On Thu, Sep 13, 2018 at 09:40:42AM +0200, Peter Zijlstra wrote:
> On Wed, Sep 12, 2018 at 09:33:17PM +0200, Jiri Olsa wrote:
>
> > # perf record -e 'sched:sched_switch,sched:sched_wakeup' perf bench sched
> > messaging
>
> > The reason for the corruptions are some of the scheduling tracepoints,
> > that have __perf_task dfined and thus allow to store data to another
> > cpu ring buffer:
> >
> > sched_waking
> > sched_wakeup
> > sched_wakeup_new
> > sched_stat_wait
> > sched_stat_sleep
> > sched_stat_iowait
> > sched_stat_blocked
>
> > And then iterates events of the 'task' and store the sample
> > for any task's event that passes tracepoint checks:
> >
> > ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
> >
> > list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
> > if (event->attr.type != PERF_TYPE_TRACEPOINT)
> > continue;
> > if (event->attr.config != entry->type)
> > continue;
> >
> > perf_swevent_event(event, count, &data, regs);
> > }
> >
> > Above code can race with same code running on another cpu,
> > ending up with 2 cpus trying to store under the same ring
> > buffer, which is not handled at the moment.
>
> It can yes, however the only way I can see this breaking is if we use
> !inherited events with a strict per-task buffer, but your record command
> doesn't use that.
>
> Now, your test-case uses inherited events, which would all share the
> buffer, however IIRC inherited events require per-task-per-cpu buffers,
that's what perf record always does when monitoring task.. there's
an event/rb for each cpu and the given task
and all events for the task (sched:*) on given cpu share that single
cpu ring buffer via PERF_EVENT_IOC_SET_OUTPUT
> because there is already no guarantee the various tasks run on the same
> CPU in the first place.
>
> This means we _should_ write to the @task's local CPU buffer, and that
> would work again.
>
> Let me try and figure out where this is going wrong.
