IB/mad: Fix possible lock-lock-timer deadlock
Roland Dreier [Mon, 7 Sep 2009 15:27:50 +0000 (08:27 -0700)]
Lockdep reported a possible deadlock with cm_id_priv->lock,
mad_agent_priv->lock and mad_agent_priv->timed_work.timer; this
happens because the mad module does

cancel_delayed_work(&mad_agent_priv->timed_work);

while holding mad_agent_priv->lock.  cancel_delayed_work() internally
does del_timer_sync(&mad_agent_priv->timed_work.timer).

This can turn into a deadlock because mad_agent_priv->lock is taken
inside cm_id_priv->lock, so we can get the following set of contexts
that deadlock each other:

 A: holding cm_id_priv->lock, waiting for mad_agent_priv->lock
 B: holding mad_agent_priv->lock, waiting for del_timer_sync()
 C: interrupt during mad_agent_priv->timed_work.timer that takes
    cm_id_priv->lock

Fix this by using the new __cancel_delayed_work() interface (which
internally does del_timer() instead of del_timer_sync()) in all the
places where we are holding a lock.

Addresses: http://bugzilla.kernel.org/show_bug.cgi?id=13757
Reported-by: Bart Van Assche <bart.vanassche@gmail.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/core/mad.c

index de922a0..bc30c00 100644 (file)
@@ -1974,7 +1974,7 @@ static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
        unsigned long delay;
 
        if (list_empty(&mad_agent_priv->wait_list)) {
-               cancel_delayed_work(&mad_agent_priv->timed_work);
+               __cancel_delayed_work(&mad_agent_priv->timed_work);
        } else {
                mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
                                         struct ib_mad_send_wr_private,
@@ -1983,7 +1983,7 @@ static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
                if (time_after(mad_agent_priv->timeout,
                               mad_send_wr->timeout)) {
                        mad_agent_priv->timeout = mad_send_wr->timeout;
-                       cancel_delayed_work(&mad_agent_priv->timed_work);
+                       __cancel_delayed_work(&mad_agent_priv->timed_work);
                        delay = mad_send_wr->timeout - jiffies;
                        if ((long)delay <= 0)
                                delay = 1;
@@ -2023,7 +2023,7 @@ static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
 
        /* Reschedule a work item if we have a shorter timeout */
        if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
-               cancel_delayed_work(&mad_agent_priv->timed_work);
+               __cancel_delayed_work(&mad_agent_priv->timed_work);
                queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
                                   &mad_agent_priv->timed_work, delay);
        }