Re: [PATCH v2 1/1] perf: Sharing PMU counters across compatible events

From: Song Liu <songliubraving@fb.com>
To: Jiri Olsa <jolsa@redhat.com>
Cc: lkml <linux-kernel@vger.kernel.org>,
	"lkp@intel.com" <lkp@intel.com>, Kernel Team <Kernel-team@fb.com>,
	Tejun Heo <tj@kernel.org>, Peter Zijlstra <peterz@infradead.org>,
	Jiri Olsa <jolsa@kernel.org>,
	Alexey Budankov <alexey.budankov@linux.intel.com>
Subject: Re: [PATCH v2 1/1] perf: Sharing PMU counters across compatible events
Date: Tue, 11 Sep 2018 13:38:49 +0000	[thread overview]
Message-ID: <F3425547-3F28-44F3-8328-1CAB0FD12A96@fb.com> (raw)
In-Reply-To: <20180910081531.GC31644@krava>

> On Sep 10, 2018, at 1:15 AM, Jiri Olsa <jolsa@redhat.com> wrote:
> 
> On Thu, Aug 30, 2018 at 06:51:07PM +0000, Song Liu wrote:
>> 
>> 
>>> On Aug 30, 2018, at 8:18 AM, Jiri Olsa <jolsa@redhat.com> wrote:
>>> 
>>> On Wed, Aug 15, 2018 at 10:03:13AM -0700, Song Liu wrote:
>>> 
>>> SNIP
>>> 
>>>> @@ -6100,7 +6333,7 @@ static void perf_output_read_group(struct perf_output_handle *handle,
>>>> 
>>>> 		if ((sub != event) &&
>>>> 		    (sub->state == PERF_EVENT_STATE_ACTIVE))
>>>> -			sub->pmu->read(sub);
>>>> +			event_pmu_read(sub);
>>>> 
>>>> 		values[n++] = perf_event_count(sub);
>>>> 		if (read_format & PERF_FORMAT_ID)
>>>> @@ -9109,7 +9342,7 @@ static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
>>>> 	if (event->state != PERF_EVENT_STATE_ACTIVE)
>>>> 		return HRTIMER_NORESTART;
>>>> 
>>>> -	event->pmu->read(event);
>>>> +	event_pmu_read(event);
>>>> 
>>>> 	perf_sample_data_init(&data, 0, event->hw.last_period);
>>>> 	regs = get_irq_regs();
>>>> @@ -10504,6 +10737,14 @@ SYSCALL_DEFINE5(perf_event_open,
>>>> 		goto err_cred;
>>>> 	}
>>>> 
>>>> +	if (perf_event_can_share(event)) {
>>>> +		event->tmp_master = perf_event_alloc(&event->attr, cpu,
>>>> +						     task, NULL, NULL,
>>>> +						     NULL, NULL, -1);
>>> 
>>> can't get around this.. I understand the need, but AFAICS you allocate
>>> the whole 'struct perf_event', just because there's count field in it
>>> otherwise the 'struct hw_perf_event' should be enough to carry all that's
>>> needed to read hw event
>>> 
>>> would it be better to move the count to 'struct hw_perf_event' and use
>>> that instead? assuming I'm not missing anything..
>>> 
>>> jirka
>> 
>> I am trying to make the master event function the same as a real event, 
>> while keep dup events as followers. This avoids "switching master" in 
>> earlier versions (and Tejun's RFC). 
> 
> yep, I understand.. still, it seems too much to allocate
> the whole 'struct perf_even't just to get separated 'count'
> variable

In theory, we only need separated counters. However, in practice, there
are other variables we need to handle for a switch_master operation. 
For example, we need make sure event->state is always set properly. So 
this optimization is not easy to implement. How about we optimize it 
after this patch gets in? 

Thanks,
Song