Support standalone metrics and metric groups for perf

Add generic support for standalone metrics specified in JSON files to perf stat. A metric is a formula that uses multiple events to compute a higher level result (e.g. IPC).

For more complex metrics we need to have micro architecture specific knowledge, so it makes sense to tie metrics to JSON event lists.

Previously metrics were always tied to an event and automatically enabled with that event. But now change it that we can have standalone metrics. They are in the same JSON data structure as events, but don't have an event name, only a metric name.

We also allow to organize the metrics in metric groups, which allows a short cut to select several related metrics at once.

This patch kit adds the code to perf to manage metric groups

The first few patches are generic bug fixes and can be applied directly. Then there is a 'weak group' feature that is useful independently from metrics. After there are metrics specific patches.

The patches are available in

git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc.git perf/metric-group-6

The actual Intel JSON metrics are available in git as a separate pull request in

git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc.git perf/intel-json-metrics-2

Some example output:

% perf list metricgroup
..
Metric Groups:

DSB:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
FLOPS:
GFLOPs
[Giga Floating Point Operations Per Second]
Frontend:
IFetch_Line_Utilization
[Rough Estimation of fraction of fetched lines bytes that were likely consumed by program instructions]
Frontend_Bandwidth:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
Memory_BW:
MLP
[Memory-Level-Parallelism (average number of L1 miss demand load when there is at least 1 such miss)]

   % perf stat -M Summary --metric-only -a sleep 1

Performance counter stats for 'system wide':

Instructions CLKS CPU_Utilization GFLOPs SMT_2T_Utilization Kernel_Utilization
317614222.0 1392930775.0 0.0 0.0 0.2 0.1

1.001497549 seconds time elapsed

% perf stat -M GFLOPs flops

Performance counter stats for 'flops':

3,999,541,471 fp_comp_ops_exe.sse_scalar_single # 1.2 GFLOPs (66.65%)
14 fp_comp_ops_exe.sse_scalar_double (66.65%)
0 fp_comp_ops_exe.sse_packed_double (66.67%)
0 fp_comp_ops_exe.sse_packed_single (66.70%)
0 simd_fp_256.packed_double (66.70%)
0 simd_fp_256.packed_single (66.67%)

3.238372845 seconds time elapsed

v1: Initial post
v2: Address all review feedback (see individual patches)
BPF now works again.
Fix some bugs in perf list printing that I added last minute last time.
v3: Address all review feedback. Some patches are split. Rebased. Not caching cpuids because it's too complicated.

From: Andi Kleen <[email protected]>

Setting up groups can be complicated due to the complicated scheduling restrictions of different PMUs. User tools usually don't understand all these restrictions. Still in many cases it is useful to set up groups and they work most of the time. However if the group is set up wrong some members will not reported any values because they never get scheduled.

Add a concept of a 'weak group': try to set up a group, but if it's not schedulable fallback to not using
a group. That gives us the best of both worlds:
groups if they work, but still a usable fallback if they don't.

In theory it would be possible to have more complex fallback strategies (e.g. try to split the group in half), but the simple fallback of not using a group seems to work for now.

So far the weak group is only implemented for perf stat, not for record.

Here's an unschedulable group (on IvyBridge with SMT on)

% perf stat -e '{branches,branch-misses,l1d.replacement,l2_lines_in.all,l2_rqsts.all_code_rd}' -a sleep 1

        73,806,067      branches
         4,848,144      branch-misses             #    6.57% of all branches
        14,754,458      l1d.replacement
        24,905,558      l2_lines_in.all
   <not supported>      l2_rqsts.all_code_rd         <------- will never report anything

With the weak group:

% perf stat -e '{branches,branch-misses,l1d.replacement,l2_lines_in.all,l2_rqsts.all_code_rd}:W' -a sleep 1

       125,366,055      branches                                                      (80.02%)
         9,208,402      branch-misses             #    7.35% of all branches          (80.01%)
        24,560,249      l1d.replacement                                               (80.00%)
        43,174,971      l2_lines_in.all                                               (80.05%)
        31,891,457      l2_rqsts.all_code_rd                                          (79.92%)

The extra event scheduled with some extra multiplexing

v2: Move fallback code to separate function.
Add comment on for_each_group_member
Adjust to new perf_evsel__close interface
v3:
Fix debug print out.

---
 tools/perf/Documentation/perf-list.txt |  1 +
 tools/perf/builtin-stat.c              | 35 ++++++++++++++++++++++++++++++++++
 tools/perf/util/evsel.h                |  1 +
 tools/perf/util/parse-events.c         |  8 +++++++-
 tools/perf/util/parse-events.l         |  2 +-
 5 files changed, 45 insertions(+), 2 deletions(-)

index f709de54707b..d432965d728d 100644
@@ -47,6 +47,7 @@ counted. The following modifiers exist:
  P - use maximum detected precise level
  S - read sample value (PERF_SAMPLE_READ)
  D - pin the event to the PMU
+ W - group is weak and will fallback to non-group if not schedulable
 
 The 'p' modifier can be used for specifying how precise the instruction
 address should be. The 'p' modifier can be specified multiple times:

index 866da7aa54bf..501c1e0272fe 100644
@@ -582,6 +582,32 @@ static bool perf_evsel__should_store_id(struct perf_evsel *counter)
 	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
 }
 
+static struct perf_evsel *reset_weak_group(struct perf_evsel *counter)
+{
+	struct perf_evsel *c2, *leader;
+	bool is_open = true;
+
+	leader = counter->leader;
+	pr_debug("Weak group for %s/%d failed\n",
+			leader->name, leader->nr_members);
+
+	/*
+	 * for_each_group_member doesn't work here because it doesn't
+	 * include the first entry.
+	 */
+	evlist__for_each_entry(evsel_list, c2) {
+		if (c2 == counter)
+			is_open = false;
+		if (c2->leader == leader) {
+			if (is_open)
+				perf_evsel__close(c2);
+			c2->leader = c2;
+			c2->nr_members = 0;
+		}
+	}
+	return leader;
+}
+
 static int __run_perf_stat(int argc, const char **argv)
 {
 	int interval = stat_config.interval;

@@ -618,6 +644,15 @@ static int __run_perf_stat(int argc, const char **argv)
 	evlist__for_each_entry(evsel_list, counter) {
 try_again:
 		if (create_perf_stat_counter(counter) < 0) {
+
+			/* Weak group failed. Reset the group. */
+			if (errno == EINVAL &&
+			    counter->leader != counter &&
+			    counter->weak_group) {
+				counter = reset_weak_group(counter);
+				goto try_again;
+			}
+
 			/*
 			 * PPC returns ENXIO for HW counters until 2.6.37
 			 * (behavior changed with commit b0a873e).

index 351d3b2d8887..f538c3530227 100644
@@ -136,6 +136,7 @@ struct perf_evsel {
 	const char *		metric_name;
 	struct perf_evsel	**metric_events;
 	bool			collect_stat;
+	bool			weak_group;
 };
 
 union u64_swap {

index f44aeba51d1f..cd80c4eac569 100644
@@ -1358,6 +1358,7 @@ struct event_modifier {
 	int exclude_GH;
 	int sample_read;
 	int pinned;
+	int weak;
 };
 
 static int get_event_modifier(struct event_modifier *mod, char *str,

@@ -1376,6 +1377,7 @@ static int get_event_modifier(struct event_modifier *mod, char *str,
 
 	int exclude = eu | ek | eh;
 	int exclude_GH = evsel ? evsel->exclude_GH : 0;
+	int weak = 0;
 
 	memset(mod, 0, sizeof(*mod));
 

@@ -1413,6 +1415,8 @@ static int get_event_modifier(struct event_modifier *mod, char *str,
 			sample_read = 1;
 		} else if (*str == 'D') {
 			pinned = 1;
+		} else if (*str == 'W') {
+			weak = 1;
 		} else
 			break;
 

@@ -1443,6 +1447,7 @@ static int get_event_modifier(struct event_modifier *mod, char *str,
 	mod->exclude_GH = exclude_GH;
 	mod->sample_read = sample_read;
 	mod->pinned = pinned;
+	mod->weak = weak;
 
 	return 0;
 }

@@ -1456,7 +1461,7 @@ static int check_modifier(char *str)
 	char *p = str;
 
