Re: [PATCH v2] perf/core: fix userpage->time_enabled of inactive events

[Song Liu]

> On Sep 29, 2021, at 10:04 AM, Athira Rajeev <atrajeev@xxxxxxxxxxxxxxxxxx> wrote:
> 
> 
> 
>> On 24-Sep-2021, at 6:58 AM, Song Liu <songliubraving@xxxxxx> wrote:
>> 
>> Users of rdpmc rely on the mmapped user page to calculate accurate
>> time_enabled. Currently, userpage->time_enabled is only updated when the
>> event is added to the pmu. As a result, inactive event (due to counter
>> multiplexing) does not have accurate userpage->time_enabled. This can
>> be reproduced with something like:
>> 
>>  /* open 20 task perf_event "cycles", to create multiplexing */
>> 
>>  fd = perf_event_open();  /* open task perf_event "cycles" */
>>  userpage = mmap(fd);     /* use mmap and rdmpc */
>> 
>>  while (true) {
>>    time_enabled_mmap = xxx; /* use logic in perf_event_mmap_page */
>>    time_enabled_read = read(fd).time_enabled;
>>    if (time_enabled_mmap > time_enabled_read)
>>        BUG();
>>  }
>> 
>> Fix this by updating userpage for inactive events in merge_sched_in.
>> 
>> Suggested-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
>> Reported-and-tested-by: Lucian Grijincu <lucian@xxxxxx>
>> Signed-off-by: Song Liu <songliubraving@xxxxxx>
>> ---
>> include/linux/perf_event.h |  4 +++-
>> kernel/events/core.c       | 49 ++++++++++++++++++++++++++++++++++----
>> 2 files changed, 48 insertions(+), 5 deletions(-)
>> 
>> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
>> index 2d510ad750edc..4aa52f7a48c16 100644
>> --- a/include/linux/perf_event.h
>> +++ b/include/linux/perf_event.h
>> @@ -683,7 +683,9 @@ struct perf_event {
>> 	/*
>> 	 * timestamp shadows the actual context timing but it can
>> 	 * be safely used in NMI interrupt context. It reflects the
>> -	 * context time as it was when the event was last scheduled in.
>> +	 * context time as it was when the event was last scheduled in,
>> +	 * or when ctx_sched_in failed to schedule the event because we
>> +	 * run out of PMC.
>> 	 *
>> 	 * ctx_time already accounts for ctx->timestamp. Therefore to
>> 	 * compute ctx_time for a sample, simply add perf_clock().
>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>> index 1cb1f9b8392e2..d73f986eef7b3 100644
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
>> @@ -3707,6 +3707,46 @@ static noinline int visit_groups_merge(struct perf_cpu_context *cpuctx,
>> 	return 0;
>> }
>> 
>> +static inline bool event_update_userpage(struct perf_event *event)
>> +{
>> +	/*
>> +	 * Checking mmap_count to avoid unnecessary work. This does leave a
>> +	 * corner case: if the event is enabled before mmap(), the first
>> +	 * time the event gets scheduled is via:
>> +	 *
>> +	 *  __perf_event_enable (or __perf_install_in_context)
>> +	 *      -> ctx_resched
>> +	 *         -> perf_event_sched_in
>> +	 *            -> ctx_sched_in
>> +	 *
>> +	 * with mmap_count of 0, so we will skip here. As a result,
>> +	 * userpage->offset is not accurate after mmap and before the
>> +	 * first rotation.
>> +	 *
>> +	 * To avoid the discrepancy of this window, the user space should
>> +	 * mmap the event before enabling it.
>> +	 */
>> +	if (likely(!atomic_read(&event->mmap_count)))
>> +		return false;
>> +
>> +	perf_event_update_time(event);
>> +	perf_set_shadow_time(event, event->ctx);
>> +	perf_event_update_userpage(event);
>> +
>> +	return true;
>> +}
>> +
>> +static inline void group_update_userpage(struct perf_event *group_event)
>> +{
>> +	struct perf_event *event;
>> +
>> +	if (!event_update_userpage(group_event))
>> +		return;
>> +
>> +	for_each_sibling_event(event, group_event)
>> +		event_update_userpage(event);
>> +}
>> +
>> static int merge_sched_in(struct perf_event *event, void *data)
>> {
>> 	struct perf_event_context *ctx = event->ctx;
>> @@ -3725,14 +3765,15 @@ static int merge_sched_in(struct perf_event *event, void *data)
>> 	}
>> 
>> 	if (event->state == PERF_EVENT_STATE_INACTIVE) {
>> +		*can_add_hw = 0;
>> 		if (event->attr.pinned) {
>> 			perf_cgroup_event_disable(event, ctx);
>> 			perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
>> +		} else {
>> +			ctx->rotate_necessary = 1;
>> +			perf_mux_hrtimer_restart(cpuctx);
>> +			group_update_userpage(event);
>> 		}
>> -
>> -		*can_add_hw = 0;
>> -		ctx->rotate_necessary = 1;
>> -		perf_mux_hrtimer_restart(cpuctx);
>> 	}
> 
> Another optimisation that is possible in merge_sched_in is we can avoid calling "perf_mux_hrtimer_restart" multiple times if already rotate_necessary is set for that context. Even though "perf_mux_hrtimer_restart" will just return if hrtimer is already active, we could avoid the overhead of calling this function multiple times if there are many groups. 
> 
> Something like below:
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 0c000cb01eeb..26eae79bd723 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -3731,8 +3731,10 @@ static int merge_sched_in(struct perf_event *event, void *data)
>                }
> 
>                *can_add_hw = 0;
> -               ctx->rotate_necessary = 1;
> -               perf_mux_hrtimer_restart(cpuctx);
> +               if (!ctx->rotate_necessary) {
> +                       ctx->rotate_necessary = 1;
> +                       perf_mux_hrtimer_restart(cpuctx);
> +               }
>        }
> 
>        return 0;

Yeah, this makes sense. Do you plan to send a patch for this? 

Thanks,
Song