With the update of the truncate code, ip->i_disksize and
inode->i_size are merely copies of each other. This means
we can remove ip->i_disksize and use inode->i_size exclusively
reducing the size of a GFS2 inode by 8 bytes.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq: (55 commits)
workqueue: mark init_workqueues() as early_initcall()
workqueue: explain for_each_*cwq_cpu() iterators
fscache: fix build on !CONFIG_SYSCTL
slow-work: kill it
gfs2: use workqueue instead of slow-work
drm: use workqueue instead of slow-work
cifs: use workqueue instead of slow-work
fscache: drop references to slow-work
fscache: convert operation to use workqueue instead of slow-work
fscache: convert object to use workqueue instead of slow-work
workqueue: fix how cpu number is stored in work->data
workqueue: fix mayday_mask handling on UP
workqueue: fix build problem on !CONFIG_SMP
workqueue: fix locking in retry path of maybe_create_worker()
async: use workqueue for worker pool
workqueue: remove WQ_SINGLE_CPU and use WQ_UNBOUND instead
workqueue: implement unbound workqueue
workqueue: prepare for WQ_UNBOUND implementation
libata: take advantage of cmwq and remove concurrency limitations
workqueue: fix worker management invocation without pending works
...
Fixed up conflicts in fs/cifs/* as per Tejun. Other trivial conflicts in
include/linux/workqueue.h, kernel/trace/Kconfig and kernel/workqueue.c
This patch implements a wait for the journal id in the case that it has
not been specified on the command line. This is to allow the future
removal of the mount.gfs2 helper. The journal id would instead be
directly communicated by gfs_controld to the file system. Here is a
comparison of the two systems:
Current:
1. mount calls mount.gfs2
2. mount.gfs2 connects to gfs_controld to retrieve the journal id
3. mount.gfs2 adds the journal id to the mount command line and calls
the mount system call
4. gfs_controld receives the status of the mount request via a uevent
Proposed:
1. mount calls the mount system call (no mount.gfs2 helper)
2. gfs_controld receives a uevent for a gfs2 fs which it doesn't know
about already
3. gfs_controld assigns a journal id to it via sysfs
4. the mount system call then completes as normal (sending a uevent
according to status)
The advantage of the proposed system is that it is completely backward
compatible with the current system both at the kernel and at the
userland levels. The "first" parameter can also be set the same way,
with the restriction that it must be set before the journal id is
assigned.
In addition, if mount becomes stuck waiting for a reply from
gfs_controld which never arrives, then it is killable and will abort the
mount gracefully.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Workqueue can now handle high concurrency. Convert gfs to use
workqueue instead of slow-work.
* Steven pointed out that recovery path might be run from allocation
path and thus requires forward progress guarantee without memory
allocation. Create and use gfs_recovery_wq with rescuer. Please
note that forward progress wasn't guaranteed with slow-work.
* Updated to use non-reentrant workqueue.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
The following patch adds a message to indicate when barriers have been
disabled due to a block device which doesn't support them. You could
already tell this via the mount options in /proc/mounts, but all the
other filesystems also log a message at the same time.
Also, the same mechanisms are used to indicate when the lock
demote interface has been used (only ever used for debugging)
which is a request from our support team.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch contains various tweaks to how log flushes and active item writeback
work. gfs2_logd is now managed by a waitqueue, and gfs2_log_reseve now waits
for gfs2_logd to do the log flushing. Multiple functions were rewritten to
remove the need to call gfs2_log_lock(). Instead of using one test to see if
gfs2_logd had work to do, there are now seperate tests to check if there
are two many buffers in the incore log or if there are two many items on the
active items list.
This patch is a port of a patch Steve Whitehouse wrote about a year ago, with
some minor changes. Since gfs2_ail1_start always submits all the active items,
it no longer needs to keep track of the first ai submitted, so this has been
removed. In gfs2_log_reserve(), the order of the calls to
prepare_to_wait_exclusive() and wake_up() when firing off the logd thread has
been switched. If it called wake_up first there was a small window for a race,
where logd could run and return before gfs2_log_reserve was ready to get woken
up. If gfs2_logd ran, but did not free up enough blocks, gfs2_log_reserve()
would be left waiting for gfs2_logd to eventualy run because it timed out.
Finally, gt_logd_secs, which controls how long to wait before gfs2_logd times
out, and flushes the log, can now be set on mount with ar_commit.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
GFS2 tracks the number of revokes and unrevokes that are part of committed
transactions via sd_log_commited_revoke. It is possible for one process to add
revokes during its transaction, while another process unrevokes them during its
transaction. If the second process finishes its transaction first,
sd_log_commited_revoke will be decremented by the number of unrevokes that the
second process did, without first being incremented by the number of revokes
the first process did. This is fine, since all started transactions must be
completed before the journal can be flushed. However, sd_log_commited_revoke
is an unsigned integer, and log_refund() causes an assertion failure if it
would go negative at the end of a transaction. This patch makes
sd_log_commited_revoke a signed integer and allows it to go negative.
__gfs2_log_flush() still checks that it mataches the actual number of revokes.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As a consequence of the previous patch, we can now remove the
loop which used to be required due to the circular dependency
between the inodes and glocks. Instead we can just invalidate
the inodes, and then clear up any glocks which are left.
