Re: PERF_EVENT_IOC_SET_OUTPUT

[Ingo Molnar <mingo@kernel.org>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Wed, Oct 02, 2013 at 12:29:56PM +0200, Frederic Weisbecker wrote:
> > On Wed, Oct 02, 2013 at 12:03:50PM +0200, Peter Zijlstra wrote:
> > > On Tue, Oct 01, 2013 at 10:11:56PM +0300, Adrian Hunter wrote:
> > > > Hi
> > > > 
> > > > It does not seem possible to use set-output between
> > > > task contexts of different types (e.g. a software event
> > > > to a hardware event)
> > > > 
> > > > If you look at perf_event_set_output():
> > > > 
> > > >           /*
> > > >            * If its not a per-cpu rb, it must be the same task.
> > > >            */
> > > >           if (output_event->cpu == -1 && output_event->ctx != event->ctx)
> > > >                   goto out;
> > > > 
> > > > ctx (perf_event_context) won't be the same for events
> > > > of different types.  Is this restriction necessary?
> > > 
> > > Hmm.. so last night I wrote me a big reply saying we couldn't do it;
> > > then this morning I reconsidered and thing that something like:
> > > 
> > >   output_event->ctx->task != event->ctx->task
> > > 
> > > should actually work.
> > > 
> > > The reason it should be OK I think is because perf_mmap() will refuse to
> > > create a buffer for inherited events that have ->cpu == -1.
> > > 
> > > My initial response was going to say that it wouldn't be possible
> > > because __perf_event_task_sched_out() could 'break' one ctx while still
> > > swapping the other, at which point the buffer would have to service two
> > > different tasks, potentially from different CPUs and with the buffers
> > > not actually being SMP safe that's a problem.
> > 
> > I don't get what you mean with breaking or swapping a ctx. But I can 
> > confirm that perf_mmap() won't allow a buffer to be remotely accessed 
> > from another CPU. Now there may be other issues than locality which 
> > I'm missing :)
> 
> The way we 'optimize' context switches between tasks with identical 
> contexts is to simply swap the context and leave the hardware alone.

Btw., this does not seem to be working very well when the perf context is 
inherited:

Baseline kernel with no perf context:

  aldebaran:~> taskset 1 perf stat --null perf bench sched pipe
  # Running sched/pipe benchmark...
  # Executed 1000000 pipe operations between two tasks

     Total time: 5.024 [sec]

       5.024951 usecs/op
         199006 ops/sec

with inherited perf contexts:

  aldebaran:~> taskset 1 perf stat perf bench sched pipe
  # Running sched/pipe benchmark...
  # Executed 1000000 pipe operations between two tasks

     Total time: 17.869 [sec]

      17.869061 usecs/op
          55962 ops/sec

+12.8 usecs of perf switching fat per context switch, on a non-debug 
kernel on a 2.8GHz CPU :-/

We should declare a hard feature stop until that is improved.

Thanks,

	Ingo