perf and cgroup event scheduling

perf and cgroup event scheduling

[Peter Zijlstra]

Hi Stephane,

while trying to make sense of various other fun event scheduling issues,
I noticed that perf_cgroup_switch() isn't always doing the right thing.

So we typically want to schedule: CPU-pinned, TASK-pinned, CPU-flexible
TASK-flexible, however the current code doesn't respect that.

In particular perf_cgroup_sched_in() is placed wrong, it should be
_before_ we schedule in the task contexts.

There's further complications with things like perf_install_in_context()
and perf_cgroup_move(), but those are either rare or 'interesting' to
fix.

Re: perf and cgroup event scheduling

[Stephane Eranian]

Peter,

On Thu, Mar 31, 2011 at 6:27 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> Hi Stephane,
>
> while trying to make sense of various other fun event scheduling issues,
> I noticed that perf_cgroup_switch() isn't always doing the right thing.
>
> So we typically want to schedule: CPU-pinned, TASK-pinned, CPU-flexible
> TASK-flexible, however the current code doesn't respect that.
>
Unfortunately, you are right ;-). I think we should try to centralize
the scheduling
of per-cpu and per-thread event in a single function that goes through
the priority
list. In certain cases, we would sched out and back in per-thread
events, but that
would be the only to maintain the priority scheme. I will look into that.


In the meantime, yesterday, I found and fixed one potential kernel crash problem
related to mixing cgroup + per-thread events. I will be posting the
patch shortly.


> In particular perf_cgroup_sched_in() is placed wrong, it should be
> _before_ we schedule in the task contexts.
>
> There's further complications with things like perf_install_in_context()
> and perf_cgroup_move(), but those are either rare or 'interesting' to
> fix.
>
>
>

Re: perf and cgroup event scheduling

[Peter Zijlstra]

On Thu, 2011-03-31 at 10:06 -0700, Stephane Eranian wrote:
> Peter,
> 
> On Thu, Mar 31, 2011 at 6:27 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > Hi Stephane,
> >
> > while trying to make sense of various other fun event scheduling issues,
> > I noticed that perf_cgroup_switch() isn't always doing the right thing.
> >
> > So we typically want to schedule: CPU-pinned, TASK-pinned, CPU-flexible
> > TASK-flexible, however the current code doesn't respect that.
> >
> Unfortunately, you are right ;-). I think we should try to centralize
> the scheduling
> of per-cpu and per-thread event in a single function that goes through
> the priority
> list. In certain cases, we would sched out and back in per-thread
> events, but that
> would be the only to maintain the priority scheme. I will look into that.

I've done something near to that here:  
  https://lkml.org/lkml/2011/3/31/232

but that patch is very fresh and not tested at all.

> 
> In the meantime, yesterday, I found and fixed one potential kernel crash problem
> related to mixing cgroup + per-thread events. I will be posting the
> patch shortly.

Looking fwd to that, Thanks!