Pull MD updates from Song:
"1. Deprecate bitmap file support, by Christoph Hellwig;
2. Fix deadlock with md sync thread, by Yu Kuai;
3. Refactor md io accounting, by Yu Kuai;
4. Various non-urgent fixes by Li Nan, Yu Kuai, and Jack Wang."
* tag 'md-next-20230729' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md: (36 commits)
md/md-bitmap: hold 'reconfig_mutex' in backlog_store()
md/md-bitmap: remove unnecessary local variable in backlog_store()
md/raid10: use dereference_rdev_and_rrdev() to get devices
md/raid10: factor out dereference_rdev_and_rrdev()
md/raid10: check replacement and rdev to prevent submit the same io twice
md/raid1: Avoid lock contention from wake_up()
md: restore 'noio_flag' for the last mddev_resume()
md: don't quiesce in mddev_suspend()
md: remove redundant check in fix_read_error()
md/raid10: optimize fix_read_error
md/raid1: prioritize adding disk to 'removed' mirror
md/md-faulty: enable io accounting
md/md-linear: enable io accounting
md/md-multipath: enable io accounting
md/raid10: switch to use md_account_bio() for io accounting
md/raid1: switch to use md_account_bio() for io accounting
raid5: fix missing io accounting in raid5_align_endio()
md: also clone new io if io accounting is disabled
md: move initialization and destruction of 'io_acct_set' to md.c
md: deprecate bitmap file support
...
Several reasons why 'reconfig_mutex' should be held:
1) rdev_for_each() is not safe to be called without the lock, because
rdev can be removed concurrently.
2) mddev_destroy_serial_pool() and mddev_create_serial_pool() should not
be called concurrently.
3) mddev_suspend() from mddev_destroy/create_serial_pool() should be
protected by the lock.
Fixes: 10c92fca63 ("md-bitmap: create and destroy wb_info_pool with the change of backlog")
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20230706083727.608914-3-yukuai1@huaweicloud.com
Signed-off-by: Song Liu <song@kernel.org>
Commit 2ae6aaf769 ("md/raid10: fix io loss while replacement replace
rdev") reads replacement first to prevent io loss. However, there are same
issue in wait_blocked_dev() and raid10_handle_discard(), too. Fix it by
using dereference_rdev_and_rrdev() to get devices.
Fixes: d30588b273 ("md/raid10: improve raid10 discard request")
Fixes: f2e7e269a7 ("md/raid10: pull the code that wait for blocked dev into one function")
Signed-off-by: Li Nan <linan122@huawei.com>
Link: https://lore.kernel.org/r/20230701080529.2684932-4-linan666@huaweicloud.com
Signed-off-by: Song Liu <song@kernel.org>
wake_up is called unconditionally in a few paths such as make_request(),
which cause lock contention under high concurrency workload like below
raid1_end_write_request
wake_up
__wake_up_common_lock
spin_lock_irqsave
Improve performance by only call wake_up() if waitqueue is not empty
Fio test script:
[global]
name=random reads and writes
ioengine=libaio
direct=1
readwrite=randrw
rwmixread=70
iodepth=64
buffered=0
filename=/dev/md0
size=1G
runtime=30
time_based
randrepeat=0
norandommap
refill_buffers
ramp_time=10
bs=4k
numjobs=400
group_reporting=1
[job1]
Test result with 2 ramdisk in raid1 on a Intel Broadwell 56 cores server.
Before this patch With this patch
READ BW=4621MB/s BW=7337MB/s
WRITE BW=1980MB/s BW=3144MB/s
The patch is inspired by Yu Kuai's change for raid10:
https://lore.kernel.org/r/20230621105728.1268542-1-yukuai1@huaweicloud.com
Cc: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20230705113227.148494-1-jinpu.wang@ionos.com
Signed-off-by: Song Liu <song@kernel.org>
memalloc_noio_save() is called for the first mddev_suspend(), and
repeated mddev_suspend() only increase 'suspended'. However,
memalloc_noio_restore() is also called for the first mddev_resume(),
which means that memory reclaim will be enabled before the last
mddev_resume() is called, while the array is still suspended.
Fix this problem by restore 'noio_flag' for the last mddev_resume().
