blob: ec053dc1e35c82cede6708175d899f0ea7a3831d [file] [log] [blame]
Thomas Gleixner91007042019-05-29 07:12:25 -07001// SPDX-License-Identifier: GPL-2.0-only
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002/*
Ingo Molnarbf9e1872009-06-02 23:37:05 +02003 * builtin-stat.c
4 *
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
7 *
8 * Sample output:
Ingo Molnarddcacfa2009-04-20 15:37:32 +02009
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020010 $ perf stat ./hackbench 10
Ingo Molnarddcacfa2009-04-20 15:37:32 +020011
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020012 Time: 0.118
Ingo Molnarddcacfa2009-04-20 15:37:32 +020013
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020014 Performance counter stats for './hackbench 10':
Ingo Molnarddcacfa2009-04-20 15:37:32 +020015
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020016 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
27
28 0.154822978 seconds time elapsed
Ingo Molnarddcacfa2009-04-20 15:37:32 +020029
Ingo Molnar52425192009-05-26 09:17:18 +020030 *
Ingo Molnar2cba3ff2011-05-19 13:30:56 +020031 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
Ingo Molnar52425192009-05-26 09:17:18 +020032 *
33 * Improvements and fixes by:
34 *
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
Jaswinder Singh Rajput6e750a8f2009-06-27 03:02:07 +053040 * Jaswinder Singh Rajput <jaswinder@kernel.org>
Ingo Molnarddcacfa2009-04-20 15:37:32 +020041 */
42
Ingo Molnar16f762a2009-05-27 09:10:38 +020043#include "builtin.h"
Arnaldo Carvalho de Meloc1a604d2019-08-29 15:20:59 -030044#include "perf.h"
Arnaldo Carvalho de Melof14d5702014-10-17 12:17:40 -030045#include "util/cgroup.h"
Josh Poimboeuf4b6ab942015-12-15 09:39:39 -060046#include <subcmd/parse-options.h>
Ingo Molnar52425192009-05-26 09:17:18 +020047#include "util/parse-events.h"
Andi Kleen4cabc3d2013-08-21 16:47:26 -070048#include "util/pmu.h"
Frederic Weisbecker8f288272009-08-16 22:05:48 +020049#include "util/event.h"
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -020050#include "util/evlist.h"
Arnaldo Carvalho de Melo69aad6f2011-01-03 16:39:04 -020051#include "util/evsel.h"
Frederic Weisbecker8f288272009-08-16 22:05:48 +020052#include "util/debug.h"
Ingo Molnara5d243d2011-04-27 05:39:24 +020053#include "util/color.h"
Xiao Guangrong0007ece2012-09-17 16:31:14 +080054#include "util/stat.h"
Liming Wang60666c62009-12-31 16:05:50 +080055#include "util/header.h"
Paul Mackerrasa12b51c2010-03-10 20:36:09 +110056#include "util/cpumap.h"
Arnaldo Carvalho de Melofd782602011-01-18 15:15:24 -020057#include "util/thread_map.h"
Jiri Olsad8095602015-08-07 12:51:03 +020058#include "util/counts.h"
Andi Kleen44b1e602016-05-30 12:49:42 -030059#include "util/group.h"
Jiri Olsa4979d0c2015-11-05 15:40:46 +010060#include "util/session.h"
Jiri Olsaba6039b62015-11-05 15:40:55 +010061#include "util/tool.h"
Arnaldo Carvalho de Meloa0675582017-04-17 16:51:59 -030062#include "util/string2.h"
Andi Kleenb18f3e32017-08-31 12:40:31 -070063#include "util/metricgroup.h"
Arnaldo Carvalho de Meloea49e012019-09-18 11:36:13 -030064#include "util/synthetic-events.h"
Arnaldo Carvalho de Meloaeb00b12019-08-22 15:40:29 -030065#include "util/target.h"
Arnaldo Carvalho de Melof3711022019-08-29 15:16:27 -030066#include "util/time-utils.h"
Jiri Olsa9660e082018-06-07 00:15:06 +020067#include "util/top.h"
Andi Kleen4804e012019-11-20 16:15:19 -080068#include "util/affinity.h"
Jiri Olsaba6039b62015-11-05 15:40:55 +010069#include "asm/bug.h"
