perf stat: Add scripting support

hi,
sending another version of stat scripting.

v8 changes:
- check for stat callbacks properly [Namhyung]
- used '#!/usr/bin/env python' for stat-cpi.py [Namhyung]
- used tuple_set_u64 for storing u64 into python tuple [Namhyung]

v7 changes:
- perf stat record/report patches already taken, posting the rest of the scripting support
- rebased to latest Arnaldo's perf/core

v6 changes:
- several patches from v4 already taken
- perf stat record can now place 'record' keyword anywhere within stat options
- placed STAT feature checking earlier into record patches so commands processing perf.data recognize stat data and skip sample_type checking
- rebased on Arnaldo's perf/stat
- added Tested-by: Kan Liang <[email protected]>

v5 changes:
- several patches from v4 already taken
- using u16 for cpu number in cpu_map_event
- renamed PERF_RECORD_HEADER_ATTR_UPDATE to PERF_RECORD_EVENT_UPDATE
- moved low hanging fuits patches to the start of the patchset
- patchset tested by Kan Liang, thanks!

v4 changes:
- added attr update event for event's cpumask
- forbig aggregation on task workloads
- some minor reorders and changelog fixes

v3 changes:
- added attr update event to handle unit,scale,name for event it fixed the uncore_imc_1/cas_count_read/ record/report
- perf report -D now displays stat related events
- some minor and changelog fixes

v2 changes:
- rebased to latest Arnaldo's perf/core
- patches 1 to 11 already merged in
- added --per-core/--per-socket/-A options for perf stat report command to allow custom aggregation in stat report, please check new examples below
- couple changelogs changes

The initial attempt defined its own formula lang and allowed triggering user's script on the end of the stat command:
http://marc.info/?l=linux-kernel&m=136742146322273&w=2

This patchset abandons the idea of new formula language and rather adds support to:
- store stat data into perf.data file
- add python support to process stat events

Basically it allows to store stat data into perf.data and post process it with python scripts in a similar way we do for sampling data.

The stat data are stored in new stat, stat-round, stat-config user events.
stat - stored for each read syscall of the counter
stat round - stored for each interval or end of the command invocation stat config - stores all the config information needed to process data
so report tool could restore the same output as record

The python script can now define 'stat__<eventname>_<modifier>' functions to get stat events data and 'stat__interval' to get stat-round data.

See CPI script example in scripts/python/stat-cpi.py.

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

thanks,
jirka

Examples:

- To record data for command stat workload:

  $ perf stat record kill
  ...

Performance counter stats for 'kill':

            0.372007      task-clock (msec)         #    0.613 CPUs utilized          
                   3      context-switches          #    0.008 M/sec                  
                   0      cpu-migrations            #    0.000 K/sec                  
                  62      page-faults               #    0.167 M/sec                  
           1,129,973      cycles                    #    3.038 GHz                    
     <not supported>      stalled-cycles-frontend  
     <not supported>      stalled-cycles-backend   
             813,313      instructions              #    0.72  insns per cycle        
             166,161      branches                  #  446.661 M/sec                  
               8,747      branch-misses             #    5.26% of all branches        

         0.000607287 seconds time elapsed

- To report perf stat data:

  $ perf stat report

Performance counter stats for '/home/jolsa/bin/perf stat record kill':

            0.372007      task-clock (msec)         #      inf CPUs utilized          
                   3      context-switches          #    0.008 M/sec                  
                   0      cpu-migrations            #    0.000 K/sec                  
                  62      page-faults               #    0.167 M/sec                  
           1,129,973      cycles                    #    3.038 GHz                    
     <not supported>      stalled-cycles-frontend  
     <not supported>      stalled-cycles-backend   
             813,313      instructions              #    0.72  insns per cycle        
             166,161      branches                  #  446.661 M/sec                  
               8,747      branch-misses             #    5.26% of all branches        

         0.000000000 seconds time elapsed

- To store system-wide period stat data:

  $ perf stat -e cycles:u,instructions:u -a -I 1000 record
  #           time             counts unit events
       1.000265471        462,311,482      cycles:u                   (100.00%)
       1.000265471        590,037,440      instructions:u           
       2.000483453        722,532,336      cycles:u                   (100.00%)
       2.000483453        848,678,197      instructions:u           
       3.000759876         75,990,880      cycles:u                   (100.00%)
       3.000759876         86,187,813      instructions:u           
  ^C     3.213960893         85,329,533      cycles:u                   (100.00%)
       3.213960893        135,954,296      instructions:u           

- To report perf stat data:

