[PATCH 0/2] dcache: don't leave negative dentries in casefolded directories

[PATCH 0/2] dcache: don't leave negative dentries in casefolded directories

[Christian Brauner]

ext4, f2fs and tmpfs keep negative dentries out of casefolded
directories. A negative dentry remembers one spelling of the name and
the next create() through it would write that spelling to disk.

So their ->lookup() doesn't cache a miss and their ->unlink() and
->rmdir() call d_invalidate() on the dentry so that d_delete() doesn't
turn it into a negative dentry.

But d_invalidate() is the wrong primitive for that. It's for a name that
went stale. It walks the subtree, prunes it and detaches every mount it
finds, under namespace_sem, and here that runs from inside the
filesystem's own directory operation. ext4_rmdir() does it with the
journal handle still open and ext4_unlink() does it even when the
unlink failed. All anyone wanted is the unhash.

So d_delete() now unhashes the dentry when the parent is casefolded,
next to the dentry-negative sysctl that does the same for every
directory. ext4, f2fs and tmpfs drop their d_invalidate() calls.

xfs's ascii-ci doesn't set S_CASEFOLD and can't so it d_drop()s instead.
I did consider setting S_CASEFOLD on ascii-ci directories and didn't.
overlayfs warns when a layer's casefold flag disagrees with the
superblock having an encoding, and xfs has none.

Nothing changes for userspace. A miss still isn't cached and a removed
name still leaves no negative dentry behind. Verified with the same
create, unlink, create, readdir and stat sequence in a plain and in a
casefolded directory on tmpfs and on ext4, before and after.

Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
---
Christian Brauner (2):
      dcache: don't leave negative dentries in casefolded directories
      xfs: unhash a removed name in ascii-ci mode instead of invalidating it

 fs/dcache.c       |  8 +++++++-
 fs/ext4/namei.c   | 18 ------------------
 fs/f2fs/namei.c   |  9 ---------
 fs/xfs/xfs_iops.c |  4 ++--
 mm/shmem.c        |  7 -------
 5 files changed, 9 insertions(+), 37 deletions(-)
---
base-commit: 7be4bc52fd9e0e55dc9a2e0232b89427d2c4b614
change-id: 20261023-work-dcache-casefold-f26205ee1bb7

[PATCH 1/2] dcache: don't leave negative dentries in casefolded directories

[Christian Brauner]

ext4, f2fs and tmpfs keep negative dentries out of casefolded
directories. A negative dentry remembers one spelling of the name and
the next create() through it would write that spelling to disk.

So their ->lookup() doesn't cache a miss and their ->unlink() and
->rmdir() call d_invalidate() on the dentry so that d_delete() doesn't
turn it into a negative dentry.

But d_invalidate() is the wrong primitive for that. It's for a name that
went stale. It walks the subtree, prunes it and detaches every mount it
finds, under namespace_sem, and here that runs from inside the
filesystem's own directory operation. ext4_rmdir() does it with the
journal handle still open and ext4_unlink() does it even when the
unlink failed. All anyone wanted is the unhash.

So d_delete() now unhashes the dentry when the parent is casefolded,
next to the dentry-negative sysctl that does the same for every
directory. ext4, f2fs and tmpfs drop their d_invalidate() calls.

Nothing changes for a casefolded directory. A miss still isn't cached
and a removed name still doesn't leave a negative dentry behind. The
target of a rename over an existing name, which tmpfs also passes
through shmem_unlink(), doesn't turn negative either. vfs_rename()
detaches its mounts and d_move() unhashes it. A failed unlink or rmdir
now leaves the dentry hashed and its mounts in place, as in any other
directory.

Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
---
 fs/dcache.c     |  8 +++++++-
 fs/ext4/namei.c | 18 ------------------
 fs/f2fs/namei.c |  9 ---------
 mm/shmem.c      |  7 -------
 4 files changed, 7 insertions(+), 35 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index efd9332c8407..f5cbb2e24e53 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -2746,6 +2746,12 @@ struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
 	return d_lookup(dir, name);
 }
 
+static __always_inline bool d_delete_drop(struct dentry *dentry)
+{
+	return dentry_negative_policy ||
+	       IS_CASEFOLDED(d_inode(dentry->d_parent));
+}
+
 /*
  * When a file is deleted, we have two options:
  * - turn this dentry into a negative dentry
@@ -2777,7 +2783,7 @@ void d_delete(struct dentry * dentry)
 	 * Are we the only user?
 	 */
 	if (dentry->d_lockref.count == 1) {
-		if (dentry_negative_policy)
+		if (d_delete_drop(dentry))
 			__d_drop(dentry);
 		dentry->d_flags &= ~DCACHE_CANT_MOUNT;
 		dentry_unlink_inode(dentry);
diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
index 6e0630a49e48..c4b1121471e6 100644
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c
@@ -3209,15 +3209,6 @@ static int ext4_rmdir(struct inode *dir, struct dentry *dentry)
 	ext4_fc_track_unlink(handle, dentry);
 	retval = ext4_mark_inode_dirty(handle, dir);
 
