[Devel] [PATCH hci-8.1 v2 0/4] dm-ploop: fix acknowledged writes lost when discard and allocation share a BAT writeback

[Andrey Zhadchenko]

Reviewed-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>


On 10/28/26 15:36, Denis V. Lunev wrote:
> The hci-ploop fstrim test caught dm-ploop losing an acknowledged write
> (VSTOR-151979). Since "dm-ploop: fix and rework md updates", an
> allocation and a discard on the same BAT page can share one pending
> index writeback, and its completion applies the written page by the
> writeback's type only. A first write that rides in a DISCARD writeback
> is on disk but never reaches the local BAT: it reads back as zeroes,
> the next write allocates the cluster again and the data is gone. A
> discard that rides in an ALLOC writeback leaves the local BAT mapped
> while disk has zero, so writes after it are lost on the next open.
> fstrim or guest TRIM racing ordinary writes is enough.
> 
> Patch 4 is the fix: the completion applies the changes of the pios
> attached to the writeback, and of the relocation done on grow, instead
> of filtering the page by type. Patches 1 and 2 fix older bugs on the
> same path that the old completion partly hid: a failed BAT writeback
> freed the wrong clusters, and the online merge of the latest snapshot
> left BAT levels above the top one. Patch 3 drops the dead PLOOP_DELAYWB
> code.
> 
> Tested on dev-cl04, 6.12.0-211.52.1.22.9.vz10 with the module built
> out of tree:
> - a reproducer racing cluster discards with first writes on one BAT
>    page, 512 clusters a round: without the series 3 of 15360 first
>    writes read back as zeroes right after the write returned and stayed
>    lost after remount; with a 200 us sleep before bat_lock in the
>    allocation path 1801 of 2560; the reverse order, a sleep in the
>    discard path, left data written after a discard unmapped on disk.
>    With the series nothing is lost in either order;
> - the ploop pytest suite, ploop and qcow2 formats, passes;
> - nothing exercises a failing BAT writeback, that path is checked by
>    reading only.
> 
> Changes in v2, after Andrey's review:
> - 2/4: also clear the deltas[] slot the top delta moved out of, as
>    notify_delta_merged() does. The md_lock in the renumbering stays: it
>    is per page, and notify_delta_merged() takes it the same way;
> - 3/4: new, drops every PLOOP_DELAYWB block instead of only the one in
>    the completion;
> - 4/4: the PLOOP_DELAYWB removal moved to 3/4, otherwise unchanged;
> - Reviewed-by added to 1/4, 2/4 and 4/4.
> The ploop pytest suite and the reproducer were rerun on v2.
> 
> v1: http://lore.virtuozzo.com/kernel/20261027163300.2166639-1-den@openvz.org/
> 
> Denis V. Lunev (4):
>    dm-ploop: fix freeing the wrong clusters when a BAT writeback fails
>    dm-ploop: renumber BAT levels when the latest snapshot is merged
>    dm-ploop: drop the dead PLOOP_DELAYWB code
>    dm-ploop: fix lost BAT updates when alloc and discard share a
>      writeback
> 
>   drivers/md/dm-ploop-cmd.c |  25 +++++-
>   drivers/md/dm-ploop-map.c | 161 +++++++++++++++++---------------------
>   drivers/md/dm-ploop.h     |   1 +
>   3 files changed, 95 insertions(+), 92 deletions(-)
> 
> 
> base-commit: b499efa17aa9fb8f20512d1831d1cb61b22e60d3