Re: [PATCH v6 3/4] perf inject: add jitdump mmap injection support

From: Adrian Hunter
Date: Mon Apr 13 2015 - 03:06:31 EST


On 13/04/15 03:37, Stephane Eranian wrote:
> On Fri, Apr 10, 2015 at 5:51 AM, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>> On 08/04/15 17:12, Stephane Eranian wrote:
>>> On Wed, Apr 8, 2015 at 5:15 AM, Adrian Hunter <adrian.hunter@xxxxxxxxx> 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.
>>
>> Yes, but won't the jitdump timestamps already be in order?
>>
> I looked at implementing your approach and it turns out to be much
> more complicated
> than what is already there.
> I do not inject anything if I don't have to. I am waiting to see a
> special MMAP of the
> jitdump file to start inject jit mmaps. That could not change in your
> scheme. I would
> need to track MMAP records and check if they correspond to a jitdump. Note that
> I may have to inject multiple jitdumps in one run of perf inject. In
> that case, in the
> deliver function, I would have to (1) either have all the jitdump MMAPs prepared
> in order for all the jitdumps I have detected or (2) go through all
> the jitdump current
> head records to see which one matches. In (2) I would have to do this
> not just when
> I see an MMAP but for any event delivered: for each new event, compare its
> timestamp to the timestamps of jitdump current record for ALL jitdump
> files. For (1),
> I need to build a global list of all jitdump records in for each
> jitdump file, i.e, another
> RB tree. And then pull them one by one out if I find an event with the
> right timestamp.
>
> I think (2) i more reasonable but requires quite some infrastructure
> work in the jitdump
> code. Most of the pieces are there, the RB tree needs to be added
> obviously. I will
> experiment with (2).

I do something like (2) for Intel PT, but I use a heap to hold the next
lowest timestamp for each "queue" of data that needs to be processed -
because I have to allow for the possibility of a large number of "queues".

>
>
>>> 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) {
>>
>> (prev_event->timestamp <= j) must be always true because otherwise you would
>> have injected the jit mmap event already.
>>
>> So perf is about to deliver an event with timestamp 'event->timestamp', so
>> you just need to deliver all your events that have not already been
>> delivered but have a timestamp < 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 can sort them all at once, but finished_round is an optimization that
>> starts delivering "early" events before having to sort the "later" events.
>> The result is still that the delivered events are in sorted order.
>>
>>> 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?
>>
>> If you were to queue the events for sorting, yes. What I suggested was
>> delivering your events directly at the "right time".
>>
>>>
>>>> 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.
>>>>
>>>>
>>>
>>>
>>
>
>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/