Ingo Molnarddcacfa2009-04-20 15:37:32 +020070
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -030071#include <linux/time64.h>
Arnaldo Carvalho de Melo7f7c5362019-07-04 11:32:27 -030072#include <linux/zalloc.h>
Andi Kleen44b1e602016-05-30 12:49:42 -030073#include <api/fs/fs.h>
Arnaldo Carvalho de Meloa43783a2017-04-18 10:46:11 -030074#include <errno.h>
Arnaldo Carvalho de Melo9607ad32017-04-19 15:49:18 -030075#include <signal.h>
Peter Zijlstra1f16c572012-10-23 13:40:14 +020076#include <stdlib.h>
Ingo Molnarddcacfa2009-04-20 15:37:32 +020077#include <sys/prctl.h>
Arnaldo Carvalho de Melofd20e812017-04-17 15:23:08 -030078#include <inttypes.h>
Stephane Eranian5af52b52010-05-18 15:00:01 +020079#include <locale.h>
Andi Kleene3b03b62016-05-05 16:04:03 -070080#include <math.h>
Arnaldo Carvalho de Melo7a8ef4c2017-04-19 20:57:47 -030081#include <sys/types.h>
82#include <sys/stat.h>
Arnaldo Carvalho de Melo42087352017-04-19 19:06:30 -030083#include <sys/wait.h>
Arnaldo Carvalho de Melo7a8ef4c2017-04-19 20:57:47 -030084#include <unistd.h>
Jiri Olsa0ce2da12018-06-05 14:13:13 +020085#include <sys/time.h>
86#include <sys/resource.h>
Mamatha Inamdar6ef81c52019-08-22 12:50:49 +053087#include <linux/err.h>
Peter Zijlstra16c8a102009-05-05 17:50:27 +020088
Arnaldo Carvalho de Melo3052ba52019-06-25 17:27:31 -030089#include <linux/ctype.h>
Jiri Olsa453fa032019-07-21 13:24:43 +020090#include <perf/evlist.h>
Arnaldo Carvalho de Melo3d689ed2017-04-17 16:10:49 -030091
Stephane Eraniand7470b62010-12-01 18:49:05 +020092#define DEFAULT_SEPARATOR " "
Kan Liangdaefd0b2017-05-26 12:05:38 -070093#define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
Stephane Eraniand7470b62010-12-01 18:49:05 +020094
Jiri Olsad4f63a42015-06-26 11:29:26 +020095static void print_counters(struct timespec *ts, int argc, const char **argv);
Stephane Eranian13370a92013-01-29 12:47:44 +010096
Andi Kleen4cabc3d2013-08-21 16:47:26 -070097/* Default events used for perf stat -T */
Jiri Olsaa4547422015-06-03 16:25:53 +020098static const char *transaction_attrs = {
99 "task-clock,"
Andi Kleen4cabc3d2013-08-21 16:47:26 -0700100 "{"
101 "instructions,"
102 "cycles,"
103 "cpu/cycles-t/,"
104 "cpu/tx-start/,"
105 "cpu/el-start/,"
106 "cpu/cycles-ct/"
107 "}"
108};
109
110/* More limited version when the CPU does not have all events. */
Jiri Olsaa4547422015-06-03 16:25:53 +0200111static const char * transaction_limited_attrs = {
112 "task-clock,"
Andi Kleen4cabc3d2013-08-21 16:47:26 -0700113 "{"
114 "instructions,"
115 "cycles,"
116 "cpu/cycles-t/,"
117 "cpu/tx-start/"
118 "}"
119};
120
Andi Kleen44b1e602016-05-30 12:49:42 -0300121static const char * topdown_attrs[] = {
122 "topdown-total-slots",
123 "topdown-slots-retired",
124 "topdown-recovery-bubbles",
125 "topdown-fetch-bubbles",
126 "topdown-slots-issued",
127 NULL,
128};
129
Kan Liangdaefd0b2017-05-26 12:05:38 -0700130static const char *smi_cost_attrs = {
131 "{"
132 "msr/aperf/,"
133 "msr/smi/,"
134 "cycles"
135 "}"
136};
137
Jiri Olsa63503db2019-07-21 13:23:52 +0200138static struct evlist *evsel_list;
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -0200139
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -0300140static struct target target = {
Namhyung Kim77a6f012012-05-07 14:09:04 +0900141 .uid = UINT_MAX,
142};
Jaswinder Singh Rajput3d632592009-06-24 18:19:34 +0530143
Jiri Olsac1a1f5d2018-06-07 00:15:09 +0200144#define METRIC_ONLY_LEN 20
145
