From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755115Ab3CONpr (ORCPT ); Fri, 15 Mar 2013 09:45:47 -0400 Received: from multi.imgtec.com ([194.200.65.239]:34593 "EHLO multi.imgtec.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754268Ab3CONof (ORCPT ); Fri, 15 Mar 2013 09:44:35 -0400 From: James Hogan To: CC: James Hogan , Peter Zijlstra , Paul Mackerras , Ingo Molnar , Arnaldo Carvalho de Melo Subject: [PATCH 4/9] metag: perf: add missing prev_count updates Date: Fri, 15 Mar 2013 13:43:43 +0000 Message-ID: <1363355028-23278-5-git-send-email-james.hogan@imgtec.com> X-Mailer: git-send-email 1.8.1.2 In-Reply-To: <1363355028-23278-1-git-send-email-james.hogan@imgtec.com> References: <1363355028-23278-1-git-send-email-james.hogan@imgtec.com> MIME-Version: 1.0 Content-Type: text/plain X-SEF-Processed: 7_3_0_01181__2013_03_15_13_44_32 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org The prev_count needs setting when changing the counter value, otherwise the calculated delta will be wrong, which for frequency sampling (dynamic period sampling) results in sampling at too high a frequency. For non-interrupting performance counters it should also be cleared when enabling the counter since the write to the PERF_COUNT register will clear the perf counter. This also includes a minor change to remove the u64 cast from the metag_pmu->write() call as metag_pmu->write() takes a u32 anyway, and in any case GCC is smart enough to optimise away the cast. Signed-off-by: James Hogan Cc: Peter Zijlstra Cc: Paul Mackerras Cc: Ingo Molnar Cc: Arnaldo Carvalho de Melo --- arch/metag/kernel/perf/perf_event.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/arch/metag/kernel/perf/perf_event.c b/arch/metag/kernel/perf/perf_event.c index a00f527..5bf984f 100644 --- a/arch/metag/kernel/perf/perf_event.c +++ b/arch/metag/kernel/perf/perf_event.c @@ -240,8 +240,10 @@ int metag_pmu_event_set_period(struct perf_event *event, if (left > (s64)metag_pmu->max_period) left = metag_pmu->max_period; - if (metag_pmu->write) - metag_pmu->write(idx, (u64)(-left) & MAX_PERIOD); + if (metag_pmu->write) { + local64_set(&hwc->prev_count, -(s32)left); + metag_pmu->write(idx, -left & MAX_PERIOD); + } perf_event_update_userpage(event); @@ -651,6 +653,12 @@ static void metag_pmu_enable_counter(struct hw_perf_event *event, int idx) * set to a specific value that needs preserving. */ tmp |= metag_in32(PERF_COUNT(idx)) & 0x00ffffff; + else + /* + * Older cores reset the counter on write, so prev_count needs + * resetting too so we can calculate a correct delta. + */ + local64_set(&event->prev_count, 0); metag_out32(tmp, PERF_COUNT(idx)); unlock: -- 1.8.1.2