Message-ID: <20230727082852.916093-3-changbin.du@huawei.com>
Date:   Thu, 27 Jul 2023 16:28:50 +0800
From:   Changbin Du <changbin.du@...wei.com>
To:     Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>,
        Arnaldo Carvalho de Melo <acme@...nel.org>
CC:     Mark Rutland <mark.rutland@....com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Jiri Olsa <jolsa@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        Ian Rogers <irogers@...gle.com>,
        Adrian Hunter <adrian.hunter@...el.com>,
        <linux-perf-users@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        Hui Wang <hw.huiwang@...wei.com>,
        Changbin Du <changbin.du@...wei.com>,
        kernel test robot <yujie.liu@...el.com>
Subject: [PATCH v4 2/4] perf: add new option '--workload-attr' to set workload sched_policy/priority/cpumask

To get consistent benchmarking results, sometimes we need to set the
sched_policy/priority/cpumask 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.

Cc: kernel test robot <yujie.liu@...el.com>
Signed-off-by: Changbin Du <changbin.du@...wei.com>

---
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                 | 116 +++++++++++++++++++++++
 tools/perf/util/evlist.h                 |   3 +
 tools/perf/util/target.h                 |   9 ++
 7 files changed, 185 insertions(+)

diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Documentation/perf-record.txt
index 137771b36578..8ee5d60225ec 100644
--- a/tools/perf/Documentation/perf-record.txt
+++ b/tools/perf/Documentation/perf-record.txt
@@ -843,6 +843,13 @@ filtered through the mask provided by -C option.
 	only, as of now.  So the applications built without the frame
 	pointer might see bogus addresses.
 
+--workload-attr <sched_policy[,priority][,cpu-list]>::
+	setup target workload (the <command>) 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[]
 
 SEE ALSO
diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/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 $?
 
+--workload-attr <sched_policy[,priority][,cpu-list]>::
+	setup target workload (the <command>) 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
 
 --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 option *opt,
 	return 0;
 }
 
+static int record_parse_workload_attr_opt(const struct option *opt,
+					  const char *arg,
+					  int unset __maybe_unused)
+{
+	struct record_opts *opts = opt->value;
+
+	return evlist__parse_workload_attr(arg, &opts->target.workload.sched_policy,
+					   &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 = {
 		.target		     = {
 			.uses_mmap   = true,
 			.default_per_cpu = true,
+			.workload.sched_policy = -1,
+			.workload.sched_priority = 0,
 		},
 		.mmap_flush          = MMAP_FLUSH_DEFAULT,
 		.nr_threads_synthesize = 1,
@@ -3333,6 +3346,12 @@ static struct record record = {
 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
 	"\n\t\t\t\tDefault: fp";
 
+const char record_workload_attr_help[] =
+	"setup target workload (the <command>) 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;
 
 static struct parse_events_option_args parse_events_option_args = {
@@ -3551,6 +3570,9 @@ static struct option __record_options[] = {
 			    "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()
 };
 
@@ -4266,6 +4288,10 @@ int cmd_record(int argc, const char **argv)
 	record__free_thread_masks(rec, rec->nr_threads);
 	rec->nr_threads = 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 = NULL;
+	}
 	return err;
 }
 
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 = true;
 
 static struct target target = {
 	.uid	= UINT_MAX,
+	.workload.sched_policy = -1,
+	.workload.sched_priority = 0,
 };
 
 #define METRIC_ONLY_LEN 20
@@ -1160,6 +1162,14 @@ static int parse_cache_level(const struct option *opt,
 	return 0;
 }
 
+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[] = {
 	OPT_BOOLEAN('T', "transaction", &transaction_run,
 		    "hardware transaction statistics"),
@@ -1220,6 +1230,9 @@ static struct option stat_options[] = {
 	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_config.ctl_fd_close);
 
+	if (target.workload.cpu_map) {
+		perf_cpu_map__put(target.workload.cpu_map);
+		target.workload.cpu_map = NULL;
+	}
+
 	return status;
 }
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 7ef43f72098e..6a9ff6a331aa 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -46,6 +46,7 @@
 #include <sys/mman.h>
 #include <sys/prctl.h>
 #include <sys/timerfd.h>
+#include <sys/resource.h>
 
 #include <linux/bitops.h>
 #include <linux/hash.h>
@@ -1398,6 +1399,118 @@ int evlist__open(struct evlist *evlist)
 	return err;
 }
 
