Skip to contentSkip to conversation

perf ftrace: Add --stats option to latency subcommand

PATCH In review Namhyung Kim 2 participants 7 messages Jan 3, 2025 – Jan 7, 2025 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH 1/3 perf ftrace: Add --stats option to latency subcommand

Sometimes users also want to see average latency as well as histogram. The --stats option is to display latency statistics at the end.

  $ sudo ./perf ftrace latency -ab --stats -T synchronize_rcu -- ...
  #   DURATION     |      COUNT | GRAPH                                          |
       0 - 1    us |          0 |                                                |
       1 - 2    us |          0 |                                                |
       2 - 4    us |          0 |                                                |
       4 - 8    us |          0 |                                                |
       8 - 16   us |          0 |                                                |
      16 - 32   us |          0 |                                                |
      32 - 64   us |          0 |                                                |
      64 - 128  us |          0 |                                                |
     128 - 256  us |          0 |                                                |
     256 - 512  us |          0 |                                                |
     512 - 1024 us |          0 |                                                |
       1 - 2    ms |          0 |                                                |
       2 - 4    ms |          0 |                                                |
       4 - 8    ms |          0 |                                                |
       8 - 16   ms |          1 | #####                                          |
      16 - 32   ms |          7 | ########################################       |
      32 - 64   ms |          0 |                                                |
      64 - 128  ms |          0 |                                                |
     128 - 256  ms |          0 |                                                |
     256 - 512  ms |          0 |                                                |
     512 - 1024 ms |          0 |                                                |
       1 - ...   s |          0 |                                                |
  # statistics  (in usec)
    total time:               171832
      avg time:                21479
      max time:                30906
      min time:                15869
         count:                    8
---
 tools/perf/Documentation/perf-ftrace.txt    |  4 ++++
 tools/perf/builtin-ftrace.c                 | 24 ++++++++++++++++++++-
 tools/perf/util/bpf_ftrace.c                | 16 ++++++++++++--
 tools/perf/util/bpf_skel/func_latency.bpf.c | 19 ++++++++++++++++
 tools/perf/util/ftrace.h                    |  4 +++-
 5 files changed, 63 insertions(+), 4 deletions(-)
tools/perf/Documentation/perf-ftrace.txt +4 -0
index eccc0483f7faecad..16b067eb5d84746d 100644
@@ -160,6 +160,10 @@ OPTIONS for 'perf ftrace latency'
 	-n/--use-nsec). The setting is ignored if the value results in more than
 	22 buckets.
 
+--stats::
+	Display latency statistics at the end.
+
+
 OPTIONS for 'perf ftrace profile'
 ---------------------------------
 
tools/perf/builtin-ftrace.c +23 -1
index 834074162a63bf2b..0ac2b20b272bd790 100644
@@ -43,6 +43,8 @@
 static volatile sig_atomic_t workload_exec_errno;
 static volatile sig_atomic_t done;
 
+static struct stats latency_stats;  /* for tracepoints */
+
 static void sig_handler(int sig __maybe_unused)
 {
 	done = true;
@@ -801,6 +803,13 @@ static void make_histogram(struct perf_ftrace *ftrace, int buckets[],
 do_inc:
 		buckets[i]++;
 
+		if (ftrace->stats) {
+			if (num >= min_latency)
+				num += min_latency;
+
+			update_stats(&latency_stats, num);
+		}
+
 next:
 		/* empty the line buffer for the next output  */
 		linebuf[0] = '\0';
@@ -894,6 +903,14 @@ static void display_histogram(struct perf_ftrace *ftrace, int buckets[])
 	printf(" | %10d | %.*s%*s |\n", buckets[NUM_BUCKET - 1],
 	       bar_len, bar, bar_total - bar_len, "");
 
+	if (ftrace->stats) {
+		printf("\n# statistics  (in %s)\n", ftrace->use_nsec ? "nsec" : "usec");
+		printf("  total time: %20.0f\n", latency_stats.mean * latency_stats.n);
+		printf("    avg time: %20.0f\n", latency_stats.mean);
+		printf("    max time: %20"PRIu64"\n", latency_stats.max);
+		printf("    min time: %20"PRIu64"\n", latency_stats.min);
+		printf("       count: %20.0f\n", latency_stats.n);
+	}
 }
 
 static int prepare_func_latency(struct perf_ftrace *ftrace)
@@ -932,6 +949,9 @@ static int prepare_func_latency(struct perf_ftrace *ftrace)
 	if (fd < 0)
 		pr_err("failed to open trace_pipe\n");
 
+	if (ftrace->stats)
+		init_stats(&latency_stats);
+
 	put_tracing_file(trace_file);
 	return fd;
 }
@@ -961,7 +981,7 @@ static int stop_func_latency(struct perf_ftrace *ftrace)
 static int read_func_latency(struct perf_ftrace *ftrace, int buckets[])
 {
 	if (ftrace->target.use_bpf)
-		return perf_ftrace__latency_read_bpf(ftrace, buckets);
+		return perf_ftrace__latency_read_bpf(ftrace, buckets, &latency_stats);
 
 	return 0;
 }
@@ -1620,6 +1640,8 @@ int cmd_ftrace(int argc, const char **argv)
 		    "Minimum latency (1st bucket). Works only with --bucket-range."),
 	OPT_UINTEGER(0, "max-latency", &ftrace.max_latency,
 		    "Maximum latency (last bucket). Works only with --bucket-range and total buckets less than 22."),
+	OPT_BOOLEAN(0, "stats", &ftrace.stats,
+		    "Collect function latency stats too"),
 	OPT_PARENT(common_options),
 	};
 	const struct option profile_options[] = {
tools/perf/util/bpf_ftrace.c +14 -2
index bc484e65fb8f69ca..ecd5e16df9730e0d 100644
@@ -11,6 +11,7 @@
 #include "util/debug.h"
 #include "util/evlist.h"
 #include "util/bpf_counter.h"
+#include "util/stat.h"
 
 #include "util/bpf_skel/func_latency.skel.h"
 
@@ -53,6 +54,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
 	}
 
 	skel->rodata->use_nsec = ftrace->use_nsec;
+	skel->rodata->stats = ftrace->stats;
 
 	set_max_rlimit();
 
@@ -86,6 +88,9 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
 		}
 	}
 
+	if (ftrace->stats)
+		skel->bss->min = INT64_MAX;
+
 	skel->links.func_begin = bpf_program__attach_kprobe(skel->progs.func_begin,
 							    false, func->name);
 	if (IS_ERR(skel->links.func_begin)) {
@@ -121,8 +126,8 @@ int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused)
 	return 0;
 }
 
-int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
-				  int buckets[])
+int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
+				  int buckets[], struct stats *stats)
 {
 	int i, fd, err;
 	u32 idx;
@@ -146,6 +151,13 @@ int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
 			buckets[idx] += hist[i];
 	}
 
+	if (ftrace->stats && skel->bss->count) {
+		stats->mean = skel->bss->total / skel->bss->count;
+		stats->n = skel->bss->count;
+		stats->max = skel->bss->max;
+		stats->min = skel->bss->min;
+	}
+
 	free(hist);
 	return 0;
 }
tools/perf/util/bpf_skel/func_latency.bpf.c +19 -0
index 4df54e1b1411b10a..9b68eea8fd444699 100644
@@ -38,12 +38,19 @@ struct {
 
 int enabled = 0;
 
+// stats
+__s64 total;
+__s64 count;
+__s64 max;
+__s64 min;
+
 const volatile int has_cpu = 0;
 const volatile int has_task = 0;
 const volatile int use_nsec = 0;
 const volatile unsigned int bucket_range;
 const volatile unsigned int min_latency;
 const volatile unsigned int max_latency;
+const volatile int stats = 0;
 
 SEC("kprobe/func")
 int BPF_PROG(func_begin)
@@ -136,6 +143,18 @@ int BPF_PROG(func_end)
 			return 0;
 
 		*hist += 1;
+
+		if (stats) {
+			delta /= cmp_base;
+
+			__sync_fetch_and_add(&total, delta);
+			__sync_fetch_and_add(&count, 1);
+
+			if (delta > max)
+				max = delta;
+			if (delta < min)
+				min = delta;
+		}
 	}
 
 	return 0;
tools/perf/util/ftrace.h +3 -1
index f218703063f74786..76038ed712085fbc 100644
@@ -7,6 +7,7 @@
 
 struct evlist;
 struct hashamp;
+struct stats;
 
 struct perf_ftrace {
 	struct evlist		*evlist;
@@ -23,6 +24,7 @@ struct perf_ftrace {
 	unsigned int		bucket_range;
 	unsigned int		min_latency;
 	unsigned int		max_latency;
+	bool			stats;
 	int			graph_depth;
 	int			func_stack_trace;
 	int			func_irq_info;
@@ -46,7 +48,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace);
 int perf_ftrace__latency_start_bpf(struct perf_ftrace *ftrace);
 int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace);
 int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
-				  int buckets[]);
+				  int buckets[], struct stats *stats);
 int perf_ftrace__latency_cleanup_bpf(struct perf_ftrace *ftrace);
 
 #else  /* !HAVE_BPF_SKEL */
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
author maintainer
#1 lore raw

PATCH 2/3 perf ftrace: Add --graph-opts option to profile subcommand

Like trace subcommand, it should be able to pass some options to control the tracing behavior for the function graph tracer. But some options are limited in order to maintain the internal behavior.

For example, it can limit the function call depth like below:

  # perf ftrace profile --graph-opts depth=5 -- myprog
---
 tools/perf/Documentation/perf-ftrace.txt | 8 ++++++++
 tools/perf/builtin-ftrace.c              | 4 ++++
 2 files changed, 12 insertions(+)
tools/perf/Documentation/perf-ftrace.txt +8 -0
index 16b067eb5d84746d..4a69a1135f07341b 100644
@@ -205,6 +205,14 @@ OPTIONS for 'perf ftrace profile'
 	Sort the result by the given field.  Available values are:
 	total, avg, max, count, name.  Default is 'total'.
 
+--graph-opts::
+	List of options allowed to set:
+
+	  - nosleep-time - Measure on-CPU time only for function_graph tracer.
+	  - noirqs       - Ignore functions that happen inside interrupt.
+	  - thresh=<n>   - Setup trace duration threshold in microseconds.
+	  - depth=<n>    - Set max depth for function graph tracer to follow.
+
 
 SEE ALSO
 --------
tools/perf/builtin-ftrace.c +4 -0
index 0ac2b20b272bd790..d747cca2403e494e 100644
@@ -1072,6 +1072,7 @@ static int prepare_func_profile(struct perf_ftrace *ftrace)
 {
 	ftrace->tracer = "function_graph";
 	ftrace->graph_tail = 1;
+	ftrace->graph_verbose = 0;
 
 	ftrace->profile_hash = hashmap__new(profile_hash, profile_equal, NULL);
 	if (ftrace->profile_hash == NULL)
@@ -1660,6 +1661,9 @@ int cmd_ftrace(int argc, const char **argv)
 	OPT_CALLBACK('s', "sort", &profile_sort, "key",
 		     "Sort result by key: total (default), avg, max, count, name.",
 		     parse_sort_key),
+	OPT_CALLBACK(0, "graph-opts", &ftrace, "options",
+		     "Graph tracer options, available options: nosleep-time,noirqs,thresh=<n>,depth=<n>",
+		     parse_graph_tracer_opts),
 	OPT_PARENT(common_options),
 	};
 	const struct option *options = ftrace_options;
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
reviewer Reviewed-by on v1 2/3
#2 lore raw

Re: [PATCH 2/3] perf ftrace: Add --graph-opts option to profile subcommand

Quoted 6 lines · Namhyung Kim

Like trace subcommand, it should be able to pass some options to control the tracing behavior for the function graph tracer. But some options are limited in order to maintain the internal behavior.

For example, it can limit the function call depth like below:

Quoted 8 lines
   # perf ftrace profile --graph-opts depth=5 -- myprog
---
  tools/perf/Documentation/perf-ftrace.txt | 8 ++++++++
  tools/perf/builtin-ftrace.c              | 4 ++++
  2 files changed, 12 insertions(+)
Quoted 41 lines
tools/perf/Documentation/perf-ftrace.txt +8 -0
index 16b067eb5d84746d..4a69a1135f07341b 100644
@@ -205,6 +205,14 @@ OPTIONS for 'perf ftrace profile'
  	Sort the result by the given field.  Available values are:
  	total, avg, max, count, name.  Default is 'total'.
 
+--graph-opts::
+	List of options allowed to set:
+
+	  - nosleep-time - Measure on-CPU time only for function_graph tracer.
+	  - noirqs       - Ignore functions that happen inside interrupt.
+	  - thresh=<n>   - Setup trace duration threshold in microseconds.
+	  - depth=<n>    - Set max depth for function graph tracer to follow.
+
 
