Re: [PATCH 12/12] xfs: add lockless xfs_buf_readahead_map fast path

On Wed, Jul 15, 2026 at 04:51:05PM +0200, Christoph Hellwig wrote:
> Readahead currently always locks the buffer, which can cause contention
> with actual users of the buffer.  Add a fast path without taking any
> locks if the buffer is uptodate and not stale.

So we're trying to cut down on b_sema contention for detecting the case
where the buffer is already uptodate?  And I guess the idea here is that
if you had per-fsblock filesystem metadata, you'd want to be able to
readahead that metadata into memory during setup for file IO without
so much locking?

Sounds reasonable to me...
Reviewed-by: "Darrick J. Wong" <[email protected]>

--D


> Signed-off-by: Christoph Hellwig <[email protected]>
> ---
>  fs/xfs/xfs_buf.c | 16 +++++++++++++---
>  1 file changed, 13 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> index f5701f74061f..2e498e9e05e9 100644
> --- a/fs/xfs/xfs_buf.c
> +++ b/fs/xfs/xfs_buf.c
> @@ -756,13 +756,23 @@ xfs_buf_readahead_map(
>  
>  	if (xfs_find_get_buf(target, map, nmaps, flags, &bp))
>  		return;
> -	if (xfs_buf_find_lock(bp, XBF_TRYLOCK))
> +
> +	/*
> +	 * Do a lockless fast path check for a valid uptodate buffer and avoid
> +	 * locking entirely in this case.
> +	 */
> +	if ((READ_ONCE(bp->b_flags) & (XBF_DONE | XBF_STALE)) == XBF_DONE)
>  		goto out_rele;
>  
> -	trace_xfs_buf_readahead(bp, 0, _RET_IP_);
> -	if (bp->b_flags & XBF_DONE)
> +	/* Otherwise lock the buffer to stabilize the state */
> +	if (!xfs_buf_trylock(bp))
> +		goto out_rele;
> +
> +	/* Let the actual reader deal with stale buffers. */
> +	if (bp->b_flags & (XBF_STALE | XBF_DONE))
>  		goto out_unlock;
>  
> +	trace_xfs_buf_readahead(bp, 0, _RET_IP_);
>  	XFS_STATS_INC(target->bt_mount, xb_get_read);
>  	bp->b_ops = ops;
>  	xfs_buf_clear_flags(bp, XBF_WRITE | XBF_DONE);
> -- 
> 2.53.0
> 
>