+int evlist__parse_workload_attr(const char *str, int *sched_policy, int *sched_priority,
+				struct perf_cpu_map **cpu_map)
+{
+	char *tok, *saveptr = NULL;
+	char *buf;
+	int ret = -1;
+
+	/* We need buffer that we know we can write to. */
+	buf = strdup(str);
+	if (!buf)
+		return -ENOMEM;
+
+	tok = strtok_r((char *)buf, ",", &saveptr);
+	do {
+		/* sched policy */
+		if (!strncmp(tok, "other", sizeof("other")))
+			*sched_policy = SCHED_OTHER;
+		else if (!strncmp(tok, "fifo", sizeof("fifo"))) {
+			*sched_policy = SCHED_FIFO;
+			/* default to lowest priority */
+			*sched_priority = 99;
+		} else if (!strncmp(tok, "rr", sizeof("rr"))) {
+			*sched_policy = SCHED_RR;
+			*sched_priority = 99;
+		} else if (!strncmp(tok, "batch", sizeof("batch")))
+			*sched_policy = SCHED_BATCH;
+		else if (!strncmp(tok, "idle", sizeof("idle")))
+			*sched_policy = SCHED_IDLE;
+		else {
+			pr_err("workload_attr: unknown sched policy %s\n", tok);
+			break;
+		}
+
+		/* sched priority */
+		tok = strtok_r(NULL, ",", &saveptr);
+		if (tok) {
+			int priority;
+			char *endptr;
+
+			priority = strtol(tok, &endptr, 0);
+			if (*endptr) {
+				pr_err("workload_attr: invalid sched priority %s\n", tok);
+				break;
+			}
+
+			if (*sched_policy == SCHED_FIFO || *sched_policy == 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 = priority;
+		}
+
+		/* cpu list */
+		tok = strtok_r(NULL, ",", &saveptr);
+		if (tok) {
+			*cpu_map = __perf_cpu_map__new(tok, ':');
+			if (!*cpu_map) {
+				pr_err("workload_attr: failed to get cpus map from %s\n", tok);
+				break;
+			}
+		}
+		ret = 0;
+	} while (0);
+
+	free(buf);
+	return ret;
+}
+
+static int setup_workload_attr(struct target *target)
+{
+	struct sched_param param;
+	int policy = target->workload.sched_policy;
+	int priority = target->workload.sched_priority;
+
+	if (policy >= 0) {
+		param.sched_priority = (policy == SCHED_FIFO || policy == SCHED_RR) ?
+					priority : 0;
+		if (sched_setscheduler(0, policy, &param) != 0) {
+			perror("failed to set the sched policy");
+			return -1;
+		}
+
+		if (policy == SCHED_OTHER) {
+			if (setpriority(PRIO_PROCESS, 0, priority) != 0) {
+				perror("failed to set the nice value");
+				return -1;
+			}
+		}
+	}
+
+	if (target->workload.cpu_map) {
+		size_t cpuset_size = -1;
+		cpu_set_t *cpu_set;
+
+		cpu_set = perf_cpu_map__2_cpuset(target->workload.cpu_map, &cpuset_size);
+		if (!cpu_set)
+			return -1;
+
+		if (sched_setaffinity(0, cpuset_size, cpu_set) != 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, void *ucontext))
 {
@@ -1464,6 +1577,9 @@ int evlist__prepare_workload(struct evlist *evlist, struct target *target, const
 			exit(ret);
 		}
 
+		if (setup_workload_attr(target) != 0)
+			exit(-1);
+
 		execvp(argv[0], (char **)argv);
 
 		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 <pthread.h>
 #include <signal.h>
 #include <unistd.h>
+#include <sched.h>
 
 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, struct callchain_param *callchain);
 int record_opts__config(struct record_opts *opts);
 
+int evlist__parse_workload_attr(const char *str, int *sched_policy, int *sched_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 @@
 
 #include <stdbool.h>
 #include <sys/types.h>
+#include <sched.h>
 
 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;
 };
 
 enum target_errno {
@@ -103,4 +110,6 @@ static inline bool target__uses_dummy_map(struct target *target)
 	return use_dummy;
 }
 
+extern const char record_workload_attr_help[];
+
 #endif /* _PERF_TARGET_H */
-- 
2.25.1

Powered by blists - more mailing lists