perf stat: Enable group read of counters

Adding perf_evsel__read_counter function to read single or group counter. After calling this function the counter's evsel::counts struct is filled with values for the counter and member of its group if there are any.

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/util/evsel.c | 100 ++++++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/evsel.h |   2 +
 tools/perf/util/stat.c  |   3 ++
 tools/perf/util/stat.h  |   5 ++-
 4 files changed, 108 insertions(+), 2 deletions(-)

tools/perf/util/evsel.c +100 -0

index 4dd0fcc06db9..89aecf3a35c7 100644
@@ -1302,6 +1302,106 @@ int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
 	return 0;
 }
 
+static int
+perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
+{
+	struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
+
+	return perf_evsel__read(evsel, cpu, thread, count);
+}
+
+static void
+perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
+		      u64 val, u64 ena, u64 run)
+{
+	struct perf_counts_values *count;
+
+	count = perf_counts(counter->counts, cpu, thread);
+
+	count->val    = val;
+	count->ena    = ena;
+	count->run    = run;
+}
+
+static int
+perf_evsel__process_group_data(struct perf_evsel *leader,
+			       int cpu, int thread, u64 *data)
+{
+	u64 read_format = leader->attr.read_format;
+	struct sample_read_value *v;
+	u64 nr, ena = 0, run = 0, i;
+
+	nr = *data++;
+
+	if (nr != (u64) leader->nr_members)
+		return -EINVAL;
+
+	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
+		ena = *data++;
+
+	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
+		run = *data++;
+
+	v = (struct sample_read_value *) data;
+
+	perf_evsel__set_count(leader, cpu, thread,
+			      v[0].value, ena, run);
+
+	for (i = 1; i < nr; i++) {
+		struct perf_evsel *counter;
+
+		counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
+		if (!counter)
+			return -EINVAL;
+
+		perf_evsel__set_count(counter, cpu, thread,
+				      v[i].value, ena, run);
+	}
+
+	return 0;
+}
+
+static int
+perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
+{
+	struct perf_stat_evsel *ps = leader->priv;
+	u64 read_format = leader->attr.read_format;
+	int size = perf_evsel__read_size(leader);
+	u64 *data = ps->group_data;
+
+	if (!(read_format & PERF_FORMAT_ID))
+		return -EINVAL;
+
+	if (!perf_evsel__is_group_leader(leader))
+		return -EINVAL;
+
+	if (!data) {
+		data = zalloc(size);
+		if (!data)
+			return -ENOMEM;
+
+		ps->group_data = data;
+	}
+
+	if (FD(leader, cpu, thread) < 0)
+		return -EINVAL;
+
+	if (readn(FD(leader, cpu, thread), data, size) <= 0)
+		return -errno;
+
+	return perf_evsel__process_group_data(leader, cpu, thread, data);
+}
+
+int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
+{
+	u64 read_format = evsel->attr.read_format;
+
+	if (read_format & PERF_FORMAT_GROUP)
+		return perf_evsel__read_group(evsel, cpu, thread);
+	else
+		return perf_evsel__read_one(evsel, cpu, thread);
+}
+
 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 			      int cpu, int thread, bool scale)
 {

tools/perf/util/evsel.h +2 -0

index fb40ca3c6519..de03c18daaf0 100644
@@ -299,6 +299,8 @@ static inline bool perf_evsel__match2(struct perf_evsel *e1,
 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
 		     struct perf_counts_values *count);
 
+int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread);
+
 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 			      int cpu, int thread, bool scale);
 

tools/perf/util/stat.c +0 -0

index 53b9a994a3dc..2c258554f94d 100644

tools/perf/util/stat.c +0 -0

 	zfree(&evsel->priv);
 }
 

tools/perf/util/stat.h +3 -2

index 7522bf10b03e..eacaf958e19d 100644
@@ -28,8 +28,9 @@ enum perf_stat_evsel_id {
 };
 
 struct perf_stat_evsel {
-	struct stats		res_stats[3];
-	enum perf_stat_evsel_id	id;
+	struct stats		 res_stats[3];
+	enum perf_stat_evsel_id	 id;
+	u64			*group_data;
 };
 
 enum aggr_mode {

Signature · 1 line

2.9.4

Em Fri, Jul 21, 2017 at 02:12:11PM +0200, Jiri Olsa escreveu:

tools/perf/util/stat.c static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) · line 133

Quoted 8 lines

+++ b/tools/perf/util/stat.c

tools/perf/util/stat.c +3 -0

@@ -128,6 +128,9 @@ static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 
 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 {
+	struct perf_stat_evsel *ps = evsel->priv;
+
+	free(ps->group_data);

Humm, are you sure you can always do that, i.e. that evsel->priv is not NULL?

