For blk-mq rotational zoned block devices (e.g. SMR HDDs), default to
having zone write plugging limit write operations to a maximum queue
depth of 1 for all zones. This significantly reduce write seek overhead
and improves SMR HDD write throughput.
For remotely connected disks with a very high network latency this
features might not be useful. However, remotely connected zoned devices
are rare at the moment, and we cannot know the round trip latency to
pick a good default for network attached devices. System administrators
can however disable this feature in that case.
For BIO based (non blk-mq) rotational zoned block devices, the device
driver (e.g. a DM target driver) can directly set an appropriate
default.
Signed-off-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In order to maintain sequential write patterns per zone with zoned block
devices, zone write plugging issues only a single write BIO per zone at
any time. This works well but has the side effect that when large
sequential write streams are issued by the user and these streams cross
zone boundaries, the device ends up receiving a discontiguous set of
write commands for different zones. The same also happens when a user
writes simultaneously at high queue depth multiple zones: the device
does not see all sequential writes per zone and receives discontiguous
writes to different zones. While this does not affect the performance of
solid state zoned block devices, when using an SMR HDD, this pattern
change from sequential writes to discontiguous writes to different zones
significantly increases head seek which results in degraded write
throughput.
In order to reduce this seek overhead for rotational media devices,
introduce a per disk zone write plugs kernel thread to issue all write
BIOs to zones. This single zone write issuing context is enabled for
any zoned block device that has a request queue flagged with the new
QUEUE_ZONED_QD1_WRITES flag.
The flag QUEUE_ZONED_QD1_WRITES is visible as the sysfs queue attribute
zoned_qd1_writes for zoned devices. For regular block devices, this
attribute is not visible. For zoned block devices, a user can override
the default value set to force the global write maximum queue depth of
1 for a zoned block device, or clear this attribute to fallback to the
default behavior of zone write plugging which limits writes to QD=1 per
sequential zone.
Writing to a zoned block device flagged with QUEUE_ZONED_QD1_WRITES is
implemented using a list of zone write plugs that have a non-empty BIO
list. Listed zone write plugs are processed by the disk zone write plugs
worker kthread in FIFO order, and all BIOs of a zone write plug are all
processed before switching to the next listed zone write plug. A newly
submitted BIO for a non-FULL zone write plug that is not yet listed
causes the addition of the zone write plug at the end of the disk list
of zone write plugs.
Since the write BIOs queued in a zone write plug BIO list are
necessarilly sequential, for rotational media, using the single zone
write plugs kthread to issue all BIOs maintains a sequential write
pattern and thus reduces seek overhead and improves write throughput.
This processing essentially result in always writing to HDDs at QD=1,
which is not an issue for HDDs operating with write caching enabled.
Performance with write cache disabled is also not degraded thanks to
the efficient write handling of modern SMR HDDs.
A disk list of zone write plugs is defined using the new struct gendisk
zone_wplugs_list, and accesses to this list is protected using the
zone_wplugs_list_lock spinlock. The per disk kthread
(zone_wplugs_worker) code is implemented by the function
disk_zone_wplugs_worker(). A reference on listed zone write plugs is
always held until all BIOs of the zone write plug are processed by the
worker kthread. BIO issuing at QD=1 is driven using a completion
structure (zone_wplugs_worker_bio_done) and calls to blk_io_wait().
With this change, performance when sequentially writing the zones of a
30 TB SMR SATA HDD connected to an AHCI adapter changes as follows
(1MiB direct I/Os, results in MB/s unit):
+--------------------+
| Write BW (MB/s) |
+------------------+----------+---------+
| Sequential write | Baseline | Patched |
| Queue Depth | 6.19-rc8 | |
+------------------+----------+---------+
| 1 | 244 | 245 |
| 2 | 244 | 245 |
| 4 | 245 | 245 |
| 8 | 242 | 245 |
| 16 | 222 | 246 |
| 32 | 211 | 245 |
| 64 | 193 | 244 |
| 128 | 112 | 246 |
+------------------+----------+---------+
With the current code (baseline), as the sequential write stream crosses
a zone boundary, higher queue depth creates a gap between the
last IO to the previous zone and the first IOs to the following zones,
causing head seeks and degrading performance. Using the disk zone
write plugs worker thread, this pattern disappears and the maximum
throughput of the drive is maintained, leading to over 100%
improvements in throughput for high queue depth write.
Using 16 fio jobs all writing to randomly chosen zones at QD=32 with 1
MiB direct IOs, write throughput also increases significantly.
