Re: [PATCH 3/3] xfs: switch (back) to a per-buftarg buffer hash

["Darrick J. Wong" <djwong@kernel.org>]

On Tue, Jan 20, 2026 at 08:06:15AM +0100, Christoph Hellwig wrote:
> On Mon, Jan 19, 2026 at 06:39:18PM -0800, Darrick J. Wong wrote:
> > On Mon, Jan 19, 2026 at 04:31:37PM +0100, Christoph Hellwig wrote:
> > > The per-AG buffer hashes were added when all buffer lookups took a
> > > per-hash look.  Since then we've made lookups entirely lockless and
> > > removed the need for a hash-wide lock for inserts and removals as
> > > well.  With this there is no need to sharding the hash, so reduce the
> > > used resources by using a per-buftarg hash for all buftargs.
> > 
> > Hey, not having all the per-ag buffer cache sounds neat!
> > 
> > > Long after writing this initially, syzbot found a problem in the
> > > buffer cache teardown order, which this happens to fix as well.
> > 
> > What did we get wrong, specifically?
> 
> Dave has a really good analysis here:
> 
> https://lore.kernel.org/linux-xfs/aLeUdemAZ5wmtZel@dread.disaster.area/

Can you Link: to that in this patch?  If I'm reading linus' most recent
exposition correctly, links to other threads are still allowed.

> > Also: Is there a simpler fix for this bug that we can stuff into old lts
> > kernels?
> 
> I can't really think of anything much simpler, just different.  It would
> require some careful reordering of the unmount path, which is always
> hairy.
> 
> > Or is this fix independent of the b_hold and lockref changes
> > in the previous patches?
> 
> In theory it is, except that the old tricks with the refcount would
> make it very difficult.  I tried to reorder this twice and failed
> both times.

<nod> I figured that might be the case seeing as you cleaned up the
confusing b_hold rules.

--D