From nobody Fri Sep 5 09:28:13 2025 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id BCB27EB64DA for ; Tue, 18 Jul 2023 03:34:35 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S229946AbjGRDed (ORCPT ); Mon, 17 Jul 2023 23:34:33 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:38844 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S229920AbjGRDeZ (ORCPT ); Mon, 17 Jul 2023 23:34:25 -0400 Received: from szxga02-in.huawei.com (szxga02-in.huawei.com [45.249.212.188]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 20979C0; Mon, 17 Jul 2023 20:34:22 -0700 (PDT) Received: from kwepemi500013.china.huawei.com (unknown [172.30.72.53]) by szxga02-in.huawei.com (SkyGuard) with ESMTP id 4R4kxb6vkPzNm84; Tue, 18 Jul 2023 11:30:59 +0800 (CST) Received: from M910t.huawei.com (10.110.54.157) by kwepemi500013.china.huawei.com (7.221.188.120) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2507.27; Tue, 18 Jul 2023 11:34:18 +0800 From: Changbin Du To: Peter Zijlstra , Ingo Molnar , Arnaldo Carvalho de Melo CC: Mark Rutland , Alexander Shishkin , Jiri Olsa , Namhyung Kim , Ian Rogers , Adrian Hunter , , , Hui Wang , Changbin Du Subject: [PATCH v3 2/3] perf: add new option '--workload-attr' to set workload sched_policy/priority/mask Date: Tue, 18 Jul 2023 11:33:54 +0800 Message-ID: <20230718033355.2960912-3-changbin.du@huawei.com> X-Mailer: git-send-email 2.25.1 In-Reply-To: <20230718033355.2960912-1-changbin.du@huawei.com> References: <20230718033355.2960912-1-changbin.du@huawei.com> MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable X-Originating-IP: [10.110.54.157] X-ClientProxiedBy: dggems701-chm.china.huawei.com (10.3.19.178) To kwepemi500013.china.huawei.com (7.221.188.120) X-CFilter-Loop: Reflected Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Type: text/plain; charset="utf-8" To get consistent benchmarking results, sometimes we need to set the sched_policy/priority/mask of the workload to reduce system noise. For example, CPU binding is required on big.little system. $ perf stat -- taskset -c 0 ls Nevertheless, the 'taskset' is also counted here. To get away of the middleman, this adds a new option '--workload-attr' to do the same jobs for stat and record commands. $ sudo perf stat --workload-attr fifo,40,0-3:7 -- ls Above will make 'ls' run on CPU #0-#3 and #7 with fifo scheduler and realtime priority is 40. Signed-off-by: Changbin Du --- v2: Use cpu list spec instead of cpu mask number. --- tools/perf/Documentation/perf-record.txt | 7 ++ tools/perf/Documentation/perf-stat.txt | 6 ++ tools/perf/builtin-record.c | 26 +++++ tools/perf/builtin-stat.c | 18 ++++ tools/perf/util/evlist.c | 117 +++++++++++++++++++++++ tools/perf/util/evlist.h | 3 + tools/perf/util/target.h | 9 ++ 7 files changed, 186 insertions(+) diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Document= ation/perf-record.txt index 680396c56bd1..9c01076f6c89 100644 --- a/tools/perf/Documentation/perf-record.txt +++ b/tools/perf/Documentation/perf-record.txt @@ -838,6 +838,13 @@ filtered through the mask provided by -C option. only, as of now. So the applications built without the frame pointer might see bogus addresses. =20 +--workload-attr :: + setup target workload (the ) attributes: + + sched_policy: other|fifo|rr|batch|idle + priority: scheduling priority for fifo|rr, nice value for other + cpu-list: CPU affinity. e.g. 1-2:4 is processors #1, #2, and #4 + include::intel-hybrid.txt[] =20 SEE ALSO diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentat= ion/perf-stat.txt index 8f789fa1242e..18d0f73458b1 100644 --- a/tools/perf/Documentation/perf-stat.txt +++ b/tools/perf/Documentation/perf-stat.txt @@ -262,6 +262,12 @@ disable events during measurements: wait -n ${perf_pid} exit $? =20 +--workload-attr :: + setup target workload (the ) attributes: + + sched_policy: other|fifo|rr|batch|idle + priority: scheduling priority for fifo|rr, nice value for other + cpu-list: CPU affinity. e.g. 1-2:4 is processors #1, #2, and #4 =20 --pre:: --post:: diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c index aec18db7ff23..c16e230afcd5 100644 --- a/tools/perf/builtin-record.c +++ b/tools/perf/builtin-record.c @@ -3289,6 +3289,17 @@ static int parse_record_synth_option(const struct op= tion *opt, return 0; } =20 +static int