibm.com Cc: Boaz Harrosh <bharrosh@panasas.com> Cc: Benny Halevy <bhalevy@tonian.com> Cc: "James E.J. Bottomley" <JBottomley@parallels.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Nicholas A. Bellinger" <nab@linux-iscsi.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Chris Mason <chris.mason@fusionio.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Jaegeuk Kim <jaegeuk.kim@samsung.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@kernel.org> Cc: Joern Engel <joern@logfs.org> Cc: Prasad Joshi <prasadjoshi.linux@gmail.com> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: KONISHI Ryusuke <konishi.ryusuke@lab.ntt.co.jp> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Ben Myers <bpm@sgi.com> Cc: xfs@oss.sgi.com Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Len Brown <len.brown@intel.com> Cc: Pavel Machek <pavel@ucw.cz> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Guo Chao <yan@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Wei Yongjun <yongjun_wei@trendmicro.com.cn> Cc: "Roger Pau Monné" <roger.pau@citrix.com> Cc: Jan Beulich <jbeulich@suse.com> Cc: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Cc: Ian Campbell <Ian.Campbell@citrix.com> Cc: Sebastian Ott <sebott@linux.vnet.ibm.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Jerome Marchand <jmarchand@redhat.com> Cc: Joe Perches <joe@perches.com> Cc: Peng Tao <tao.peng@emc.com> Cc: Andy Adamson <andros@netapp.com> Cc: fanchaoting <fanchaoting@cn.fujitsu.com> Cc: Jie Liu <jeff.liu@oracle.com> Cc: Sunil Mushran <sunil.mushran@gmail.com> Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Namjae Jeon <namjae.jeon@samsung.com> Cc: Pankaj Kumar <pankaj.km@samsung.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Mel Gorman <mgorman@suse.de>6
2013-10-11 22:44:27 +00:00
logical_sector = raid_bio->bi_iter.bi_sector &
~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
sector = raid5_compute_sector(conf, logical_sector,
0, &dd_idx, NULL);
last_sector = bio_end_sector(raid_bio);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
for (; logical_sector < last_sector;
logical_sector += RAID5_STRIPE_SECTORS(conf),
sector += RAID5_STRIPE_SECTORS(conf),
scnt++) {
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
if (scnt < offset)
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
/* already done this stripe */
continue;
sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
if (!sh) {
/* failed to get a stripe - must wait */
conf->retry_read_aligned = raid_bio;
conf->retry_read_offset = scnt;
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
return handled;
}
if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
raid5_release_stripe(sh);
conf->retry_read_aligned = raid_bio;
conf->retry_read_offset = scnt;
return handled;
}
set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
handle_stripe(sh);
raid5_release_stripe(sh);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
handled++;
}
bio_endio(raid_bio);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
if (atomic_dec_and_test(&conf->active_aligned_reads))
wake_up(&conf->wait_for_quiescent);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
return handled;
}
static int handle_active_stripes(struct r5conf *conf, int group,
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
struct r5worker *worker,
struct list_head *temp_inactive_list)
__releases(&conf->device_lock)
__acquires(&conf->device_lock)
{
struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int i, batch_size = 0, hash;
bool release_inactive = false;
while (batch_size < MAX_STRIPE_BATCH &&
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
(sh = __get_priority_stripe(conf, group)) != NULL)
batch[batch_size++] = sh;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (batch_size == 0) {
for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
if (!list_empty(temp_inactive_list + i))
break;
if (i == NR_STRIPE_HASH_LOCKS) {
spin_unlock_irq(&conf->device_lock);
log_flush_stripe_to_raid(conf);
spin_lock_irq(&conf->device_lock);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
return batch_size;
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
release_inactive = true;
}
spin_unlock_irq(&conf->device_lock);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
release_inactive_stripe_list(conf, temp_inactive_list,
NR_STRIPE_HASH_LOCKS);
r5l_flush_stripe_to_raid(conf->log);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
if (release_inactive) {
spin_lock_irq(&conf->device_lock);
return 0;
}
for (i = 0; i < batch_size; i++)
handle_stripe(batch[i]);
log_write_stripe_run(conf);
cond_resched();
spin_lock_irq(&conf->device_lock);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
for (i = 0; i < batch_size; i++) {
hash = batch[i]->hash_lock_index;
__release_stripe(conf, batch[i], &temp_inactive_list[hash]);
}
return batch_size;
}
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
static void raid5_do_work(struct work_struct *work)
{
struct r5worker *worker = container_of(work, struct r5worker, work);
struct r5worker_group *group = worker->group;
struct r5conf *conf = group->conf;
struct mddev *mddev = conf->mddev;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
int group_id = group - conf->worker_groups;
int handled;
struct blk_plug plug;
pr_debug("+++ raid5worker active\n");
blk_start_plug(&plug);
handled = 0;
spin_lock_irq(&conf->device_lock);
while (1) {
int batch_size, released;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
released = release_stripe_list(conf, worker->temp_inactive_list);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
batch_size = handle_active_stripes(conf, group_id, worker,
worker->temp_inactive_list);
worker->working = false;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
if (!batch_size && !released)
break;
handled += batch_size;
wait_event_lock_irq(mddev->sb_wait,
!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
conf->device_lock);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
}
pr_debug("%d stripes handled\n", handled);
spin_unlock_irq(&conf->device_lock);
flush_deferred_bios(conf);
r5l_flush_stripe_to_raid(conf->log);
async_tx_issue_pending_all();
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
blk_finish_plug(&plug);
pr_debug("--- raid5worker inactive\n");
}
/*
* This is our raid5 kernel thread.
*
* We scan the hash table for stripes which can be handled now.
* During the scan, completed stripes are saved for us by the interrupt
* handler, so that they will not have to wait for our next wakeup.
*/
static void raid5d(struct md_thread *thread)
{
struct mddev *mddev = thread->mddev;
struct r5conf *conf = mddev->private;
int handled;
struct blk_plug plug;
pr_debug("+++ raid5d active\n");
md_check_recovery(mddev);
blk_start_plug(&plug);
handled = 0;
spin_lock_irq(&conf->device_lock);
while (1) {
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
struct bio *bio;
int batch_size, released;
unsigned int offset;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
released = release_stripe_list(conf, conf->temp_inactive_list);
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
if (released)
clear_bit(R5_DID_ALLOC, &conf->cache_state);
if (
!list_empty(&conf->bitmap_list)) {
/* Now is a good time to flush some bitmap updates */
conf->seq_flush++;
spin_unlock_irq(&conf->device_lock);
md_bitmap_unplug(mddev->bitmap);
spin_lock_irq(&conf->device_lock);
conf->seq_write = conf->seq_flush;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
activate_bit_delay(conf, conf->temp_inactive_list);
}
raid5_activate_delayed(conf);
while ((bio = remove_bio_from_retry(conf, &offset))) {
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
int ok;
spin_unlock_irq(&conf->device_lock);
ok = retry_aligned_read(conf, bio, offset);
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
spin_lock_irq(&conf->device_lock);
if (!ok)
break;
handled++;
}
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
conf->temp_inactive_list);
if (!batch_size && !released)
break;
handled += batch_size;
if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
spin_unlock_irq(&conf->device_lock);
md: make it easier to wait for bad blocks to be acknowledged. It is only safe to choose not to write to a bad block if that bad block is safely recorded in metadata - i.e. if it has been 'acknowledged'. If it hasn't we need to wait for the acknowledgement. We support that using rdev->blocked wait and md_wait_for_blocked_rdev by introducing a new device flag 'BlockedBadBlock'. This flag is only advisory. It is cleared whenever we acknowledge a bad block, so that a waiter can re-check the particular bad blocks that it is interested it. It should be set by a caller when they find they need to wait. This (set after test) is inherently racy, but as md_wait_for_blocked_rdev already has a timeout, losing the race will have minimal impact. When we clear "Blocked" was also clear "BlockedBadBlocks" incase it was set incorrectly (see above race). We also modify the way we manage 'Blocked' to fit better with the new handling of 'BlockedBadBlocks' and to make it consistent between externally managed and internally managed metadata. This requires that each raidXd loop checks if the metadata needs to be written and triggers a write (md_check_recovery) if needed. Otherwise a queued write request might cause raidXd to wait for the metadata to write, and only that thread can write it. Before writing metadata, we set FaultRecorded for all devices that are Faulty, then after writing the metadata we clear Blocked for any device for which the Fault was certainly Recorded. The 'faulty' device flag now appears in sysfs if the device is faulty *or* it has unacknowledged bad blocks. So user-space which does not understand bad blocks can continue to function correctly. User space which does, should not assume a device is faulty until it sees the 'faulty' flag, and then sees the list of unacknowledged bad blocks is empty. Signed-off-by: NeilBrown <neilb@suse.de>
2011-07-28 01:31:48 +00:00
md_check_recovery(mddev);
spin_lock_irq(&conf->device_lock);
}
}
pr_debug("%d stripes handled\n", handled);
spin_unlock_irq(&conf->device_lock);
if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
mutex_trylock(&conf->cache_size_mutex)) {
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
grow_one_stripe(conf, __GFP_NOWARN);
/* Set flag even if allocation failed. This helps
* slow down allocation requests when mem is short
*/
set_bit(R5_DID_ALLOC, &conf->cache_state);
mutex_unlock(&conf->cache_size_mutex);
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
}
flush_deferred_bios(conf);
raid5: log reclaim support This is the reclaim support for raid5 log. A stripe write will have following steps: 1. reconstruct the stripe, read data/calculate parity. ops_run_io prepares to write data/parity to raid disks 2. hijack ops_run_io. stripe data/parity is appending to log disk 3. flush log disk cache 4. ops_run_io run again and do normal operation. stripe data/parity is written in raid array disks. raid core can return io to upper layer. 5. flush cache of all raid array disks 6. update super block 7. log disk space used by the stripe can be reused In practice, several stripes consist of an io_unit and we will batch several io_unit in different steps, but the whole process doesn't change. It's possible io return just after data/parity hit log disk, but then read IO will need read from log disk. For simplicity, IO return happens at step 4, where read IO can directly read from raid disks. Currently reclaim run if there is specific reclaimable space (1/4 disk size or 10G) or we are out of space. Reclaim is just to free log disk spaces, it doesn't impact data consistency. The size based force reclaim is to make sure log isn't too big, so recovery doesn't scan log too much. Recovery make sure raid disks and log disk have the same data of a stripe. If crash happens before 4, recovery might/might not recovery stripe's data/parity depending on if data/parity and its checksum matches. In either case, this doesn't change the syntax of an IO write. After step 3, stripe is guaranteed recoverable, because stripe's data/parity is persistent in log disk. In some cases, log disk content and raid disks content of a stripe are the same, but recovery will still copy log disk content to raid disks, this doesn't impact data consistency. space reuse happens after superblock update and cache flush. There is one situation we want to avoid. A broken meta in the middle of a log causes recovery can't find meta at the head of log. If operations require meta at the head persistent in log, we must make sure meta before it persistent in log too. The case is stripe data/parity is in log and we start write stripe to raid disks (before step 4). stripe data/parity must be persistent in log before we do the write to raid disks. The solution is we restrictly maintain io_unit list order. In this case, we only write stripes of an io_unit to raid disks till the io_unit is the first one whose data/parity is in log. The io_unit list order is important for other cases too. For example, some io_unit are reclaimable and others not. They can be mixed in the list, we shouldn't reuse space of an unreclaimable io_unit. Includes fixes to problems which were... Reported-by: kbuild test robot <fengguang.wu@intel.com> Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: NeilBrown <neilb@suse.com>
2015-08-13 21:32:00 +00:00
r5l_flush_stripe_to_raid(conf->log);
async_tx_issue_pending_all();
blk_finish_plug(&plug);
pr_debug("--- raid5d inactive\n");
}
static ssize_t
raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
{
struct r5conf *conf;
int ret = 0;
spin_lock(&mddev->lock);
conf = mddev->private;
if (conf)
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
ret = sprintf(page, "%d\n", conf->min_nr_stripes);
spin_unlock(&mddev->lock);
return ret;
}
int
raid5_set_cache_size(struct mddev *mddev, int size)
{
int result = 0;
struct r5conf *conf = mddev->private;
if (size <= 16 || size > 32768)
return -EINVAL;
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
conf->min_nr_stripes = size;
mutex_lock(&conf->cache_size_mutex);
while (size < conf->max_nr_stripes &&
drop_one_stripe(conf))
;
mutex_unlock(&conf->cache_size_mutex);
md_allow_write(mddev);
mutex_lock(&conf->cache_size_mutex);
while (size > conf->max_nr_stripes)
if (!grow_one_stripe(conf, GFP_KERNEL)) {
conf->min_nr_stripes = conf->max_nr_stripes;
result = -ENOMEM;
break;
}
mutex_unlock(&conf->cache_size_mutex);
return result;
}
EXPORT_SYMBOL(raid5_set_cache_size);
static ssize_t
raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
{
struct r5conf *conf;
unsigned long new;
int err;
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtoul(page, 10, &new))
return -EINVAL;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf)
err = -ENODEV;
else
err = raid5_set_cache_size(mddev, new);
mddev_unlock(mddev);
return err ?: len;
}
static struct md_sysfs_entry
raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
raid5_show_stripe_cache_size,
raid5_store_stripe_cache_size);
static ssize_t
raid5_show_rmw_level(struct mddev *mddev, char *page)
{
struct r5conf *conf = mddev->private;
if (conf)
return sprintf(page, "%d\n", conf->rmw_level);
else
return 0;
}
static ssize_t
raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
{
struct r5conf *conf = mddev->private;
unsigned long new;
if (!conf)
return -ENODEV;
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtoul(page, 10, &new))
return -EINVAL;
if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
return -EINVAL;
if (new != PARITY_DISABLE_RMW &&
new != PARITY_ENABLE_RMW &&
new != PARITY_PREFER_RMW)
return -EINVAL;
conf->rmw_level = new;
return len;
}
static struct md_sysfs_entry
raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
raid5_show_rmw_level,
raid5_store_rmw_level);
static ssize_t
raid5_show_stripe_size(struct mddev *mddev, char *page)
{
struct r5conf *conf;
int ret = 0;
spin_lock(&mddev->lock);
conf = mddev->private;
if (conf)
ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
spin_unlock(&mddev->lock);
return ret;
}
#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
static ssize_t
raid5_store_stripe_size(struct mddev *mddev, const char *page, size_t len)
{
struct r5conf *conf;
unsigned long new;
int err;
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtoul(page, 10, &new))
return -EINVAL;
/*
* The value should not be bigger than PAGE_SIZE. It requires to
* be multiple of DEFAULT_STRIPE_SIZE.
