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

From: Jiri Olsa
Date: Fri Dec 07 2018 - 03:45:02 EST


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@xxxxxxxxxxxxx> 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