Re: [PATCH V5 01/12] perf/core: Add aux_pause, aux_resume, aux_start_paused

[Adrian Hunter]

On 9/02/24 02:13, Andi Kleen wrote:
>> +static void __perf_event_aux_pause(struct perf_event *event, bool pause)
>> +{
>> +	if (pause) {
>> +		if (!READ_ONCE(event->aux_paused)) {
>> +			WRITE_ONCE(event->aux_paused, 1);
>> +			event->pmu->stop(event, PERF_EF_PAUSE);
>> +		}
>> +	} else {
>> +		if (READ_ONCE(event->aux_paused)) {
>> +			WRITE_ONCE(event->aux_paused, 0);
>> +			event->pmu->start(event, PERF_EF_RESUME);
>> +		}
> 
> This doesn't look atomic. Either the READ/WRITE once are not needed,
> or you need an actually atomic construct.

Yes READ_ONCE / WRITE_ONCE is not really needed here.

> 
>> +
>> +	rb = ring_buffer_get(event);
>> +	if (!rb)
>> +		return;
>> +
>> +	local_irq_save(flags);
>> +	/* Guard against NMI, NMI loses here */
>> +	if (READ_ONCE(rb->aux_in_pause_resume))
>> +		goto out_restore;
>> +	WRITE_ONCE(rb->aux_in_pause_resume, 1);
> 
> 
>> +	barrier();
>> +	__perf_event_aux_pause(event, pause);
>> +	barrier();
>> +	WRITE_ONCE(rb->aux_in_pause_resume, 0);
> 
> Dito.
> 

The writes to rb->aux_in_pause_resume must be done
only once.  It might be possible to get away without
WRITE_ONCE(), but really the compiler should be informed
not to make assumptions.