 	/* The sizeof includes 0 byte as well. */
-	if (strlen(str) > (sizeof("ukhGHpppPSDI") - 1))
+	if (strlen(str) > (sizeof("ukhGHpppPSDIW") - 1))
 		return -1;
 
 	while (*p) {

@@ -1496,6 +1501,7 @@ int parse_events__modifier_event(struct list_head *list, char *str, bool add)
 		evsel->exclude_GH          = mod.exclude_GH;
 		evsel->sample_read         = mod.sample_read;
 		evsel->precise_max         = mod.precise_max;
+		evsel->weak_group	   = mod.weak;
 
 		if (perf_evsel__is_group_leader(evsel))
 			evsel->attr.pinned = mod.pinned;

index c42edeac451f..fdb5bb52f01f 100644
@@ -161,7 +161,7 @@ name		[a-zA-Z_*?][a-zA-Z0-9_*?.]*
 name_minus	[a-zA-Z_*?][a-zA-Z0-9\-_*?.:]*
 drv_cfg_term	[a-zA-Z0-9_\.]+(=[a-zA-Z0-9_*?\.:]+)?
 /* If you add a modifier you need to update check_modifier() */
-modifier_event	[ukhpPGHSDI]+
+modifier_event	[ukhpPGHSDIW]+
 modifier_bp	[rwx]{1,3}
 
 %%

2.9.5

thanks,
jirka

just realized we support this in stat only and the doc indicates it's global.. maybe the perf-list.txt line could have the '(perf stat only)' suffix ;-)

jirka

Em Fri, Sep 01, 2017 at 07:00:39PM +0200, Jiri Olsa escreveu:

I'll try to add something to that effect.

- Arnaldo

From: Andi Kleen <[email protected]>

Some enhancements to the JSON parser to prepare for metrics support

- Parse the new MetricGroup field
- Support JSON events with no event name, that have only MetricName.

---
 tools/perf/pmu-events/jevents.c    | 24 ++++++++++++++++++------
 tools/perf/pmu-events/jevents.h    |  2 +-
 tools/perf/pmu-events/pmu-events.h |  1 +
 3 files changed, 20 insertions(+), 7 deletions(-)

index d51dc9ca8861..9eb7047bafe4 100644
@@ -292,7 +292,7 @@ static int print_events_table_entry(void *data, char *name, char *event,
 				    char *desc, char *long_desc,
 				    char *pmu, char *unit, char *perpkg,
 				    char *metric_expr,
-				    char *metric_name)
+				    char *metric_name, char *metric_group)
 {
 	struct perf_entry_data *pd = data;
 	FILE *outfp = pd->outfp;

@@ -304,8 +304,10 @@ static int print_events_table_entry(void *data, char *name, char *event,
 	 */
 	fprintf(outfp, "{\n");
 
-	fprintf(outfp, "\t.name = \"%s\",\n", name);
-	fprintf(outfp, "\t.event = \"%s\",\n", event);
+	if (name)
+		fprintf(outfp, "\t.name = \"%s\",\n", name);
+	if (event)
+		fprintf(outfp, "\t.event = \"%s\",\n", event);
 	fprintf(outfp, "\t.desc = \"%s\",\n", desc);
 	fprintf(outfp, "\t.topic = \"%s\",\n", topic);
 	if (long_desc && long_desc[0])

@@ -320,6 +322,8 @@ static int print_events_table_entry(void *data, char *name, char *event,
 		fprintf(outfp, "\t.metric_expr = \"%s\",\n", metric_expr);
 	if (metric_name)
 		fprintf(outfp, "\t.metric_name = \"%s\",\n", metric_name);
+	if (metric_group)
+		fprintf(outfp, "\t.metric_group = \"%s\",\n", metric_group);
 	fprintf(outfp, "},\n");
 
 	return 0;

@@ -357,6 +361,9 @@ static char *real_event(const char *name, char *event)
 {
 	int i;
 
+	if (!name)
+		return NULL;
+
 	for (i = 0; fixed[i].name; i++)
 		if (!strcasecmp(name, fixed[i].name))
 			return (char *)fixed[i].event;

@@ -369,7 +376,7 @@ int json_events(const char *fn,
 		      char *long_desc,
 		      char *pmu, char *unit, char *perpkg,
 		      char *metric_expr,
-		      char *metric_name),
+		      char *metric_name, char *metric_group),
 	  void *data)
 {
 	int err = -EIO;

@@ -397,6 +404,7 @@ int json_events(const char *fn,
 		char *unit = NULL;
 		char *metric_expr = NULL;
 		char *metric_name = NULL;
+		char *metric_group = NULL;
 		unsigned long long eventcode = 0;
 		struct msrmap *msr = NULL;
 		jsmntok_t *msrval = NULL;

@@ -476,6 +484,8 @@ int json_events(const char *fn,
 				addfield(map, &perpkg, "", "", val);
 			} else if (json_streq(map, field, "MetricName")) {
 				addfield(map, &metric_name, "", "", val);
+			} else if (json_streq(map, field, "MetricGroup")) {
+				addfield(map, &metric_group, "", "", val);
 			} else if (json_streq(map, field, "MetricExpr")) {
 				addfield(map, &metric_expr, "", "", val);
 				for (s = metric_expr; *s; s++)

@@ -501,10 +511,11 @@ int json_events(const char *fn,
 			addfield(map, &event, ",", filter, NULL);
 		if (msr != NULL)
 			addfield(map, &event, ",", msr->pname, msrval);
-		fixname(name);
+		if (name)
+			fixname(name);
 
 		err = func(data, name, real_event(name, event), desc, long_desc,
-				pmu, unit, perpkg, metric_expr, metric_name);
+			   pmu, unit, perpkg, metric_expr, metric_name, metric_group);
 		free(event);
 		free(desc);
 		free(name);

@@ -516,6 +527,7 @@ int json_events(const char *fn,
 		free(unit);
 		free(metric_expr);
 		free(metric_name);
+		free(metric_group);
 		if (err)
 			break;
 		tok += j;

index 611fac01913d..557994754410 100644
@@ -6,7 +6,7 @@ int json_events(const char *fn,
 				char *long_desc,
 				char *pmu,
 				char *unit, char *perpkg, char *metric_expr,
-				char *metric_name),
+				char *metric_name, char *metric_group),
 		void *data);
 char *get_cpu_str(void);
 

index 569eab3688dd..94fa1720f6fd 100644
@@ -15,6 +15,7 @@ struct pmu_event {
 	const char *perpkg;
 	const char *metric_expr;
 	const char *metric_name;
+	const char *metric_group;
 };
 
 /*

2.9.5

From: Andi Kleen <[email protected]>

The perf stat shadow metric printing already supports generic metrics. Factor out the code doing that into a separate function that can be re-used in a later patch.

No behavior changes.

v2: Fix indentation

---
 tools/perf/util/stat-shadow.c | 69 ++++++++++++++++++++++++++-----------------
 1 file changed, 42 insertions(+), 27 deletions(-)

index a04cf56d3517..96aa6cbf24d6 100644
@@ -627,6 +627,46 @@ static void print_smi_cost(int cpu, struct perf_evsel *evsel,
 	out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
 }
 
+static void generic_metric(const char *metric_expr,
+			   struct perf_evsel **metric_events,
+			   char *name,
+			   const char *metric_name,
+			   double avg,
+			   int cpu,
+			   int ctx,
+			   struct perf_stat_output_ctx *out)
+{
+	print_metric_t print_metric = out->print_metric;
+	struct parse_ctx pctx;
+	double ratio;
+	int i;
+	void *ctxp = out->ctx;
+
+	expr__ctx_init(&pctx);
+	expr__add_id(&pctx, name, avg);
+	for (i = 0; metric_events[i]; i++) {
+		struct saved_value *v;
+
+		v = saved_value_lookup(metric_events[i], cpu, ctx, false);
+		if (!v)
+			break;
+		expr__add_id(&pctx, metric_events[i]->name, avg_stats(&v->stats));
+	}
+	if (!metric_events[i]) {
+		const char *p = metric_expr;
+
+		if (expr__parse(&ratio, &pctx, &p) == 0)
+			print_metric(ctxp, NULL, "%8.1f",
+				metric_name ?
+				metric_name :
+				out->force_header ?  name : "",
+				ratio);
+		else
+			print_metric(ctxp, NULL, NULL, "", 0);
+	} else
+		print_metric(ctxp, NULL, NULL, "", 0);
+}
+
 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 				   double avg, int cpu,
 				   struct perf_stat_output_ctx *out)