*/
if (new % DEFAULT_STRIPE_SIZE != 0 || new > PAGE_SIZE || new == 0)
return -EINVAL;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf) {
err = -ENODEV;
goto out_unlock;
}
if (new == conf->stripe_size)
goto out_unlock;
pr_debug("md/raid: change stripe_size from %lu to %lu\n",
conf->stripe_size, new);
mddev_suspend(mddev);
conf->stripe_size = new;
conf->stripe_shift = ilog2(new) - 9;
conf->stripe_sectors = new >> 9;
mddev_resume(mddev);
out_unlock:
mddev_unlock(mddev);
return err ?: len;
}
static struct md_sysfs_entry
raid5_stripe_size = __ATTR(stripe_size, 0644,
raid5_show_stripe_size,
raid5_store_stripe_size);
#else
static struct md_sysfs_entry
raid5_stripe_size = __ATTR(stripe_size, 0444,
raid5_show_stripe_size,
NULL);
#endif
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
static ssize_t
raid5_show_preread_threshold(struct mddev *mddev, char *page)
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
{
struct r5conf *conf;
int ret = 0;
spin_lock(&mddev->lock);
conf = mddev->private;
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (conf)
ret = sprintf(page, "%d\n", conf->bypass_threshold);
spin_unlock(&mddev->lock);
return ret;
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
}
static ssize_t
raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
{
struct r5conf *conf;
unsigned long new;
int err;
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtoul(page, 10, &new))
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
return -EINVAL;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf)
err = -ENODEV;
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
else if (new > conf->min_nr_stripes)
err = -EINVAL;
else
conf->bypass_threshold = new;
mddev_unlock(mddev);
return err ?: len;
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
}
static struct md_sysfs_entry
raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
S_IRUGO | S_IWUSR,
raid5_show_preread_threshold,
raid5_store_preread_threshold);
static ssize_t
raid5_show_skip_copy(struct mddev *mddev, char *page)
{
struct r5conf *conf;
int ret = 0;
spin_lock(&mddev->lock);
conf = mddev->private;
if (conf)
ret = sprintf(page, "%d\n", conf->skip_copy);
spin_unlock(&mddev->lock);
return ret;
}
static ssize_t
raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
{
struct r5conf *conf;
unsigned long new;
int err;
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtoul(page, 10, &new))
return -EINVAL;
new = !!new;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf)
err = -ENODEV;
else if (new != conf->skip_copy) {
mddev_suspend(mddev);
conf->skip_copy = new;
if (new)
mddev->queue->backing_dev_info->capabilities |=
BDI_CAP_STABLE_WRITES;
else
mddev->queue->backing_dev_info->capabilities &=
~BDI_CAP_STABLE_WRITES;
mddev_resume(mddev);
}
mddev_unlock(mddev);
return err ?: len;
}
static struct md_sysfs_entry
raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
raid5_show_skip_copy,
raid5_store_skip_copy);
static ssize_t
stripe_cache_active_show(struct mddev *mddev, char *page)
{
struct r5conf *conf = mddev->private;
if (conf)
return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
else
return 0;
}
static struct md_sysfs_entry
raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
static ssize_t
raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
{
struct r5conf *conf;
int ret = 0;
spin_lock(&mddev->lock);
conf = mddev->private;
if (conf)
ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
spin_unlock(&mddev->lock);
return ret;
}
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
static int alloc_thread_groups(struct r5conf *conf, int cnt,
int *group_cnt,
struct r5worker_group **worker_groups);
static ssize_t
raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
{
struct r5conf *conf;
unsigned int new;
int err;
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
struct r5worker_group *new_groups, *old_groups;
int group_cnt;
if (len >= PAGE_SIZE)
return -EINVAL;
if (kstrtouint(page, 10, &new))
return -EINVAL;
/* 8192 should be big enough */
if (new > 8192)
return -EINVAL;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf)
err = -ENODEV;
else if (new != conf->worker_cnt_per_group) {
mddev_suspend(mddev);
old_groups = conf->worker_groups;
if (old_groups)
flush_workqueue(raid5_wq);
md/raid5: Before freeing old multi-thread worker, it should flush them. When changing group_thread_cnt from sysfs entry, the kernel can oops. The kernel messages are: [ 740.961389] BUG: unable to handle kernel NULL pointer dereference at 0000000000000008 [ 740.961444] IP: [<ffffffff81062570>] process_one_work+0x30/0x500 [ 740.961476] PGD b9013067 PUD b651e067 PMD 0 [ 740.961503] Oops: 0000 [#1] SMP [ 740.961525] Modules linked in: netconsole e1000e ptp pps_core [ 740.961577] CPU: 0 PID: 3683 Comm: kworker/u8:5 Not tainted 3.12.0+ #23 [ 740.961602] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 740.961646] task: ffff88013abe0000 ti: ffff88013a246000 task.ti: ffff88013a246000 [ 740.961673] RIP: 0010:[<ffffffff81062570>] [<ffffffff81062570>] process_one_work+0x30/0x500 [ 740.961708] RSP: 0018:ffff88013a247e08 EFLAGS: 00010086 [ 740.961730] RAX: ffff8800b912b400 RBX: ffff88013a61e680 RCX: ffff8800b912b400 [ 740.961757] RDX: ffff8800b912b600 RSI: ffff8800b912b600 RDI: ffff88013a61e680 [ 740.961782] RBP: ffff88013a247e48 R08: ffff88013a246000 R09: 000000000002c09d [ 740.961808] R10: 000000000000010f R11: 0000000000000000 R12: ffff88013b00cc00 [ 740.961833] R13: 0000000000000000 R14: ffff88013b00cf80 R15: ffff88013a61e6b0 [ 740.961861] FS: 0000000000000000(0000) GS:ffff88013fc00000(0000) knlGS:0000000000000000 [ 740.961893] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 740.962001] CR2: 00000000000000b8 CR3: 00000000b24fe000 CR4: 00000000000407f0 [ 740.962001] Stack: [ 740.962001] 0000000000000008 ffff8800b912b600 ffff88013b00cc00 ffff88013a61e680 [ 740.962001] ffff88013b00cc00 ffff88013b00cc18 ffff88013b00cf80 ffff88013a61e6b0 [ 740.962001] ffff88013a247eb8 ffffffff810639c6 0000000000012a80 ffff88013a247fd8 [ 740.962001] Call Trace: [ 740.962001] [<ffffffff810639c6>] worker_thread+0x206/0x3f0 [ 740.962001] [<ffffffff810637c0>] ? manage_workers+0x2c0/0x2c0 [ 740.962001] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 740.962001] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 740.962001] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 740.962001] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 740.962001] Code: 89 e5 41 57 41 56 41 55 45 31 ed 41 54 53 48 89 fb 48 83 ec 18 48 8b 06 4c 8b 67 48 48 89 c1 30 c9 a8 04 4c 0f 45 e9 80 7f 58 00 <49> 8b 45 08 44 8b b0 00 01 00 00 78 0c 41 f6 44 24 10 04 0f 84 [ 740.962001] RIP [<ffffffff81062570>] process_one_work+0x30/0x500 [ 740.962001] RSP <ffff88013a247e08> [ 740.962001] CR2: 0000000000000008 [ 740.962001] ---[ end trace 39181460000748de ]--- [ 740.962001] Kernel panic - not syncing: Fatal exception This can happen if there are some stripes left, fewer than MAX_STRIPE_BATCH. A worker is queued to handle them. But before calling raid5_do_work, raid5d handles those stripes making conf->active_stripe = 0. So mddev_suspend() can return. We might then free old worker resources before the queued raid5_do_work() handled them. When it runs, it crashes. raid5d() raid5_store_group_thread_cnt() queue_work mddev_suspend() handle_strips active_stripe=0 free(old worker resources) process_one_work raid5_do_work To avoid this, we should only flush the worker resources before freeing them. This fixes a bug introduced in 3.12 so is suitable for the 3.12.x stable series. Cc: stable@vger.kernel.org (3.12) Fixes: b721420e8719131896b009b11edbbd27 Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:19 +00:00
err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
if (!err) {
spin_lock_irq(&conf->device_lock);
conf->group_cnt = group_cnt;
conf->worker_cnt_per_group = new;
conf->worker_groups = new_groups;
spin_unlock_irq(&conf->device_lock);
if (old_groups)
kfree(old_groups[0].workers);
kfree(old_groups);
}
mddev_resume(mddev);
}
mddev_unlock(mddev);
return err ?: len;
}
static struct md_sysfs_entry
raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
raid5_show_group_thread_cnt,
raid5_store_group_thread_cnt);
static struct attribute *raid5_attrs[] = {
&raid5_stripecache_size.attr,
&raid5_stripecache_active.attr,
md: introduce get_priority_stripe() to improve raid456 write performance Improve write performance by preventing the delayed_list from dumping all its stripes onto the handle_list in one shot. Delayed stripes are now further delayed by being held on the 'hold_list'. The 'hold_list' is bypassed when: * a STRIPE_IO_STARTED stripe is found at the head of 'handle_list' * 'handle_list' is empty and i/o is being done to satisfy full stripe-width write requests * 'bypass_count' is less than 'bypass_threshold'. By default the threshold is 1, i.e. every other stripe handled is a preread stripe provided the top two conditions are false. Benchmark data: System: 2x Xeon 5150, 4x SATA, mem=1GB Baseline: 2.6.24-rc7 Configuration: mdadm --create /dev/md0 /dev/sd[b-e] -n 4 -l 5 --assume-clean Test1: dd if=/dev/zero of=/dev/md0 bs=1024k count=2048 * patched: +33% (stripe_cache_size = 256), +25% (stripe_cache_size = 512) Test2: tiobench --size 2048 --numruns 5 --block 4096 --block 131072 (XFS) * patched: +13% * patched + preread_bypass_threshold = 0: +37% Changes since v1: * reduce bypass_threshold from (chunk_size / sectors_per_chunk) to (1) and make it configurable. This defaults to fairness and modest performance gains out of the box. Changes since v2: * [neilb@suse.de]: kill STRIPE_PRIO_HI and preread_needed as they are not necessary, the important change was clearing STRIPE_DELAYED in add_stripe_bio and this has been moved out to make_request for the hang fix. * [neilb@suse.de]: simplify get_priority_stripe * [dan.j.williams@intel.com]: reset the bypass_count when ->hold_list is sampled empty (+11%) * [dan.j.williams@intel.com]: decrement the bypass_count at the detection of stripes being naturally promoted off of hold_list +2%. Note, resetting bypass_count instead of decrementing on these events yields +4% but that is probably too aggressive. Changes since v3: * cosmetic fixups Tested-by: James W. Laferriere <babydr@baby-dragons.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 09:15:53 +00:00
&raid5_preread_bypass_threshold.attr,
&raid5_group_thread_cnt.attr,
&raid5_skip_copy.attr,
&raid5_rmw_level.attr,
&raid5_stripe_size.attr,
&r5c_journal_mode.attr,
&ppl_write_hint.attr,
NULL,
};
static struct attribute_group raid5_attrs_group = {
.name = NULL,
.attrs = raid5_attrs,
};
static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
struct r5worker_group **worker_groups)
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
{
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int i, j, k;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
ssize_t size;
struct r5worker *workers;
if (cnt == 0) {
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
*group_cnt = 0;
*worker_groups = NULL;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
return 0;
}
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
*group_cnt = num_possible_nodes();
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
size = sizeof(struct r5worker) * cnt;
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
workers = kcalloc(size, *group_cnt, GFP_NOIO);
*worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
GFP_NOIO);
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
if (!*worker_groups || !workers) {
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
kfree(workers);
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
kfree(*worker_groups);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
return -ENOMEM;
}
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
for (i = 0; i < *group_cnt; i++) {
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
struct r5worker_group *group;
group = &(*worker_groups)[i];
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
INIT_LIST_HEAD(&group->handle_list);
INIT_LIST_HEAD(&group->loprio_list);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
group->conf = conf;
group->workers = workers + i * cnt;
for (j = 0; j < cnt; j++) {
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
struct r5worker *worker = group->workers + j;
worker->group = group;
INIT_WORK(&worker->work, raid5_do_work);
for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
INIT_LIST_HEAD(worker->temp_inactive_list + k);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
}
}
return 0;
}
static void free_thread_groups(struct r5conf *conf)
{
if (conf->worker_groups)
kfree(conf->worker_groups[0].workers);
kfree(conf->worker_groups);
conf->worker_groups = NULL;
}
static sector_t
raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
{
struct r5conf *conf = mddev->private;
if (!sectors)
sectors = mddev->dev_sectors;
if (!raid_disks)
/* size is defined by the smallest of previous and new size */
raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
sectors &= ~((sector_t)conf->chunk_sectors - 1);
sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
return sectors * (raid_disks - conf->max_degraded);
}
static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
{
safe_put_page(percpu->spare_page);
percpu->spare_page = NULL;
kvfree(percpu->scribble);
percpu->scribble = NULL;
}
static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
{
if (conf->level == 6 && !percpu->spare_page) {
percpu->spare_page = alloc_page(GFP_KERNEL);
if (!percpu->spare_page)
return -ENOMEM;
}
if (scribble_alloc(percpu,
max(conf->raid_disks,
conf->previous_raid_disks),
max(conf->chunk_sectors,
conf->prev_chunk_sectors)
/ RAID5_STRIPE_SECTORS(conf))) {
