Merge branch 'for-4.6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq

Pull workqueue fix from Tejun Heo:
 "So, it turns out we had a silly bug in the most fundamental part of
  workqueue for a very long time.  AFAICS, this dates back to pre-git
  era and has quite likely been there from the time workqueue was first
  introduced.

  A work item uses its PENDING bit to synchronize multiple queuers.
  Anyone who wins the PENDING bit owns the pending state of the work
  item.  Whether a queuer wins or loses the race, one thing should be
  guaranteed - there will soon be at least one execution of the work
  item - where "after" means that the execution instance would be able
  to see all the changes that the queuer has made prior to the queueing
  attempt.

  Unfortunately, we were missing a smp_mb() after clearing PENDING for
  execution, so nothing guaranteed visibility of the changes that a
  queueing loser has made, which manifested as a reproducible blk-mq
  stall.

  Lots of kudos to Roman for debugging the problem.  The patch for
  -stable is the minimal one.  For v3.7, Peter is working on a patch to
  make the code path slightly more efficient and less fragile"

* 'for-4.6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
  workqueue: fix ghost PENDING flag while doing MQ IO

Author: torvalds

kernel/workqueue.c

@@ -666,6 +666,35 @@ static void set_work_pool_and_clear_pending(struct work_struct *work,
 	 */
 	smp_wmb();
 	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
+	/*
+	 * The following mb guarantees that previous clear of a PENDING bit
+	 * will not be reordered with any speculative LOADS or STORES from
+	 * work->current_func, which is executed afterwards.  This possible
+	 * reordering can lead to a missed execution on attempt to qeueue
+	 * the same @work.  E.g. consider this case:
+	 *
+	 *   CPU#0                         CPU#1
+	 *   ----------------------------  --------------------------------
+	 *
+	 * 1  STORE event_indicated
+	 * 2  queue_work_on() {
+	 * 3    test_and_set_bit(PENDING)
+	 * 4 }                             set_..._and_clear_pending() {
+	 * 5                                 set_work_data() # clear bit
+	 * 6                                 smp_mb()
+	 * 7                               work->current_func() {
+	 * 8				      LOAD event_indicated
+	 *				   }
+	 *
+	 * Without an explicit full barrier speculative LOAD on line 8 can
+	 * be executed before CPU#0 does STORE on line 1.  If that happens,
+	 * CPU#0 observes the PENDING bit is still set and new execution of
+	 * a @work is not queued in a hope, that CPU#1 will eventually
+	 * finish the queued @work.  Meanwhile CPU#1 does not see
+	 * event_indicated is set, because speculative LOAD was executed
+	 * before actual STORE.
+	 */
+	smp_mb();
 }
 
 static void clear_work_data(struct work_struct *work)