Messages in this thread Patch in this message | | | Date | Wed, 3 Aug 2011 13:00:12 -0400 | From | Vince Weaver <> | Subject | Re: [perf] enable raw OFFCORE_EVENTS for non-perf userspace |
| |
and in case we don't want to be nice to userspace and want to brutally suppress anyone not using the "perf" tool, here's a patch that returns EINVAL for any raw access that sets config1.
I actually tested to make sure the patch doesn't break PAPI, unlike my last attempt.
Currently we "block" access to setting the config1 value from userspace by silently failing. This can cause unexpected behavior (especially if the kernel is using the OFFCORE_RSP_0 msr internally) and also makes it impossible to test if OFFCORE_RESPONSE events are being blocked or not.
This hackish patch causes an EINVAL to propogate back to userspace by setting an invalid value to the config1 field.
I personally think this patch is a step backward, and the right thing to do is give users full access to their own hardware. But the current state of affairs is bad for tools like PAPI too.
diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index 4ee3abf..72baf25 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -608,8 +608,11 @@ static int x86_setup_perfctr(struct perf_event *event) * Do not allow config1 (extended registers) to propagate, * there's no sane user-space generalization yet: */ - if (attr->type == PERF_TYPE_RAW) - return 0; + if (attr->type == PERF_TYPE_RAW) { + /* I HATE THIS PATCH AND EVERYTHING IT STANDS FOR. -- VMW */ + event->attr.config1=0xffffffff; /* invalid, will be caught */ + return x86_pmu_extra_regs(event->attr.config, event); + } if (attr->type == PERF_TYPE_HW_CACHE) return set_ext_hw_attr(hwc, event);
| |