-	/* VFS negative dentries are incompatible with Encoding and
-	 * Case-insensitiveness. Eventually we'll want avoid
-	 * invalidating the dentries here, alongside with returning the
-	 * negative dentries at ext4_lookup(), when it is better
-	 * supported by the VFS for the CI case.
-	 */
-	if (IS_ENABLED(CONFIG_UNICODE) && IS_CASEFOLDED(dir))
-		d_invalidate(dentry);
-
 end_rmdir:
 	brelse(bh);
 	if (handle)
@@ -3320,15 +3311,6 @@ static int ext4_unlink(struct inode *dir, struct dentry *dentry)
 
 	retval = __ext4_unlink(dir, &dentry->d_name, d_inode(dentry), dentry);
 
-	/* VFS negative dentries are incompatible with Encoding and
-	 * Case-insensitiveness. Eventually we'll want avoid
-	 * invalidating the dentries here, alongside with returning the
-	 * negative dentries at ext4_lookup(), when it is  better
-	 * supported by the VFS for the CI case.
-	 */
-	if (IS_ENABLED(CONFIG_UNICODE) && IS_CASEFOLDED(dir))
-		d_invalidate(dentry);
-
 out_trace:
 	trace_ext4_unlink_exit(dentry, retval);
 	return retval;
diff --git a/fs/f2fs/namei.c b/fs/f2fs/namei.c
index 233c07fec171..30f6b8448eee 100644
--- a/fs/f2fs/namei.c
+++ b/fs/f2fs/namei.c
@@ -631,15 +631,6 @@ static int f2fs_unlink(struct inode *dir, struct dentry *dentry)
 	if (err)
 		goto out;
 
-	/* VFS negative dentries are incompatible with Encoding and
-	 * Case-insensitiveness. Eventually we'll want avoid
-	 * invalidating the dentries here, alongside with returning the
-	 * negative dentries at f2fs_lookup(), when it is better
-	 * supported by the VFS for the CI case.
-	 */
-	if (IS_ENABLED(CONFIG_UNICODE) && IS_CASEFOLDED(dir))
-		d_invalidate(dentry);
-
 	if (IS_DIRSYNC(dir))
 		err = f2fs_sync_fs(F2FS_I_SB(dir)->sb, 1);
 out:
diff --git a/mm/shmem.c b/mm/shmem.c
index 0d891ade18b5..5a353ac32573 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -3953,13 +3953,6 @@ static int shmem_unlink(struct inode *dir, struct dentry *dentry)
 	inode_inc_iversion(dir);
 	simple_unlink(dir, dentry);
 
-	/*
-	 * For now, VFS can't deal with case-insensitive negative dentries, so
-	 * we invalidate them
-	 */
-	if (IS_ENABLED(CONFIG_UNICODE) && IS_CASEFOLDED(dir))
-		d_invalidate(dentry);
-
 	return 0;
 }
 

-- 
2.53.0

[PATCH 2/2] xfs: unhash a removed name in ascii-ci mode instead of invalidating it

[Christian Brauner]

xfs_vn_unlink() uses d_invalidate() on the dentry of a removed name on a
case-insensitive filesystem, so that d_delete() doesn't leave a negative
dentry behind. The next create() would otherwise reuse with the wrong
spelling. But as the comment already states all it wants is the unhash.

But d_invalidate() is the wrong primitive for that. It's for a name that
went stale. It walks the subtree, prunes it and detaches every mount it
finds, under namespace_sem, and here that runs from inside the
filesystem's own directory operation. There is nothing below a file or
an empty directory to prune, and vfs_unlink() and vfs_rmdir() detach the
mounts a few lines later as for every other filesystem.

The previous patch moved ext4, f2fs and tmpfs over to d_delete()
unhashing the dentry for them. ascii-ci doesn't set S_CASEFOLD, so the
VFS can't tell these directories apart, and the feature is deprecated
anyway. So just d_drop() the dentry. d_delete() finds it unhashed and
turns it negative without hashing it, exactly as before.

Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
---
 fs/xfs/xfs_iops.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/xfs/xfs_iops.c b/fs/xfs/xfs_iops.c
index d307ec97fac8..bed8372e4bcb 100644
--- a/fs/xfs/xfs_iops.c
+++ b/fs/xfs/xfs_iops.c
@@ -416,10 +416,10 @@ xfs_vn_unlink(
 	/*
 	 * With unlink, the VFS makes the dentry "negative": no inode,
 	 * but still hashed. This is incompatible with case-insensitive
-	 * mode, so invalidate (unhash) the dentry in CI-mode.
+	 * mode, so unhash the dentry in CI-mode.
 	 */
 	if (xfs_has_asciici(XFS_M(dir->i_sb)))
-		d_invalidate(dentry);
+		d_drop(dentry);
 	return 0;
 }
 

-- 
2.53.0

Re: [PATCH 1/2] dcache: don't leave negative dentries in casefolded directories

[Eric Biggers]

On Fri, Oct 23, 2026 at 02:44:22PM +0200, Christian Brauner wrote:
> +static __always_inline bool d_delete_drop(struct dentry *dentry)
> +{
> +	return dentry_negative_policy ||
> +	       IS_CASEFOLDED(d_inode(dentry->d_parent));
> +}

Should this also check IS_ENABLED(CONFIG_UNICODE), so there will be no
overhead when CONFIG_UNICODE=n?

- Eric