Program received signal SIGSEGV, Segmentation fault.
0x000000000054ca04 in perf_evsel__free_stat_priv (evsel=0x2454be0) at util/stat.c:133
133 free(ps->group_data);
Missing separate debuginfos, use: dnf debuginfo-install audit-libs-2.7.7-1.fc25.x86_64 elfutils-libelf-0.169-1.fc25.x86_64 elfutils-libs-0.169-1.fc25.x86_64 libunwind-1.2-1.fc25.x86_64 perl-libs-5.24.1-386.fc25.x86_64 python-libs-2.7.13-2.fc25.x86_64 slang-2.3.0-7.fc25.x86_64 (gdb) bt

#0  0x000000000054ca04 in perf_evsel__free_stat_priv (evsel=0x2454be0) at util/stat.c:133
#1  0x000000000054cc93 in perf_evlist__free_stats (evlist=0x24541a0) at util/stat.c:189
#2  0x0000000000460311 in cmd_script (argc=0, argv=0x7fffffffe170) at builtin-script.c:3075
#3  0x00000000004be56c in run_builtin (p=0xa359f8 <commands+408>, argc=1, argv=0x7fffffffe170) at perf.c:296
#4  0x00000000004be7d9 in handle_internal_command (argc=1, argv=0x7fffffffe170) at perf.c:348
#5  0x00000000004be92b in run_argv (argcp=0x7fffffffdfcc, argv=0x7fffffffdfc0) at perf.c:392
#6  0x00000000004bed05 in main (argc=1, argv=0x7fffffffe170) at perf.c:530

(gdb)

I'm dropping the series, please retest with 'perf script'.

- Arnaldo

tools/perf/util/stat.c static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) · line 133

Quoted 11 lines

tools/perf/util/stat.c +3 -0

index 53b9a994a3dc..2c258554f94d 100644
@@ -128,6 +128,9 @@ static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 
 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 {
+	struct perf_stat_evsel *ps = evsel->priv;
+
+	free(ps->group_data);

Humm, are you sure you can always do that, i.e. that evsel->priv is not NULL?

Program received signal SIGSEGV, Segmentation fault.
0x000000000054ca04 in perf_evsel__free_stat_priv (evsel=0x2454be0) at util/stat.c:133
133 free(ps->group_data);
Missing separate debuginfos, use: dnf debuginfo-install audit-libs-2.7.7-1.fc25.x86_64 elfutils-libelf-0.169-1.fc25.x86_64 elfutils-libs-0.169-1.fc25.x86_64 libunwind-1.2-1.fc25.x86_64 perl-libs-5.24.1-386.fc25.x86_64 python-libs-2.7.13-2.fc25.x86_64 slang-2.3.0-7.fc25.x86_64 (gdb) bt

#0  0x000000000054ca04 in perf_evsel__free_stat_priv (evsel=0x2454be0) at util/stat.c:133
#1  0x000000000054cc93 in perf_evlist__free_stats (evlist=0x24541a0) at util/stat.c:189
#2  0x0000000000460311 in cmd_script (argc=0, argv=0x7fffffffe170) at builtin-script.c:3075
#3  0x00000000004be56c in run_builtin (p=0xa359f8 <commands+408>, argc=1, argv=0x7fffffffe170) at perf.c:296
#4  0x00000000004be7d9 in handle_internal_command (argc=1, argv=0x7fffffffe170) at perf.c:348
#5  0x00000000004be92b in run_argv (argcp=0x7fffffffdfcc, argv=0x7fffffffdfc0) at perf.c:392
#6  0x00000000004bed05 in main (argc=1, argv=0x7fffffffe170) at perf.c:530

(gdb)

I'm dropping the series, please retest with 'perf script'.

Make perf stat use group read if there are groups defined. The group read will get the values for all member of groups within a single syscall instead of calling read syscall for every event.

We can see considerable less amount of kernel cycles spent on single group read, than reading each event separately, like for following perf stat command:

  # perf stat -e {cycles,instructions} -I 10 -a sleep 1

Monitored with "perf stat -r 5 -e '{cycles:u,cycles:k}'"

Before:

        24,325,676      cycles:u
       297,040,775      cycles:k

       1.038554134 seconds time elapsed

After:

        25,034,418      cycles:u
       158,256,395      cycles:k

       1.036864497 seconds time elapsed

The perf_evsel__open fallback changes contributed by Andi Kleen.