Stephane Eraniand07f0b12013-06-04 17:44:26 +0200146static volatile pid_t child_pid = -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +0200147static int detailed_run = 0;
Andi Kleen4cabc3d2013-08-21 16:47:26 -0700148static bool transaction_run;
Andi Kleen44b1e602016-05-30 12:49:42 -0300149static bool topdown_run = false;
Kan Liangdaefd0b2017-05-26 12:05:38 -0700150static bool smi_cost = false;
151static bool smi_reset = false;
Stephane Eraniand7470b62010-12-01 18:49:05 +0200152static int big_num_opt = -1;
Lin Ming43bece72011-08-17 18:42:07 +0800153static bool group = false;
Peter Zijlstra1f16c572012-10-23 13:40:14 +0200154static const char *pre_cmd = NULL;
155static const char *post_cmd = NULL;
156static bool sync_run = false;
Frederik Deweerdta7e191c2013-03-01 13:02:27 -0500157static bool forever = false;
Andi Kleen44b1e602016-05-30 12:49:42 -0300158static bool force_metric_only = false;
Stephane Eranian13370a92013-01-29 12:47:44 +0100159static struct timespec ref_time;
Jiri Olsae0547312015-11-05 15:40:45 +0100160static bool append_file;
yuzhoujiandb06a262018-01-29 10:25:22 +0100161static bool interval_count;
Jiri Olsae0547312015-11-05 15:40:45 +0100162static const char *output_name;
163static int output_fd;
Stephane Eranian5af52b52010-05-18 15:00:01 +0200164
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100165struct perf_stat {
166 bool record;
Jiri Olsa8ceb41d2017-01-23 22:07:59 +0100167 struct perf_data data;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100168 struct perf_session *session;
169 u64 bytes_written;
Jiri Olsaba6039b62015-11-05 15:40:55 +0100170 struct perf_tool tool;
Jiri Olsa1975d362015-11-05 15:40:56 +0100171 bool maps_allocated;
Jiri Olsaf8548392019-07-21 13:23:49 +0200172 struct perf_cpu_map *cpus;
Jiri Olsa9749b902019-07-21 13:23:50 +0200173 struct perf_thread_map *threads;
Jiri Olsa89af4e02015-11-05 15:41:02 +0100174 enum aggr_mode aggr_mode;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100175};
176
177static struct perf_stat perf_stat;
178#define STAT_RECORD perf_stat.record
179
Liming Wang60666c62009-12-31 16:05:50 +0800180static volatile int done = 0;
181
Jiri Olsa421a50f2015-07-21 14:31:22 +0200182static struct perf_stat_config stat_config = {
Jiri Olsa26893a62018-08-30 08:32:40 +0200183 .aggr_mode = AGGR_GLOBAL,
184 .scale = true,
185 .unit_width = 4, /* strlen("unit") */
186 .run_count = 1,
187 .metric_only_len = METRIC_ONLY_LEN,
188 .walltime_nsecs_stats = &walltime_nsecs_stats,
Jiri Olsa34ff0862018-08-30 08:32:47 +0200189 .big_num = true,
Jiri Olsa421a50f2015-07-21 14:31:22 +0200190};
191
Stephane Eranian13370a92013-01-29 12:47:44 +0100192static inline void diff_timespec(struct timespec *r, struct timespec *a,
193 struct timespec *b)
194{
195 r->tv_sec = a->tv_sec - b->tv_sec;
196 if (a->tv_nsec < b->tv_nsec) {
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300197 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
Stephane Eranian13370a92013-01-29 12:47:44 +0100198 r->tv_sec--;
199 } else {
200 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
201 }
202}
203
Jiri Olsa254ecbc72015-06-26 11:29:13 +0200204static void perf_stat__reset_stats(void)
205{
Jin Yao56739442017-12-05 22:03:07 +0800206 int i;
207
Jiri Olsa254ecbc72015-06-26 11:29:13 +0200208 perf_evlist__reset_stats(evsel_list);
Jiri Olsaf87027b2015-06-03 16:25:59 +0200209 perf_stat__reset_shadow_stats();
Jin Yao56739442017-12-05 22:03:07 +0800210
211 for (i = 0; i < stat_config.stats_num; i++)
