]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
ksmbd: fix use-after-free of a deferred file_lock on SMB2_CLOSE then SMB2_CANCEL
authorGil Portnoy <dddhkts1@gmail.com>
Wed, 10 Jun 2026 10:53:14 +0000 (19:53 +0900)
committerSteve French <stfrench@microsoft.com>
Tue, 16 Jun 2026 23:57:21 +0000 (18:57 -0500)
Commit f580d27e8928 ("ksmbd: fix use-after-free of a deferred file_lock on
double SMB2_CANCEL") made smb2_cancel() skip a work whose state is
KSMBD_WORK_CANCELLED, so its cancel_fn cannot be fired a second time. But
KSMBD_WORK has three states (ACTIVE, CANCELLED, CLOSED), and the same
freeing producer path is reached for CLOSED too:

  SMB2_CLOSE on the locking handle -> set_close_state_blocked_works() sets
  the deferred work's state to KSMBD_WORK_CLOSED and wakes the smb2_lock()
  worker. The worker takes the non-ACTIVE early-exit, locks_free_lock()s
  the file_lock and, because the state is not KSMBD_WORK_CANCELLED, takes
  the STATUS_RANGE_NOT_LOCKED branch with "goto out2" -- which, like the
  cancelled branch, skips release_async_work(). The work stays on
  conn->async_requests with a live cancel_fn = smb2_remove_blocked_lock
  pointing at the freed file_lock.

A subsequent SMB2_CANCEL for the same AsyncId then passes the
KSMBD_WORK_CANCELLED-only guard (its state is KSMBD_WORK_CLOSED), so
smb2_cancel() fires cancel_fn again over the freed file_lock -- the same
use-after-free fixed, via SMB2_CLOSE instead of a first SMB2_CANCEL:

  BUG: KASAN: slab-use-after-free in __locks_delete_block
    __locks_delete_block
    locks_delete_block
    ksmbd_vfs_posix_lock_unblock
    smb2_remove_blocked_lock
    smb2_cancel                 <- 2nd SMB2_CANCEL fires cancel_fn
    handle_ksmbd_work
  Allocated by ...: locks_alloc_lock <- smb2_lock
  Freed by ...:     locks_free_lock  <- smb2_lock (non-ACTIVE early-exit)
  ... cache file_lock_cache of size 192

Reproduced on mainline 7.1-rc7 (which already contains f580d27e8928) with
KASAN by an authenticated SMB client; the double-SMB2_CANCEL control is
silent on that kernel, so the splat is attributable to the CLOSE trigger.

Only an ACTIVE deferred work may have its cancel_fn fired: both terminal
states (CANCELLED and CLOSED) reach the smb2_lock() early-exit that frees
the file_lock and skips release_async_work(). Guard on KSMBD_WORK_ACTIVE
so any non-active work is skipped.

Fixes: f580d27e8928 ("ksmbd: fix use-after-free of a deferred file_lock on double SMB2_CANCEL")
Cc: stable@vger.kernel.org
Signed-off-by: Gil Portnoy <dddhkts1@gmail.com>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
fs/smb/server/smb2pdu.c

index 327ec625d42b79892e220ebdabcfa8660d090dcd..06c92af288fcb97bb25734182d44d57b6d72c23e 100644 (file)
@@ -7349,14 +7349,14 @@ int smb2_cancel(struct ksmbd_work *work)
                                continue;
 
                        /*
-                        * A cancelled deferred byte-range lock frees its
-                        * file_lock and takes the smb2_lock() early-exit that
-                        * skips release_async_work(), so the work stays on
-                        * conn->async_requests with a live cancel_fn pointing
-                        * at the freed file_lock.  Re-firing it on a second
-                        * SMB2_CANCEL is a use-after-free.
+                        * Only an ACTIVE deferred work may have its cancel_fn
+                        * fired.  A CANCELLED or CLOSED work already took the
+                        * smb2_lock() non-ACTIVE early-exit that frees the
+                        * file_lock and skips release_async_work(), so it is
+                        * still on conn->async_requests with a live cancel_fn
+                        * pointing at the freed file_lock.
                         */
-                       if (iter->state == KSMBD_WORK_CANCELLED)
+                       if (iter->state != KSMBD_WORK_ACTIVE)
                                break;
 
                        ksmbd_debug(SMB,