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/builtin-stat.c | 30 +++++++++++++++++++++++++++---
 tools/perf/util/counts.h  |  1 +
 tools/perf/util/evsel.c   | 10 ++++++++++
 3 files changed, 38 insertions(+), 3 deletions(-)

tools/perf/builtin-stat.c +2 -1

index 48ac53b199fc..866da7aa54bf 100644
@@ -213,10 +213,20 @@ static void perf_stat__reset_stats(void)
 static int create_perf_stat_counter(struct perf_evsel *evsel)
 {
 	struct perf_event_attr *attr = &evsel->attr;
+	struct perf_evsel *leader = evsel->leader;
 
-	if (stat_config.scale)
+	if (stat_config.scale) {
 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |

tools/perf/builtin-stat.c +16 -2

 	attr->inherit = !no_inherit;
 

@@ -333,13 +343,21 @@ static int read_counter(struct perf_evsel *counter)
 			struct perf_counts_values *count;
 
 			count = perf_counts(counter->counts, cpu, thread);
-			if (perf_evsel__read(counter, cpu, thread, count)) {
+
+			/*
+			 * The leader's group read loads data into its group members
+			 * (via perf_evsel__read_counter) and sets threir count->loaded.
+			 */
+			if (!count->loaded &&
+			    perf_evsel__read_counter(counter, cpu, thread)) {
 				counter->counts->scaled = -1;
 				perf_counts(counter->counts, cpu, thread)->ena = 0;
 				perf_counts(counter->counts, cpu, thread)->run = 0;
 				return -1;
 			}
 
+			count->loaded = false;
+
 			if (STAT_RECORD) {
 				if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 					pr_err("failed to write stat event\n");

@@ -559,6 +577,11 @@ static int store_counter_ids(struct perf_evsel *counter)
 	return __store_counter_ids(counter, cpus, threads);
 }
 
+static bool perf_evsel__should_store_id(struct perf_evsel *counter)
+{
+	return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
+}
+
 static int __run_perf_stat(int argc, const char **argv)
 {
 	int interval = stat_config.interval;

@@ -631,7 +654,8 @@ static int __run_perf_stat(int argc, const char **argv)
 		if (l > unit_width)
 			unit_width = l;
 
-		if (STAT_RECORD && store_counter_ids(counter))
+		if (perf_evsel__should_store_id(counter) &&
+		    store_counter_ids(counter))
 			return -1;
 	}
 

tools/perf/util/counts.h +1 -0

index 34d8baaf558a..cb45a6aecf9d 100644
@@ -12,6 +12,7 @@ struct perf_counts_values {
 		};
 		u64 values[3];
 	};
+	bool	loaded;
 };
 
 struct perf_counts {

tools/perf/util/evsel.c +10 -0

index 89aecf3a35c7..3735c9e0080d 100644
@@ -49,6 +49,7 @@ static struct {
 	bool clockid_wrong;
 	bool lbr_flags;
 	bool write_backward;
+	bool group_read;
 } perf_missing_features;
 
 static clockid_t clockid;

@@ -1321,6 +1322,7 @@ perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
 	count->val    = val;
 	count->ena    = ena;
 	count->run    = run;
+	count->loaded = true;
 }
 
 static int

@@ -1677,6 +1679,8 @@ int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 	if (perf_missing_features.lbr_flags)
 		evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
+	if (perf_missing_features.group_read && evsel->attr.inherit)
+		evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
 retry_sample_id:
 	if (perf_missing_features.sample_id_all)
 		evsel->attr.sample_id_all = 0;

@@ -1832,6 +1836,12 @@ int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 		perf_missing_features.lbr_flags = true;
 		pr_debug2("switching off branch sample type no (cycles/flags)\n");
 		goto fallback_missing_features;
+	} else if (!perf_missing_features.group_read &&
+		    evsel->attr.inherit &&
+		   (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
+		perf_missing_features.group_read = true;
+		pr_debug2("switching off group read\n");
+		goto fallback_missing_features;
 	}
 out_close:
 	do {

Signature · 1 line

2.9.4

Quoted 18 lines

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/builtin-stat.c | 30 +++++++++++++++++++++++++++---
 tools/perf/util/counts.h  |  1 +
 tools/perf/util/evsel.c   | 10 ++++++++++
 3 files changed, 38 insertions(+), 3 deletions(-)

tools/perf/builtin-stat.c +1 -0

index 48ac53b199fc..866da7aa54bf 100644
@@ -213,10 +213,20 @@ static void perf_stat__reset_stats(void)
 static int create_perf_stat_counter(struct perf_evsel *evsel)
 {
 	struct perf_event_attr *attr = &evsel->attr;
+	struct perf_evsel *leader = evsel->leader;
 

tools/perf/builtin-stat.c static void perf_stat__reset_stats(void) · line 229

Quoted 13 lines

tools/perf/builtin-stat.c +10 -1

-	if (stat_config.scale)
+	if (stat_config.scale) {
 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
+	}
+
+	/*
+	 * The event is part of non trivial group, let's enable
+	 * the group read (for leader) and ID retrieval for all
+	 * members.
+	 */
+	if (leader->nr_members > 1)
+		attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;

I just wonder ID is really necessary. Doesn't it have same order we can traverse with the for_each_group_member()?

hi,
sending changes to enable group read of perf counters for perf stat command. It allows us to read whole group of counters within single read syscall.

Also available in here:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/stat_group

Not sure why we haven't supported yet, but anyway it was unavailable for some time due to a bug which was fixed
just recently via:
ba5213ae6b88 ("perf/core: Correct event creation with PERF_FORMAT_GROUP")