  $ perf stat report
  #           time             counts unit events
       1.000265471        462,311,482      cycles:u                   (100.00%)
       1.000265471        590,037,440      instructions:u           
       2.000483453        722,532,336      cycles:u                   (100.00%)
       2.000483453        848,678,197      instructions:u           
       3.000759876         75,990,880      cycles:u                   (100.00%)
       3.000759876         86,187,813      instructions:u           
       3.213960893         85,329,533      cycles:u                   (100.00%)
       3.213960893        135,954,296      instructions:u           

- To run stat-cpi.py script over perf.data:

  $ perf script -s scripts/python/stat-cpi.py 
         1.000265: cpu -1, thread -1 -> cpi 0.783529 (462311482/590037440)
         2.000483: cpu -1, thread -1 -> cpi 0.851362 (722532336/848678197)
         3.000760: cpu -1, thread -1 -> cpi 0.881689 (75990880/86187813)
         3.213961: cpu -1, thread -1 -> cpi 0.627634 (85329533/135954296)

- To pipe data from stat to stat-cpi script:

  $ perf stat -e cycles:u,instructions:u -A -C 0 -I 1000 record | perf script -s scripts/python/stat-cpi.py 
         1.000192: cpu 0, thread -1 -> cpi 0.739535 (23921908/32347236)
         2.000376: cpu 0, thread -1 -> cpi 1.663482 (2519340/1514498)
         3.000621: cpu 0, thread -1 -> cpi 1.396308 (16162767/11575362)
         4.000700: cpu 0, thread -1 -> cpi 1.092246 (20077258/18381624)
         5.000867: cpu 0, thread -1 -> cpi 0.473816 (45157586/95306156)
         6.001034: cpu 0, thread -1 -> cpi 0.532792 (43701668/82023818)
         7.001195: cpu 0, thread -1 -> cpi 1.122059 (29890042/26638561)

- Raw script stat data output:

  $ perf stat -e cycles:u,instructions:u -A -C 0 -I 1000 record | perf --no-pager script
  CPU   THREAD             VAL             ENA             RUN            TIME EVENT
    0       -1        12302059      1000811347      1000810712      1000198821 cycles:u
    0       -1         2565362      1000823218      1000823218      1000198821 instructions:u
    0       -1        14453353      1000812704      1000812704      2000382283 cycles:u
    0       -1         4600932      1000799342      1000799342      2000382283 instructions:u
    0       -1        15245106      1000774425      1000774425      3000538255 cycles:u
    0       -1         2624324      1000769310      1000769310      3000538255 instructions:u

- To display different aggregation in report:

  $ perf stat -e cycles -a -I 1000 record sleep 3 
  #           time             counts unit events
       1.000223609        703,427,617      cycles                   
       2.000443651        609,975,307      cycles                   
       3.000569616        668,479,597      cycles                   
       3.000735323          1,155,816      cycles                 

  $ perf stat report
  #           time             counts unit events
       1.000223609        703,427,617      cycles                   
       2.000443651        609,975,307      cycles                   
       3.000569616        668,479,597      cycles                   
       3.000735323          1,155,816      cycles                   

  $ perf stat report --per-core
  #           time core         cpus             counts unit events
       1.000223609 S0-C0           2        327,612,412      cycles                   
       1.000223609 S0-C1           2        375,815,205      cycles                   
       2.000443651 S0-C0           2        287,462,177      cycles                   
       2.000443651 S0-C1           2        322,513,130      cycles                   
       3.000569616 S0-C0           2        271,571,908      cycles                   
       3.000569616 S0-C1           2        396,907,689      cycles                   
       3.000735323 S0-C0           2            694,977      cycles                   
       3.000735323 S0-C1           2            460,839      cycles                   

  $ perf stat report --per-socket
  #           time socket cpus             counts unit events
       1.000223609 S0        4        703,427,617      cycles                   
       2.000443651 S0        4        609,975,307      cycles                   
       3.000569616 S0        4        668,479,597      cycles                   
       3.000735323 S0        4          1,155,816      cycles                   

  $ perf stat report -A
  #           time CPU                counts unit events
       1.000223609 CPU0           205,431,505      cycles                   
       1.000223609 CPU1           122,180,907      cycles                   
       1.000223609 CPU2           176,649,682      cycles                   
       1.000223609 CPU3           199,165,523      cycles                   
       2.000443651 CPU0           148,447,922      cycles                   
       2.000443651 CPU1           139,014,255      cycles                   
       2.000443651 CPU2           204,436,559      cycles                   
       2.000443651 CPU3           118,076,571      cycles                   
       3.000569616 CPU0           149,788,954      cycles                   
       3.000569616 CPU1           121,782,954      cycles                   
       3.000569616 CPU2           247,277,700      cycles                   
       3.000569616 CPU3           149,629,989      cycles                   
       3.000735323 CPU0               269,675      cycles                   
       3.000735323 CPU1               425,302      cycles                   
       3.000735323 CPU2               364,169      cycles                   
       3.000735323 CPU3                96,670      cycles                   

---

For pipe sessions we need to keep sample_type zero, because script's perf_evsel__check_attr is triggered by sample_type != 0, and the check would fail on stat session.

