Message ID | 20110117161742.5feb3761@kryten (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
On Mon, 17 Jan 2011 16:17:42 +1100 Anton Blanchard <anton@samba.org> wrote: > > When profiling a benchmark that is almost 100% userspace, I noticed some > wildly inaccurate profiles that showed almost all time spent in the kernel. > Closer examination shows we were programming a tiny number of cycles into > the PMU after each overflow (about ~200 away from the next overflow). This > gets us stuck in a loop which we eventually break out of by throttling the > PMU (there are regular throttle/unthrottle events in the log). > > It looks like we aren't setting event->hw.last_period to something same > and the frequency to period calculations in perf are going haywire. With > the following patch we find the correct period after a few interrupts and > stay there. I also see no more throttle events. > > Signed-off-by: Anton Blanchard <anton@samba.org> > --- > > diff --git a/arch/powerpc/kernel/perf_event.c b/arch/powerpc/kernel/perf_event.c > index 5674807..ab6f6be 100644 > --- a/arch/powerpc/kernel/perf_event.c > +++ b/arch/powerpc/kernel/perf_event.c > @@ -1212,6 +1212,7 @@ static void record_and_restart(struct perf_event *event, unsigned long val, > if (left <= 0) > left = period; > record = 1; > + event->hw.last_period = event->hw.sample_period; > } > if (left < 0x80000000LL) > val = 0x80000000LL - left; > _______________________________________________ > Linuxppc-dev mailing list > Linuxppc-dev@lists.ozlabs.org > https://lists.ozlabs.org/listinfo/linuxppc-dev > Does perf_event_fsl_emb.c need this as well (it has almost the same record_and_restart code)? -Scott
On Mon, 2011-01-17 at 11:32 -0600, Scott Wood wrote: > > diff --git a/arch/powerpc/kernel/perf_event.c b/arch/powerpc/kernel/perf_event.c > > index 5674807..ab6f6be 100644 > > --- a/arch/powerpc/kernel/perf_event.c > > +++ b/arch/powerpc/kernel/perf_event.c > > @@ -1212,6 +1212,7 @@ static void record_and_restart(struct perf_event *event, unsigned long val, > > if (left <= 0) > > left = period; > > record = 1; > > + event->hw.last_period = event->hw.sample_period; > > } > > if (left < 0x80000000LL) > > val = 0x80000000LL - left; > Does perf_event_fsl_emb.c need this as well (it has almost the same > record_and_restart code)? I would think so.
diff --git a/arch/powerpc/kernel/perf_event.c b/arch/powerpc/kernel/perf_event.c index 5674807..ab6f6be 100644 --- a/arch/powerpc/kernel/perf_event.c +++ b/arch/powerpc/kernel/perf_event.c @@ -1212,6 +1212,7 @@ static void record_and_restart(struct perf_event *event, unsigned long val, if (left <= 0) left = period; record = 1; + event->hw.last_period = event->hw.sample_period; } if (left < 0x80000000LL) val = 0x80000000LL - left;
When profiling a benchmark that is almost 100% userspace, I noticed some wildly inaccurate profiles that showed almost all time spent in the kernel. Closer examination shows we were programming a tiny number of cycles into the PMU after each overflow (about ~200 away from the next overflow). This gets us stuck in a loop which we eventually break out of by throttling the PMU (there are regular throttle/unthrottle events in the log). It looks like we aren't setting event->hw.last_period to something same and the frequency to period calculations in perf are going haywire. With the following patch we find the correct period after a few interrupts and stay there. I also see no more throttle events. Signed-off-by: Anton Blanchard <anton@samba.org> ---