Message ID | 20130111191117.GA9407@us.ibm.com (mailing list archive) |
---|---|
State | Accepted, archived |
Headers | show |
On Fri, Jan 11, 2013 at 11:11:17AM -0800, Sukadev Bhattiprolu wrote: > If we disable a perf event because we exceeded the specified ->event_limit, > power_pmu_stop() sets the PERF_HES_STOPPED flag on the event. > > If the application then re-enables the event using PERF_EVENT_IOC_ENABLE > ioctl, we don't seem to ever clear this STOPPED flag. Consequently, the > user space is never notified of the event. > > Following message has more background and test case. > > http://lists.eecs.utk.edu/pipermail/ptools-perfapi/2012-October/002528.html > > The problem reported there does not seem to occur on x86. My unverified theory: > > Both x86 and Power clear the event->hw.state flag to 0 in their ->pmu_start() > operations. On X86 x86_pmu_start() is called from x86_pmu_enable(). But on > Power, power_pmu_start() is not called from power_pmu_enable(). This code has changed a lot since I worked on it, but it seems like x86 has the STOPPED flag set whenever the event isn't currently active on a hardware counter, whereas we have it set only when the event has been throttled. > Used the following test cases to verify that this patch works on latest PAPI. > > $ papi.git/src/ctests/nonthread PAPI_TOT_CYC@5000000 > > $ papi.git/src/ctests/overflow_single_event > > Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> > --- > arch/powerpc/perf/core-book3s.c | 8 ++++++++ > 1 files changed, 8 insertions(+), 0 deletions(-) > > diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c > index aa2465e..a6faada 100644 > --- a/arch/powerpc/perf/core-book3s.c > +++ b/arch/powerpc/perf/core-book3s.c > @@ -880,8 +880,16 @@ static int power_pmu_add(struct perf_event *event, int ef_flags) > cpuhw->events[n0] = event->hw.config; > cpuhw->flags[n0] = event->hw.event_base; > > + /* > + * If this event was disabled in record_and_restart() because we > + * exceeded the ->event_limit, this is probably a good time to > + * re-enable the event ? If we don't reenable the event, we will > + * never notify the user again about this event. > + */ The comment seems a bit tentative. :) If the PERF_EF_START bit is set then we are being told to restart the event. > if (!(ef_flags & PERF_EF_START)) > event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE; > + else > + event->hw.state &= ~PERF_HES_STOPPED; This looks fine, though I think you could equally well just set event->hw.state to 0 in the else clause. That would clear the UPTODATE flag too, which is appropriate since we are about to put the event on a hardware counter. Paul.
Paul Mackerras [paulus@samba.org] wrote: | > + /* | > + * If this event was disabled in record_and_restart() because we | > + * exceeded the ->event_limit, this is probably a good time to | > + * re-enable the event ? If we don't reenable the event, we will | > + * never notify the user again about this event. | > + */ | | The comment seems a bit tentative. :) If the PERF_EF_START bit is set | then we are being told to restart the event. | | > if (!(ef_flags & PERF_EF_START)) | > event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE; | > + else | > + event->hw.state &= ~PERF_HES_STOPPED; | | This looks fine, though I think you could equally well just set | event->hw.state to 0 in the else clause. That would clear the | UPTODATE flag too, which is appropriate since we are about to put the | event on a hardware counter. Agree. I submitted a new patch with better comments and patch description and cleared the state to 0. Thanks for the review. Sukadev
diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c index aa2465e..a6faada 100644 --- a/arch/powerpc/perf/core-book3s.c +++ b/arch/powerpc/perf/core-book3s.c @@ -880,8 +880,16 @@ static int power_pmu_add(struct perf_event *event, int ef_flags) cpuhw->events[n0] = event->hw.config; cpuhw->flags[n0] = event->hw.event_base; + /* + * If this event was disabled in record_and_restart() because we + * exceeded the ->event_limit, this is probably a good time to + * re-enable the event ? If we don't reenable the event, we will + * never notify the user again about this event. + */ if (!(ef_flags & PERF_EF_START)) event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE; + else + event->hw.state &= ~PERF_HES_STOPPED; /* * If group events scheduling transaction was started,
If we disable a perf event because we exceeded the specified ->event_limit, power_pmu_stop() sets the PERF_HES_STOPPED flag on the event. If the application then re-enables the event using PERF_EVENT_IOC_ENABLE ioctl, we don't seem to ever clear this STOPPED flag. Consequently, the user space is never notified of the event. Following message has more background and test case. http://lists.eecs.utk.edu/pipermail/ptools-perfapi/2012-October/002528.html The problem reported there does not seem to occur on x86. My unverified theory: Both x86 and Power clear the event->hw.state flag to 0 in their ->pmu_start() operations. On X86 x86_pmu_start() is called from x86_pmu_enable(). But on Power, power_pmu_start() is not called from power_pmu_enable(). Used the following test cases to verify that this patch works on latest PAPI. $ papi.git/src/ctests/nonthread PAPI_TOT_CYC@5000000 $ papi.git/src/ctests/overflow_single_event Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> --- arch/powerpc/perf/core-book3s.c | 8 ++++++++ 1 files changed, 8 insertions(+), 0 deletions(-)