  SEE ALSO
  --------
tools/perf/builtin-ftrace.c +4 -0
index 0ac2b20b272bd790..d747cca2403e494e 100644
@@ -1072,6 +1072,7 @@ static int prepare_func_profile(struct perf_ftrace *ftrace)
  {
  	ftrace->tracer = "function_graph";
  	ftrace->graph_tail = 1;
+	ftrace->graph_verbose = 0;
 
  	ftrace->profile_hash = hashmap__new(profile_hash, profile_equal, NULL);
  	if (ftrace->profile_hash == NULL)
@@ -1660,6 +1661,9 @@ int cmd_ftrace(int argc, const char **argv)
  	OPT_CALLBACK('s', "sort", &profile_sort, "key",
  		     "Sort result by key: total (default), avg, max, count, name.",
  		     parse_sort_key),
+	OPT_CALLBACK(0, "graph-opts", &ftrace, "options",
+		     "Graph tracer options, available options: nosleep-time,noirqs,thresh=<n>,depth=<n>",
+		     parse_graph_tracer_opts),
  	OPT_PARENT(common_options),
  	};
  	const struct option *options = ftrace_options;
author maintainer
#3 lore raw

PATCH 3/3 perf test: Update ftrace test to use --graph-opts

In reply to Namhyung Kim

I found it failed on machines with limited memory because 16M byte per-cpu buffer is too big. The reason it added the option is not to miss tracing data. Thus we can limit the data size by reducing the function call depth instead of increasing the buffer size to handle the whole data.

As it used the same option in the test_ftrace_trace() and it was able to find the sleep function, it should work with the profile subcommand.

Get rid of other grep commands which might be affected by the depth change.

Cc Thomas Richter <[email protected]>
---
 tools/perf/tests/shell/ftrace.sh | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
tools/perf/tests/shell/ftrace.sh +1 -4
index 2df05052c324a21c..c243731d2fbf67aa 100755
@@ -67,11 +67,8 @@ test_ftrace_latency() {
 
 test_ftrace_profile() {
     echo "perf ftrace profile test"
-    perf ftrace profile -m 16M sleep 0.1 > "${output}"
+    perf ftrace profile --graph-opts depth=5 sleep 0.1 > "${output}"
     grep ^# "${output}"
-    grep sleep "${output}"
-    grep schedule "${output}"
-    grep execve "${output}"
     time_re="[[:space:]]+1[[:digit:]]{5}\.[[:digit:]]{3}"
     # 100283.000 100283.000 100283.000          1   __x64_sys_clock_nanosleep
     # Check for one *clock_nanosleep line with a Count of just 1 that takes a bit more than 0.1 seconds
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
reviewer Reviewed-by on v1 3/3
#4 lore raw

Re: [PATCH 3/3] perf test: Update ftrace test to use --graph-opts

Quoted 9 lines · Namhyung Kim

I found it failed on machines with limited memory because 16M byte per-cpu buffer is too big. The reason it added the option is not to miss tracing data. Thus we can limit the data size by reducing the function call depth instead of increasing the buffer size to handle the whole data.

As it used the same option in the test_ftrace_trace() and it was able to find the sleep function, it should work with the profile subcommand.

Quoted 9 lines

Get rid of other grep commands which might be affected by the depth change.

Cc Thomas Richter <[email protected]>
---
  tools/perf/tests/shell/ftrace.sh | 5 +----
  1 file changed, 1 insertion(+), 4 deletions(-)
Quoted 17 lines
tools/perf/tests/shell/ftrace.sh +1 -4
index 2df05052c324a21c..c243731d2fbf67aa 100755
@@ -67,11 +67,8 @@ test_ftrace_latency() {
 
  test_ftrace_profile() {
      echo "perf ftrace profile test"
-    perf ftrace profile -m 16M sleep 0.1 > "${output}"
+    perf ftrace profile --graph-opts depth=5 sleep 0.1 > "${output}"
      grep ^# "${output}"
-    grep sleep "${output}"
-    grep schedule "${output}"
-    grep execve "${output}"
      time_re="[[:space:]]+1[[:digit:]]{5}\.[[:digit:]]{3}"
      # 100283.000 100283.000 100283.000          1   __x64_sys_clock_nanosleep
      # Check for one *clock_nanosleep line with a Count of just 1 that takes a bit more than 0.1 seconds
reviewer Reviewed-by on v1 1/3
#5 lore raw
In reply to Namhyung Kim
Quoted 248 lines · Namhyung Kim

Sometimes users also want to see average latency as well as histogram. The --stats option is to display latency statistics at the end.

   $ sudo ./perf ftrace latency -ab --stats -T synchronize_rcu -- ...
   #   DURATION     |      COUNT | GRAPH                                          |
        0 - 1    us |          0 |                                                |
        1 - 2    us |          0 |                                                |
        2 - 4    us |          0 |                                                |
        4 - 8    us |          0 |                                                |
        8 - 16   us |          0 |                                                |
       16 - 32   us |          0 |                                                |
       32 - 64   us |          0 |                                                |
       64 - 128  us |          0 |                                                |
      128 - 256  us |          0 |                                                |
      256 - 512  us |          0 |                                                |
      512 - 1024 us |          0 |                                                |
        1 - 2    ms |          0 |                                                |
        2 - 4    ms |          0 |                                                |
        4 - 8    ms |          0 |                                                |
        8 - 16   ms |          1 | #####                                          |
       16 - 32   ms |          7 | ########################################       |
       32 - 64   ms |          0 |                                                |
       64 - 128  ms |          0 |                                                |
      128 - 256  ms |          0 |                                                |
      256 - 512  ms |          0 |                                                |
      512 - 1024 ms |          0 |                                                |
        1 - ...   s |          0 |                                                |
   # statistics  (in usec)
     total time:               171832
       avg time:                21479
       max time:                30906
       min time:                15869
          count:                    8
---
  tools/perf/Documentation/perf-ftrace.txt    |  4 ++++
  tools/perf/builtin-ftrace.c                 | 24 ++++++++++++++++++++-
  tools/perf/util/bpf_ftrace.c                | 16 ++++++++++++--
  tools/perf/util/bpf_skel/func_latency.bpf.c | 19 ++++++++++++++++
  tools/perf/util/ftrace.h                    |  4 +++-
  5 files changed, 63 insertions(+), 4 deletions(-)
tools/perf/Documentation/perf-ftrace.txt +4 -0
index eccc0483f7faecad..16b067eb5d84746d 100644
@@ -160,6 +160,10 @@ OPTIONS for 'perf ftrace latency'
  	-n/--use-nsec). The setting is ignored if the value results in more than
  	22 buckets.
 
+--stats::
+	Display latency statistics at the end.
+
+
  OPTIONS for 'perf ftrace profile'
  ---------------------------------
 
tools/perf/builtin-ftrace.c +23 -1
index 834074162a63bf2b..0ac2b20b272bd790 100644
@@ -43,6 +43,8 @@
  static volatile sig_atomic_t workload_exec_errno;
  static volatile sig_atomic_t done;
 
+static struct stats latency_stats;  /* for tracepoints */
+
  static void sig_handler(int sig __maybe_unused)
  {
  	done = true;
@@ -801,6 +803,13 @@ static void make_histogram(struct perf_ftrace *ftrace, int buckets[],
  do_inc:
  		buckets[i]++;
 
+		if (ftrace->stats) {
+			if (num >= min_latency)
+				num += min_latency;
+
+			update_stats(&latency_stats, num);
+		}
+
  next:
  		/* empty the line buffer for the next output  */
  		linebuf[0] = '\0';
@@ -894,6 +903,14 @@ static void display_histogram(struct perf_ftrace *ftrace, int buckets[])
  	printf(" | %10d | %.*s%*s |\n", buckets[NUM_BUCKET - 1],
  	       bar_len, bar, bar_total - bar_len, "");
 
+	if (ftrace->stats) {
+		printf("\n# statistics  (in %s)\n", ftrace->use_nsec ? "nsec" : "usec");
+		printf("  total time: %20.0f\n", latency_stats.mean * latency_stats.n);
+		printf("    avg time: %20.0f\n", latency_stats.mean);
+		printf("    max time: %20"PRIu64"\n", latency_stats.max);
+		printf("    min time: %20"PRIu64"\n", latency_stats.min);
+		printf("       count: %20.0f\n", latency_stats.n);
+	}
  }
 
  static int prepare_func_latency(struct perf_ftrace *ftrace)
@@ -932,6 +949,9 @@ static int prepare_func_latency(struct perf_ftrace *ftrace)
  	if (fd < 0)
  		pr_err("failed to open trace_pipe\n");
 
+	if (ftrace->stats)
+		init_stats(&latency_stats);
+
  	put_tracing_file(trace_file);
  	return fd;
  }
@@ -961,7 +981,7 @@ static int stop_func_latency(struct perf_ftrace *ftrace)
  static int read_func_latency(struct perf_ftrace *ftrace, int buckets[])
  {
  	if (ftrace->target.use_bpf)
-		return perf_ftrace__latency_read_bpf(ftrace, buckets);
+		return perf_ftrace__latency_read_bpf(ftrace, buckets, &latency_stats);
 
  	return 0;
  }
@@ -1620,6 +1640,8 @@ int cmd_ftrace(int argc, const char **argv)
  		    "Minimum latency (1st bucket). Works only with --bucket-range."),
  	OPT_UINTEGER(0, "max-latency", &ftrace.max_latency,
  		    "Maximum latency (last bucket). Works only with --bucket-range and total buckets less than 22."),
+	OPT_BOOLEAN(0, "stats", &ftrace.stats,
+		    "Collect function latency stats too"),
  	OPT_PARENT(common_options),
  	};
  	const struct option profile_options[] = {
tools/perf/util/bpf_ftrace.c +14 -2
index bc484e65fb8f69ca..ecd5e16df9730e0d 100644
@@ -11,6 +11,7 @@
  #include "util/debug.h"
  #include "util/evlist.h"
  #include "util/bpf_counter.h"
+#include "util/stat.h"
 
