Re: [PATCH 1/8] perf: Allow to block process in syscall tracepoints

From: Jiri Olsa <jolsa@redhat.com>
To: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Jiri Olsa <jolsa@kernel.org>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	lkml <linux-kernel@vger.kernel.org>,
	Ingo Molnar <mingo@kernel.org>,
	Namhyung Kim <namhyung@kernel.org>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	"Luis Claudio R. Goncalves" <lclaudio@uudg.org>,
	ldv@altlinux.org, esyr@redhat.com,
	Frederic Weisbecker <fweisbec@gmail.com>
Subject: Re: [PATCH 1/8] perf: Allow to block process in syscall tracepoints
Date: Fri, 7 Dec 2018 09:44:53 +0100	[thread overview]
Message-ID: <20181207084453.GA1727@krava> (raw)
In-Reply-To: <20181206131946.2c47f556@vmware.local.home>

On Thu, Dec 06, 2018 at 01:19:46PM -0500, Steven Rostedt wrote:
> On Thu, 6 Dec 2018 09:34:00 +0100
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > > 
> > > I don't understand this.. why are we using schedule_timeout() and all
> > > that?  
> > 
> > Urgh.. in fact, the more I look at this the more I hate it.
> > 
> > We want to block in __perf_output_begin(), but we cannot because both
> > tracepoints and perf will have preemptability disabled down there.
> > 
> > So what we do is fail the event, fake the lost count and go all the way
> > up that callstack, detect the failure and then poll-wait and retry.
> > 
> > And only do this for a few special events...  *yuck*
> 
> Since this is a special case, we should add a new option to the perf
> system call that, 1 states that it wants the traced process to block
> (and must have PTRACE permission to do so) and 2, after it reads from
> the buffer, it needs to check a bit that says "this process is blocked,
> please wake it up" and then do another perf call to kick the process to
> continue.

so instead of polling the traced process would properly wait for tracer
to kick him again after it reads/frees the buffer

I guess we could use the control mmap page (struct perf_event_mmap_page)
to communicate the 'we are block-ed' message to the tracer and have new
ioctl to wake the waiting process

jirka

> 
> I really dislike the polling too. But because this is not a default
> case, and is a new feature, we can add more infrastructure to make it
> work properly, instead of trying to hack the current method into
> something that does something poorly.
> 
> -- Steve