Also we no longer need the rwsem since there is no longer any
danger of the inode invalidation calling back into the glock
code (and from there back into the inode code).
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Since the start of GFS2, an "extra" inode has been used to store
the metadata belonging to each inode. The only reason for using
this inode was to have an extra address space, the other fields
were unused. This means that the memory usage was rather inefficient.
The reason for keeping each inode's metadata in a separate address
space is that when glocks are requested on remote nodes, we need to
be able to efficiently locate the data and metadata which relating
to that glock (inode) in order to sync or sync and invalidate it
(depending on the remotely requested lock mode).
This patch adds a new type of glock, which has in addition to
its normal fields, has an address space. This applies to all
inode and rgrp glocks (but to no other glock types which remain
as before). As a result, we no longer need to have the second
inode.
This results in three major improvements:
1. A saving of approx 25% of memory used in caching inodes
2. A removal of the circular dependency between inodes and glocks
3. No confusion between "normal" and "metadata" inodes in super.c
Although the first of these is the more immediately apparent, the
second is just as important as it now enables a number of clean
ups at umount time. Those will be the subject of future patches.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch adds a wait on umount between the point at which we
dispose of all glocks and the point at which we unmount the
lock protocol. This ensures that we've received all the replies
to our unlock requests before we stop the locking.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-by: Fabio M. Di Nitto <fdinitto@redhat.com>
Currently gfs2 issues barrier unconditionally. There are various reasons
to disable them, be that just for testing or for stupid devices flushing
large battert backed caches. Add a nobarrier option that matches xfs and
btrfs for this. Also add a symmetric barrier option to turn it back on
at remount time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
GFS2 now has three new mount options, statfs_quantum, quota_quantum and
statfs_percent. statfs_quantum and quota_quantum simply allow you to
set the tunables of the same name. Setting setting statfs_quantum to 0
will also turn on the statfs_slow tunable. statfs_percent accepts an
integer between 0 and 100. Numbers between 1 and 100 will cause GFS2 to
do any early sync when the local number of blocks free changes by at
least statfs_percent from the totoal number of blocks free. Setting
statfs_percent to 0 disables this.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The /sys/fs/gfs2/<fsname>/lock_module/id file has been unused for
some time now, so we can remove it. We still accept the mount option
though, as userspace still sends that.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The inum structure used throughout GFS2 has two fields. One
no_addr is the disk block number of the inode in question and
is used everywhere as the inode number. The other, no_formal_ino,
is used only as the generation number for NFS.
Historically the no_formal_ino field was set using a complicated
system of one global and one per-node file containing inode numbers
in order to ensure that each no_formal_ino was unique. Also this
code made no provision for what would happen when eventually the
(64 bit) numbers ran out. Now I know that is pretty unlikely to
happen given the large space of numbers, but it is possible
nevertheless.
The only guarantee required for no_formal_ino is that, for any
single inode, the same number doesn't get reused too quickly.
We already have a generation number which is kept in the inode
and initialised from a counter in the resource group (almost
no overhead, since we have to touch the resource group anyway
in order to allocate an inode in the first place). Aside from
ensuring that we never use the value 0 in the no_formal_ino
field, we can use that counter directly.
As a result of that change, we lose about 200 lines of code and
also gain about 10 creates/sec on the postmark benchmark (on
my test machine).
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch adds "-o errors=panic" and "-o errors=withdraw" to the
gfs2 mount options. The "errors=withdraw" option is today's
current behaviour, meaning to withdraw from the file system if a
non-serious gfs2 error occurs. The new "errors=panic" option
tells gfs2 to force a kernel panic if a non-serious gfs2 file
system error occurs. This may be useful, for example, where
fabric-level fencing is used that has no way to reboot (such as
fence_scsi).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
When a file is deleted from a gfs2 filesystem on one node, a dcache
entry for it may still exist on other nodes in the cluster. If this
happens, gfs2 will be unable to free this file on disk. Because of this,
it's possible to have a gfs2 filesystem with no files on it and no free
space. With this patch, when a node receives a callback notifying it
that the file is being deleted on another node, it schedules a new
workqueue thread to remove the file's dcache entry.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch increases the frequency with which gfs2 looks
for unlinked, but still allocated inodes. Its the equivalent
operation to ext3's orphan list, but done with bitmaps in
the resource groups.
This also fixes a bug where a field in the rgrp was too small.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
During block allocation, it is useful to know if sections of disk
are full on a finer grained basis than a single resource group.
This can make a performance difference when resource groups have
larger numbers of bitmap blocks, since we no longer have to search
them all block by block in each individual bitmap.
The full flag is set on a per-bitmap basis when it has been
searched and found to have no free space. It is then skipped in
subsequent searches until the flag is reset. The resetting
occurs if we have to drop the glock on the resource group for any
reason, or if we deallocate some blocks within that resource
group and thus free up some space.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch improves the error handling in the case where we
discover that the summary information in the resource group
doesn't match the bitmap information while in the process of
allocating blocks. Originally this resulted in a kernel bug,
but this patch changes that so that we return -EIO and print
some messages explaining what went wrong, and how to fix it.
We also remember locally not to try and allocate from the
same rgrp again, so that a subsequent allocation in a
different rgrp should succeed.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes a race condition where we can receive recovery
requests part way through processing a umount. This was causing
problems since the recovery thread had already gone away.
