linux/drivers/md/dm-zoned-reclaim.c

569 lines
13 KiB
C
Raw Normal View History

dm zoned: drive-managed zoned block device target The dm-zoned device mapper target provides transparent write access to zoned block devices (ZBC and ZAC compliant block devices). dm-zoned hides to the device user (a file system or an application doing raw block device accesses) any constraint imposed on write requests by the device, equivalent to a drive-managed zoned block device model. Write requests are processed using a combination of on-disk buffering using the device conventional zones and direct in-place processing for requests aligned to a zone sequential write pointer position. A background reclaim process implemented using dm_kcopyd_copy ensures that conventional zones are always available for executing unaligned write requests. The reclaim process overhead is minimized by managing buffer zones in a least-recently-written order and first targeting the oldest buffer zones. Doing so, blocks under regular write access (such as metadata blocks of a file system) remain stored in conventional zones, resulting in no apparent overhead. dm-zoned implementation focus on simplicity and on minimizing overhead (CPU, memory and storage overhead). For a 14TB host-managed disk with 256 MB zones, dm-zoned memory usage per disk instance is at most about 3 MB and as little as 5 zones will be used internally for storing metadata and performing buffer zone reclaim operations. This is achieved using zone level indirection rather than a full block indirection system for managing block movement between zones. dm-zoned primary target is host-managed zoned block devices but it can also be used with host-aware device models to mitigate potential device-side performance degradation due to excessive random writing. Zoned block devices can be formatted and checked for use with the dm-zoned target using the dmzadm utility available at: https://github.com/hgst/dm-zoned-tools Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com> [Mike Snitzer partly refactored Damien's original work to cleanup the code] Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2017-06-07 06:55:39 +00:00
/*
* Copyright (C) 2017 Western Digital Corporation or its affiliates.
*
* This file is released under the GPL.
*/
#include "dm-zoned.h"
#include <linux/module.h>
#define DM_MSG_PREFIX "zoned reclaim"
struct dmz_reclaim {
struct dmz_metadata *metadata;
struct dmz_dev *dev;
struct delayed_work work;
struct workqueue_struct *wq;
struct dm_kcopyd_client *kc;
struct dm_kcopyd_throttle kc_throttle;
int kc_err;
unsigned long flags;
/* Last target access time */
unsigned long atime;
};
/*
* Reclaim state flags.
*/
enum {
DMZ_RECLAIM_KCOPY,
};
/*
* Number of seconds of target BIO inactivity to consider the target idle.
*/
#define DMZ_IDLE_PERIOD (10UL * HZ)
/*
* Percentage of unmapped (free) random zones below which reclaim starts
* even if the target is busy.
*/
#define DMZ_RECLAIM_LOW_UNMAP_RND 30
/*
* Percentage of unmapped (free) random zones above which reclaim will
* stop if the target is busy.
*/
#define DMZ_RECLAIM_HIGH_UNMAP_RND 50
/*
* Align a sequential zone write pointer to chunk_block.
*/
static int dmz_reclaim_align_wp(struct dmz_reclaim *zrc, struct dm_zone *zone,
sector_t block)
{
struct dmz_metadata *zmd = zrc->metadata;
sector_t wp_block = zone->wp_block;
unsigned int nr_blocks;
int ret;
if (wp_block == block)
return 0;
if (wp_block > block)
return -EIO;
/*
* Zeroout the space between the write
* pointer and the requested position.