free_scratch_buffer(conf, percpu);
return -ENOMEM;
}
return 0;
}
static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
{
struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
return 0;
}
static void raid5_free_percpu(struct r5conf *conf)
{
if (!conf->percpu)
return;
cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
free_percpu(conf->percpu);
}
static void free_conf(struct r5conf *conf)
{
int i;
log_exit(conf);
unregister_shrinker(&conf->shrinker);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
free_thread_groups(conf);
shrink_stripes(conf);
raid5_free_percpu(conf);
for (i = 0; i < conf->pool_size; i++)
if (conf->disks[i].extra_page)
put_page(conf->disks[i].extra_page);
kfree(conf->disks);
bioset_exit(&conf->bio_split);
kfree(conf->stripe_hashtbl);
kfree(conf->pending_data);
kfree(conf);
}
static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
{
struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
if (alloc_scratch_buffer(conf, percpu)) {
pr_warn("%s: failed memory allocation for cpu%u\n",
__func__, cpu);
return -ENOMEM;
}
return 0;
}
static int raid5_alloc_percpu(struct r5conf *conf)
{
int err = 0;
conf->percpu = alloc_percpu(struct raid5_percpu);
if (!conf->percpu)
return -ENOMEM;
err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
if (!err) {
conf->scribble_disks = max(conf->raid_disks,
conf->previous_raid_disks);
conf->scribble_sectors = max(conf->chunk_sectors,
conf->prev_chunk_sectors);
}
return err;
}
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
static unsigned long raid5_cache_scan(struct shrinker *shrink,
struct shrink_control *sc)
{
struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
unsigned long ret = SHRINK_STOP;
if (mutex_trylock(&conf->cache_size_mutex)) {
ret= 0;
while (ret < sc->nr_to_scan &&
conf->max_nr_stripes > conf->min_nr_stripes) {
if (drop_one_stripe(conf) == 0) {
ret = SHRINK_STOP;
break;
}
ret++;
}
mutex_unlock(&conf->cache_size_mutex);
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
}
return ret;
}
static unsigned long raid5_cache_count(struct shrinker *shrink,
struct shrink_control *sc)
{
struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
if (conf->max_nr_stripes < conf->min_nr_stripes)
/* unlikely, but not impossible */
return 0;
return conf->max_nr_stripes - conf->min_nr_stripes;
}
static struct r5conf *setup_conf(struct mddev *mddev)
{
struct r5conf *conf;
int raid_disk, memory, max_disks;
struct md_rdev *rdev;
struct disk_info *disk;
char pers_name[6];
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
int i;
int group_cnt;
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
struct r5worker_group *new_group;
int ret;
if (mddev->new_level != 5
&& mddev->new_level != 4
&& mddev->new_level != 6) {
pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
mdname(mddev), mddev->new_level);
return ERR_PTR(-EIO);
}
if ((mddev->new_level == 5
&& !algorithm_valid_raid5(mddev->new_layout)) ||
(mddev->new_level == 6
&& !algorithm_valid_raid6(mddev->new_layout))) {
pr_warn("md/raid:%s: layout %d not supported\n",
mdname(mddev), mddev->new_layout);
return ERR_PTR(-EIO);
}
if (mddev->new_level == 6 && mddev->raid_disks < 4) {
pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
mdname(mddev), mddev->raid_disks);
return ERR_PTR(-EINVAL);
}
if (!mddev->new_chunk_sectors ||
(mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
!is_power_of_2(mddev->new_chunk_sectors)) {
pr_warn("md/raid:%s: invalid chunk size %d\n",
mdname(mddev), mddev->new_chunk_sectors << 9);
return ERR_PTR(-EINVAL);
}
conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
if (conf == NULL)
goto abort;
md/raid5: set default stripe_size as 4096 In RAID5, if issued bio size is bigger than stripe_size, it will be split in the unit of stripe_size and process them one by one. Even for size less then stripe_size, RAID5 also request data from disk at least of stripe_size. Nowdays, stripe_size is equal to the value of PAGE_SIZE. Since filesystem usually issue bio in the unit of 4KB, there is no problem for PAGE_SIZE as 4KB. But, for 64KB PAGE_SIZE, bio from filesystem requests 4KB data while RAID5 issue IO at least stripe_size (64KB) each time. That will waste resource of disk bandwidth and compute xor. To avoding the waste, we want to make stripe_size configurable. This patch just set default stripe_size as 4096. User can also set the value bigger than 4KB for some special requirements, such as we know the issued io size is more than 4KB. To evaluate the new feature, we create raid5 device '/dev/md5' with 4 SSD disk and test it on arm64 machine with 64KB PAGE_SIZE. 1) We format /dev/md5 with mkfs.ext4 and mount ext4 with default configure on /mnt directory. Then, trying to test it by dbench with command: dbench -D /mnt -t 1000 10. Result show as: 'stripe_size = 64KB' Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 9805011 0.021 64.728 Close 7202525 0.001 0.120 Rename 415213 0.051 44.681 Unlink 1980066 0.079 93.147 Deltree 240 1.793 6.516 Mkdir 120 0.004 0.007 Qpathinfo 8887512 0.007 37.114 Qfileinfo 1557262 0.001 0.030 Qfsinfo 1629582 0.012 0.152 Sfileinfo 798756 0.040 57.641 Find 3436004 0.019 57.782 WriteX 4887239 0.021 57.638 ReadX 15370483 0.005 37.818 LockX 31934 0.003 0.022 UnlockX 31933 0.001 0.021 Flush 687205 13.302 530.088 Throughput 307.799 MB/sec 10 clients 10 procs max_latency=530.091 ms ------------------------------------------------------- 'stripe_size = 4KB' Operation Count AvgLat MaxLat ---------------------------------------- NTCreateX 11999166 0.021 36.380 Close 8814128 0.001 0.122 Rename 508113 0.051 29.169 Unlink 2423242 0.070 38.141 Deltree 300 1.885 7.155 Mkdir 150 0.004 0.006 Qpathinfo 10875921 0.007 35.485 Qfileinfo 1905837 0.001 0.032 Qfsinfo 1994304 0.012 0.125 Sfileinfo 977450 0.029 26.489 Find 4204952 0.019 9.361 WriteX 5981890 0.019 27.804 ReadX 18809742 0.004 33.491 LockX 39074 0.003 0.025 UnlockX 39074 0.001 0.014 Flush 841022 10.712 458.848 Throughput 376.777 MB/sec 10 clients 10 procs max_latency=458.852 ms ------------------------------------------------------- It show that setting stripe_size as 4KB has higher thoughput, i.e. (376.777 vs 307.799) and has smaller latency than that setting as 64KB. 2) We try to evaluate IO throughput for /dev/md5 by fio with config: [4KB randwrite] direct=1 numjob=2 iodepth=64 ioengine=libaio filename=/dev/md5 bs=4KB rw=randwrite [64KB write] direct=1 numjob=2 iodepth=64 ioengine=libaio filename=/dev/md5 bs=1MB rw=write The result as follow: + + | stripe_size(64KB) | stripe_size(4KB) +----------------------------------------------------+ 4KB randwrite | 15MB/s | 100MB/s +----------------------------------------------------+ 1MB write | 1000MB/s | 700MB/s The result show that when size of io is bigger than 4KB (64KB), 64KB stripe_size has much higher IOPS. But for 4KB randwrite, that means, size of io issued to device are smaller, 4KB stripe_size have better performance. Normally, default value (4096) can get relatively good performance. But if each issued io is bigger than 4096, setting value more than 4096 may get better performance. Here, we just set default stripe_size as 4096, and we will try to support setting different stripe_size by sysfs interface in the following patch. Signed-off-by: Yufen Yu <yuyufen@huawei.com> Signed-off-by: Song Liu <songliubraving@fb.com>
2020-07-18 09:29:08 +00:00
#if PAGE_SIZE != DEFAULT_STRIPE_SIZE
conf->stripe_size = DEFAULT_STRIPE_SIZE;
conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
#endif
INIT_LIST_HEAD(&conf->free_list);
INIT_LIST_HEAD(&conf->pending_list);
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
conf->pending_data = kcalloc(PENDING_IO_MAX,
sizeof(struct r5pending_data),
GFP_KERNEL);
if (!conf->pending_data)
goto abort;
for (i = 0; i < PENDING_IO_MAX; i++)
list_add(&conf->pending_data[i].sibling, &conf->free_list);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
/* Don't enable multi-threading by default*/
if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
conf->group_cnt = group_cnt;
conf->worker_cnt_per_group = 0;
md/raid5: Use conf->device_lock protect changing of multi-thread resources. When we change group_thread_cnt from sysfs entry, it can OOPS. The kernel messages are: [ 135.299021] BUG: unable to handle kernel NULL pointer dereference at (null) [ 135.299073] IP: [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299107] PGD 0 [ 135.299122] Oops: 0000 [#1] SMP [ 135.299144] Modules linked in: netconsole e1000e ptp pps_core [ 135.299188] CPU: 3 PID: 2225 Comm: md0_raid5 Not tainted 3.12.0+ #24 [ 135.299214] Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./To be filled by O.E.M., BIOS 080015 11/09/2011 [ 135.299255] task: ffff8800b9638f80 ti: ffff8800b77a4000 task.ti: ffff8800b77a4000 [ 135.299283] RIP: 0010:[<ffffffff815188ab>] [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.299323] RSP: 0018:ffff8800b77a5c48 EFLAGS: 00010002 [ 135.299344] RAX: ffff880037bb5c70 RBX: 0000000000000000 RCX: 0000000000000008 [ 135.299371] RDX: ffff880037bb5cb8 RSI: 0000000000000001 RDI: ffff880037bb5c00 [ 135.299398] RBP: ffff8800b77a5d08 R08: 0000000000000001 R09: 0000000000000000 [ 135.299425] R10: ffff8800b77a5c98 R11: 00000000ffffffff R12: ffff880037bb5c00 [ 135.299452] R13: 0000000000000000 R14: 0000000000000000 R15: ffff880037bb5c70 [ 135.299479] FS: 0000000000000000(0000) GS:ffff88013fd80000(0000) knlGS:0000000000000000 [ 135.299510] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b [ 135.299532] CR2: 0000000000000000 CR3: 0000000001c0b000 CR4: 00000000000407e0 [ 135.299559] Stack: [ 135.299570] ffff8800b77a5c88 ffffffff8107383e ffff8800b77a5c88 ffff880037a64300 [ 135.299611] 000000000000ec08 ffff880037bb5cb8 ffff8800b77a5c98 ffffffffffffffd8 [ 135.299654] 000000000000ec08 ffff880037bb5c60 ffff8800b77a5c98 ffff8800b77a5c98 [ 135.299696] Call Trace: [ 135.299711] [<ffffffff8107383e>] ? __wake_up+0x4e/0x70 [ 135.299733] [<ffffffff81518f88>] raid5d+0x4c8/0x680 [ 135.299756] [<ffffffff817174ed>] ? schedule_timeout+0x15d/0x1f0 [ 135.299781] [<ffffffff81524c9f>] md_thread+0x11f/0x170 [ 135.299804] [<ffffffff81069cd0>] ? wake_up_bit+0x40/0x40 [ 135.299826] [<ffffffff81524b80>] ? md_rdev_init+0x110/0x110 [ 135.299850] [<ffffffff81069656>] kthread+0xc6/0xd0 [ 135.299871] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299899] [<ffffffff81722ffc>] ret_from_fork+0x7c/0xb0 [ 135.299923] [<ffffffff81069590>] ? kthread_freezable_should_stop+0x70/0x70 [ 135.299951] Code: ff ff ff 0f 84 d7 fe ff ff e9 5c fe ff ff 66 90 41 8b b4 24 d8 01 00 00 45 31 ed 85 f6 0f 8e 7b fd ff ff 49 8b 9c 24 d0 01 00 00 <48> 3b 1b 49 89 dd 0f 85 67 fd ff ff 48 8d 43 28 31 d2 eb 17 90 [ 135.300005] RIP [<ffffffff815188ab>] handle_active_stripes+0x32b/0x440 [ 135.300005] RSP <ffff8800b77a5c48> [ 135.300005] CR2: 0000000000000000 [ 135.300005] ---[ end trace 504854e5bb7562ed ]--- [ 135.300005] Kernel panic - not syncing: Fatal exception This is because raid5d() can be running when the multi-thread resources are changed via system. We see need to provide locking. mddev->device_lock is suitable, but we cannot simple call alloc_thread_groups under this lock as we cannot allocate memory while holding a spinlock. So change alloc_thread_groups() to allocate and return the data structures, then raid5_store_group_thread_cnt() can take the lock while updating the pointers to the data structures. This fixes a bug introduced in 3.12 and so is suitable for the 3.12.x stable series. Fixes: b721420e8719131896b009b11edbbd27 Cc: stable@vger.kernel.org (3.12) Signed-off-by: Jianpeng Ma <majianpeng@gmail.com> Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Shaohua Li <shli@kernel.org>
2013-11-14 04:16:20 +00:00
conf->worker_groups = new_group;
} else
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
goto abort;
spin_lock_init(&conf->device_lock);
seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
mutex_init(&conf->cache_size_mutex);
init_waitqueue_head(&conf->wait_for_quiescent);
init_waitqueue_head(&conf->wait_for_stripe);
init_waitqueue_head(&conf->wait_for_overlap);
INIT_LIST_HEAD(&conf->handle_list);
INIT_LIST_HEAD(&conf->loprio_list);
INIT_LIST_HEAD(&conf->hold_list);
INIT_LIST_HEAD(&conf->delayed_list);
INIT_LIST_HEAD(&conf->bitmap_list);
init_llist_head(&conf->released_stripes);
atomic_set(&conf->active_stripes, 0);
atomic_set(&conf->preread_active_stripes, 0);
atomic_set(&conf->active_aligned_reads, 0);
spin_lock_init(&conf->pending_bios_lock);
conf->batch_bio_dispatch = true;
rdev_for_each(rdev, mddev) {
if (test_bit(Journal, &rdev->flags))
continue;
if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
conf->batch_bio_dispatch = false;
break;
}
}
conf->bypass_threshold = BYPASS_THRESHOLD;
conf->recovery_disabled = mddev->recovery_disabled - 1;
conf->raid_disks = mddev->raid_disks;
if (mddev->reshape_position == MaxSector)
conf->previous_raid_disks = mddev->raid_disks;
else
conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
max_disks = max(conf->raid_disks, conf->previous_raid_disks);
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:03:40 +00:00
conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
GFP_KERNEL);
if (!conf->disks)
goto abort;
for (i = 0; i < max_disks; i++) {
conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
if (!conf->disks[i].extra_page)
goto abort;
}
ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
if (ret)
goto abort;
conf->mddev = mddev;
if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
goto abort;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
/* We init hash_locks[0] separately to that it can be used
* as the reference lock in the spin_lock_nest_lock() call
* in lock_all_device_hash_locks_irq in order to convince
* lockdep that we know what we are doing.
