workqueue: fix spurious CPU locality WARN from process_one_work()
Tejun Heo [Sun, 22 Jul 2012 17:16:34 +0000 (10:16 -0700)]
25511a4776 "workqueue: reimplement CPU online rebinding to handle idle
workers" added CPU locality sanity check in process_one_work().  It
triggers if a worker is executing on a different CPU without UNBOUND
or REBIND set.

This works for all normal workers but rescuers can trigger this
spuriously when they're serving the unbound or a disassociated
global_cwq - rescuers don't have either flag set and thus its
gcwq->cpu can be a different value including %WORK_CPU_UNBOUND.

Fix it by additionally testing %GCWQ_DISASSOCIATED.

bug 1237774

Change-Id: I7ac841fb02f97c354f096df1715669acca0265b5
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
LKML-Refence: <20120721213656.GA7783@linux.vnet.ibm.com>
Signed-off-by: Bo Yan <byan@nvidia.com>
Reviewed-on: http://git-master/r/226758
Reviewed-by: Yu-Huan Hsu <yhsu@nvidia.com>
GVS: Gerrit_Virtual_Submit

kernel/workqueue.c

index c1e3432..ca02bb5 100644 (file)
@@ -1946,7 +1946,13 @@ __acquires(&gcwq->lock)
         */
        struct lockdep_map lockdep_map = work->lockdep_map;
 #endif
+       /*
+        * Ensure we're on the correct CPU.  DISASSOCIATED test is
+        * necessary to avoid spurious warnings from rescuers servicing the
+        * unbound or a disassociated gcwq.
+        */
        WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) &&
+                    !(gcwq->flags & GCWQ_DISASSOCIATED) &&
                     raw_smp_processor_id() != gcwq->cpu);
 
        /*