*/
nr_blocks = block - wp_block;
ret = blkdev_issue_zeroout(zrc->dev->bdev,
dmz_start_sect(zmd, zone) + dmz_blk2sect(wp_block),
dmz_blk2sect(nr_blocks), GFP_NOIO, 0);
dm zoned: drive-managed zoned block device target The dm-zoned device mapper target provides transparent write access to zoned block devices (ZBC and ZAC compliant block devices). dm-zoned hides to the device user (a file system or an application doing raw block device accesses) any constraint imposed on write requests by the device, equivalent to a drive-managed zoned block device model. Write requests are processed using a combination of on-disk buffering using the device conventional zones and direct in-place processing for requests aligned to a zone sequential write pointer position. A background reclaim process implemented using dm_kcopyd_copy ensures that conventional zones are always available for executing unaligned write requests. The reclaim process overhead is minimized by managing buffer zones in a least-recently-written order and first targeting the oldest buffer zones. Doing so, blocks under regular write access (such as metadata blocks of a file system) remain stored in conventional zones, resulting in no apparent overhead. dm-zoned implementation focus on simplicity and on minimizing overhead (CPU, memory and storage overhead). For a 14TB host-managed disk with 256 MB zones, dm-zoned memory usage per disk instance is at most about 3 MB and as little as 5 zones will be used internally for storing metadata and performing buffer zone reclaim operations. This is achieved using zone level indirection rather than a full block indirection system for managing block movement between zones. dm-zoned primary target is host-managed zoned block devices but it can also be used with host-aware device models to mitigate potential device-side performance degradation due to excessive random writing. Zoned block devices can be formatted and checked for use with the dm-zoned target using the dmzadm utility available at: https://github.com/hgst/dm-zoned-tools Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com> [Mike Snitzer partly refactored Damien's original work to cleanup the code] Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2017-06-07 06:55:39 +00:00
if (ret) {
dmz_dev_err(zrc->dev,
"Align zone %u wp %llu to %llu (wp+%u) blocks failed %d",
dmz_id(zmd, zone), (unsigned long long)wp_block,
(unsigned long long)block, nr_blocks, ret);
return ret;
}
zone->wp_block = block;
return 0;
}
/*
* dm_kcopyd_copy end notification.
*/
static void dmz_reclaim_kcopy_end(int read_err, unsigned long write_err,
void *context)
{
struct dmz_reclaim *zrc = context;
if (read_err || write_err)
zrc->kc_err = -EIO;
else
zrc->kc_err = 0;
clear_bit_unlock(DMZ_RECLAIM_KCOPY, &zrc->flags);
smp_mb__after_atomic();
wake_up_bit(&zrc->flags, DMZ_RECLAIM_KCOPY);
}
/*
* Copy valid blocks of src_zone into dst_zone.
*/
static int dmz_reclaim_copy(struct dmz_reclaim *zrc,
struct dm_zone *src_zone, struct dm_zone *dst_zone)
{
struct dmz_metadata *zmd = zrc->metadata;
struct dmz_dev *dev = zrc->dev;
struct dm_io_region src, dst;
sector_t block = 0, end_block;
sector_t nr_blocks;
sector_t src_zone_block;
sector_t dst_zone_block;
unsigned long flags = 0;
int ret;
if (dmz_is_seq(src_zone))
end_block = src_zone->wp_block;
else
end_block = dev->zone_nr_blocks;
src_zone_block = dmz_start_block(zmd, src_zone);
dst_zone_block = dmz_start_block(zmd, dst_zone);
if (dmz_is_seq(dst_zone))
set_bit(DM_KCOPYD_WRITE_SEQ, &flags);
while (block < end_block) {
/* Get a valid region from the source zone */
ret = dmz_first_valid_block(zmd, src_zone, &block);
if (ret <= 0)
return ret;
nr_blocks = ret;
/*
* If we are writing in a sequential zone, we must make sure
* that writes are sequential. So Zeroout any eventual hole
* between writes.
*/
if (dmz_is_seq(dst_zone)) {
ret = dmz_reclaim_align_wp(zrc, dst_zone, block);
if (ret)
return ret;
}
src.bdev = dev->bdev;
src.sector = dmz_blk2sect(src_zone_block + block);
src.count = dmz_blk2sect(nr_blocks);
dst.bdev = dev->bdev;
dst.sector = dmz_blk2sect(dst_zone_block + block);
dst.count = src.count;
/* Copy the valid region */
set_bit(DMZ_RECLAIM_KCOPY, &zrc->flags);
dm_kcopyd_copy(zrc->kc, &src, 1, &dst, flags,
dmz_reclaim_kcopy_end, zrc);
dm zoned: drive-managed zoned block device target The dm-zoned device mapper target provides transparent write access to zoned block devices (ZBC and ZAC compliant block devices). dm-zoned hides to the device user (a file system or an application doing raw block device accesses) any constraint imposed on write requests by the device, equivalent to a drive-managed zoned block device model. Write requests are processed using a combination of on-disk buffering using the device conventional zones and direct in-place processing for requests aligned to a zone sequential write pointer position. A background reclaim process implemented using dm_kcopyd_copy ensures that conventional zones are always available for executing unaligned write requests. The reclaim process overhead is minimized by managing buffer zones in a least-recently-written order and first targeting the oldest buffer zones. Doing so, blocks under regular write access (such as metadata blocks of a file system) remain stored in conventional zones, resulting in no apparent overhead. dm-zoned implementation focus on simplicity and on minimizing overhead (CPU, memory and storage overhead). For a 14TB host-managed disk with 256 MB zones, dm-zoned memory usage per disk instance is at most about 3 MB and as little as 5 zones will be used internally for storing metadata and performing buffer zone reclaim operations. This is achieved using zone level indirection rather than a full block indirection system for managing block movement between zones. dm-zoned primary target is host-managed zoned block devices but it can also be used with host-aware device models to mitigate potential device-side performance degradation due to excessive random writing. Zoned block devices can be formatted and checked for use with the dm-zoned target using the dmzadm utility available at: https://github.com/hgst/dm-zoned-tools Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com> [Mike Snitzer partly refactored Damien's original work to cleanup the code] Signed-off-by: Mike Snitzer <snitzer@redhat.com>
2017-06-07 06:55:39 +00:00
/* Wait for copy to complete */
wait_on_bit_io(&zrc->flags, DMZ_RECLAIM_KCOPY,
TASK_UNINTERRUPTIBLE);
if (zrc->kc_err)
return zrc->kc_err;
block += nr_blocks;
if (dmz_is_seq(dst_zone))
dst_zone->wp_block = block;
}
return 0;
}
/*
* Move valid blocks of dzone buffer zone into dzone (after its write pointer)
* and free the buffer zone.
