Re: [RFC][PATCH] perf: Rewrite core context handling

From: Song Liu <songliubraving@fb.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>,
	Ingo Molnar <mingo@kernel.org>,
	lkml <linux-kernel@vger.kernel.org>,
	"acme@kernel.org" <acme@kernel.org>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Jiri Olsa <jolsa@redhat.com>,
	Stephane Eranian <eranian@google.com>,
	"Thomas Gleixner" <tglx@linutronix.de>,
	"mark.rutland@arm.com" <mark.rutland@arm.com>,
	"megha.dey@intel.com" <megha.dey@intel.com>,
	"frederic@kernel.org" <frederic@kernel.org>
Subject: Re: [RFC][PATCH] perf: Rewrite core context handling
Date: Wed, 17 Oct 2018 18:57:25 +0000	[thread overview]
Message-ID: <26D16201-2CEE-4D4A-B2FD-8F472B48A273@fb.com> (raw)
In-Reply-To: <20181017183310.GB2603@hirez.programming.kicks-ass.net>

> On Oct 17, 2018, at 11:33 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> 
> On Wed, Oct 17, 2018 at 07:19:55PM +0200, Peter Zijlstra wrote:
>> On Wed, Oct 17, 2018 at 04:43:27PM +0000, Song Liu wrote:
>> 
>>>> That makes task and cpu contexts wildly different, which will complicate
>>>> matters I feel.
>>>> 
>>> 
>>> I think we only need different logic when adding events to the task/cpu 
>>> contexts. The ctx_sched_in() and ctx_sched_out() will need some extra
>>> logic to filter out events that are not being scheduled (don't schedule
>>> events on PMU-a when rotating PMU-b). This logic will be the same for 
>>> task and cpu context. The difference is, the CPU context will not have
>>> such events, because we never added such event to CPU context. 
>>> 
>>> Does this make sense? I could try draft a RFC to see how difficult it is. 
>> 
>> I'm not sure it saves much, if we have multiple per-cpu contexts we get
>> to re-introduce the active_ctx_list and loose the simplification for the
>> online status.
>> 
>> Plus that fundamental assymetry -- which would bother my OCD forever
>> more :-)
> 
> Worse, the whole syscall that installs the events will come apart. The
> locking for the two cases is different :/

I agree... I didn't get into details of locking. I just consider these all
as part of "adding event to context". 

I believe this patch should give close to the optimal performance. However, 
I do feel it makes the logic more complicate. Before this patch, perf_cpu_context
and perf_event_context don't need to know much about multiple PMUs. With
this patch, the two extra *_pmu_context are necessary for performance (and
maybe also for correctness). 

If we take first a baby step, how about adding more perf_event_ctx to 
task_struct->perf_event_ctxp? We need one sw perf_event_ctx and a few hw 
perf_event_ctx (one for each hw PMU). (I haven't checked whether it is OK
to allocate these when attaching events). (And I guess you don't really 
like this..)

On the other hand, this patch makes it possible to create groups of events
from different hw PMUs. I guess that will be useful. 

Thanks,
Song