*/
spin_lock_init(conf->hash_locks);
for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
spin_lock_init(conf->hash_locks + i);
for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
INIT_LIST_HEAD(conf->inactive_list + i);
for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
INIT_LIST_HEAD(conf->temp_inactive_list + i);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
atomic_set(&conf->r5c_cached_full_stripes, 0);
INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
atomic_set(&conf->r5c_cached_partial_stripes, 0);
INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
atomic_set(&conf->r5c_flushing_full_stripes, 0);
atomic_set(&conf->r5c_flushing_partial_stripes, 0);
md/r5cache: caching phase of r5cache As described in previous patch, write back cache operates in two phases: caching and writing-out. The caching phase works as: 1. write data to journal (r5c_handle_stripe_dirtying, r5c_cache_data) 2. call bio_endio (r5c_handle_data_cached, r5c_return_dev_pending_writes). Then the writing-out phase is as: 1. Mark the stripe as write-out (r5c_make_stripe_write_out) 2. Calcualte parity (reconstruct or RMW) 3. Write parity (and maybe some other data) to journal device 4. Write data and parity to RAID disks This patch implements caching phase. The cache is integrated with stripe cache of raid456. It leverages code of r5l_log to write data to journal device. Writing-out phase of the cache is implemented in the next patch. With r5cache, write operation does not wait for parity calculation and write out, so the write latency is lower (1 write to journal device vs. read and then write to raid disks). Also, r5cache will reduce RAID overhead (multipile IO due to read-modify-write of parity) and provide more opportunities of full stripe writes. This patch adds 2 flags to stripe_head.state: - STRIPE_R5C_PARTIAL_STRIPE, - STRIPE_R5C_FULL_STRIPE, Instead of inactive_list, stripes with cached data are tracked in r5conf->r5c_full_stripe_list and r5conf->r5c_partial_stripe_list. STRIPE_R5C_FULL_STRIPE and STRIPE_R5C_PARTIAL_STRIPE are flags for stripes in these lists. Note: stripes in r5c_full/partial_stripe_list are not considered as "active". For RMW, the code allocates an extra page for each data block being updated. This is stored in r5dev->orig_page and the old data is read into it. Then the prexor calculation subtracts ->orig_page from the parity block, and the reconstruct calculation adds the ->page data back into the parity block. r5cache naturally excludes SkipCopy. When the array has write back cache, async_copy_data() will not skip copy. There are some known limitations of the cache implementation: 1. Write cache only covers full page writes (R5_OVERWRITE). Writes of smaller granularity are write through. 2. Only one log io (sh->log_io) for each stripe at anytime. Later writes for the same stripe have to wait. This can be improved by moving log_io to r5dev. 3. With writeback cache, read path must enter state machine, which is a significant bottleneck for some workloads. 4. There is no per stripe checkpoint (with r5l_payload_flush) in the log, so recovery code has to replay more than necessary data (sometimes all the log from last_checkpoint). This reduces availability of the array. This patch includes a fix proposed by ZhengYuan Liu <liuzhengyuan@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:39 +00:00
conf->level = mddev->new_level;
conf->chunk_sectors = mddev->new_chunk_sectors;
if (raid5_alloc_percpu(conf) != 0)
goto abort;
pr_debug("raid456: run(%s) called.\n", mdname(mddev));
rdev_for_each(rdev, mddev) {
raid_disk = rdev->raid_disk;
if (raid_disk >= max_disks
|| raid_disk < 0 || test_bit(Journal, &rdev->flags))
continue;
disk = conf->disks + raid_disk;
if (test_bit(Replacement, &rdev->flags)) {
if (disk->replacement)
goto abort;
disk->replacement = rdev;
} else {
if (disk->rdev)
goto abort;
disk->rdev = rdev;
}
if (test_bit(In_sync, &rdev->flags)) {
char b[BDEVNAME_SIZE];
pr_info("md/raid:%s: device %s operational as raid disk %d\n",
mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
} else if (rdev->saved_raid_disk != raid_disk)
/* Cannot rely on bitmap to complete recovery */
conf->fullsync = 1;
}
conf->level = mddev->new_level;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (conf->level == 6) {
conf->max_degraded = 2;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
if (raid6_call.xor_syndrome)
conf->rmw_level = PARITY_ENABLE_RMW;
else
conf->rmw_level = PARITY_DISABLE_RMW;
} else {
conf->max_degraded = 1;
md/raid5: activate raid6 rmw feature Glue it altogehter. The raid6 rmw path should work the same as the already existing raid5 logic. So emulate the prexor handling/flags and split functions as needed. 1) Enable xor_syndrome() in the async layer. 2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome at the start of a rmw run as we did it before for the single parity. 3) Take care of rmw run in ops_run_reconstruct6(). Again process only the changed pages to get syndrome back into sync. 4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw run. The lower layers will calculate start & end pages from that and call the xor_syndrome() correspondingly. 5) Adapt the several places where we ignored Q handling up to now. Performance numbers for a single E5630 system with a mix of 10 7200k desktop/server disks. 300 seconds random write with 8 threads onto a 3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4) bsize rmw_level=1 rmw_level=0 rmw_level=1 rmw_level=0 skip_copy=1 skip_copy=1 skip_copy=0 skip_copy=0 4K 115 KB/s 141 KB/s 165 KB/s 140 KB/s 8K 225 KB/s 275 KB/s 324 KB/s 274 KB/s 16K 434 KB/s 536 KB/s 640 KB/s 534 KB/s 32K 751 KB/s 1,051 KB/s 1,234 KB/s 1,045 KB/s 64K 1,339 KB/s 1,958 KB/s 2,282 KB/s 1,962 KB/s 128K 2,673 KB/s 3,862 KB/s 4,113 KB/s 3,898 KB/s 256K 7,685 KB/s 7,539 KB/s 7,557 KB/s 7,638 KB/s 512K 19,556 KB/s 19,558 KB/s 19,652 KB/s 19,688 Kb/s Signed-off-by: Markus Stockhausen <stockhausen@collogia.de> Signed-off-by: NeilBrown <neilb@suse.de>
2014-12-15 01:57:05 +00:00
conf->rmw_level = PARITY_ENABLE_RMW;
}
conf->algorithm = mddev->new_layout;
conf->reshape_progress = mddev->reshape_position;
if (conf->reshape_progress != MaxSector) {
conf->prev_chunk_sectors = mddev->chunk_sectors;
conf->prev_algo = mddev->layout;
} else {
conf->prev_chunk_sectors = conf->chunk_sectors;
conf->prev_algo = conf->algorithm;
}
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
conf->min_nr_stripes = NR_STRIPES;
if (mddev->reshape_position != MaxSector) {
int stripes = max_t(int,
((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
conf->min_nr_stripes = max(NR_STRIPES, stripes);
if (conf->min_nr_stripes != NR_STRIPES)
pr_info("md/raid:%s: force stripe size %d for reshape\n",
mdname(mddev), conf->min_nr_stripes);
}
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
if (grow_stripes(conf, conf->min_nr_stripes)) {
pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
mdname(mddev), memory);
goto abort;
} else
pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
/*
* Losing a stripe head costs more than the time to refill it,
* it reduces the queue depth and so can hurt throughput.
* So set it rather large, scaled by number of devices.
*/
conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
conf->shrinker.scan_objects = raid5_cache_scan;
conf->shrinker.count_objects = raid5_cache_count;
conf->shrinker.batch = 128;
conf->shrinker.flags = 0;
if (register_shrinker(&conf->shrinker)) {
pr_warn("md/raid:%s: couldn't register shrinker.\n",
mdname(mddev));
goto abort;
}
sprintf(pers_name, "raid%d", mddev->new_level);
conf->thread = md_register_thread(raid5d, mddev, pers_name);
if (!conf->thread) {
pr_warn("md/raid:%s: couldn't allocate thread.\n",
mdname(mddev));
goto abort;
}
return conf;
abort:
if (conf) {
free_conf(conf);
return ERR_PTR(-EIO);
} else
return ERR_PTR(-ENOMEM);
}
static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
{
switch (algo) {
case ALGORITHM_PARITY_0:
if (raid_disk < max_degraded)
return 1;
break;
case ALGORITHM_PARITY_N:
if (raid_disk >= raid_disks - max_degraded)
return 1;
break;
case ALGORITHM_PARITY_0_6:
if (raid_disk == 0 ||
raid_disk == raid_disks - 1)
return 1;
break;
case ALGORITHM_LEFT_ASYMMETRIC_6:
case ALGORITHM_RIGHT_ASYMMETRIC_6:
case ALGORITHM_LEFT_SYMMETRIC_6:
case ALGORITHM_RIGHT_SYMMETRIC_6:
if (raid_disk == raid_disks - 1)
return 1;
}
return 0;
}
static int raid5_run(struct mddev *mddev)
{
struct r5conf *conf;
int working_disks = 0;
int dirty_parity_disks = 0;
struct md_rdev *rdev;
struct md_rdev *journal_dev = NULL;
sector_t reshape_offset = 0;
int i;
long long min_offset_diff = 0;
int first = 1;
if (mddev_init_writes_pending(mddev) < 0)
return -ENOMEM;
if (mddev->recovery_cp != MaxSector)
pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
mdname(mddev));
rdev_for_each(rdev, mddev) {
long long diff;
if (test_bit(Journal, &rdev->flags)) {
journal_dev = rdev;
continue;
}
if (rdev->raid_disk < 0)
continue;
diff = (rdev->new_data_offset - rdev->data_offset);
if (first) {
min_offset_diff = diff;
first = 0;
} else if (mddev->reshape_backwards &&
diff < min_offset_diff)
min_offset_diff = diff;
else if (!mddev->reshape_backwards &&
diff > min_offset_diff)
min_offset_diff = diff;
}
if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
(mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
mdname(mddev));
return -EINVAL;
}
if (mddev->reshape_position != MaxSector) {
/* Check that we can continue the reshape.
* Difficulties arise if the stripe we would write to
* next is at or after the stripe we would read from next.
* For a reshape that changes the number of devices, this
* is only possible for a very short time, and mdadm makes
* sure that time appears to have past before assembling
* the array. So we fail if that time hasn't passed.
* For a reshape that keeps the number of devices the same
* mdadm must be monitoring the reshape can keeping the
* critical areas read-only and backed up. It will start
* the array in read-only mode, so we check for that.
*/
sector_t here_new, here_old;
int old_disks;
int max_degraded = (mddev->level == 6 ? 2 : 1);
int chunk_sectors;
int new_data_disks;
if (journal_dev) {
pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
mdname(mddev));
return -EINVAL;
}
if (mddev->new_level != mddev->level) {
pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
mdname(mddev));
return -EINVAL;
}
old_disks = mddev->raid_disks - mddev->delta_disks;
/* reshape_position must be on a new-stripe boundary, and one
* further up in new geometry must map after here in old
* geometry.
