Re: [PATCH 1/3] xfs: don't keep a reference for buffers on the LRU

[Christoph Hellwig <hch@lst.de>]

On Mon, Jan 19, 2026 at 06:53:15PM -0800, Darrick J. Wong wrote:
> So under the current code, a cached buffer gets created with b_hold=1
> and b_lru_ref=1.  Calling xfs_buf_hold can bump up b_hold.  Calling
> xfs_buf_rele immediately will either transfer ownership of the buf to
> the lru (if it wasn't on the lru ref) or decrement b_hold.
> 
> Higher level xfs code might boost b_lru_ref for buffers that it thinks
> we should try to hang on to (e.g. AG headers).

We actually do this b_lru_ref boost for a lot, if not most of the
metadata buffers.  It is used to manage relative eviction priority.

> xfs_buftarg_isolate will decrement b_lru_ref unless it was already zero,
> in which case it'll actually free the buffer.  If xfs_buf_rele finds a
> buffer with b_lru_ref==0 it'll drop b_hold and try to free the buffer if
> b_hold drops to zero.
> 
> Right?

Yes.

> > Switch to not having a reference for buffers in the LRU, and use a
> > separate negative hold value to mark buffers as dead.  This simplifies
> > xfs_buf_rele, which now just deal with the last "real" reference,
> > and prepares for using the lockref primitive.
> 
> And now, b_hold is the number of higher-level owners of the buffer.  If
> that drops to zero the buffer gets put on the lru list if it hasn't
> already run out of lru refs, in which case it's freed directly.  If a
> buffer on the lru list runs out of lru refs then it'll get freed.
> b_hodl < 0 means "headed for destruction".
> 
> Is that also correct?

Yes.