workqueue: UNBOUND -> REBIND morphing in rebind_workers() should be atomic
authorLai Jiangshan <laijs@cn.fujitsu.com>
Sat, 1 Sep 2012 16:28:19 +0000 (00:28 +0800)
committerBen Hutchings <ben@decadent.org.uk>
Wed, 19 Sep 2012 14:04:58 +0000 (15:04 +0100)
commit 96e65306b81351b656835c15931d1d237b252f27 upstream.

The compiler may compile the following code into TWO write/modify
instructions.

worker->flags &= ~WORKER_UNBOUND;
worker->flags |= WORKER_REBIND;

so the other CPU may temporarily see worker->flags which doesn't have
either WORKER_UNBOUND or WORKER_REBIND set and perform local wakeup
prematurely.

Fix it by using single explicit assignment via ACCESS_ONCE().

Because idle workers have another WORKER_NOT_RUNNING flag, this bug
doesn't exist for them; however, update it to use the same pattern for
consistency.

tj: Applied the change to idle workers too and updated comments and
    patch description a bit.

stable: Idle worker rebinding doesn't apply for -stable and
        WORKER_UNBOUND used to be WORKER_ROGUE.  Updated accordingly.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/workqueue.c

index a650bee..979d4de 100644 (file)
@@ -3437,14 +3437,17 @@ static int __cpuinit trustee_thread(void *__gcwq)
 
        for_each_busy_worker(worker, i, pos, gcwq) {
                struct work_struct *rebind_work = &worker->rebind_work;
+               unsigned long worker_flags = worker->flags;
 
                /*
                 * Rebind_work may race with future cpu hotplug
                 * operations.  Use a separate flag to mark that
-                * rebinding is scheduled.
+                * rebinding is scheduled.  The morphing should
+                * be atomic.
                 */
-               worker->flags |= WORKER_REBIND;
-               worker->flags &= ~WORKER_ROGUE;
+               worker_flags |= WORKER_REBIND;
+               worker_flags &= ~WORKER_ROGUE;
+               ACCESS_ONCE(worker->flags) = worker_flags;
 
                /* queue rebind_work, wq doesn't matter, use the default one */
                if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,