@@ -819,33 +859,8 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 		else
 			print_metric(ctxp, NULL, NULL, name, 0);
 	} else if (evsel->metric_expr) {
-		struct parse_ctx pctx;
-		int i;
-
-		expr__ctx_init(&pctx);
-		expr__add_id(&pctx, evsel->name, avg);
-		for (i = 0; evsel->metric_events[i]; i++) {
-			struct saved_value *v;
-
-			v = saved_value_lookup(evsel->metric_events[i], cpu, ctx, false);
-			if (!v)
-				break;
-			expr__add_id(&pctx, evsel->metric_events[i]->name,
-					     avg_stats(&v->stats));
-		}
-		if (!evsel->metric_events[i]) {
-			const char *p = evsel->metric_expr;
-
-			if (expr__parse(&ratio, &pctx, &p) == 0)
-				print_metric(ctxp, NULL, "%8.1f",
-					evsel->metric_name ?
-					evsel->metric_name :
-					out->force_header ?  evsel->name : "",
-					ratio);
-			else
-				print_metric(ctxp, NULL, NULL, "", 0);
-		} else
-			print_metric(ctxp, NULL, NULL, "", 0);
+		generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
+				evsel->metric_name, avg, cpu, ctx, out);
 	} else if (runtime_nsecs_stats[cpu].n != 0) {
 		char unit = 'M';
 		char unit_buf[10];

2.9.5

From: Andi Kleen <[email protected]>

Print the generic metric header even when the expression evaluation failed. Otherwise an expression that fails on the first collections due to division by zero may suddenly reappear later without an header.

---
 tools/perf/util/stat-shadow.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

index 96aa6cbf24d6..8c7ab29169b9 100644
@@ -662,7 +662,9 @@ static void generic_metric(const char *metric_expr,
 				out->force_header ?  name : "",
 				ratio);
 		else
-			print_metric(ctxp, NULL, NULL, "", 0);
+			print_metric(ctxp, NULL, NULL,
+				     out->force_header ?
+				     (metric_name ? metric_name : name) : "", 0);
 	} else
 		print_metric(ctxp, NULL, NULL, "", 0);
 }

2.9.5

From: Andi Kleen <[email protected]>

Extract the code to get the per cpu json alias into a separate function for reuse. No behavior changes.

---
 tools/perf/util/pmu.c | 49 +++++++++++++++++++++++++++++++++----------------
 tools/perf/util/pmu.h |  2 ++
 2 files changed, 35 insertions(+), 16 deletions(-)

index ac16a9db1fb5..ed25d7f88731 100644
@@ -516,16 +516,8 @@ char * __weak get_cpuid_str(void)
 	return NULL;
 }
 
-/*
- * From the pmu_events_map, find the table of PMU events that corresponds
- * to the current running CPU. Then, add all PMU events from that table
- * as aliases.
- */
-static void pmu_add_cpu_aliases(struct list_head *head, const char *name)
+static char *perf_pmu__getcpuid(void)
 {
-	int i;
-	struct pmu_events_map *map;
-	struct pmu_event *pe;
 	char *cpuid;
 	static bool printed;
 

@@ -535,22 +527,50 @@ static void pmu_add_cpu_aliases(struct list_head *head, const char *name)
 	if (!cpuid)
 		cpuid = get_cpuid_str();
 	if (!cpuid)
-		return;
+		return NULL;
 
 	if (!printed) {
 		pr_debug("Using CPUID %s\n", cpuid);
 		printed = true;
 	}
+	return cpuid;
+}
+
+struct pmu_events_map *perf_pmu__find_map(void)
+{
+	struct pmu_events_map *map;
+	char *cpuid = perf_pmu__getcpuid();
+	int i;
 
 	i = 0;
-	while (1) {
+	for (;;) {
 		map = &pmu_events_map[i++];
-		if (!map->table)
-			goto out;
+		if (!map->table) {
+			map = NULL;
+			break;
+		}
 
 		if (!strcmp(map->cpuid, cpuid))
 			break;
 	}
+	free(cpuid);
+	return map;
+}
+
+/*
+ * From the pmu_events_map, find the table of PMU events that corresponds
+ * to the current running CPU. Then, add all PMU events from that table
+ * as aliases.
+ */
+static void pmu_add_cpu_aliases(struct list_head *head, const char *name)
+{
+	int i;
+	struct pmu_events_map *map;
+	struct pmu_event *pe;
+
+	map = perf_pmu__find_map();
+	if (!map)
+		return;
 
 	/*
 	 * Found a matching PMU events table. Create aliases

@@ -575,9 +595,6 @@ static void pmu_add_cpu_aliases(struct list_head *head, const char *name)
 				(char *)pe->metric_expr,
 				(char *)pe->metric_name);
 	}
-
-out:
-	free(cpuid);
 }
 
 struct perf_event_attr * __weak

index 389e9729331f..060f6abba8ed 100644
@@ -90,4 +90,6 @@ int perf_pmu__test(void);
 
 struct perf_event_attr *perf_pmu__get_default_config(struct perf_pmu *pmu);
 
+struct pmu_events_map *perf_pmu__find_map(void);
+
 #endif /* __PMU_H */

2.9.5

From: Andi Kleen <[email protected]>

Add generic support for standalone metrics specified in JSON files to perf stat. A metric is a formula that uses multiple events to compute a higher level result (e.g. IPC).

Previously metrics were always tied to an event and automatically enabled with that event. But now change it that we can have standalone metrics. They are in the same JSON data structure as events, but don't have an event name.

We also allow to organize the metrics in metric groups, which allows a short cut to select several related metrics at once.

Add a new -M / --metrics option to perf stat that adds the metrics or metric groups specified.

Add the core code to manage and parse the metric groups. They are collected from the JSON data structures into a separate rblist. When computing shadow values look for metrics in that list. Then they are computed using the existing saved values infrastructure in stat-shadow.c

The actual JSON metrics are in a separate pull request.

% perf stat -M Summary --metric-only -a sleep 1

Performance counter stats for 'system wide':

Instructions CLKS CPU_Utilization GFLOPs SMT_2T_Utilization Kernel_Utilization
317614222.0 1392930775.0 0.0 0.0 0.2 0.1

       1.001497549 seconds time elapsed

% perf stat -M GFLOPs flops

Performance counter stats for 'flops':

     3,999,541,471      fp_comp_ops_exe.sse_scalar_single #      1.2 GFLOPs                   (66.65%)
                14      fp_comp_ops_exe.sse_scalar_double                                     (66.65%)
                 0      fp_comp_ops_exe.sse_packed_double                                     (66.67%)
                 0      fp_comp_ops_exe.sse_packed_single                                     (66.70%)
                 0      simd_fp_256.packed_double                                     (66.70%)
                 0      simd_fp_256.packed_single                                     (66.67%)
                 0      duration_time

       3.238372845 seconds time elapsed

v2: Add missing header file
v3: Move find_map to pmu.c

---
 tools/perf/Documentation/perf-stat.txt |   7 +
 tools/perf/builtin-stat.c              |  18 +-
 tools/perf/util/Build                  |   1 +
 tools/perf/util/metricgroup.c          | 313 +++++++++++++++++++++++++++++++++
 tools/perf/util/metricgroup.h          |  31 ++++
 tools/perf/util/pmu.c                  |   5 +-
 tools/perf/util/stat-shadow.c          |  22 ++-
 tools/perf/util/stat.h                 |   4 +-
 8 files changed, 395 insertions(+), 6 deletions(-)
 create mode 100644 tools/perf/util/metricgroup.c
 create mode 100644 tools/perf/util/metricgroup.h

index c37d61682dfb..823fce7674bb 100644
@@ -199,6 +199,13 @@ Aggregate counts per processor socket for system-wide mode measurements.
 --per-core::
 Aggregate counts per physical processor for system-wide mode measurements.
 
