Re: [PATCH] sched: Fix SCHED_HRTICK bug leading to late preemption of tasks

[Joonwoo Park]

On Mon, Sep 19, 2016 at 10:21:58AM +0200, Peter Zijlstra wrote:
> On Fri, Sep 16, 2016 at 06:28:51PM -0700, Joonwoo Park wrote:
> > From: Srivatsa Vaddagiri <vatsa@xxxxxxxxxxxxxx>
> > 
> > SCHED_HRTICK feature is useful to preempt SCHED_FAIR tasks on-the-dot
> 
> Right, but I always found the overhead of the thing too high to be
> really useful.
> 
> How come you're using this?

This patch was in our internal tree for decades so I unfortunately cannot
find actual usecase or history.
But I guess it was about excessive latency when there are number of CPU
bound tasks running on a CPU but on different cfs_rqs and CONFIG_HZ = 100.

See how I recreated :

* run 4 cpu hogs on the same cgroup [1] :
 dd-960   [000] d..3   110.651060: sched_switch: prev_comm=dd prev_pid=960 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=959 next_prio=120
 dd-959   [000] d..3   110.652566: sched_switch: prev_comm=dd prev_pid=959 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=961 next_prio=120
 dd-961   [000] d..3   110.654072: sched_switch: prev_comm=dd prev_pid=961 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=962 next_prio=120
 dd-962   [000] d..3   110.655578: sched_switch: prev_comm=dd prev_pid=962 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=960 next_prio=120
  preempt every 1.5ms slice by hrtick.

* run 4 CPU hogs on 4 different cgroups [2] :
 dd-964   [000] d..3    24.169873: sched_switch: prev_comm=dd prev_pid=964 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=966 next_prio=120
 dd-966   [000] d..3    24.179873: sched_switch: prev_comm=dd prev_pid=966 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=965 next_prio=120
 dd-965   [000] d..3    24.189873: sched_switch: prev_comm=dd prev_pid=965 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=967 next_prio=120
 dd-967   [000] d..3    24.199873: sched_switch: prev_comm=dd prev_pid=967 prev_prio=120 prev_state=R+ ==> next_comm=dd next_pid=964 next_prio=120
  preempt every 10ms by scheduler tick so that all tasks suffers from 40ms preemption latency.

[1] : 
 dd if=/dev/zero of=/dev/zero &
 dd if=/dev/zero of=/dev/zero &
 dd if=/dev/zero of=/dev/zero &
 dd if=/dev/zero of=/dev/zero &

[2] :
 mount -t cgroup -o cpu cpu /sys/fs/cgroup
 mkdir /sys/fs/cgroup/grp1
 mkdir /sys/fs/cgroup/grp2
 mkdir /sys/fs/cgroup/grp3
 mkdir /sys/fs/cgroup/grp4
 dd if=/dev/zero of=/dev/zero &
 echo $! > /sys/fs/cgroup/grp1/tasks 
 dd if=/dev/zero of=/dev/zero &
 echo $! > /sys/fs/cgroup/grp2/tasks 
 dd if=/dev/zero of=/dev/zero &
 echo $! > /sys/fs/cgroup/grp3/tasks 
 dd if=/dev/zero of=/dev/zero &
 echo $! > /sys/fs/cgroup/grp4/tasks 

I could confirm this patch makes the latter behaves as same as the former in terms of preemption latency.

> 
> 
> >  joonwoop: Do we also need to update or remove if-statement inside
> >  hrtick_update()?
> 
> >  I guess not because hrtick_update() doesn't want to start hrtick when cfs_rq
> >  has large number of nr_running where slice is longer than sched_latency.
> 
> Right, you want that to match with whatever sched_slice() does.

Cool.  Thank you!

Thanks,
Joonwoo

> 
> > +++ b/kernel/sched/fair.c
> > @@ -4458,7 +4458,7 @@ static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
> >  
> >  	WARN_ON(task_rq(p) != rq);
> >  
> > -	if (cfs_rq->nr_running > 1) {
> > +	if (rq->cfs.h_nr_running > 1) {
> >  		u64 slice = sched_slice(cfs_rq, se);
> >  		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
> >  		s64 delta = slice - ran;
> 
> Yeah, that looks right. I don't think I've ever tried hrtick with
> cgroups enabled...