Looking in more detail at the recovery code, it was really trying
to implement a slight variation on a work queue, and that happens to
align nicely with the recently introduced slow-work subsystem. As a
result I've updated the code to use slow-work, rather than its own home
grown variety of work queue.
When using the wait_on_bit() function, I noticed that the wait function
that was supplied as an argument was appearing in the WCHAN field, so
I've updated the function names in order to produce more meaningful
output.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
It has always been possible to adjust the gfs2 log commit
interval, but only from the sysfs interface. This adds a
mount option, commit=<nn>, which will be familar to ext3
users.
The sysfs interface continues to be available as well, although
this might be removed in the future.
Also this patch cleans up some duplicated structures in the GFS2
sysfs code.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
After calling out to the dlm, GFS2 sets the new state of a glock to
gl_target in gdlm_ast(). However, gl_target is not always the lock
state that was requested. If a conversion from shared to exclusive
fails, finish_xmote() will call do_xmote() with LM_ST_UNLOCKED, instead
of gl->gl_target, so that it can reacquire the lock in exlusive the next
time around. In this case, setting the lock to gl_target in gdlm_ast()
will make GFS2 think that it has the glock in exclusive mode, when
really, it doesn't have the glock locked at all. This patch adds a new
field to the gfs2_glock structure, gl_req, to track the mode that was
requested.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Since we have a UUID, we ought to expose it to the user via sysfs
and uevents. We already have the fs name in both of these places
(a combination of the lock proto and lock table name) so if we add
the UUID as well, we have a full set.
For older filesystems (i.e. those created before mkfs.gfs2 was writing
UUIDs by default) the sysfs file will appear zero length, and no UUID
env var will be added to the uevents.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch allows GFS2 to generate discard requests for blocks which are
no longer useful to the filesystem (i.e. those which have been freed as
the result of an unlink operation). The requests are generated at the
time which those blocks become available for reuse in the filesystem.
In order to use this new feature, you have to specify the "discard"
mount option. The code coalesces adjacent blocks into a single extent
when generating the discard requests, thus generating the minimum
number.
If an error occurs when the request has been sent to the block device,
then it will print a message and turn off the requests for that
filesystem. If the problem is temporary, then you can use remount to
turn the option back on again. There is also a nodiscard mount option
so that you can use remount to turn discard requests off, if required.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The time stamp field is unused in the glock now that we are
using a shrinker, so that we can remove it and save sizeof(unsigned long)
bytes in each glock.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is the big patch that I've been working on for some time
now. There are many reasons for wanting to make this change
such as:
o Reducing overhead by eliminating duplicated fields between structures
o Simplifcation of the code (reduces the code size by a fair bit)
o The locking interface is now the DLM interface itself as proposed
some time ago.
o Fewer lookups of glocks when processing replies from the DLM
o Fewer memory allocations/deallocations for each glock
o Scope to do further optimisations in the future (but this patch is
more than big enough for now!)
Please note that (a) this patch relates to the lock_dlm module and
not the DLM itself, that is still a separate module; and (b) that
we retain the ability to build GFS2 as a standalone single node
filesystem with out requiring the DLM.
This patch needs a lot of testing, hence my keeping it I restarted
my -git tree after the last merge window. That way, this has the maximum
exposure before its merged. This is (modulo a few minor bug fixes) the
same patch that I've been posting on and off the the last three months
and its passed a number of different tests so far.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We only really need a single spin lock for the quota data, so
lets just use the lru lock for now.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Abhijith Das <adas@redhat.com>
Deallocation of gfs2_quota_data objects now happens on-demand through a
shrinker instead of routinely deallocating through the quotad daemon.
Signed-off-by: Abhijith Das <adas@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch removes the two daemons, gfs2_scand and gfs2_glockd
and replaces them with a shrinker which is called from the VM.
The net result is that GFS2 responds better when there is memory
pressure, since it shrinks the glock cache at the same rate
as the VFS shrinks the dcache and icache. There are no longer
any time based criteria for shrinking glocks, they are kept
until such time as the VM asks for more memory and then we
demote just as many glocks as required.
There are potential future changes to this code, including the
possibility of sorting the glocks which are to be written back
into inode number order, to get a better I/O ordering. It would
be very useful to have an elevator based workqueue implementation
for this, as that would automatically deal with the read I/O cases
at the same time.
This patch is my answer to Andrew Morton's remark, made during
the initial review of GFS2, asking why GFS2 needs so many kernel
threads, the answer being that it doesn't :-) This patch is a
net loss of about 200 lines of code.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Following on from the recent clean up of gfs2_quotad, this patch moves
the processing of "truncate in progress" inodes from the glock workqueue
into gfs2_quotad. This fixes a hang due to the "truncate in progress"
processing requiring glocks in order to complete.
It might seem odd to use gfs2_quotad for this particular item, but
we have to use a pre-existing thread since creating a thread implies
a GFP_KERNEL memory allocation which is not allowed from the glock
workqueue context. Of the existing threads, gfs2_logd and gfs2_recoverd
may deadlock if used for this operation. gfs2_scand and gfs2_glockd are
both scheduled for removal at some (hopefully not too distant) future
point. That leaves only gfs2_quotad whose workload is generally fairly
light and is easily adapted for this extra task.
