Re: [PATCH] trace: add ability to set a target task for events (v2)

From: Frederic Weisbecker <fweisbec@gmail.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Andrew Vagin <avagin@openvz.org>,
	linux-kernel@vger.kernel.org, Ingo Molnar <mingo@kernel.org>,
	Steven Rostedt <rostedt@goodmis.org>,
	Paul Mackerras <paulus@samba.org>,
	Arnaldo Carvalho de Melo <acme@ghostprotocols.net>,
	Arun Sharma <asharma@fb.com>
Subject: Re: [PATCH] trace: add ability to set a target task for events (v2)
Date: Wed, 11 Jul 2012 16:36:59 +0200	[thread overview]
Message-ID: <20120711143656.GB17991@somewhere> (raw)
In-Reply-To: <1342017221.3462.159.camel@twins>

On Wed, Jul 11, 2012 at 04:33:41PM +0200, Peter Zijlstra wrote:
> On Wed, 2012-07-11 at 16:31 +0200, Frederic Weisbecker wrote:
> > On Wed, Jul 11, 2012 at 06:14:58PM +0400, Andrew Vagin wrote:
> > > A few events are interesting not only for a current task.
> > > For example, sched_stat_* are interesting for a task, which
> > > wake up. For this reason, it will be good, if such events will
> > > be delivered to a target task too.
> > > 
> > > Now a target task can be set by using __perf_task().
> > > 
> > > The original idea and a draft patch belongs to Peter Zijlstra.
> > > 
> > > I need this events for profiling sleep times.  sched_switch is used for
> > > getting callchains and sched_stat_* is used for getting time periods.
> > > This events are combined in user space, then it can be analized by
> > > perf tools.
> > 
> > We've talked about that numerous times. But I still don't really
> > understand why you're not using sched switch events and compute
> > the difference between schedule in and schedule out.
> > 
> > I think you said that's because you got too much events with sched
> > switch. Are you loosing events? Otherwise I don't see why it's
> > a problem.
> > 
> > Also the sched_stat_sleep event produce an event which period equals the
> > time slept. Internally, perf split this into as many events as that period
> > because the requested period for trace events is 1 by default. We probably
> > should allow to send events with a higher number than the one requested. This
> > this produce sometimes a huge pile of events, and that even often result in
> > tons of lost events. We definetly need to fix that.
> > 
> > In the meantime you'll certainly get saner results by just recording
> > sched switch events.
> 
> Not really, there's an arbitrary large delay between wakeup and getting
> scheduled back in, which is unrelated to the cause that you went to
> sleep.
> 
> The wants the time between going to sleep and getting woken up,
> sched_switch simply doesn't give you that.

In this case he can just record sched wakeup as well. With sched_switch
+ sched_wakeup, he'll unlikely lose events.

With sched_stat_sleep he will lose events, unless we fix this period
demux thing.