Re: [PATCH] xfs: fix skipped flushing items not counted in xfsaild_push()

[Dave Chinner <dgc@kernel.org>]

On Wed, Sep 23, 2026 at 09:44:32AM -0400, Brian Foster wrote:
> On Wed, Sep 23, 2026 at 08:12:46AM +1000, Dave Chinner wrote:
> > On Tue, Sep 22, 2026 at 04:14:29PM +0800, MingTao Huang wrote:
> > > From: MingTao Huang <mintaohuang@tencent.com>
> > > 
> > > Commit f3f7ae68a4ea ("xfs: skip flushing log items during push")
> > > introduced a fast path in xfsaild_push() that uses
> > > test_bit(XFS_LI_FLUSHING) to skip log items already being
> > > flushed.  However, the fast path jumps directly to the
> > > next_item label, bypassing flushing++, count++, and the
> > > ail_last_pushed_lsn update.  This causes three problems:

[....]

> > >    letting the loop scan an order of magnitude more items per round.
> > >    Each cluster flush adds a buffer to ail_buf_list, and the resulting
> > >    oversized list causes xfs_buf_delwri_submit_nowait() -- which runs
> > >    list_sort() plus per-buffer trylock and IO submission -- to take so
> > >    long that the watchdog fires.
> > 
> > How? count bounds the delwri list to 1000 buffers at most. That's
> > very intentional, and the number is based on benchmarking results.
> > 
> 
> I'm confused by some of the reasoning here. Maybe I'm missing something,
> but that count check looks like it only limits the current pass.  The
> flushing value is basically just input into the thread throttling (i.e.
> schedule timeout) behavior.

It's much more complex than that.

The count bounds the number of buffers we can add to the list each
pass.  We also submit the entire list for IO once per pass, hence we
fill and drain the list once per pass. Hence in normal
circumstances, the AIL buffer list should never grow very much
beyond 1000 buffers.

Further, using xfs_buf_delwri_submit_nowait() doesn't mean that it
won't block, it just won't wait for IO completion.  We can still
block on IO submission when the request queue fills up.  i.e. the
AIL push algorithm is designed to throttle writeback processing at
IO submission time, not via the higher level backoff loops.

If the disk is really that fast, then all the items in FLUSHING
state should be completing -really fast- and being removed from the
AIL just as fast. Hence we should not be spinning on the same items
over and over again if we are never blocking in delwri buffer
submission. If we are spinning on the same items, that implies a
problem with IO completion (ie not removing flushing items), not the
AIL push algorithm.

The AIl control loop is also intended to sleep unconditionally when
it reaches the target LSN. Therefore, not sleeping for a long time
implies that the target is continually being moved forward faster
than the AIL can push items to the disk. This also implies that the
disk is sufficiently fast or has sufficiently deep request queues
that it never blocks on IO submission.

Spinning without AIL level backoff or IO submission blocking
occuring therefore implies journal reservations must be full (i.e.
lots of userspace modifcation concurrency on a small journal) or we
are under severe memory pressure triggering repeated full AIL
pushes.  These are really the only two ways to guarantee the push
target always keeps ahead of the aild pushing and so keeps it
permanently busy.

Hence my request for details about the workload, fs geometry, etc,
so I have some idea of how we are getting into this state in the
first place.

> So ISTM that if the majority of the AIL is
> flushing, we'd historically want to schedule out and wait for I/O to
> complete, whereas what this patch describes is a scenario where we now
> spin around until enough items clear off the list in that particular
> situation.

We will only "spin" if pushing is not making progress. If IO is
being submitted, the the LI_FLUSHING state is transient, and the
items will be removed from the AIL when the IO completes. We skip
over them because waiting on LI_FLUSHING items when we are not yet
at the push target LSN means leaving the disk idle when we could be
pushing more writeback to it. i.e. we take longer to submit all the
items we need to get to the target and so end up with lower disk
utilisation, longer journal reservation latency and lower
performance/throughput.

i.e. we only want to throttle processing on IO submission,
not on the number of items we have flushed to IO buffers. Only when
we hit AIL congestion do we want to back off at the AIL level.

> Prior to the optimization in ~2024, it looks like we'd account all these
> flushing items and back off.

Yes, that was broken behaviour, and not what I'd originally intended
for the flush accounting. That commit was when I realised that I'd
overlooked the fact that ITEM_FLUSHING was reporting two completely
different things for inode items. i.e. I realised that the
predominant behaviour being accounted was not the feedback metric
the control loop was designed to use.

The control loop attempts to maximise speed and efficiency of
IO submission - it is not intended to maximise/optimise the number
of items that get processed. Inodes reporting ITEM_FLUSHING when
XFS_IFLUSHING was set was essentially just counting the number of
inodes we attempt to push, and had nothing correlation to IO
submission behaviour.

When the inode is first pushed to the cluster buffer it has
XFS_IFLUSHING set, and if the buffer is successfully queued, it
returns ITEM_SUCCESS. However, if the buffer was already queued (i.e some
other inode has already been pushed to it this cycle), then it will
return ITEM_FLUSHING.

This ITEM_FLUSHING return value indicates that newly dirty inodes
have entered the AIL between when the inode cluster buffer was first
flushed and queued in this processing loop (and returned
ITEM_SUCESS) and now.

