fsnotify: Fix busy inodes during unmount

Detaching of mark connector from fsnotify_put_mark() can race with
unmounting of the filesystem like:

  CPU1				CPU2
fsnotify_put_mark()
  spin_lock(&conn->lock);
  ...
  inode = fsnotify_detach_connector_from_object(conn)
  spin_unlock(&conn->lock);
				generic_shutdown_super()
				  fsnotify_unmount_inodes()
				    sees connector detached for inode
				      -> nothing to do
				  evict_inode()
				    barfs on pending inode reference
  iput(inode);

Resulting in "Busy inodes after unmount" message and possible kernel
oops. Make fsnotify_unmount_inodes() properly wait for outstanding inode
references from detached connectors.

Note that the accounting of outstanding inode references in the
superblock can cause some cacheline contention on the counter. OTOH it
happens only during deletion of the last notification mark from an inode
(or during unlinking of watched inode) and that is not too bad. I have
measured time to create & delete inotify watch 100000 times from 64
processes in parallel (each process having its own inotify group and its
own file on a shared superblock) on a 64 CPU machine. Average and
standard deviation of 15 runs look like:

	Avg		Stddev
Vanilla	9.817400	0.276165
Fixed	9.710467	0.228294

So there's no statistically significant difference.

Fixes: 6b3f05d24d35 ("fsnotify: Detach mark from object list when last reference is dropped")
CC: stable@vger.kernel.org
Signed-off-by: Jan Kara <jack@suse.cz>

1 files changed, 3 insertions, 0 deletions

diff --git a/include/linux/fs.h b/include/linux/fs.h
index 6da94deb957f..00b23b21e78a 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -1437,6 +1437,9 @@ struct super_block {
 	/* Number of inodes with nlink == 0 but still referenced */
 	atomic_long_t s_remove_count;
 
+	/* Pending fsnotify inode refs */
+	atomic_long_t s_fsnotify_inode_refs;
+
 	/* Being remounted read-only */
 	int s_readonly_remount;