]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/commitdiff
locks: use cmpxchg to assign i_flctx pointer
authorJeff Layton <jeff.layton@primarydata.com>
Fri, 3 Apr 2015 13:04:04 +0000 (09:04 -0400)
committerJeff Layton <jeff.layton@primarydata.com>
Fri, 3 Apr 2015 13:04:04 +0000 (09:04 -0400)
During the v3.20/v4.0 cycle, I had originally had the code manage the
inode->i_flctx pointer using a compare-and-swap operation instead of the
i_lock.

Sasha Levin though hit a problem while testing with trinity that made me
believe that that wasn't safe. At the time, changing the code to protect
the i_flctx pointer seemed to fix the issue, but I now think that was
just coincidence.

The issue was likely the same race that Kirill Shutemov hit while
testing the pre-rc1 v4.0 kernel and that Linus spotted. Due to the way
that the spinlock was dropped in the middle of flock_lock_file, you
could end up with multiple flock locks for the same struct file on the
inode.

Reinstate the use of a CAS operation to assign this pointer since it's
likely to be more efficient and gets the i_lock completely out of the
file locking business.

Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
fs/locks.c

index 16cae1a00851bb7ef56d33522308b284c6f8424a..52b780fb5258ba0bea1b06edad7586a62a900f41 100644 (file)
@@ -223,14 +223,7 @@ locks_get_lock_context(struct inode *inode, int type)
         * Assign the pointer if it's not already assigned. If it is, then
         * free the context we just allocated.
         */
-       spin_lock(&inode->i_lock);
-       if (likely(!inode->i_flctx)) {
-               inode->i_flctx = new;
-               new = NULL;
-       }
-       spin_unlock(&inode->i_lock);
-
-       if (new)
+       if (cmpxchg(&inode->i_flctx, NULL, new))
                kmem_cache_free(flctx_cache, new);
 out:
        return inode->i_flctx;