Re: [PATCH 1/2] kernel/events: Add option to notify through signals on wakeup

[Naveen N. Rao]

On 2017/07/12 01:46PM, Peter Zijlstra wrote:
> On Mon, Jun 19, 2017 at 08:01:07PM +0530, Naveen N. Rao wrote:
> 
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 6c4e523dc1e2..812fcfc767f4 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -2679,7 +2679,8 @@ static int _perf_event_refresh(struct perf_event *event, int refresh)
> >  	/*
> >  	 * not supported on inherited events
> 
> That comment wants updating to explain the signal crud..

Ok sure. Will do.

> 
> >  	 */
> > -	if (event->attr.inherit || !is_sampling_event(event))
> > +	if (event->attr.inherit || event->attr.signal_on_wakeup ||
> > +			!is_sampling_event(event))
> >  		return -EINVAL;
> >  
> >  	atomic_add(refresh, &event->event_limit);
> 
> > @@ -7362,12 +7362,15 @@ static int __perf_event_overflow(struct perf_event *event,
> >  	 * events
> >  	 */
> >  
> > +	if (!event->attr.signal_on_wakeup) {
> > +		int events = atomic_read(&event->event_limit);
> > +		event->pending_kill = POLL_IN;
> > +		if (events && atomic_dec_and_test(&event->event_limit)) {
> > +			ret = 1;
> > +			event->pending_kill = POLL_HUP;
> >  
> > +			perf_event_disable_inatomic(event);
> > +		}
> >  	}
> 
> So even without event_limit (IOC_REFRESH) this would've generated
> SIGIO:POLL_IN.

Yes, this still does if signal_on_wakeup isn't set. However, if 
signal_on_wakeup is set, we follow the ring buffer notification which 
will generate a POLL_IN controlled by {wakeup_events, wakeup_watermark}.


> 
> >  
> >  	READ_ONCE(event->overflow_handler)(event, data, regs);
> > @@ -10408,6 +10411,7 @@ perf_event_exit_event(struct perf_event *child_event,
> >  		perf_group_detach(child_event);
> >  	list_del_event(child_event, child_ctx);
> >  	child_event->state = PERF_EVENT_STATE_EXIT; /* is_event_hup() */
> > +	child_event->pending_kill = POLL_HUP;
> 
> This looks like an undocumented change..

Yes, sorry - I should have added a note.

This comes from Jiri's feedback that if user chooses to have a signal 
delivered on ring buffer wakeup, we should also send a HUP on exit.

> 
> >  	raw_spin_unlock_irq(&child_ctx->lock);
> >  
> >  	/*
> > diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> > index 2831480c63a2..4e7c728569a8 100644
> > --- a/kernel/events/ring_buffer.c
> > +++ b/kernel/events/ring_buffer.c
> > @@ -21,6 +21,9 @@ static void perf_output_wakeup(struct perf_output_handle *handle)
> >  {
> >  	atomic_set(&handle->rb->poll, POLLIN);
> >  
> > +	if (handle->event->attr.signal_on_wakeup)
> > +		handle->event->pending_kill = POLL_IN;
> > +
> 
> And this is the bit that generates SIGIO:POLL_IN on wakeup.

Yes.


Thanks,
Naveen

> 
> >  	handle->event->pending_wakeup = 1;
> >  	irq_work_queue(&handle->event->pending);
> >  }
> > -- 
> > 2.13.1
> > 
>