Also, as a result of this change, it opens the way for a future patch to
make the reading of the inode's information asynchronous with respect to
the glock workqueue, which is another improvement that has been on the list
for some time now.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is a clean up of gfs2_quotad prior to giving it an
extra job to do in addition to the current portfolio of updating
the quota and statfs information from time to time.
As a result it has been moved into quota.c allowing one of the
functions it calls to be made static. Also the clean up allows
the two existing functions to have separate timeouts and also
to coexist with its future role of dealing with the "truncate in
progress" inode flag.
The (pointless) setting of gfs2_quotad_secs is removed since we
arrange to only wake up quotad when one of the two timers expires.
In addition the struct gfs2_quota_data is moved into a slab cache,
mainly for easier debugging. It should also be possible to use
a shrinker in the future, rather than the current scheme of scanning
the quota data entries from time to time.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch moves the final field so that we can get rid
of struct gfs2_rgrpd_host, as promised some time ago. Also
by rearranging the fields slightly, we are able to reduce
the size of the gfs2_rgrpd structure at the same time.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This moves one of the fields of struct gfs2_rgrpd_host into
the struct gfs2_rgrpd with the eventual aim of removing
the struct rgrpd_host completely.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The final field in gfs2_dinode_host was the i_flags field. Thats
renamed to i_diskflags in order to avoid confusion with the existing
inode flags, and moved into the inode proper at a suitable location
to avoid creating a "hole".
At that point struct gfs2_dinode_host is no longer needed and as
promised (quite some time ago!) it can now be removed completely.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch moved the i_size field from the gfs2_dinode_host and
following the ext3 convention renames it i_disksize.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This moves the directory entry count into the proper inode.
Potentially we could get this to share the space used by
something else in the future, but this is one more step
on the way to removing the gfs2_dinode_host structure.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This moves the generation number from the gfs2_dinode_host
into the gfs2_inode structure. Eventually the plan is to get
rid of the gfs2_dinode_host structure completely.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch adds barrier support to GFS2. There is not a lot of change
really... we just add the barrier flag when we write journal header
blocks. If the underlying device refuses to support them, we fall back
to the previous way of doing things (wait for the I/O and hope) since
there is nothing else we can do. There is no user configuration,
barriers will always be on unless the device refuses to support them.
This seems a reasonable solution to me since this is a correctness
issue.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Until now, we've used the same scheme as GFS1 for atime. This has failed
since atime is a per vfsmnt flag, not a per fs flag and as such the
"noatime" flag was not getting passed down to the filesystems. This
patch removes all the "special casing" around atime updates and we
simply use the VFS's atime code.
The net result is that GFS2 will now support all the same atime related
mount options of any other filesystem on a per-vfsmnt basis. We do lose
the "lazy atime" updates, but we gain "relatime". We could add lazy
atime to the VFS at a later date, if there is a requirement for that
variant still - I suspect relatime will be enough.
Also we lose about 100 lines of code after this patch has been applied,
and I have a suspicion that it will speed things up a bit, even when
atime is "on". So it seems like a nice clean up as well.
From a user perspective, everything stays the same except the loss of
the per-fs atime quantum tweekable (ought to be per-vfsmnt at the very
least, and to be honest I don't think anybody ever used it) and that a
number of options which were ignored before now work correctly.
Please let me know if you've got any comments. I'm pushing this out
early so that you can all see what my plans are.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The following patch shrinks the gfs2_args structure which is embedded in
every GFS2 superblock. It cuts down the size of the options to a single
unsigned int (the 13 bits of bitfields will be rounded up to that size
by the compiler) from the current 11 unsigned ints. So on x86 thats 44
bytes shrinking to 4 bytes, in each and every GFS2 superblock.
Signed-off-by: Steven Whitehouse <swhitho@redhat.com>
This patch is intended to fix the issues reported in bz #457798. Instead
of having the metafs as a separate filesystem, it becomes a second root
of gfs2. As a result it will appear as type gfs2 in /proc/mounts, but it
is still possible (for backwards compatibility purposes) to mount it as
type gfs2meta. A new mount flag "meta" is introduced so that its possible
to tell the two cases apart in /proc/mounts.
As a result it becomes possible to mount type gfs2 with -o meta and
get the same result as mounting type gfs2meta. So it is possible to
mount just the metafs on its own. Currently if you do this, its then
impossible to mount the "normal" root of the gfs2 filesystem without
first unmounting the metafs root. I'm not sure if thats a feature or
a bug :-)
Either way, this is a great improvement on the previous scheme and I've
verified that it works ok with bind mounts on both the "normal" root
and the metafs root in various combinations.
There were also a bunch of functions in super.c which didn't belong there,
so this moves them into ops_fstype.c where they can be static. Hopefully
the mount/umount sequence is now more obvious as a result.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Alexander Viro <aviro@redhat.com>
The ability to mark files for direct i/o access when opened
normally is both unused and pointless, so this patch removes
support for that feature.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch removes the "recent list" which is used during allocation
and replaces it with the (already existing) mru list used during
deletion. The "recent list" was not a true mru list leading to a number
of inefficiencies including a "next" function which made scanning the
list an order N^2 operation wrt to the number of list elements.
This should increase allocation performance with large numbers of rgrps.
Its also a useful preparation and cleanup before some further changes
which are planned in this area.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch implements a number of cleanups to the core of the
GFS2 glock code. As a result a lot of code is removed. It looks
like a really big change, but actually a large part of this patch
is either removing or moving existing code.
