IPoIB: fix mcast_dev_flush/mcast_restart_task race
authorDoug Ledford <dledford@redhat.com>
Wed, 10 Dec 2014 16:47:01 +0000 (11:47 -0500)
committerRoland Dreier <roland@purestorage.com>
Tue, 16 Dec 2014 02:11:14 +0000 (18:11 -0800)
Our mcast_dev_flush routine and our mcast_restart_task can race
against each other.  In particular, they both hold the priv->lock
while manipulating the rbtree and while removing mcast entries from
the multicast_list and while adding entries to the remove_list, but
they also both drop their locks prior to doing the actual removes.
The mcast_dev_flush routine is run entirely under the rtnl lock and so
has at least some locking.  The actual race condition is like this:

Thread 1                                Thread 2
ifconfig ib0 up
  start multicast join for broadcast
  multicast join completes for broadcast
  start to add more multicast joins
    call mcast_restart_task to add new entries
                                        ifconfig ib0 down
  mcast_dev_flush
    mcast_leave(mcast A)
    mcast_leave(mcast A)

As mcast_leave calls ib_sa_multicast_leave, and as member in
core/multicast.c is ref counted, we run into an unbalanced refcount
issue.  To avoid stomping on each others removes, take the rtnl lock
specifically when we are deleting the entries from the remove list.

Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
drivers/infiniband/ulp/ipoib/ipoib_multicast.c

index a52c9f3f7e420d733b2ca165f84dad6bf8750485..41325960e4e0410ab91c0f982e9a654fecfc0f1a 100644 (file)
@@ -802,7 +802,10 @@ void ipoib_mcast_dev_flush(struct net_device *dev)
 
        spin_unlock_irqrestore(&priv->lock, flags);
 
-       /* seperate between the wait to the leave*/
+       /*
+        * make sure the in-flight joins have finished before we attempt
+        * to leave
+        */
        list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
                if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
                        wait_for_completion(&mcast->done);
@@ -923,14 +926,38 @@ void ipoib_mcast_restart_task(struct work_struct *work)
        netif_addr_unlock(dev);
        local_irq_restore(flags);
 
-       /* We have to cancel outside of the spinlock */
+       /*
+        * make sure the in-flight joins have finished before we attempt
+        * to leave
+        */
+       list_for_each_entry_safe(mcast, tmcast, &remove_list, list)
+               if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
+                       wait_for_completion(&mcast->done);
+
+       /*
+        * We have to cancel outside of the spinlock, but we have to
+        * take the rtnl lock or else we race with the removal of
+        * entries from the remove list in mcast_dev_flush as part
+        * of ipoib_stop() which will call mcast_stop_thread with
+        * flush == 1 while holding the rtnl lock, and the
+        * flush_workqueue won't complete until this restart_mcast_task
+        * completes.  So do like the carrier on task and attempt to
+        * take the rtnl lock, but if we can't before the ADMIN_UP flag
+        * goes away, then just return and know that the remove list will
+        * get flushed later by mcast_dev_flush.
+        */
+       while (!rtnl_trylock()) {
+               if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
+                       return;
+               else
+                       msleep(20);
+       }
        list_for_each_entry_safe(mcast, tmcast, &remove_list, list) {
                ipoib_mcast_leave(mcast->dev, mcast);
                ipoib_mcast_free(mcast);
        }
-
-       if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
-               ipoib_mcast_start_thread(dev);
+       ipoib_mcast_start_thread(dev);
+       rtnl_unlock();
 }
 
 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG