[PATCH v3 0/2] perf stat: add interval counter printing

[PATCH v3 0/2] perf stat: add interval counter printing

[Stephane Eranian]

This small patchset extends perf stat with the -I option.

The -I option enables interval printing. It expects
an interval duration in milliseconds. Minimum is
100ms. Once, activated perf stat prints events deltas
since last printout. All modes are supported.

$ perf stat -I 1000 -e cycles noploop 10
noploop for 10 seconds
1.000086918         2385155642 cycles                    #    0.000 GHz
2.000267937         2392279774 cycles                    #    0.000 GHz
3.000385400         2390971450 cycles                    #    0.000 GHz
4.000504408         2390996752 cycles                    #    0.000 GHz
5.000626878         2390853097 cycles                    #    0.000 GHz

The output format makes it easy to feed into a plotting program
such as gnuplot when the -I option is used in combination with the -x
option:

$ perf stat -x, -I 1000 -e cycles noploop 10
noploop for 10 seconds
1.000084113,2378775498,cycles
2.000245798,2391056897,cycles
3.000354445,2392089414,cycles
4.000459115,2390936603,cycles
5.000565341,2392108173,cycles

The first patch adds the code to keep previous counts. The second
is the actual extension of perf stat.

In v2, we changed the option from -T to -I based on feedback
from Andi Kleen (who is using -T in an upcoming patch).

In v3, added compute_deltas() to evsel.c to factorize
deltas computation based on Jiri's comments.

Signed-off-by: Stephane Eranian <eranian@google.com>

Stephane Eranian (2):
  perf tools: add evsel prev_raw_count field
  perf stat: add interval printing

 tools/perf/Documentation/perf-stat.txt |    4 +
 tools/perf/builtin-stat.c              |  157 ++++++++++++++++++++++++++++----
 tools/perf/util/evsel.c                |   26 ++++++
 tools/perf/util/evsel.h                |    1 +
 4 files changed, 172 insertions(+), 16 deletions(-)

-- 
1.7.9.5

[PATCH v3 1/2] perf tools: add evsel prev_raw_count field

[Stephane Eranian]

This field will be used by commands which print
counter deltas on regular timer intervals,
such as perf stat -I.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 tools/perf/util/evsel.c |   26 ++++++++++++++++++++++++++
 tools/perf/util/evsel.h |    1 +
 2 files changed, 27 insertions(+)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 1b16dd1..371e93d 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -631,6 +631,28 @@ void perf_evsel__delete(struct perf_evsel *evsel)
 	free(evsel);
 }
 
+static inline void compute_deltas(struct perf_evsel *evsel,
+				  int cpu,
+				  struct perf_counts_values *count)
+{
+	struct perf_counts_values tmp;
+
+	if (!evsel->prev_raw_counts)
+		return;
+
+	if (cpu == -1) {
+		tmp = evsel->prev_raw_counts->aggr;
+		evsel->prev_raw_counts->aggr = *count;
+	} else {
+		tmp = evsel->prev_raw_counts->cpu[cpu];
+		evsel->prev_raw_counts->cpu[cpu] = *count;
+	}
+
+	count->val = count->val - tmp.val;
+	count->ena = count->ena - tmp.ena;
+	count->run = count->run - tmp.run;
+}
+
 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 			      int cpu, int thread, bool scale)
 {
@@ -646,6 +668,8 @@ int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 		return -errno;
 
+	compute_deltas(evsel, cpu, &count);
+
 	if (scale) {
 		if (count.run == 0)
 			count.val = 0;
@@ -684,6 +708,8 @@ int __perf_evsel__read(struct perf_evsel *evsel,
 		}
 	}
 
+	compute_deltas(evsel, -1, aggr);
+
 	evsel->counts->scaled = 0;
 	if (scale) {
 		if (aggr->run == 0) {
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 3d2b801..63ea570 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -53,6 +53,7 @@ struct perf_evsel {
 	struct xyarray		*sample_id;
 	u64			*id;
 	struct perf_counts	*counts;
+	struct perf_counts	*prev_raw_counts;
 	int			idx;
 	u32			ids;
 	struct hists		hists;
-- 
1.7.9.5

[PATCH v3 2/2] perf stat: add interval printing

[Stephane Eranian]

This patch adds a new printing mode for perf stat.
It allows internval printing. That means perf stat
can now print event deltas at regular time interval.
This is useful to detect phases in programs.

The -I option enables interval printing. It expects
an interval duration in milliseconds. Minimum is
100ms. Once, activated perf stat prints events deltas
since last printout. All modes are supported.

$ perf stat -I 1000 -e cycles noploop 10
noploop for 10 seconds
1.000086918         2385155642 cycles                    #    0.000 GHz
2.000267937         2392279774 cycles                    #    0.000 GHz
3.000385400         2390971450 cycles                    #    0.000 GHz
4.000504408         2390996752 cycles                    #    0.000 GHz
5.000626878         2390853097 cycles                    #    0.000 GHz

The output format makes it easy to feed into a plotting program
such as gnuplot when the -I option is used in combination with the -x
option:

$ perf stat -x, -I 1000 -e cycles noploop 10
noploop for 10 seconds
1.000084113,2378775498,cycles
2.000245798,2391056897,cycles
3.000354445,2392089414,cycles
4.000459115,2390936603,cycles
5.000565341,2392108173,cycles

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 tools/perf/Documentation/perf-stat.txt |    4 +
 tools/perf/builtin-stat.c              |  157 ++++++++++++++++++++++++++++----
 2 files changed, 145 insertions(+), 16 deletions(-)

diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
index cf0c310..5289da3 100644
--- a/tools/perf/Documentation/perf-stat.txt
+++ b/tools/perf/Documentation/perf-stat.txt
@@ -114,6 +114,10 @@ Log output to fd, instead of stderr.  Complementary to --output, and mutually ex
 
 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
 
+-I msecs::
+--interval-print msecs::
+	print count deltas every N milliseconds (minimum: 100ms)
+	example: perf stat -I 1000 -e cycles -a sleep 5
 
 EXAMPLES
 --------
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index c247fac..67aaff2 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -65,6 +65,10 @@
 #define CNTR_NOT_SUPPORTED	"<not supported>"
 #define CNTR_NOT_COUNTED	"<not counted>"
 
+static void print_stat(int argc, const char **argv);
+static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
+static void print_counter(struct perf_evsel *counter, char *prefix);
+
 static struct perf_evlist	*evsel_list;
 
 static struct perf_target	target = {
@@ -87,6 +91,8 @@ static FILE			*output				= NULL;
 static const char		*pre_cmd			= NULL;
 static const char		*post_cmd			= NULL;
 static bool			sync_run			= false;
+static int			interval			= 0;
+static struct timespec		ref_time;
 
 static volatile int done = 0;
 
@@ -94,6 +100,28 @@ struct perf_stat {
 	struct stats	  res_stats[3];
 };
 
+static inline void diff_timespec(struct timespec *r, struct timespec *a,
+				 struct timespec *b)
+{
+	r->tv_sec = a->tv_sec - b->tv_sec;
+	if (a->tv_nsec < b->tv_nsec) {
+		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
+		r->tv_sec--;
+	} else {
+		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
+	}
+}
+
+static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
+{
+	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
+}
+
+static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
+{
+	return perf_evsel__cpus(evsel)->nr;
+}
+
 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
 {
 	evsel->priv = zalloc(sizeof(struct perf_stat));
@@ -106,14 +134,27 @@ static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
 	evsel->priv = NULL;
 }
 
-static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
+static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
 {
-	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
+	void *addr;
+	size_t sz;
+
+	sz = sizeof(*evsel->counts) +
+	     (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
+
+	addr = zalloc(sz);
+	if (!addr)
+		return -ENOMEM;
+
+	evsel->prev_raw_counts =  addr;
+
+	return 0;
 }
 
-static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
+static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
 {
-	return perf_evsel__cpus(evsel)->nr;
+	free(evsel->prev_raw_counts);
+	evsel->prev_raw_counts = NULL;
 }
 
 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
@@ -269,15 +310,57 @@ static int read_counter(struct perf_evsel *counter)
 	return 0;
 }
 
