Re: [PATCH v2 1/2] workqueue: add quiescent state between work items

[Jiri Pirko <jiri@resnulli.us>]

Wed, Oct 08, 2014 at 05:24:11AM CEST, paulmck@linux.vnet.ibm.com wrote:
>On Tue, Oct 07, 2014 at 01:45:28PM -0400, Joe Lawrence wrote:
>> On Tue, 7 Oct 2014 06:43:29 -0700
>> "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
>> 
>> > On Tue, Oct 07, 2014 at 09:29:42AM +0200, Jiri Pirko wrote:
>> [ ... snip ... ]
>> > > 
>> > > Paul, Tehun, how do you propose to fix this on older kernels which do
>> > > not have rcu_note_voluntary_context_switch? I'm particullary interested
>> > > in 3.10.
>> > 
>> > Hello, Jiri,
>> > 
>> > Older kernels can instead use rcu_note_context_switch().
>> 
>> Hi Paul,
>> 
>> Does 4a81e8328d37 ("rcu: Reduce overhead of cond_resched() checks for
>> RCU") affect a backport to 3.10?
>> 
>> I noticed that rcu_note_context_switch added a call to
>> rcu_momentary_dyntick_idle in that change, which is only present in
>> v3.16+.
>> 
>> Would rcu_note_context_switch be effective by itself on a 3.10 kernel?
>
>Should be fine.  There is more overhead than current mainline, but that
>should not be in the noise compared to executing a work-queue item.
>
>							Thanx, Paul
>

I cooked up following patch. Please tell me if it is fine or not. I can
also send it oficially so it can be included into stable trees:

Subject: workqueue: Add quiescent state between work items

Similar to the stop_machine deadlock scenario on !PREEMPT kernels
addressed in b22ce2785d97 "workqueue: cond_resched() after processing
each work item", kworker threads requeueing back-to-back with zero jiffy
delay can stall RCU. The cond_resched call introduced in that fix will
yield only iff there are other higher priority tasks to run, so force a
quiescent RCU state between work items.

Signed-off-by: Jiri Pirko <jiri@resnulli.us>
---
 kernel/workqueue.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index e9719c7..14a7163 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -2196,8 +2196,10 @@ __acquires(&pool->lock)
 	 * 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.
+	 * stop_machine. At the same time, report a quiescent RCU state so
+	 * the same condition doesn't freeze RCU.
 	 */
+	rcu_note_context_switch(raw_smp_processor_id());
 	cond_resched();
 
 	spin_lock_irq(&pool->lock);
-- 
1.9.3