record_parse_workload_attr_opt(const struct option *opt, + const char *arg, + int unset __maybe_unused) +{ + struct record_opts *opts =3D opt->value; + + return evlist__parse_workload_attr(arg, &opts->target.workload.sched_poli= cy, + &opts->target.workload.sched_priority, + &opts->target.workload.cpu_map); +} + /* * XXX Ideally would be local to cmd_record() and passed to a record__new * because we need to have access to it in record__exit, that is called @@ -3309,6 +3320,8 @@ static struct record record =3D { .target =3D { .uses_mmap =3D true, .default_per_cpu =3D true, + .workload.sched_policy =3D -1, + .workload.sched_priority =3D 0, }, .mmap_flush =3D MMAP_FLUSH_DEFAULT, .nr_threads_synthesize =3D 1, @@ -3333,6 +3346,12 @@ static struct record record =3D { const char record_callchain_help[] =3D CALLCHAIN_RECORD_HELP "\n\t\t\t\tDefault: fp"; =20 +const char record_workload_attr_help[] =3D + "setup target workload (the ) attributes:\n\n" + HELP_PAD "sched_policy: other|fifo|rr|batch|idle\n" + HELP_PAD "priority: scheduling priority for fifo|rr, nice value for other= \n" + HELP_PAD "cpu-list: CPU affinity. e.g. 1-2:4 is processors #1, #2, and #4= "; + static bool dry_run; =20 static struct parse_events_option_args parse_events_option_args =3D { @@ -3551,6 +3570,9 @@ static struct option __record_options[] =3D { "write collected trace data into several data files using parallel = threads", record__parse_threads), OPT_BOOLEAN(0, "off-cpu", &record.off_cpu, "Enable off-cpu analysis"), + OPT_CALLBACK(0, "workload-attr", &record.opts, + "sched_policy[,priority][,cpu-list]", record_workload_attr_help, + &record_parse_workload_attr_opt), OPT_END() }; =20 @@ -4266,6 +4288,10 @@ int cmd_record(int argc, const char **argv) record__free_thread_masks(rec, rec->nr_threads); rec->nr_threads =3D 0; evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.= ctl_fd_close); + if (rec->opts.target.workload.cpu_map) { + perf_cpu_map__put(rec->opts.target.workload.cpu_map); + rec->opts.target.workload.cpu_map =3D NULL; + } return err; } =20 diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 07b48f6df48e..bb755cb44103 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -108,6 +108,8 @@ static bool all_counters_use_bpf =3D true; =20 static struct target target =3D { .uid =3D UINT_MAX, + .workload.sched_policy =3D -1, + .workload.sched_priority =3D 0, }; =20 #define METRIC_ONLY_LEN 20 @@ -1160,6 +1162,14 @@ static int parse_cache_level(const struct option *op= t, return 0; } =20 +static int parse_workload_attr_opt(const struct option *opt __maybe_unused= , const char *arg, + int unset __maybe_unused) +{ + return evlist__parse_workload_attr(arg, &target.workload.sched_policy, + &target.workload.sched_priority, + &target.workload.cpu_map); +} + static struct option stat_options[] =3D { OPT_BOOLEAN('T', "transaction", &transaction_run, "hardware transaction statistics"), @@ -1220,6 +1230,9 @@ static struct option stat_options[] =3D { OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), OPT_INTEGER(0, "log-fd", &output_fd, "log output to fd, instead of stderr"), + OPT_CALLBACK(0, "workload-attr", &stat_config, + "sched_policy[,priority][,cpu-list]", record_workload_attr_help, + &parse_workload_attr_opt), OPT_STRING(0, "pre", &pre_cmd, "command", "command to run prior to the measured command"), OPT_STRING(0, "post", &post_cmd, "command", @@ -2893,5 +2906,10 @@ int cmd_stat(int argc, const char **argv) metricgroup__rblist_exit(&stat_config.metric_events); evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_c= onfig.ctl_fd_close); =20 + if (target.workload.cpu_map) { + perf_cpu_map__put(target.workload.cpu_map); + target.workload.cpu_map =3D NULL; + } + return status; } diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c index 7ef43f72098e..cbdc2184f8a1 100644 --- a/tools/perf/util/evlist.c +++ b/tools/perf/util/evlist.c @@ -46,6 +46,7 @@ #include #include #include +#include =20 #include #include @@ -58,6 +59,7 @@ #include #include #include +#include =20 #include =20 @@ -1398,6 +1400,118 @@ int evlist__open(struct evlist *evlist) return err; } =20 +int evlist__parse_workload_attr(const char *str, int *sched_policy, int *s= ched_priority, + struct perf_cpu_map **cpu_map) +{ + char *tok, *saveptr =3D NULL; + char *buf; + int ret =3D -1; + + /* We need