CacheFiles: Fix a race in cachefiles_delete_object() vs rename

cachefiles_delete_object() can race with rename.  It gets the parent directory
of the object it's asked to delete, then locks it - but rename may have changed
the object's parent between the get and the completion of the lock.

However, if such a circumstance is detected, we abandon our attempt to delete
the object - since it's no longer in the index key path, it won't be seen
again by lookups of that key.  The assumption is that cachefilesd may have
culled it by renaming it to the graveyard for later destruction.

Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

diff --git a/fs/cachefiles/namei.c b/fs/cachefiles/namei.c
index 14ac4806e2913f8c37e76d95dc1d0274f4bbd027..eeb4986ea7db7cbb65b0b48685ae0a0d3f6a1a7e 100644 (file)
--- a/fs/cachefiles/namei.c
+++ b/fs/cachefiles/namei.c
@@ -348,7 +348,17 @@ int cachefiles_delete_object(struct cachefiles_cache *cache,
        dir = dget_parent(object->dentry);
 
        mutex_lock_nested(&dir->d_inode->i_mutex, I_MUTEX_PARENT);
-       ret = cachefiles_bury_object(cache, dir, object->dentry);
+
+       /* we need to check that our parent is _still_ our parent - it may have
+        * been renamed */
+       if (dir == object->dentry->d_parent) {
+               ret = cachefiles_bury_object(cache, dir, object->dentry);
+       } else {
+               /* it got moved, presumably by cachefilesd culling it, so it's
+                * no longer in the key path and we can ignore it */
+               mutex_unlock(&dir->d_inode->i_mutex);
+               ret = 0;
+       }
 
        dput(dir);
        _leave(" = %d", ret);