  #include "util/bpf_skel/func_latency.skel.h"
 
@@ -53,6 +54,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
  	}
 
  	skel->rodata->use_nsec = ftrace->use_nsec;
+	skel->rodata->stats = ftrace->stats;
 
  	set_max_rlimit();
 
@@ -86,6 +88,9 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
  		}
  	}
 
+	if (ftrace->stats)
+		skel->bss->min = INT64_MAX;
+
  	skel->links.func_begin = bpf_program__attach_kprobe(skel->progs.func_begin,
  							    false, func->name);
  	if (IS_ERR(skel->links.func_begin)) {
@@ -121,8 +126,8 @@ int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused)
  	return 0;
  }
 
-int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
-				  int buckets[])
+int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
+				  int buckets[], struct stats *stats)
  {
  	int i, fd, err;
  	u32 idx;
@@ -146,6 +151,13 @@ int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
  			buckets[idx] += hist[i];
  	}
 
+	if (ftrace->stats && skel->bss->count) {
+		stats->mean = skel->bss->total / skel->bss->count;
+		stats->n = skel->bss->count;
+		stats->max = skel->bss->max;
+		stats->min = skel->bss->min;
+	}
+
  	free(hist);
  	return 0;
  }
tools/perf/util/bpf_skel/func_latency.bpf.c +19 -0
index 4df54e1b1411b10a..9b68eea8fd444699 100644
@@ -38,12 +38,19 @@ struct {
 
  int enabled = 0;
 
+// stats
+__s64 total;
+__s64 count;
+__s64 max;
+__s64 min;
+
  const volatile int has_cpu = 0;
  const volatile int has_task = 0;
  const volatile int use_nsec = 0;
  const volatile unsigned int bucket_range;
  const volatile unsigned int min_latency;
  const volatile unsigned int max_latency;
+const volatile int stats = 0;
 
  SEC("kprobe/func")
  int BPF_PROG(func_begin)
@@ -136,6 +143,18 @@ int BPF_PROG(func_end)
  			return 0;
 
  		*hist += 1;
+
+		if (stats) {
+			delta /= cmp_base;
+
+			__sync_fetch_and_add(&total, delta);
+			__sync_fetch_and_add(&count, 1);
+
+			if (delta > max)
+				max = delta;
+			if (delta < min)
+				min = delta;
+		}
  	}
 
  	return 0;
tools/perf/util/ftrace.h +2 -0
index f218703063f74786..76038ed712085fbc 100644
@@ -7,6 +7,7 @@
 
  struct evlist;
  struct hashamp;
+struct stats;
 
  struct perf_ftrace {
  	struct evlist		*evlist;
@@ -23,6 +24,7 @@ struct perf_ftrace {
  	unsigned int		bucket_range;
  	unsigned int		min_latency;
  	unsigned int		max_latency;
+	bool			stats;
  	int			graph_depth;
  	int			func_stack_trace;
  	int			func_irq_info;

tools/perf/util/ftrace.h int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace); · line 51

Quoted 5 lines
tools/perf/util/ftrace.h +1 -1
@@ -46,7 +48,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_start_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
-				  int buckets[]);
+				  int buckets[], struct stats *stats);

Hi Namhyung,

There's a stub below for !HAVE_BPF_SKEL that needs updating too.

Also a minor point, but it seems quite low impact so it could be always on and then get rid of the --stats option.

Other than that:

Quoted 3 lines
tools/perf/util/ftrace.h +0 -0
  int perf_ftrace__latency_cleanup_bpf(struct perf_ftrace *ftrace);
 
  #else  /* !HAVE_BPF_SKEL */
author maintainer
#6 lore raw

Hello,

Quoted 236 lines · James Clark
Quoted 225 lines · Namhyung Kim

Sometimes users also want to see average latency as well as histogram. The --stats option is to display latency statistics at the end.