* If the chunk sizes are different, then as we perform reshape
* in units of the largest of the two, reshape_position needs
* be a multiple of the largest chunk size times new data disks.
*/
here_new = mddev->reshape_position;
chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
new_data_disks = mddev->raid_disks - max_degraded;
if (sector_div(here_new, chunk_sectors * new_data_disks)) {
pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
mdname(mddev));
return -EINVAL;
}
reshape_offset = here_new * chunk_sectors;
/* here_new is the stripe we will write to */
here_old = mddev->reshape_position;
sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
/* here_old is the first stripe that we might need to read
* from */
if (mddev->delta_disks == 0) {
/* We cannot be sure it is safe to start an in-place
* reshape. It is only safe if user-space is monitoring
* and taking constant backups.
* mdadm always starts a situation like this in
* readonly mode so it can take control before
* allowing any writes. So just check for that.
*/
if (abs(min_offset_diff) >= mddev->chunk_sectors &&
abs(min_offset_diff) >= mddev->new_chunk_sectors)
/* not really in-place - so OK */;
else if (mddev->ro == 0) {
pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
mdname(mddev));
return -EINVAL;
}
} else if (mddev->reshape_backwards
? (here_new * chunk_sectors + min_offset_diff <=
here_old * chunk_sectors)
: (here_new * chunk_sectors >=
here_old * chunk_sectors + (-min_offset_diff))) {
/* Reading from the same stripe as writing to - bad */
pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
mdname(mddev));
return -EINVAL;
}
pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
/* OK, we should be able to continue; */
} else {
BUG_ON(mddev->level != mddev->new_level);
BUG_ON(mddev->layout != mddev->new_layout);
BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
BUG_ON(mddev->delta_disks != 0);
}
if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
test_bit(MD_HAS_PPL, &mddev->flags)) {
pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
mdname(mddev));
clear_bit(MD_HAS_PPL, &mddev->flags);
clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
}
if (mddev->private == NULL)
conf = setup_conf(mddev);
else
conf = mddev->private;
if (IS_ERR(conf))
return PTR_ERR(conf);
if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
if (!journal_dev) {
pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
mdname(mddev));
mddev->ro = 1;
set_disk_ro(mddev->gendisk, 1);
} else if (mddev->recovery_cp == MaxSector)
set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
}
conf->min_offset_diff = min_offset_diff;
mddev->thread = conf->thread;
conf->thread = NULL;
mddev->private = conf;
for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
i++) {
rdev = conf->disks[i].rdev;
if (!rdev && conf->disks[i].replacement) {
/* The replacement is all we have yet */
rdev = conf->disks[i].replacement;
conf->disks[i].replacement = NULL;
clear_bit(Replacement, &rdev->flags);
conf->disks[i].rdev = rdev;
}
if (!rdev)
continue;
if (conf->disks[i].replacement &&
conf->reshape_progress != MaxSector) {
/* replacements and reshape simply do not mix. */
pr_warn("md: cannot handle concurrent replacement and reshape.\n");
goto abort;
}
if (test_bit(In_sync, &rdev->flags)) {
working_disks++;
continue;
}
/* This disc is not fully in-sync. However if it
* just stored parity (beyond the recovery_offset),
* when we don't need to be concerned about the
* array being dirty.
* When reshape goes 'backwards', we never have
* partially completed devices, so we only need
* to worry about reshape going forwards.
*/
/* Hack because v0.91 doesn't store recovery_offset properly. */
if (mddev->major_version == 0 &&
mddev->minor_version > 90)
rdev->recovery_offset = reshape_offset;
if (rdev->recovery_offset < reshape_offset) {
/* We need to check old and new layout */
if (!only_parity(rdev->raid_disk,
conf->algorithm,
conf->raid_disks,
conf->max_degraded))
continue;
}
if (!only_parity(rdev->raid_disk,
conf->prev_algo,
conf->previous_raid_disks,
conf->max_degraded))
continue;
dirty_parity_disks++;
}
/*
* 0 for a fully functional array, 1 or 2 for a degraded array.
*/
mddev->degraded = raid5_calc_degraded(conf);
if (has_failed(conf)) {
pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
mdname(mddev), mddev->degraded, conf->raid_disks);
goto abort;
}
/* device size must be a multiple of chunk size */
mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
mddev->resync_max_sectors = mddev->dev_sectors;
if (mddev->degraded > dirty_parity_disks &&
mddev->recovery_cp != MaxSector) {
if (test_bit(MD_HAS_PPL, &mddev->flags))
pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
mdname(mddev));
else if (mddev->ok_start_degraded)
pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
mdname(mddev));
else {
pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
mdname(mddev));
goto abort;
}
}
pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
mdname(mddev), conf->level,
mddev->raid_disks-mddev->degraded, mddev->raid_disks,
mddev->new_layout);
print_raid5_conf(conf);
if (conf->reshape_progress != MaxSector) {
conf->reshape_safe = conf->reshape_progress;
atomic_set(&conf->reshape_stripes, 0);
clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
mddev->sync_thread = md_register_thread(md_do_sync, mddev,
"reshape");
if (!mddev->sync_thread)
goto abort;
}
/* Ok, everything is just fine now */
if (mddev->to_remove == &raid5_attrs_group)
mddev->to_remove = NULL;
else if (mddev->kobj.sd &&
sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
pr_warn("raid5: failed to create sysfs attributes for %s\n",
mdname(mddev));
md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
if (mddev->queue) {
int chunk_size;
/* read-ahead size must cover two whole stripes, which
* is 2 * (datadisks) * chunksize where 'n' is the
* number of raid devices
*/
int data_disks = conf->previous_raid_disks - conf->max_degraded;
int stripe = data_disks *
((mddev->chunk_sectors << 9) / PAGE_SIZE);
if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
chunk_size = mddev->chunk_sectors << 9;
blk_queue_io_min(mddev->queue, chunk_size);
blk_queue_io_opt(mddev->queue, chunk_size *
(conf->raid_disks - conf->max_degraded));
mddev->queue->limits.raid_partial_stripes_expensive = 1;
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
/*
* We can only discard a whole stripe. It doesn't make sense to
* discard data disk but write parity disk
*/
stripe = stripe * PAGE_SIZE;
/* Round up to power of 2, as discard handling
* currently assumes that */
while ((stripe-1) & stripe)
stripe = (stripe | (stripe-1)) + 1;
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
mddev->queue->limits.discard_alignment = stripe;
mddev->queue->limits.discard_granularity = stripe;
blk_queue_max_write_same_sectors(mddev->queue, 0);
blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
rdev_for_each(rdev, mddev) {
disk_stack_limits(mddev->gendisk, rdev->bdev,
rdev->data_offset << 9);
disk_stack_limits(mddev->gendisk, rdev->bdev,
rdev->new_data_offset << 9);
}
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
/*
* zeroing is required, otherwise data
* could be lost. Consider a scenario: discard a stripe
* (the stripe could be inconsistent if
* discard_zeroes_data is 0); write one disk of the
* stripe (the stripe could be inconsistent again
* depending on which disks are used to calculate
* parity); the disk is broken; The stripe data of this
* disk is lost.
*
* We only allow DISCARD if the sysadmin has confirmed that
* only safe devices are in use by setting a module parameter.
* A better idea might be to turn DISCARD into WRITE_ZEROES
* requests, as that is required to be safe.
*/
if (devices_handle_discard_safely &&
mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
mddev->queue->limits.discard_granularity >= stripe)
blk_queue_flag_set(QUEUE_FLAG_DISCARD,
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
mddev->queue);
else
blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
MD: raid5 trim support Discard for raid4/5/6 has limitation. If discard request size is small, we do discard for one disk, but we need calculate parity and write parity disk. To correctly calculate parity, zero_after_discard must be guaranteed. Even it's true, we need do discard for one disk but write another disks, which makes the parity disks wear out fast. This doesn't make sense. So an efficient discard for raid4/5/6 should discard all data disks and parity disks, which requires the write pattern to be (A, A+chunk_size, A+chunk_size*2...). If A's size is smaller than chunk_size, such pattern is almost impossible in practice. So in this patch, I only handle the case that A's size equals to chunk_size. That is discard request should be aligned to stripe size and its size is multiple of stripe size. Since we can only handle request with specific alignment and size (or part of the request fitting stripes), we can't guarantee zero_after_discard even zero_after_discard is true in low level drives. The block layer doesn't send down correctly aligned requests even correct discard alignment is set, so I must filter out. For raid4/5/6 parity calculation, if data is 0, parity is 0. So if zero_after_discard is true for all disks, data is consistent after discard. Otherwise, data might be lost. Let's consider a scenario: discard a stripe, write data to one disk and write parity disk. The stripe could be still inconsistent till then depending on using data from other data disks or parity disks to calculate new parity. If the disk is broken, we can't restore it. So in this patch, we only enable discard support if all disks have zero_after_discard. If discard fails in one disk, we face the similar inconsistent issue above. The patch will make discard follow the same path as normal write request. If discard fails, a resync will be scheduled to make the data consistent. This isn't good to have extra writes, but data consistency is important. If a subsequent read/write request hits raid5 cache of a discarded stripe, the discarded dev page should have zero filled, so the data is consistent. This patch will always zero dev page for discarded request stripe. This isn't optimal because discard request doesn't need such payload. Next patch will avoid it. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2012-10-11 02:49:05 +00:00
mddev->queue);
blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
}
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
goto abort;
return 0;
abort:
md_unregister_thread(&mddev->thread);
print_raid5_conf(conf);
free_conf(conf);
mddev->private = NULL;
pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
return -EIO;
}
static void raid5_free(struct mddev *mddev, void *priv)
{
struct r5conf *conf = priv;
free_conf(conf);
mddev->to_remove = &raid5_attrs_group;
}
static void raid5_status(struct seq_file *seq, struct mddev *mddev)
{
struct r5conf *conf = mddev->private;
int i;
seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
conf->chunk_sectors / 2, mddev->layout);
seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
rcu_read_lock();
for (i = 0; i < conf->raid_disks; i++) {
struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
}
rcu_read_unlock();
seq_printf (seq, "]");
}
static void print_raid5_conf (struct r5conf *conf)
{
int i;
struct disk_info *tmp;
pr_debug("RAID conf printout:\n");
if (!conf) {
pr_debug("(conf==NULL)\n");
return;
}
pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
conf->raid_disks,
conf->raid_disks - conf->mddev->degraded);
for (i = 0; i < conf->raid_disks; i++) {
char b[BDEVNAME_SIZE];
tmp = conf->disks + i;
if (tmp->rdev)
pr_debug(" disk %d, o:%d, dev:%s\n",
i, !test_bit(Faulty, &tmp->rdev->flags),
bdevname(tmp->rdev->bdev, b));
}
}
static int raid5_spare_active(struct mddev *mddev)
{
int i;
struct r5conf *conf = mddev->private;
struct disk_info *tmp;
int count = 0;
unsigned long flags;
for (i = 0; i < conf->raid_disks; i++) {
tmp = conf->disks + i;
if (tmp->replacement
&& tmp->replacement->recovery_offset == MaxSector
&& !test_bit(Faulty, &tmp->replacement->flags)
&& !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
/* Replacement has just become active. */
if (!tmp->rdev
|| !test_and_clear_bit(In_sync, &tmp->rdev->flags))
count++;
if (tmp->rdev) {
/* Replaced device not technically faulty,
* but we need to be sure it gets removed
* and never re-added.
*/
set_bit(Faulty, &tmp->rdev->flags);
sysfs_notify_dirent_safe(
tmp->rdev->sysfs_state);
}
sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
} else if (tmp->rdev
md: Don't update ->recovery_offset when reshaping an array to fewer devices. When an array is reshaped to have fewer devices, the reshape proceeds from the end of the devices to the beginning. If a device happens to be non-In_sync (which is possible but rare) we would normally update the ->recovery_offset as the reshape progresses. However that would be wrong as the recover_offset records that the early part of the device is in_sync, while in fact it would only be the later part that is in_sync, and in any case the offset number would be measured from the wrong end of the device. Relatedly, if after a reshape a spare is discovered to not be recoverred all the way to the end, not allow spare_active to incorporate it in the array. This becomes relevant in the following sample scenario: A 4 drive RAID5 is converted to a 6 drive RAID6 in a combined operation. The RAID5->RAID6 conversion will cause a 5 drive to be included as a spare, then the 5drive -> 6drive reshape will effectively rebuild that spare as it progresses. The 6th drive is treated as in_sync the whole time as there is never any case that we might consider reading from it, but must not because there is no valid data. If we interrupt this reshape part-way through and reverse it to return to a 5-drive RAID6 (or event a 4-drive RAID5), we don't want to update the recovery_offset - as that would be wrong - and we don't want to include that spare as active in the 5-drive RAID6 when the reversed reshape completed and it will be mostly out-of-sync still. Signed-off-by: NeilBrown <neilb@suse.de>
2010-06-16 07:01:25 +00:00
&& tmp->rdev->recovery_offset == MaxSector
&& !test_bit(Faulty, &tmp->rdev->flags)
&& !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
count++;
sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
}
}
spin_lock_irqsave(&conf->device_lock, flags);
mddev->degraded = raid5_calc_degraded(conf);
spin_unlock_irqrestore(&conf->device_lock, flags);
print_raid5_conf(conf);
return count;
}
static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
{
struct r5conf *conf = mddev->private;
int err = 0;
int number = rdev->raid_disk;
struct md_rdev **rdevp;
struct disk_info *p = conf->disks + number;
print_raid5_conf(conf);
if (test_bit(Journal, &rdev->flags) && conf->log) {
/*
* we can't wait pending write here, as this is called in
* raid5d, wait will deadlock.
