[v4.2+ regression] fd7a4bed sched, rt: Convert switched_{from, to}_rt() / prio_changed_rt() to balance callbacks

[v4.2+ regression] fd7a4bed sched, rt: Convert switched_{from, to}_rt() / prio_changed_rt() to balance callbacks

[Mike Galbraith]

Hi Peter,

I bumped into an odd futextest regression, and finally bisected it to
$subject.  I haven't poked at it yet, chasing down and confirming the
little bugger munched the day.

#!/bin/sh

cd /home/git/futextest/functional
trap "killall -KILL -q cpuhog;exit;" INT

for i in `seq 0 7`; do taskset -c $i cpuhog& done
sleep 1
for i in `seq 1000`
do
	echo iteration $i
	./futex_requeue_pi -c -b -o -t 500000
done
killall cpuhog


Without the cpuhogs, it won't hang.

	-Mike

Re: [v4.2+ regression] fd7a4bed sched, rt: Convert switched_{from, to}_rt() / prio_changed_rt() to balance callbacks

[Mike Galbraith]

On Mon, 2015-09-28 at 04:17 +0200, Mike Galbraith wrote:
> Hi Peter,
> 
> I bumped into an odd futextest regression, and finally bisected it to
> $subject.  I haven't poked at it yet, chasing down and confirming the
> little bugger munched the day.

homer:/home/git/futextest/functional # ./futex_requeue_pi -c -b -o -t 500000 -v 2 2>&1|egrep 'exit|Blocker'
         INFO: Waiter 0: exiting with 0
         INFO: Waiter 1: exiting with 0
         INFO: Waiter 2: exiting with 0
         INFO: Waiter 3: exiting with 0
         INFO: Waiter 4: exiting with 0
         INFO: Waiter 5: exiting with 0
         INFO: Waiter 6: exiting with 0
         INFO: Waiter 7: exiting with 0
         INFO: Waiter 8: exiting with 0
         INFO: Waiter 9: exiting with 0
         INFO: Waker: exiting with 0
         INFO: Blocker: Calling futex_wait()

Well now, it _seems_ you're innocent Peter.  Waker can call futex_wake()
before blocker calls futex_wait(), leaving poor blocker stranded.

Adding atomic_inc(&waiters_blocked) to blocker, and telling wakers to
expect one more when a blocker exists seems to have fixed it up.

	-Mike

Re: [v4.2+ regression] fd7a4bed sched, rt: Convert switched_{from, to}_rt() / prio_changed_rt() to balance callbacks

[Peter Zijlstra]

On Tue, Sep 29, 2015 at 10:19:51AM +0200, Mike Galbraith wrote:
> On Mon, 2015-09-28 at 04:17 +0200, Mike Galbraith wrote:
> > Hi Peter,
> > 
> > I bumped into an odd futextest regression, and finally bisected it to
> > $subject.  I haven't poked at it yet, chasing down and confirming the
> > little bugger munched the day.

> Well now, it _seems_ you're innocent Peter. 


Whee! thanks for looking.