blob: 688dea7cb08f7337bc6c620f98aa3e996ce4d581 [file] [log] [blame]
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001/*
Ingo Molnarbf9e1872009-06-02 23:37:05 +02002 * builtin-stat.c
3 *
4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
6 *
7 * Sample output:
Ingo Molnarddcacfa2009-04-20 15:37:32 +02008
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02009 $ perf stat ./hackbench 10
Ingo Molnarddcacfa2009-04-20 15:37:32 +020010
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020011 Time: 0.118
Ingo Molnarddcacfa2009-04-20 15:37:32 +020012
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020013 Performance counter stats for './hackbench 10':
Ingo Molnarddcacfa2009-04-20 15:37:32 +020014
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020015 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
Ingo Molnarddcacfa2009-04-20 15:37:32 +020028
Ingo Molnar52425192009-05-26 09:17:18 +020029 *
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020030 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
Ingo Molnar52425192009-05-26 09:17:18 +020031 *
32 * Improvements and fixes by:
33 *
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
Jaswinder Singh Rajput6e750a8f2009-06-27 03:02:07 +053039 * Jaswinder Singh Rajput <jaswinder@kernel.org>
Ingo Molnar52425192009-05-26 09:17:18 +020040 *
41 * Released under the GPL v2. (and only v2, not any later version)
Ingo Molnarddcacfa2009-04-20 15:37:32 +020042 */
43
Peter Zijlstra1a482f32009-05-23 18:28:58 +020044#include "perf.h"
Ingo Molnar16f762a2009-05-27 09:10:38 +020045#include "builtin.h"
Arnaldo Carvalho de Melof14d5702014-10-17 12:17:40 -030046#include "util/cgroup.h"
Ingo Molnar148be2c2009-04-27 08:02:14 +020047#include "util/util.h"
Josh Poimboeuf4b6ab942015-12-15 09:39:39 -060048#include <subcmd/parse-options.h>
Ingo Molnar52425192009-05-26 09:17:18 +020049#include "util/parse-events.h"
Andi Kleen4cabc3d2013-08-21 16:47:26 -070050#include "util/pmu.h"
Frederic Weisbecker8f288272009-08-16 22:05:48 +020051#include "util/event.h"
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -020052#include "util/evlist.h"
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -020053#include "util/evsel.h"
Frederic Weisbecker8f288272009-08-16 22:05:48 +020054#include "util/debug.h"
Mathieu Poirier5d8bb1e2016-09-16 09:50:03 -060055#include "util/drv_configs.h"
Ingo Molnara5d243d2011-04-27 05:39:24 +020056#include "util/color.h"
Xiao Guangrong0007ece2012-09-17 16:31:14 +080057#include "util/stat.h"
Liming Wang60666c62009-12-31 16:05:50 +080058#include "util/header.h"
Paul Mackerrasa12b51c2010-03-10 20:36:09 +110059#include "util/cpumap.h"
Zhang, Yanmind6d901c2010-03-18 11:36:05 -030060#include "util/thread.h"
Arnaldo Carvalho de Melofd782602011-01-18 15:15:24 -020061#include "util/thread_map.h"
Jiri Olsad8095602015-08-07 12:51:03 +020062#include "util/counts.h"
Andi Kleen44b1e602016-05-30 12:49:42 -030063#include "util/group.h"
Jiri Olsa4979d0c2015-11-05 15:40:46 +010064#include "util/session.h"
Jiri Olsaba6039b62015-11-05 15:40:55 +010065#include "util/tool.h"
Andi Kleen44b1e602016-05-30 12:49:42 -030066#include "util/group.h"
Jiri Olsaba6039b62015-11-05 15:40:55 +010067#include "asm/bug.h"
Ingo Molnarddcacfa2009-04-20 15:37:32 +020068
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -030069#include <linux/time64.h>
Andi Kleen44b1e602016-05-30 12:49:42 -030070#include <api/fs/fs.h>
Peter Zijlstra1f16c572012-10-23 13:40:14 +020071#include <stdlib.h>
Ingo Molnarddcacfa2009-04-20 15:37:32 +020072#include <sys/prctl.h>
Stephane Eranian5af52b52010-05-18 15:00:01 +020073#include <locale.h>
Andi Kleene3b03b62016-05-05 16:04:03 -070074#include <math.h>
Peter Zijlstra16c8a102009-05-05 17:50:27 +020075
Stephane Eraniand7470b62010-12-01 18:49:05 +020076#define DEFAULT_SEPARATOR " "
David Ahern2cee77c2011-05-30 08:55:59 -060077#define CNTR_NOT_SUPPORTED "<not supported>"
78#define CNTR_NOT_COUNTED "<not counted>"
Stephane Eraniand7470b62010-12-01 18:49:05 +020079
Jiri Olsad4f63a42015-06-26 11:29:26 +020080static void print_counters(struct timespec *ts, int argc, const char **argv);
Stephane Eranian13370a92013-01-29 12:47:44 +010081
Andi Kleen4cabc3d2013-08-21 16:47:26 -070082/* Default events used for perf stat -T */
Jiri Olsaa4547422015-06-03 16:25:53 +020083static const char *transaction_attrs = {
84 "task-clock,"
Andi Kleen4cabc3d2013-08-21 16:47:26 -070085 "{"
86 "instructions,"
87 "cycles,"
88 "cpu/cycles-t/,"
89 "cpu/tx-start/,"
90 "cpu/el-start/,"
91 "cpu/cycles-ct/"
92 "}"
93};
94
95/* More limited version when the CPU does not have all events. */
Jiri Olsaa4547422015-06-03 16:25:53 +020096static const char * transaction_limited_attrs = {
97 "task-clock,"
Andi Kleen4cabc3d2013-08-21 16:47:26 -070098 "{"
99 "instructions,"
100 "cycles,"
101 "cpu/cycles-t/,"
102 "cpu/tx-start/"
103 "}"
104};
105
Andi Kleen44b1e602016-05-30 12:49:42 -0300106static const char * topdown_attrs[] = {
107 "topdown-total-slots",
108 "topdown-slots-retired",
109 "topdown-recovery-bubbles",
110 "topdown-fetch-bubbles",
111 "topdown-slots-issued",
112 NULL,
113};
114
Robert Richter666e6d42012-04-05 18:26:27 +0200115static struct perf_evlist *evsel_list;
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -0200116
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -0300117static struct target target = {
Namhyung Kim77a6f012012-05-07 14:09:04 +0900118 .uid = UINT_MAX,
119};
Jaswinder Singh Rajput3d632592009-06-24 18:19:34 +0530120
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100121typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
122
Jaswinder Singh Rajput3d632592009-06-24 18:19:34 +0530123static int run_count = 1;
Stephane Eranian2e6cdf92010-05-12 10:40:01 +0200124static bool no_inherit = false;
Stephane Eraniand07f0b12013-06-04 17:44:26 +0200125static volatile pid_t child_pid = -1;
Ian Munsiec0555642010-04-13 18:37:33 +1000126static bool null_run = false;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +0200127static int detailed_run = 0;
Andi Kleen4cabc3d2013-08-21 16:47:26 -0700128static bool transaction_run;
Andi Kleen44b1e602016-05-30 12:49:42 -0300129static bool topdown_run = false;
Arnaldo Carvalho de Melo201e0b02010-12-01 17:53:27 -0200130static bool big_num = true;
Stephane Eraniand7470b62010-12-01 18:49:05 +0200131static int big_num_opt = -1;
Stephane Eraniand7470b62010-12-01 18:49:05 +0200132static const char *csv_sep = NULL;
133static bool csv_output = false;
Lin Ming43bece72011-08-17 18:42:07 +0800134static bool group = false;
Peter Zijlstra1f16c572012-10-23 13:40:14 +0200135static const char *pre_cmd = NULL;
136static const char *post_cmd = NULL;
137static bool sync_run = false;
Andi Kleen41191682013-08-02 17:41:11 -0700138static unsigned int initial_delay = 0;
Stephane Eranian410136f2013-11-12 17:58:49 +0100139static unsigned int unit_width = 4; /* strlen("unit") */
Frederik Deweerdta7e191c2013-03-01 13:02:27 -0500140static bool forever = false;
Andi Kleen54b50912016-03-03 15:57:36 -0800141static bool metric_only = false;
Andi Kleen44b1e602016-05-30 12:49:42 -0300142static bool force_metric_only = false;
Stephane Eranian13370a92013-01-29 12:47:44 +0100143static struct timespec ref_time;
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100144static struct cpu_map *aggr_map;
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100145static aggr_get_id_t aggr_get_id;
Jiri Olsae0547312015-11-05 15:40:45 +0100146static bool append_file;
147static const char *output_name;
148static int output_fd;
Stephane Eranian5af52b52010-05-18 15:00:01 +0200149
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100150struct perf_stat {
151 bool record;
152 struct perf_data_file file;
153 struct perf_session *session;
154 u64 bytes_written;
Jiri Olsaba6039b62015-11-05 15:40:55 +0100155 struct perf_tool tool;
Jiri Olsa1975d362015-11-05 15:40:56 +0100156 bool maps_allocated;
157 struct cpu_map *cpus;
158 struct thread_map *threads;
Jiri Olsa89af4e02015-11-05 15:41:02 +0100159 enum aggr_mode aggr_mode;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100160};
161
162static struct perf_stat perf_stat;
163#define STAT_RECORD perf_stat.record
164
Liming Wang60666c62009-12-31 16:05:50 +0800165static volatile int done = 0;
166
Jiri Olsa421a50f2015-07-21 14:31:22 +0200167static struct perf_stat_config stat_config = {
168 .aggr_mode = AGGR_GLOBAL,
Jiri Olsa711a5722015-07-21 14:31:23 +0200169 .scale = true,
Jiri Olsa421a50f2015-07-21 14:31:22 +0200170};
171
Stephane Eranian13370a92013-01-29 12:47:44 +0100172static inline void diff_timespec(struct timespec *r, struct timespec *a,
173 struct timespec *b)
174{
175 r->tv_sec = a->tv_sec - b->tv_sec;
176 if (a->tv_nsec < b->tv_nsec) {
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300177 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
Stephane Eranian13370a92013-01-29 12:47:44 +0100178 r->tv_sec--;
179 } else {
180 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
181 }
182}
183
Jiri Olsa254ecbc2015-06-26 11:29:13 +0200184static void perf_stat__reset_stats(void)
185{
186 perf_evlist__reset_stats(evsel_list);
Jiri Olsaf87027b2015-06-03 16:25:59 +0200187 perf_stat__reset_shadow_stats();
Jiri Olsa1eda3b22015-06-03 16:25:55 +0200188}
189
Jiri Olsacac21422012-11-12 18:34:00 +0100190static int create_perf_stat_counter(struct perf_evsel *evsel)
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200191{
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -0200192 struct perf_event_attr *attr = &evsel->attr;
Arnaldo Carvalho de Melo727ab042011-10-25 10:42:19 -0200193
Jiri Olsa711a5722015-07-21 14:31:23 +0200194 if (stat_config.scale)
Ingo Molnara21ca2c2009-06-06 09:58:57 +0200195 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
196 PERF_FORMAT_TOTAL_TIME_RUNNING;
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200197
Arnaldo Carvalho de Melo48290602011-01-03 17:48:12 -0200198 attr->inherit = !no_inherit;
Arnaldo Carvalho de Melo5d2cd902011-04-14 11:20:14 -0300199
Jiri Olsa6acd8e92015-11-25 16:36:54 +0100200 /*
201 * Some events get initialized with sample_(period/type) set,
202 * like tracepoints. Clear it up for counting.
203 */
204 attr->sample_period = 0;
Jiri Olsa6db1a5c2016-01-05 22:09:05 +0100205
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100206 /*
207 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
208 * while avoiding that older tools show confusing messages.
Jiri Olsa6db1a5c2016-01-05 22:09:05 +0100209 *
210 * However for pipe sessions we need to keep it zero,
211 * because script's perf_evsel__check_attr is triggered
212 * by attr->sample_type != 0, and we can't run it on
213 * stat sessions.
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100214 */
Jiri Olsa6db1a5c2016-01-05 22:09:05 +0100215 if (!(STAT_RECORD && perf_stat.file.is_pipe))
216 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
Jiri Olsa6acd8e92015-11-25 16:36:54 +0100217
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100218 /*
219 * Disabling all counters initially, they will be enabled
220 * either manually by us or by kernel via enable_on_exec
221 * set later.
222 */
Jiri Olsac8280ce2015-12-03 10:06:45 +0100223 if (perf_evsel__is_group_leader(evsel)) {
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100224 attr->disabled = 1;
225
Jiri Olsac8280ce2015-12-03 10:06:45 +0100226 /*
227 * In case of initial_delay we enable tracee
228 * events manually.
229 */
230 if (target__none(&target) && !initial_delay)
231 attr->enable_on_exec = 1;
232 }
233
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -0300234 if (target__has_cpu(&target))
Arnaldo Carvalho de Melo594ac612012-12-13 13:13:07 -0300235 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
Stephane Eranian5622c072012-04-27 14:45:38 +0200236
Arnaldo Carvalho de Melo594ac612012-12-13 13:13:07 -0300237 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200238}
239
Ingo Molnarc04f5e52009-05-29 09:10:54 +0200240/*
241 * Does the counter have nsecs as a unit?