* neilb: there is no locking about new writes here,
* so this cannot be safe.
*/
if (atomic_read(&conf->active_stripes) ||
atomic_read(&conf->r5c_cached_full_stripes) ||
atomic_read(&conf->r5c_cached_partial_stripes)) {
return -EBUSY;
}
log_exit(conf);
return 0;
}
if (rdev == p->rdev)
rdevp = &p->rdev;
else if (rdev == p->replacement)
rdevp = &p->replacement;
else
return 0;
if (number >= conf->raid_disks &&
conf->reshape_progress == MaxSector)
clear_bit(In_sync, &rdev->flags);
if (test_bit(In_sync, &rdev->flags) ||
atomic_read(&rdev->nr_pending)) {
err = -EBUSY;
goto abort;
}
/* Only remove non-faulty devices if recovery
* isn't possible.
*/
if (!test_bit(Faulty, &rdev->flags) &&
mddev->recovery_disabled != conf->recovery_disabled &&
!has_failed(conf) &&
(!p->replacement || p->replacement == rdev) &&
number < conf->raid_disks) {
err = -EBUSY;
goto abort;
}
*rdevp = NULL;
if (!test_bit(RemoveSynchronized, &rdev->flags)) {
synchronize_rcu();
if (atomic_read(&rdev->nr_pending)) {
/* lost the race, try later */
err = -EBUSY;
*rdevp = rdev;
}
}
if (!err) {
err = log_modify(conf, rdev, false);
if (err)
goto abort;
}
if (p->replacement) {
/* We must have just cleared 'rdev' */
p->rdev = p->replacement;
clear_bit(Replacement, &p->replacement->flags);
smp_mb(); /* Make sure other CPUs may see both as identical
* but will never see neither - if they are careful
*/
p->replacement = NULL;
if (!err)
err = log_modify(conf, p->rdev, true);
}
clear_bit(WantReplacement, &rdev->flags);
abort:
print_raid5_conf(conf);
return err;
}
static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
{
struct r5conf *conf = mddev->private;
int ret, err = -EEXIST;
int disk;
struct disk_info *p;
int first = 0;
int last = conf->raid_disks - 1;
if (test_bit(Journal, &rdev->flags)) {
if (conf->log)
return -EBUSY;
rdev->raid_disk = 0;
/*
* The array is in readonly mode if journal is missing, so no
* write requests running. We should be safe
*/
ret = log_init(conf, rdev, false);
if (ret)
return ret;
ret = r5l_start(conf->log);
if (ret)
return ret;
return 0;
}
if (mddev->recovery_disabled == conf->recovery_disabled)
return -EBUSY;
if (rdev->saved_raid_disk < 0 && has_failed(conf))
/* no point adding a device */
return -EINVAL;
if (rdev->raid_disk >= 0)
first = last = rdev->raid_disk;
/*
* find the disk ... but prefer rdev->saved_raid_disk
* if possible.
*/
if (rdev->saved_raid_disk >= 0 &&
rdev->saved_raid_disk >= first &&
conf->disks[rdev->saved_raid_disk].rdev == NULL)
first = rdev->saved_raid_disk;
for (disk = first; disk <= last; disk++) {
p = conf->disks + disk;
if (p->rdev == NULL) {
clear_bit(In_sync, &rdev->flags);
rdev->raid_disk = disk;
if (rdev->saved_raid_disk != disk)
conf->fullsync = 1;
rcu_assign_pointer(p->rdev, rdev);
err = log_modify(conf, rdev, true);
goto out;
}
}
for (disk = first; disk <= last; disk++) {
p = conf->disks + disk;
if (test_bit(WantReplacement, &p->rdev->flags) &&
p->replacement == NULL) {
clear_bit(In_sync, &rdev->flags);
set_bit(Replacement, &rdev->flags);
rdev->raid_disk = disk;
err = 0;
conf->fullsync = 1;
rcu_assign_pointer(p->replacement, rdev);
break;
}
}
out:
print_raid5_conf(conf);
return err;
}
static int raid5_resize(struct mddev *mddev, sector_t sectors)
{
/* no resync is happening, and there is enough space
* on all devices, so we can resize.
* We need to make sure resync covers any new space.
* If the array is shrinking we should possibly wait until
* any io in the removed space completes, but it hardly seems
* worth it.
*/
sector_t newsize;
struct r5conf *conf = mddev->private;
if (raid5_has_log(conf) || raid5_has_ppl(conf))
return -EINVAL;
sectors &= ~((sector_t)conf->chunk_sectors - 1);
newsize = raid5_size(mddev, sectors, mddev->raid_disks);
if (mddev->external_size &&
mddev->array_sectors > newsize)
return -EINVAL;
if (mddev->bitmap) {
int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
if (ret)
return ret;
}
md_set_array_sectors(mddev, newsize);
if (sectors > mddev->dev_sectors &&
mddev->recovery_cp > mddev->dev_sectors) {
mddev->recovery_cp = mddev->dev_sectors;
set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
}
mddev->dev_sectors = sectors;
mddev->resync_max_sectors = sectors;
return 0;
}
static int check_stripe_cache(struct mddev *mddev)
{
/* Can only proceed if there are plenty of stripe_heads.
* We need a minimum of one full stripe,, and for sensible progress
* it is best to have about 4 times that.
* If we require 4 times, then the default 256 4K stripe_heads will
* allow for chunk sizes up to 256K, which is probably OK.
* If the chunk size is greater, user-space should request more
* stripe_heads first.
*/
struct r5conf *conf = mddev->private;
if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
> conf->min_nr_stripes ||
((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
md/raid5: allow the stripe_cache to grow and shrink. The default setting of 256 stripe_heads is probably much too small for many configurations. So it is best to make it auto-configure. Shrinking the cache under memory pressure is easy. The only interesting part here is that we put a fairly high cost ('seeks') on shrinking the cache as the cost is greater than just having to read more data, it reduces parallelism. Growing the cache on demand needs to be done carefully. If we allow fast growth, that can upset memory balance as lots of dirty memory can quickly turn into lots of memory queued in the stripe_cache. It is important for the raid5 block device to appear congested to allow write-throttling to work. So we only add stripes slowly. We set a flag when an allocation fails because all stripes are in use, allocate at a convenient time when that flag is set, and don't allow it to be set again until at least one stripe_head has been released for re-use. This means that a spurt of requests will only cause one stripe_head to be allocated, but a steady stream of requests will slowly increase the cache size - until memory pressure puts it back again. It could take hours to reach a steady state. The value written to, and displayed in, stripe_cache_size is used as a minimum. The cache can grow above this and shrink back down to it. The actual size is not directly visible, though it can be deduced to some extent by watching stripe_cache_active. Signed-off-by: NeilBrown <neilb@suse.de>
2015-02-26 01:47:56 +00:00
> conf->min_nr_stripes) {
pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
mdname(mddev),
((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
/ RAID5_STRIPE_SIZE(conf))*4);
return 0;
}
return 1;
}
static int check_reshape(struct mddev *mddev)
{
struct r5conf *conf = mddev->private;
if (raid5_has_log(conf) || raid5_has_ppl(conf))
return -EINVAL;
if (mddev->delta_disks == 0 &&
mddev->new_layout == mddev->layout &&
mddev->new_chunk_sectors == mddev->chunk_sectors)
return 0; /* nothing to do */
if (has_failed(conf))
return -EINVAL;
if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
/* We might be able to shrink, but the devices must
* be made bigger first.
* For raid6, 4 is the minimum size.
* Otherwise 2 is the minimum
*/
int min = 2;
if (mddev->level == 6)
min = 4;
if (mddev->raid_disks + mddev->delta_disks < min)
return -EINVAL;
}
if (!check_stripe_cache(mddev))
return -ENOSPC;
if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
mddev->delta_disks > 0)
if (resize_chunks(conf,
conf->previous_raid_disks
+ max(0, mddev->delta_disks),
max(mddev->new_chunk_sectors,
mddev->chunk_sectors)
) < 0)
return -ENOMEM;
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
return 0; /* never bother to shrink */
return resize_stripes(conf, (conf->previous_raid_disks
+ mddev->delta_disks));
}
static int raid5_start_reshape(struct mddev *mddev)
{
struct r5conf *conf = mddev->private;
struct md_rdev *rdev;
int spares = 0;
unsigned long flags;
if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
return -EBUSY;
if (!check_stripe_cache(mddev))
return -ENOSPC;
if (has_failed(conf))
return -EINVAL;
rdev_for_each(rdev, mddev) {
if (!test_bit(In_sync, &rdev->flags)
&& !test_bit(Faulty, &rdev->flags))
spares++;
}
if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
/* Not enough devices even to make a degraded array
* of that size
*/
return -EINVAL;
/* Refuse to reduce size of the array. Any reductions in
* array size must be through explicit setting of array_size
* attribute.
*/
if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
< mddev->array_sectors) {
pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
mdname(mddev));
return -EINVAL;
}
atomic_set(&conf->reshape_stripes, 0);
spin_lock_irq(&conf->device_lock);
write_seqcount_begin(&conf->gen_lock);
conf->previous_raid_disks = conf->raid_disks;
conf->raid_disks += mddev->delta_disks;
conf->prev_chunk_sectors = conf->chunk_sectors;
conf->chunk_sectors = mddev->new_chunk_sectors;
conf->prev_algo = conf->algorithm;
conf->algorithm = mddev->new_layout;
conf->generation++;
/* Code that selects data_offset needs to see the generation update
* if reshape_progress has been set - so a memory barrier needed.
*/
smp_mb();
if (mddev->reshape_backwards)
conf->reshape_progress = raid5_size(mddev, 0, 0);
else
conf->reshape_progress = 0;
conf->reshape_safe = conf->reshape_progress;
write_seqcount_end(&conf->gen_lock);
spin_unlock_irq(&conf->device_lock);
/* Now make sure any requests that proceeded on the assumption
* the reshape wasn't running - like Discard or Read - have
* completed.
*/
mddev_suspend(mddev);
mddev_resume(mddev);
/* Add some new drives, as many as will fit.
* We know there are enough to make the newly sized array work.
* Don't add devices if we are reducing the number of
* devices in the array. This is because it is not possible
* to correctly record the "partially reconstructed" state of
* such devices during the reshape and confusion could result.
*/
if (mddev->delta_disks >= 0) {
rdev_for_each(rdev, mddev)
if (rdev->raid_disk < 0 &&
!test_bit(Faulty, &rdev->flags)) {
if (raid5_add_disk(mddev, rdev) == 0) {
if (rdev->raid_disk
>= conf->previous_raid_disks)
set_bit(In_sync, &rdev->flags);
else
rdev->recovery_offset = 0;
/* Failure here is OK */
sysfs_link_rdev(mddev, rdev);
}
} else if (rdev->raid_disk >= conf->previous_raid_disks
&& !test_bit(Faulty, &rdev->flags)) {
/* This is a spare that was manually added */
set_bit(In_sync, &rdev->flags);
}
/* When a reshape changes the number of devices,
* ->degraded is measured against the larger of the
* pre and post number of devices.
*/
spin_lock_irqsave(&conf->device_lock, flags);
mddev->degraded = raid5_calc_degraded(conf);
spin_unlock_irqrestore(&conf->device_lock, flags);
}
mddev->raid_disks = conf->raid_disks;
mddev->reshape_position = conf->reshape_progress;
set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
mddev->sync_thread = md_register_thread(md_do_sync, mddev,
"reshape");
if (!mddev->sync_thread) {
mddev->recovery = 0;
spin_lock_irq(&conf->device_lock);
write_seqcount_begin(&conf->gen_lock);
mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
mddev->new_chunk_sectors =
conf->chunk_sectors = conf->prev_chunk_sectors;
mddev->new_layout = conf->algorithm = conf->prev_algo;
rdev_for_each(rdev, mddev)
rdev->new_data_offset = rdev->data_offset;
smp_wmb();
conf->generation --;
conf->reshape_progress = MaxSector;
mddev->reshape_position = MaxSector;
write_seqcount_end(&conf->gen_lock);
spin_unlock_irq(&conf->device_lock);
return -EAGAIN;
}
conf->reshape_checkpoint = jiffies;
md_wakeup_thread(mddev->sync_thread);
md_new_event(mddev);
return 0;
}
/* This is called from the reshape thread and should make any
* changes needed in 'conf'
*/
static void end_reshape(struct r5conf *conf)
{
if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
struct md_rdev *rdev;
spin_lock_irq(&conf->device_lock);
conf->previous_raid_disks = conf->raid_disks;
md_finish_reshape(conf->mddev);
smp_wmb();
conf->reshape_progress = MaxSector;
conf->mddev->reshape_position = MaxSector;
rdev_for_each(rdev, conf->mddev)
if (rdev->raid_disk >= 0 &&
!test_bit(Journal, &rdev->flags) &&
!test_bit(In_sync, &rdev->flags))
rdev->recovery_offset = MaxSector;
spin_unlock_irq(&conf->device_lock);
wake_up(&conf->wait_for_overlap);
/* read-ahead size must cover two whole stripes, which is
* 2 * (datadisks) * chunksize where 'n' is the number of raid devices
*/
if (conf->mddev->queue) {
int data_disks = conf->raid_disks - conf->max_degraded;
int stripe = data_disks * ((conf->chunk_sectors << 9)
/ PAGE_SIZE);
if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
}
}
}
/* This is called from the raid5d thread with mddev_lock held.