212 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
Jiri Olsa1eda3b22015-06-03 16:25:55 +0200213}
214
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100215static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100216 union perf_event *event,
217 struct perf_sample *sample __maybe_unused,
218 struct machine *machine __maybe_unused)
219{
Jiri Olsa8ceb41d2017-01-23 22:07:59 +0100220 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100221 pr_err("failed to write perf data, error: %m\n");
222 return -1;
223 }
224
225 perf_stat.bytes_written += event->header.size;
226 return 0;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100227}
228
Jiri Olsa1975d362015-11-05 15:40:56 +0100229static int write_stat_round_event(u64 tm, u64 type)
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100230{
Jiri Olsa1975d362015-11-05 15:40:56 +0100231 return perf_event__synthesize_stat_round(NULL, tm, type,
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100232 process_synthesized_event,
233 NULL);
234}
235
236#define WRITE_STAT_ROUND_EVENT(time, interval) \
237 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
238
Jiri Olsa8cd36f32019-09-02 22:04:12 +0200239#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
Jiri Olsa5a6ea812015-11-05 15:40:51 +0100240
241static int
Jiri Olsa32dcd022019-07-21 13:23:51 +0200242perf_evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
Jiri Olsa5a6ea812015-11-05 15:40:51 +0100243 struct perf_counts_values *count)
244{
245 struct perf_sample_id *sid = SID(counter, cpu, thread);
246
247 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
248 process_synthesized_event, NULL);
249}
250
Jiri Olsa32dcd022019-07-21 13:23:51 +0200251static int read_single_counter(struct evsel *counter, int cpu,
Andi Kleenf0fbb112019-03-26 15:18:21 -0700252 int thread, struct timespec *rs)
253{
254 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
255 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
256 struct perf_counts_values *count =
257 perf_counts(counter->counts, cpu, thread);
258 count->ena = count->run = val;
259 count->val = val;
260 return 0;
261 }
262 return perf_evsel__read_counter(counter, cpu, thread);
263}
264
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200265/*
266 * Read out the results of a single counter:
267 * do not aggregate counts across CPUs in system-wide mode
268 */
Andi Kleen4b49ab72019-11-20 16:15:20 -0800269static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200270{
Jiri Olsaa2f354e2019-08-22 13:11:41 +0200271 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
Andi Kleen4b49ab72019-11-20 16:15:20 -0800272 int thread;
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200273
Suzuki K. Poulose3b4331d2015-02-13 18:40:58 +0000274 if (!counter->supported)
275 return -ENOENT;
276
Jiri Olsa648b5af2019-08-06 11:35:19 +0200277 if (counter->core.system_wide)
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100278 nthreads = 1;
279
280 for (thread = 0; thread < nthreads; thread++) {
Andi Kleen4b49ab72019-11-20 16:15:20 -0800281 struct perf_counts_values *count;
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200282
Andi Kleen4b49ab72019-11-20 16:15:20 -0800283 count = perf_counts(counter->counts, cpu, thread);
Jiri Olsa82bf3112017-07-26 14:02:06 +0200284
Andi Kleen4b49ab72019-11-20 16:15:20 -0800285 /*
286 * The leader's group read loads data into its group members
287 * (via perf_evsel__read_counter()) and sets their count->loaded.
