workqueue: cond_resched() after processing each work item
authorTejun Heo <tj@kernel.org>
Wed, 28 Aug 2013 21:33:37 +0000 (17:33 -0400)
committerTejun Heo <tj@kernel.org>
Thu, 29 Aug 2013 13:19:28 +0000 (09:19 -0400)
If !PREEMPT, a kworker running work items back to back can hog CPU.
This becomes dangerous when a self-requeueing work item which is
waiting for something to happen races against stop_machine.  Such
self-requeueing work item would requeue itself indefinitely hogging
the kworker and CPU it's running on while stop_machine would wait for
that CPU to enter stop_machine while preventing anything else from
happening on all other CPUs.  The two would deadlock.

Jamie Liu reports that this deadlock scenario exists around
scsi_requeue_run_queue() and libata port multiplier support, where one
port may exclude command processing from other ports.  With the right
timing, scsi_requeue_run_queue() can end up requeueing itself trying
to execute an IO which is asked to be retried while another device has
an exclusive access, which in turn can't make forward progress due to
stop_machine.

Fix it by invoking cond_resched() after executing each work item.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Jamie Liu <jamieliu@google.com>
References: http://thread.gmane.org/gmane.linux.kernel/1552567
Cc: stable@vger.kernel.org
--
 kernel/workqueue.c |    9 +++++++++
 1 file changed, 9 insertions(+)

kernel/workqueue.c

index 7f5d4be220345b0899d38ced692f2458b9278ce7..e93f7b9067d80a9a407df8eaed869e3b3c54299a 100644 (file)
@@ -2201,6 +2201,15 @@ __acquires(&pool->lock)
                dump_stack();
        }
 
+       /*
+        * The following prevents a kworker from hogging CPU on !PREEMPT
+        * kernels, where a requeueing work item waiting for something to
+        * happen could deadlock with stop_machine as such work item could
+        * indefinitely requeue itself while all other CPUs are trapped in
+        * stop_machine.
+        */
+       cond_resched();
+
        spin_lock_irq(&pool->lock);
 
        /* clear cpu intensive status */