* It makes config changes to the device.
*/
static void raid5_finish_reshape(struct mddev *mddev)
{
struct r5conf *conf = mddev->private;
if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
md: fix a potential deadlock of raid5/raid10 reshape There is a potential deadlock if mount/umount happens when raid5_finish_reshape() tries to grow the size of emulated disk. How the deadlock happens? 1) The raid5 resync thread finished reshape (expanding array). 2) The mount or umount thread holds VFS sb->s_umount lock and tries to write through critical data into raid5 emulated block device. So it waits for raid5 kernel thread handling stripes in order to finish it I/Os. 3) In the routine of raid5 kernel thread, md_check_recovery() will be called first in order to reap the raid5 resync thread. That is, raid5_finish_reshape() will be called. In this function, it will try to update conf and call VFS revalidate_disk() to grow the raid5 emulated block device. It will try to acquire VFS sb->s_umount lock. The raid5 kernel thread cannot continue, so no one can handle mount/ umount I/Os (stripes). Once the write-through I/Os cannot be finished, mount/umount will not release sb->s_umount lock. The deadlock happens. The raid5 kernel thread is an emulated block device. It is responible to handle I/Os (stripes) from upper layers. The emulated block device should not request any I/Os on itself. That is, it should not call VFS layer functions. (If it did, it will try to acquire VFS locks to guarantee the I/Os sequence.) So we have the resync thread to send resync I/O requests and to wait for the results. For solving this potential deadlock, we can put the size growth of the emulated block device as the final step of reshape thread. 2017/12/29: Thanks to Guoqing Jiang <gqjiang@suse.com>, we confirmed that there is the same deadlock issue in raid10. It's reproducible and can be fixed by this patch. For raid10.c, we can remove the similar code to prevent deadlock as well since they has been called before. Reported-by: Alex Wu <alexwu@synology.com> Reviewed-by: Alex Wu <alexwu@synology.com> Reviewed-by: Chung-Chiang Cheng <cccheng@synology.com> Signed-off-by: BingJing Chang <bingjingc@synology.com> Signed-off-by: Shaohua Li <sh.li@alibaba-inc.com>
2018-02-22 05:34:46 +00:00
if (mddev->delta_disks <= 0) {
int d;
spin_lock_irq(&conf->device_lock);
mddev->degraded = raid5_calc_degraded(conf);
spin_unlock_irq(&conf->device_lock);
for (d = conf->raid_disks ;
d < conf->raid_disks - mddev->delta_disks;
d++) {
struct md_rdev *rdev = conf->disks[d].rdev;
if (rdev)
clear_bit(In_sync, &rdev->flags);
rdev = conf->disks[d].replacement;
if (rdev)
clear_bit(In_sync, &rdev->flags);
}
}
mddev->layout = conf->algorithm;
mddev->chunk_sectors = conf->chunk_sectors;
mddev->reshape_position = MaxSector;
mddev->delta_disks = 0;
mddev->reshape_backwards = 0;
}
}
static void raid5_quiesce(struct mddev *mddev, int quiesce)
{
struct r5conf *conf = mddev->private;
if (quiesce) {
/* stop all writes */
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
lock_all_device_hash_locks_irq(conf);
/* '2' tells resync/reshape to pause so that all
* active stripes can drain
*/
md/r5cache: write-out phase and reclaim support There are two limited resources, stripe cache and journal disk space. For better performance, we priotize reclaim of full stripe writes. To free up more journal space, we free earliest data on the journal. In current implementation, reclaim happens when: 1. Periodically (every R5C_RECLAIM_WAKEUP_INTERVAL, 30 seconds) reclaim if there is no reclaim in the past 5 seconds. 2. when there are R5C_FULL_STRIPE_FLUSH_BATCH (256) cached full stripes, or cached stripes is enough for a full stripe (chunk size / 4k) (r5c_check_cached_full_stripe) 3. when there is pressure on stripe cache (r5c_check_stripe_cache_usage) 4. when there is pressure on journal space (r5l_write_stripe, r5c_cache_data) r5c_do_reclaim() contains new logic of reclaim. For stripe cache: When stripe cache pressure is high (more than 3/4 stripes are cached, or there is empty inactive lists), flush all full stripe. If fewer than R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) full stripes are flushed, flush some paritial stripes. When stripe cache pressure is moderate (1/2 to 3/4 of stripes are cached), flush all full stripes. For log space: To avoid deadlock due to log space, we need to reserve enough space to flush cached data. The size of required log space depends on total number of cached stripes (stripe_in_journal_count). In current implementation, the writing-out phase automatically include pending data writes with parity writes (similar to write through case). Therefore, we need up to (conf->raid_disks + 1) pages for each cached stripe (1 page for meta data, raid_disks pages for all data and parity). r5c_log_required_to_flush_cache() calculates log space required to flush cache. In the following, we refer to the space calculated by r5c_log_required_to_flush_cache() as reclaim_required_space. Two flags are added to r5conf->cache_state: R5C_LOG_TIGHT and R5C_LOG_CRITICAL. R5C_LOG_TIGHT is set when free space on the log device is less than 3x of reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log device is less than 2x of reclaim_required_space. r5c_cache keeps all data in cache (not fully committed to RAID) in a list (stripe_in_journal_list). These stripes are in the order of their first appearance on the journal. So the log tail (last_checkpoint) should point to the journal_start of the first item in the list. When R5C_LOG_TIGHT is set, r5l_reclaim_thread starts flushing out stripes at the head of stripe_in_journal. When R5C_LOG_CRITICAL is set, the state machine only writes data that are already in the log device (in stripe_in_journal_list). This patch includes a fix to improve performance by Shaohua Li <shli@fb.com>. Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Shaohua Li <shli@fb.com>
2016-11-17 23:24:40 +00:00
r5c_flush_cache(conf, INT_MAX);
conf->quiesce = 2;
wait_event_cmd(conf->wait_for_quiescent,
[PATCH] md: allow reads that have bypassed the cache to be retried on failure If a bypass-the-cache read fails, we simply try again through the cache. If it fails again it will trigger normal recovery precedures. update 1: From: NeilBrown <neilb@suse.de> 1/ chunk_aligned_read and retry_aligned_read assume that data_disks == raid_disks - 1 which is not true for raid6. So when an aligned read request bypasses the cache, we can get the wrong data. 2/ The cloned bio is being used-after-free in raid5_align_endio (to test BIO_UPTODATE). 3/ We forgot to add rdev->data_offset when submitting a bio for aligned-read 4/ clone_bio calls blk_recount_segments and then we change bi_bdev, so we need to invalidate the segment counts. 5/ We don't de-reference the rdev when the read completes. This means we need to record the rdev to so it is still available in the end_io routine. Fortunately bi_next in the original bio is unused at this point so we can stuff it in there. 6/ We leak a cloned bio if the target rdev is not usable. From: NeilBrown <neilb@suse.de> update 2: 1/ When aligned requests fail (read error) they need to be retried via the normal method (stripe cache). As we cannot be sure that we can process a single read in one go (we may not be able to allocate all the stripes needed) we store a bio-being-retried and a list of bioes-that-still-need-to-be-retried. When find a bio that needs to be retried, we should add it to the list, not to single-bio... 2/ We were never incrementing 'scnt' when resubmitting failed aligned requests. [akpm@osdl.org: build fix] Signed-off-by: Neil Brown <neilb@suse.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 10:20:47 +00:00
atomic_read(&conf->active_stripes) == 0 &&
atomic_read(&conf->active_aligned_reads) == 0,
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
unlock_all_device_hash_locks_irq(conf),
lock_all_device_hash_locks_irq(conf));
conf->quiesce = 1;
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
unlock_all_device_hash_locks_irq(conf);
/* allow reshape to continue */
wake_up(&conf->wait_for_overlap);
} else {
/* re-enable writes */
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
lock_all_device_hash_locks_irq(conf);
conf->quiesce = 0;
wake_up(&conf->wait_for_quiescent);
wake_up(&conf->wait_for_overlap);
raid5: relieve lock contention in get_active_stripe() get_active_stripe() is the last place we have lock contention. It has two paths. One is stripe isn't found and new stripe is allocated, the other is stripe is found. The first path basically calls __find_stripe and init_stripe. It accesses conf->generation, conf->previous_raid_disks, conf->raid_disks, conf->prev_chunk_sectors, conf->chunk_sectors, conf->max_degraded, conf->prev_algo, conf->algorithm, the stripe_hashtbl and inactive_list. Except stripe_hashtbl and inactive_list, other fields are changed very rarely. With this patch, we split inactive_list and add new hash locks. Each free stripe belongs to a specific inactive list. Which inactive list is determined by stripe's lock_hash. Note, even a stripe hasn't a sector assigned, it has a lock_hash assigned. Stripe's inactive list is protected by a hash lock, which is determined by it's lock_hash too. The lock_hash is derivied from current stripe_hashtbl hash, which guarantees any stripe_hashtbl list will be assigned to a specific lock_hash, so we can use new hash lock to protect stripe_hashtbl list too. The goal of the new hash locks introduced is we can only use the new locks in the first path of get_active_stripe(). Since we have several hash locks, lock contention is relieved significantly. The first path of get_active_stripe() accesses other fields, since they are changed rarely, changing them now need take conf->device_lock and all hash locks. For a slow path, this isn't a problem. If we need lock device_lock and hash lock, we always lock hash lock first. The tricky part is release_stripe and friends. We need take device_lock first. Neil's suggestion is we put inactive stripes to a temporary list and readd it to inactive_list after device_lock is released. In this way, we add stripes to temporary list with device_lock hold and remove stripes from the list with hash lock hold. So we don't allow concurrent access to the temporary list, which means we need allocate temporary list for all participants of release_stripe. One downside is free stripes are maintained in their inactive list, they can't across between the lists. By default, we have total 256 stripes and 8 lists, so each list will have 32 stripes. It's possible one list has free stripe but other list hasn't. The chance should be rare because stripes allocation are even distributed. And we can always allocate more stripes for cache, several mega bytes memory isn't a big deal. This completely removes the lock contention of the first path of get_active_stripe(). It slows down the second code path a little bit though because we now need takes two locks, but since the hash lock isn't contended, the overhead should be quite small (several atomic instructions). The second path of get_active_stripe() (basically sequential write or big request size randwrite) still has lock contentions. Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-11-14 04:16:17 +00:00
unlock_all_device_hash_locks_irq(conf);
}
log_quiesce(conf, quiesce);
}
static void *raid45_takeover_raid0(struct mddev *mddev, int level)
{
struct r0conf *raid0_conf = mddev->private;
sector_t sectors;
/* for raid0 takeover only one zone is supported */
if (raid0_conf->nr_strip_zones > 1) {
pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
mdname(mddev));
return ERR_PTR(-EINVAL);
}
sectors = raid0_conf->strip_zone[0].zone_end;
sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
mddev->dev_sectors = sectors;
mddev->new_level = level;
mddev->new_layout = ALGORITHM_PARITY_N;
mddev->new_chunk_sectors = mddev->chunk_sectors;
mddev->raid_disks += 1;
mddev->delta_disks = 1;
/* make sure it will be not marked as dirty */
mddev->recovery_cp = MaxSector;
return setup_conf(mddev);
}
static void *raid5_takeover_raid1(struct mddev *mddev)
{
int chunksect;
void *ret;
if (mddev->raid_disks != 2 ||
mddev->degraded > 1)
return ERR_PTR(-EINVAL);
/* Should check if there are write-behind devices? */
chunksect = 64*2; /* 64K by default */
/* The array must be an exact multiple of chunksize */
while (chunksect && (mddev->array_sectors & (chunksect-1)))
chunksect >>= 1;
if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
/* array size does not allow a suitable chunk size */
return ERR_PTR(-EINVAL);
mddev->new_level = 5;
mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
mddev->new_chunk_sectors = chunksect;
ret = setup_conf(mddev);
if (!IS_ERR(ret))
mddev_clear_unsupported_flags(mddev,
UNSUPPORTED_MDDEV_FLAGS);
return ret;
}
static void *raid5_takeover_raid6(struct mddev *mddev)
{
int new_layout;
switch (mddev->layout) {
case ALGORITHM_LEFT_ASYMMETRIC_6:
new_layout = ALGORITHM_LEFT_ASYMMETRIC;
break;
case ALGORITHM_RIGHT_ASYMMETRIC_6:
new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
break;
case ALGORITHM_LEFT_SYMMETRIC_6:
new_layout = ALGORITHM_LEFT_SYMMETRIC;
break;
case ALGORITHM_RIGHT_SYMMETRIC_6:
new_layout = ALGORITHM_RIGHT_SYMMETRIC;
break;
case ALGORITHM_PARITY_0_6:
new_layout = ALGORITHM_PARITY_0;
break;
case ALGORITHM_PARITY_N:
new_layout = ALGORITHM_PARITY_N;
break;
default:
return ERR_PTR(-EINVAL);
}
mddev->new_level = 5;
mddev->new_layout = new_layout;
mddev->delta_disks = -1;
mddev->raid_disks -= 1;
return setup_conf(mddev);
}
static int raid5_check_reshape(struct mddev *mddev)
{
/* For a 2-drive array, the layout and chunk size can be changed
* immediately as not restriping is needed.