+-M::
+--metrics::
+Print metrics or metricgroups specified in a comma separated list.
+For a group all metrics from the group are added.
+The events from the metrics are automatically measured.
+See perf list output for the possble metrics and metricgroups.
+
 -A::
 --no-aggr::
 Do not aggregate counts across all monitored CPUs.

index 501c1e0272fe..045efd7d7785 100644
@@ -65,6 +65,7 @@
 #include "util/tool.h"
 #include "util/group.h"
 #include "util/string2.h"
+#include "util/metricgroup.h"
 #include "asm/bug.h"
 
 #include <linux/time64.h>

@@ -133,6 +134,8 @@ static const char *smi_cost_attrs = {
 
 static struct perf_evlist	*evsel_list;
 
+static struct rblist		 metric_events;
+
 static struct target target = {
 	.uid	= UINT_MAX,
 };

@@ -1234,7 +1237,7 @@ static void printout(int id, int nr, struct perf_evsel *counter, double uval,
 
 	perf_stat__print_shadow_stats(counter, uval,
 				first_shadow_cpu(counter, id),
-				&out);
+				&out, &metric_events);
 	if (!csv_output && !metric_only) {
 		print_noise(counter, noise);
 		print_running(run, ena);

@@ -1565,7 +1568,8 @@ static void print_metric_headers(const char *prefix, bool no_indent)
 		os.evsel = counter;
 		perf_stat__print_shadow_stats(counter, 0,
 					      0,
-					      &out);
+					      &out,
+					      &metric_events);
 	}
 	fputc('\n', stat_config.output);
 }

@@ -1789,6 +1793,13 @@ static int enable_metric_only(const struct option *opt __maybe_unused,
 	return 0;
 }
 
+static int parse_metric_groups(const struct option *opt,
+			       const char *str,
+			       int unset __maybe_unused)
+{
+	return metricgroup__parse_groups(opt, str, &metric_events);
+}
+
 static const struct option stat_options[] = {
 	OPT_BOOLEAN('T', "transaction", &transaction_run,
 		    "hardware transaction statistics"),

@@ -1854,6 +1865,9 @@ static const struct option stat_options[] = {
 			"measure topdown level 1 statistics"),
 	OPT_BOOLEAN(0, "smi-cost", &smi_cost,
 			"measure SMI cost"),
+	OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
+		     "monitor specified metrics or metric groups (separated by ,)",
+		     parse_metric_groups),
 	OPT_END()
 };
 

index 94518c1bf8b6..71ab8466714d 100644
@@ -34,6 +34,7 @@ libperf-y += dso.o
 libperf-y += symbol.o
 libperf-y += symbol_fprintf.o
 libperf-y += color.o
+libperf-y += metricgroup.o
 libperf-y += header.o
 libperf-y += callchain.o
 libperf-y += values.o

new file mode 100644
index 000000000000..06854e125ee7
@@ -0,0 +1,31 @@
+#ifndef METRICGROUP_H
+#define METRICGROUP_H 1
+
+#include "linux/list.h"
+#include "rblist.h"
+#include <subcmd/parse-options.h>
+#include "evlist.h"
+#include "strbuf.h"
+
+struct metric_event {
+	struct rb_node nd;
+	struct perf_evsel *evsel;
+	struct list_head head; /* list of metric_expr */
+};
+
+struct metric_expr {
+	struct list_head nd;
+	const char *metric_expr;
+	const char *metric_name;
+	struct perf_evsel **metric_events;
+};
+
+struct metric_event *metricgroup__lookup(struct rblist *metric_events,
+					 struct perf_evsel *evsel,
+					 bool create);
+int metricgroup__parse_groups(const struct option *opt,
+			const char *str,
+			struct rblist *metric_events);
+
+void metricgroup__print(bool metrics, bool groups, char *filter, bool raw);
+#endif

index ed25d7f88731..7070638ab600 100644
@@ -580,8 +580,11 @@ static void pmu_add_cpu_aliases(struct list_head *head, const char *name)
 		const char *pname;
 
 		pe = &map->table[i++];
-		if (!pe->name)
+		if (!pe->name) {
+			if (pe->metric_group || pe->metric_name)
+				continue;
 			break;
+		}
 
 		pname = pe->pmu ? pe->pmu : "cpu";
 		if (strncmp(pname, name, strlen(pname)))

index 8c7ab29169b9..42e6c17be7ff 100644
@@ -6,6 +6,7 @@
 #include "rblist.h"
 #include "evlist.h"
 #include "expr.h"
+#include "metricgroup.h"
 
 enum {
 	CTX_BIT_USER	= 1 << 0,

@@ -671,13 +672,16 @@ static void generic_metric(const char *metric_expr,
 
 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 				   double avg, int cpu,
-				   struct perf_stat_output_ctx *out)
+				   struct perf_stat_output_ctx *out,
+				   struct rblist *metric_events)
 {
 	void *ctxp = out->ctx;
 	print_metric_t print_metric = out->print_metric;
 	double total, ratio = 0.0, total2;
 	const char *color = NULL;
 	int ctx = evsel_context(evsel);
+	struct metric_event *me;
+	int num = 1;
 
 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);

@@ -880,6 +884,20 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
 		print_smi_cost(cpu, evsel, out);
 	} else {
-		print_metric(ctxp, NULL, NULL, NULL, 0);
+		num = 0;
 	}
+
+	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
+		struct metric_expr *mexp;
+
+		list_for_each_entry (mexp, &me->head, nd) {
+			if (num++ > 0)
+				out->new_line(ctxp);
+			generic_metric(mexp->metric_expr, mexp->metric_events,
+					evsel->name, mexp->metric_name,
+					avg, cpu, ctx, out);
+		}
+	}
+	if (num == 0)
+		print_metric(ctxp, NULL, NULL, NULL, 0);
 }

index eacaf958e19d..47915df346fb 100644
@@ -91,9 +91,11 @@ struct perf_stat_output_ctx {
 	bool force_header;
 };
 
+struct rblist;
 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 				   double avg, int cpu,
-				   struct perf_stat_output_ctx *out);
+				   struct perf_stat_output_ctx *out,
+				   struct rblist *metric_events);
 void perf_stat__collect_metric_expr(struct perf_evlist *);
 
 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw);

2.9.5

Em Thu, Aug 31, 2017 at 12:40:31PM -0700, Andi Kleen escreveu:

Yeah, I noticed when trying to test it :-\

Was this pull req submitted?

- Arnaldo

[root@jouet ~]# perf stat -M Summary --metric-only -a sleep 1 Cannot find metric or group `Summary'

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@jouet ~]# perf stat -M GFLOPs flops
Cannot find metric or group `GFLOPs'

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@jouet ~]#

Not yet. I was waiting to finish review.

It makes perf crash until you apply the metricgroup patchkit first.

But you can get it here

git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-misc.git perf/intel-json-metrics-2

-Andi

Em Mon, Sep 04, 2017 at 10:37:25AM -0700, Andi Kleen escreveu:

Ok, so I tried installing the Broadwell ones (one of my test machines) and got into:

[root@jouet ~]# perf stat -M Pipeline flops
bpf: builtin compilation failed: -95, try external compiler
ERROR: problems with path cpu/uops_executed.c: No such file or directory
Cannot set up events {uops_retired.retire_slots,inst_retired.any}:W,{inst_retired.any,cycles}:W,{uops_executed.thread,cpu/uops_executed.core,cmask=1/,uops_executed.cycles_ge_1_uop_exec}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@jouet ~]# perf stat -M Pipeline flops

Wasn't this already fixed?

This is with these patches applied:

[acme@jouet linux]$ git log --oneline -5
5e6bf24a1929 (HEAD) perf stat: Support JSON metrics in perf stat
b1e1f50913c8 perf pmu: Extract function to get JSON alias map
af8aca46b528 perf stat: Print generic metric header even for failed expressions
15669491ee54 perf stat: Factor out generic metric printing
7d8900f2984c perf vendor events: Support metric_group and no event name in JSON parser
[acme@jouet linux]$ git log --oneline -10
5e6bf24a1929 (HEAD) perf stat: Support JSON metrics in perf stat
b1e1f50913c8 perf pmu: Extract function to get JSON alias map
af8aca46b528 perf stat: Print generic metric header even for failed expressions
15669491ee54 perf stat: Factor out generic metric printing
7d8900f2984c perf vendor events: Support metric_group and no event name in JSON parser
8f6ea3aac2e2 perf tools: Support weak groups in 'perf stat'
b5ef5ea5e8ed perf sched timehist: Add pid and tid options
eba9fac01761 (tag: perf-core-for-mingo-4.14-20170901, acme/perf/core) perf annotate browser: Help for cycling thru hottest instructions with TAB/shift+TAB 63ce8449bc10 perf stat: Only auto-merge events that are PMU aliases
fc33dccba395 perf test: Add test case for PERF_SAMPLE_PHYS_ADDR
[acme@jouet linux]$

Will try with all of them...