+--------------------+
| Write BW (MB/s) |
+------------------+----------+---------+
| Random write | Baseline | Patched |
| Number of zones | 6.19-rc7 | |
+------------------+----------+---------+
| 1 | 191 | 192 |
| 2 | 101 | 128 |
| 4 | 115 | 123 |
| 8 | 90 | 120 |
| 16 | 64 | 115 |
| 32 | 58 | 105 |
| 64 | 56 | 101 |
| 128 | 55 | 99 |
+------------------+----------+---------+
Tests using XFS shows that buffered write speed with 8 jobs writing
files increases by 12% to 35% depending on the workload.
+--------------------+
| Write BW (MB/s) |
+------------------+----------+---------+
| Workload | Baseline | Patched |
| | 6.19-rc7 | |
+------------------+----------+---------+
| 256MiB file size | 212 | 238 |
+------------------+----------+---------+
| 4MiB .. 128 MiB | 213 | 243 |
| random file size | | |
+------------------+----------+---------+
| 2MiB .. 8 MiB | 179 | 242 |
| random file size | | |
+------------------+----------+---------+
Performance gains are even more significant when using an HBA that
limits the maximum size of commands to a small value, e.g. HBAs
controlled with the mpi3mr driver limit commands to a maximum of 1 MiB.
In such case, the write throughput gains are over 40%.
+--------------------+
| Write BW (MB/s) |
+------------------+----------+---------+
| Workload | Baseline | Patched |
| | 6.19-rc7 | |
+------------------+----------+---------+
| 256MiB file size | 175 | 245 |
+------------------+----------+---------+
| 4MiB .. 128 MiB | 174 | 244 |
| random file size | | |
+------------------+----------+---------+
| 2MiB .. 8 MiB | 171 | 243 |
| random file size | | |
+------------------+----------+---------+
Signed-off-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use trylock instead of blocking lock acquisition for update_nr_hwq_lock
in queue_requests_store() and elv_iosched_store() to avoid circular lock
dependency with kernfs active reference during concurrent disk deletion:
update_nr_hwq_lock -> kn->active (via del_gendisk -> kobject_del)
kn->active -> update_nr_hwq_lock (via sysfs write path)
Return -EBUSY when the lock is not immediately available.
Reported-and-tested-by: Yi Zhang <yi.zhang@redhat.com>
Closes: https://lore.kernel.org/linux-block/CAHj4cs-em-4acsHabMdT=jJhXkCzjnprD-aQH1OgrZo4nTnmMw@mail.gmail.com/
Fixes: 626ff4f8eb ("blk-mq: convert to serialize updating nr_requests with update_nr_hwq_lock")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Yi Zhang <yi.zhang@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Creating debugfs entries can trigger fs reclaim, which can enter back
into the block layer request_queue. This can cause deadlock if the
queue is frozen.
Previously, a WARN_ON_ONCE check was used in debugfs_create_files()
to detect this condition, but it was racy since the queue can be frozen
from another context at any time.
Introduce blk_debugfs_lock()/blk_debugfs_unlock() helpers that combine
the debugfs_mutex with memalloc_noio_save()/restore() to prevent fs
reclaim from triggering block I/O. Also add blk_debugfs_lock_nomemsave()
and blk_debugfs_unlock_nomemrestore() variants for callers that don't
need NOIO protection (e.g., debugfs removal or read-only operations).
Replace all raw debugfs_mutex lock/unlock pairs with these helpers,
using the _nomemsave/_nomemrestore variants where appropriate.
Reported-by: Yi Zhang <yi.zhang@redhat.com>
Closes: https://lore.kernel.org/all/CAHj4cs9gNKEYAPagD9JADfO5UH+OiCr4P7OO2wjpfOYeM-RV=A@mail.gmail.com/
Reported-by: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Closes: https://lore.kernel.org/all/aYWQR7CtYdk3K39g@shinmob/
Suggested-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Yu Kuai <yukuai@fnnas.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add a new field async_depth to request_queue and related APIs, this is
currently not used, following patches will convert elevators to use
this instead of internal async_depth.
Signed-off-by: Yu Kuai <yukuai@fnnas.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To move implementation details inside blk-wbt.c, prepare to fix possible
deadlock to call wbt_init() while queue is frozen in the next patch.
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Yu Kuai <yukuai@fnnas.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When wbt_enable_default() is moved out of queue freezing in elevator_change(),
it can cause the wbt inflight counter to become negative (-1), leading to hung
tasks in the writeback path. Tasks get stuck in wbt_wait() because the counter
is in an inconsistent state.