288 */
289 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
290 read_single_counter(counter, cpu, thread, rs)) {
291 counter->counts->scaled = -1;
292 perf_counts(counter->counts, cpu, thread)->ena = 0;
293 perf_counts(counter->counts, cpu, thread)->run = 0;
294 return -1;
295 }
296
297 perf_counts__set_loaded(counter->counts, cpu, thread, false);
298
299 if (STAT_RECORD) {
300 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
301 pr_err("failed to write stat event\n");
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100302 return -1;
Stephane Eraniandb49a712017-04-12 11:23:01 -0700303 }
Andi Kleen4b49ab72019-11-20 16:15:20 -0800304 }
Jiri Olsa5a6ea812015-11-05 15:40:51 +0100305
Andi Kleen4b49ab72019-11-20 16:15:20 -0800306 if (verbose > 1) {
307 fprintf(stat_config.output,
308 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
309 perf_evsel__name(counter),
310 cpu,
311 count->val, count->ena, count->run);
Jiri Olsa9bf1a522014-11-21 10:31:09 +0100312 }
Stephane Eranianf5b4a9c32010-11-16 11:05:01 +0200313 }
Arnaldo Carvalho de Meloc52b12e2011-01-03 17:45:52 -0200314
315 return 0;
Ingo Molnar2996f5d2009-05-29 09:10:54 +0200316}
317
Andi Kleenf0fbb112019-03-26 15:18:21 -0700318static void read_counters(struct timespec *rs)
Jiri Olsa106a94a2015-06-26 11:29:19 +0200319{
Jiri Olsa32dcd022019-07-21 13:23:51 +0200320 struct evsel *counter;
Andi Kleen4b49ab72019-11-20 16:15:20 -0800321 struct affinity affinity;
322 int i, ncpus, cpu;
323
324 if (affinity__setup(&affinity) < 0)
325 return;
326
327 ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
328 if (!target__has_cpu(&target) || target__has_per_thread(&target))
329 ncpus = 1;
330 evlist__for_each_cpu(evsel_list, i, cpu) {
331 if (i >= ncpus)
332 break;
333 affinity__set(&affinity, cpu);
334
335 evlist__for_each_entry(evsel_list, counter) {
336 if (evsel__cpu_iter_skip(counter, cpu))
337 continue;
338 if (!counter->err) {
339 counter->err = read_counter_cpu(counter, rs,
340 counter->cpu_iter - 1);
341 }
342 }
343 }
344 affinity__cleanup(&affinity);
Jiri Olsa106a94a2015-06-26 11:29:19 +0200345
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -0300346 evlist__for_each_entry(evsel_list, counter) {
Andi Kleen4b49ab72019-11-20 16:15:20 -0800347 if (counter->err)
Andi Kleen245bad82015-09-01 15:52:46 -0700348 pr_debug("failed to read counter %s\n", counter->name);
Andi Kleen4b49ab72019-11-20 16:15:20 -0800349 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
Jiri Olsa3b3eb042015-06-26 11:29:20 +0200350 pr_warning("failed to process counter %s\n", counter->name);
Andi Kleen4b49ab72019-11-20 16:15:20 -0800351 counter->err = 0;
Jiri Olsa106a94a2015-06-26 11:29:19 +0200352 }
353}
354
Jiri Olsaba411a92015-06-26 11:29:24 +0200355static void process_interval(void)
Stephane Eranian13370a92013-01-29 12:47:44 +0100356{
Stephane Eranian13370a92013-01-29 12:47:44 +0100357 struct timespec ts, rs;
Stephane Eranian13370a92013-01-29 12:47:44 +0100358
Stephane Eranian13370a92013-01-29 12:47:44 +0100359 clock_gettime(CLOCK_MONOTONIC, &ts);
360 diff_timespec(&rs, &ts, &ref_time);
Stephane Eranian13370a92013-01-29 12:47:44 +0100361
Andi Kleenf0fbb112019-03-26 15:18:21 -0700362 read_counters(&rs);
363
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100364 if (STAT_RECORD) {
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -0300365 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