- Arnaldo

Em Tue, Sep 05, 2017 at 03:09:19PM -0300, Arnaldo Carvalho de Melo escreveu:

So there is no 'flops' binary here, unsure if the problem is related to that, if so, we need a fix.

But that doesn't seem to be the case, as with a metric that works I get a sensible error message:

[root@jouet ~]# perf stat -M Memory_BW --metric-only -a flops Workload failed: No such file or directory [root@jouet ~]#

Anyway, testing with some other metric I got some results:

[root@jouet ~]# perf stat -M FLOPS --metric-only -a sleep 1
Cannot set up events {fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@jouet ~]# perf stat -M Memory_BW --metric-only -a sleep 1

Performance counter stats for 'system wide':

MLP

     2.9            

       1.001544615 seconds time elapsed

[root@jouet ~]#

Em Tue, Sep 05, 2017 at 03:16:12PM -0300, Arnaldo Carvalho de Melo escreveu:

further info:

[acme@jouet linux]$ perf stat -v -M GFLOPs sleep 1
Using CPUID GenuineIntel-6-3D
metric expr ( 1*( fp_arith_inst_retired.scalar_single + fp_arith_inst_retired.scalar_double ) + 2* fp_arith_inst_retired.128b_packed_double + 4*( fp_arith_inst_retired.128b_packed_single + fp_arith_inst_retired.256b_packed_double ) + 8* fp_arith_inst_retired.256b_packed_single ) / 1000000000 / duration_time for GFLOPs found event fp_arith_inst_retired.scalar_single
found event fp_arith_inst_retired.scalar_double
found event fp_arith_inst_retired.128b_packed_double
found event fp_arith_inst_retired.128b_packed_single
found event fp_arith_inst_retired.256b_packed_double
found event fp_arith_inst_retired.256b_packed_single
found event duration_time
adding {fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W fp_arith_inst_retired.scalar_single -> cpu/umask=0x2,period=2000003,event=0xc7/
fp_arith_inst_retired.scalar_double -> cpu/umask=0x1,period=2000003,event=0xc7/
fp_arith_inst_retired.128b_packed_double -> cpu/umask=0x4,period=2000003,event=0xc7/
fp_arith_inst_retired.128b_packed_single -> cpu/umask=0x8,period=2000003,event=0xc7/
fp_arith_inst_retired.256b_packed_double -> cpu/umask=0x10,period=2000003,event=0xc7/
fp_arith_inst_retired.256b_packed_single -> cpu/umask=0x20,period=2000003,event=0xc7/
Cannot set up events {fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[acme@jouet linux]$

Humm, so it uses that duration_time and that gets added only later?

/me tries with all the patches:

Yeah:

[acme@jouet linux]$ perf stat -M GFLOPs sleep 1

Performance counter stats for 'sleep 1':

                 0      fp_arith_inst_retired.scalar_single:u                                   
                 3      fp_arith_inst_retired.scalar_double:u                                   
                 0      fp_arith_inst_retired.128b_packed_double:u                                   
                 0      fp_arith_inst_retired.128b_packed_single:u                                   
                 0      fp_arith_inst_retired.256b_packed_double:u                                     (40.86%)
     <not counted>      fp_arith_inst_retired.256b_packed_single:u                                     (0.00%)
                 0      duration_time:u                                             

       1.003398909 seconds time elapsed

[acme@jouet linux]$

With --metric-only I get nothing, what am I doing wrong?

[acme@jouet linux]$ perf stat -M GFLOPs --metric-only sleep 1

Performance counter stats for 'sleep 1':

GFLOPs

       1.003217768 seconds time elapsed

[acme@jouet linux]$

I just rebuilt the tools completely from scratch, deleting the build dir, to see if this was some build artifact, didn't help, I still get just the time elapsed.

Now with -v:

[acme@jouet linux]$ perf stat -v -M GFLOPs --metric-only sleep 1
Using CPUID GenuineIntel-6-3D
metric expr ( 1*( fp_arith_inst_retired.scalar_single + fp_arith_inst_retired.scalar_double ) + 2* fp_arith_inst_retired.128b_packed_double + 4*( fp_arith_inst_retired.128b_packed_single + fp_arith_inst_retired.256b_packed_double ) + 8* fp_arith_inst_retired.256b_packed_single ) / 1000000000 / duration_time for GFLOPs found event fp_arith_inst_retired.scalar_single
found event fp_arith_inst_retired.scalar_double
found event fp_arith_inst_retired.128b_packed_double
found event fp_arith_inst_retired.128b_packed_single
found event fp_arith_inst_retired.256b_packed_double
found event fp_arith_inst_retired.256b_packed_single
found event duration_time
adding {fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W fp_arith_inst_retired.scalar_single -> cpu/umask=0x2,period=2000003,event=0xc7/
fp_arith_inst_retired.scalar_double -> cpu/umask=0x1,period=2000003,event=0xc7/
fp_arith_inst_retired.128b_packed_double -> cpu/umask=0x4,period=2000003,event=0xc7/
fp_arith_inst_retired.128b_packed_single -> cpu/umask=0x8,period=2000003,event=0xc7/
fp_arith_inst_retired.256b_packed_double -> cpu/umask=0x10,period=2000003,event=0xc7/
fp_arith_inst_retired.256b_packed_single -> cpu/umask=0x20,period=2000003,event=0xc7/
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Weak group for fp_arith_inst_retired.scalar_single:u/7 failed
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
Warning:
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel samples
fp_arith_inst_retired.scalar_single:u: 0 2137416 524337
fp_arith_inst_retired.scalar_double:u: 0 1525387 1525387
fp_arith_inst_retired.128b_packed_double:u: 0 2429065 2429065
fp_arith_inst_retired.128b_packed_single:u: 0 2429065 2429065
fp_arith_inst_retired.256b_packed_double:u: 0 2429065 1904728
fp_arith_inst_retired.256b_packed_single:u: 0 2429065 903678
duration_time:u: 0 2394375 2394375

Performance counter stats for 'sleep 1':

fp_arith_inst_retired.scalar_single not found
GFLOPs
fp_arith_inst_retired.scalar_single not found

       1.005367355 seconds time elapsed

[acme@jouet linux]$

We need to improve these messages...

As root for that Summary more:

[acme@jouet linux]$ sudo ~acme/bin/perf stat -M Summary --metric-only -a sleep 1

Performance counter stats for 'system wide':

Instructions CPI CLKS CPU_Utilization GFLOPs SMT_2T_Utilization Kernel_Utilization
134707420.0 0.0 698757419.0 0.1 0.0 0.3 0.4

       1.003365814 seconds time elapsed

[acme@jouet linux]$

- Arnaldo

I would suggest to apply all the patches from the patchkit before you test. Some of the later ones add/fix stuff needed by some of the metrics.

Yes that was supposed to be fixed with

commit 77d0871c76bad1093a3d86870fe76dd1ad0ca397

Author: Andi Kleen <[email protected]>
Date:   Fri Aug 11 16:26:19 2017 -0700

perf bpf: Tighten detection of BPF events

perf stat -e cpu/uops_executed.core,cmask=1/

I'll check.

-Andi

Em Tue, Sep 05, 2017 at 11:19:52AM -0700, Andi Kleen escreveu:

Ok, I couldn't reproduce it anymore, after you check this, please let me know if I should pull that perf/intel-json-metrics-2 branch so that someone wanting to test this can have it all in one place, ok?

- Arnaldo

Sure please pull.

The only missing thing are metrics for Skylake Server, but I can submit those later.

-Andi

Em Tue, Sep 05, 2017 at 12:52:35PM -0700, Andi Kleen escreveu:

Ok, so I looked at the commit logs and found them rather dull, how is that option to fake a CPU so that the tool thinks it is running on some specific machine (broadwell, skylake, etc) so that I can augment those with the output of 'perf list metricgroup' for each of them?