The issue occurs because wbt_enable_default() could race with IO submission,
allowing the counter to be decremented before proper initialization. This manifests
as:
rq_wait[0]:
inflight: -1
has_waiters: True
rwb_enabled() checks the state, which can be updated exactly between wbt_wait()
(rq_qos_throttle()) and wbt_track()(rq_qos_track()), then the inflight counter
will become negative.
And results in hung task warnings like:
task:kworker/u24:39 state:D stack:0 pid:14767
Call Trace:
rq_qos_wait+0xb4/0x150
wbt_wait+0xa9/0x100
__rq_qos_throttle+0x24/0x40
blk_mq_submit_bio+0x672/0x7b0
...
Fix this by:
1. Splitting wbt_enable_default() into:
- __wbt_enable_default(): Returns true if wbt_init() should be called
- wbt_enable_default(): Wrapper for existing callers (no init)
- wbt_init_enable_default(): New function that checks and inits WBT
2. Using wbt_init_enable_default() in blk_register_queue() to ensure
proper initialization during queue registration
3. Move wbt_init() out of wbt_enable_default() which is only for enabling
disabled wbt from bfq and iocost, and wbt_init() isn't needed. Then the
original lock warning can be avoided.
4. Removing the ELEVATOR_FLAG_ENABLE_WBT_ON_EXIT flag and its handling
code since it's no longer needed
This ensures WBT is properly initialized before any IO can be submitted,
preventing the counter from going negative.
Cc: Nilay Shroff <nilay@linux.ibm.com>
Cc: Yu Kuai <yukuai@fnnas.com>
Cc: Guangwu Zhang <guazhang@redhat.com>
Fixes: 78c271344b ("block: move wbt_enable_default() out of queue freezing from sched ->exit()")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Freezing the request queue from inside sysfs store callbacks may cause a
deadlock in combination with the dm-multipath driver and the
queue_if_no_path option. Additionally, freezing the request queue slows
down system boot on systems where sysfs attributes are set synchronously.
Fix this by removing the blk_mq_freeze_queue() / blk_mq_unfreeze_queue()
calls from the store callbacks that do not strictly need these callbacks.
Add the __data_racy annotation to request_queue.rq_timeout to suppress
KCSAN data race reports about the rq_timeout reads.
This patch may cause a small delay in applying the new settings.
For all the attributes affected by this patch, I/O will complete
correctly whether the old or the new value of the attribute is used.
This patch affects the following sysfs attributes:
* io_poll_delay
* io_timeout
* nomerges
* read_ahead_kb
* rq_affinity
Here is an example of a deadlock triggered by running test srp/002
if this patch is not applied:
task:multipathd
Call Trace:
<TASK>
__schedule+0x8c1/0x1bf0
schedule+0xdd/0x270
schedule_preempt_disabled+0x1c/0x30
__mutex_lock+0xb89/0x1650
mutex_lock_nested+0x1f/0x30
dm_table_set_restrictions+0x823/0xdf0
__bind+0x166/0x590
dm_swap_table+0x2a7/0x490
do_resume+0x1b1/0x610
dev_suspend+0x55/0x1a0
ctl_ioctl+0x3a5/0x7e0
dm_ctl_ioctl+0x12/0x20
__x64_sys_ioctl+0x127/0x1a0
x64_sys_call+0xe2b/0x17d0
do_syscall_64+0x96/0x3a0
entry_SYSCALL_64_after_hwframe+0x4b/0x53
</TASK>
task:(udev-worker)
Call Trace:
<TASK>
__schedule+0x8c1/0x1bf0
schedule+0xdd/0x270
blk_mq_freeze_queue_wait+0xf2/0x140
blk_mq_freeze_queue_nomemsave+0x23/0x30
queue_ra_store+0x14e/0x290
queue_attr_store+0x23e/0x2c0
sysfs_kf_write+0xde/0x140
kernfs_fop_write_iter+0x3b2/0x630
vfs_write+0x4fd/0x1390
ksys_write+0xfd/0x230
__x64_sys_write+0x76/0xc0
x64_sys_call+0x276/0x17d0
do_syscall_64+0x96/0x3a0
entry_SYSCALL_64_after_hwframe+0x4b/0x53
</TASK>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Nilay Shroff <nilay@linux.ibm.com>
Cc: Martin Wilck <mwilck@suse.com>
Cc: Benjamin Marzinski <bmarzins@redhat.com>
Cc: stable@vger.kernel.org
Fixes: af28141498 ("block: freeze the queue in queue_attr_store")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
request_queue->nr_requests can be changed by:
a) switch elevator by updating nr_hw_queues
b) switch elevator by elevator sysfs attribute
c) configue queue sysfs attribute nr_requests
Current lock order is:
1) update_nr_hwq_lock, case a,b
2) freeze_queue
3) elevator_lock, case a,b,c
And update nr_requests is seriablized by elevator_lock() already,
however, in the case c, we'll have to allocate new sched_tags if
nr_requests grow, and do this with elevator_lock held and queue
freezed has the risk of deadlock.