*/
static int dmz_reclaim_buf(struct dmz_reclaim *zrc, struct dm_zone *dzone)
{
struct dm_zone *bzone = dzone->bzone;
sector_t chunk_block = dzone->wp_block;
struct dmz_metadata *zmd = zrc->metadata;
int ret;
dmz_dev_debug(zrc->dev,
"Chunk %u, move buf zone %u (weight %u) to data zone %u (weight %u)",
dzone->chunk, dmz_id(zmd, bzone), dmz_weight(bzone),
dmz_id(zmd, dzone), dmz_weight(dzone));
/* Flush data zone into the buffer zone */
ret = dmz_reclaim_copy(zrc, bzone, dzone);
if (ret < 0)
return ret;
dmz_lock_flush(zmd);
/* Validate copied blocks */
ret = dmz_merge_valid_blocks(zmd, bzone, dzone, chunk_block);
if (ret == 0) {
/* Free the buffer zone */
dmz_invalidate_blocks(zmd, bzone, 0, zrc->dev->zone_nr_blocks);
dmz_lock_map(zmd);
dmz_unmap_zone(zmd, bzone);
dmz_unlock_zone_reclaim(dzone);
dmz_free_zone(zmd, bzone);
dmz_unlock_map(zmd);
}
dmz_unlock_flush(zmd);
return 0;
}
/*
* Merge valid blocks of dzone into its buffer zone and free dzone.
*/
static int dmz_reclaim_seq_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
{
unsigned int chunk = dzone->chunk;
struct dm_zone *bzone = dzone->bzone;
struct dmz_metadata *zmd = zrc->metadata;
int ret = 0;
dmz_dev_debug(zrc->dev,
"Chunk %u, move data zone %u (weight %u) to buf zone %u (weight %u)",
chunk, dmz_id(zmd, dzone), dmz_weight(dzone),
dmz_id(zmd, bzone), dmz_weight(bzone));
/* Flush data zone into the buffer zone */
ret = dmz_reclaim_copy(zrc, dzone, bzone);
if (ret < 0)
return ret;
dmz_lock_flush(zmd);
/* Validate copied blocks */
ret = dmz_merge_valid_blocks(zmd, dzone, bzone, 0);
if (ret == 0) {
/*
* Free the data zone and remap the chunk to
* the buffer zone.
*/
dmz_invalidate_blocks(zmd, dzone, 0, zrc->dev->zone_nr_blocks);
dmz_lock_map(zmd);
dmz_unmap_zone(zmd, bzone);
dmz_unmap_zone(zmd, dzone);
dmz_unlock_zone_reclaim(dzone);
dmz_free_zone(zmd, dzone);
dmz_map_zone(zmd, bzone, chunk);
dmz_unlock_map(zmd);
}
dmz_unlock_flush(zmd);
return 0;
}
/*
* Move valid blocks of the random data zone dzone into a free sequential zone.
* Once blocks are moved, remap the zone chunk to the sequential zone.
