lkml.org 
[lkml]   [2010]   [Dec]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3/3] perf record/report: Process events in order
On Mon, 6 Dec 2010, Ian Munsie wrote:

> Excerpts from Thomas Gleixner's message of Mon Dec 06 20:20:06 +1100 2010:
> > See https://lkml.org/lkml/2010/12/5/45
> >
> > Slightly different, but the same idea :)
>
> Yeah I noted that in my cover email. I meant to get my patches out
> before the weekend, but hadn't quite finished cleaning them up to ensure
> they didn't break perf when running on a kernel without sample_id_all.
>
> > > + case PERF_RECORD_HEADER_ATTR:
> > > + return ops->attr(event, s);
> > > + case PERF_RECORD_HEADER_EVENT_TYPE:
> > > + return ops->event_type(event, s);
> > > + case PERF_RECORD_HEADER_TRACING_DATA:
> > > + return ops->tracing_data(event, s);
> > > + case PERF_RECORD_HEADER_BUILD_ID:
> > > + return ops->build_id(event, s);
> >
> > These can be processed unordered.
>
> I just moved them into this routine to keep all the dispatching in one
> place, whether delayed or not. These particular events will still be
> processed immediately when encountered in the file. Only >=
> PERF_RECORD_MMAP && <= PERF_RECORD_SAMPLE will be delayed via the
> perf_session__process_timed function.

Gah. This is nasty. I really prefer the explicit split of instant
processed and possibly delayed events. That makes the code clear and
easy to extend. I just want to add a new event type at the right place
and not worry about magic comparisions in some other place.

> > > {
> > > + if (ops->ordered_samples && sample->time == -1ULL) {
> > > + dump_printf("Event missing timestamp, switching to unordered processing\n");
> > > + flush_sample_queue(s, ops);
> > > + ops->ordered_samples = false;
> >
> > Why ? The events injected by perf record itself have no timestamps and
> > do not need them. So why disabling ordered_samples ?
>
> For instance, suppose we ran this on an old kernel without support for
> timestamps on every event (so timestamps are only on sample events):
>
> perf record -T
> perf report
>
> If perf report tried to process the events in order, all the events
> without timestamps would be processed first -- including the
> PERF_RECORD_EXIT event, which would cause every sample not to be
> attributed. Falling back means we should get no worse than the old
> behaviour, while an upgraded kernel will provide the timestamps and
> should not fall back.

Ok, but you'll break existing code which does only care about sample
ordering if you do that at the session level unconditionally.

Thanks,

tglx


\
 
 \ /
  Last update: 2010-12-06 14:07    [W:0.037 / U:0.392 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site