lkml.org 
[lkml]   [2015]   [Apr]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v6 3/4] perf inject: add jitdump mmap injection support
From
On Wed, Apr 8, 2015 at 5:15 AM, Adrian Hunter <adrian.hunter@intel.com> wrote:
> On 06/04/15 22:41, Stephane Eranian wrote:
>> > + if (inject.build_ids) {
>> > + /*
>> > + * to make sure the mmap records are ordered correctly
>> > + * and so that the correct especially due to jitted code
>> > + * mmaps. We cannot generate the buildid hit list and
>> > + * inject the jit mmaps at the same time for now.
>> > + */
>> > + inject.tool.ordered_events = true;
>> > + inject.tool.ordering_requires_timestamps = true;
>> > + }
>> > +
>> > + if (inject.jit_mode) {
>> > + inject.tool.mmap2 = perf_event__repipe_mmap2;
>> > + inject.tool.mmap = perf_event__repipe_mmap;
>>
>> As suggested above, why not make your own tool fns e.g.
>>
>> inject.tool.mmap2 = perf_event__jit_mode_mmap2;
>> inject.tool.mmap = perf_event__jit_mode_mmap;
>>
>>
>> > + inject.tool.ordered_events = true;
>> > + inject.tool.ordering_requires_timestamps = true;
>>
>> You are taking advantage of a bug in perf-inject, that is the
>> "finished_round" events are not being processed. Really they should be
>> processed and you should inject in time order.
>>
>> I am not sure I understand.
>> Yes, I am trying to have inject reorder the samples instead of perf report.
>> You are likely to run inject once, and report many times. Also avoids a
>> warning in report about out-of-order events.
>
> Well forgetting about "finished_round", it seems to me you need to intercept
> all the delivered events (which will be in time order) and inject your own
> events at the right time.
>
> At the moment it seems to me you are injecting all your events in one go
> when you see the special jitdump mmap. So I would not expect the injected
> events to be ordered with respect to other events that come later. But
> maybe I misunderstand that?
>
My understanding is that if I set ordered_events = true, then events
are not delivered immediately to the callbacks. They are queued and
sorted and then passed to the callbacks. And yes, there is the finished
round mechanism of which I don't quite fully understand the logic in this
case.

> As I confusingly tried to suggest earlier, one way to see all the
> delivered events is to hook the ordered_events "deliver" callback. That
> will mean injecting one mmap event at a time.
>
> Here is just an idea.
>
> struct perf_inject {
> ...
> ordered_events__deliver_t deliver;
> };
>
> int cmd_inject(int argc, const char **argv, const char *prefix __maybe_unused)
> {
> ...
> inject.deliver = inject.session->ordered_events.deliver;
> inject.session->ordered_events.deliver = inject_jit_mmap;
> ...
> }
>
ok on that.

> int inject_jit_mmap(struct ordered_events *oe, struct ordered_event *event)
> {
> struct perf_session *session = container_of(oe, struct perf_session, ordered_events);
> struct perf_inject *inject = container_of(session->tool, struct perf_inject, tool);
>
> /* Is it time to inject an event */
> if (jit_next_timestamp(inject) < event->timestamp) {
> /* Yes, so inject it by delivery */
> perf_session__deliver_synth_event(...);
> }
> inject->deliver(oe, event);
> }
>
That suggests I have buffered all the MMAPs synthesized from the jitdump and
I have some sorted queue based on jitdump timestamps. The test would have to
be more sophisticated that this. You'd want to insert at the right
time, i.e., you'd
have to track the previous and next timestamp in the stream:

retry:
j = jit_next_timestamp(inject->jit);
if (prev_event->timestamp <= j && j > next_event->timestamp) {
deliver_synth_event(inject->jit);
jit_next(inject->jit);
goto retry;
}

All of this is required because the finished round logic is such that perf
cannot sort all records at once. It does sort them by chunks. Thus I need
to make sure I inject in the right chunk otherwise it won't work. Am I reading
this right?

> You would also need to inject any remaining events right at the end.
>
> Note the advantage of "delivering" events is that the same approach works whatever the tool.
>
>


\
 
 \ /
  Last update: 2015-04-08 16:41    [W:0.146 / U:0.148 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site