- Arnaldo

PERF_CPUID=GenuineIntel-...

See the mapfile.csv for valid codes

But it's quite a few.

-Andi

Em Tue, Sep 05, 2017 at 01:37:08PM -0700, Andi Kleen escreveu:

yeah, I'm testing on the ones I have access and on a Skylake machine I'm having trouble, see below.

That error message is not that helpful, what could be at play here? The kernel running there is a recent RT variant for the RHEL kernel, with lots of backports in the perf codebase, so I wasn't expecting that to be an issue, please take a look.

I'll try later with a more recent kernel, maybe there are kernel patches supporting Skylake that aren't in this kernel :-\

[root@seventh ~]# uname -a
Linux seventh 3.10.0-703.rt56.630.el7.x86_64 #1 SMP PREEMPT RT Sat Aug 19 22:32:17 EDT 2017 x86_64 x86_64 x86_64 GNU/Linux [root@seventh ~]# grep "model name" /proc/cpuinfo | head -1
model name : Intel(R) Core(TM) i5-7500 CPU @ 3.40GHz
[root@seventh ~]# perf list metricgroup

List of pre-defined events (to be used in -e):

Metric Groups:

DSB
FLOPS
Frontend
Frontend_Bandwidth
Memory_BW
Memory_Bound
Memory_Lat
Pipeline
Ports_Utilization
Power
SMT
Summary
TLB
TopDownL1
Unknown_Branches
[root@seventh ~]# perf stat --metric-only -M Summary -a sleep 1
Cannot set up events {inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W,{cpu_clk_thread_unhalted.one_thread_active,cpu_clk_thread_unhalted.ref_xclk_any}:W,{cpu_clk_unhalted.ref_tsc:u,cpu_clk_unhalted.ref_tsc}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@seventh ~]#

[root@seventh ~]# perf stat -vv --metric-only -M Summary -a sleep 1
Using CPUID GenuineIntel-6-9E
metric expr 1 / inst_retired.any / cycles for CPI
found event inst_retired.any
found event cycles
metric expr cpu_clk_unhalted.thread for CLKS
found event cpu_clk_unhalted.thread
metric expr inst_retired.any for Instructions
found event inst_retired.any
metric expr cpu_clk_unhalted.ref_tsc / msr@tsc@ for CPU_Utilization
found event cpu_clk_unhalted.ref_tsc
found event msr/tsc/
metric expr ( 1*( fp_arith_inst_retired.scalar_single + fp_arith_inst_retired.scalar_double ) + 2* fp_arith_inst_retired.128b_packed_double + 4*( fp_arith_inst_retired.128b_packed_single + fp_arith_inst_retired.256b_packed_double ) + 8* fp_arith_inst_retired.256b_packed_single ) / 1000000000 / duration_time for GFLOPs
found event fp_arith_inst_retired.scalar_single
found event fp_arith_inst_retired.scalar_double
found event fp_arith_inst_retired.128b_packed_double
found event fp_arith_inst_retired.128b_packed_single
found event fp_arith_inst_retired.256b_packed_double
found event fp_arith_inst_retired.256b_packed_single
found event duration_time
metric expr 1 - cpu_clk_thread_unhalted.one_thread_active / ( cpu_clk_thread_unhalted.ref_xclk_any / 2 ) if #smt_on else 0 for SMT_2T_Utilization
found event cpu_clk_thread_unhalted.one_thread_active
found event cpu_clk_thread_unhalted.ref_xclk_any
metric expr cpu_clk_unhalted.ref_tsc:u / cpu_clk_unhalted.ref_tsc for Kernel_Utilization
found event cpu_clk_unhalted.ref_tsc:u
found event cpu_clk_unhalted.ref_tsc
adding {inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W,{cpu_clk_thread_unhalted.one_thread_active,cpu_clk_thread_unhalted.ref_xclk_any}:W,{cpu_clk_unhalted.ref_tsc:u,cpu_clk_unhalted.ref_tsc}:W intel_pt default config: tsc,mtc,mtc_period=3,psb_period=3
inst_retired.any -> cpu/event=0xc0/
cpu_clk_unhalted.thread -> cpu/event=0x3c/
inst_retired.any -> cpu/event=0xc0/
cpu_clk_unhalted.ref_tsc -> cpu/umask=0x3,period=2000003,event=0/
Cannot set up events {inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W,{cpu_clk_thread_unhalted.one_thread_active,cpu_clk_thread_unhalted.ref_xclk_any}:W,{cpu_clk_unhalted.ref_tsc:u,cpu_clk_unhalted.ref_tsc}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@seventh ~]#

Humm, it doesn't even get to try sys_perf_event_open() it seems to be bailing out while reading some sysfs event_source files, a better error output seems in demand:

[root@seventh ~]# perf trace -e perf_* perf stat --metric-only -M Summary -a sleep 1
Cannot set up events {inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W,{cpu_clk_thread_unhalted.one_thread_active,cpu_clk_thread_unhalted.ref_xclk_any}:W,{cpu_clk_unhalted.ref_tsc:u,cpu_clk_unhalted.ref_tsc}:W

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@seventh ~]#

[root@seventh linux]# strace perf stat --metric-only -M Summary -a sleep 1 |& tail -30
getdents(3, /* 0 entries */, 32768) = 0
close(3) = 0
openat(AT_FDCWD, "/sys/bus/event_source/devices/", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 3

getdents(3, /* 8 entries */, 32768)     = 232
getdents(3, /* 0 entries */, 32768)     = 0
close(3)                                = 0
stat("/sys/bus/event_source/devices/msr/format", 0x7ffd36f79480) = -1 ENOENT (No such file or directory)
stat("/sys/bus/event_source/devices/msr/type", 0x7ffd36f79470) = -1 ENOENT (No such file or directory)
openat(AT_FDCWD, "/sys/bus/event_source/devices/", O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC) = 3
getdents(3, /* 8 entries */, 32768)     = 232
getdents(3, /* 0 entries */, 32768)     = 0
close(3)                                = 0

write(2, "Cannot set up events {inst_retir"..., 529Cannot set up events {inst_retired.any,cycles}:W,{cpu_clk_unhalted.thread}:W,{inst_retired.any}:W,{cpu_clk_unhalted.ref_tsc,msr/tsc/}:W,{fp_arith_inst_retired.scalar_single,fp_arith_inst_retired.scalar_double,fp_arith_inst_retired.128b_packed_double,fp_arith_inst_retired.128b_packed_single,fp_arith_inst_retired.256b_packed_double,fp_arith_inst_retired.256b_packed_single,duration_time}:W,{cpu_clk_thread_unhalted.one_thread_active,cpu_clk_thread_unhalted.ref_xclk_any}:W,{cpu_clk_unhalted.ref_tsc:u,cpu_clk_unhalted.ref_tsc}:W ) = 529
write(2, "\n Usage: perf stat [<options>] ["..., 43
Usage: perf stat [<options>] [<command>]
) = 43
write(2, "\n", 1
) = 1
write(2, " ", 4 ) = 4
write(2, "-M", 2-M) = 2
write(2, ", ", 2, ) = 2
write(2, "--metrics", 9--metrics) = 9
write(2, " <metric/metric group list>", 27 <metric/metric group list>) = 27
write(2, "\n", 1
) = 1
write(2, " monito"..., 86 monitor specified metrics or metric groups (separated by ,)
) = 86
exit_group(129) = ?
+++ exited with 129 +++
[root@seventh linux]#

The problem is that your kernel is missing the msr// PMU.

It should work with a newer kernel.

I can look at better error messages.

-Andi

Em Fri, Sep 08, 2017 at 12:08:48PM -0700, Andi Kleen escreveu:

Yes:

[root@seventh ~]# uname -r
4.12.0-rc6+
[root@seventh ~]# perf stat --metric-only -M Summary -a sleep 1

Performance counter stats for 'system wide':

Instructions CPI CLKS CPU_Utilization GFLOPs SMT_2T_Utilization Kernel_Utilization
34021097.0 0.0 119424171.0 0.0 0.0 0.0 0.0

       1.001001793 seconds time elapsed

[root@seventh ~]#

Please do,

Thanks!