*/
static int dmz_reclaim_rnd_data(struct dmz_reclaim *zrc, struct dm_zone *dzone)
{
unsigned int chunk = dzone->chunk;
struct dm_zone *szone = NULL;
struct dmz_metadata *zmd = zrc->metadata;
int ret;
/* Get a free sequential zone */
dmz_lock_map(zmd);
szone = dmz_alloc_zone(zmd, DMZ_ALLOC_RECLAIM);
dmz_unlock_map(zmd);
if (!szone)
return -ENOSPC;
dmz_dev_debug(zrc->dev,
"Chunk %u, move rnd zone %u (weight %u) to seq zone %u",
chunk, dmz_id(zmd, dzone), dmz_weight(dzone),
dmz_id(zmd, szone));
/* Flush the random data zone into the sequential zone */
ret = dmz_reclaim_copy(zrc, dzone, szone);
dmz_lock_flush(zmd);
if (ret == 0) {
/* Validate copied blocks */
ret = dmz_copy_valid_blocks(zmd, dzone, szone);
}
if (ret) {
/* Free the sequential zone */
dmz_lock_map(zmd);
dmz_free_zone(zmd, szone);
dmz_unlock_map(zmd);
} else {
/* Free the data zone and remap the chunk */
dmz_invalidate_blocks(zmd, dzone, 0, zrc->dev->zone_nr_blocks);
dmz_lock_map(zmd);
dmz_unmap_zone(zmd, dzone);
dmz_unlock_zone_reclaim(dzone);
dmz_free_zone(zmd, dzone);
dmz_map_zone(zmd, szone, chunk);
dmz_unlock_map(zmd);
}
dmz_unlock_flush(zmd);
return 0;
}
/*
* Reclaim an empty zone.
*/
static void dmz_reclaim_empty(struct dmz_reclaim *zrc, struct dm_zone *dzone)
{
struct dmz_metadata *zmd = zrc->metadata;
dmz_lock_flush(zmd);
dmz_lock_map(zmd);
dmz_unmap_zone(zmd, dzone);
dmz_unlock_zone_reclaim(dzone);
dmz_free_zone(zmd, dzone);
dmz_unlock_map(zmd);
dmz_unlock_flush(zmd);
}
/*
* Find a candidate zone for reclaim and process it.
*/
static void dmz_reclaim(struct dmz_reclaim *zrc)
{
struct dmz_metadata *zmd = zrc->metadata;
struct dm_zone *dzone;
struct dm_zone *rzone;
unsigned long start;
int ret;
/* Get a data zone */
dzone = dmz_get_zone_for_reclaim(zmd);
if (!dzone)
return;
start = jiffies;
if (dmz_is_rnd(dzone)) {
if (!dmz_weight(dzone)) {
/* Empty zone */
dmz_reclaim_empty(zrc, dzone);
ret = 0;
} else {
/*
* Reclaim the random data zone by moving its
* valid data blocks to a free sequential zone.
*/
ret = dmz_reclaim_rnd_data(zrc, dzone);
}
rzone = dzone;
} else {
struct dm_zone *bzone = dzone->bzone;
sector_t chunk_block = 0;
ret = dmz_first_valid_block(zmd, bzone, &chunk_block);
if (ret < 0)
goto out;
if (ret == 0 || chunk_block >= dzone->wp_block) {
/*
* The buffer zone is empty or its valid blocks are
* after the data zone write pointer.
*/
ret = dmz_reclaim_buf(zrc, dzone);
rzone = bzone;
} else {
/*
* Reclaim the data zone by merging it into the
* buffer zone so that the buffer zone itself can
* be later reclaimed.
*/
ret = dmz_reclaim_seq_data(zrc, dzone);
rzone = dzone;
}
}
out:
if (ret) {
dmz_unlock_zone_reclaim(dzone);
return;
}
(void) dmz_flush_metadata(zrc->metadata);
dmz_dev_debug(zrc->dev, "Reclaimed zone %u in %u ms",
dmz_id(zmd, rzone), jiffies_to_msecs(jiffies - start));
}
/*
* Test if the target device is idle.
*/
static inline int dmz_target_idle(struct dmz_reclaim *zrc)
{
return time_is_before_jiffies(zrc->atime + DMZ_IDLE_PERIOD);
}
/*
* Test if reclaim is necessary.
*/
static bool dmz_should_reclaim(struct dmz_reclaim *zrc)
{
struct dmz_metadata *zmd = zrc->metadata;
unsigned int nr_rnd = dmz_nr_rnd_zones(zmd);
unsigned int nr_unmap_rnd = dmz_nr_unmap_rnd_zones(zmd);
unsigned int p_unmap_rnd = nr_unmap_rnd * 100 / nr_rnd;
/* Reclaim when idle */
if (dmz_target_idle(zrc) && nr_unmap_rnd < nr_rnd)
return true;
/* If there are still plenty of random zones, do not reclaim */
if (p_unmap_rnd >= DMZ_RECLAIM_HIGH_UNMAP_RND)
return false;
/*
* If the percentage of unmappped random zones is low,
* reclaim even if the target is busy.