I was tempted to keep it zero unconditionally, but the pipe session is sufficient. In perf.data session we are guarded by HEADER_STAT feature.

---
 tools/perf/builtin-stat.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

index 9805e03ab163..7f568244662b 100644
@@ -184,11 +184,18 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
 	 * like tracepoints. Clear it up for counting.
 	 */
 	attr->sample_period = 0;
+
 	/*
 	 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 	 * while avoiding that older tools show confusing messages.
+	 *
+	 * However for pipe sessions we need to keep it zero,
+	 * because script's perf_evsel__check_attr is triggered
+	 * by attr->sample_type != 0, and we can't run it on
+	 * stat sessions.
 	 */
-	attr->sample_type   = PERF_SAMPLE_IDENTIFIER;
+	if (!(STAT_RECORD && perf_stat.file.is_pipe))
+		attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 
 	/*
 	 * Disabling all counters initially, they will be enabled

2.4.3

Adding processing of cpu/threads maps. Configuring session's evlist with these maps.

---
 tools/perf/builtin-script.c | 67 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 67 insertions(+)

index bcc3542d9df5..aa6d7cf87dab 100644
@@ -18,7 +18,11 @@
 #include "util/sort.h"
 #include "util/data.h"
 #include "util/auxtrace.h"
+#include "util/cpumap.h"
+#include "util/thread_map.h"
+#include "util/stat.h"
 #include <linux/bitmap.h>
+#include "asm/bug.h"
 
 static char const		*script_name;
 static char const		*generate_script_lang;

@@ -606,6 +610,9 @@ struct perf_script {
 	bool			show_task_events;
 	bool			show_mmap_events;
 	bool			show_switch_events;
+	bool			allocated;
+	struct cpu_map		*cpus;
+	struct thread_map	*threads;
 };
 
 static void process_event(struct perf_script *script __maybe_unused, union perf_event *event,

@@ -1682,6 +1689,63 @@ static void script__setup_sample_type(struct perf_script *script)
 	}
 }
 
+static int set_maps(struct perf_script *script)
+{
+	struct perf_evlist *evlist = script->session->evlist;
+
+	if (!script->cpus || !script->threads)
+		return 0;
+
+	if (WARN_ONCE(script->allocated, "stats double allocation\n"))
+		return -EINVAL;
+
+	perf_evlist__set_maps(evlist, script->cpus, script->threads);
+
+	if (perf_evlist__alloc_stats(evlist, true))
+		return -ENOMEM;
+
+	script->allocated = true;
+	return 0;
+}
+
+static
+int process_thread_map_event(struct perf_tool *tool,
+			     union perf_event *event,
+			     struct perf_session *session __maybe_unused)
+{
+	struct perf_script *script = container_of(tool, struct perf_script, tool);
+
+	if (script->threads) {
+		pr_warning("Extra thread map event, ignoring.\n");
+		return 0;
+	}
+
+	script->threads = thread_map__new_event(&event->thread_map);
+	if (!script->threads)
+		return -ENOMEM;
+
+	return set_maps(script);
+}
+
+static
+int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
+			  union perf_event *event,
+			  struct perf_session *session __maybe_unused)
+{
+	struct perf_script *script = container_of(tool, struct perf_script, tool);
+
+	if (script->cpus) {
+		pr_warning("Extra cpu map event, ignoring.\n");
+		return 0;
+	}
+
+	script->cpus = cpu_map__new_data(&event->cpu_map.data);
+	if (!script->cpus)
+		return -ENOMEM;
+
+	return set_maps(script);
+}
+
 int cmd_script(int argc, const char **argv, const char *prefix __maybe_unused)
 {
 	bool show_full_info = false;

@@ -1710,6 +1774,8 @@ int cmd_script(int argc, const char **argv, const char *prefix __maybe_unused)
 			.auxtrace_info	 = perf_event__process_auxtrace_info,
 			.auxtrace	 = perf_event__process_auxtrace,
 			.auxtrace_error	 = perf_event__process_auxtrace_error,
+			.thread_map	 = process_thread_map_event,
+			.cpu_map	 = process_cpu_map_event,
 			.ordered_events	 = true,
 			.ordering_requires_timestamps = true,
 		},

@@ -2063,6 +2129,7 @@ int cmd_script(int argc, const char **argv, const char *prefix __maybe_unused)
 	flush_scripting();
 
 out_delete:
+	perf_evlist__free_stats(session->evlist);
 	perf_session__delete(session);
 
 	if (script_started)

2.4.3

Adding processing of stat config event and initialize stat_config object.

