Re: [PATCH 1/1] perf/core: find auxiliary events in running pmus list

[Peter Zijlstra]

You forgot to Cc Alexander, who wrote most of the AUX bits.

On Wed, Feb 24, 2016 at 01:20:36PM -0800, kan.liang@xxxxxxxxx wrote:
> From: Kan Liang <kan.liang@xxxxxxxxx>
> 
> perf_event_aux funciton goes through pmus list to find proper auxiliary
> events to output. The pmus list consists of all possible pmus in the
> system, that may or may not be running at the moment, while the
> auxiliary events must be from the running pmus. Therefore searching
> non-running pmus is unnecessary and expensive especially when there are
> many non-running pmus on the list.
> 
> For example, the brk test case in lkp triggers many mmap operations,
> at the time, perf with cycles:pp is also running on the system. As a
> result, many perf_event_aux are invoked, and each would search the whole
> pmus list. If we enable the uncore support (even when uncore event are
> not really used), dozens of uncore pmus will be added into pmus list,
> which can significantly decrease brk_test's ops_per_sec. Based on our
> test, the ops_per_sec without uncore patch is 2647573, while the
> ops_per_sec with uncore patch is only 1768444, which is a 33.2%
> reduction.
> 
> This patch introduces a running_pmus list which only tracks the running
> pmus in the system. The perf_event_aux uses running_pmus list instead of
> pmus list to find auxiliary events.
> 
> Reported-by: Huang, Ying <ying.huang@xxxxxxxxxxxxxxx>
> Signed-off-by: Kan Liang <kan.liang@xxxxxxxxx>
> ---
>  kernel/events/core.c | 57 +++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 56 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 94c47e3..e33a0de 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -335,6 +335,14 @@ static LIST_HEAD(pmus);
>  static DEFINE_MUTEX(pmus_lock);
>  static struct srcu_struct pmus_srcu;
>  
> +struct running_pmu {
> +	struct list_head	entry;
> +	struct pmu		*pmu;
> +	int			nr_event;
> +};
> +static LIST_HEAD(running_pmus);
> +static DEFINE_MUTEX(running_pmus_lock);
> +
>  /*
>   * perf event paranoia level:
>   *  -1 - not paranoid at all
> @@ -3511,6 +3519,23 @@ static void unaccount_event_cpu(struct perf_event *event, int cpu)
>  		atomic_dec(&per_cpu(perf_cgroup_events, cpu));
>  }
>  
> +static void unaccount_running_pmu(struct perf_event *event)
> +{
> +	struct running_pmu *pmu, *tmp;
> +
> +	mutex_lock(&running_pmus_lock);
> +
> +	list_for_each_entry_safe(pmu, tmp, &running_pmus, entry) {
> +		if ((pmu->pmu == event->pmu) && !(--pmu->nr_event)) {
> +			list_del_rcu(&pmu->entry);
> +			kfree(pmu);
> +			break;
> +		}
> +	}
> +
> +	mutex_unlock(&running_pmus_lock);
> +}
> +
>  static void unaccount_event(struct perf_event *event)
>  {
>  	bool dec = false;
> @@ -3541,6 +3566,8 @@ static void unaccount_event(struct perf_event *event)
>  		static_key_slow_dec_deferred(&perf_sched_events);
>  
>  	unaccount_event_cpu(event, event->cpu);
> +
> +	unaccount_running_pmu(event);
>  }
>  
>  /*
> @@ -5616,6 +5643,7 @@ perf_event_aux(perf_event_aux_output_cb output, void *data,
>  {
>  	struct perf_cpu_context *cpuctx;
>  	struct perf_event_context *ctx;
> +	struct running_pmu *running_pmu;
>  	struct pmu *pmu;
>  	int ctxn;
>  
> @@ -5631,7 +5659,9 @@ perf_event_aux(perf_event_aux_output_cb output, void *data,
>  	}
>  
>  	rcu_read_lock();
> -	list_for_each_entry_rcu(pmu, &pmus, entry) {
> +
> +	list_for_each_entry_rcu(running_pmu, &running_pmus, entry) {
> +		pmu = running_pmu->pmu;
>  		cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
>  		if (cpuctx->unique_pmu != pmu)
>  			goto next;
> @@ -7740,6 +7770,29 @@ static void account_event_cpu(struct perf_event *event, int cpu)
>  		atomic_inc(&per_cpu(perf_cgroup_events, cpu));
>  }
>  
> +static void account_running_pmu(struct perf_event *event)
> +{
> +	struct running_pmu *pmu;
> +
> +	mutex_lock(&running_pmus_lock);
> +
> +	list_for_each_entry(pmu, &running_pmus, entry) {
> +		if (pmu->pmu == event->pmu) {
> +			pmu->nr_event++;
> +			goto out;
> +		}
> +	}
> +
> +	pmu = kzalloc(sizeof(struct running_pmu), GFP_KERNEL);
> +	if (pmu != NULL) {
> +		pmu->nr_event++;
> +		pmu->pmu = event->pmu;
> +		list_add_rcu(&pmu->entry, &running_pmus);
> +	}
> +out:
> +	mutex_unlock(&running_pmus_lock);
> +}
> +
>  static void account_event(struct perf_event *event)
>  {
>  	bool inc = false;
> @@ -7772,6 +7825,8 @@ static void account_event(struct perf_event *event)
>  		static_key_slow_inc(&perf_sched_events.key);
>  
>  	account_event_cpu(event, event->cpu);
> +
> +	account_running_pmu(event);
>  }
>  
>  /*
> -- 
> 2.5.0
>