- Arnaldo

From: Andi Kleen <[email protected]>

Add code to perf list to print metric groups, and metrics that don't have an event name. The metricgroup code collects the eventgroups and events into a rblist, and then prints them according to the configured filters.

The metricgroups are printed by default, but can be limited by perf list metric or perf list metricgroup

% perf list metricgroup
..
Metric Groups:

DSB:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
FLOPS:
GFLOPs
[Giga Floating Point Operations Per Second]
Frontend:
IFetch_Line_Utilization
[Rough Estimation of fraction of fetched lines bytes that were likely consumed by program instructions]
Frontend_Bandwidth:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
Memory_BW:
MLP
[Memory-Level-Parallelism (average number of L1 miss demand load when there is at least 1 such miss)]

v2: Check return value of asprintf to fix warning on FC26 Fix key in lookup/addition for the groups list

---
 tools/perf/Documentation/perf-list.txt |   7 +-
 tools/perf/builtin-list.c              |   7 ++
 tools/perf/util/metricgroup.c          | 176 +++++++++++++++++++++++++++++++++
 tools/perf/util/parse-events.c         |   3 +
 4 files changed, 192 insertions(+), 1 deletion(-)

index d432965d728d..ca6369fd83c1 100644
@@ -8,7 +8,8 @@ perf-list - List all symbolic event types
 SYNOPSIS
 --------
 [verse]
-'perf list' [--no-desc] [--long-desc] [hw|sw|cache|tracepoint|pmu|sdt|event_glob]
+'perf list' [--no-desc] [--long-desc]
+            [hw|sw|cache|tracepoint|pmu|sdt|metric|metricgroup|event_glob]
 
 DESCRIPTION
 -----------

@@ -247,6 +248,10 @@ To limit the list use:
 
 . 'sdt' to list all Statically Defined Tracepoint events.
 
+. 'metric' to list metrics
+
+. 'metricgroup' to list metricgroups with metrics.
+
 . If none of the above is matched, it will apply the supplied glob to all
   events, printing the ones that match.
 

index 4bf2cb4d25aa..b2d2ad3dd478 100644
@@ -15,6 +15,7 @@
 #include "util/cache.h"
 #include "util/pmu.h"
 #include "util/debug.h"
+#include "util/metricgroup.h"
 #include <subcmd/parse-options.h>
 
 static bool desc_flag = true;

@@ -79,6 +80,10 @@ int cmd_list(int argc, const char **argv)
 						long_desc_flag, details_flag);
 		else if (strcmp(argv[i], "sdt") == 0)
 			print_sdt_events(NULL, NULL, raw_dump);
+		else if (strcmp(argv[i], "metric") == 0)
+			metricgroup__print(true, false, NULL, raw_dump);
+		else if (strcmp(argv[i], "metricgroup") == 0)
+			metricgroup__print(false, true, NULL, raw_dump);
 		else if ((sep = strchr(argv[i], ':')) != NULL) {
 			int sep_idx;
 

@@ -96,6 +101,7 @@ int cmd_list(int argc, const char **argv)
 			s[sep_idx] = '\0';
 			print_tracepoint_events(s, s + sep_idx + 1, raw_dump);
 			print_sdt_events(s, s + sep_idx + 1, raw_dump);
+			metricgroup__print(true, true, s, raw_dump);
 			free(s);
 		} else {
 			if (asprintf(&s, "*%s*", argv[i]) < 0) {

@@ -112,6 +118,7 @@ int cmd_list(int argc, const char **argv)
 						details_flag);
 			print_tracepoint_events(NULL, s, raw_dump);
 			print_sdt_events(NULL, s, raw_dump);
+			metricgroup__print(true, true, NULL, raw_dump);
 			free(s);
 		}
 	}

index cd80c4eac569..97507d5c37dc 100644
@@ -28,6 +28,7 @@
 #include "probe-file.h"
 #include "asm/bug.h"
 #include "util/parse-branch-options.h"
+#include "metricgroup.h"
 
 #define MAX_NAME_LEN 100
 

@@ -2372,6 +2373,8 @@ void print_events(const char *event_glob, bool name_only, bool quiet_flag,
 	print_tracepoint_events(NULL, NULL, name_only);
 
 	print_sdt_events(NULL, NULL, name_only);
+
+	metricgroup__print(true, true, NULL, name_only);
 }
 
 int parse_events__is_hardcoded_term(struct parse_events_term *term)

2.9.5

Em Thu, Aug 31, 2017 at 12:40:32PM -0700, Andi Kleen escreveu:

Andi, I just noticed that metric groups appear when we pass some substring to search events that looks unrelated, can you please take a look at this?

Thanks,

- Arnaldo

[root@jouet ~]# perf list energy-cores
oList of pre-defined events (to be used in -e):

power/energy-cores/ [Kernel PMU event]

Metric Groups:

DSB:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
FLOPS:
GFLOPs
[Giga Floating Point Operations Per Second]
Frontend:
IFetch_Line_Utilization
[Rough Estimation of fraction of fetched lines bytes that were likely consumed by program instructions]
Frontend_Bandwidth:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
Memory_BW:
MLP
[Memory-Level-Parallelism (average number of L1 miss demand load when there is at least 1 such miss)]
Memory_Bound:
Load_Miss_Real_Latency
[Actual Average Latency for L1 data-cache miss demand loads]
MLP
[Memory-Level-Parallelism (average number of L1 miss demand load when there is at least 1 such miss)]
Memory_Lat:
Load_Miss_Real_Latency
[Actual Average Latency for L1 data-cache miss demand loads]
Pipeline:
CPI

From: Andi Kleen <[email protected]>

We don't need to use ctx to look up events for saved values. The context is already part of the evsel pointer, which is the primary key.

---
 tools/perf/util/stat-shadow.c | 16 +++++-----------
 1 file changed, 5 insertions(+), 11 deletions(-)

index 42e6c17be7ff..664f49a9b012 100644
@@ -56,7 +56,6 @@ struct saved_value {
 	struct rb_node rb_node;
 	struct perf_evsel *evsel;
 	int cpu;
-	int ctx;
 	struct stats stats;
 };
 

@@ -67,8 +66,6 @@ static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
 					     rb_node);
 	const struct saved_value *b = entry;
 
-	if (a->ctx != b->ctx)
-		return a->ctx - b->ctx;
 	if (a->cpu != b->cpu)
 		return a->cpu - b->cpu;
 	if (a->evsel == b->evsel)

@@ -90,13 +87,12 @@ static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
 }
 
 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
-					      int cpu, int ctx,
+					      int cpu,
 					      bool create)
 {
 	struct rb_node *nd;
 	struct saved_value dm = {
 		.cpu = cpu,
-		.ctx = ctx,
 		.evsel = evsel,
 	};
 	nd = rblist__find(&runtime_saved_values, &dm);

@@ -232,8 +228,7 @@ void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
 		update_stats(&runtime_aperf_stats[ctx][cpu], count[0]);
 
 	if (counter->collect_stat) {
-		struct saved_value *v = saved_value_lookup(counter, cpu, ctx,
-							   true);
+		struct saved_value *v = saved_value_lookup(counter, cpu, true);
 		update_stats(&v->stats, count[0]);
 	}
 }

@@ -634,7 +629,6 @@ static void generic_metric(const char *metric_expr,
 			   const char *metric_name,
 			   double avg,
 			   int cpu,
-			   int ctx,
 			   struct perf_stat_output_ctx *out)
 {
 	print_metric_t print_metric = out->print_metric;

@@ -648,7 +642,7 @@ static void generic_metric(const char *metric_expr,
 	for (i = 0; metric_events[i]; i++) {
 		struct saved_value *v;
 
-		v = saved_value_lookup(metric_events[i], cpu, ctx, false);
+		v = saved_value_lookup(metric_events[i], cpu, false);
 		if (!v)
 			break;
 		expr__add_id(&pctx, metric_events[i]->name, avg_stats(&v->stats));

@@ -866,7 +860,7 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 			print_metric(ctxp, NULL, NULL, name, 0);
 	} else if (evsel->metric_expr) {
 		generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
-				evsel->metric_name, avg, cpu, ctx, out);
+				evsel->metric_name, avg, cpu, out);
 	} else if (runtime_nsecs_stats[cpu].n != 0) {
 		char unit = 'M';
 		char unit_buf[10];

@@ -895,7 +889,7 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
 				out->new_line(ctxp);
 			generic_metric(mexp->metric_expr, mexp->metric_events,
 					evsel->name, mexp->metric_name,
-					avg, cpu, ctx, out);
+					avg, cpu, out);
 		}
 	}
 	if (num == 0)

2.9.5

From: Andi Kleen <[email protected]>

Some of the metrics formulas (like GFLOPs) need to know how long the measurement period is. Support an internal event called duration_time, which reports time in second. It maps to the dummy event, but is
special cased for statistics to report the walltime duration.

