RE:(2) [PATCH v2 2/2] xfs: skip the fsync flush when a later flush has completed
From: Daejun Park
Date: Wed Nov 11 2026 - 21:39:57 EST
On Wed, Nov 11, 2026 at 09:22:19AM -0500, Brian Foster wrote:
> Hmm.. that's kind of interesting. So IIUC unless dm can pass through a
> preflush+data bio, it will split out the flushes and ignore any error
> status in favor of the packaged data I/O..?
Yes. dm sends the empty flush first and then reissues the data without
REQ_PREFLUSH. __dm_io_complete() drops the flush error when it reissues
the data.
I sent a fix to dm-devel. With the fix, a failed preflush fails the
iclog write, and XFS shuts down:
https://lore.kernel.org/r/20261112-dm-flush-data-error-v1-1-c83d3f8f3ac4@xxxxxxxxxxx
[...]
> But if you have to post this again I think it would be useful to
> describe this caveat in the commit log description as well. If there
> happens to be some currently unknown flush failiure scenario that
> materializes in the future and we don't currently hit due to spamming
> enough flushes, that would be a useful clue about a subtle behavior
> change.
Thanks, agreed. If I post a v3, the dm paragraph below the --- will move
into the commit log of patch 2.
Best regards,
Daejun