All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH] kernfs: clear i_nlink before deactivating removed nodes
@ 2026-11-20 20:59 Christian Brauner
  2026-11-21  1:21 ` Shakeel Butt
  0 siblings, 1 reply; 2+ messages in thread
From: Christian Brauner @ 2026-11-20 20:59 UTC (permalink / raw)
  To: linux-fsdevel
  Cc: Greg Kroah-Hartman, Tejun Heo, T.J. Mercier, driver-core,
	Shakeel Butt, Jori Koolstra, Ian Kent, Meta kernel team, stable,
	Christian Brauner (Amutable)

__kernfs_remove() deactivates the whole subtree first and only clears
i_nlink on a node's inodes later, once it unlinks the node from its
parent. In between kernfs_dop_revalidate() already fails for the node,
so a lookup that runs into one of its cached dentries kills the dentry
while i_nlink is still 1. For a file of a cgroup that is being removed:

  rmdir("cg")                          stat("cg/cgroup.events")
  css_clear_dir()
    __kernfs_remove(cgroup.events)
      deactivate the node
                                       kernfs_dop_revalidate() == 0
                                       d_invalidate()
                                       dput()
                                         dentry_unlink_inode()
                                           i_nlink == 1, no IN_DELETE_SELF
      kernfs_unlink_sibling()
      kernfs_clear_inode_nlink()
        i_nlink == 0, but the inode has no dentry left

The inotify watch keeps the inode alive, so nothing ever sends
IN_DELETE_SELF for it and the watch only goes away once the watcher
closes the inotify fd.

The window covers everything between the two loops in
__kernfs_remove(). kernfs_drain() drops kernfs_rwsem and waits for the
active references and open files of the node, and the clear itself
waits for kernfs_iattr_rwsem behind every concurrent permission check
and stat(). In a VM with a few threads stat()ing a cgroup and one of
its files while the cgroup is removed, IN_DELETE_SELF for the watched
file got lost in 16 to 33 out of 150 rmdirs, and it didn't show up even
after dropping the dcache.

So clear i_nlink in the loop that marks the subtree as being removed,
before a node turns inactive. kernfs_rwsem is held for writing there, so
no new inode can show up for these nodes while the loop runs, and once
a node is inactive kernfs_get_inode_locked() refuses to attach one. Two
overlapping removals may clear the same inode twice, which is fine.

Fixes: eea5d2bb34ba ("kernfs: Send IN_DELETE_SELF and IN_IGNORED")
Cc: stable@vger.kernel.org # v7.1+
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
---
 fs/kernfs/dir.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/fs/kernfs/dir.c b/fs/kernfs/dir.c
index fb1dcf35d900d..aab5f5f12036d 100644
--- a/fs/kernfs/dir.c
+++ b/fs/kernfs/dir.c
@@ -1604,10 +1604,18 @@ static void __kernfs_remove(struct kernfs_node *kn)
 
 	pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn));
 
-	/* prevent new usage by marking all nodes removing and deactivating */
+	/*
+	 * Prevent new usage by marking all nodes removing and deactivating.
+	 *
+	 * Clear i_nlink before a node turns inactive.  From then on
+	 * kernfs_dop_revalidate() fails and a lookup can kill the node's
+	 * dentries, and only a dentry dying with i_nlink at 0 sends
+	 * IN_DELETE_SELF.
+	 */
 	down_write(&kernfs_root(kn)->kernfs_iattr_rwsem);
 	pos = NULL;
 	while ((pos = kernfs_next_descendant_post(pos, kn))) {
+		kernfs_clear_inode_nlink(pos);
 		pos->flags |= KERNFS_REMOVING;
 		if (kernfs_active(pos))
 			atomic_add(KN_DEACTIVATED_BIAS, &pos->active);
@@ -1638,8 +1646,6 @@ static void __kernfs_remove(struct kernfs_node *kn)
 
 			down_write(&kernfs_root(kn)->kernfs_iattr_rwsem);
 
-			kernfs_clear_inode_nlink(pos);
-
 			/* update timestamps on the parent */
 			if (ps_iattr) {
 				ktime_get_real_ts64(&ps_iattr->ia_ctime);

---
base-commit: ccf2a4692fed0fa77751d32d2690585127a152c5
change-id: 20261120-work-kernfs-delete_self-a1fdb64ebcce


^ permalink raw reply related	[flat|nested] 2+ messages in thread

* Re: [PATCH] kernfs: clear i_nlink before deactivating removed nodes
  2026-11-20 20:59 [PATCH] kernfs: clear i_nlink before deactivating removed nodes Christian Brauner
@ 2026-11-21  1:21 ` Shakeel Butt
  0 siblings, 0 replies; 2+ messages in thread
From: Shakeel Butt @ 2026-11-21  1:21 UTC (permalink / raw)
  To: Christian Brauner
  Cc: linux-fsdevel, Greg Kroah-Hartman, Tejun Heo, T.J. Mercier,
	driver-core, Jori Koolstra, Ian Kent, Meta kernel team, stable

On Fri, Nov 20, 2026 at 09:59:59PM +0100, Christian Brauner wrote:
> __kernfs_remove() deactivates the whole subtree first and only clears
> i_nlink on a node's inodes later, once it unlinks the node from its
> parent. In between kernfs_dop_revalidate() already fails for the node,
> so a lookup that runs into one of its cached dentries kills the dentry
> while i_nlink is still 1. For a file of a cgroup that is being removed:
> 
>   rmdir("cg")                          stat("cg/cgroup.events")
>   css_clear_dir()
>     __kernfs_remove(cgroup.events)
>       deactivate the node
>                                        kernfs_dop_revalidate() == 0
>                                        d_invalidate()
>                                        dput()
>                                          dentry_unlink_inode()
>                                            i_nlink == 1, no IN_DELETE_SELF
>       kernfs_unlink_sibling()
>       kernfs_clear_inode_nlink()
>         i_nlink == 0, but the inode has no dentry left
> 
> The inotify watch keeps the inode alive, so nothing ever sends
> IN_DELETE_SELF for it and the watch only goes away once the watcher
> closes the inotify fd.
> 
> The window covers everything between the two loops in
> __kernfs_remove(). kernfs_drain() drops kernfs_rwsem and waits for the
> active references and open files of the node, and the clear itself
> waits for kernfs_iattr_rwsem behind every concurrent permission check
> and stat(). In a VM with a few threads stat()ing a cgroup and one of
> its files while the cgroup is removed, IN_DELETE_SELF for the watched
> file got lost in 16 to 33 out of 150 rmdirs, and it didn't show up even
> after dropping the dcache.
> 
> So clear i_nlink in the loop that marks the subtree as being removed,
> before a node turns inactive. kernfs_rwsem is held for writing there, so
> no new inode can show up for these nodes while the loop runs, and once
> a node is inactive kernfs_get_inode_locked() refuses to attach one. Two
> overlapping removals may clear the same inode twice, which is fine.
> 
> Fixes: eea5d2bb34ba ("kernfs: Send IN_DELETE_SELF and IN_IGNORED")
> Cc: stable@vger.kernel.org # v7.1+
> Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>

Reviewed-by: Shakeel Butt <shakeel.butt@linux.dev>

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-21  1:21 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-20 20:59 [PATCH] kernfs: clear i_nlink before deactivating removed nodes Christian Brauner
2026-11-21  1:21 ` Shakeel Butt

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.