Re: [PATCH 3/4] xfs: switch (back) to a per-buftarg buffer hash

[Dave Chinner <dgc@kernel.org>]

On Tue, Mar 17, 2026 at 02:40:54PM +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.
> 
> Long after writing this initially, syzbot found a problem in the buffer
> cache teardown order, which this happens to fix as well by doing the
> entire buffer cache teardown in one places instead of splitting it
> between destroying the buftarg and the perag structures.
> 
> Link: https://lore.kernel.org/linux-xfs/aLeUdemAZ5wmtZel@dread.disaster.area/
> Reported-by: syzbot+0391d34e801643e2809b@syzkaller.appspotmail.com
> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
> Tested-by: syzbot+0391d34e801643e2809b@syzkaller.appspotmail.com
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  fs/xfs/libxfs/xfs_ag.c | 13 ++---------
>  fs/xfs/libxfs/xfs_ag.h |  2 --
>  fs/xfs/xfs_buf.c       | 51 +++++++++++-------------------------------
>  fs/xfs/xfs_buf.h       | 10 +--------
>  fs/xfs/xfs_buf_mem.c   | 11 ++-------
>  5 files changed, 18 insertions(+), 69 deletions(-)

Looks fine from a logic POV - the LRU life cycle now sits inside the
hash table life cycle.

I'd also suggest that the minimum size of the rhashtable should be
increased - it will always have a higher minimum population as a
global table as a set of per-ag tables. We should try to avoid
thrashing on resizing when the filesystem is mostly idle and/or
under memory pressure, so I think a larger min size should be
specified along with this globalisation change...

-Dave.
-- 
Dave Chinner
dgc@kernel.org