Re: [PATCH] perf/core: fix multiplexing event scheduling issue

[Peter Zijlstra]

On Thu, Oct 17, 2019 at 05:27:46PM -0700, Stephane Eranian wrote:
> @@ -2153,6 +2157,7 @@ __perf_remove_from_context(struct perf_event *event,
>  			   void *info)
>  {
>  	unsigned long flags = (unsigned long)info;
> +	int was_necessary = ctx->rotate_necessary;
>  
>  	if (ctx->is_active & EVENT_TIME) {
>  		update_context_time(ctx);
> @@ -2171,6 +2176,37 @@ __perf_remove_from_context(struct perf_event *event,
>  			cpuctx->task_ctx = NULL;
>  		}
>  	}
> +
> +	/*
> +	 * sanity check that event_sched_out() does not and will not
> +	 * change the state of ctx->rotate_necessary
> +	 */
> +	WARN_ON(was_necessary != event->ctx->rotate_necessary);

It doesn't... why is this important to check?

> +	/*
> +	 * if we remove an event AND we were multiplexing then, that means
> +	 * we had more events than we have counters for, and thus, at least,
> +	 * one event was in INACTIVE state. Now, that we removed an event,
> +	 * we need to resched to give a chance to all events to get scheduled,
> +	 * otherwise some may get stuck.
> +	 *
> +	 * By the time this function is called the event is usually in the OFF
> +	 * state.
> +	 * Note that this is not a duplicate of the same code in _perf_event_disable()
> +	 * because the call path are different. Some events may be simply disabled

It is the exact same code twice though; IIRC this C language has a
feature to help with that.

> +	 * others removed. There is a way to get removed and not be disabled first.
> +	 */
> +	if (ctx->rotate_necessary && ctx->nr_events) {
> +		int type = get_event_type(event);
> +		/*
> +		 * In case we removed a pinned event, then we need to
> +		 * resched for both pinned and flexible events. The
> +		 * opposite is not true. A pinned event can never be
> +		 * inactive due to multiplexing.
> +		 */
> +		if (type & EVENT_PINNED)
> +			type |= EVENT_FLEXIBLE;
> +		ctx_resched(cpuctx, cpuctx->task_ctx, type);
> +	}

What you're relying on is that ->rotate_necessary implies ->is_active
and there's pending events. And if we tighten ->rotate_necessary you can
remove the && ->nr_events.

> @@ -2232,6 +2270,35 @@ static void __perf_event_disable(struct perf_event *event,
>  		event_sched_out(event, cpuctx, ctx);
>  
>  	perf_event_set_state(event, PERF_EVENT_STATE_OFF);
> +	/*
> +	 * sanity check that event_sched_out() does not and will not
> +	 * change the state of ctx->rotate_necessary
> +	 */
> +	WARN_ON_ONCE(was_necessary != event->ctx->rotate_necessary);
> +
> +	/*
> +	 * if we disable an event AND we were multiplexing then, that means
> +	 * we had more events than we have counters for, and thus, at least,
> +	 * one event was in INACTIVE state. Now, that we disabled an event,
> +	 * we need to resched to give a chance to all events to be scheduled,
> +	 * otherwise some may get stuck.
> +	 *
> +	 * Note that this is not a duplicate of the same code in
> +	 * __perf_remove_from_context()
> +	 * because events can be disabled without being removed.

It _IS_ a duplicate, it is the _exact_ same code twice. What you're
trying to say is that we need it in both places, but that's something
else entirely.

> +	 */
> +	if (ctx->rotate_necessary && ctx->nr_events) {
> +		int type = get_event_type(event);
> +		/*
> +		 * In case we removed a pinned event, then we need to
> +		 * resched for both pinned and flexible events. The
> +		 * opposite is not true. A pinned event can never be
> +		 * inactive due to multiplexing.
> +		 */
> +		if (type & EVENT_PINNED)
> +			type |= EVENT_FLEXIBLE;
> +		ctx_resched(cpuctx, cpuctx->task_ctx, type);
> +	}
>  }