---
 tools/perf/builtin-script.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

index aa6d7cf87dab..a90bc0b81e70 100644
@@ -36,6 +36,7 @@ static bool			print_flags;
 static bool			nanosecs;
 static const char		*cpu_list;
 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
+static struct perf_stat_config	stat_config;
 
 unsigned int scripting_max_stack = PERF_MAX_STACK_DEPTH;
 

@@ -1689,6 +1690,14 @@ static void script__setup_sample_type(struct perf_script *script)
 	}
 }
 
+static int process_stat_config_event(struct perf_tool *tool __maybe_unused,
+				     union perf_event *event,
+				     struct perf_session *session __maybe_unused)
+{
+	perf_event__read_stat_config(&stat_config, &event->stat_config);
+	return 0;
+}
+
 static int set_maps(struct perf_script *script)
 {
 	struct perf_evlist *evlist = script->session->evlist;

@@ -1774,6 +1783,7 @@ int cmd_script(int argc, const char **argv, const char *prefix __maybe_unused)
 			.auxtrace_info	 = perf_event__process_auxtrace_info,
 			.auxtrace	 = perf_event__process_auxtrace,
 			.auxtrace_error	 = perf_event__process_auxtrace_error,
+			.stat_config	 = process_stat_config_event,
 			.thread_map	 = process_thread_map_event,
 			.cpu_map	 = process_cpu_map_event,
 			.ordered_events	 = true,

2.4.3

Python and perl scripting code will define those
callbacks and get stat data.

---
 tools/perf/util/trace-event.h | 4 ++++
 1 file changed, 4 insertions(+)

index b85ee55cca0c..0ebc9dab2c7c 100644
@@ -65,6 +65,7 @@ int tracing_data_put(struct tracing_data *tdata);
 struct addr_location;
 
 struct perf_session;
+struct perf_stat_config;
 
 struct scripting_ops {
 	const char *name;

@@ -75,6 +76,9 @@ struct scripting_ops {
 			       struct perf_sample *sample,
 			       struct perf_evsel *evsel,
 			       struct addr_location *al);
+	void (*process_stat) (struct perf_stat_config *config,
+			      struct perf_evsel *evsel, u64 time);
+	void (*process_stat_interval) (u64 time);
 	int (*generate_script) (struct pevent *pevent, const char *outfile);
 };
 

2.4.3

Implement struct scripting_ops::(process_stat|process_stat_interval) handlers - calling scripting handlers from stat events handlers.

---
 tools/perf/builtin-script.c | 33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

index a90bc0b81e70..521e3e0e7625 100644
@@ -221,6 +221,9 @@ static int perf_evsel__check_attr(struct perf_evsel *evsel,
 	struct perf_event_attr *attr = &evsel->attr;
 	bool allow_user_set;
 
+	if (perf_header__has_feat(&session->header, HEADER_STAT))
+		return 0;
+
 	allow_user_set = perf_header__has_feat(&session->header,
 					       HEADER_AUXTRACE);
 

@@ -674,6 +677,18 @@ static void process_event(struct perf_script *script __maybe_unused, union perf_
 
 static struct scripting_ops	*scripting_ops;
 
+static void process_stat(struct perf_evsel *counter, u64 time)
+{
+	if (scripting_ops && scripting_ops->process_stat)
+		scripting_ops->process_stat(&stat_config, counter, time);
+}
+
+static void process_stat_interval(u64 time)
+{
+	if (scripting_ops && scripting_ops->process_stat_interval)
+		scripting_ops->process_stat_interval(time);
+}
+
 static void setup_scripting(void)
 {
 	setup_perl_scripting();

@@ -1690,6 +1705,22 @@ static void script__setup_sample_type(struct perf_script *script)
 	}
 }
 
+static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
+				    union perf_event *event,
+				    struct perf_session *session)
+{
+	struct stat_round_event *round = &event->stat_round;
+	struct perf_evsel *counter;
+
+	evlist__for_each(session->evlist, counter) {
+		perf_stat_process_counter(&stat_config, counter);
+		process_stat(counter, round->time);
+	}
+
+	process_stat_interval(round->time);
+	return 0;
+}
+
 static int process_stat_config_event(struct perf_tool *tool __maybe_unused,
 				     union perf_event *event,
 				     struct perf_session *session __maybe_unused)