Jiri Olsa7aad0c32015-11-05 15:40:52 +0100366 pr_err("failed to write stat round event\n");
367 }
368
Andi Kleenb90f1332017-08-31 12:40:36 -0700369 init_stats(&walltime_nsecs_stats);
370 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
Jiri Olsad4f63a42015-06-26 11:29:26 +0200371 print_counters(&rs, 0, NULL);
Stephane Eranian13370a92013-01-29 12:47:44 +0100372}
373
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100374static void enable_counters(void)
Andi Kleen41191682013-08-02 17:41:11 -0700375{
Jiri Olsa728c0ee2018-08-30 08:32:11 +0200376 if (stat_config.initial_delay)
377 usleep(stat_config.initial_delay * USEC_PER_MSEC);
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100378
379 /*
380 * We need to enable counters only if:
381 * - we don't have tracee (attaching to task or cpu)
382 * - we have initial delay configured
383 */
Jiri Olsa728c0ee2018-08-30 08:32:11 +0200384 if (!target__none(&target) || stat_config.initial_delay)
Jiri Olsa1c87f162019-07-21 13:24:08 +0200385 evlist__enable(evsel_list);
Andi Kleen41191682013-08-02 17:41:11 -0700386}
387
Mark Rutland3df33ef2016-08-09 14:04:29 +0100388static void disable_counters(void)
389{
390 /*
391 * If we don't have tracee (attaching to task or cpu), counters may
392 * still be running. To get accurate group ratios, we must stop groups
393 * from counting before reading their constituent counters.
394 */
395 if (!target__none(&target))
Jiri Olsae74676d2019-07-21 13:24:09 +0200396 evlist__disable(evsel_list);
Mark Rutland3df33ef2016-08-09 14:04:29 +0100397}
398
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300399static volatile int workload_exec_errno;
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300400
401/*
402 * perf_evlist__prepare_workload will send a SIGUSR1
403 * if the fork fails, since we asked by setting its
404 * want_signal to true.
405 */
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300406static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
407 void *ucontext __maybe_unused)
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300408{
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300409 workload_exec_errno = info->si_value.sival_int;
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300410}
411
Jiri Olsa32dcd022019-07-21 13:23:51 +0200412static bool perf_evsel__should_store_id(struct evsel *counter)
Jiri Olsa82bf3112017-07-26 14:02:06 +0200413{
Jiri Olsa1fc632c2019-07-21 13:24:29 +0200414 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
Jiri Olsa82bf3112017-07-26 14:02:06 +0200415}
416
Jiri Olsacbb5df7e2018-10-22 11:30:15 +0200417static bool is_target_alive(struct target *_target,
Jiri Olsa9749b902019-07-21 13:23:50 +0200418 struct perf_thread_map *threads)
Jiri Olsacbb5df7e2018-10-22 11:30:15 +0200419{
420 struct stat st;
421 int i;
422
423 if (!target__has_task(_target))
424 return true;
425
426 for (i = 0; i < threads->nr; i++) {
427 char path[PATH_MAX];
428
429 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
430 threads->map[i].pid);
431
432 if (!stat(path, &st))
433 return true;
434 }
435
436 return false;
437}
438
Andi Kleene0e6a6c2019-11-20 16:15:18 -0800439enum counter_recovery {
440 COUNTER_SKIP,
441 COUNTER_RETRY,
442 COUNTER_FATAL,
443};
444
445static enum counter_recovery stat_handle_error(struct evsel *counter)
446{
447 char msg[BUFSIZ];
448 /*
449 * PPC returns ENXIO for HW counters until 2.6.37
450 * (behavior changed with commit b0a873e).
