Messages in this thread | | | Subject | Re: [PATCH] perf_events: fix time tracking in samples | From | Peter Zijlstra <> | Date | Wed, 20 Oct 2010 13:13:16 +0200 |
| |
On Tue, 2010-10-19 at 21:03 +0200, Stephane Eranian wrote:
> >> Ok, I missed that. But I don't understand why you need the lock to > >> udpate the time. The lower-level clock is lockless if I recall. Can't you > >> use an atomic ops in update_context_time()? > > > > atomic ops would slow down those code paths, also, I don't think you can > > fully get the ordering between ->tstamp_$foo and ->total_time_$foo just > > right. > > > > I don't get that. Could you give an example?
Take update_context_time(), it has:
now = perf_clock(); ctx->time += now - ctx->timestamp; ctx->timestamp = now;
If you interleave two of those you get:
ctx->timestamp = T0;
now = perf_clock(); /* T1 */ ctx->time += now - ctx->timestamp; now = perf_clock(); /* T2 */ ctx->time += now - ctx->timestamp; ctx->timestamp = now; ctx->timestamp = now;
So at this point you would expect timestamp = T2 and time += T2-T0.
Except that: time += T1 - T0 + T2 - T0 != T2 - T0 and timestamp = T1
You can of course write it as something like x86_perf_event_update(), but then there's trying to keep total_time_running and total_time_enabled in sync.
> > Not sure, but barring 64bit atomics for all these, 32bit archs and NMI > > are going to be 'interesting' > > > > Every sample needs to be correct, otherwise you run the risk of introducing > bias. > > I think if the tradeoffs is correctness vs. speed, I'd choose correctness.
Well, yes, but it sucks, esp. since its only relevant to PERF_SAMPLE_READ.
| |