Hence use update_nr_hwq_lock instead, make it possible to allocate
memory if tags grow, meanwhile also prevent nr_requests to be changed
concurrently.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
queue_requests_store() is the only caller of
blk_mq_update_nr_requests(), and blk_mq_update_nr_requests() is the
only caller of blk_mq_tag_update_depth(), however, they all have
checkings for nr_requests input by user.
Make code cleaner by moving all the checkings to the top function:
1) nr_requests > reserved tags;
2) if there is elevator, 4 <= nr_requests <= 2048;
3) if elevator is none, 4 <= nr_requests <= tag_set->queue_depth;
Meanwhile, case 2 is the only case tags can grow and -ENOMEM might be
returned.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_queue_attr_visible() already checked queue_is_mq(), no need to
check this again in queue_requests_store().
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_quiesce_queue() does not wait for pending I/O to finish. Freezing
a queue waits for pending I/O to finish. Hence move the comment that
refers to waiting for pending I/O above the call that freezes the
request queue. This patch moves this comment back to the position where
it was when this comment was introduced. See also commit c125311d96
("blk-wbt: don't maintain inflight counts if disabled").
Cc: Christoph Hellwig <hch@lst.de>
Cc: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Link: https://lore.kernel.org/r/20250825151424.1653910-1-bvanassche@acm.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The commit 245618f8e4 ("block: protect wbt_lat_usec using
q->elevator_lock") protected wbt_enable_default() with
q->elevator_lock; however, it also placed wbt_enable_default()
before blk_queue_flag_set(QUEUE_FLAG_REGISTERED, q);, resulting
in wbt failing to be enabled.
Moreover, the protection of wbt_enable_default() by q->elevator_lock
was removed in commit 78c271344b ("block: move wbt_enable_default()
out of queue freezing from sched ->exit()"), so we can directly fix
this issue by placing wbt_enable_default() after
blk_queue_flag_set(QUEUE_FLAG_REGISTERED, q);.
Additionally, this issue also causes the inability to read the
wbt_lat_usec file, and the scenario is as follows:
root@q:/sys/block/sda/queue# cat wbt_lat_usec
cat: wbt_lat_usec: Invalid argument
root@q:/data00/sjc/linux# ls /sys/kernel/debug/block/sda/rqos
cannot access '/sys/kernel/debug/block/sda/rqos': No such file or directory
root@q:/data00/sjc/linux# find /sys -name wbt
/sys/kernel/debug/tracing/events/wbt
After testing with this patch, wbt can be enabled normally.
Signed-off-by: Julian Sun <sunjunchao@bytedance.com>
Cc: stable@vger.kernel.org
Fixes: 245618f8e4 ("block: protect wbt_lat_usec using q->elevator_lock")
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20250812154257.57540-1-sunjunchao@bytedance.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Device-mapper can call add_disk() multiple times for the same gendisk
due to its two-phase creation process (dm create + dm load). This leads
to kobject double initialization errors when the underlying iSCSI devices
become temporarily unavailable and then reappear.
However, if the first add_disk() call fails and is retried, the queue_kobj
gets initialized twice, causing:
kobject: kobject (ffff88810c27bb90): tried to init an initialized object,
something is seriously wrong.