There are some new bits too though, such as the new run_queue()
function which is considerably streamlined. Highlights of this
patch include:
o Fixes a cluster coherency bug during SH -> EX lock conversions
o Removes the "glmutex" code in favour of a single bit lock
o Removes the ->go_xmote_bh() for inodes since it was duplicating
->go_lock()
o We now only use the ->lm_lock() function for both locks and
unlocks (i.e. unlock is a lock with target mode LM_ST_UNLOCKED)
o The fast path is considerably shortly, giving performance gains
especially with lock_nolock
o The glock_workqueue is now used for all the callbacks from the DLM
which allows us to simplify the lock_dlm module (see following patch)
o The way is now open to make further changes such as eliminating the two
threads (gfs2_glockd and gfs2_scand) in favour of a more efficient
scheme.
This patch has undergone extensive testing with various test suites
so it should be pretty stable by now.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Bob Peterson <rpeterso@redhat.com>
This patch fixes a GFS2 filesystem consistency error reported from
function do_strip. The problem was caused by a timing window
that allowed two vfs inodes to be created in memory that point
to the same file. The problem is fixed by making the vfs's
iget_test, iget_set mechanism check and set a new bit in the
in-core gfs2_inode structure while the vfs inode spin_lock is held.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
GFS2 wasn't invalidating its cache before it called into the lock manager
with a request that could potentially drop a lock. This was leaving a
window where the lock could be actually be held by another node, but the
file's page cache would still appear valid, causing coherency problems.
This patch moves the cache invalidation to before the lock manager call
when dropping a lock. It also adds the option to the lock_dlm lock
manager to not use conversion mode deadlock avoidance, which, on a
conversion from shared to exclusive, could internally drop the lock, and
then reacquire in. GFS2 now asks lock_dlm to not do this. Instead, GFS2
manually drops the lock and reacquires it.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The blocks counter is almost a duplicate of the i_blocks
field in the VFS inode. The only difference is that i_blocks
can be only 32bits long for 32bit arch without large single file
support. Since GFS2 doesn't handle the non-large single file
case (for 32 bit anyway) this adds a new config dependency on
64BIT || LSF. This has always been the case, however we've never
explicitly said so before.
Even if we do add support for the non-LSF case, we will still
not require this field to be duplicated since we will not be
able to access oversized files anyway.
So the net result of all this is that we shave 8 bytes from a gfs2_inode
and get our config deps correct.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We don't need to keep track of when we last allocated data
and metadata separately since the only thing thats important
when searching for a free block is whether its free or not,
which is independent from what type of block it is.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
There were three fields being used to keep track of the location
of the most recently allocated block for each inode. These have
been merged into a single field in order to better keep the
data and metadata for an inode close on disk, and also to reduce
the space required for storage.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is kind of trivial in the greater scheme of things, but
this removes three counters that AFAICT are never used.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch forms a pair with the previous patch which shrunk
di_height. Like that patch di_depth is renamed i_depth and moved
into struct gfs2_inode directly. Also the field goes from 16 bits
to 8 bits since it is also limited to a max value which is rather
small (17 in this case). In addition we also now validate the field
against this maximum value when its read in.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch further reduces GFS2's memory requirements by
eliminating the 64-bit version number fields in lieu of
a couple bits.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The functions in lm.c were just wrappers which were mostly
only used in one other file. By moving the functions to
the files where they are being used, they can be marked
static and also this will usually result in them being inlined
since they are often only used from one point in the code.
A couple of really trivial functions have been inlined by hand
into the function which called them as it makes the code clearer
to do that.
We also gain from one fewer function call in the glock lock and
unlock paths.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch further reduces the memory needs of GFS2 by
eliminating the gl_req_bh variable from struct gfs2_glock.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch reduces memory by replacing the int variable
gl_waiters2 by a single bit in the gl_flags.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch reduces the memory required by GFS2 by combining
the rd_flags and rg_flags (in core only).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is performance related. When we're doing a log flush,
I noticed we were calling buf_lo_incore_commit twice: once for
data bufs and once for metadata bufs. Since this is the same
function and does the same thing in both cases, there should be
no reason to call it twice. Since we only need to call it once,
we can also make it faster by removing it from the generic "lops"
code and making it a stand-along static function.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch improves the calculation of the tree height in order to reduce
the number of operations which are carried out on each call to gfs2_block_map.
In the common case, we now make a single comparison, rather than calculating
the required tree height from scratch each time. Also in the case that the
tree does need some extra height, we start from the current height rather from
zero when we work out what the new height ought to be.
In addition the di_height field is moved into the inode proper and reduced
in size to a u8 since the value must be between 0 and GFS2_MAX_META_HEIGHT (10).
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The gl_owner_pid field is used to get the lock owning task by its pid, so make
it in a proper manner, i.e. by using the struct pid pointer and pid_task()
function.