242 */
Arnaldo Carvalho de Melodaec78a2011-01-03 16:49:44 -0200243static inline int nsec_counter(struct perf_evsel *evsel)
Ingo Molnarc04f5e52009-05-29 09:10:54 +0200244{
Arnaldo Carvalho de Melodaec78a2011-01-03 16:49:44 -0200245 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
246 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
Ingo Molnarc04f5e52009-05-29 09:10:54 +0200247 return 1;
248
249 return 0;
250}
251
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100252static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100253 union perf_event *event,
254 struct perf_sample *sample __maybe_unused,
255 struct machine *machine __maybe_unused)
256{
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100257 if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
258 pr_err("failed to write perf data, error: %m\n");
259 return -1;
260 }
261
262 perf_stat.bytes_written += event->header.size;
263 return 0;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100264}
265
Jiri Olsa1975d362015-11-05 15:40:56 +0100266static int write_stat_round_event(u64 tm, u64 type)
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100267{
Jiri Olsa1975d362015-11-05 15:40:56 +0100268 return perf_event__synthesize_stat_round(NULL, tm, type,
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100269 process_synthesized_event,
270 NULL);
271}
272
273#define WRITE_STAT_ROUND_EVENT(time, interval) \
274 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
275
Jiri Olsa5a6ea812015-11-05 15:40:51 +0100276#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
277
278static int
279perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
280 struct perf_counts_values *count)
281{
282 struct perf_sample_id *sid = SID(counter, cpu, thread);
283
284 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
285 process_synthesized_event, NULL);
286}
287
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200288/*
289 * Read out the results of a single counter:
290 * do not aggregate counts across CPUs in system-wide mode
291 */
Arnaldo Carvalho de Meloc52b12e2011-01-03 17:45:52 -0200292static int read_counter(struct perf_evsel *counter)
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200293{
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100294 int nthreads = thread_map__nr(evsel_list->threads);
Mark Rutland00e727b2016-07-15 11:08:10 +0100295 int ncpus, cpu, thread;
296
297 if (target__has_cpu(&target))
298 ncpus = perf_evsel__nr_cpus(counter);
299 else
300 ncpus = 1;
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200301
Suzuki K. Poulose3b4331d2015-02-13 18:40:58 +0000302 if (!counter->supported)
303 return -ENOENT;
304
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100305 if (counter->system_wide)
306 nthreads = 1;
307
308 for (thread = 0; thread < nthreads; thread++) {
309 for (cpu = 0; cpu < ncpus; cpu++) {
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200310 struct perf_counts_values *count;
311
312 count = perf_counts(counter->counts, cpu, thread);
313 if (perf_evsel__read(counter, cpu, thread, count))
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100314 return -1;
Jiri Olsa5a6ea812015-11-05 15:40:51 +0100315
316 if (STAT_RECORD) {
317 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
318 pr_err("failed to write stat event\n");
319 return -1;
320 }
321 }
Andi Kleen0b1abbf2016-04-27 13:00:51 -0700322
323 if (verbose > 1) {
324 fprintf(stat_config.output,
325 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
326 perf_evsel__name(counter),
327 cpu,
328 count->val, count->ena, count->run);
329 }
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100330 }
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200331 }
Arnaldo Carvalho de Meloc52b12e2011-01-03 17:45:52 -0200332
333 return 0;
Ingo Molnar2996f5d2009-05-29 09:10:54 +0200334}
335
Mark Rutland3df33ef2016-08-09 14:04:29 +0100336static void read_counters(void)
Jiri Olsa106a94a2015-06-26 11:29:19 +0200337{
338 struct perf_evsel *counter;
Jiri Olsa106a94a2015-06-26 11:29:19 +0200339
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -0300340 evlist__for_each_entry(evsel_list, counter) {
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200341 if (read_counter(counter))
Andi Kleen245bad82015-09-01 15:52:46 -0700342 pr_debug("failed to read counter %s\n", counter->name);
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200343
Jiri Olsaf80010e2015-07-21 14:31:27 +0200344 if (perf_stat_process_counter(&stat_config, counter))
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200345 pr_warning("failed to process counter %s\n", counter->name);
Jiri Olsa106a94a2015-06-26 11:29:19 +0200346 }
347}
348
Jiri Olsaba411a92015-06-26 11:29:24 +0200349static void process_interval(void)
Stephane Eranian13370a92013-01-29 12:47:44 +0100350{
Stephane Eranian13370a92013-01-29 12:47:44 +0100351 struct timespec ts, rs;
Stephane Eranian13370a92013-01-29 12:47:44 +0100352
Mark Rutland3df33ef2016-08-09 14:04:29 +0100353 read_counters();
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100354
Stephane Eranian13370a92013-01-29 12:47:44 +0100355 clock_gettime(CLOCK_MONOTONIC, &ts);
356 diff_timespec(&rs, &ts, &ref_time);
Stephane Eranian13370a92013-01-29 12:47:44 +0100357
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100358 if (STAT_RECORD) {
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -0300359 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100360 pr_err("failed to write stat round event\n");
361 }
362
Jiri Olsad4f63a42015-06-26 11:29:26 +0200363 print_counters(&rs, 0, NULL);
Stephane Eranian13370a92013-01-29 12:47:44 +0100364}
365
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100366static void enable_counters(void)
Andi Kleen41191682013-08-02 17:41:11 -0700367{
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100368 if (initial_delay)
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300369 usleep(initial_delay * USEC_PER_MSEC);
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100370
371 /*
372 * We need to enable counters only if:
373 * - we don't have tracee (attaching to task or cpu)
374 * - we have initial delay configured
375 */
376 if (!target__none(&target) || initial_delay)
Jiri Olsaab46db02015-12-03 10:06:43 +0100377 perf_evlist__enable(evsel_list);
Andi Kleen41191682013-08-02 17:41:11 -0700378}
379
Mark Rutland3df33ef2016-08-09 14:04:29 +0100380static void disable_counters(void)
381{
382 /*
383 * If we don't have tracee (attaching to task or cpu), counters may
384 * still be running. To get accurate group ratios, we must stop groups
385 * from counting before reading their constituent counters.
386 */
387 if (!target__none(&target))
388 perf_evlist__disable(evsel_list);
389}
390
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300391static volatile int workload_exec_errno;
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300392
393/*
394 * perf_evlist__prepare_workload will send a SIGUSR1
395 * if the fork fails, since we asked by setting its
396 * want_signal to true.
397 */
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300398static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
399 void *ucontext __maybe_unused)
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300400{
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300401 workload_exec_errno = info->si_value.sival_int;
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300402}
403
Jiri Olsa7b60a7e2015-11-05 15:40:54 +0100404static bool has_unit(struct perf_evsel *counter)
405{
406 return counter->unit && *counter->unit;
407}
408
409static bool has_scale(struct perf_evsel *counter)
410{
411 return counter->scale != 1;
412}
413
Jiri Olsa664c98d2015-11-05 15:40:50 +0100414static int perf_stat_synthesize_config(bool is_pipe)
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100415{
Jiri Olsa7b60a7e2015-11-05 15:40:54 +0100416 struct perf_evsel *counter;
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100417 int err;
418
Jiri Olsa664c98d2015-11-05 15:40:50 +0100419 if (is_pipe) {
420 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
421 process_synthesized_event);
422 if (err < 0) {
423 pr_err("Couldn't synthesize attrs.\n");
424 return err;
425 }
426 }
427
Jiri Olsa7b60a7e2015-11-05 15:40:54 +0100428 /*
429 * Synthesize other events stuff not carried within
430 * attr event - unit, scale, name
431 */
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -0300432 evlist__for_each_entry(evsel_list, counter) {
Jiri Olsa7b60a7e2015-11-05 15:40:54 +0100433 if (!counter->supported)
434 continue;
435
436 /*
437 * Synthesize unit and scale only if it's defined.
438 */
439 if (has_unit(counter)) {
440 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
441 if (err < 0) {
442 pr_err("Couldn't synthesize evsel unit.\n");
443 return err;
444 }
445 }
446
447 if (has_scale(counter)) {
448 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
449 if (err < 0) {
450 pr_err("Couldn't synthesize evsel scale.\n");
451 return err;
452 }
453 }
454
455 if (counter->own_cpus) {
456 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
457 if (err < 0) {
458 pr_err("Couldn't synthesize evsel scale.\n");
459 return err;
460 }
461 }
462
463 /*
464 * Name is needed only for pipe output,
465 * perf.data carries event names.
466 */
467 if (is_pipe) {
468 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
469 if (err < 0) {
470 pr_err("Couldn't synthesize evsel name.\n");
471 return err;
472 }
473 }
474 }
475
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100476 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
477 process_synthesized_event,
478 NULL);
479 if (err < 0) {
480 pr_err("Couldn't synthesize thread map.\n");
481 return err;
482 }
483
484 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
485 process_synthesized_event, NULL);
486 if (err < 0) {
487 pr_err("Couldn't synthesize thread map.\n");
488 return err;
489 }
490
491 err = perf_event__synthesize_stat_config(NULL, &stat_config,
492 process_synthesized_event, NULL);
493 if (err < 0) {
494 pr_err("Couldn't synthesize config.\n");
495 return err;
496 }
497
498 return 0;
499}
500
Jiri Olsa2af46462015-11-05 15:40:49 +0100501#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
502
503static int __store_counter_ids(struct perf_evsel *counter,
504 struct cpu_map *cpus,
505 struct thread_map *threads)
506{
507 int cpu, thread;
508
509 for (cpu = 0; cpu < cpus->nr; cpu++) {
510 for (thread = 0; thread < threads->nr; thread++) {
511 int fd = FD(counter, cpu, thread);
512
513 if (perf_evlist__id_add_fd(evsel_list, counter,
514 cpu, thread, fd) < 0)
515 return -1;
516 }
517 }
518
519 return 0;
520}
521
522static int store_counter_ids(struct perf_evsel *counter)
523{
524 struct cpu_map *cpus = counter->cpus;
525 struct thread_map *threads = counter->threads;
526
527 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
528 return -ENOMEM;
529
530 return __store_counter_ids(counter, cpus, threads);
531}
532
Namhyung Kimacf28922013-03-11 16:43:18 +0900533static int __run_perf_stat(int argc, const char **argv)
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200534{
Jiri Olsaec0d3d12015-07-21 14:31:25 +0200535 int interval = stat_config.interval;
Arnaldo Carvalho de Melo56e52e82012-12-13 15:10:58 -0300536 char msg[512];
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200537 unsigned long long t0, t1;
Jiri Olsacac21422012-11-12 18:34:00 +0100538 struct perf_evsel *counter;
Stephane Eranian13370a92013-01-29 12:47:44 +0100539 struct timespec ts;
Stephane Eranian410136f2013-11-12 17:58:49 +0100540 size_t l;
Ingo Molnar42202dd2009-06-13 14:57:28 +0200541 int status = 0;
Zhang, Yanmin6be28502010-03-18 11:36:03 -0300542 const bool forks = (argc > 0);
Jiri Olsa664c98d2015-11-05 15:40:50 +0100543 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
Mathieu Poirier5d8bb1e2016-09-16 09:50:03 -0600544 struct perf_evsel_config_term *err_term;
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200545
Stephane Eranian13370a92013-01-29 12:47:44 +0100546 if (interval) {
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300547 ts.tv_sec = interval / USEC_PER_MSEC;
548 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
Stephane Eranian13370a92013-01-29 12:47:44 +0100549 } else {
550 ts.tv_sec = 1;
551 ts.tv_nsec = 0;
552 }
553
Liming Wang60666c62009-12-31 16:05:50 +0800554 if (forks) {
Jiri Olsa664c98d2015-11-05 15:40:50 +0100555 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
Arnaldo Carvalho de Melo735f7e02014-01-03 14:56:49 -0300556 workload_exec_failed_signal) < 0) {
Namhyung Kimacf28922013-03-11 16:43:18 +0900557 perror("failed to prepare workload");
558 return -1;
Liming Wang60666c62009-12-31 16:05:50 +0800559 }
Namhyung Kimd20a47e2013-09-30 18:01:11 +0900560 child_pid = evsel_list->workload.pid;
Paul Mackerras051ae7f2009-06-29 21:13:21 +1000561 }
562
Jiri Olsa6a4bb042012-08-08 12:22:36 +0200563 if (group)
Arnaldo Carvalho de Melo63dab222012-08-14 16:35:48 -0300564 perf_evlist__set_leader(evsel_list);
Jiri Olsa6a4bb042012-08-08 12:22:36 +0200565
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -0300566 evlist__for_each_entry(evsel_list, counter) {
Arnaldo Carvalho de Melo42ef8a72016-05-12 16:25:18 -0300567try_again:
Jiri Olsacac21422012-11-12 18:34:00 +0100568 if (create_perf_stat_counter(counter) < 0) {
David Ahern979987a2012-05-08 09:29:16 -0600569 /*
570 * PPC returns ENXIO for HW counters until 2.6.37
571 * (behavior changed with commit b0a873e).