@@ -1783,6 +1814,8 @@ int cmd_script(int argc, const char **argv, const char *prefix __maybe_unused)
 			.auxtrace_info	 = perf_event__process_auxtrace_info,
 			.auxtrace	 = perf_event__process_auxtrace,
 			.auxtrace_error	 = perf_event__process_auxtrace_error,
+			.stat		 = perf_event__process_stat_event,
+			.stat_round	 = process_stat_round_event,
 			.stat_config	 = process_stat_config_event,
 			.thread_map	 = process_thread_map_event,
 			.cpu_map	 = process_cpu_map_event,

2.4.3

If no script is specified for stat data, display
stat events in raw form.

  $ perf stat record ls

SNIP

Performance counter stats for 'ls':

            0.851585      task-clock (msec)         #    0.717 CPUs utilized
                   0      context-switches          #    0.000 K/sec
                   0      cpu-migrations            #    0.000 K/sec
                 114      page-faults               #    0.134 M/sec
           2,620,918      cycles                    #    3.078 GHz
     <not supported>      stalled-cycles-frontend
     <not supported>      stalled-cycles-backend
           2,714,111      instructions              #    1.04  insns per cycle
             542,434      branches                  #  636.970 M/sec
              15,946      branch-misses             #    2.94% of all branches

         0.001186954 seconds time elapsed

  $ perf script
  CPU   THREAD             VAL             ENA             RUN            TIME EVENT
   -1    26185          851585          851585          851585         1186954 task-clock
   -1    26185               0          851585          851585         1186954 context-switches
   -1    26185               0          851585          851585         1186954 cpu-migrations
   -1    26185             114          851585          851585         1186954 page-faults
   -1    26185         2620918          853340          853340         1186954 cycles
   -1    26185               0               0               0         1186954 stalled-cycles-frontend
   -1    26185               0               0               0         1186954 stalled-cycles-backend
   -1    26185         2714111          853340          853340         1186954 instructions
   -1    26185          542434          853340          853340         1186954 branches
   -1    26185           15946          853340          853340         1186954 branch-misses

---
 tools/perf/builtin-script.c | 36 ++++++++++++++++++++++++++++++++++++
 1 file changed, 36 insertions(+)

index 521e3e0e7625..c298cdc618e2 100644
@@ -677,10 +677,46 @@ static void process_event(struct perf_script *script __maybe_unused, union perf_
 
 static struct scripting_ops	*scripting_ops;
 
+static void __process_stat(struct perf_evsel *counter, u64 time)
+{
+	int nthreads = thread_map__nr(counter->threads);
+	int ncpus = perf_evsel__nr_cpus(counter);
+	int cpu, thread;
+	static int header_printed;
+
+	if (counter->system_wide)
+		nthreads = 1;
+
+	if (!header_printed) {
+		printf("%3s %8s %15s %15s %15s %15s %s\n",
+		       "CPU", "THREAD", "VAL", "ENA", "RUN", "TIME", "EVENT");
+		header_printed = 1;
+	}
+
+	for (thread = 0; thread < nthreads; thread++) {
+		for (cpu = 0; cpu < ncpus; cpu++) {
+			struct perf_counts_values *counts;
+
+			counts = perf_counts(counter->counts, cpu, thread);
+
+			printf("%3d %8d %15" PRIu64 " %15" PRIu64 " %15" PRIu64 " %15" PRIu64 " %s\n",
+				counter->cpus->map[cpu],
+				thread_map__pid(counter->threads, thread),
+				counts->val,
+				counts->ena,
+				counts->run,
+				time,
+				perf_evsel__name(counter));
+		}
+	}
+}
+
 static void process_stat(struct perf_evsel *counter, u64 time)
 {
 	if (scripting_ops && scripting_ops->process_stat)
 		scripting_ops->process_stat(&stat_config, counter, time);
+	else
+		__process_stat(counter, time);
 }
 
 static void process_stat_interval(u64 time)

2.4.3

Em Tue, Jan 05, 2016 at 10:09:10PM +0100, Jiri Olsa escreveu:

The CPU field is getting misformatted:

[acme@zoo linux]$
[acme@zoo linux]$ perf stat record usleep 1

Performance counter stats for 'usleep 1':