IOWs, ITEM_FLUSHING is supposed to be an indication that the the
item is under active modification (i.e. that it has been flushed
multiple times this processing loop), and so there is active
access/modification vs writeback contention on the item.

The other case that inode item push returns ITEM_FLUSHING is if the
inode already has XFS_IFLUSHING set on it. This indicates that the
inode has already been flushed to the cluster buffer via a another
inode push on the same cluster buffer, but by itself it tells us
-nothing- about whether that inode is being actively modified.

That is because xfs_iflush_cluster() gathers all inodes in the
buffer, regardless of where they are in the AIL. So may be older,
some newer, and so encountering a XFS_IFLUSHING inode carries no
signal about whether it is being actively modified, nor does it tell
us that the queued cluster buffer has been updated whilst it was
already queued.

Hence, from the POV of the AIL control loop, this latter
ITEM_FLUSHING metric is pure noise. We did not queue new IO, we did
not update a buffer already queued for IO, and we do not know if the
item has been recently modified, yet we still accounted it as
"flushing".

Hence the LI_FLUSHING tag: it separates the signal the control loop
wants (how many objects we pushed multiple times in this processing
loop) from the noise (how many inodes we've encountered that are
already either queued for submission, under IO or queued for
completion processing from the current and/or any previous
processing loop iteration).

IOWs, the AIL push loop backoffs are not really about waiting for IO
completion - IO completion occurs as a side effect of backing off.
The purpose is to reduce the amount of unnecessary writeback and log
forces we do for objects in the AIL that are under active
modification.

By allowing time for such objects to be relogged and moved forward
in the AIL instead of written back, we minimise the amount of
repeated metadata IO we will need to reach the target LSN. Relogging
is far more IO efficient (sequential journal IO) that metadata
writeback (random small write IO), and this is one of the key IO
optimisations the back-offs are trying to achieve.

TL;DR: the flush backoff is supposed to address the case where the
AIL is contending with active modifications to items that need
writeback. It backs off to allow the IO to complete and allow the
actively modified objects to be relogged and moved well away from
the current LSN the AIL is processing. This avoids repeated
mod->journal->writeback->mod->journal->writeback... cycles,
encouraging mod->journal->mod->journal->... behaviour instead.

> Also FWIW, looking back at that commit it
> doesn't say anything about bounding I/O (not to say this isn't a natural
> side effect of xfaild throttling).

It's not a side effect of xfsaild throttling. It's a natural
behaviour from the 1:1 "fill queue, drain queue" IO processing loop.
I thought that was obvious from the "block on IO submission" part of
the commit message and never needed more explanation, but I guess it
wasn't.

> That commit seems mainly focused on
> reducing ->iop_push() call overhead and the implementation only focuses
> on inode items, so altogether this strikes me as more of an (perfectly
> valid) optimization than fundamental behavior change in how the
> processing thread or ail accounting should work.

Commit messages can never tell the whole story. They have to walk a
line between describing the problem being solved and reviewers
needing to understand all the subtle intricacies of how a tiny tweak
fixes a problem with a complex algorithm....

Reading that commit message back now, the iop_push() cpu consumption
was the measurable symptom that exposed the issue (i.e excessive CPU
usage processing a million 'no-op' ITEM_FLUSHING pushes every
second). However, the second half of the commit message is all about
how the LI_FLUSHING accounting change exposed other mitigations we'd
made to handle the noise the inode flushing accounting generated,
and why they weren't necessary anymore.

> > > We hit this as a soft lockup during stress testing on an internal
> > > kernel that includes commit f3f7ae68a4ea ("xfs: skip flushing log
> > > items during push").  The xfsaild kthread was stuck for
> > > 22 seconds inside xfs_buf_delwri_submit_nowait(), called from
> > > xfsaild_push(), processing an excessively large ail_buf_list:
> > 
> > Tell me how the delwri list got so long that it gets stuck inside
> > xfs_buf_delwri_submit_nowait(). Whatever caused that is the bug that
> > we need to understand and fix - changing accounting to make gross
> > behavioural changes that result in exceedingly inefficient CPU usage
> > and break IO optimisations is not the way to address a list length
> > that has apparently excceeded the bounds built into the code...
> > 
> 
> I think the presumption being made here is that the list is growing
> excessively because xfsaild is failing to back off when the majority of
> (inode) items on the list are in the flushing state.

I don't want to presume anything.

i.e. I don't know how the problem manifests yet, and the proposed
solution is not viable. We need to find the root cause of the issue
before going any further.

> I guess it would be interesting to know how long the buf list actually
> is, and whether the soft lockup warning is purely due to buf list size,
> or more of a combination of ail list size/state, buf list size, and
> xfsaild thread spinning behavior (we also changed the default timeout to
> 0).

Yes.

> More of a side note to the discussion and related to the actual patch...
> if we did go with something like this ISTM to make more sense to let
> xfsaild_push_item() detect and return the flushing state (i.e. similar
> to how we handle failed state) so the accounting all exists in one
> place, rather than duplicating it for the optimization. Just my .02.

I'm not sure it does - see my comment above about ITEM_FLUSHING
being returned by inode items meaning two very different things, and
LI_FLUSHING being used to remove the noisy/useless one from the
control loop.

Cheers,

Dave.
-- 
Dave Chinner
dgc@kernel.org