   $ sudo ./perf ftrace latency -ab --stats -T synchronize_rcu -- ...
   #   DURATION     |      COUNT | GRAPH                                          |
        0 - 1    us |          0 |                                                |
        1 - 2    us |          0 |                                                |
        2 - 4    us |          0 |                                                |
        4 - 8    us |          0 |                                                |
        8 - 16   us |          0 |                                                |
       16 - 32   us |          0 |                                                |
       32 - 64   us |          0 |                                                |
       64 - 128  us |          0 |                                                |
      128 - 256  us |          0 |                                                |
      256 - 512  us |          0 |                                                |
      512 - 1024 us |          0 |                                                |
        1 - 2    ms |          0 |                                                |
        2 - 4    ms |          0 |                                                |
        4 - 8    ms |          0 |                                                |
        8 - 16   ms |          1 | #####                                          |
       16 - 32   ms |          7 | ########################################       |
       32 - 64   ms |          0 |                                                |
       64 - 128  ms |          0 |                                                |
      128 - 256  ms |          0 |                                                |
      256 - 512  ms |          0 |                                                |
      512 - 1024 ms |          0 |                                                |
        1 - ...   s |          0 |                                                |
   # statistics  (in usec)
     total time:               171832
       avg time:                21479
       max time:                30906
       min time:                15869
          count:                    8
---
  tools/perf/Documentation/perf-ftrace.txt    |  4 ++++
  tools/perf/builtin-ftrace.c                 | 24 ++++++++++++++++++++-
  tools/perf/util/bpf_ftrace.c                | 16 ++++++++++++--
  tools/perf/util/bpf_skel/func_latency.bpf.c | 19 ++++++++++++++++
  tools/perf/util/ftrace.h                    |  4 +++-
  5 files changed, 63 insertions(+), 4 deletions(-)
tools/perf/Documentation/perf-ftrace.txt +4 -0
index eccc0483f7faecad..16b067eb5d84746d 100644
@@ -160,6 +160,10 @@ OPTIONS for 'perf ftrace latency'
  	-n/--use-nsec). The setting is ignored if the value results in more than
  	22 buckets.
+--stats::
+	Display latency statistics at the end.
+
+
  OPTIONS for 'perf ftrace profile'
  ---------------------------------
tools/perf/builtin-ftrace.c +23 -1
index 834074162a63bf2b..0ac2b20b272bd790 100644
@@ -43,6 +43,8 @@
  static volatile sig_atomic_t workload_exec_errno;
  static volatile sig_atomic_t done;
+static struct stats latency_stats;  /* for tracepoints */
+
  static void sig_handler(int sig __maybe_unused)
  {
  	done = true;
@@ -801,6 +803,13 @@ static void make_histogram(struct perf_ftrace *ftrace, int buckets[],
  do_inc:
  		buckets[i]++;
+		if (ftrace->stats) {
+			if (num >= min_latency)
+				num += min_latency;
+
+			update_stats(&latency_stats, num);
+		}
+
  next:
  		/* empty the line buffer for the next output  */
  		linebuf[0] = '\0';
@@ -894,6 +903,14 @@ static void display_histogram(struct perf_ftrace *ftrace, int buckets[])
  	printf(" | %10d | %.*s%*s |\n", buckets[NUM_BUCKET - 1],
  	       bar_len, bar, bar_total - bar_len, "");
+	if (ftrace->stats) {
+		printf("\n# statistics  (in %s)\n", ftrace->use_nsec ? "nsec" : "usec");
+		printf("  total time: %20.0f\n", latency_stats.mean * latency_stats.n);
+		printf("    avg time: %20.0f\n", latency_stats.mean);
+		printf("    max time: %20"PRIu64"\n", latency_stats.max);
+		printf("    min time: %20"PRIu64"\n", latency_stats.min);
+		printf("       count: %20.0f\n", latency_stats.n);
+	}
  }
  static int prepare_func_latency(struct perf_ftrace *ftrace)
@@ -932,6 +949,9 @@ static int prepare_func_latency(struct perf_ftrace *ftrace)
  	if (fd < 0)
  		pr_err("failed to open trace_pipe\n");
+	if (ftrace->stats)