Fixes: 78f57ef9d5 ("md: use memalloc scope APIs in mddev_suspend()/mddev_resume()")
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20230628012931.88911-3-yukuai1@huaweicloud.com
Signed-off-by: Song Liu <song@kernel.org>
Some levels doesn't implement "pers->quiesce", for example
raid0_quiesce() is empty, and now that all levels will drop 'active_io'
until io is done, wait for 'active_io' to be 0 is enough to make sure all
normal io is done, and percpu_ref_kill() for 'active_io' will make sure
no new normal io can be dispatched. There is no need to call
"pers->quiesce" anymore from mddev_suspend().
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20230628012931.88911-2-yukuai1@huaweicloud.com
Signed-off-by: Song Liu <song@kernel.org>
Two problems can be fixed this way:
1) 'active_io' will represent inflight io instead of io that is
dispatching.
2) If io accounting is enabled or disabled while io is still inflight,
bio_start_io_acct() and bio_end_io_acct() is not balanced and io
inflight counter will be leaked.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230621165110.1498313-5-yukuai1@huaweicloud.com
Currently, 'active_io' is grabbed before make_reqeust() is called, and
it's dropped immediately make_reqeust() returns. Hence 'active_io'
actually means io is dispatching, not io is inflight.
For raid0 and raid456 that io accounting is enabled, 'active_io' will
also be grabbed when bio is cloned for io accounting, and this 'active_io'
is dropped until io is done.
Always clone new bio so that 'active_io' will mean that io is inflight,
raid1 and raid10 will switch to use this method in later patches.
Now that bio will be cloned even if io accounting is disabled, also
rename related structure from '*_acct_*' to '*_clone_*'.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230621165110.1498313-3-yukuai1@huaweicloud.com
'io_acct_set' is only used for raid0 and raid456, prepare to use it for
raid1 and raid10, so that io accounting from different levels can be
consistent.
By the way, follow up patches will also use this io clone mechanism to
make sure 'active_io' represents in flight io, not io that is dispatching,
so that mddev_suspend will wait for io to be done as designed.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230621165110.1498313-2-yukuai1@huaweicloud.com
The support for bitmaps on files is a very bad idea abusing various kernel
APIs, and fundamentally requires the file to not be on the actual array
without a way to check that this is actually the case. Add a deprecation
warning to see if we might be able to eventually drop it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230615064840.629492-12-hch@lst.de
The support for write intent bitmaps in files on an external files in md
is a hot mess that abuses ->bmap to map file offsets into physical device
objects, and also abuses buffer_heads in a creative way.
Make this code optional so that MD can be built into future kernels
without buffer_head support, and so that we can eventually deprecate it.
Note this does not affect the internal bitmap support, which has none of
the problems.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230615064840.629492-11-hch@lst.de
The md driver allocates pages for storing the bitmap file data, which
are not page cache pages, and then stores the page granularity file
offset in page->index, which is a field that isn't really valid except
for page cache pages.
Use a separate index for the superblock, and use the scheme used at
read size to recalculate the index for the bitmap pages instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230615064840.629492-10-hch@lst.de
For md_check_recovery():
1) if 'MD_RECOVERY_RUNING' is not set, register new sync_thread.
2) if 'MD_RECOVERY_RUNING' is set:
a) if 'MD_RECOVERY_DONE' is not set, don't do anything, wait for
md_do_sync() to be done.
b) if 'MD_RECOVERY_DONE' is set, unregister sync_thread. Current code
expects that sync_thread is not NULL, otherwise new sync_thread will
be registered, which will corrupt the array.
Make sure md_check_recovery() won't register new sync_thread if
'MD_RECOVERY_RUNING' is still set, and a new WARN_ON_ONCE() is added for
the above corruption,
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230529132037.2124527-7-yukuai1@huaweicloud.com
Our test found a following deadlock in raid10:
1) Issue a normal write, and such write failed:
raid10_end_write_request
set_bit(R10BIO_WriteError, &r10_bio->state)
one_write_done
reschedule_retry
// later from md thread
raid10d
handle_write_completed
list_add(&r10_bio->retry_list, &conf->bio_end_io_list)
// later from md thread
raid10d
if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
list_move(conf->bio_end_io_list.prev, &tmp)
r10_bio = list_first_entry(&tmp, struct r10bio, retry_list)
raid_end_bio_io(r10_bio)
Dependency chain 1: normal io is waiting for updating superblock
2) Trigger a recovery:
raid10_sync_request
raise_barrier
Dependency chain 2: sync thread is waiting for normal io
3) echo idle/frozen to sync_action:
action_store
mddev_lock
md_unregister_thread
kthread_stop
Dependency chain 3: drop 'reconfig_mutex' is waiting for sync thread
4) md thread can't update superblock:
raid10d
md_check_recovery
if (mddev_trylock(mddev))
md_update_sb
Dependency chain 4: update superblock is waiting for 'reconfig_mutex'
Hence cyclic dependency exist, in order to fix the problem, we must
break one of them. Dependency 1 and 2 can't be broken because they are
foundation design. Dependency 4 may be possible if it can be guaranteed
that no io can be inflight, however, this requires a new mechanism which
seems complex. Dependency 3 is a good choice, because idle/frozen only
requires sync thread to finish, which can be done asynchronously that is
already implemented, and 'reconfig_mutex' is not needed anymore.