The pid_task() becomes exported for the gfs2 module.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The gl_owner_pid field is used to get the holder task by its pid and check
whether the current is a holder, so make it in a proper manner, i.e. via the
struct pid * manipulations.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch removes a vestigial variable "i_spin" from the gfs2_inode
structure. This not only saves us memory (>300000 of these in memory
for the oom test) it also saves us time because we don't have to
spend time initializing it (i.e. slightly better performance).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
It is possible to reduce the size of GFS2 inodes by taking the i_alloc
structure out of the gfs2_inode. This patch allocates the i_alloc
structure whenever its needed, and frees it afterward. This decreases
the amount of low memory we use at the expense of requiring a memory
allocation for each page or partial page that we write. A quick test
with postmark shows that the overhead is not measurable and I also note
that OCFS2 use the same approach.
In the future I'd like to solve the problem by shrinking down the size
of the members of the i_alloc structure, but for now, this reduces the
immediate problem of using too much low-memory on x86 and doesn't add
too much overhead.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch eliminates the unneeded sd_statfs_mutex mutex but preserves
the ordering as discussed.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch saves a little time when gfs2 writes to the journals by
keeping a mapping between logical and physical blocks on disk.
That's better than constantly looking up indirect pointers in
buffers, when the journals are several levels of indirection
(which they typically are).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We only care about the content of the jindex in two cases,
one is when we mount the fs and the other is when we need
to recover another journal. In both cases we have to update
the jindex anyway, so there is no point in updating it
periodically between times, so this removes it to simplify
gfs2_logd.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch changes the counter which keeps track of the free
blocks in the journal to an atomic_t in preparation for the
following patch which will update the log reservation code.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The only reason for adding glocks to the journal was to keep track
of which locks required a log flush prior to release. We add a
flag to the glock to allow this check to be made in a simpler way.
This reduces the size of a glock (by 12 bytes on i386, 24 on x86_64)
and means that we can avoid extra work during the journal flush.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The HIF_MUTEX and HIF_PROMOTE flags were set on the glock holders
depending upon which of the two waiters lists they were going to
be queued upon. They were then tested when the holders were taken
off the lists to ensure that the right type of holder was being
dequeued.
Since we are already using separate lists, there doesn't seem a
lot of point having these flags as well, and since setting them
and testing them is in the fast path for locking and unlocking
glock, this patch removes them.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This call to reclaim glocks is not needed, and in particular we don't want it
in the fast path for locking glocks. The limit was entirely arbitrary anyway
and we can't expect users to adjust things like this, the remaining code will
do the right thing on its own.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This adds a function "gfs2_is_writeback()" along the lines of the
existing "gfs2_is_jdata()" in order to clean up the code and make
the various tests for the inode mode more obvious. It also fixes
the PageChecked() logic where we were resetting the flag too early
in the case of an error path.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The i_cache was designed to keep references to the indirect blocks
used during block mapping so that they didn't have to be looked
up continually. The idea failed because there are too many places
where the i_cache needs to be freed, and this has in the past been
the cause of many bugs.
In addition there was no performance benefit being gained since the
disk blocks in question were cached anyway. So this patch removes
it in order to simplify the code to prepare for other changes which
would otherwise have had to add further support for this feature.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This cleans up the mmap() code path for GFS2 by implementing the
page_mkwrite function for GFS2. We are thus able to use the
generic filemap_fault function for our ->fault() implementation.
This now means that shared writable mappings will be much more
efficiently shared across the cluster if there is a reasonable
proportion of read activity (the greater proportion, the better).
As a side effect, it also reduces the size of the code, removes
special cases from readpage and readpages, and makes the code
path easier to follow.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Fix a race condition where multiple glock demote requests are sent to
a node back-to-back. This patch does a check inside handle_callback()
to see whether a demote request is in progress. If true, it sets a flag
to make sure run_queue() will loop again to handle the new request,
instead of erronously setting gl_demote_state to a different state.
Signed-off-by: S. Wendy Cheng <wcheng@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch cleans up the code for writing journaled data into the log.
It also removes the need to allocate a small "tag" structure for each
block written into the log. Instead we just keep count of the outstanding
I/O so that we can be sure that its all been written at the correct time.
Another result of this patch is that a number of ll_rw_block() calls
have become submit_bh() calls, closing some races at the same time.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Both the revoke structure and the bufdata structure are quite similar.
They are basically small tags which are put on lists. In addition to
which the revoke structure is always allocated when there is a bufdata
structure which is (or can be) freed. As such it should be possible to
reduce the number of frees and allocations by using the same structure
for both purposes.
This patch is the first step along that path. It replaces existing uses
of the revoke structure with the bufdata structure.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The following patch removes the ordered write processing from
databuf_lo_before_commit() and moves it to log.c. This has the effect of
greatly simplyfying databuf_lo_before_commit() and well as potentially
making the ordered write code more efficient.
As a side effect of this, its now possible to remove ordered buffers
from the ordered buffer list at any time, so we now make use of this in
invalidatepage and releasepage to ensure timely release of these
buffers.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We only need a single gfs2_scand process rather than the one
per filesystem which we had previously. As a result the parameter
determining the frequency of gfs2_scand runs becomes a module
parameter rather than a mount parameter as it was before.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch passes all my nasty tests that were causing the code to
fail under one circumstance or another. Here is a complete summary
of all changes from today's git tree, in order of appearance:
1. There are now separate variables for metadata buffer accounting.
2. Variable sd_log_num_hdrs is no longer needed, since the header
accounting is taken care of by the reserve/refund sequence.