572 */
Anton Blanchard38f6ae12011-12-02 09:38:33 +1100573 if (errno == EINVAL || errno == ENOSYS ||
David Ahern979987a2012-05-08 09:29:16 -0600574 errno == ENOENT || errno == EOPNOTSUPP ||
575 errno == ENXIO) {
David Ahernc63ca0c2011-04-29 16:04:15 -0600576 if (verbose)
577 ui__warning("%s event is not supported by the kernel.\n",
Arnaldo Carvalho de Melo7289f832012-06-12 12:34:58 -0300578 perf_evsel__name(counter));
David Ahern2cee77c2011-05-30 08:55:59 -0600579 counter->supported = false;
Kan Liangcb5ef602015-06-11 02:32:40 -0400580
581 if ((counter->leader != counter) ||
582 !(counter->leader->nr_members > 1))
583 continue;
Arnaldo Carvalho de Melo42ef8a72016-05-12 16:25:18 -0300584 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
585 if (verbose)
586 ui__warning("%s\n", msg);
587 goto try_again;
588 }
Ingo Molnarede70292011-04-28 08:48:42 +0200589
Arnaldo Carvalho de Melo56e52e82012-12-13 15:10:58 -0300590 perf_evsel__open_strerror(counter, &target,
591 errno, msg, sizeof(msg));
592 ui__error("%s\n", msg);
593
Arnaldo Carvalho de Melo48290602011-01-03 17:48:12 -0200594 if (child_pid != -1)
595 kill(child_pid, SIGTERM);
David Ahernfceda7f2012-08-26 12:24:44 -0600596
Arnaldo Carvalho de Melo48290602011-01-03 17:48:12 -0200597 return -1;
598 }
David Ahern2cee77c2011-05-30 08:55:59 -0600599 counter->supported = true;
Stephane Eranian410136f2013-11-12 17:58:49 +0100600
601 l = strlen(counter->unit);
602 if (l > unit_width)
603 unit_width = l;
Jiri Olsa2af46462015-11-05 15:40:49 +0100604
605 if (STAT_RECORD && store_counter_ids(counter))
606 return -1;
Arnaldo Carvalho de Melo084ab9f2010-03-22 13:10:28 -0300607 }
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200608
Arnaldo Carvalho de Melo23d4aad2015-03-24 19:23:47 -0300609 if (perf_evlist__apply_filters(evsel_list, &counter)) {
610 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
611 counter->filter, perf_evsel__name(counter), errno,
Arnaldo Carvalho de Meloc8b5f2c2016-07-06 11:56:20 -0300612 str_error_r(errno, msg, sizeof(msg)));
Frederic Weisbeckercfd748a2011-03-14 16:40:30 +0100613 return -1;
614 }
615
Mathieu Poirier5d8bb1e2016-09-16 09:50:03 -0600616 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
617 error("failed to set config \"%s\" on event %s with %d (%s)\n",
618 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
619 str_error_r(errno, msg, sizeof(msg)));
620 return -1;
621 }
622
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100623 if (STAT_RECORD) {
624 int err, fd = perf_data_file__fd(&perf_stat.file);
625
Jiri Olsa664c98d2015-11-05 15:40:50 +0100626 if (is_pipe) {
627 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
628 } else {
629 err = perf_session__write_header(perf_stat.session, evsel_list,
630 fd, false);
631 }
632
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100633 if (err < 0)
634 return err;
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100635
Jiri Olsa664c98d2015-11-05 15:40:50 +0100636 err = perf_stat_synthesize_config(is_pipe);
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100637 if (err < 0)
638 return err;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100639 }
640
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200641 /*
642 * Enable counters and exec the command:
643 */
644 t0 = rdclock();
Stephane Eranian13370a92013-01-29 12:47:44 +0100645 clock_gettime(CLOCK_MONOTONIC, &ref_time);
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200646
Liming Wang60666c62009-12-31 16:05:50 +0800647 if (forks) {
Namhyung Kimacf28922013-03-11 16:43:18 +0900648 perf_evlist__start_workload(evsel_list);
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100649 enable_counters();
Namhyung Kimacf28922013-03-11 16:43:18 +0900650
Stephane Eranian13370a92013-01-29 12:47:44 +0100651 if (interval) {
652 while (!waitpid(child_pid, &status, WNOHANG)) {
653 nanosleep(&ts, NULL);
Jiri Olsaba411a92015-06-26 11:29:24 +0200654 process_interval();
Stephane Eranian13370a92013-01-29 12:47:44 +0100655 }
656 }
Liming Wang60666c62009-12-31 16:05:50 +0800657 wait(&status);
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300658
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300659 if (workload_exec_errno) {
Arnaldo Carvalho de Meloc8b5f2c2016-07-06 11:56:20 -0300660 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300661 pr_err("Workload failed: %s\n", emsg);
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300662 return -1;
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300663 }
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300664
Andi Kleen33e49ea2011-09-15 14:31:40 -0700665 if (WIFSIGNALED(status))
666 psignal(WTERMSIG(status), argv[0]);
Liming Wang60666c62009-12-31 16:05:50 +0800667 } else {
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100668 enable_counters();
Stephane Eranian13370a92013-01-29 12:47:44 +0100669 while (!done) {
670 nanosleep(&ts, NULL);
671 if (interval)
Jiri Olsaba411a92015-06-26 11:29:24 +0200672 process_interval();
Stephane Eranian13370a92013-01-29 12:47:44 +0100673 }
Liming Wang60666c62009-12-31 16:05:50 +0800674 }
Ingo Molnar44db76c2009-06-03 19:36:07 +0200675
Mark Rutland3df33ef2016-08-09 14:04:29 +0100676 disable_counters();
677
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200678 t1 = rdclock();
679
Peter Zijlstra9e9772c2009-09-04 15:36:08 +0200680 update_stats(&walltime_nsecs_stats, t1 - t0);
Ingo Molnar42202dd2009-06-13 14:57:28 +0200681
Mark Rutland3df33ef2016-08-09 14:04:29 +0100682 /*
683 * Closing a group leader splits the group, and as we only disable
684 * group leaders, results in remaining events becoming enabled. To
685 * avoid arbitrary skew, we must read all counters before closing any
686 * group leaders.
687 */
688 read_counters();
689 perf_evlist__close(evsel_list);
Arnaldo Carvalho de Meloc52b12e2011-01-03 17:45:52 -0200690
Ingo Molnar42202dd2009-06-13 14:57:28 +0200691 return WEXITSTATUS(status);
692}
693
Arnaldo Carvalho de Melo41cde472014-01-03 17:34:42 -0300694static int run_perf_stat(int argc, const char **argv)
Peter Zijlstra1f16c572012-10-23 13:40:14 +0200695{
696 int ret;
697
698 if (pre_cmd) {
699 ret = system(pre_cmd);
700 if (ret)
701 return ret;
702 }
703
704 if (sync_run)
705 sync();
706
707 ret = __run_perf_stat(argc, argv);
708 if (ret)
709 return ret;
710
711 if (post_cmd) {
712 ret = system(post_cmd);
713 if (ret)
714 return ret;
715 }
716
717 return ret;
718}
719
Andi Kleend73515c2015-03-11 07:16:27 -0700720static void print_running(u64 run, u64 ena)
721{
722 if (csv_output) {
Jiri Olsa58215222015-07-21 14:31:24 +0200723 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
Andi Kleend73515c2015-03-11 07:16:27 -0700724 csv_sep,
725 run,
726 csv_sep,
727 ena ? 100.0 * run / ena : 100.0);
728 } else if (run != ena) {
Jiri Olsa58215222015-07-21 14:31:24 +0200729 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
Andi Kleend73515c2015-03-11 07:16:27 -0700730 }
731}
732
Ingo Molnarf99844c2011-04-27 05:35:39 +0200733static void print_noise_pct(double total, double avg)
734{
Xiao Guangrong0007ece2012-09-17 16:31:14 +0800735 double pct = rel_stddev_stats(total, avg);
Ingo Molnarf99844c2011-04-27 05:35:39 +0200736
Zhengyu He3ae9a34d2011-06-23 13:45:42 -0700737 if (csv_output)
Jiri Olsa58215222015-07-21 14:31:24 +0200738 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
Jim Cromiea1bca6c2011-09-07 17:14:02 -0600739 else if (pct)
Jiri Olsa58215222015-07-21 14:31:24 +0200740 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
Ingo Molnarf99844c2011-04-27 05:35:39 +0200741}
742
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -0200743static void print_noise(struct perf_evsel *evsel, double avg)
Ingo Molnar42202dd2009-06-13 14:57:28 +0200744{
Jiri Olsa581cc8a2015-10-16 12:41:03 +0200745 struct perf_stat_evsel *ps;
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -0200746
Peter Zijlstra849abde2009-09-04 18:23:38 +0200747 if (run_count == 1)
748 return;
749
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -0200750 ps = evsel->priv;
Ingo Molnarf99844c2011-04-27 05:35:39 +0200751 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
Ingo Molnar42202dd2009-06-13 14:57:28 +0200752}
753
Stephane Eranian12c08a92013-02-14 13:57:29 +0100754static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
Ingo Molnar42202dd2009-06-13 14:57:28 +0200755{
Jiri Olsa421a50f2015-07-21 14:31:22 +0200756 switch (stat_config.aggr_mode) {
Stephane Eranian12c08a92013-02-14 13:57:29 +0100757 case AGGR_CORE:
Jiri Olsa58215222015-07-21 14:31:24 +0200758 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
Stephane Eranian12c08a92013-02-14 13:57:29 +0100759 cpu_map__id_to_socket(id),
760 csv_output ? 0 : -8,
761 cpu_map__id_to_cpu(id),
762 csv_sep,
763 csv_output ? 0 : 4,
764 nr,
765 csv_sep);
766 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100767 case AGGR_SOCKET:
Jiri Olsa58215222015-07-21 14:31:24 +0200768 fprintf(stat_config.output, "S%*d%s%*d%s",
Stephane Eraniand7e7a452013-02-06 15:46:02 +0100769 csv_output ? 0 : -5,
Stephane Eranian12c08a92013-02-14 13:57:29 +0100770 id,
Stephane Eraniand7e7a452013-02-06 15:46:02 +0100771 csv_sep,
772 csv_output ? 0 : 4,
773 nr,
774 csv_sep);
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100775 break;
776 case AGGR_NONE:
Jiri Olsa58215222015-07-21 14:31:24 +0200777 fprintf(stat_config.output, "CPU%*d%s",
Stephane Eraniand7470b62010-12-01 18:49:05 +0200778 csv_output ? 0 : -4,
Stephane Eranian12c08a92013-02-14 13:57:29 +0100779 perf_evsel__cpus(evsel)->map[id], csv_sep);
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100780 break;
Jiri Olsa32b8af82015-06-26 11:29:27 +0200781 case AGGR_THREAD:
Jiri Olsa58215222015-07-21 14:31:24 +0200782 fprintf(stat_config.output, "%*s-%*d%s",
Jiri Olsa32b8af82015-06-26 11:29:27 +0200783 csv_output ? 0 : 16,
784 thread_map__comm(evsel->threads, id),
785 csv_output ? 0 : -8,
786 thread_map__pid(evsel->threads, id),
787 csv_sep);
788 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100789 case AGGR_GLOBAL:
Jiri Olsa208df992015-10-16 12:41:04 +0200790 case AGGR_UNSET:
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100791 default:
792 break;
793 }
794}
Stephane Eraniand7470b62010-12-01 18:49:05 +0200795
Andi Kleen140aead2016-01-30 09:06:49 -0800796struct outstate {
797 FILE *fh;
798 bool newline;
Andi Kleenf9483392016-01-30 09:06:50 -0800799 const char *prefix;
Andi Kleen92a61f62016-02-29 14:36:21 -0800800 int nfields;
Andi Kleen44d49a62016-02-29 14:36:22 -0800801 int id, nr;
802 struct perf_evsel *evsel;
Andi Kleen140aead2016-01-30 09:06:49 -0800803};
804
805#define METRIC_LEN 35
806
807static void new_line_std(void *ctx)
808{
809 struct outstate *os = ctx;
810
811 os->newline = true;
812}
813
814static void do_new_line_std(struct outstate *os)
815{
816 fputc('\n', os->fh);
Andi Kleenf9483392016-01-30 09:06:50 -0800817 fputs(os->prefix, os->fh);
Andi Kleen44d49a62016-02-29 14:36:22 -0800818 aggr_printout(os->evsel, os->id, os->nr);
Andi Kleen140aead2016-01-30 09:06:49 -0800819 if (stat_config.aggr_mode == AGGR_NONE)
820 fprintf(os->fh, " ");
Andi Kleen140aead2016-01-30 09:06:49 -0800821 fprintf(os->fh, " ");
822}
823
824static void print_metric_std(void *ctx, const char *color, const char *fmt,
825 const char *unit, double val)
826{
827 struct outstate *os = ctx;
828 FILE *out = os->fh;
829 int n;
830 bool newline = os->newline;
831
832 os->newline = false;
833
834 if (unit == NULL || fmt == NULL) {
835 fprintf(out, "%-*s", METRIC_LEN, "");
836 return;
837 }
838
839 if (newline)
840 do_new_line_std(os);
841
842 n = fprintf(out, " # ");
843 if (color)
844 n += color_fprintf(out, color, fmt, val);
845 else
846 n += fprintf(out, fmt, val);
847 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
848}
849
Andi Kleen92a61f62016-02-29 14:36:21 -0800850static void new_line_csv(void *ctx)
851{
852 struct outstate *os = ctx;
853 int i;
854
855 fputc('\n', os->fh);
856 if (os->prefix)
857 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
Andi Kleen44d49a62016-02-29 14:36:22 -0800858 aggr_printout(os->evsel, os->id, os->nr);
Andi Kleen92a61f62016-02-29 14:36:21 -0800859 for (i = 0; i < os->nfields; i++)
860 fputs(csv_sep, os->fh);
861}
862
863static void print_metric_csv(void *ctx,
864 const char *color __maybe_unused,
865 const char *fmt, const char *unit, double val)
866{
867 struct outstate *os = ctx;
868 FILE *out = os->fh;
869 char buf[64], *vals, *ends;
870
871 if (unit == NULL || fmt == NULL) {
872 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
873 return;
874 }
875 snprintf(buf, sizeof(buf), fmt, val);
876 vals = buf;
877 while (isspace(*vals))
878 vals++;
879 ends = vals;
880 while (isdigit(*ends) || *ends == '.')
881 ends++;
882 *ends = 0;
883 while (isspace(*unit))
884 unit++;
885 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
886}
887
Andi Kleen54b50912016-03-03 15:57:36 -0800888#define METRIC_ONLY_LEN 20
889
890/* Filter out some columns that don't work well in metrics only mode */
891
892static bool valid_only_metric(const char *unit)
893{
894 if (!unit)
895 return false;
896 if (strstr(unit, "/sec") ||
897 strstr(unit, "hz") ||
898 strstr(unit, "Hz") ||
899 strstr(unit, "CPUs utilized"))
900 return false;
901 return true;
902}
903
904static const char *fixunit(char *buf, struct perf_evsel *evsel,
905 const char *unit)
906{
907 if (!strncmp(unit, "of all", 6)) {
908 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
909 unit);
910 return buf;
911 }
912 return unit;
913}
914
915static void print_metric_only(void *ctx, const char *color, const char *fmt,
916 const char *unit, double val)
917{
918 struct outstate *os = ctx;
919 FILE *out = os->fh;
920 int n;
921 char buf[1024];
922 unsigned mlen = METRIC_ONLY_LEN;
923
924 if (!valid_only_metric(unit))
925 return;
926 unit = fixunit(buf, os->evsel, unit);
927 if (color)
928 n = color_fprintf(out, color, fmt, val);
929 else
930 n = fprintf(out, fmt, val);
931 if (n > METRIC_ONLY_LEN)
932 n = METRIC_ONLY_LEN;
933 if (mlen < strlen(unit))
934 mlen = strlen(unit) + 1;
935 fprintf(out, "%*s", mlen - n, "");
936}
937
938static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
939 const char *fmt,
940 const char *unit, double val)
941{
942 struct outstate *os = ctx;
943 FILE *out = os->fh;
944 char buf[64], *vals, *ends;
945 char tbuf[1024];
946
947 if (!valid_only_metric(unit))
948 return;
949 unit = fixunit(tbuf, os->evsel, unit);
950 snprintf(buf, sizeof buf, fmt, val);
951 vals = buf;
952 while (isspace(*vals))
953 vals++;
954 ends = vals;
955 while (isdigit(*ends) || *ends == '.')
956 ends++;
957 *ends = 0;
958 fprintf(out, "%s%s", vals, csv_sep);
959}
960
961static void new_line_metric(void *ctx __maybe_unused)
962{
963}
964
965static void print_metric_header(void *ctx, const char *color __maybe_unused,
966 const char *fmt __maybe_unused,
967 const char *unit, double val __maybe_unused)
968{
969 struct outstate *os = ctx;
970 char tbuf[1024];
971
972 if (!valid_only_metric(unit))
973 return;
974 unit = fixunit(tbuf, os->evsel, unit);
975 if (csv_output)
976 fprintf(os->fh, "%s%s", unit, csv_sep);
977 else
978 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
979}
980
Andi Kleenda88c7f2014-09-24 13:50:46 -0700981static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100982{
Jiri Olsa58215222015-07-21 14:31:24 +0200983 FILE *output = stat_config.output;
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300984 double msecs = avg / NSEC_PER_MSEC;
Stephane Eranian410136f2013-11-12 17:58:49 +0100985 const char *fmt_v, *fmt_n;
David Ahern4bbe5a62013-09-28 14:28:00 -0600986 char name[25];
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100987
Stephane Eranian410136f2013-11-12 17:58:49 +0100988 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
989 fmt_n = csv_output ? "%s" : "%-25s";
990
Andi Kleenda88c7f2014-09-24 13:50:46 -0700991 aggr_printout(evsel, id, nr);
Stephane Eranian86ee6e12013-02-14 13:57:27 +0100992
David Ahern4bbe5a62013-09-28 14:28:00 -0600993 scnprintf(name, sizeof(name), "%s%s",
994 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
Stephane Eranian410136f2013-11-12 17:58:49 +0100995
996 fprintf(output, fmt_v, msecs, csv_sep);
997
998 if (csv_output)
999 fprintf(output, "%s%s", evsel->unit, csv_sep);
1000 else
1001 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1002
1003 fprintf(output, fmt_n, name);
Stephane Eraniand7470b62010-12-01 18:49:05 +02001004
Stephane Eranian023695d2011-02-14 11:20:01 +02001005 if (evsel->cgrp)
Stephane Eranian4aa90152011-08-15 22:22:33 +02001006 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
Ingo Molnar42202dd2009-06-13 14:57:28 +02001007}
1008
Andi Kleen44d49a62016-02-29 14:36:22 -08001009static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1010{
1011 int i;
1012
1013 if (!aggr_get_id)
1014 return 0;
1015
1016 if (stat_config.aggr_mode == AGGR_NONE)
1017 return id;
1018
1019 if (stat_config.aggr_mode == AGGR_GLOBAL)
1020 return 0;
1021
1022 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1023 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1024
1025 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1026 return cpu2;
1027 }
1028 return 0;
1029}
1030
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001031static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1032{
Jiri Olsa58215222015-07-21 14:31:24 +02001033 FILE *output = stat_config.output;
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001034 double sc = evsel->scale;
1035 const char *fmt;
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001036
1037 if (csv_output) {
Andi Kleene3b03b62016-05-05 16:04:03 -07001038 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001039 } else {
1040 if (big_num)
Andi Kleene3b03b62016-05-05 16:04:03 -07001041 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001042 else
Andi Kleene3b03b62016-05-05 16:04:03 -07001043 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001044 }
1045
1046 aggr_printout(evsel, id, nr);
1047
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001048 fprintf(output, fmt, avg, csv_sep);
1049
1050 if (evsel->unit)
1051 fprintf(output, "%-*s%s",
1052 csv_output ? 0 : unit_width,
1053 evsel->unit, csv_sep);
1054
1055 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1056
1057 if (evsel->cgrp)
1058 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
Andi Kleeneedfcb42015-11-02 17:50:21 -08001059}
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001060
Andi Kleenf9483392016-01-30 09:06:50 -08001061static void printout(int id, int nr, struct perf_evsel *counter, double uval,
Andi Kleencb110f42016-01-30 09:06:51 -08001062 char *prefix, u64 run, u64 ena, double noise)
Andi Kleeneedfcb42015-11-02 17:50:21 -08001063{
Andi Kleen140aead2016-01-30 09:06:49 -08001064 struct perf_stat_output_ctx out;
Andi Kleenf9483392016-01-30 09:06:50 -08001065 struct outstate os = {
1066 .fh = stat_config.output,
Andi Kleen44d49a62016-02-29 14:36:22 -08001067 .prefix = prefix ? prefix : "",
1068 .id = id,
1069 .nr = nr,
1070 .evsel = counter,
Andi Kleenf9483392016-01-30 09:06:50 -08001071 };
Andi Kleen140aead2016-01-30 09:06:49 -08001072 print_metric_t pm = print_metric_std;
1073 void (*nl)(void *);
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001074
Andi Kleen54b50912016-03-03 15:57:36 -08001075 if (metric_only) {
1076 nl = new_line_metric;
1077 if (csv_output)
1078 pm = print_metric_only_csv;
1079 else
1080 pm = print_metric_only;
1081 } else
1082 nl = new_line_std;
Andi Kleeneedfcb42015-11-02 17:50:21 -08001083
Andi Kleen54b50912016-03-03 15:57:36 -08001084 if (csv_output && !metric_only) {
Andi Kleen92a61f62016-02-29 14:36:21 -08001085 static int aggr_fields[] = {
1086 [AGGR_GLOBAL] = 0,
1087 [AGGR_THREAD] = 1,
1088 [AGGR_NONE] = 1,
1089 [AGGR_SOCKET] = 2,
1090 [AGGR_CORE] = 2,
1091 };
1092
1093 pm = print_metric_csv;
1094 nl = new_line_csv;
1095 os.nfields = 3;
1096 os.nfields += aggr_fields[stat_config.aggr_mode];
1097 if (counter->cgrp)
1098 os.nfields++;
1099 }
Andi Kleenb002f3b2016-02-17 14:44:00 -08001100 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
Andi Kleen54b50912016-03-03 15:57:36 -08001101 if (metric_only) {
1102 pm(&os, NULL, "", "", 0);
1103 return;
1104 }
Andi Kleencb110f42016-01-30 09:06:51 -08001105 aggr_printout(counter, id, nr);
1106
1107 fprintf(stat_config.output, "%*s%s",
1108 csv_output ? 0 : 18,
1109 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1110 csv_sep);
1111
1112 fprintf(stat_config.output, "%-*s%s",
1113 csv_output ? 0 : unit_width,
1114 counter->unit, csv_sep);
1115
1116 fprintf(stat_config.output, "%*s",
1117 csv_output ? 0 : -25,
1118 perf_evsel__name(counter));
1119
1120 if (counter->cgrp)
1121 fprintf(stat_config.output, "%s%s",
1122 csv_sep, counter->cgrp->name);
1123
Andi Kleen92a61f62016-02-29 14:36:21 -08001124 if (!csv_output)
1125 pm(&os, NULL, NULL, "", 0);
1126 print_noise(counter, noise);
Andi Kleencb110f42016-01-30 09:06:51 -08001127 print_running(run, ena);
Andi Kleen92a61f62016-02-29 14:36:21 -08001128 if (csv_output)
1129 pm(&os, NULL, NULL, "", 0);
Andi Kleencb110f42016-01-30 09:06:51 -08001130 return;
1131 }
1132
Andi Kleen54b50912016-03-03 15:57:36 -08001133 if (metric_only)
1134 /* nothing */;
1135 else if (nsec_counter(counter))
Andi Kleeneedfcb42015-11-02 17:50:21 -08001136 nsec_printout(id, nr, counter, uval);
1137 else
1138 abs_printout(id, nr, counter, uval);
1139
Andi Kleen140aead2016-01-30 09:06:49 -08001140 out.print_metric = pm;
1141 out.new_line = nl;
1142 out.ctx = &os;
1143
Andi Kleen54b50912016-03-03 15:57:36 -08001144 if (csv_output && !metric_only) {
Andi Kleen92a61f62016-02-29 14:36:21 -08001145 print_noise(counter, noise);
1146 print_running(run, ena);
1147 }
1148
1149 perf_stat__print_shadow_stats(counter, uval,
Andi Kleen44d49a62016-02-29 14:36:22 -08001150 first_shadow_cpu(counter, id),
Andi Kleen140aead2016-01-30 09:06:49 -08001151 &out);
Andi Kleen54b50912016-03-03 15:57:36 -08001152 if (!csv_output && !metric_only) {
Andi Kleen92a61f62016-02-29 14:36:21 -08001153 print_noise(counter, noise);
1154 print_running(run, ena);
1155 }
Jiri Olsa556b1fb2015-06-03 16:25:56 +02001156}
1157
Andi Kleen44d49a62016-02-29 14:36:22 -08001158static void aggr_update_shadow(void)
1159{
1160 int cpu, s2, id, s;
1161 u64 val;
1162 struct perf_evsel *counter;
1163
1164 for (s = 0; s < aggr_map->nr; s++) {
1165 id = aggr_map->map[s];
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001166 evlist__for_each_entry(evsel_list, counter) {
Andi Kleen44d49a62016-02-29 14:36:22 -08001167 val = 0;
1168 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1169 s2 = aggr_get_id(evsel_list->cpus, cpu);
1170 if (s2 != id)
1171 continue;
1172 val += perf_counts(counter->counts, cpu, 0)->val;
1173 }
1174 val = val * counter->scale;
1175 perf_stat__update_shadow_stats(counter, &val,
1176 first_shadow_cpu(counter, id));
1177 }
1178 }
1179}
1180
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001181static void print_aggr(char *prefix)
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001182{
Jiri Olsa58215222015-07-21 14:31:24 +02001183 FILE *output = stat_config.output;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001184 struct perf_evsel *counter;
Kan Liang601083c2015-07-02 03:08:43 -04001185 int cpu, s, s2, id, nr;
Stephane Eranian410136f2013-11-12 17:58:49 +01001186 double uval;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001187 u64 ena, run, val;
Andi Kleen54b50912016-03-03 15:57:36 -08001188 bool first;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001189
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001190 if (!(aggr_map || aggr_get_id))
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001191 return;
1192
Andi Kleen44d49a62016-02-29 14:36:22 -08001193 aggr_update_shadow();
1194
Andi Kleen54b50912016-03-03 15:57:36 -08001195 /*
1196 * With metric_only everything is on a single line.
1197 * Without each counter has its own line.