This patch switch 'idle' and 'frozen' to wait sync thread to be done
asynchronously, and this patch also add a sequence counter to record how
many times sync thread is done, so that 'idle' won't keep waiting on new
started sync thread.
Noted that raid456 has similiar deadlock([1]), and it's verified[2] this
deadlock can be fixed by this patch as well.
[1] https://lore.kernel.org/linux-raid/5ed54ffc-ce82-bf66-4eff-390cb23bc1ac@molgen.mpg.de/T/#t
[2] https://lore.kernel.org/linux-raid/e9067438-d713-f5f3-0d3d-9e6b0e9efa0e@huaweicloud.com/
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230529132037.2124527-5-yukuai1@huaweicloud.com
Currently, for idle and frozen, action_store will hold 'reconfig_mutex'
and call md_reap_sync_thread() to stop sync thread, however, this will
cause deadlock (explained in the next patch). In order to fix the
problem, following patch will release 'reconfig_mutex' and wait on
'resync_wait', like md_set_readonly() and do_md_stop() does.
Consider that action_store() will set/clear 'MD_RECOVERY_FROZEN'
unconditionally, which might cause unexpected problems, for example,
frozen just set 'MD_RECOVERY_FROZEN' and is still in progress, while
'idle' clear 'MD_RECOVERY_FROZEN' and new sync thread is started, which
might starve in progress frozen. A mutex is added to synchronize idle
and frozen from action_store().
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230529132037.2124527-4-yukuai1@huaweicloud.com
This reverts commit 9dfbdafda3.
Because it will introduce a defect that sync_thread can be running while
MD_RECOVERY_RUNNING is cleared, which will cause some unexpected problems,
for example:
list_add corruption. prev->next should be next (ffff0001ac1daba0), but was ffff0000ce1a02a0. (prev=ffff0000ce1a02a0).
Call trace:
__list_add_valid+0xfc/0x140
insert_work+0x78/0x1a0
__queue_work+0x500/0xcf4
queue_work_on+0xe8/0x12c
md_check_recovery+0xa34/0xf30
raid10d+0xb8/0x900 [raid10]
md_thread+0x16c/0x2cc
kthread+0x1a4/0x1ec
ret_from_fork+0x10/0x18
This is because work is requeued while it's still inside workqueue:
t1: t2:
action_store
mddev_lock
if (mddev->sync_thread)
mddev_unlock
md_unregister_thread
// first sync_thread is done
md_check_recovery
mddev_try_lock
/*
* once MD_RECOVERY_DONE is set, new sync_thread
* can start.
*/
set_bit(MD_RECOVERY_RUNNING, &mddev->recovery)
INIT_WORK(&mddev->del_work, md_start_sync)
queue_work(md_misc_wq, &mddev->del_work)
test_and_set_bit(WORK_STRUCT_PENDING_BIT, ...)
// set pending bit
insert_work
list_add_tail
mddev_unlock
mddev_lock_nointr
md_reap_sync_thread
// MD_RECOVERY_RUNNING is cleared
mddev_unlock
t3:
// before queued work started from t2
md_check_recovery
// MD_RECOVERY_RUNNING is not set, a new sync_thread can be started
INIT_WORK(&mddev->del_work, md_start_sync)
work->data = 0
// work pending bit is cleared
queue_work(md_misc_wq, &mddev->del_work)
insert_work
list_add_tail
// list is corrupted
The above commit is reverted to fix the problem, the deadlock this
commit tries to fix will be fixed in following patches.