          0.356474      task-clock (msec)         #    0.516 CPUs utilized          
                 1      context-switches          #    0.003 M/sec                  
                 0      cpu-migrations            #    0.000 K/sec                  
                54      page-faults               #    0.151 M/sec                  
         1,044,123      cycles                    #    2.929 GHz                    
           716,911      stalled-cycles-frontend   #   68.66% frontend cycles idle   
   <not supported>      stalled-cycles-backend   
           694,609      instructions              #    0.67  insns per cycle        

# 1.03 stalled cycles per insn
140,037 branches # 392.839 M/sec
7,920 branch-misses # 5.66% of all branches

0.000690200 seconds time elapsed

[acme@zoo linux]$
[acme@zoo linux]$ perf script

CPU   THREAD             VAL             ENA             RUN            TIME EVENT
65535     4883          356474          356474          356474          690200 task-clock
65535     4883               1          356474          356474          690200 context-switches
65535     4883               0          356474          356474          690200 cpu-migrations
65535     4883              54          356474          356474          690200 page-faults
65535     4883         1044123          360329          360329          690200 cycles
65535     4883          716911          360329          360329          690200 stalled-cycles-frontend
65535     4883               0               0               0          690200 stalled-cycles-backend
65535     4883          694609          360329          360329          690200 instructions
65535     4883          140037          360329          360329          690200 branches
65535     4883            7920          360329          360329          690200 branch-misses

[acme@zoo linux]$

On Tue, Jan 05, 2016 at 07:49:27PM -0300, Arnaldo Carvalho de Melo wrote:

SNIP

there's wrong conversion from u16 (-1) cpu data to int cpu_map->map[0]

attached patch fixes that for me, not sure why I did not see that before.. I'll send new version shortly

thanks,
jirka

---

index a0717b93d8f5..581aec244ac1 100644
@@ -189,7 +189,7 @@ static struct cpu_map *cpu_map__from_entries(struct cpu_map_entries *cpus)
 		unsigned i;
 
 		for (i = 0; i < cpus->nr; i++)
-			map->map[i] = (int)cpus->cpu[i];
+			map->map[i] = (int)(s16)cpus->cpu[i];
 	}
 
 	return map;

Add support to get stat events data in perf python scripts.

The python script shall implement following
new interface to process stat data:

def stat__<event_name>_[<modifier>](cpu, thread, time, val, ena, run):

- is called for every stat event for given counter, if user monitors 'cycles,instructions:u" following callbacks should be defined:

def stat__cycles(cpu, thread, time, val, ena, run):
def stat__instructions_u(cpu, thread, time, val, ena, run):

def stat__interval(time):

- is called for every interval with its time,
in non interval mode it's called after last
stat event with total measured time in ns

The rest of the current interface stays untouched..

Please check example CPI metrics script in following patch with command line examples in changelogs.

---
 .../util/scripting-engines/trace-event-python.c    | 115 +++++++++++++++++++--
 1 file changed, 109 insertions(+), 6 deletions(-)

index a8e825fca42a..5ebe8842557c 100644
@@ -41,6 +41,9 @@
 #include "../thread-stack.h"
 #include "../trace-event.h"
 #include "../machine.h"
+#include "thread_map.h"
+#include "cpumap.h"
+#include "stat.h"
 
 PyMODINIT_FUNC initperf_trace_context(void);
 

@@ -859,6 +862,104 @@ static void python_process_event(union perf_event *event,
 	}
 }
 
