jbd2: fix theoretical race in jbd2__journal_restart

Once we decrement transaction->t_updates, if this is the last handle
holding the transaction from closing, and once we release the
t_handle_lock spinlock, it's possible for the transaction to commit
and be released.  In practice with normal kernels, this probably won't
happen, since the commit happens in a separate kernel thread and it's
unlikely this could all happen within the space of a few CPU cycles.

On the other hand, with a real-time kernel, this could potentially
happen, so save the tid found in transaction->t_tid before we release
t_handle_lock.  It would require an insane configuration, such as one
where the jbd2 thread was set to a very high real-time priority,
perhaps because a high priority real-time thread is trying to read or
write to a file system.  But some people who use real-time kernels
have been known to do insane things, including controlling
laser-wielding industrial robots.  :-)

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index dd422e6804183a5539f0174ce5ac004c2cb05956..383b0fbc6e19da3b1c218e5c3f81a2a715768baf 100644 (file)
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -652,10 +652,10 @@ int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
        }
        if (atomic_dec_and_test(&transaction->t_updates))
                wake_up(&journal->j_wait_updates);
+       tid = transaction->t_tid;
        spin_unlock(&transaction->t_handle_lock);
 
        jbd_debug(2, "restarting handle %p\n", handle);
-       tid = transaction->t_tid;
        need_to_start = !tid_geq(journal->j_commit_request, tid);
        read_unlock(&journal->j_state_lock);
        if (need_to_start)