Re: [PATCH v2 1/8] perf/ring_buffer: Add comment for barriers on AUX ring buffer

[Peter Zijlstra]

On Wed, Jun 02, 2021 at 06:30:00PM +0800, Leo Yan wrote:
> AUX ring buffer applies almost the same barriers as perf ring buffer,
> but there has an exception for ordering between writing the AUX trace
> data and updating user_page::aux_head.
> 
> This patch adds comment for how to use the barriers on AUX ring buffer,
> and gives comment to ask the drivers to flush the trace data into AUX
> ring buffer prior to updating user_page::aux_head.
> 
> Signed-off-by: Leo Yan <leo.yan@xxxxxxxxxx>

Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

> ---
>  kernel/events/ring_buffer.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> index 52868716ec35..5cf6579be05e 100644
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -509,6 +509,15 @@ void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
>  		perf_event_aux_event(handle->event, aux_head, size,
>  				     handle->aux_flags);
>  
> +	/*
> +	 * See perf_output_put_handle(), AUX ring buffer applies the same
> +	 * barrier pairing as the perf ring buffer; except for B, since
> +	 * AUX ring buffer is written by hardware trace, we cannot simply
> +	 * use the generic memory barrier (like smp_wmb()) prior to update
> +	 * user_page::aux_head, the hardware trace driver takes the
> +	 * responsibility to ensure the trace data has been flushed into
> +	 * the AUX buffer before calling perf_aux_output_end().
> +	 */
>  	WRITE_ONCE(rb->user_page->aux_head, rb->aux_head);
>  	if (rb_need_aux_wakeup(rb))
>  		wakeup = true;
> -- 
> 2.25.1
>