Prior to commit 4149be7bda7e, sys_flock() would allocate the file_lock
struct it was going to use to pass parameters, call ->flock() and then call
locks_free_lock() to get rid of it - which had the side effect of calling
locks_release_private() and thus ->fl_release_private().
With commit 4149be7bda7e, however, this is no longer the case: the struct
is now allocated on the stack, and locks_free_lock() is no longer called -
and thus any remaining private data doesn't get cleaned up either.
This causes afs flock to cause oops. Kasan catches this as a UAF by the
list_del_init() in afs_fl_release_private() for the file_lock record
produced by afs_fl_copy_lock() as the original record didn't get delisted.
It can be reproduced using the generic/504 xfstest.
Fix this by reinstating the locks_release_private() call in sys_flock().
I'm not sure if this would affect any other filesystems. If not, then the
release could be done in afs_flock() instead.
Changes
=======
ver #2)
- Don't need to call ->fl_release_private() after calling the security
hook, only after calling ->flock().
Fixes: 4149be7bda7e ("fs/lock: Don't allocate file_lock in flock_make_lock().")
cc: Kuniyuki Iwashima <kuniyu@xxxxxxxxxx>
cc: Chuck Lever <chuck.lever@xxxxxxxxxx>
cc: Jeff Layton <jlayton@xxxxxxxxxx>
cc: Marc Dionne <marc.dionne@xxxxxxxxxxxx>
cc: linux-afs@xxxxxxxxxxxxxxxxxxx
cc: linux-fsdevel@xxxxxxxxxxxxxxx
Link: https://lore.kernel.org/r/166075758809.3532462.13307935588777587536.stgit@xxxxxxxxxxxxxxxxxxxxxx/ # v1
---
fs/locks.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/fs/locks.c b/fs/locks.c
index c266cfdc3291..607f94a0e789 100644
--- a/fs/locks.c
+++ b/fs/locks.c
@@ -2129,6 +2129,7 @@ SYSCALL_DEFINE2(flock, unsigned int, fd, unsigned int, cmd)
else
error = locks_lock_file_wait(f.file, &fl);
+ locks_release_private(&fl);
out_putf:
fdput(f);