1198 */
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001199 for (s = 0; s < aggr_map->nr; s++) {
Andi Kleen54b50912016-03-03 15:57:36 -08001200 if (prefix && metric_only)
1201 fprintf(output, "%s", prefix);
1202
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001203 id = aggr_map->map[s];
Andi Kleen54b50912016-03-03 15:57:36 -08001204 first = true;
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001205 evlist__for_each_entry(evsel_list, counter) {
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001206 val = ena = run = 0;
1207 nr = 0;
1208 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
Kan Liang601083c2015-07-02 03:08:43 -04001209 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001210 if (s2 != id)
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001211 continue;
Jiri Olsaa6fa0032015-06-26 11:29:11 +02001212 val += perf_counts(counter->counts, cpu, 0)->val;
1213 ena += perf_counts(counter->counts, cpu, 0)->ena;
1214 run += perf_counts(counter->counts, cpu, 0)->run;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001215 nr++;
1216 }
Andi Kleen54b50912016-03-03 15:57:36 -08001217 if (first && metric_only) {
1218 first = false;
1219 aggr_printout(counter, id, nr);
1220 }
1221 if (prefix && !metric_only)
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001222 fprintf(output, "%s", prefix);
1223
Stephane Eranian410136f2013-11-12 17:58:49 +01001224 uval = val * counter->scale;
Andi Kleencb110f42016-01-30 09:06:51 -08001225 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
Andi Kleen54b50912016-03-03 15:57:36 -08001226 if (!metric_only)
1227 fputc('\n', output);
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001228 }
Andi Kleen54b50912016-03-03 15:57:36 -08001229 if (metric_only)
1230 fputc('\n', output);
Stephane Eraniand7e7a452013-02-06 15:46:02 +01001231 }
1232}
1233
Jiri Olsa32b8af82015-06-26 11:29:27 +02001234static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1235{
Jiri Olsa58215222015-07-21 14:31:24 +02001236 FILE *output = stat_config.output;
Jiri Olsa32b8af82015-06-26 11:29:27 +02001237 int nthreads = thread_map__nr(counter->threads);
1238 int ncpus = cpu_map__nr(counter->cpus);
1239 int cpu, thread;
1240 double uval;
1241
1242 for (thread = 0; thread < nthreads; thread++) {
1243 u64 ena = 0, run = 0, val = 0;
1244
1245 for (cpu = 0; cpu < ncpus; cpu++) {
1246 val += perf_counts(counter->counts, cpu, thread)->val;
1247 ena += perf_counts(counter->counts, cpu, thread)->ena;
1248 run += perf_counts(counter->counts, cpu, thread)->run;
1249 }
1250
1251 if (prefix)
1252 fprintf(output, "%s", prefix);
1253
1254 uval = val * counter->scale;
Andi Kleencb110f42016-01-30 09:06:51 -08001255 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
Jiri Olsa32b8af82015-06-26 11:29:27 +02001256 fputc('\n', output);
1257 }
1258}
1259
Ingo Molnar42202dd2009-06-13 14:57:28 +02001260/*
1261 * Print out the results of a single counter:
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001262 * aggregated counts in system-wide mode
Ingo Molnar42202dd2009-06-13 14:57:28 +02001263 */
Stephane Eranian13370a92013-01-29 12:47:44 +01001264static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
Ingo Molnar42202dd2009-06-13 14:57:28 +02001265{
Jiri Olsa58215222015-07-21 14:31:24 +02001266 FILE *output = stat_config.output;
Jiri Olsa581cc8a2015-10-16 12:41:03 +02001267 struct perf_stat_evsel *ps = counter->priv;
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -02001268 double avg = avg_stats(&ps->res_stats[0]);
Stephane Eranian410136f2013-11-12 17:58:49 +01001269 double uval;
Andi Kleend73515c2015-03-11 07:16:27 -07001270 double avg_enabled, avg_running;
1271
1272 avg_enabled = avg_stats(&ps->res_stats[1]);
1273 avg_running = avg_stats(&ps->res_stats[2]);
Ingo Molnar42202dd2009-06-13 14:57:28 +02001274
Andi Kleen54b50912016-03-03 15:57:36 -08001275 if (prefix && !metric_only)
Stephane Eranian13370a92013-01-29 12:47:44 +01001276 fprintf(output, "%s", prefix);
1277
Stephane Eranian410136f2013-11-12 17:58:49 +01001278 uval = avg * counter->scale;
Andi Kleencb110f42016-01-30 09:06:51 -08001279 printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
Andi Kleen54b50912016-03-03 15:57:36 -08001280 if (!metric_only)
1281 fprintf(output, "\n");
Ingo Molnar42202dd2009-06-13 14:57:28 +02001282}
1283
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001284/*
1285 * Print out the results of a single counter:
1286 * does not use aggregated count in system-wide
1287 */
Stephane Eranian13370a92013-01-29 12:47:44 +01001288static void print_counter(struct perf_evsel *counter, char *prefix)
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001289{
Jiri Olsa58215222015-07-21 14:31:24 +02001290 FILE *output = stat_config.output;
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001291 u64 ena, run, val;
Stephane Eranian410136f2013-11-12 17:58:49 +01001292 double uval;
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001293 int cpu;
1294
Yan, Zheng7ae92e72012-09-10 15:53:50 +08001295 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
Jiri Olsaa6fa0032015-06-26 11:29:11 +02001296 val = perf_counts(counter->counts, cpu, 0)->val;
1297 ena = perf_counts(counter->counts, cpu, 0)->ena;
1298 run = perf_counts(counter->counts, cpu, 0)->run;
Stephane Eranian13370a92013-01-29 12:47:44 +01001299
1300 if (prefix)
1301 fprintf(output, "%s", prefix);
1302
Stephane Eranian410136f2013-11-12 17:58:49 +01001303 uval = val * counter->scale;
Andi Kleencb110f42016-01-30 09:06:51 -08001304 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001305
Stephane Eranian4aa90152011-08-15 22:22:33 +02001306 fputc('\n', output);
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001307 }
1308}
1309
Andi Kleen206cab62016-03-03 15:57:37 -08001310static void print_no_aggr_metric(char *prefix)
1311{
1312 int cpu;
1313 int nrcpus = 0;
1314 struct perf_evsel *counter;
1315 u64 ena, run, val;
1316 double uval;
1317
1318 nrcpus = evsel_list->cpus->nr;
1319 for (cpu = 0; cpu < nrcpus; cpu++) {
1320 bool first = true;
1321
1322 if (prefix)
1323 fputs(prefix, stat_config.output);
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001324 evlist__for_each_entry(evsel_list, counter) {
Andi Kleen206cab62016-03-03 15:57:37 -08001325 if (first) {
1326 aggr_printout(counter, cpu, 0);
1327 first = false;
1328 }
1329 val = perf_counts(counter->counts, cpu, 0)->val;
1330 ena = perf_counts(counter->counts, cpu, 0)->ena;
1331 run = perf_counts(counter->counts, cpu, 0)->run;
1332
1333 uval = val * counter->scale;
1334 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1335 }
1336 fputc('\n', stat_config.output);
1337 }
1338}
1339
Andi Kleen54b50912016-03-03 15:57:36 -08001340static int aggr_header_lens[] = {
1341 [AGGR_CORE] = 18,
1342 [AGGR_SOCKET] = 12,
Andi Kleen206cab62016-03-03 15:57:37 -08001343 [AGGR_NONE] = 6,
Andi Kleen54b50912016-03-03 15:57:36 -08001344 [AGGR_THREAD] = 24,
1345 [AGGR_GLOBAL] = 0,
1346};
1347
Andi Kleenc51fd632016-05-24 12:52:39 -07001348static const char *aggr_header_csv[] = {
1349 [AGGR_CORE] = "core,cpus,",
1350 [AGGR_SOCKET] = "socket,cpus",
1351 [AGGR_NONE] = "cpu,",
1352 [AGGR_THREAD] = "comm-pid,",
1353 [AGGR_GLOBAL] = ""
1354};
1355
Andi Kleen41c8ca22016-05-24 12:52:38 -07001356static void print_metric_headers(const char *prefix, bool no_indent)
Andi Kleen54b50912016-03-03 15:57:36 -08001357{
1358 struct perf_stat_output_ctx out;
1359 struct perf_evsel *counter;
1360 struct outstate os = {
1361 .fh = stat_config.output
1362 };
1363
1364 if (prefix)
1365 fprintf(stat_config.output, "%s", prefix);
1366
Andi Kleen41c8ca22016-05-24 12:52:38 -07001367 if (!csv_output && !no_indent)
Andi Kleen54b50912016-03-03 15:57:36 -08001368 fprintf(stat_config.output, "%*s",
1369 aggr_header_lens[stat_config.aggr_mode], "");
Andi Kleenc51fd632016-05-24 12:52:39 -07001370 if (csv_output) {
1371 if (stat_config.interval)
1372 fputs("time,", stat_config.output);
1373 fputs(aggr_header_csv[stat_config.aggr_mode],
1374 stat_config.output);
1375 }
Andi Kleen54b50912016-03-03 15:57:36 -08001376
1377 /* Print metrics headers only */
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001378 evlist__for_each_entry(evsel_list, counter) {
Andi Kleen54b50912016-03-03 15:57:36 -08001379 os.evsel = counter;
1380 out.ctx = &os;
1381 out.print_metric = print_metric_header;
1382 out.new_line = new_line_metric;
1383 os.evsel = counter;
1384 perf_stat__print_shadow_stats(counter, 0,
1385 0,
1386 &out);
1387 }
1388 fputc('\n', stat_config.output);
1389}
1390
Jiri Olsad4f63a42015-06-26 11:29:26 +02001391static void print_interval(char *prefix, struct timespec *ts)
Ingo Molnar42202dd2009-06-13 14:57:28 +02001392{
Jiri Olsa58215222015-07-21 14:31:24 +02001393 FILE *output = stat_config.output;
Jiri Olsad4f63a42015-06-26 11:29:26 +02001394 static int num_print_interval;
1395
1396 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1397
Andi Kleen41c8ca22016-05-24 12:52:38 -07001398 if (num_print_interval == 0 && !csv_output) {
Jiri Olsa421a50f2015-07-21 14:31:22 +02001399 switch (stat_config.aggr_mode) {
Jiri Olsad4f63a42015-06-26 11:29:26 +02001400 case AGGR_SOCKET:
Andi Kleen41c8ca22016-05-24 12:52:38 -07001401 fprintf(output, "# time socket cpus");
1402 if (!metric_only)
1403 fprintf(output, " counts %*s events\n", unit_width, "unit");
Jiri Olsad4f63a42015-06-26 11:29:26 +02001404 break;
1405 case AGGR_CORE:
Andi Kleen41c8ca22016-05-24 12:52:38 -07001406 fprintf(output, "# time core cpus");
1407 if (!metric_only)
1408 fprintf(output, " counts %*s events\n", unit_width, "unit");
Jiri Olsad4f63a42015-06-26 11:29:26 +02001409 break;
1410 case AGGR_NONE:
Andi Kleen41c8ca22016-05-24 12:52:38 -07001411 fprintf(output, "# time CPU");
1412 if (!metric_only)
1413 fprintf(output, " counts %*s events\n", unit_width, "unit");
Jiri Olsad4f63a42015-06-26 11:29:26 +02001414 break;
Jiri Olsa32b8af82015-06-26 11:29:27 +02001415 case AGGR_THREAD:
Andi Kleen41c8ca22016-05-24 12:52:38 -07001416 fprintf(output, "# time comm-pid");
1417 if (!metric_only)
1418 fprintf(output, " counts %*s events\n", unit_width, "unit");
Jiri Olsa32b8af82015-06-26 11:29:27 +02001419 break;
Jiri Olsad4f63a42015-06-26 11:29:26 +02001420 case AGGR_GLOBAL:
1421 default:
Andi Kleen41c8ca22016-05-24 12:52:38 -07001422 fprintf(output, "# time");
1423 if (!metric_only)
1424 fprintf(output, " counts %*s events\n", unit_width, "unit");
Jiri Olsa208df992015-10-16 12:41:04 +02001425 case AGGR_UNSET:
1426 break;
Jiri Olsad4f63a42015-06-26 11:29:26 +02001427 }
1428 }
1429
Andi Kleen41c8ca22016-05-24 12:52:38 -07001430 if (num_print_interval == 0 && metric_only)
1431 print_metric_headers(" ", true);
Jiri Olsad4f63a42015-06-26 11:29:26 +02001432 if (++num_print_interval == 25)
1433 num_print_interval = 0;
1434}
1435
1436static void print_header(int argc, const char **argv)
1437{
Jiri Olsa58215222015-07-21 14:31:24 +02001438 FILE *output = stat_config.output;
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -02001439 int i;
Ingo Molnar42202dd2009-06-13 14:57:28 +02001440
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001441 fflush(stdout);
1442
Stephane Eraniand7470b62010-12-01 18:49:05 +02001443 if (!csv_output) {
Stephane Eranian4aa90152011-08-15 22:22:33 +02001444 fprintf(output, "\n");
1445 fprintf(output, " Performance counter stats for ");
David Ahern62d3b612013-09-28 14:27:58 -06001446 if (target.system_wide)
1447 fprintf(output, "\'system wide");
1448 else if (target.cpu_list)
1449 fprintf(output, "\'CPU(s) %s", target.cpu_list);
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03001450 else if (!target__has_task(&target)) {
Jiri Olsaba6039b62015-11-05 15:40:55 +01001451 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1452 for (i = 1; argv && (i < argc); i++)
Stephane Eranian4aa90152011-08-15 22:22:33 +02001453 fprintf(output, " %s", argv[i]);
Namhyung Kim20f946b2012-04-26 14:15:16 +09001454 } else if (target.pid)
1455 fprintf(output, "process id \'%s", target.pid);
Stephane Eraniand7470b62010-12-01 18:49:05 +02001456 else
Namhyung Kim20f946b2012-04-26 14:15:16 +09001457 fprintf(output, "thread id \'%s", target.tid);
Ingo Molnar44db76c2009-06-03 19:36:07 +02001458
Stephane Eranian4aa90152011-08-15 22:22:33 +02001459 fprintf(output, "\'");
Stephane Eraniand7470b62010-12-01 18:49:05 +02001460 if (run_count > 1)
Stephane Eranian4aa90152011-08-15 22:22:33 +02001461 fprintf(output, " (%d runs)", run_count);
1462 fprintf(output, ":\n\n");
Stephane Eraniand7470b62010-12-01 18:49:05 +02001463 }
Jiri Olsad4f63a42015-06-26 11:29:26 +02001464}
1465
1466static void print_footer(void)
1467{
Jiri Olsa58215222015-07-21 14:31:24 +02001468 FILE *output = stat_config.output;
1469
Jiri Olsad4f63a42015-06-26 11:29:26 +02001470 if (!null_run)
1471 fprintf(output, "\n");
1472 fprintf(output, " %17.9f seconds time elapsed",
