Re: [PATCH v1 4/5] perf: Introduce address range filtering

From: Alexander Shishkin
Date: Mon Apr 25 2016 - 12:03:21 EST


Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Fri, Apr 22, 2016 at 07:19:11PM +0300, Alexander Shishkin wrote:
>> +static int perf_event_restart(struct perf_event *event)
>> +{
>> + struct stop_event_data sd = {
>> + .event = event,
>> + .restart = 1,
>> + };
>> + int ret = 0;
>> +
>> + do {
>> + if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
>> + return 0;
>> +
>> + /* matches smp_wmb() in event_sched_in() */
>> + smp_rmb();
>> +
>> + ret = cpu_function_call(READ_ONCE(event->oncpu),
>> + __perf_event_stop, &sd);
>> + } while (ret == -EAGAIN);
>> +
>> + return ret;
>> +}
>
> A few concerns with this;

(awesome, and now I see where the comments are actually needed)

> - if I understand things right; this will loose a bit of trace when
> people tickle the ioctl(), right? This might want a note of sorts,

I don't actually see how, if they're doing the ioctl() on an enabled
event, it'll generate unfiltered data until it gets scheduled out or it
receives this cross call, which will apply the new filters. The
__perf_event_stop(), which restarts the event will run on the target
cpu, so it doesn't race with the workload.

> - you need to handle event->oncpu == -1,

Ok, we're in practice ignoring this but it deserves a comment. The
cpu_function_call() here should return -ENXIO in this case and we'll
just fall through, which is fine, because that means that the event is
inactive and the filters will apply when it goes active again.

> - can we race with an actual stop, and accidentally re-enable the
> thing? If not, this might be a good place for a comment explaining
> how.

We can (mmap in workload vs munmap(AUX) in trace consumer), but the
event's start will fail in perf_aux_output_begin() because of
rb::aux_mmap_count being zero. The actual stop is only used in aux
unmapping, so like this we should be covered.

Thanks,
--
Alex