+static void get_handler_name(char *str, size_t size,
+			     struct perf_evsel *evsel)
+{
+	char *p = str;
+
+	scnprintf(str, size, "stat__%s", perf_evsel__name(evsel));
+
+	while ((p = strchr(p, ':'))) {
+		*p = '_';
+		p++;
+	}
+}
+
+static void
+process_stat(struct perf_evsel *counter, int cpu, int thread, u64 time,
+	     struct perf_counts_values *count)
+{
+	PyObject *handler, *t;
+	static char handler_name[256];
+	int n = 0;
+
+	t = PyTuple_New(MAX_FIELDS);
+	if (!t)
+		Py_FatalError("couldn't create Python tuple");
+
+	get_handler_name(handler_name, sizeof(handler_name),
+			 counter);
+
+	handler = get_handler(handler_name);
+	if (!handler) {
+		pr_debug("can't find python handler %s\n", handler_name);
+		return;
+	}
+
+	PyTuple_SetItem(t, n++, PyInt_FromLong(cpu));
+	PyTuple_SetItem(t, n++, PyInt_FromLong(thread));
+
+	tuple_set_u64(t, n++, time);
+	tuple_set_u64(t, n++, count->val);
+	tuple_set_u64(t, n++, count->ena);
+	tuple_set_u64(t, n++, count->run);
+
+	if (_PyTuple_Resize(&t, n) == -1)
+		Py_FatalError("error resizing Python tuple");
+
+	call_object(handler, t, handler_name);
+
+	Py_DECREF(t);
+}
+
+static void python_process_stat(struct perf_stat_config *config,
+				struct perf_evsel *counter, u64 time)
+{
+	struct thread_map *threads = counter->threads;
+	struct cpu_map *cpus = counter->cpus;
+	int cpu, thread;
+
+	if (config->aggr_mode == AGGR_GLOBAL) {
+		process_stat(counter, -1, -1, time,
+			     &counter->counts->aggr);
+		return;
+	}
+
+	for (thread = 0; thread < threads->nr; thread++) {
+		for (cpu = 0; cpu < cpus->nr; cpu++) {
+			process_stat(counter, cpus->map[cpu],
+				     thread_map__pid(threads, thread), time,
+				     perf_counts(counter->counts, cpu, thread));
+		}
+	}
+}
+
+static void python_process_stat_interval(u64 time)
+{
+	PyObject *handler, *t;
+	static const char handler_name[] = "stat__interval";
+	int n = 0;
+
+	t = PyTuple_New(MAX_FIELDS);
+	if (!t)
+		Py_FatalError("couldn't create Python tuple");
+
+	handler = get_handler(handler_name);
+	if (!handler) {
+		pr_debug("can't find python handler %s\n", handler_name);
+		return;
+	}
+
+	tuple_set_u64(t, n++, time);
+
+	if (_PyTuple_Resize(&t, n) == -1)
+		Py_FatalError("error resizing Python tuple");
+
+	call_object(handler, t, handler_name);
+
+	Py_DECREF(t);
+}
+
 static int run_start_sub(void)
 {
 	main_module = PyImport_AddModule("__main__");

@@ -1201,10 +1302,12 @@ static int python_generate_script(struct pevent *pevent, const char *outfile)
 }
 
 struct scripting_ops python_scripting_ops = {
-	.name = "Python",
-	.start_script = python_start_script,
-	.flush_script = python_flush_script,
-	.stop_script = python_stop_script,
-	.process_event = python_process_event,
-	.generate_script = python_generate_script,
+	.name			= "Python",
+	.start_script		= python_start_script,
+	.flush_script		= python_flush_script,
+	.stop_script		= python_stop_script,
+	.process_event		= python_process_event,
+	.process_stat		= python_process_stat,
+	.process_stat_interval	= python_process_stat_interval,
+	.generate_script	= python_generate_script,
 };

2.4.3

Adding stat-cpi.py as an example of how to do stat scripting. It computes the CPI metrics from cycles and instructions events.

Following stat record/report/script combinations could be used:

- get CPI for given workload

    $ perf stat -e cycles,instructions record ls

SNIP

Performance counter stats for 'ls':

2,904,431 cycles
3,346,878 instructions # 1.15 insns per cycle

0.001782686 seconds time elapsed

$ perf script -s ./scripts/python/stat-cpi.py
0.001783: cpu -1, thread -1 -> cpi 0.867803 (2904431/3346878)

$ perf stat -e cycles,instructions record ls | perf script -s ./scripts/python/stat-cpi.py

SNIP

0.001730: cpu -1, thread -1 -> cpi 0.869026 (2928292/3369627)

- get CPI systemwide:

    $ perf stat -e cycles,instructions -a -I 1000 record sleep 3
    #           time             counts unit events
         1.000158618        594,274,711      cycles                     (100.00%)
         1.000158618        441,898,250      instructions
         2.000350973        567,649,705      cycles                     (100.00%)
         2.000350973        432,669,206      instructions
         3.000559210        561,940,430      cycles                     (100.00%)
         3.000559210        420,403,465      instructions
         3.000670798            780,105      cycles                     (100.00%)
         3.000670798            326,516      instructions

    $ perf script -s ./scripts/python/stat-cpi.py
           1.000159: cpu -1, thread -1 -> cpi 1.344823 (594274711/441898250)
           2.000351: cpu -1, thread -1 -> cpi 1.311972 (567649705/432669206)
           3.000559: cpu -1, thread -1 -> cpi 1.336669 (561940430/420403465)
           3.000671: cpu -1, thread -1 -> cpi 2.389178 (780105/326516)

    $ perf stat -e cycles,instructions -a -I 1000 record sleep 3 | perf script -s ./scripts/python/stat-cpi.py
           1.000202: cpu -1, thread -1 -> cpi 1.035091 (940778881/908885530)
           2.000392: cpu -1, thread -1 -> cpi 1.442600 (627493992/434974455)
           3.000545: cpu -1, thread -1 -> cpi 1.353612 (741463930/547766890)
           3.000622: cpu -1, thread -1 -> cpi 2.642110 (784083/296764)