Call Trace:
<TASK>
dump_stack_lvl+0x5b/0x80
kobject_init.cold+0x43/0x51
blk_register_queue+0x46/0x280
add_disk_fwnode+0xb5/0x280
dm_setup_md_queue+0x194/0x1c0
table_load+0x297/0x2d0
ctl_ioctl+0x2a2/0x480
dm_ctl_ioctl+0xe/0x20
__x64_sys_ioctl+0xc7/0x110
do_syscall_64+0x72/0x390
entry_SYSCALL_64_after_hwframe+0x76/0x7e
Fix this by separating kobject initialization from sysfs registration:
- Initialize queue_kobj early during gendisk allocation
- add_disk() only adds the already-initialized kobject to sysfs
- del_gendisk() removes from sysfs but doesn't destroy the kobject
- Final cleanup happens when the disk is released
Fixes: 2bd85221a6 ("block: untangle request_queue refcounting from sysfs")
Reported-by: Li Lingfeng <lilingfeng3@huawei.com>
Closes: https://lore.kernel.org/all/83591d0b-2467-433c-bce0-5581298eb161@huawei.com/
Signed-off-by: Zheng Qixing <zhengqixing@huawei.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Link: https://lore.kernel.org/r/20250808053609.3237836-1-zhengqixing@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull fallocate updates from Christian Brauner:
"fallocate() currently supports creating preallocated files
efficiently. However, on most filesystems fallocate() will preallocate
blocks in an unwriten state even if FALLOC_FL_ZERO_RANGE is specified.
The extent state must later be converted to a written state when the
user writes data into this range, which can trigger numerous metadata
changes and journal I/O. This may leads to significant write
amplification and performance degradation in synchronous write mode.
At the moment, the only method to avoid this is to create an empty
file and write zero data into it (for example, using 'dd' with a large
block size). However, this method is slow and consumes a considerable
amount of disk bandwidth.
Now that more and more flash-based storage devices are available it is
possible to efficiently write zeros to SSDs using the unmap write
zeroes command if the devices do not write physical zeroes to the
media.
For example, if SCSI SSDs support the UMMAP bit or NVMe SSDs support
the DEAC bit[1], the write zeroes command does not write actual data
to the device, instead, NVMe converts the zeroed range to a
deallocated state, which works fast and consumes almost no disk write
bandwidth.
This series implements the BLK_FEAT_WRITE_ZEROES_UNMAP feature and
BLK_FLAG_WRITE_ZEROES_UNMAP_DISABLED flag for SCSI, NVMe and
device-mapper drivers, and add the FALLOC_FL_WRITE_ZEROES and
STATX_ATTR_WRITE_ZEROES_UNMAP support for ext4 and raw bdev devices.
fallocate() is subsequently extended with the FALLOC_FL_WRITE_ZEROES
flag. FALLOC_FL_WRITE_ZEROES zeroes a specified file range in such a
way that subsequent writes to that range do not require further
changes to the file mapping metadata. This flag is beneficial for
subsequent pure overwriting within this range, as it can save on block
allocation and, consequently, significant metadata changes"
* tag 'vfs-6.17-rc1.fallocate' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
ext4: add FALLOC_FL_WRITE_ZEROES support
block: add FALLOC_FL_WRITE_ZEROES support
block: factor out common part in blkdev_fallocate()
fs: introduce FALLOC_FL_WRITE_ZEROES to fallocate
dm: clear unmap write zeroes limits when disabling write zeroes
scsi: sd: set max_hw_wzeroes_unmap_sectors if device supports SD_ZERO_*_UNMAP
nvmet: set WZDS and DRB if device enables unmap write zeroes operation
nvme: set max_hw_wzeroes_unmap_sectors if device supports DEAC bit
block: introduce max_{hw|user}_wzeroes_unmap_sectors to queue limits
The kobject for the queue, `disk->queue_kobj`, is initialized with a
reference count of 1 via `kobject_init()` in `blk_register_queue()`.
While `kobject_del()` is called during the unregister path to remove
the kobject from sysfs, the initial reference is never released.
Add a call to `kobject_put()` in `blk_unregister_queue()` to properly
decrement the reference count and fix the leak.
Fixes: 2bd85221a6 ("block: untangle request_queue refcounting from sysfs")
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20250711083009.2574432-1-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, disks primarily implement the write zeroes command (aka
REQ_OP_WRITE_ZEROES) through two mechanisms: the first involves
physically writing zeros to the disk media (e.g., HDDs), while the
second performs an unmap operation on the logical blocks, effectively
putting them into a deallocated state (e.g., SSDs). The first method is
generally slow, while the second method is typically very fast.
For example, on certain NVMe SSDs that support NVME_NS_DEAC, submitting
REQ_OP_WRITE_ZEROES requests with the NVME_WZ_DEAC bit can accelerate
the write zeros operation by placing disk blocks into a deallocated
state, which opportunistically avoids writing zeroes to media while
still guaranteeing that subsequent reads from the specified block range
will return zeroed data. This is a best-effort optimization, not a
mandatory requirement, some devices may partially fall back to writing
physical zeroes due to factors such as misalignment or being asked to
clear a block range smaller than the device's internal allocation unit.
