Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752242AbdLEHMg (ORCPT ); Tue, 5 Dec 2017 02:12:36 -0500 Received: from mail-lf0-f66.google.com ([209.85.215.66]:38616 "EHLO mail-lf0-f66.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751112AbdLEHMc (ORCPT ); Tue, 5 Dec 2017 02:12:32 -0500 X-Google-Smtp-Source: AGs4zMbkxQ4pZkCNSnGRFT1QpqLRARfu2RC6ko8ywy9PEB/roUagS4MUwBc2XMkgFOnPUyeJk8JgVoJ+slPEbIahjHk= MIME-Version: 1.0 In-Reply-To: <6fdad34d-1612-0447-e58e-5c748f92668d@linux.intel.com> References: <20171016183222.25750-1-ganapatrao.kulkarni@cavium.com> <20171016183222.25750-4-ganapatrao.kulkarni@cavium.com> <6fdad34d-1612-0447-e58e-5c748f92668d@linux.intel.com> From: Ganapatrao Kulkarni Date: Tue, 5 Dec 2017 12:42:30 +0530 Message-ID: Subject: Re: [PATCH v9 3/5] perf utils: use pmu->is_uncore to detect PMU UNCORE devices To: "Jin, Yao" Cc: Ganapatrao Kulkarni , linux-kernel@vger.kernel.org, linux-arm-kernel@lists.infradead.org, Will Deacon , catalin.marinas@arm.com, mark.rutland@arm.com, Arnaldo Carvalho de Melo , Alexander Shishkin , Peter Zijlstra , Ingo Molnar , jnair@caviumnetworks.com, Zhangshaokun , Jonathan.Cameron@huawei.com, Robert Richter , "Jin, Yao" Content-Type: text/plain; charset="UTF-8" Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 5945 Lines: 182 thanks Jin Yao for point this out. looks like logic of leveraging uncore device type(which i have changed in v9) does not go well with some json events of x86. can you please try below diff(logic used till v8), which keeps the original logic of identifying core/cpu PMUs. diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index 5ad8a18..57e38fd 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -538,6 +538,34 @@ static bool pmu_is_uncore(const char *name) } /* + * PMU CORE devices have different name other than cpu in sysfs on some + * platforms. looking for possible sysfs files to identify as core device. + */ +static int is_pmu_core(const char *name) +{ + struct stat st; + char path[PATH_MAX]; + const char *sysfs = sysfs__mountpoint(); + + if (!sysfs) + return 0; + + /* Look for cpu sysfs (x86 and others) */ + scnprintf(path, PATH_MAX, "%s/bus/event_source/devices/cpu", sysfs); + if ((stat(path, &st) == 0) && + (strncmp(name, "cpu", strlen("cpu")) == 0)) + return 1; + + /* Look for cpu sysfs (specific to arm) */ + scnprintf(path, PATH_MAX, "%s/bus/event_source/devices/%s/cpus", + sysfs, name); + if (stat(path, &st) == 0) + return 1; + + return 0; +} + +/* * Return the CPU id as a raw string. * * Each architecture should provide a more precise id string that @@ -641,7 +669,7 @@ static void pmu_add_cpu_aliases(struct list_head *head, struct perf_pmu *pmu) break; } - if (pmu->is_uncore) { + if (!is_pmu_core(name)) { /* check for uncore devices */ if (pe->pmu == NULL) continue; i have tried this diff on my x86 PC(haswell) and looks to be ok. please confirm your testing on skylake. gkulkarni@gkFc25>perf>> ./perf stat --per-thread -p 12663 -M CPI,IPC sleep 1 Performance counter stats for process id '12663': bash-12663 278,886 inst_retired.any:u bash-12663 482,284 cycles:u bash-12663 278,886 inst_retired.any:u bash-12663 483,597 cpu_clk_unhalted.thread:u 1.000923760 seconds time elapsed On Tue, Dec 5, 2017 at 7:42 AM, Jin, Yao wrote: > Hi Kulkarni, Arnaldo, > > This patch has been merged in perf/core branch today. > > But I see a regression issue when I run the 'perf stat'. > > With bisect checking, I locate to this patch. > > commit ad8737a08973f5dca632bdd63cf2abc99670e540 > Author: Ganapatrao Kulkarni > Date: Tue Oct 17 00:02:20 2017 +0530 > > perf pmu: Use pmu->is_uncore to detect UNCORE devices > > For example (on Intel skylake desktop), > > 1. The correct output should be (without this patch): > > root@skl:/tmp# perf stat --per-thread -p 1754 -M CPI,IPC > ^C > Performance counter stats for process id '1754': > > vmstat-1754 1,882,798 inst_retired.any # > 0.8 CPI > vmstat-1754 1,589,720 cycles > vmstat-1754 1,882,798 inst_retired.any # > 1.2 IPC > vmstat-1754 1,589,720 cpu_clk_unhalted.thread > > 2.647443167 seconds time elapsed > > 2. With this patch, the output will be: > > root@skl:/tmp# perf stat --per-thread -p 1754 -M CPI,IPC > ^C > Performance counter stats for process id '1754': > > vmstat-1754 1,945,589 inst_retired.any > vmstat-1754 inst_retired.any > vmstat-1754 1,609,892 cycles > vmstat-1754 1,945,589 inst_retired.any > vmstat-1754 inst_retired.any > vmstat-1754 1,609,892 cpu_clk_unhalted.thread > vmstat-1754 cpu_clk_unhalted.thread > > 3.051274166 seconds time elapsed > > Could you please help to take a look? > > Thanks > Jin Yao > > > On 10/17/2017 2:32 AM, Ganapatrao Kulkarni wrote: >> >> PMU CORE devices are identified using sysfs filename cpu, however >> on some platforms(like arm/arm64), PMU CORE sysfs name is not cpu. >> Hence cpu cannot be used to differentiate PMU CORE/UNCORE devices. >> >> commit: >> 66ec1191 ("perf pmu: Unbreak perf record for arm/arm64 with events with >> explicit PMU") >> >> has introduced pmu->is_uncore, which is set to PMU UNCORE devices only. >> Adding changes to use pmu->is_uncore to identify UNCORE devices. >> >> Acked-by: Will Deacon >> Tested-by: Shaokun Zhang >> Signed-off-by: Ganapatrao Kulkarni >> --- >> tools/perf/util/pmu.c | 11 +++++++---- >> 1 file changed, 7 insertions(+), 4 deletions(-) >> >> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c >> index 8b17db5..9110718 100644 >> --- a/tools/perf/util/pmu.c >> +++ b/tools/perf/util/pmu.c >> @@ -603,7 +603,6 @@ static void pmu_add_cpu_aliases(struct list_head >> *head, struct perf_pmu *pmu) >> */ >> i = 0; >> while (1) { >> - const char *pname; >> pe = &map->table[i++]; >> if (!pe->name) { >> @@ -612,9 +611,13 @@ static void pmu_add_cpu_aliases(struct list_head >> *head, struct perf_pmu *pmu) >> break; >> } >> - pname = pe->pmu ? pe->pmu : "cpu"; >> - if (strncmp(pname, name, strlen(pname))) >> - continue; >> + if (pmu->is_uncore) { >> + /* check for uncore devices */ >> + if (pe->pmu == NULL) >> + continue; >> + if (strncmp(pe->pmu, name, strlen(pe->pmu))) >> + continue; >> + } >> /* need type casts to override 'const' */ >> __perf_pmu__new_alias(head, NULL, (char *)pe->name, >> > thanks Ganapat