3. Fixed a tiny grammatical problem in a comment.
4. Added a new function "calc_reserved" to calculate the reserved
log space. This isn't entirely necessary, but it has two benefits:
First, it simplifies the gfs2_log_refund function greatly.
Second, it allows for easier debugging because I could sprinkle the
code with calls to this function to make sure the accounting is
proper (by adding asserts and printks) at strategic point of the code.
5. In log_pull_tail there apparently was a kludge to fix up the
accounting based on a "pull" parameter. The buffer accounting is
now done properly, so the kludge was removed.
6. File sync operations were making a call to gfs2_log_flush that
writes another journal header. Since that header was unplanned
for (reserved) by the reserve/refund sequence, the free space had
to be decremented so that when log_pull_tail gets called, the free
space is be adjusted properly. (Did I hear you call that a kludge?
well, maybe, but a lot more justifiable than the one I removed).
7. In the gfs2_log_shutdown code, it optionally syncs the log by
specifying the PULL parameter to log_write_header. I'm not sure
this is necessary anymore. It just seems to me there could be
cases where shutdown is called while there are outstanding log
buffers.
8. In the (data)buf_lo_before_commit functions, I changed some offset
values from being calculated on the fly to being constants. That
simplified some code and we might as well let the compiler do the
calculation once rather than redoing those cycles at run time.
9. This version has my rewritten databuf_lo_add function.
This version is much more like its predecessor, buf_lo_add, which
makes it easier to understand. Again, this might not be necessary,
but it seems as if this one works as well as the previous one,
maybe even better, so I decided to leave it in.
10. In databuf_lo_before_commit, a previous data corruption problem
was caused by going off the end of the buffer. The proper solution
is to have the proper limit in place, rather than stopping earlier.
(Thus my previous attempt to fix it is wrong).
If you don't wrap the buffer, you're stopping too early and that
causes more log buffer accounting problems.
11. In lops.h there are two new (previously mentioned) constants for
figuring out the data offset for the journal buffers.
12. There are also two new functions, buf_limit and databuf_limit to
calculate how many entries will fit in the buffer.
13. In function gfs2_meta_wipe, it needs to distinguish between pinned
metadata buffers and journaled data buffers for proper journal buffer
accounting. It can't use the JDATA gfs2_inode flag because it's
sometimes passed the "real" inode and sometimes the "metadata
inode" and the inode flags will be random bits in a metadata
gfs2_inode. It needs to base its decision on which was passed in.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Under certain circumstances its possible (though rather unlikely) that
inodes which were unlinked by one node while still open on another might
get "lost" in the sense that they don't get deallocated if the node
which held the inode open crashed before it was unlinked.
This patch adds the recovery code which allows automatic deallocation of
the inode if its found during block allocation (the sensible time to
look for such inodes since we are scanning the rgrp's bitmaps anyway at
this time, so it adds no overhead to do this).
Since the inode will have had its i_nlink set to zero, all we need to
trigger recovery is a lookup and an iput(), and the normal deallocation
code takes care of the rest.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes some sign issues which were accidentally introduced
into the quota & statfs code during the endianess annotation process.
Also included is a general clean up which moves all of the _host
structures out of gfs2_ondisk.h (where they should not have been to
start with) and into the places where they are actually used (often only
one place). Also those _host structures which are not required any more
are removed entirely (which is the eventual plan for all of them).
The conversion routines from ondisk.c are also moved into the places
where they are actually used, which for almost every one, was just one
single place, so all those are now static functions. This also cleans up
the end of gfs2_ondisk.h which no longer needs the #ifdef __KERNEL__.
The net result is a reduction of about 100 lines of code, many functions
now marked static plus the bug fixes as mentioned above. For good
measure I ran the code through sparse after making these changes to
check that there are no warnings generated.
This fixes Red Hat bz #239686
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes an error in the quota code where a 'struct
gfs2_quota_lvb*' was being passed to gfs2_adjust_quota() instead of a
'struct gfs2_quota_data*'. Also moved 'struct gfs2_quota_lvb' from
fs/gfs2/incore.h to include/linux/gfs2_ondisk.h as per Steve's suggestion.
Signed-off-by: Abhijith Das <adas@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch cleans up the inode number handling code. The main difference
is that instead of looking up the inodes using a struct gfs2_inum_host
we now use just the no_addr member of this structure. The tests relating
to no_formal_ino can then be done by the calling code. This has
advantages in that we want to do different things in different code
paths if the no_formal_ino doesn't match. In the NFS patch we want to
return -ESTALE, but in the ->lookup() path, its a bug in the fs if the
no_formal_ino doesn't match and thus we can withdraw in this case.
In order to later fix bz #201012, we need to be able to look up an inode
without knowing no_formal_ino, as the only information that is known to
us is the on-disk location of the inode in question.
This patch will also help us to fix bz #236099 at a later date by
cleaning up a lot of the code in that area.
There are no user visible changes as a result of this patch and there
are no changes to the on-disk format either.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The patch below consists of the following changes (in code order):
1. I fixed a minor compiler warning regarding the printing of
a kernel symbol address.
2. I implemented a suggestion from Dave Teigland that moves
the debugfs information for gfs2 into a subdirectory so
we can easily expand our use of debugfs in the future.
The current code keeps the glock information in:
/debug/gfs2/<fs>
With the patch, the new code keeps the glock information in:
/debug/gfs2/<fs>/glock
That will allow us to create more debugfs files in the future.