Therefore, the speed of this operation is not guaranteed.
It is difficult to determine whether the storage device supports unmap
write zeroes operation. We cannot determine this by only querying
bdev_limits(bdev)->max_write_zeroes_sectors. Therefore, first, add a new
hardware queue limit parameters, max_hw_wzeroes_unmap_sectors, to
indicate whether a device supports this unmap write zeroes operation.
Then, add two new counterpart software queue limits,
max_wzeroes_unmap_sectors and max_user_wzeroes_unmap_sectors, which
allow users to disable this operation if the speed is very slow on some
sepcial devices.
Finally, for the stacked devices cases, initialize these two parameters
to UINT_MAX. This operation should be enabled by both the stacking
driver and all underlying devices.
Thanks to Martin K. Petersen for optimizing the documentation of the
write_zeroes_unmap sysfs interface.
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
Link: https://lore.kernel.org/20250619111806.3546162-2-yi.zhang@huaweicloud.com
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: "Martin K. Petersen" <martin.petersen@oracle.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
When blk_unregister_queue() is called from add_disk() failure path,
there is race in registering/unregistering elevator queue kobject
from the two code paths, because commit 559dc11143 ("block: move
elv_register[unregister]_queue out of elevator_lock") moves elevator
queue register/unregister out of elevator lock.
Fix the race by removing elevator after deleting disk->queue_kobj,
because kobject_del(&disk->queue_kobj) drains in-progress sysfs
show()/store() of all attributes.
Fixes: 559dc11143 ("block: move elv_register[unregister]_queue out of elevator_lock")
Reported-by: Nilay Shroff <nilay@linux.ibm.com>
Suggested-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Link: https://lore.kernel.org/r/20250508085807.3175112-3-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_freeze_queue() can't be called on quiesced queue, otherwise it may
never return if there is any queued requests.
Fix it by removing quiesce queue around elevator_set_none() because
elevator_switch() does quiesce queue in case that we need to switch
to none really.
Fixes: 1e44bedbc9 ("block: unifying elevator change")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Link: https://lore.kernel.org/r/20250508085807.3175112-2-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
scheduler's ->exit() is called with queue frozen and elevator lock is held, and
wbt_enable_default() can't be called with queue frozen, otherwise the
following lockdep warning is triggered:
#6 (&q->rq_qos_mutex){+.+.}-{4:4}:
#5 (&eq->sysfs_lock){+.+.}-{4:4}:
#4 (&q->elevator_lock){+.+.}-{4:4}:
#3 (&q->q_usage_counter(io)#3){++++}-{0:0}:
#2 (fs_reclaim){+.+.}-{0:0}:
#1 (&sb->s_type->i_mutex_key#3){+.+.}-{4:4}:
#0 (&q->debugfs_mutex){+.+.}-{4:4}:
Fix the issue by moving wbt_enable_default() out of bfq's exit(), and
call it from elevator_change_done().
Meantime add disk->rqos_state_mutex for covering wbt state change, which
matches the purpose more than ->elevator_lock.
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20250505141805.2751237-26-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Elevator change is one well-define behavior:
- tear down current elevator if it exists
- setup new elevator
It is supposed to cover any case for changing elevator by single
internal API, typically the following cases:
- setup default elevator in add_disk()
- switch to none in del_disk()
- reset elevator in blk_mq_update_nr_hw_queues()
- switch elevator in sysfs `store` elevator attribute
This patch uses elevator_change() to cover all above cases:
- every elevator switch is serialized with each other: add_disk/del_disk/
store elevator is serialized already, blk_mq_update_nr_hw_queues() uses
srcu for syncing with the other three cases
- for both add_disk()/del_disk(), queue freeze works at atomic mode
or has been froze, so the freeze in elevator_change() won't add extra
delay
- `struct elev_change_ctx` instance holds any info for changing elevator
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20250505141805.2751237-17-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Thomas WeißschuhandJens Axboe
Damien Le MoalandJens Axboe
Ming LeiandJens Axboe
Yu KuaiandJens Axboe
Bart Van AsscheandJens Axboe
Yu KuaiandJens Axboe
Julian SunandJens Axboe
Zheng QixingandJens Axboe
Linus Torvalds
Zhang YiandChristian Brauner
Christoph HellwigandJens Axboe
Keith BuschandJens Axboe