2011-05-04 01:47:59

by Paul Mackerras

Subject: [PATCH] workqueue: Don't spin forever in worker_maybe_bind_and_lock

On a 48-thread POWER7 box, I often see the system hang when offlining
processors. What happens is that we get a rescuer thread trying to
move to some processor at the same time that a cpu offline operation
is happening for that processor, and we end up with one cpu spinning in
worker_maybe_bind_and_lock() and all of the rest of the online cpus
spinning inside the stop_machine code. The rescuer thread is
continually calling set_cpus_allowed_ptr() which is continually
failing because the cpu it is trying to move to is no longer in the
cpu_active_mask. The result is a deadlock.

This fixes worker_maybe_bind_and_lock so that it stops trying to move
to a cpu if that cpu is no longer in the cpu_active_mask, and instead
returns to its caller. With this I no longer see the deadlocks when
offlining cpus.

Signed-off-by: Paul Mackerras <[email protected]>
---
kernel/workqueue.c | 2 ++
1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 8859a41..12faf78 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -1289,6 +1289,8 @@ __acquires(&gcwq->lock)
cpumask_equal(&current->cpus_allowed,
get_cpu_mask(gcwq->cpu)))
return true;
+ if (!cpumask_test_cpu(gcwq->cpu, cpu_active_mask))
+ return false;
spin_unlock_irq(&gcwq->lock);

/* CPU has come up in between, retry migration */
--
1.7.4.1

2011-05-04 08:26:39

by Tejun Heo

Subject: Re: [PATCH] workqueue: Don't spin forever in worker_maybe_bind_and_lock

On Wed, May 04, 2011 at 11:47:49AM +1000, Paul Mackerras wrote:
> On a 48-thread POWER7 box, I often see the system hang when offlining
> processors. What happens is that we get a rescuer thread trying to
> move to some processor at the same time that a cpu offline operation
> is happening for that processor, and we end up with one cpu spinning in
> worker_maybe_bind_and_lock() and all of the rest of the online cpus
> spinning inside the stop_machine code. The rescuer thread is
> continually calling set_cpus_allowed_ptr() which is continually
> failing because the cpu it is trying to move to is no longer in the
> cpu_active_mask. The result is a deadlock.
>
> This fixes worker_maybe_bind_and_lock so that it stops trying to move
> to a cpu if that cpu is no longer in the cpu_active_mask, and instead
> returns to its caller. With this I no longer see the deadlocks when
> offlining cpus.
>
> Signed-off-by: Paul Mackerras <[email protected]>

Hmm.. fix for the problem has already been merged into mainline and
scheduled for -stable. Can you please verify the following fixes the
problem?

Thank you.

>From 5035b20fa5cd146b66f5f89619c20a4177fb736d Mon Sep 17 00:00:00 2001
From: Tejun Heo <[email protected]>
Date: Fri, 29 Apr 2011 18:08:37 +0200
Subject: [PATCH] workqueue: fix deadlock in worker_maybe_bind_and_lock()

If a rescuer and stop_machine() bringing down a CPU race with each
other, they may deadlock on non-preemptive kernel. The CPU won't
accept a new task, so the rescuer can't migrate to the target CPU,
while stop_machine() can't proceed because the rescuer is holding one
of the CPU retrying migration. GCWQ_DISASSOCIATED is never cleared
and worker_maybe_bind_and_lock() retries indefinitely.

This problem can be reproduced semi reliably while the system is
entering suspend.

http://thread.gmane.org/gmane.linux.kernel/1122051

A lot of kudos to Thilo-Alexander for reporting this tricky issue and
painstaking testing.

stable: This affects all kernels with cmwq, so all kernels since and
including v2.6.36 need this fix.

Signed-off-by: Tejun Heo <[email protected]>
Reported-by: Thilo-Alexander Ginkel <[email protected]>
Tested-by: Thilo-Alexander Ginkel <[email protected]>
Cc: [email protected]
---
kernel/workqueue.c | 8 +++++++-
1 files changed, 7 insertions(+), 1 deletions(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 04ef830..e3378e8 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -1291,8 +1291,14 @@ __acquires(&gcwq->lock)
return true;
spin_unlock_irq(&gcwq->lock);

- /* CPU has come up inbetween, retry migration */
+ /*
+ * We've raced with CPU hot[un]plug. Give it a breather
+ * and retry migration. cond_resched() is required here;
+ * otherwise, we might deadlock against cpu_stop trying to
+ * bring down the CPU on non-preemptive kernel.
+ */
cpu_relax();
+ cond_resched();
}
}

--
1.7.1