+		init_stats(&latency_stats);
+
  	put_tracing_file(trace_file);
  	return fd;
  }
@@ -961,7 +981,7 @@ static int stop_func_latency(struct perf_ftrace *ftrace)
  static int read_func_latency(struct perf_ftrace *ftrace, int buckets[])
  {
  	if (ftrace->target.use_bpf)
-		return perf_ftrace__latency_read_bpf(ftrace, buckets);
+		return perf_ftrace__latency_read_bpf(ftrace, buckets, &latency_stats);
  	return 0;
  }
@@ -1620,6 +1640,8 @@ int cmd_ftrace(int argc, const char **argv)
  		    "Minimum latency (1st bucket). Works only with --bucket-range."),
  	OPT_UINTEGER(0, "max-latency", &ftrace.max_latency,
  		    "Maximum latency (last bucket). Works only with --bucket-range and total buckets less than 22."),
+	OPT_BOOLEAN(0, "stats", &ftrace.stats,
+		    "Collect function latency stats too"),
  	OPT_PARENT(common_options),
  	};
  	const struct option profile_options[] = {
tools/perf/util/bpf_ftrace.c +14 -2
index bc484e65fb8f69ca..ecd5e16df9730e0d 100644
@@ -11,6 +11,7 @@
  #include "util/debug.h"
  #include "util/evlist.h"
  #include "util/bpf_counter.h"
+#include "util/stat.h"
  #include "util/bpf_skel/func_latency.skel.h"
@@ -53,6 +54,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
  	}
  	skel->rodata->use_nsec = ftrace->use_nsec;
+	skel->rodata->stats = ftrace->stats;
  	set_max_rlimit();
@@ -86,6 +88,9 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace)
  		}
  	}
+	if (ftrace->stats)
+		skel->bss->min = INT64_MAX;
+
  	skel->links.func_begin = bpf_program__attach_kprobe(skel->progs.func_begin,
  							    false, func->name);
  	if (IS_ERR(skel->links.func_begin)) {
@@ -121,8 +126,8 @@ int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused)
  	return 0;
  }
-int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
-				  int buckets[])
+int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
+				  int buckets[], struct stats *stats)
  {
  	int i, fd, err;
  	u32 idx;
@@ -146,6 +151,13 @@ int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused,
  			buckets[idx] += hist[i];
  	}
+	if (ftrace->stats && skel->bss->count) {
+		stats->mean = skel->bss->total / skel->bss->count;
+		stats->n = skel->bss->count;
+		stats->max = skel->bss->max;
+		stats->min = skel->bss->min;
+	}
+
  	free(hist);
  	return 0;
  }
tools/perf/util/bpf_skel/func_latency.bpf.c +19 -0
index 4df54e1b1411b10a..9b68eea8fd444699 100644
@@ -38,12 +38,19 @@ struct {
  int enabled = 0;
+// stats
+__s64 total;
+__s64 count;
+__s64 max;
+__s64 min;
+
  const volatile int has_cpu = 0;
  const volatile int has_task = 0;
  const volatile int use_nsec = 0;
  const volatile unsigned int bucket_range;
  const volatile unsigned int min_latency;
  const volatile unsigned int max_latency;
+const volatile int stats = 0;
  SEC("kprobe/func")
  int BPF_PROG(func_begin)
@@ -136,6 +143,18 @@ int BPF_PROG(func_end)
  			return 0;
  		*hist += 1;
+
+		if (stats) {
+			delta /= cmp_base;
+
+			__sync_fetch_and_add(&total, delta);
+			__sync_fetch_and_add(&count, 1);
+
+			if (delta > max)
+				max = delta;
+			if (delta < min)
+				min = delta;
+		}
  	}
  	return 0;
tools/perf/util/ftrace.h +2 -0
index f218703063f74786..76038ed712085fbc 100644
@@ -7,6 +7,7 @@
  struct evlist;
  struct hashamp;
+struct stats;
  struct perf_ftrace {
  	struct evlist		*evlist;
@@ -23,6 +24,7 @@ struct perf_ftrace {
  	unsigned int		bucket_range;
  	unsigned int		min_latency;
  	unsigned int		max_latency;
+	bool			stats;
  	int			graph_depth;
  	int			func_stack_trace;
  	int			func_irq_info;

tools/perf/util/ftrace.h int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace); · line 51

Quoted 5 lines
tools/perf/util/ftrace.h +1 -1
@@ -46,7 +48,7 @@ int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_start_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace);
  int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace,
-				  int buckets[]);
+				  int buckets[], struct stats *stats);

Hi Namhyung,

Quoted 1 line

There's a stub below for !HAVE_BPF_SKEL that needs updating too.

Oh, right. I'll add that too.

Quoted 3 lines

Also a minor point, but it seems quite low impact so it could be always on and then get rid of the --stats option.

Yeah, probably it's better to add it unconditionally.

Quoted 4 lines

Other than that:

Thanks for your review!
Namhyung

Quoted 4 lines
Quoted 3 lines
tools/perf/util/ftrace.h +0 -0
  int perf_ftrace__latency_cleanup_bpf(struct perf_ftrace *ftrace);
  #else  /* !HAVE_BPF_SKEL */