---
 tools/perf/scripts/python/stat-cpi.py | 77 +++++++++++++++++++++++++++++++++++
 1 file changed, 77 insertions(+)
 create mode 100644 tools/perf/scripts/python/stat-cpi.py

new file mode 100644
index 000000000000..8b60f343dd07
@@ -0,0 +1,77 @@
+#!/usr/bin/env python
+
+data    = {}
+times   = []
+threads = []
+cpus    = []
+
+def get_key(time, event, cpu, thread):
+    return "%d-%s-%d-%d" % (time, event, cpu, thread)
+
+def store_key(time, cpu, thread):
+    if (time not in times):
+        times.append(time)
+
+    if (cpu not in cpus):
+        cpus.append(cpu)
+
+    if (thread not in threads):
+        threads.append(thread)
+
+def store(time, event, cpu, thread, val, ena, run):
+    #print "event %s cpu %d, thread %d, time %d, val %d, ena %d, run %d" % \
+    #      (event, cpu, thread, time, val, ena, run)
+
+    store_key(time, cpu, thread)
+    key = get_key(time, event, cpu, thread)
+    data[key] = [ val, ena, run]
+
+def get(time, event, cpu, thread):
+    key = get_key(time, event, cpu, thread)
+    return data[key][0]
+
+def stat__cycles_k(cpu, thread, time, val, ena, run):
+    store(time, "cycles", cpu, thread, val, ena, run);
+
+def stat__instructions_k(cpu, thread, time, val, ena, run):
+    store(time, "instructions", cpu, thread, val, ena, run);
+
+def stat__cycles_u(cpu, thread, time, val, ena, run):
+    store(time, "cycles", cpu, thread, val, ena, run);
+
+def stat__instructions_u(cpu, thread, time, val, ena, run):
+    store(time, "instructions", cpu, thread, val, ena, run);
+
+def stat__cycles(cpu, thread, time, val, ena, run):
+    store(time, "cycles", cpu, thread, val, ena, run);
+
+def stat__instructions(cpu, thread, time, val, ena, run):
+    store(time, "instructions", cpu, thread, val, ena, run);
+
+def stat__interval(time):
+    for cpu in cpus:
+        for thread in threads:
+            cyc = get(time, "cycles", cpu, thread)
+            ins = get(time, "instructions", cpu, thread)
+            cpi = 0
+
+            if ins != 0:
+                cpi = cyc/float(ins)
+
+            print "%15f: cpu %d, thread %d -> cpi %f (%d/%d)" % (time/(float(1000000000)), cpu, thread, cpi, cyc, ins)
+
+def trace_end():
+    pass
+# XXX trace_end callback could be used as an alternative place
+#     to compute same values as in the script above:
+#
+#    for time in times:
+#        for cpu in cpus:
+#            for thread in threads:
+#                cyc = get(time, "cycles", cpu, thread)
+#                ins = get(time, "instructions", cpu, thread)
+#
+#                if ins != 0:
+#                    cpi = cyc/float(ins)
+#
+#                print "time %.9f, cpu %d, thread %d -> cpi %f" % (time/(float(1000000000)), cpu, thread, cpi)

2.4.3

Em Tue, Jan 05, 2016 at 10:09:12PM +0100, Jiri Olsa escreveu:

Can you please state what CPI means and why someone would want to use this?

I applied the first 6 patches of this kit to my perf/core branch, will continue tomorrow when I'll probably try to push what is there to Ingo.

Thanks,

- Arnaldo

Em Tue, Jan 05, 2016 at 10:09:04PM +0100, Jiri Olsa escreveu:

While testing this I noticed I failed to notice that after adding all these new user meta events we have:

[acme@zoo linux]$ perf stat -e cycles record usleep 1

Performance counter stats for 'usleep 1':

         1,049,367      cycles                                                      

       0.002616357 seconds time elapsed

  [acme@zoo linux]$ oldperf report --stdio
  0xf0 [0x28]: failed to process type: 73
  Error:
  failed to process sample
  # To display the perf.data header info, please use
  # --header/--header-only options.
  #
  [acme@zoo linux]$ 

While it should just have ignored the user meta events if doesn't support and state that the perf.data file had no samples :-\

So I think I'll make perf_session__process_user_event() return 0 for events it doesn't support, just adding a pr_debug(), ok?

This will be left to after processing your current batch, just need to remember to do it before adding more user meta events.

- Arnaldo

ok, but please put there something more visible.. so it's obvious it was hit

jirka