So far it is not printed, but only used internally for metrics.

---
 tools/perf/util/parse-events.l |  1 +
 tools/perf/util/stat-shadow.c  | 17 +++++++++++++----
 2 files changed, 14 insertions(+), 4 deletions(-)

index fdb5bb52f01f..ea2426daf7e8 100644
@@ -288,6 +288,7 @@ cpu-migrations|migrations			{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COU
 alignment-faults				{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_ALIGNMENT_FAULTS); }
 emulation-faults				{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_EMULATION_FAULTS); }
 dummy						{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_DUMMY); }
+duration_time					{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_DUMMY); }
 bpf-output					{ return sym(yyscanner, PERF_TYPE_SOFTWARE, PERF_COUNT_SW_BPF_OUTPUT); }
 
 	/*

index 664f49a9b012..a2c12d1ef32a 100644
@@ -641,11 +641,20 @@ static void generic_metric(const char *metric_expr,
 	expr__add_id(&pctx, name, avg);
 	for (i = 0; metric_events[i]; i++) {
 		struct saved_value *v;
+		struct stats *stats;
+		double scale;
 
-		v = saved_value_lookup(metric_events[i], cpu, false);
-		if (!v)
-			break;
-		expr__add_id(&pctx, metric_events[i]->name, avg_stats(&v->stats));
+		if (!strcmp(metric_events[i]->name, "duration_time")) {
+			stats = &walltime_nsecs_stats;
+			scale = 1e-9;
+		} else {
+			v = saved_value_lookup(metric_events[i], cpu, false);
+			if (!v)
+				break;
+			stats = &v->stats;
+			scale = 1.0;
+		}
+		expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
 	}
 	if (!metric_events[i]) {
 		const char *p = metric_expr;

2.9.5

From: Andi Kleen <[email protected]>

Some perf stat metrics use an internal "duration_time" metric. It is not correctly printed however. So hide it during output to avoid confusing users with 0 counts.

---
 tools/perf/builtin-stat.c | 27 ++++++++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)

index 045efd7d7785..bf6ae0144ecf 100644
@@ -195,6 +195,11 @@ static struct perf_stat_config stat_config = {
 	.scale		= true,
 };
 
+static bool is_duration_time(struct perf_evsel *evsel)
+{
+	return !strcmp(evsel->name, "duration_time");
+}
+
 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 				 struct timespec *b)
 {

@@ -1363,6 +1368,9 @@ static void print_aggr(char *prefix)
 		ad.id = id = aggr_map->map[s];
 		first = true;
 		evlist__for_each_entry(evsel_list, counter) {
+			if (is_duration_time(counter))
+				continue;
+
 			ad.val = ad.ena = ad.run = 0;
 			ad.nr = 0;
 			if (!collect_data(counter, aggr_cb, &ad))

@@ -1506,6 +1514,8 @@ static void print_no_aggr_metric(char *prefix)
 		if (prefix)
 			fputs(prefix, stat_config.output);
 		evlist__for_each_entry(evsel_list, counter) {
+			if (is_duration_time(counter))
+				continue;
 			if (first) {
 				aggr_printout(counter, cpu, 0);
 				first = false;

@@ -1560,6 +1570,8 @@ static void print_metric_headers(const char *prefix, bool no_indent)
 
 	/* Print metrics headers only */
 	evlist__for_each_entry(evsel_list, counter) {
+		if (is_duration_time(counter))
+			continue;
 		os.evsel = counter;
 		out.ctx = &os;
 		out.print_metric = print_metric_header;

@@ -1707,12 +1719,18 @@ static void print_counters(struct timespec *ts, int argc, const char **argv)
 		print_aggr(prefix);
 		break;
 	case AGGR_THREAD:
-		evlist__for_each_entry(evsel_list, counter)
+		evlist__for_each_entry(evsel_list, counter) {
+			if (is_duration_time(counter))
+				continue;
 			print_aggr_thread(counter, prefix);
+		}
 		break;
 	case AGGR_GLOBAL:
-		evlist__for_each_entry(evsel_list, counter)
+		evlist__for_each_entry(evsel_list, counter) {
+			if (is_duration_time(counter))
+				continue;
 			print_counter_aggr(counter, prefix);
+		}
 		if (metric_only)
 			fputc('\n', stat_config.output);
 		break;

@@ -1720,8 +1738,11 @@ static void print_counters(struct timespec *ts, int argc, const char **argv)
 		if (metric_only)
 			print_no_aggr_metric(prefix);
 		else {
-			evlist__for_each_entry(evsel_list, counter)
+			evlist__for_each_entry(evsel_list, counter) {
+				if (is_duration_time(counter))
+					continue;
 				print_counter(counter, prefix);
+			}
 		}
 		break;
 	case AGGR_UNSET:

2.9.5

From: Andi Kleen <[email protected]>

Some metrics (like GFLOPs) need walltime_nsecs_stats for each interval. Compute it for each interval instead of only at the end.

Pointed out by Jiri.

---
 tools/perf/builtin-stat.c | 2 ++
 1 file changed, 2 insertions(+)

index bf6ae0144ecf..60206a1e03aa 100644
@@ -415,6 +415,8 @@ static void process_interval(void)
 			pr_err("failed to write stat round event\n");
 	}
 
+	init_stats(&walltime_nsecs_stats);
+	update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
 	print_counters(&rs, 0, NULL);
 }
 

2.9.5

On Thu, Aug 31, 2017 at 12:40:25PM -0700, Andi Kleen wrote:

SNIP

looks like some events are probably crossing some
output boundaries we have:

[jolsa@krava perf]$ sudo ./perf stat -M SMT -I 1000
# time counts unit events

     1.000565706        408,879,985      inst_retired.any          #      0.7 CoreIPC                  (66.68%)
     1.000565706      1,120,999,114      cpu_clk_unhalted.thread_any                                     (66.68%)
     1.000565706        701,285,312      cycles                                                        (66.68%)
     1.000565706      1,148,325,740      cpu_clk_unhalted.thread_any # 574162870.0 CORE_CLKS             (66.67%)
     1.000565706        711,565,247      cpu_clk_unhalted.thread                                       (66.66%)
     1.000565706         24,057,590      cpu_clk_thread_unhalted.one_thread_active #      0.3 SMT_2T_Utilization       (66.67%)
     1.000565706         65,753,475      cpu_clk_thread_unhalted.ref_xclk_any                                     (66.67%)
^C     1.349436822         21,198,385      inst_retired.any          #      0.1 CoreIPC                  (66.70%)
     1.349436822        112,740,282      cpu_clk_unhalted.thread_any                                     (66.70%)
     1.349436822         84,509,414      cycles                                                        (66.70%)
     1.349436822        108,181,315      cpu_clk_unhalted.thread_any # 54090657.5 CORE_CLKS              (66.62%)
     1.349436822         79,700,353      cpu_clk_unhalted.thread                                       (66.61%)
     1.349436822          3,911,698      cpu_clk_thread_unhalted.one_thread_active #      0.8 SMT_2T_Utilization       (66.69%)
     1.349436822         14,739,671      cpu_clk_thread_unhalted.ref_xclk_any                                     (66.69%)

could you please check on that and maybe shift the alignment for the longest name?

thanks,
jirka

other than this the rest of the patchset looks ok to me

thanks,
jirka

To make everything align would require shifting everything. So most usages which don't have that wide events wouldn't fit into 80 character columns anymore. That would be worse.

Or do a two pass output that computes lengths first. But that's fairly complicated and would prefer to not tackle this right now.

Again it's only a few cases where it is visible at all.

-Andi

that's what I had in mind.. but could be dealt with later

jirka