+static void print_interval(void)
+{
+	struct perf_evsel *counter;
+	struct perf_stat *ps;
+	struct timespec ts, rs;
+	char prefix[64];
+
+	if (no_aggr) {
+		list_for_each_entry(counter, &evsel_list->entries, node) {
+			ps = counter->priv;
+			memset(ps->res_stats, 0, sizeof(ps->res_stats));
+			read_counter(counter);
+		}
+	} else {
+		list_for_each_entry(counter, &evsel_list->entries, node) {
+			ps = counter->priv;
+			memset(ps->res_stats, 0, sizeof(ps->res_stats));
+			read_counter_aggr(counter);
+		}
+	}
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	diff_timespec(&rs, &ts, &ref_time);
+	sprintf(prefix, "%lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
+
+	if (no_aggr) {
+		list_for_each_entry(counter, &evsel_list->entries, node)
+			print_counter(counter, prefix);
+	} else {
+		list_for_each_entry(counter, &evsel_list->entries, node)
+			print_counter_aggr(counter, prefix);
+	}
+}
+
 static int __run_perf_stat(int argc __maybe_unused, const char **argv)
 {
 	unsigned long long t0, t1;
 	struct perf_evsel *counter;
-	int status = 0;
+	struct timespec ts;
+	int status = 0, ret;
 	int child_ready_pipe[2], go_pipe[2];
 	const bool forks = (argc > 0);
 	char buf;
 
+	if (interval) {
+		ts.tv_sec  = interval / 1000;
+		ts.tv_nsec = (interval % 1000) * 1000000;
+	} else {
+		ts.tv_sec  = 1;
+		ts.tv_nsec = 0;
+	}
+
 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 		perror("failed to create pipes");
 		return -1;
@@ -377,14 +460,26 @@ static int __run_perf_stat(int argc __maybe_unused, const char **argv)
 	 * Enable counters and exec the command:
 	 */
 	t0 = rdclock();
+	clock_gettime(CLOCK_MONOTONIC, &ref_time);
 
 	if (forks) {
 		close(go_pipe[1]);
+		if (interval) {
+			while (!waitpid(child_pid, &status, WNOHANG)) {
+				ret = nanosleep(&ts, NULL);
+				if (ret == 0 && interval)
+					print_interval();
+			}
+		}
 		wait(&status);
 		if (WIFSIGNALED(status))
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
-		while(!done) sleep(1);
+		while (!done) {
+			ret = nanosleep(&ts, NULL);
+			if (ret == 0 && interval)
+				print_interval();
+		}
 	}
 
 	t1 = rdclock();
@@ -470,7 +565,7 @@ static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
 	if (evsel->cgrp)
 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 
-	if (csv_output)
+	if (csv_output || interval)
 		return;
 
 	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
@@ -684,12 +779,11 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 	if (evsel->cgrp)
 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 
-	if (csv_output)
+	if (csv_output || interval)
 		return;
 
 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
 		total = avg_stats(&runtime_cycles_stats[cpu]);
-
 		if (total)
 			ratio = avg / total;
 
@@ -783,12 +877,15 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
  * Print out the results of a single counter:
  * aggregated counts in system-wide mode
  */
-static void print_counter_aggr(struct perf_evsel *counter)
+static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 {
 	struct perf_stat *ps = counter->priv;
 	double avg = avg_stats(&ps->res_stats[0]);
 	int scaled = counter->counts->scaled;
 
+	if (prefix)
+		fprintf(output, "%s", prefix);
+
 	if (scaled == -1) {
 		fprintf(output, "%*s%s%*s",
 			csv_output ? 0 : 18,
@@ -831,7 +928,7 @@ static void print_counter_aggr(struct perf_evsel *counter)
  * Print out the results of a single counter:
  * does not use aggregated count in system-wide
  */
-static void print_counter(struct perf_evsel *counter)
+static void print_counter(struct perf_evsel *counter, char *prefix)
 {
 	u64 ena, run, val;
 	int cpu;
@@ -840,6 +937,10 @@ static void print_counter(struct perf_evsel *counter)
 		val = counter->counts->cpu[cpu].val;
 		ena = counter->counts->cpu[cpu].ena;
 		run = counter->counts->cpu[cpu].run;
+
+		if (prefix)
+			fprintf(output, "%s", prefix);
+
 		if (run == 0 || ena == 0) {
 			fprintf(output, "CPU%*d%s%*s%s%*s",
 				csv_output ? 0 : -4,
@@ -877,6 +978,8 @@ static void print_counter(struct perf_evsel *counter)
 static void print_stat(int argc, const char **argv)
 {
 	struct perf_evsel *counter;
+	struct timespec ts, rs;
+	char prefix[64] = { 0, };
 	int i;
 
 	fflush(stdout);
@@ -899,12 +1002,18 @@ static void print_stat(int argc, const char **argv)
 		fprintf(output, ":\n\n");
 	}
 