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -03001473 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
Jiri Olsad4f63a42015-06-26 11:29:26 +02001474 if (run_count > 1) {
1475 fprintf(output, " ");
1476 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1477 avg_stats(&walltime_nsecs_stats));
1478 }
1479 fprintf(output, "\n\n");
1480}
1481
1482static void print_counters(struct timespec *ts, int argc, const char **argv)
1483{
Jiri Olsaec0d3d12015-07-21 14:31:25 +02001484 int interval = stat_config.interval;
Jiri Olsad4f63a42015-06-26 11:29:26 +02001485 struct perf_evsel *counter;
1486 char buf[64], *prefix = NULL;
1487
Jiri Olsa664c98d2015-11-05 15:40:50 +01001488 /* Do not print anything if we record to the pipe. */
1489 if (STAT_RECORD && perf_stat.file.is_pipe)
1490 return;
1491
Jiri Olsad4f63a42015-06-26 11:29:26 +02001492 if (interval)
1493 print_interval(prefix = buf, ts);
1494 else
1495 print_header(argc, argv);
Ingo Molnar2996f5d2009-05-29 09:10:54 +02001496
Andi Kleen54b50912016-03-03 15:57:36 -08001497 if (metric_only) {
1498 static int num_print_iv;
1499
Andi Kleen41c8ca22016-05-24 12:52:38 -07001500 if (num_print_iv == 0 && !interval)
1501 print_metric_headers(prefix, false);
Andi Kleen54b50912016-03-03 15:57:36 -08001502 if (num_print_iv++ == 25)
1503 num_print_iv = 0;
1504 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1505 fprintf(stat_config.output, "%s", prefix);
1506 }
1507
Jiri Olsa421a50f2015-07-21 14:31:22 +02001508 switch (stat_config.aggr_mode) {
Stephane Eranian12c08a92013-02-14 13:57:29 +01001509 case AGGR_CORE:
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001510 case AGGR_SOCKET:
Jiri Olsad4f63a42015-06-26 11:29:26 +02001511 print_aggr(prefix);
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001512 break;
Jiri Olsa32b8af82015-06-26 11:29:27 +02001513 case AGGR_THREAD:
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001514 evlist__for_each_entry(evsel_list, counter)
Jiri Olsa32b8af82015-06-26 11:29:27 +02001515 print_aggr_thread(counter, prefix);
1516 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001517 case AGGR_GLOBAL:
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001518 evlist__for_each_entry(evsel_list, counter)
Jiri Olsad4f63a42015-06-26 11:29:26 +02001519 print_counter_aggr(counter, prefix);
Andi Kleen54b50912016-03-03 15:57:36 -08001520 if (metric_only)
1521 fputc('\n', stat_config.output);
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001522 break;
1523 case AGGR_NONE:
Andi Kleen206cab62016-03-03 15:57:37 -08001524 if (metric_only)
1525 print_no_aggr_metric(prefix);
1526 else {
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001527 evlist__for_each_entry(evsel_list, counter)
Andi Kleen206cab62016-03-03 15:57:37 -08001528 print_counter(counter, prefix);
1529 }
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001530 break;
Jiri Olsa208df992015-10-16 12:41:04 +02001531 case AGGR_UNSET:
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001532 default:
1533 break;
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +02001534 }
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001535
Jiri Olsad4f63a42015-06-26 11:29:26 +02001536 if (!interval && !csv_output)
1537 print_footer();
1538
Jiri Olsa58215222015-07-21 14:31:24 +02001539 fflush(stat_config.output);
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001540}
1541
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02001542static volatile int signr = -1;
1543
Ingo Molnar52425192009-05-26 09:17:18 +02001544static void skip_signal(int signo)
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001545{
Jiri Olsaec0d3d12015-07-21 14:31:25 +02001546 if ((child_pid == -1) || stat_config.interval)
Liming Wang60666c62009-12-31 16:05:50 +08001547 done = 1;
1548
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02001549 signr = signo;
Stephane Eraniand07f0b12013-06-04 17:44:26 +02001550 /*
1551 * render child_pid harmless
1552 * won't send SIGTERM to a random
1553 * process in case of race condition
1554 * and fast PID recycling
1555 */
1556 child_pid = -1;
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02001557}
1558
1559static void sig_atexit(void)
1560{
Stephane Eraniand07f0b12013-06-04 17:44:26 +02001561 sigset_t set, oset;
1562
1563 /*
1564 * avoid race condition with SIGCHLD handler
1565 * in skip_signal() which is modifying child_pid
1566 * goal is to avoid send SIGTERM to a random
1567 * process
1568 */
1569 sigemptyset(&set);
1570 sigaddset(&set, SIGCHLD);
1571 sigprocmask(SIG_BLOCK, &set, &oset);
1572
Chris Wilson933da832009-10-04 01:35:01 +01001573 if (child_pid != -1)
1574 kill(child_pid, SIGTERM);
1575
Stephane Eraniand07f0b12013-06-04 17:44:26 +02001576 sigprocmask(SIG_SETMASK, &oset, NULL);
1577
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02001578 if (signr == -1)
1579 return;
1580
1581 signal(signr, SIG_DFL);
1582 kill(getpid(), signr);
Ingo Molnar52425192009-05-26 09:17:18 +02001583}
Ingo Molnarddcacfa2009-04-20 15:37:32 +02001584
Irina Tirdea1d037ca2012-09-11 01:15:03 +03001585static int stat__set_big_num(const struct option *opt __maybe_unused,
1586 const char *s __maybe_unused, int unset)
Stephane Eraniand7470b62010-12-01 18:49:05 +02001587{
1588 big_num_opt = unset ? 0 : 1;
1589 return 0;
1590}
1591
Andi Kleen44b1e602016-05-30 12:49:42 -03001592static int enable_metric_only(const struct option *opt __maybe_unused,
1593 const char *s __maybe_unused, int unset)
1594{
1595 force_metric_only = true;
1596 metric_only = !unset;
1597 return 0;
1598}
1599
Jiri Olsae0547312015-11-05 15:40:45 +01001600static const struct option stat_options[] = {
1601 OPT_BOOLEAN('T', "transaction", &transaction_run,
1602 "hardware transaction statistics"),
1603 OPT_CALLBACK('e', "event", &evsel_list, "event",
1604 "event selector. use 'perf list' to list available events",
1605 parse_events_option),
1606 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1607 "event filter", parse_filter),
1608 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1609 "child tasks do not inherit counters"),
1610 OPT_STRING('p', "pid", &target.pid, "pid",
1611 "stat events on existing process id"),
1612 OPT_STRING('t', "tid", &target.tid, "tid",
1613 "stat events on existing thread id"),
1614 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1615 "system-wide collection from all CPUs"),
1616 OPT_BOOLEAN('g', "group", &group,
1617 "put the counters into a counter group"),
1618 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1619 OPT_INCR('v', "verbose", &verbose,
1620 "be more verbose (show counter open errors, etc)"),
1621 OPT_INTEGER('r', "repeat", &run_count,
1622 "repeat command and print average + stddev (max: 100, forever: 0)"),
1623 OPT_BOOLEAN('n', "null", &null_run,
1624 "null run - dont start any counters"),
1625 OPT_INCR('d', "detailed", &detailed_run,
1626 "detailed run - start a lot of events"),
1627 OPT_BOOLEAN('S', "sync", &sync_run,
1628 "call sync() before starting a run"),
1629 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1630 "print large numbers with thousands\' separators",
1631 stat__set_big_num),
1632 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1633 "list of cpus to monitor in system-wide"),
1634 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1635 "disable CPU count aggregation", AGGR_NONE),
1636 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1637 "print counts with custom separator"),
1638 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1639 "monitor event in cgroup name only", parse_cgroups),
1640 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1641 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1642 OPT_INTEGER(0, "log-fd", &output_fd,
1643 "log output to fd, instead of stderr"),
1644 OPT_STRING(0, "pre", &pre_cmd, "command",
1645 "command to run prior to the measured command"),
1646 OPT_STRING(0, "post", &post_cmd, "command",
1647 "command to run after to the measured command"),
1648 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1649 "print counts at regular interval in ms (>= 10)"),
1650 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1651 "aggregate counts per processor socket", AGGR_SOCKET),
1652 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1653 "aggregate counts per physical processor core", AGGR_CORE),
1654 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1655 "aggregate counts per thread", AGGR_THREAD),
1656 OPT_UINTEGER('D', "delay", &initial_delay,
1657 "ms to wait before starting measurement after program start"),
Andi Kleen44b1e602016-05-30 12:49:42 -03001658 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1659 "Only print computed metrics. No raw values", enable_metric_only),
1660 OPT_BOOLEAN(0, "topdown", &topdown_run,
1661 "measure topdown level 1 statistics"),
Jiri Olsae0547312015-11-05 15:40:45 +01001662 OPT_END()
1663};
1664
Jiri Olsa1fe7a302015-10-16 12:41:15 +02001665static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1666{
1667 return cpu_map__get_socket(map, cpu, NULL);
1668}
1669
1670static int perf_stat__get_core(struct cpu_map *map, int cpu)
1671{
1672 return cpu_map__get_core(map, cpu, NULL);
1673}
1674
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001675static int cpu_map__get_max(struct cpu_map *map)
1676{
1677 int i, max = -1;
1678
1679 for (i = 0; i < map->nr; i++) {
1680 if (map->map[i] > max)
1681 max = map->map[i];
1682 }
1683
1684 return max;
1685}
1686
1687static struct cpu_map *cpus_aggr_map;
1688
1689static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1690{
1691 int cpu;
1692
1693 if (idx >= map->nr)
1694 return -1;
1695
1696 cpu = map->map[idx];
1697
1698 if (cpus_aggr_map->map[cpu] == -1)
1699 cpus_aggr_map->map[cpu] = get_id(map, idx);
1700
1701 return cpus_aggr_map->map[cpu];
1702}
1703
1704static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1705{
1706 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1707}
1708
1709static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1710{
1711 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1712}
1713
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001714static int perf_stat_init_aggr_mode(void)
1715{
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001716 int nr;
1717
Jiri Olsa421a50f2015-07-21 14:31:22 +02001718 switch (stat_config.aggr_mode) {
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001719 case AGGR_SOCKET:
1720 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1721 perror("cannot build socket map");
1722 return -1;
1723 }
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001724 aggr_get_id = perf_stat__get_socket_cached;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001725 break;
Stephane Eranian12c08a92013-02-14 13:57:29 +01001726 case AGGR_CORE:
1727 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1728 perror("cannot build core map");
1729 return -1;
1730 }
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001731 aggr_get_id = perf_stat__get_core_cached;
Stephane Eranian12c08a92013-02-14 13:57:29 +01001732 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001733 case AGGR_NONE:
1734 case AGGR_GLOBAL:
Jiri Olsa32b8af82015-06-26 11:29:27 +02001735 case AGGR_THREAD:
Jiri Olsa208df992015-10-16 12:41:04 +02001736 case AGGR_UNSET:
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001737 default:
1738 break;
1739 }
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001740
1741 /*
1742 * The evsel_list->cpus is the base we operate on,
1743 * taking the highest cpu number to be the size of
1744 * the aggregation translate cpumap.
1745 */
1746 nr = cpu_map__get_max(evsel_list->cpus);
1747 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1748 return cpus_aggr_map ? 0 : -ENOMEM;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001749}
1750
Masami Hiramatsu544c2ae2015-12-09 11:11:27 +09001751static void perf_stat__exit_aggr_mode(void)
1752{
1753 cpu_map__put(aggr_map);
1754 cpu_map__put(cpus_aggr_map);
1755 aggr_map = NULL;
1756 cpus_aggr_map = NULL;
1757}
1758
Jiri Olsa68d702f2015-11-05 15:40:58 +01001759static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1760{
1761 int cpu;
1762
1763 if (idx > map->nr)
1764 return -1;
1765
1766 cpu = map->map[idx];
1767
1768 if (cpu >= env->nr_cpus_online)
1769 return -1;
1770
1771 return cpu;
1772}
1773
1774static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1775{
1776 struct perf_env *env = data;
1777 int cpu = perf_env__get_cpu(env, map, idx);
1778
1779 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1780}
1781
1782static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1783{
1784 struct perf_env *env = data;
1785 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1786
1787 if (cpu != -1) {
1788 int socket_id = env->cpu[cpu].socket_id;
1789
1790 /*
1791 * Encode socket in upper 16 bits
1792 * core_id is relative to socket, and
1793 * we need a global id. So we combine
1794 * socket + core id.