buffer that we know we can write to. */ + buf =3D strdup(str); + if (!buf) + return -ENOMEM; + + tok =3D strtok_r((char *)buf, ",", &saveptr); + do { + /* sched policy */ + if (!strncmp(tok, "other", sizeof("other"))) + *sched_policy =3D SCHED_OTHER; + else if (!strncmp(tok, "fifo", sizeof("fifo"))) { + *sched_policy =3D SCHED_FIFO; + /* default to lowest priority */ + *sched_priority =3D 99; + } else if (!strncmp(tok, "rr", sizeof("rr"))) { + *sched_policy =3D SCHED_RR; + *sched_priority =3D 99; + } else if (!strncmp(tok, "batch", sizeof("batch"))) + *sched_policy =3D SCHED_BATCH; + else if (!strncmp(tok, "idle", sizeof("idle"))) + *sched_policy =3D SCHED_IDLE; + else { + pr_err("workload_attr: unknown sched policy %s\n", tok); + break; + } + + /* sched priority */ + tok =3D strtok_r(NULL, ",", &saveptr); + if (tok) { + int priority; + char *endptr; + + priority =3D strtol(tok, &endptr, 0); + if (*endptr) { + pr_err("workload_attr: invalid sched priority %s\n", tok); + break; + } + + if (*sched_policy =3D=3D SCHED_FIFO || *sched_policy =3D=3D SCHED_RR) { + if (priority < 1 || priority > 99) { + pr_err("workload_attr: invalid priority %d for fifo and rr, allowed [= 1,99]\n", + priority); + break; + } + } + *sched_priority =3D priority; + } + + /* cpu list */ + tok =3D strtok_r(NULL, ",", &saveptr); + if (tok) { + *cpu_map =3D __perf_cpu_map__new(tok, ':'); + if (!*cpu_map) { + pr_err("workload_attr: failed to get cpus map from %s\n", tok); + break; + } + } + ret =3D 0; + } while (0); + + free(buf); + return ret; +} + +static int setup_workload_attr(struct target *target) +{ + struct sched_param param; + int policy =3D target->workload.sched_policy; + int priority =3D target->workload.sched_priority; + + if (policy >=3D 0) { + param.sched_priority =3D (policy =3D=3D SCHED_FIFO || policy =3D=3D SCHE= D_RR) ? + priority : 0; + if (sched_setscheduler(0, policy, ¶m) !=3D 0) { + perror("failed to set the sched policy"); + return -1; + } + + if (policy =3D=3D SCHED_OTHER) { + if (setpriority(PRIO_PROCESS, 0, priority) !=3D 0) { + perror("failed to set the nice value"); + return -1; + } + } + } + + if (target->workload.cpu_map) { + size_t cpuset_size =3D -1; + cpu_set_t *cpu_set; + + cpu_set =3D perf_cpu_map__2_cpuset(target->workload.cpu_map, &cpuset_siz= e); + if (!cpu_set) + return -1; + + if (sched_setaffinity(0, cpuset_size, cpu_set) !=3D 0) { + perror("failed to set the sched affinity"); + CPU_FREE(cpu_set); + return -1; + } + CPU_FREE(cpu_set); + } + + return 0; +} + int evlist__prepare_workload(struct evlist *evlist, struct target *target,= const char *argv[], bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, v= oid *ucontext)) { @@ -1464,6 +1578,9 @@ int evlist__prepare_workload(struct evlist *evlist, s= truct target *target, const exit(ret); } =20 + if (setup_workload_attr(target) !=3D 0) + exit(-1); + execvp(argv[0], (char **)argv); =20 if (exec_error) { diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h index 664c6bf7b3e0..761ae102dda6 100644 --- a/tools/perf/util/evlist.h +++ b/tools/perf/util/evlist.h @@ -15,6 +15,7 @@ #include #include #include +#include =20 struct pollfd; struct thread_map; @@ -180,6 +181,8 @@ void evlist__set_id_pos(struct evlist *evlist); void evlist__config(struct evlist *evlist, struct record_opts *opts, struc= t callchain_param *callchain); int record_opts__config(struct record_opts *opts); =20 +int evlist__parse_workload_attr(const char *str, int *sched_policy, int *s= ched_priority, + struct perf_cpu_map **cpu_set); int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[], bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext)); diff --git a/tools/perf/util/target.h b/tools/perf/util/target.h index d582cae8e105..16fe33dda2ef 100644 --- a/tools/perf/util/target.h +++ b/tools/perf/util/target.h @@ -4,6 +4,7 @@ =20 #include #include +#include =20 struct target { const char *pid; @@ -19,6 +20,12 @@ struct target { bool use_bpf; int initial_delay; const char *attr_map; + + struct { + int sched_policy; + int sched_priority; + struct perf_cpu_map *cpu_map; + } workload; }; =20 enum target_errno { @@ -103,4 +110,6 @@ static inline bool target__uses_dummy_map(struct target= *target) return use_dummy; } =20 +extern const char record_workload_attr_help[]; + #endif /* _PERF_TARGET_H */ --=20 2.25.1