*/
return p_unmap_rnd <= DMZ_RECLAIM_LOW_UNMAP_RND;
}
/*
* Reclaim work function.
*/
static void dmz_reclaim_work(struct work_struct *work)
{
struct dmz_reclaim *zrc = container_of(work, struct dmz_reclaim, work.work);
struct dmz_metadata *zmd = zrc->metadata;
unsigned int nr_rnd, nr_unmap_rnd;
unsigned int p_unmap_rnd;
if (!dmz_should_reclaim(zrc)) {
mod_delayed_work(zrc->wq, &zrc->work, DMZ_IDLE_PERIOD);
return;
}
/*
* We need to start reclaiming random zones: set up zone copy
* throttling to either go fast if we are very low on random zones
* and slower if there are still some free random zones to avoid
* as much as possible to negatively impact the user workload.
*/
nr_rnd = dmz_nr_rnd_zones(zmd);
nr_unmap_rnd = dmz_nr_unmap_rnd_zones(zmd);
p_unmap_rnd = nr_unmap_rnd * 100 / nr_rnd;
if (dmz_target_idle(zrc) || p_unmap_rnd < DMZ_RECLAIM_LOW_UNMAP_RND / 2) {
/* Idle or very low percentage: go fast */
zrc->kc_throttle.throttle = 100;
} else {
/* Busy but we still have some random zone: throttle */
zrc->kc_throttle.throttle = min(75U, 100U - p_unmap_rnd / 2);
}
dmz_dev_debug(zrc->dev,
"Reclaim (%u): %s, %u%% free rnd zones (%u/%u)",
zrc->kc_throttle.throttle,
(dmz_target_idle(zrc) ? "Idle" : "Busy"),
p_unmap_rnd, nr_unmap_rnd, nr_rnd);
dmz_reclaim(zrc);
dmz_schedule_reclaim(zrc);
}
/*
* Initialize reclaim.
*/
int dmz_ctr_reclaim(struct dmz_dev *dev, struct dmz_metadata *zmd,
struct dmz_reclaim **reclaim)
{
struct dmz_reclaim *zrc;
int ret;
zrc = kzalloc(sizeof(struct dmz_reclaim), GFP_KERNEL);
if (!zrc)
return -ENOMEM;
zrc->dev = dev;
zrc->metadata = zmd;
zrc->atime = jiffies;
/* Reclaim kcopyd client */
zrc->kc = dm_kcopyd_client_create(&zrc->kc_throttle);
if (IS_ERR(zrc->kc)) {
ret = PTR_ERR(zrc->kc);
zrc->kc = NULL;
goto err;
}
/* Reclaim work */
INIT_DELAYED_WORK(&zrc->work, dmz_reclaim_work);
zrc->wq = alloc_ordered_workqueue("dmz_rwq_%s", WQ_MEM_RECLAIM,
dev->name);
if (!zrc->wq) {
ret = -ENOMEM;
goto err;
}
*reclaim = zrc;
queue_delayed_work(zrc->wq, &zrc->work, 0);
return 0;
err:
if (zrc->kc)
dm_kcopyd_client_destroy(zrc->kc);
kfree(zrc);
return ret;
}
/*
* Terminate reclaim.
*/
void dmz_dtr_reclaim(struct dmz_reclaim *zrc)
{
cancel_delayed_work_sync(&zrc->work);
destroy_workqueue(zrc->wq);
dm_kcopyd_client_destroy(zrc->kc);
kfree(zrc);
}
/*
* Suspend reclaim.
*/
void dmz_suspend_reclaim(struct dmz_reclaim *zrc)
{
cancel_delayed_work_sync(&zrc->work);
}
/*
* Resume reclaim.
*/
void dmz_resume_reclaim(struct dmz_reclaim *zrc)
{
queue_delayed_work(zrc->wq, &zrc->work, DMZ_IDLE_PERIOD);
}
/*
* BIO accounting.
*/
void dmz_reclaim_bio_acc(struct dmz_reclaim *zrc)
{
zrc->atime = jiffies;
}
/*
* Start reclaim if necessary.
*/
void dmz_schedule_reclaim(struct dmz_reclaim *zrc)
{
if (dmz_should_reclaim(zrc))
mod_delayed_work(zrc->wq, &zrc->work, 0);
}