1795 */
1796 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1797 }
1798
1799 return core;
1800}
1801
1802static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1803 struct cpu_map **sockp)
1804{
1805 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1806}
1807
1808static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1809 struct cpu_map **corep)
1810{
1811 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1812}
1813
1814static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1815{
1816 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1817}
1818
1819static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1820{
1821 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1822}
1823
1824static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1825{
1826 struct perf_env *env = &st->session->header.env;
1827
1828 switch (stat_config.aggr_mode) {
1829 case AGGR_SOCKET:
1830 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1831 perror("cannot build socket map");
1832 return -1;
1833 }
1834 aggr_get_id = perf_stat__get_socket_file;
1835 break;
1836 case AGGR_CORE:
1837 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1838 perror("cannot build core map");
1839 return -1;
1840 }
1841 aggr_get_id = perf_stat__get_core_file;
1842 break;
1843 case AGGR_NONE:
1844 case AGGR_GLOBAL:
1845 case AGGR_THREAD:
1846 case AGGR_UNSET:
1847 default:
1848 break;
1849 }
1850
1851 return 0;
1852}
1853
Andi Kleen44b1e602016-05-30 12:49:42 -03001854static int topdown_filter_events(const char **attr, char **str, bool use_group)
1855{
1856 int off = 0;
1857 int i;
1858 int len = 0;
1859 char *s;
1860
1861 for (i = 0; attr[i]; i++) {
1862 if (pmu_have_event("cpu", attr[i])) {
1863 len += strlen(attr[i]) + 1;
1864 attr[i - off] = attr[i];
1865 } else
1866 off++;
1867 }
1868 attr[i - off] = NULL;
1869
1870 *str = malloc(len + 1 + 2);
1871 if (!*str)
1872 return -1;
1873 s = *str;
1874 if (i - off == 0) {
1875 *s = 0;
1876 return 0;
1877 }
1878 if (use_group)
1879 *s++ = '{';
1880 for (i = 0; attr[i]; i++) {
1881 strcpy(s, attr[i]);
1882 s += strlen(s);
1883 *s++ = ',';
1884 }
1885 if (use_group) {
1886 s[-1] = '}';
1887 *s = 0;
1888 } else
1889 s[-1] = 0;
1890 return 0;
1891}
1892
1893__weak bool arch_topdown_check_group(bool *warn)
1894{
1895 *warn = false;
1896 return false;
1897}
1898
1899__weak void arch_topdown_group_warn(void)
1900{
1901}
1902
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001903/*
1904 * Add default attributes, if there were no attributes specified or
1905 * if -d/--detailed, -d -d or -d -d -d is used:
1906 */
1907static int add_default_attributes(void)
1908{
Andi Kleen44b1e602016-05-30 12:49:42 -03001909 int err;
Andi Kleen9dec4472016-02-26 16:27:56 -08001910 struct perf_event_attr default_attrs0[] = {
Arnaldo Carvalho de Melob070a542012-10-01 15:20:58 -03001911
1912 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1913 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1914 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1915 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1916
1917 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
Andi Kleen9dec4472016-02-26 16:27:56 -08001918};
1919 struct perf_event_attr frontend_attrs[] = {
Arnaldo Carvalho de Melob070a542012-10-01 15:20:58 -03001920 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
Andi Kleen9dec4472016-02-26 16:27:56 -08001921};
1922 struct perf_event_attr backend_attrs[] = {
Arnaldo Carvalho de Melob070a542012-10-01 15:20:58 -03001923 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
Andi Kleen9dec4472016-02-26 16:27:56 -08001924};
1925 struct perf_event_attr default_attrs1[] = {
Arnaldo Carvalho de Melob070a542012-10-01 15:20:58 -03001926 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1927 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1928 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1929
1930};
1931
1932/*
1933 * Detailed stats (-d), covering the L1 and last level data caches:
1934 */
1935 struct perf_event_attr detailed_attrs[] = {
1936
1937 { .type = PERF_TYPE_HW_CACHE,
1938 .config =
1939 PERF_COUNT_HW_CACHE_L1D << 0 |
1940 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1941 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1942
1943 { .type = PERF_TYPE_HW_CACHE,
1944 .config =
1945 PERF_COUNT_HW_CACHE_L1D << 0 |
1946 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1947 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1948
1949 { .type = PERF_TYPE_HW_CACHE,
1950 .config =
1951 PERF_COUNT_HW_CACHE_LL << 0 |
1952 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1953 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1954
1955 { .type = PERF_TYPE_HW_CACHE,
1956 .config =
1957 PERF_COUNT_HW_CACHE_LL << 0 |
1958 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1959 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1960};
1961
1962/*
1963 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1964 */
1965 struct perf_event_attr very_detailed_attrs[] = {
1966
1967 { .type = PERF_TYPE_HW_CACHE,
1968 .config =
1969 PERF_COUNT_HW_CACHE_L1I << 0 |
1970 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1971 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1972
1973 { .type = PERF_TYPE_HW_CACHE,
1974 .config =
1975 PERF_COUNT_HW_CACHE_L1I << 0 |
1976 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1977 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1978
1979 { .type = PERF_TYPE_HW_CACHE,
1980 .config =
1981 PERF_COUNT_HW_CACHE_DTLB << 0 |
1982 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1983 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1984
1985 { .type = PERF_TYPE_HW_CACHE,
1986 .config =
1987 PERF_COUNT_HW_CACHE_DTLB << 0 |
1988 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1989 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1990
1991 { .type = PERF_TYPE_HW_CACHE,
1992 .config =
1993 PERF_COUNT_HW_CACHE_ITLB << 0 |
1994 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1995 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1996
1997 { .type = PERF_TYPE_HW_CACHE,
1998 .config =
1999 PERF_COUNT_HW_CACHE_ITLB << 0 |
2000 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2001 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2002
2003};
2004
2005/*
2006 * Very, very detailed stats (-d -d -d), adding prefetch events:
2007 */
2008 struct perf_event_attr very_very_detailed_attrs[] = {
2009
2010 { .type = PERF_TYPE_HW_CACHE,
2011 .config =
2012 PERF_COUNT_HW_CACHE_L1D << 0 |
2013 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2014 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2015
2016 { .type = PERF_TYPE_HW_CACHE,
2017 .config =
2018 PERF_COUNT_HW_CACHE_L1D << 0 |
2019 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2020 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2021};
2022
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002023 /* Set attrs if no event is selected and !null_run: */
2024 if (null_run)
2025 return 0;
2026
Andi Kleen4cabc3d2013-08-21 16:47:26 -07002027 if (transaction_run) {
Andi Kleen4cabc3d2013-08-21 16:47:26 -07002028 if (pmu_have_event("cpu", "cycles-ct") &&
2029 pmu_have_event("cpu", "el-start"))
Jiri Olsaa4547422015-06-03 16:25:53 +02002030 err = parse_events(evsel_list, transaction_attrs, NULL);
Andi Kleen4cabc3d2013-08-21 16:47:26 -07002031 else
Jiri Olsaa4547422015-06-03 16:25:53 +02002032 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2033 if (err) {
Andi Kleen4cabc3d2013-08-21 16:47:26 -07002034 fprintf(stderr, "Cannot set up transaction events\n");
2035 return -1;
2036 }
2037 return 0;
2038 }
2039
Andi Kleen44b1e602016-05-30 12:49:42 -03002040 if (topdown_run) {
2041 char *str = NULL;
2042 bool warn = false;
2043
2044 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2045 stat_config.aggr_mode != AGGR_CORE) {
2046 pr_err("top down event configuration requires --per-core mode\n");
2047 return -1;
2048 }
2049 stat_config.aggr_mode = AGGR_CORE;
2050 if (nr_cgroups || !target__has_cpu(&target)) {
2051 pr_err("top down event configuration requires system-wide mode (-a)\n");
2052 return -1;
2053 }
2054
2055 if (!force_metric_only)
2056 metric_only = true;
2057 if (topdown_filter_events(topdown_attrs, &str,
2058 arch_topdown_check_group(&warn)) < 0) {
2059 pr_err("Out of memory\n");
2060 return -1;
2061 }
2062 if (topdown_attrs[0] && str) {
2063 if (warn)
2064 arch_topdown_group_warn();
2065 err = parse_events(evsel_list, str, NULL);
2066 if (err) {
2067 fprintf(stderr,
2068 "Cannot set up top down events %s: %d\n",
2069 str, err);
2070 free(str);
2071 return -1;
2072 }
2073 } else {
2074 fprintf(stderr, "System does not support topdown\n");
2075 return -1;
2076 }
2077 free(str);
2078 }
2079
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002080 if (!evsel_list->nr_entries) {
Namhyung Kima1f3d562016-05-13 15:01:03 +09002081 if (target__has_cpu(&target))
2082 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2083
Andi Kleen9dec4472016-02-26 16:27:56 -08002084 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2085 return -1;
2086 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2087 if (perf_evlist__add_default_attrs(evsel_list,
2088 frontend_attrs) < 0)
2089 return -1;
2090 }
2091 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2092 if (perf_evlist__add_default_attrs(evsel_list,
2093 backend_attrs) < 0)
2094 return -1;
2095 }
2096 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02002097 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002098 }
2099
2100 /* Detailed events get appended to the event list: */
2101
2102 if (detailed_run < 1)
2103 return 0;
2104
2105 /* Append detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03002106 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02002107 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002108
2109 if (detailed_run < 2)
2110 return 0;
2111
2112 /* Append very detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03002113 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02002114 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002115
2116 if (detailed_run < 3)
2117 return 0;
2118
2119 /* Append very, very detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03002120 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002121}
2122
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01002123static const char * const stat_record_usage[] = {
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002124 "perf stat record [<options>]",
2125 NULL,
2126};
2127
Jiri Olsa3ba78bd2015-11-05 15:40:47 +01002128static void init_features(struct perf_session *session)
2129{
2130 int feat;
2131
2132 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2133 perf_header__set_feat(&session->header, feat);
2134
2135 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2136 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2137 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2138 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2139}
2140
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002141static int __cmd_record(int argc, const char **argv)
2142{
2143 struct perf_session *session;
2144 struct perf_data_file *file = &perf_stat.file;
2145
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01002146 argc = parse_options(argc, argv, stat_options, stat_record_usage,
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002147 PARSE_OPT_STOP_AT_NON_OPTION);
2148
2149 if (output_name)
2150 file->path = output_name;
2151
Jiri Olsae9d6db8e82015-11-05 15:40:53 +01002152 if (run_count != 1 || forever) {
2153 pr_err("Cannot use -r option with perf stat record.\n");
2154 return -1;
2155 }
2156
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002157 session = perf_session__new(file, false, NULL);
2158 if (session == NULL) {
2159 pr_err("Perf session creation failed.\n");
2160 return -1;
2161 }
2162
Jiri Olsa3ba78bd2015-11-05 15:40:47 +01002163 init_features(session);
2164
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002165 session->evlist = evsel_list;
2166 perf_stat.session = session;
2167 perf_stat.record = true;
2168 return argc;
2169}
2170
Jiri Olsaa56f9392015-11-05 15:40:59 +01002171static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2172 union perf_event *event,
2173 struct perf_session *session)
2174{
Andi Kleene3b03b62016-05-05 16:04:03 -07002175 struct stat_round_event *stat_round = &event->stat_round;
Jiri Olsaa56f9392015-11-05 15:40:59 +01002176 struct perf_evsel *counter;
2177 struct timespec tsh, *ts = NULL;
2178 const char **argv = session->header.env.cmdline_argv;
2179 int argc = session->header.env.nr_cmdline;
2180
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03002181 evlist__for_each_entry(evsel_list, counter)
Jiri Olsaa56f9392015-11-05 15:40:59 +01002182 perf_stat_process_counter(&stat_config, counter);
2183
Andi Kleene3b03b62016-05-05 16:04:03 -07002184 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2185 update_stats(&walltime_nsecs_stats, stat_round->time);
Jiri Olsaa56f9392015-11-05 15:40:59 +01002186
Andi Kleene3b03b62016-05-05 16:04:03 -07002187 if (stat_config.interval && stat_round->time) {
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -03002188 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2189 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
Jiri Olsaa56f9392015-11-05 15:40:59 +01002190 ts = &tsh;
2191 }
2192
2193 print_counters(ts, argc, argv);
2194 return 0;
2195}
2196
Jiri Olsa62ba18b2015-11-05 15:40:57 +01002197static
2198int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2199 union perf_event *event,
2200 struct perf_session *session __maybe_unused)
2201{
Jiri Olsa68d702f2015-11-05 15:40:58 +01002202 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2203
Jiri Olsa62ba18b2015-11-05 15:40:57 +01002204 perf_event__read_stat_config(&stat_config, &event->stat_config);
Jiri Olsa68d702f2015-11-05 15:40:58 +01002205
Jiri Olsa89af4e02015-11-05 15:41:02 +01002206 if (cpu_map__empty(st->cpus)) {
2207 if (st->aggr_mode != AGGR_UNSET)
2208 pr_warning("warning: processing task data, aggregation mode not set\n");
2209 return 0;
2210 }
2211
2212 if (st->aggr_mode != AGGR_UNSET)
2213 stat_config.aggr_mode = st->aggr_mode;
2214
Jiri Olsa68d702f2015-11-05 15:40:58 +01002215 if (perf_stat.file.is_pipe)
2216 perf_stat_init_aggr_mode();
2217 else
2218 perf_stat_init_aggr_mode_file(st);
2219
Jiri Olsa62ba18b2015-11-05 15:40:57 +01002220 return 0;
2221}
2222
Jiri Olsa1975d362015-11-05 15:40:56 +01002223static int set_maps(struct perf_stat *st)
2224{
2225 if (!st->cpus || !st->threads)
2226 return 0;
2227
2228 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2229 return -EINVAL;
2230
2231 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2232
2233 if (perf_evlist__alloc_stats(evsel_list, true))
2234 return -ENOMEM;
2235
2236 st->maps_allocated = true;
2237 return 0;
2238}
2239
2240static
2241int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2242 union perf_event *event,
2243 struct perf_session *session __maybe_unused)
2244{
2245 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2246
2247 if (st->threads) {
2248 pr_warning("Extra thread map event, ignoring.\n");
2249 return 0;
2250 }
2251
2252 st->threads = thread_map__new_event(&event->thread_map);
2253 if (!st->threads)
2254 return -ENOMEM;
2255
2256 return set_maps(st);
2257}
2258
2259static
2260int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2261 union perf_event *event,
2262 struct perf_session *session __maybe_unused)
2263{
2264 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2265 struct cpu_map *cpus;
2266
2267 if (st->cpus) {
2268 pr_warning("Extra cpu map event, ignoring.\n");
2269 return 0;
2270 }
2271
2272 cpus = cpu_map__new_data(&event->cpu_map.data);
2273 if (!cpus)
2274 return -ENOMEM;
2275
2276 st->cpus = cpus;
2277 return set_maps(st);
2278}
2279
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01002280static const char * const stat_report_usage[] = {
Jiri Olsaba6039b62015-11-05 15:40:55 +01002281 "perf stat report [<options>]",
2282 NULL,
2283};
2284
2285static struct perf_stat perf_stat = {
2286 .tool = {
2287 .attr = perf_event__process_attr,
Jiri Olsafa6ea782015-11-05 15:41:00 +01002288 .event_update = perf_event__process_event_update,
Jiri Olsa1975d362015-11-05 15:40:56 +01002289 .thread_map = process_thread_map_event,
2290 .cpu_map = process_cpu_map_event,
Jiri Olsa62ba18b2015-11-05 15:40:57 +01002291 .stat_config = process_stat_config_event,
Jiri Olsaa56f9392015-11-05 15:40:59 +01002292 .stat = perf_event__process_stat_event,
2293 .stat_round = process_stat_round_event,
Jiri Olsaba6039b62015-11-05 15:40:55 +01002294 },
Jiri Olsa89af4e02015-11-05 15:41:02 +01002295 .aggr_mode = AGGR_UNSET,
Jiri Olsaba6039b62015-11-05 15:40:55 +01002296};
2297
2298static int __cmd_report(int argc, const char **argv)
2299{
2300 struct perf_session *session;
2301 const struct option options[] = {
2302 OPT_STRING('i', "input", &input_name, "file", "input file name"),
Jiri Olsa89af4e02015-11-05 15:41:02 +01002303 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2304 "aggregate counts per processor socket", AGGR_SOCKET),
2305 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2306 "aggregate counts per physical processor core", AGGR_CORE),
2307 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2308 "disable CPU count aggregation", AGGR_NONE),
Jiri Olsaba6039b62015-11-05 15:40:55 +01002309 OPT_END()
2310 };
2311 struct stat st;
2312 int ret;
2313
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01002314 argc = parse_options(argc, argv, options, stat_report_usage, 0);
Jiri Olsaba6039b62015-11-05 15:40:55 +01002315
2316 if (!input_name || !strlen(input_name)) {
2317 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2318 input_name = "-";
2319 else
2320 input_name = "perf.data";
2321 }
2322
2323 perf_stat.file.path = input_name;
2324 perf_stat.file.mode = PERF_DATA_MODE_READ;
2325
2326 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2327 if (session == NULL)
2328 return -1;
2329
2330 perf_stat.session = session;
2331 stat_config.output = stderr;
2332 evsel_list = session->evlist;
2333
2334 ret = perf_session__process_events(session);
2335 if (ret)
2336 return ret;
2337
2338 perf_session__delete(session);
2339 return 0;
2340}
2341
Irina Tirdea1d037ca2012-09-11 01:15:03 +03002342int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
Ingo Molnar52425192009-05-26 09:17:18 +02002343{
Arnaldo Carvalho de Melob070a542012-10-01 15:20:58 -03002344 const char * const stat_usage[] = {
2345 "perf stat [<options>] [<command>]",
2346 NULL
2347 };
Namhyung Kimcc03c542013-11-01 16:33:15 +09002348 int status = -EINVAL, run_idx;
Stephane Eranian4aa90152011-08-15 22:22:33 +02002349 const char *mode;
Jiri Olsa58215222015-07-21 14:31:24 +02002350 FILE *output = stderr;
Jiri Olsaec0d3d12015-07-21 14:31:25 +02002351 unsigned int interval;
Jiri Olsaba6039b62015-11-05 15:40:55 +01002352 const char * const stat_subcommands[] = { "record", "report" };
Ingo Molnar42202dd2009-06-13 14:57:28 +02002353
Stephane Eranian5af52b52010-05-18 15:00:01 +02002354 setlocale(LC_ALL, "");
2355
Namhyung Kim334fe7a2013-03-11 16:43:12 +09002356 evsel_list = perf_evlist__new();
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -02002357 if (evsel_list == NULL)
2358 return -ENOMEM;
2359
Wang Nan1669e502016-02-19 11:43:58 +00002360 parse_events__shrink_config_terms();
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002361 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2362 (const char **) stat_usage,
2363 PARSE_OPT_STOP_AT_NON_OPTION);
Andi Kleenfb4605b2016-03-01 10:57:52 -08002364 perf_stat__init_shadow_stats();
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002365
Jiri Olsa6edb78a2015-11-05 15:41:01 +01002366 if (csv_sep) {
2367 csv_output = true;
2368 if (!strcmp(csv_sep, "\\t"))
2369 csv_sep = "\t";
2370 } else
2371 csv_sep = DEFAULT_SEPARATOR;
2372
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002373 if (argc && !strncmp(argv[0], "rec", 3)) {
2374 argc = __cmd_record(argc, argv);
2375 if (argc < 0)
2376 return -1;
Jiri Olsaba6039b62015-11-05 15:40:55 +01002377 } else if (argc && !strncmp(argv[0], "rep", 3))
2378 return __cmd_report(argc, argv);
Stephane Eraniand7470b62010-12-01 18:49:05 +02002379
Jiri Olsaec0d3d12015-07-21 14:31:25 +02002380 interval = stat_config.interval;
2381
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002382 /*
2383 * For record command the -o is already taken care of.
