Re: [PATCH 4/4] xfs: don't decrement the buffer LRU count for in-use buffers

[Brian Foster <bfoster@redhat.com>]

On Mon, Mar 23, 2026 at 08:50:54AM +0100, Christoph Hellwig wrote:
> XFS buffers are added to the LRU when they are unused, but are only
> removed from the LRU lazily when the LRU list scan finds a used buffer.
> So far this only happen when the LRU counter hits 0, which is suboptimal
> as buffers that were added to the LRU, but are in use again still consume
> LRU scanning resources and are aged while actually in use.
> 
> Fix this by checking for in-use buffers and removing the from the LRU
> before decrementing the LRU counter.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>
> ---

Reviewed-by: Brian Foster <bfoster@redhat.com>

>  fs/xfs/xfs_buf.c | 22 ++++++++++++----------
>  1 file changed, 12 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> index e4b65d0c9ef0..ee8c3944015a 100644
> --- a/fs/xfs/xfs_buf.c
> +++ b/fs/xfs/xfs_buf.c
> @@ -1524,23 +1524,25 @@ xfs_buftarg_isolate(
>  		return LRU_SKIP;
>  
>  	/*
> -	 * Decrement the b_lru_ref count unless the value is already
> -	 * zero. If the value is already zero, we need to reclaim the
> -	 * buffer, otherwise it gets another trip through the LRU.
> +	 * If the buffer is in use, remove it from the LRU for now.  We can't
> +	 * free it while someone is using it, and we should also not count
> +	 * eviction passed for it, just as if it hadn't been added to the LRU
> +	 * yet.
>  	 */
> -	if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
> +	if (bp->b_lockref.count > 0) {
> +		list_lru_isolate(lru, &bp->b_lru);
>  		spin_unlock(&bp->b_lockref.lock);
> -		return LRU_ROTATE;
> +		return LRU_REMOVED;
>  	}
>  
>  	/*
> -	 * If the buffer is in use, remove it from the LRU for now as we can't
> -	 * free it.  It will be freed when the last reference drops.
> +	 * Decrement the b_lru_ref count unless the value is already
> +	 * zero. If the value is already zero, we need to reclaim the
> +	 * buffer, otherwise it gets another trip through the LRU.
>  	 */
> -	if (bp->b_lockref.count > 0) {
> -		list_lru_isolate(lru, &bp->b_lru);
> +	if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
>  		spin_unlock(&bp->b_lockref.lock);
> -		return LRU_REMOVED;
> +		return LRU_ROTATE;
>  	}
>  
>  	lockref_mark_dead(&bp->b_lockref);
> -- 
> 2.47.3
> 
>