[Devel] [PATCH hci-8.1 v7 0/8] block/blk-cbt: don't lose writes held across a dm suspend

[Pavel Tikhomirov <ptikhomirov@virtuozzo.com>]

Fix CBT missing writes that request-based dm holds across a suspend,
which made incremental backups taken with ploop-backup silently lose
data.

Claude tested full and incremental ploop-backup under write load (a
random writer, and fio with 4 jobs x iodepth 64): the incremental backup
lost blocks without the fix, and was consistent in 30/30 and 10/10 runs
with it. CBT start/stop and device resize under fio ran without
warnings. With writes throttled by io.max on a loop device, CBT recorded
24 of 64 written blocks without the throttle patch, and all 64 with it.
A RWF_NOWAIT write into a missing bitmap page now fails with EAGAIN
instead of sleeping, and succeeds when retried without RWF_NOWAIT.

https://virtuozzo.atlassian.net/browse/VSTOR-149268

Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
---
Changes in v7:
- New patch: use list_for_each_entry_safe() in blk_cbt_update_size(),
  which read the next entry through a cbt it had just queued for
  release (sashiko review).
- Take the SRCU read lock in blk_cbt_bio_queue() only after its early
  exits, so that reads and queues without a CBT don't pay for it.
- Clarify the commit message of the unsigned shift patch and the comment
  of cbt_pages_present() (sashiko review).
- Link to v6: http://lore.virtuozzo.com/20261027-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v6-0-f9870cf5c7eb@virtuozzo.com

Changes in v6:
- New patch: load a map slot once with READ_ONCE() in CBT_PAGE(), as a
  racing CBT_PAGE_MISSED store could make it return (struct page *)0x1,
  and install a map page with smp_store_release(), so that a lockless
  reader never sees it before its zeroing (sashiko review).
- Store CBT_PAGE_MISSED with cmpxchg() instead of under cbt->lock, since
  taking a spinlock_t in the flush IPI is not allowed: it may sleep on
  PREEMPT_RT (sashiko review).
- Skip queueing a page allocation that is already queued, instead of
  taking cbt->lock for every write while the page is missing (sashiko
  review).
- Link to v5: http://lore.virtuozzo.com/20261027-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v5-0-88dccc74b48b@virtuozzo.com

Changes in v5:
- New patch: store CBT_PAGE_MISSED under cbt->lock, as the cache flush
  could overwrite a page installed meanwhile; the allocation queued at
  submission in v4 made this much more likely (sashiko review).
- New patch: walk cbt_list under SRCU, so that a sleeping bitmap page
  allocation no longer drops the RCU lock in the middle of the walk, which
  let a concurrent CBT stop or resize free the next cbt under it, a use
  after free (the pre-existing problem noted in v4).
- New patch: don't sleep in CBT for a REQ_NOWAIT write, fail it with
  BLK_STS_AGAIN if a bitmap page is missing (also noted in v4).
- Link to v4: http://lore.virtuozzo.com/20261027-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v4-0-b4d92264f6a4@virtuozzo.com

Changes in v4:
- Only queue the bitmap page allocation at submission instead of
  allocating there: the sleeping allocation dropped the RCU lock in the
  middle of the cbt_list walk, and could put an extent into the cache of
  a cbt that a resize had just replaced (sashiko review).
- Link to v3: http://lore.virtuozzo.com/20261027-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v3-0-e1bc61a829a0@virtuozzo.com

Changes in v3:
- New patch: account a write delayed by blk-throttle in CBT, as the
  throttle dispatch work skipped CBT (sashiko review).
- New patch: shift the CBT block size as unsigned, and use the unsigned
  shift in the new cbt_set_present() too (sashiko review).
- Link to v2: http://lore.virtuozzo.com/20261027-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v2-1-c64ee6178730@virtuozzo.com

Changes in v2:
- No GFP_ATOMIC allocation at dispatch: bitmap pages are now allocated
  at bio submission, and the dispatch hook only sets bits in present pages.
- A page missing at dispatch (a write submitted before BLKCBTSTART) is
  allocated with GFP_NOIO from the kblockd workqueue, and dm-rq returns
  BLK_STS_RESOURCE so that blk-mq dispatches the request again later,
  instead of setting CBT_ERROR when an atomic allocation fails.
- Move the hook from dm_start_request() to dm_mq_queue_rq(), before the
  request is started, so that it can return BLK_STS_RESOURCE.
- Link to v1: http://lore.virtuozzo.com/20261016-ptikhomirov-vstor-149268-cbt-account-at-rq-start-v1-1-1f20b151431d@virtuozzo.com

---
Pavel Tikhomirov (8):
      block/blk-cbt: don't read a released cbt in blk_cbt_update_size()
      block/blk-cbt: shift the block size as unsigned
      block/blk-throttle: account a throttled write in CBT
      block/blk-cbt: read a map slot once, publish a map page with release
      block/blk-cbt: don't overwrite a map page with CBT_PAGE_MISSED
      block/blk-cbt: walk cbt_list under SRCU
      block/blk-cbt: don't sleep in CBT for a REQ_NOWAIT write
      block/blk-cbt: account a write request also when dm starts it

 block/blk-cbt.c        | 308 ++++++++++++++++++++++++++++++++++++++++++-------
 block/blk-core.c       |   5 +-
 block/blk-throttle.c   |   5 +-
 drivers/md/dm-rq.c     |   3 +
 include/linux/blkdev.h |  12 +-
 5 files changed, 288 insertions(+), 45 deletions(-)
---
base-commit: f5d1fdcfffe101fe2bafa7df09d95d2e811c5c5e
change-id: 20261016-ptikhomirov-vstor-149268-cbt-account-at-rq-start-a592cf2010aa

Best regards,
--  
Pavel Tikhomirov <ptikhomirov@virtuozzo.com>

_______________________________________________
Devel mailing list
Devel@openvz.org
https://lists.openvz.org/mailman/listinfo/devel