3. This fixes a bug whereby a failed mount attempt causes the
debugfs file to not be deleted. Failed mount attempts should
always clean up after themselves, including deleting the
debugfs file and/or directory.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
In Testing the previously posted and accepted patch for
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=228540
I uncovered some gfs2 badness. It turns out that the current
gfs2 code saves off a process pointer when glocks is taken
in both the glock and glock holder structures. Those
structures will persist in memory long after the process has
ended; pointers to poisoned memory.
This problem isn't caused by the 228540 fix; the new capability
introduced by the fix just uncovered the problem.
I wrote this patch that avoids saving process pointers
and instead saves off the process pid. Rather than
referencing the bad pointers, it now does process lookups.
There is special code that makes the output nicer for
printing holder information for processes that have ended.
This patch also adds a stub for the new "sprint_symbol"
function that exists in Andrew Morton's -mm patch set, but
won't go into the base kernel until 2.6.22, since it adds
functionality but doesn't fix a bug.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch prevents the printing of a warning message in cases where
the fs is functioning normally by handing off responsibility for
unlinked, but still open inodes, to another node for eventual deallocation.
Also, there is now an improved system for ensuring that such requests
to other nodes do not get lost. The callback on the iopen lock is
only ever called when i_nlink == 0 and when a node is unable to deallocate
it due to it still being in use on another node. When a node receives
the callback therefore, it knows that i_nlink must be zero, so we mark
it as such (in gfs2_drop_inode) in order that it will then attempt
deallocation of the inode itself.
As an additional benefit, queuing a demote request no longer requires
a memory allocation. This simplifies the code for dealing with gfs2_holders
as it removes one special case.
There are two new fields in struct gfs2_glock. gl_demote_state is the
state which the remote node has requested and gl_demote_time is the
time when the request came in. Both fields are only valid when the
GLF_DEMOTE flag is set in gl_flags.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The attached patch resolves bz 228540. This adds the capability
for gfs2 to dump gfs2 locks through the debugfs file system.
This used to exist in gfs1 as "gfs_tool lockdump" but it's missing from
gfs2 because all the ioctls were stripped out. Please see the bugzilla
for more history about the fix. This patch is also attached to the bugzilla
record.
The patch is against Steve Whitehouse's latest nmw git tree kernel
(2.6.21-rc1) and has been tested on system trin-10.
Signed-off-by: Robert Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch doesn't make any changes to the ordering of the various
operations related to glocking, but it does tidy up the calls to the
glops.c functions to make the structure more obvious.
The two functions: gfs2_glock_xmote_th() and gfs2_glock_drop_th() can be
made static within glock.c since they are called by every set of glock
operations. The xmote_th and drop_th glock operations are then made
conditional upon those two routines existing and called from the
previously mentioned functions in glock.c respectively.
Also it can be seen that the go_sync operation isn't needed since it can
easily be replaced by calls to xmote_bh and drop_bh respectively. This
results in no longer (confusingly) calling back into routines in glock.c
from glops.c and also reducing the glock operations by one member.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The "greedy" code was an attempt to retain glocks for a minimum length
of time when they relate to mmap()ed files. The current implementation
of this feature is not, however, ideal in that it required allocating
memory in order to do this and its overly complicated.
It also misses the mark by ignoring the other I/O operations which are
just as likely to suffer from the same problem. So the plan is to remove
this now and then add the functionality back as part of the glock state
machine at a later date (and thus take into account all the possible
users of this feature)
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Here is something I spotted (while looking for something entirely
different) the other day.
Rather than using a completion in each and every struct gfs2_holder,
this removes it in favour of hashed wait queues, thus saving a
considerable amount of memory both on the stack (where a number of
gfs2_holder structures are allocated) and in particular in the
gfs2_inode which has 8 gfs2_holder structures embedded within it.
As a result on x86_64 the gfs2_inode shrinks from 2488 bytes to
1912 bytes, a saving of 576 bytes per inode (no thats not a typo!).
In actual practice we get a much better result than that since
now that a gfs2_inode is under the 2048 byte barrier, we get two
per 4k slab page effectively halving the amount of memory required
to store gfs2_inodes.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This removes the extra filldir callback which gfs2 was using to
enclose an attempt at readahead for inodes during readdir. The
code was too complicated and also hurts performance badly in the
case that the getdents64/readdir call isn't being followed by
stat() and it wasn't even getting it right all the time when it
was.
As a result, on my test box an "ls" of a directory containing 250000
files fell from about 7mins (freshly mounted, so nothing cached) to
between about 15 to 25 seconds. When the directory content was cached,
the time taken fell from about 3mins to about 4 or 5 seconds.
Interestingly in the cached case, running "ls -l" once reduced the time
taken for subsequent runs of "ls" to about 6 secs even without this
patch. Now it turns out that there was a special case of glocks being
used for prefetching the metadata, but because of the timeouts for these
locks (set to 10 secs) the metadata was being timed out before it was
being used and this the prefetch code was constantly trying to prefetch
the same data over and over.
Calling "ls -l" meant that the inodes were brought into memory and once
the inodes are cached, the glocks are not disposed of until the inodes
are pushed out of the cache, thus extending the lifetime of the glocks,
and thus bringing down the time for subsequent runs of "ls"
considerably.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