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230529132037.2124527-2-yukuai1@huaweicloud.com
blk_mq_run_queue() runs the queue asynchronously if BLK_MQ_F_BLOCKING
has been set. This is suboptimal since running the queue asynchronously
is slower than running the queue synchronously. This patch modifies
blk_mq_run_queue() as follows if BLK_MQ_F_BLOCKING has been set:
- Run the queue synchronously if it is allowed to sleep.
- Run the queue asynchronously if it is not allowed to sleep.
Additionally, blk_mq_run_hw_queue(hctx, false) calls are modified into
blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING) if the caller
may be invoked from atomic context.
The following caller chains have been reviewed:
blk_mq_run_hw_queue(hctx, false)
blk_mq_get_tag() /* may sleep, hence the functions it calls may also sleep */
blk_execute_rq() /* may sleep */
blk_mq_run_hw_queues(q, async=false)
blk_freeze_queue_start() /* may sleep */
blk_mq_requeue_work() /* may sleep */
scsi_kick_queue()
scsi_requeue_run_queue() /* may sleep */
scsi_run_host_queues()
scsi_ioctl_reset() /* may sleep */
blk_mq_insert_requests(hctx, ctx, list, run_queue_async=false)
blk_mq_dispatch_plug_list(plug, from_sched=false)
blk_mq_flush_plug_list(plug, from_schedule=false)
__blk_flush_plug(plug, from_schedule=false)
blk_add_rq_to_plug()
blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */
blk_mq_plug_issue_direct()
blk_mq_flush_plug_list() /* see above */
blk_mq_dispatch_plug_list(plug, from_sched=false)
blk_mq_flush_plug_list() /* see above */
blk_mq_try_issue_directly()
blk_mq_submit_bio() /* may sleep if REQ_NOWAIT has not been set */
blk_mq_try_issue_list_directly(hctx, list)
blk_mq_insert_requests() /* see above */
Cc: Christoph Hellwig <hch@lst.de>
Cc: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20230721172731.955724-4-bvanassche@acm.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk_mq_kick_requeue_list() calls blk_mq_run_hw_queues() asynchronously.
Leave out the direct blk_mq_run_hw_queues() call. This patch causes
scsi_run_queue() to call blk_mq_run_hw_queues() asynchronously instead
of synchronously. Since scsi_run_queue() is not called from the hot I/O
submission path, this patch does not affect the hot path.
This patch prepares for allowing blk_mq_run_hw_queue() to sleep if
BLK_MQ_F_BLOCKING has been set. scsi_run_queue() may be called from
atomic context and must not sleep. Hence the removal of the
blk_mq_run_hw_queues(q, false) call. See also scsi_unblock_requests().
Cc: "Martin K. Petersen" <martin.petersen@oracle.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/20230721172731.955724-3-bvanassche@acm.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is no good reason to pass the bio to bio_try_merge_hw_seg. Just
pass the current bvec and rename the function to bvec_try_merge_hw_page.
This will allow reusing this function for supporting multi-page integrity
payload bvecs.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jinyoung Choi <j-young.choi@samsung.com>
Link: https://lore.kernel.org/r/20230724165433.117645-9-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The update of bi_size is the only thing in __bio_try_merge_page that
needs a bio. Move it to the callers, and merge __bio_try_merge_page
and page_is_mergeable into a single bvec_try_merge_page that only takes
the current bvec instead of a full bio. This will allow reusing this
function for supporting multi-page integrity payload bvecs.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jinyoung Choi <j-young.choi@samsung.com>
Link: https://lore.kernel.org/r/20230724165433.117645-8-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Checking for availability in bi_size in a function that attempts to
merge into an existing segment is a bit odd, as the limit also applies
when adding a new segment. This code works fine as we always call
__bio_try_merge_page, but contributes to sub-optimal calling conventions
and doesn't lead to clear code.
Move it to two of the callers instead, the third one already has a more
strict check that includes max_hw_segments anyway.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jinyoung Choi <j-young.choi@samsung.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Link: https://lore.kernel.org/r/20230724165433.117645-6-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__bio_try_merge_page is a way too low-level helper to assert that the
bio is not cloned. Move the check into bio_add_page and
bio_iov_iter_get_pages instead, which are the high level entry points
that should enforce this variant. bio_add_hw_page already this
check, coverig the third (indirect) caller of __bio_try_merge_page.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jinyoung Choi <j-young.choi@samsung.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Link: https://lore.kernel.org/r/20230724165433.117645-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
