GFS2: Ensure workqueue is scheduled after noexp request
authorBob Peterson <rpeterso@redhat.com>
Wed, 12 Mar 2014 14:32:20 +0000 (10:32 -0400)
committerSteven Whitehouse <swhiteho@redhat.com>
Wed, 12 Mar 2014 14:45:48 +0000 (14:45 +0000)
This patch closes a small timing window whereby a request to hold the
transaction glock can get stuck. The problem is that after the DLM has
granted the lock, it can get into a state whereby it doesn't transition
the glock to a held state, due to not having requeued the glock state
machine to finish the transition.

Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
fs/gfs2/glock.c

index 52f747858f553325704d7d9305ff8a298bc70094..aec7f73832f080c3fd5288cdc3722162b72257bd 100644 (file)
@@ -1047,9 +1047,13 @@ int gfs2_glock_nq(struct gfs2_holder *gh)
 
        spin_lock(&gl->gl_spin);
        add_to_queue(gh);
-       if ((LM_FLAG_NOEXP & gh->gh_flags) &&
-           test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
+       if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
+                    test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
                set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
+               gl->gl_lockref.count++;
+               if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
+                       gl->gl_lockref.count--;
+       }
        run_queue(gl, 1);
        spin_unlock(&gl->gl_spin);