2384 */
2385 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
Stephane Eranian4aa90152011-08-15 22:22:33 +02002386 output = NULL;
2387
Jim Cromie56f3bae2011-09-07 17:14:00 -06002388 if (output_name && output_fd) {
2389 fprintf(stderr, "cannot use both --output and --log-fd\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002390 parse_options_usage(stat_usage, stat_options, "o", 1);
2391 parse_options_usage(NULL, stat_options, "log-fd", 0);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002392 goto out;
Jim Cromie56f3bae2011-09-07 17:14:00 -06002393 }
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02002394
Andi Kleen54b50912016-03-03 15:57:36 -08002395 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2396 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2397 goto out;
2398 }
2399
Andi Kleen54b50912016-03-03 15:57:36 -08002400 if (metric_only && run_count > 1) {
2401 fprintf(stderr, "--metric-only is not supported with -r\n");
2402 goto out;
2403 }
2404
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02002405 if (output_fd < 0) {
2406 fprintf(stderr, "argument to --log-fd must be a > 0\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002407 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002408 goto out;
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02002409 }
2410
Stephane Eranian4aa90152011-08-15 22:22:33 +02002411 if (!output) {
2412 struct timespec tm;
2413 mode = append_file ? "a" : "w";
2414
2415 output = fopen(output_name, mode);
2416 if (!output) {
2417 perror("failed to create output file");
David Ahernfceda7f2012-08-26 12:24:44 -06002418 return -1;
Stephane Eranian4aa90152011-08-15 22:22:33 +02002419 }
2420 clock_gettime(CLOCK_REALTIME, &tm);
2421 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02002422 } else if (output_fd > 0) {
Jim Cromie56f3bae2011-09-07 17:14:00 -06002423 mode = append_file ? "a" : "w";
2424 output = fdopen(output_fd, mode);
2425 if (!output) {
2426 perror("Failed opening logfd");
2427 return -errno;
2428 }
Stephane Eranian4aa90152011-08-15 22:22:33 +02002429 }
2430
Jiri Olsa58215222015-07-21 14:31:24 +02002431 stat_config.output = output;
2432
Stephane Eraniand7470b62010-12-01 18:49:05 +02002433 /*
2434 * let the spreadsheet do the pretty-printing
2435 */
2436 if (csv_output) {
Jim Cromie61a9f322011-09-07 17:14:04 -06002437 /* User explicitly passed -B? */
Stephane Eraniand7470b62010-12-01 18:49:05 +02002438 if (big_num_opt == 1) {
2439 fprintf(stderr, "-B option not supported with -x\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002440 parse_options_usage(stat_usage, stat_options, "B", 1);
2441 parse_options_usage(NULL, stat_options, "x", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002442 goto out;
Stephane Eraniand7470b62010-12-01 18:49:05 +02002443 } else /* Nope, so disable big number formatting */
2444 big_num = false;
2445 } else if (big_num_opt == 0) /* User passed --no-big-num */
2446 big_num = false;
2447
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002448 if (!argc && target__none(&target))
Jiri Olsae0547312015-11-05 15:40:45 +01002449 usage_with_options(stat_usage, stat_options);
David Ahernac3063b2013-09-30 07:37:37 -06002450
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002451 if (run_count < 0) {
Namhyung Kimcc03c542013-11-01 16:33:15 +09002452 pr_err("Run count must be a positive number\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002453 parse_options_usage(stat_usage, stat_options, "r", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002454 goto out;
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002455 } else if (run_count == 0) {
2456 forever = true;
2457 run_count = 1;
2458 }
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002459
Jiri Olsa421a50f2015-07-21 14:31:22 +02002460 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
Jiri Olsa32b8af82015-06-26 11:29:27 +02002461 fprintf(stderr, "The --per-thread option is only available "
2462 "when monitoring via -p -t options.\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002463 parse_options_usage(NULL, stat_options, "p", 1);
2464 parse_options_usage(NULL, stat_options, "t", 1);
Jiri Olsa32b8af82015-06-26 11:29:27 +02002465 goto out;
2466 }
2467
2468 /*
2469 * no_aggr, cgroup are for system-wide only
2470 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2471 */
Jiri Olsa421a50f2015-07-21 14:31:22 +02002472 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2473 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002474 !target__has_cpu(&target)) {
Stephane Eranian023695d2011-02-14 11:20:01 +02002475 fprintf(stderr, "both cgroup and no-aggregation "
2476 "modes only available in system-wide mode\n");
2477
Jiri Olsae0547312015-11-05 15:40:45 +01002478 parse_options_usage(stat_usage, stat_options, "G", 1);
2479 parse_options_usage(NULL, stat_options, "A", 1);
2480 parse_options_usage(NULL, stat_options, "a", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002481 goto out;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01002482 }
2483
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002484 if (add_default_attributes())
2485 goto out;
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002486
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002487 target__validate(&target);
Arnaldo Carvalho de Melo5c98d4662011-01-03 17:53:33 -02002488
Namhyung Kim77a6f012012-05-07 14:09:04 +09002489 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002490 if (target__has_task(&target)) {
Namhyung Kim77a6f012012-05-07 14:09:04 +09002491 pr_err("Problems finding threads of monitor\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002492 parse_options_usage(stat_usage, stat_options, "p", 1);
2493 parse_options_usage(NULL, stat_options, "t", 1);
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002494 } else if (target__has_cpu(&target)) {
Namhyung Kim77a6f012012-05-07 14:09:04 +09002495 perror("failed to parse CPUs map");
Jiri Olsae0547312015-11-05 15:40:45 +01002496 parse_options_usage(stat_usage, stat_options, "C", 1);
2497 parse_options_usage(NULL, stat_options, "a", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002498 }
2499 goto out;
Arnaldo Carvalho de Melo60d567e2011-01-03 17:49:48 -02002500 }
Jiri Olsa32b8af82015-06-26 11:29:27 +02002501
2502 /*
2503 * Initialize thread_map with comm names,
2504 * so we could print it out on output.
2505 */
Jiri Olsa421a50f2015-07-21 14:31:22 +02002506 if (stat_config.aggr_mode == AGGR_THREAD)
Jiri Olsa32b8af82015-06-26 11:29:27 +02002507 thread_map__read_comms(evsel_list->threads);
2508
Stephane Eranian13370a92013-01-29 12:47:44 +01002509 if (interval && interval < 100) {
Kan Liang19afd102015-10-02 05:04:34 -04002510 if (interval < 10) {
2511 pr_err("print interval must be >= 10ms\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002512 parse_options_usage(stat_usage, stat_options, "I", 1);
Kan Liang19afd102015-10-02 05:04:34 -04002513 goto out;
2514 } else
2515 pr_warning("print interval < 100ms. "
2516 "The overhead percentage could be high in some cases. "
2517 "Please proceed with caution.\n");
Stephane Eranian13370a92013-01-29 12:47:44 +01002518 }
Stephane Eranianc45c6ea2010-05-28 12:00:01 +02002519
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002520 if (perf_evlist__alloc_stats(evsel_list, interval))
Arnaldo Carvalho de Melo03ad9742014-01-03 15:56:06 -03002521 goto out;
Zhang, Yanmind6d901c2010-03-18 11:36:05 -03002522
Stephane Eranian86ee6e12013-02-14 13:57:27 +01002523 if (perf_stat_init_aggr_mode())
Arnaldo Carvalho de Melo03ad9742014-01-03 15:56:06 -03002524 goto out;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01002525
Ingo Molnar58d7e992009-05-15 11:03:23 +02002526 /*
2527 * We dont want to block the signals - that would cause
2528 * child tasks to inherit that and Ctrl-C would not work.
2529 * What we want is for Ctrl-C to work in the exec()-ed
2530 * task, but being ignored by perf stat itself:
2531 */
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02002532 atexit(sig_atexit);
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002533 if (!forever)
2534 signal(SIGINT, skip_signal);
Stephane Eranian13370a92013-01-29 12:47:44 +01002535 signal(SIGCHLD, skip_signal);
Ingo Molnar58d7e992009-05-15 11:03:23 +02002536 signal(SIGALRM, skip_signal);
2537 signal(SIGABRT, skip_signal);
2538
Ingo Molnar42202dd2009-06-13 14:57:28 +02002539 status = 0;
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002540 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
Ingo Molnar42202dd2009-06-13 14:57:28 +02002541 if (run_count != 1 && verbose)
Stephane Eranian4aa90152011-08-15 22:22:33 +02002542 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2543 run_idx + 1);
Ingo Molnarf9cef0a2011-04-28 18:17:11 +02002544
Ingo Molnar42202dd2009-06-13 14:57:28 +02002545 status = run_perf_stat(argc, argv);
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002546 if (forever && status != -1) {
Jiri Olsad4f63a42015-06-26 11:29:26 +02002547 print_counters(NULL, argc, argv);
Jiri Olsa254ecbc2015-06-26 11:29:13 +02002548 perf_stat__reset_stats();
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002549 }
Ingo Molnar42202dd2009-06-13 14:57:28 +02002550 }
2551
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002552 if (!forever && status != -1 && !interval)
Jiri Olsad4f63a42015-06-26 11:29:26 +02002553 print_counters(NULL, argc, argv);
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002554
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002555 if (STAT_RECORD) {
2556 /*
2557 * We synthesize the kernel mmap record just so that older tools
2558 * don't emit warnings about not being able to resolve symbols
2559 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2560 * a saner message about no samples being in the perf.data file.
2561 *
2562 * This also serves to suppress a warning about f_header.data.size == 0
Jiri Olsa8b99b1a2015-11-05 15:40:48 +01002563 * in header.c at the moment 'perf stat record' gets introduced, which
2564 * is not really needed once we start adding the stat specific PERF_RECORD_
2565 * records, but the need to suppress the kptr_restrict messages in older
2566 * tools remain -acme
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002567 */
2568 int fd = perf_data_file__fd(&perf_stat.file);
2569 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2570 process_synthesized_event,
2571 &perf_stat.session->machines.host);
2572 if (err) {
2573 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2574 "older tools may produce warnings about this file\n.");
2575 }
2576
Jiri Olsa7aad0c32015-11-05 15:40:52 +01002577 if (!interval) {
2578 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2579 pr_err("failed to write stat round event\n");
2580 }
2581
Jiri Olsa664c98d2015-11-05 15:40:50 +01002582 if (!perf_stat.file.is_pipe) {
2583 perf_stat.session->header.data_size += perf_stat.bytes_written;
2584 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2585 }
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002586
2587 perf_session__delete(perf_stat.session);
2588 }
2589
Masami Hiramatsu544c2ae2015-12-09 11:11:27 +09002590 perf_stat__exit_aggr_mode();
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002591 perf_evlist__free_stats(evsel_list);
Arnaldo Carvalho de Melo0015e2e2011-02-01 16:18:10 -02002592out:
2593 perf_evlist__delete(evsel_list);
Ingo Molnar42202dd2009-06-13 14:57:28 +02002594 return status;
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002595}