Messages in this thread | | | Date | Fri, 5 Aug 2011 11:49:42 +0200 | From | Ingo Molnar <> | Subject | Re: [perf] enable raw OFFCORE_EVENTS for non-perf userspace |
| |
* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> On Wed, 2011-08-03 at 12:05 -0400, Vince Weaver wrote: > > Hello > > > > I propose we just enable raw OFFCORE_EVENT support and get it over with. > > > > There is a lot of demand for this from PAPI users, and so we encourage > > them to apply the below patch. PAPI supports this out of the box. > > > > The current "block" against using this feature *DOES NOT WORK*. > > It silently fails if you try to use the config1 field to set it. > > > > Even worse, if some previous user has set the OFFCORE_RSP_0 msr > > (say by running "perf stat -e LLC-load-misses") then > > the msr *stays set* and if you try to set the config1 field on your own > > it looks like it worked, but instead it is using whatever value the > > kernel last used. > > > > So there's a lot of userspace confusion about this, and you can't even > > reliably tell if the feature is turned off or not because it fails > > silently in unpredictable ways. > > > > diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c > > index 4ee3abf..28f9ca9 100644 > > --- a/arch/x86/kernel/cpu/perf_event.c > > +++ b/arch/x86/kernel/cpu/perf_event.c > > @@ -604,12 +604,8 @@ static int x86_setup_perfctr(struct perf_event *event) > > return -EOPNOTSUPP; > > } > > > > - /* > > - * Do not allow config1 (extended registers) to propagate, > > - * there's no sane user-space generalization yet: > > - */ > > if (attr->type == PERF_TYPE_RAW) > > - return 0; > > + return x86_pmu_extra_regs(event->attr.config, event); > > > > if (attr->type == PERF_TYPE_HW_CACHE) > > return set_ext_hw_attr(hwc, event); > > > I'm inclined to merge this, aside from snb, the offcore stuff is > actually quite usable now. Ingo can we somehow persuade you?
Sure, i think we are almost there, the only missing bit i see is to add the now generalized NUMA node bits to tools/perf/ so that it can be tested easily and then we can enable the raw bits as well.
It was supposed to be added alongside of:
89d6c0b: perf, arch: Add generic NODE cache events
what happened to that? If someone wants to accelerate the raw events enabling then that should be an easy thing to fix.
Thanks,
Ingo
| |