+	if (interval) {
+		clock_gettime(CLOCK_MONOTONIC, &ts);
+		diff_timespec(&rs, &ts, &ref_time);
+		sprintf(prefix, "%lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
+	}
+
 	if (no_aggr) {
 		list_for_each_entry(counter, &evsel_list->entries, node)
-			print_counter(counter);
+			print_counter(counter, prefix);
 	} else {
 		list_for_each_entry(counter, &evsel_list->entries, node)
-			print_counter_aggr(counter);
+			print_counter_aggr(counter, prefix);
 	}
 
 	if (!csv_output) {
@@ -925,7 +1034,7 @@ static volatile int signr = -1;
 
 static void skip_signal(int signo)
 {
-	if(child_pid == -1)
+	if((child_pid == -1) || interval)
 		done = 1;
 
 	signr = signo;
@@ -1145,6 +1254,8 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 			"command to run prior to the measured command"),
 	OPT_STRING(0, "post", &post_cmd, "command",
 			"command to run after to the measured command"),
+	OPT_INTEGER('I', "interval-print", &interval,
+		    "print counts at regular interval in ms (>= 100)"),
 	OPT_END()
 	};
 	const char * const stat_usage[] = {
@@ -1245,12 +1356,23 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 		usage_with_options(stat_usage, options);
 		return -1;
 	}
+	if (interval < 0 || (interval > 0 && interval < 100)) {
+		pr_err("print interval must be >= 100ms\n");
+		usage_with_options(stat_usage, options);
+		return -1;
+	}
 
 	list_for_each_entry(pos, &evsel_list->entries, node) {
 		if (perf_evsel__alloc_stat_priv(pos) < 0 ||
 		    perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0)
 			goto out_free_fd;
 	}
+	if (interval) {
+		list_for_each_entry(pos, &evsel_list->entries, node) {
+			if (perf_evsel__alloc_prev_raw_counts(pos) < 0)
+				goto out_free_fd;
+		}
+	}
 
 	/*
 	 * We dont want to block the signals - that would cause
@@ -1260,6 +1382,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 	 */
 	atexit(sig_atexit);
 	signal(SIGINT,  skip_signal);
+	signal(SIGCHLD, skip_signal);
 	signal(SIGALRM, skip_signal);
 	signal(SIGABRT, skip_signal);
 
@@ -1272,11 +1395,13 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 		status = run_perf_stat(argc, argv);
 	}
 
-	if (status != -1)
+	if (status != -1 && !interval)
 		print_stat(argc, argv);
 out_free_fd:
-	list_for_each_entry(pos, &evsel_list->entries, node)
+	list_for_each_entry(pos, &evsel_list->entries, node) {
 		perf_evsel__free_stat_priv(pos);
+		perf_evsel__free_prev_raw_counts(pos);
+	}
 	perf_evlist__delete_maps(evsel_list);
 out:
 	perf_evlist__delete(evsel_list);
-- 
1.7.9.5

Re: [PATCH v3 2/2] perf stat: add interval printing

[Andi Kleen]

>  	if (forks) {
>  		close(go_pipe[1]);
> +		if (interval) {
> +			while (!waitpid(child_pid, &status, WNOHANG)) {
> +				ret = nanosleep(&ts, NULL);
> +				if (ret == 0 && interval)
> +					print_interval();

It would be good to make sure that if someone Ctrl-Cs
you still print the latest interval(even if shorter)

The rest looks good to me.

-Andi

Re: [PATCH v3 2/2] perf stat: add interval printing

[Stephane Eranian]

On Fri, Jan 18, 2013 at 7:30 PM, Andi Kleen <ak@linux.intel.com> wrote:
>>       if (forks) {
>>               close(go_pipe[1]);
>> +             if (interval) {
>> +                     while (!waitpid(child_pid, &status, WNOHANG)) {
>> +                             ret = nanosleep(&ts, NULL);
>> +                             if (ret == 0 && interval)
>> +                                     print_interval();
>
> It would be good to make sure that if someone Ctrl-Cs
> you still print the latest interval(even if shorter)
>
Ok, I can remove the test. The interval is redundant here anyway.
Will post V4 shortly.