Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932650AbaJVVEd (ORCPT ); Wed, 22 Oct 2014 17:04:33 -0400 Received: from mail-oi0-f48.google.com ([209.85.218.48]:44372 "EHLO mail-oi0-f48.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932270AbaJVVEb (ORCPT ); Wed, 22 Oct 2014 17:04:31 -0400 MIME-Version: 1.0 In-Reply-To: <20141022091239.GG12706@worktop.programming.kicks-ass.net> References: <1412872486-2930-1-git-send-email-eranian@google.com> <20141021112501.GY23531@worktop.programming.kicks-ass.net> <20141021130320.GE12706@worktop.programming.kicks-ass.net> <20141022091239.GG12706@worktop.programming.kicks-ass.net> Date: Wed, 22 Oct 2014 23:04:31 +0200 Message-ID: Subject: Re: [PATCH v2 0/12] perf/x86: implement HT leak workaround for SNB/IVB/HSW From: Stephane Eranian To: Peter Zijlstra Cc: LKML , "mingo@elte.hu" , "ak@linux.intel.com" , Jiri Olsa , "Liang, Kan" , Borislav Petkov , Maria Dimakopoulou Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Wed, Oct 22, 2014 at 11:12 AM, Peter Zijlstra wrote: > On Tue, Oct 21, 2014 at 03:08:32PM +0200, Stephane Eranian wrote: >> On Tue, Oct 21, 2014 at 3:03 PM, Peter Zijlstra wrote: >> > On Tue, Oct 21, 2014 at 02:28:06PM +0200, Stephane Eranian wrote: >> >> Peter, >> >> >> >> On Tue, Oct 21, 2014 at 1:25 PM, Peter Zijlstra wrote: >> >> > >> >> > >> >> > lkml.kernel.org/r/CABPqkBRbst4sgpgE5O_VXt-CSC0VD=aP2KWA0e3Uy64tw7df3A@mail.gmail.com >> >> > >> >> > I missed that 3 lines if they were in here. >> >> > >> >> I did not put them in there because there is another problem. >> >> If you partition the generic counters 2 and 2, then some CPUs will not >> >> be able to measure some events. >> >> Unfortunately, there is no way to partition the 4 counters such that >> >> all the events can be measured by >> >> each CPU. Some events or precise sampling requires counter 2 for >> >> instance (like prec_dist). >> >> That's why I did not put this fix in. >> > >> > Ah, I wasn't thinking about a hard partition, just a limit on the number >> > of exclusive counters any one CPU can claim such as to not starve. Or is >> > that what you were talking about? I feel not being able to starve >> > another CPU is more important than a better utilization bound for >> > counter scheduling. >> >> So you're saying, just limit number of used counters to 2 regardless >> of which one they are. > > used as in marked exclusive and forced empty on the other side. > >> So sometimes, this will avoid the problem aforementioned and sometimes >> not. We can try that. > > How will this sometimes not avoid the starvation issue? Here is a simple case: Limiting each HT to only 2 counters, can be any, 2 out of 4 possible. HT0: you measure a MEM* in ctr2, it is started first, and it keeps running HT1: you measure PREC_DIST with PEBS (it requires ctr2) HT0 is measuring a corrupting event on ctr2, this prevents ctr2 on HT1 from being used. HT1 is starved, it cannot measure PREC_DIST Yes you have a quota of 2 out of 4 counters. The quota dynamic or static can help mitigate the starvation. The only way to eliminate it is to force multiplexing even though you are using fewer counters than actually avail. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/