* For larger arrays we record the new value - after validation
* to be used by a reshape pass.
*/
struct r5conf *conf = mddev->private;
int new_chunk = mddev->new_chunk_sectors;
if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
return -EINVAL;
if (new_chunk > 0) {
if (!is_power_of_2(new_chunk))
return -EINVAL;
if (new_chunk < (PAGE_SIZE>>9))
return -EINVAL;
if (mddev->array_sectors & (new_chunk-1))
/* not factor of array size */
return -EINVAL;
}
/* They look valid */
if (mddev->raid_disks == 2) {
/* can make the change immediately */
if (mddev->new_layout >= 0) {
conf->algorithm = mddev->new_layout;
mddev->layout = mddev->new_layout;
}
if (new_chunk > 0) {
conf->chunk_sectors = new_chunk ;
mddev->chunk_sectors = new_chunk;
}
set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
md_wakeup_thread(mddev->thread);
}
return check_reshape(mddev);
}
static int raid6_check_reshape(struct mddev *mddev)
{
int new_chunk = mddev->new_chunk_sectors;
if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
return -EINVAL;
if (new_chunk > 0) {
if (!is_power_of_2(new_chunk))
return -EINVAL;
if (new_chunk < (PAGE_SIZE >> 9))
return -EINVAL;
if (mddev->array_sectors & (new_chunk-1))
/* not factor of array size */
return -EINVAL;
}
/* They look valid */
return check_reshape(mddev);
}
static void *raid5_takeover(struct mddev *mddev)
{
/* raid5 can take over:
* raid0 - if there is only one strip zone - make it a raid4 layout
* raid1 - if there are two drives. We need to know the chunk size
* raid4 - trivial - just use a raid4 layout.
* raid6 - Providing it is a *_6 layout
*/
if (mddev->level == 0)
return raid45_takeover_raid0(mddev, 5);
if (mddev->level == 1)
return raid5_takeover_raid1(mddev);
if (mddev->level == 4) {
mddev->new_layout = ALGORITHM_PARITY_N;
mddev->new_level = 5;
return setup_conf(mddev);
}
if (mddev->level == 6)
return raid5_takeover_raid6(mddev);
return ERR_PTR(-EINVAL);
}
static void *raid4_takeover(struct mddev *mddev)
{
/* raid4 can take over:
* raid0 - if there is only one strip zone
* raid5 - if layout is right
*/
if (mddev->level == 0)
return raid45_takeover_raid0(mddev, 4);
if (mddev->level == 5 &&
mddev->layout == ALGORITHM_PARITY_N) {
mddev->new_layout = 0;
mddev->new_level = 4;
return setup_conf(mddev);
}
return ERR_PTR(-EINVAL);
}
static struct md_personality raid5_personality;
static void *raid6_takeover(struct mddev *mddev)
{
/* Currently can only take over a raid5. We map the
* personality to an equivalent raid6 personality
* with the Q block at the end.
*/
int new_layout;
if (mddev->pers != &raid5_personality)
return ERR_PTR(-EINVAL);
if (mddev->degraded > 1)
return ERR_PTR(-EINVAL);
if (mddev->raid_disks > 253)
return ERR_PTR(-EINVAL);
if (mddev->raid_disks < 3)
return ERR_PTR(-EINVAL);
switch (mddev->layout) {
case ALGORITHM_LEFT_ASYMMETRIC:
new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
break;
case ALGORITHM_RIGHT_ASYMMETRIC:
new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
break;
case ALGORITHM_LEFT_SYMMETRIC:
new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
break;
case ALGORITHM_RIGHT_SYMMETRIC:
new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
break;
case ALGORITHM_PARITY_0:
new_layout = ALGORITHM_PARITY_0_6;
break;
case ALGORITHM_PARITY_N:
new_layout = ALGORITHM_PARITY_N;
break;
default:
return ERR_PTR(-EINVAL);
}
mddev->new_level = 6;
mddev->new_layout = new_layout;
mddev->delta_disks = 1;
mddev->raid_disks += 1;
return setup_conf(mddev);
}
static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
{
struct r5conf *conf;
int err;
err = mddev_lock(mddev);
if (err)
return err;
conf = mddev->private;
if (!conf) {
mddev_unlock(mddev);
return -ENODEV;
}
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (strncmp(buf, "ppl", 3) == 0) {
/* ppl only works with RAID 5 */
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
if (!raid5_has_ppl(conf) && conf->level == 5) {
err = log_init(conf, NULL, true);
if (!err) {
err = resize_stripes(conf, conf->pool_size);
if (err)
log_exit(conf);
}
} else
err = -EINVAL;
} else if (strncmp(buf, "resync", 6) == 0) {
if (raid5_has_ppl(conf)) {
mddev_suspend(mddev);
log_exit(conf);
mddev_resume(mddev);
raid5-ppl: use resize_stripes() when enabling or disabling ppl Use resize_stripes() instead of raid5_reset_stripe_cache() to allocate or free sh->ppl_page at runtime for all stripes in the stripe cache. raid5_reset_stripe_cache() required suspending the mddev and could deadlock because of GFP_KERNEL allocations. Move the 'newsize' check to check_reshape() to allow reallocating the stripes with the same number of disks. Allocate sh->ppl_page in alloc_stripe() instead of grow_buffers(). Pass 'struct r5conf *conf' as a parameter to alloc_stripe() because it is needed to check whether to allocate ppl_page. Add free_stripe() and use it to free stripes rather than directly call kmem_cache_free(). Also free sh->ppl_page in free_stripe(). Set MD_HAS_PPL at the end of ppl_init_log() instead of explicitly setting it in advance and add another parameter to log_init() to allow calling ppl_init_log() without the bit set. Don't try to calculate partial parity or add a stripe to log if it does not have ppl_page set. Enabling ppl can now be performed without suspending the mddev, because the log won't be used until new stripes are allocated with ppl_page. Calling mddev_suspend/resume is still necessary when disabling ppl, because we want all stripes to finish before stopping the log, but resize_stripes() can be called after mddev_resume() when ppl is no longer active. Suggested-by: NeilBrown <neilb@suse.com> Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com> Signed-off-by: Shaohua Li <shli@fb.com>
2017-04-04 11:13:57 +00:00
err = resize_stripes(conf, conf->pool_size);
} else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
r5l_log_disk_error(conf)) {
bool journal_dev_exists = false;
struct md_rdev *rdev;
rdev_for_each(rdev, mddev)
if (test_bit(Journal, &rdev->flags)) {
journal_dev_exists = true;
break;
}
if (!journal_dev_exists) {
mddev_suspend(mddev);
clear_bit(MD_HAS_JOURNAL, &mddev->flags);
mddev_resume(mddev);
} else /* need remove journal device first */
err = -EBUSY;
} else
err = -EINVAL;
} else {
err = -EINVAL;
}
if (!err)
md_update_sb(mddev, 1);
mddev_unlock(mddev);
return err;
}
static int raid5_start(struct mddev *mddev)
{
struct r5conf *conf = mddev->private;
return r5l_start(conf->log);
}
static struct md_personality raid6_personality =
{
.name = "raid6",
.level = 6,
.owner = THIS_MODULE,
.make_request = raid5_make_request,
.run = raid5_run,
.start = raid5_start,
.free = raid5_free,
.status = raid5_status,
.error_handler = raid5_error,
.hot_add_disk = raid5_add_disk,
.hot_remove_disk= raid5_remove_disk,
.spare_active = raid5_spare_active,
.sync_request = raid5_sync_request,
.resize = raid5_resize,
.size = raid5_size,
.check_reshape = raid6_check_reshape,
.start_reshape = raid5_start_reshape,
.finish_reshape = raid5_finish_reshape,
.quiesce = raid5_quiesce,
.takeover = raid6_takeover,
.change_consistency_policy = raid5_change_consistency_policy,
};
static struct md_personality raid5_personality =
{
.name = "raid5",
.level = 5,
.owner = THIS_MODULE,
.make_request = raid5_make_request,
.run = raid5_run,
.start = raid5_start,
.free = raid5_free,
.status = raid5_status,
.error_handler = raid5_error,
.hot_add_disk = raid5_add_disk,
.hot_remove_disk= raid5_remove_disk,
.spare_active = raid5_spare_active,
.sync_request = raid5_sync_request,
.resize = raid5_resize,
.size = raid5_size,
.check_reshape = raid5_check_reshape,
.start_reshape = raid5_start_reshape,
.finish_reshape = raid5_finish_reshape,
.quiesce = raid5_quiesce,
.takeover = raid5_takeover,
.change_consistency_policy = raid5_change_consistency_policy,
};
static struct md_personality raid4_personality =
{
.name = "raid4",
.level = 4,
.owner = THIS_MODULE,
.make_request = raid5_make_request,
.run = raid5_run,
.start = raid5_start,
.free = raid5_free,
.status = raid5_status,
.error_handler = raid5_error,
.hot_add_disk = raid5_add_disk,
.hot_remove_disk= raid5_remove_disk,
.spare_active = raid5_spare_active,
.sync_request = raid5_sync_request,
.resize = raid5_resize,
.size = raid5_size,
.check_reshape = raid5_check_reshape,
.start_reshape = raid5_start_reshape,
.finish_reshape = raid5_finish_reshape,
.quiesce = raid5_quiesce,
.takeover = raid4_takeover,
.change_consistency_policy = raid5_change_consistency_policy,
};
static int __init raid5_init(void)
{
int ret;
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
raid5_wq = alloc_workqueue("raid5wq",
WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
if (!raid5_wq)
return -ENOMEM;
ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
"md/raid5:prepare",
raid456_cpu_up_prepare,
raid456_cpu_dead);
if (ret) {
destroy_workqueue(raid5_wq);
return ret;
}
register_md_personality(&raid6_personality);
register_md_personality(&raid5_personality);
register_md_personality(&raid4_personality);
return 0;
}
static void raid5_exit(void)
{
unregister_md_personality(&raid6_personality);
unregister_md_personality(&raid5_personality);
unregister_md_personality(&raid4_personality);
cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
raid5: offload stripe handle to workqueue This is another attempt to create multiple threads to handle raid5 stripes. This time I use workqueue. raid5 handles request (especially write) in stripe unit. A stripe is page size aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a state machine for the corresponding stripe, which includes reading some disks of the stripe, calculating parity, and writing some disks of the stripe. The state machine is running in raid5d thread currently. Since there is only one thread, it doesn't scale well for high speed storage. An obvious solution is multi-threading. To get better performance, we have some requirements: a. locality. stripe corresponding to request submitted from one cpu is better handled in thread in local cpu or local node. local cpu is preferred but some times could be a bottleneck, for example, parity calculation is too heavy. local node running has wide adaptability. b. configurablity. Different setup of raid5 array might need diffent configuration. Especially the thread number. More threads don't always mean better performance because of lock contentions. My original implementation is creating some kernel threads. There are interfaces to control which cpu's stripe each thread should handle. And userspace can set affinity of the threads. This provides biggest flexibility and configurability. But it's hard to use and apparently a new thread pool implementation is disfavor. Recent workqueue improvement is quite promising. unbound workqueue will be bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to do affinity setting. For example, we can only include one HT sibling in affinity. Since work is non-reentrant by default, and we can control running thread number by limiting dispatched work_struct number. In this patch, I created several stripe worker group. A group is a numa node. stripes from cpus of one node will be added to a group list. Workqueue thread of one node will only handle stripes of worker group of the node. In this way, stripe handling has numa node locality. And as I said, we can control thread number by limiting dispatched work_struct number. The work_struct callback function handles several stripes in one run. A typical work queue usage is to run one unit in each work_struct. In raid5 case, the unit is a stripe. But we can't do that: a. Though handling a stripe doesn't need lock because of reference accounting and stripe isn't in any list, queuing a work_struct for each stripe will make workqueue lock contended very heavily. b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we might dispatch request. If each work_struct only handles one stripe, such block plug is meaningless. This implementation can't do very fine grained configuration. But the numa binding is most popular usage model, should be enough for most workloads. Note: since we have only one stripe queue, switching to multi-thread might decrease request size dispatching down to low level layer. The impact depends on thread number, raid configuration and workload. So multi-thread raid5 might not be proper for all setups. Changes V1 -> V2: 1. remove WQ_NON_REENTRANT 2. disabling multi-threading by default 3. Add more descriptions in changelog Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: NeilBrown <neilb@suse.de>
2013-08-28 06:30:16 +00:00
destroy_workqueue(raid5_wq);
}
module_init(raid5_init);
module_exit(raid5_exit);
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
MODULE_ALIAS("md-personality-4"); /* RAID5 */
MODULE_ALIAS("md-raid5");
MODULE_ALIAS("md-raid4");
MODULE_ALIAS("md-level-5");
MODULE_ALIAS("md-level-4");
MODULE_ALIAS("md-personality-8"); /* RAID6 */
MODULE_ALIAS("md-raid6");
MODULE_ALIAS("md-level-6");
/* This used to be two separate modules, they were: */
MODULE_ALIAS("raid5");
MODULE_ALIAS("raid6");