On Mon, May 6, 2013 at 8:52 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > On Mon, May 06, 2013 at 07:44:19PM +0200, Stephane Eranian wrote: >> On Thu, May 2, 2013 at 9:37 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: >> > On Wed, Apr 24, 2013 at 04:04:54PM -0700, Andi Kleen wrote: >> >> From: Andi Kleen <ak@xxxxxxxxxxxxxxx> >> >> >> >> The PEBS documentation in the Intel SDM 18.6.1.1 states: >> >> >> >> """ >> >> PEBS events are only valid when the following fields of IA32_PERFEVTSELx are all >> >> zero: AnyThread, Edge, Invert, CMask. >> >> """ >> >> >> >> Since we had problems with this earlier, don't allow cmask, any, edge, invert >> >> as raw events, except for the ones explicitly listed as pebs_aliases. >> > >> > If its a simple matter of crap in crap out without affecting anything else we >> > shouldn't do anything. >> > >> The problem here is that you are sampling an instruction which did not cause >> the event you are measuring. Remember that using cmask, changes the >> nature of what's being measured (from event to cycles). > > Yeah.. I don't see the problem though. If you're using cmask and the like > you're supposed to know wth you're doing; which includes knowing your cpu and > what it thinks of such an event. > But that implies that you'd know that on Intel precise mode uses PEBS and that PEBS does not take cmask events. That seems to contradict the philosophy of perf_events where the kernel does the work for you. > The only reason to disallow events is if they (badly) interact with other > counters. -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html