[PATCH v2] dcache: touch up predicts in __d_lookup_rcu()

[PATCH v2] dcache: touch up predicts in __d_lookup_rcu()

[Mateusz Guzik]

Rationale is that if the parent dentry is the same and the length is the
same, then you have to be unlucky for the name to not match.

At the same time the dentry was literally just found on the hash, so you
have to be even more unlucky to determine it is unhashed.

While here add commentary while d_unhashed() is necessary. It was
already removed once and brought back in:
2e321806b681b192 ("Revert "vfs: remove unnecessary d_unhashed() check from __d_lookup_rcu"")

Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
---

- move and precit on d_unhashed as well
- add commentary on it

this obsoletes https://lore.kernel.org/linux-fsdevel/20251127122412.4131818-1-mjguzik@gmail.com/T/#u

 fs/dcache.c | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index 23d1752c29e6..dc2fff4811d1 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2342,11 +2342,20 @@ struct dentry *__d_lookup_rcu(const struct dentry *parent,
 		seq = raw_seqcount_begin(&dentry->d_seq);
 		if (dentry->d_parent != parent)
 			continue;
-		if (d_unhashed(dentry))
-			continue;
 		if (dentry->d_name.hash_len != hashlen)
 			continue;
-		if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
+		if (unlikely(dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0))
+			continue;
+		/*
+		 * Check for the dentry being unhashed.
+		 *
+		 * As tempting as it is, we *can't* skip it because of a race window
+		 * between us finding the dentry before it gets unhashed and loading
+		 * the sequence counter after unhashing is finished.
+		 *
+		 * We can at least predict on it.
+		 */
+		if (unlikely(d_unhashed(dentry)))
 			continue;
 		*seqp = seq;
 		return dentry;
-- 
2.34.1

Re: [PATCH v2] dcache: touch up predicts in __d_lookup_rcu()

[Jan Kara]

On Thu 27-11-25 14:15:26, Mateusz Guzik wrote:
> Rationale is that if the parent dentry is the same and the length is the
> same, then you have to be unlucky for the name to not match.
> 
> At the same time the dentry was literally just found on the hash, so you
> have to be even more unlucky to determine it is unhashed.
> 
> While here add commentary while d_unhashed() is necessary. It was
> already removed once and brought back in:
> 2e321806b681b192 ("Revert "vfs: remove unnecessary d_unhashed() check from __d_lookup_rcu"")
> 
> Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>

Looks good. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
> 
> - move and precit on d_unhashed as well
> - add commentary on it
> 
> this obsoletes https://lore.kernel.org/linux-fsdevel/20251127122412.4131818-1-mjguzik@gmail.com/T/#u
> 
>  fs/dcache.c | 15 ++++++++++++---
>  1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index 23d1752c29e6..dc2fff4811d1 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -2342,11 +2342,20 @@ struct dentry *__d_lookup_rcu(const struct dentry *parent,
>  		seq = raw_seqcount_begin(&dentry->d_seq);
>  		if (dentry->d_parent != parent)
>  			continue;
> -		if (d_unhashed(dentry))
> -			continue;
>  		if (dentry->d_name.hash_len != hashlen)
>  			continue;
> -		if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
> +		if (unlikely(dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0))
> +			continue;
> +		/*
> +		 * Check for the dentry being unhashed.
> +		 *
> +		 * As tempting as it is, we *can't* skip it because of a race window
> +		 * between us finding the dentry before it gets unhashed and loading
> +		 * the sequence counter after unhashing is finished.
> +		 *
> +		 * We can at least predict on it.
> +		 */
> +		if (unlikely(d_unhashed(dentry)))
>  			continue;
>  		*seqp = seq;
>  		return dentry;
> -- 
> 2.34.1
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [PATCH v2] dcache: touch up predicts in __d_lookup_rcu()

[Christian Brauner]

On Thu, 27 Nov 2025 14:15:26 +0100, Mateusz Guzik wrote:
> Rationale is that if the parent dentry is the same and the length is the
> same, then you have to be unlucky for the name to not match.
> 
> At the same time the dentry was literally just found on the hash, so you
> have to be even more unlucky to determine it is unhashed.
> 
> While here add commentary while d_unhashed() is necessary. It was
> already removed once and brought back in:
> 2e321806b681b192 ("Revert "vfs: remove unnecessary d_unhashed() check from __d_lookup_rcu"")
> 
> [...]

Applied to the vfs-6.19.inode branch of the vfs/vfs.git tree.
Patches in the vfs-6.19.inode branch should appear in linux-next soon.

Please report any outstanding bugs that were missed during review in a
new review to the original patch series allowing us to drop it.

It's encouraged to provide Acked-bys and Reviewed-bys even though the
patch has now been applied. If possible patch trailers will be updated.

Note that commit hashes shown below are subject to change due to rebase,
trailer updates or similar. If in doubt, please check the listed branch.

tree:   https://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs.git
branch: vfs-6.19.inode

[1/1] dcache: touch up predicts in __d_lookup_rcu()
      https://git.kernel.org/vfs/vfs/c/ca0d620b0afa