451 */
452 if (errno == EINVAL || errno == ENOSYS ||
453 errno == ENOENT || errno == EOPNOTSUPP ||
454 errno == ENXIO) {
455 if (verbose > 0)
456 ui__warning("%s event is not supported by the kernel.\n",
457 perf_evsel__name(counter));
458 counter->supported = false;
Andi Kleen4804e012019-11-20 16:15:19 -0800459 /*
460 * errored is a sticky flag that means one of the counter's
461 * cpu event had a problem and needs to be reexamined.
462 */
463 counter->errored = true;
Andi Kleene0e6a6c2019-11-20 16:15:18 -0800464
465 if ((counter->leader != counter) ||
466 !(counter->leader->core.nr_members > 1))
467 return COUNTER_SKIP;
468 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
469 if (verbose > 0)
470 ui__warning("%s\n", msg);
471 return COUNTER_RETRY;
472 } else if (target__has_per_thread(&target) &&
473 evsel_list->core.threads &&
474 evsel_list->core.threads->err_thread != -1) {
475 /*
476 * For global --per-thread case, skip current
477 * error thread.
478 */
479 if (!thread_map__remove(evsel_list->core.threads,
480 evsel_list->core.threads->err_thread)) {
481 evsel_list->core.threads->err_thread = -1;
482 return COUNTER_RETRY;
483 }
484 }
485
486 perf_evsel__open_strerror(counter, &target,
487 errno, msg, sizeof(msg));
488 ui__error("%s\n", msg);
489
490 if (child_pid != -1)
491 kill(child_pid, SIGTERM);
492 return COUNTER_FATAL;
493}
494
Jiri Olsae55c14a2018-04-23 11:08:21 +0200495static int __run_perf_stat(int argc, const char **argv, int run_idx)
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200496{
Jiri Olsaec0d3d12015-07-21 14:31:25 +0200497 int interval = stat_config.interval;
yuzhoujiandb06a262018-01-29 10:25:22 +0100498 int times = stat_config.times;
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100499 int timeout = stat_config.timeout;
Arnaldo Carvalho de Melod6195a62017-02-13 16:45:24 -0300500 char msg[BUFSIZ];
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200501 unsigned long long t0, t1;
Jiri Olsa32dcd022019-07-21 13:23:51 +0200502 struct evsel *counter;
Stephane Eranian13370a92013-01-29 12:47:44 +0100503 struct timespec ts;
Stephane Eranian410136f2013-11-12 17:58:49 +0100504 size_t l;
Ingo Molnar42202dd2009-06-13 14:57:28 +0200505 int status = 0;
Zhang, Yanmin6be28502010-03-18 11:36:03 -0300506 const bool forks = (argc > 0);
Jiri Olsa8ceb41d2017-01-23 22:07:59 +0100507 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
Andi Kleen4804e012019-11-20 16:15:19 -0800508 struct affinity affinity;
509 int i, cpu;
510 bool second_pass = false;
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200511
Stephane Eranian13370a92013-01-29 12:47:44 +0100512 if (interval) {
Arnaldo Carvalho de Melo310ebb92016-08-08 14:57:04 -0300513 ts.tv_sec = interval / USEC_PER_MSEC;
514 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100515 } else if (timeout) {
516 ts.tv_sec = timeout / USEC_PER_MSEC;
517 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
Stephane Eranian13370a92013-01-29 12:47:44 +0100518 } else {
519 ts.tv_sec = 1;
520 ts.tv_nsec = 0;
521 }
522
Liming Wang60666c62009-12-31 16:05:50 +0800523 if (forks) {
Jiri Olsa664c98d2015-11-05 15:40:50 +0100524 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
Arnaldo Carvalho de Melo735f7e02014-01-03 14:56:49 -0300525 workload_exec_failed_signal) < 0) {
Namhyung Kimacf28922013-03-11 16:43:18 +0900526 perror("failed to prepare workload");
527 return -1;
Liming Wang60666c62009-12-31 16:05:50 +0800528 }
Namhyung Kimd20a47e2013-09-30 18:01:11 +0900529 child_pid = evsel_list->workload.pid;
Paul Mackerras051ae7f2009-06-29 21:13:21 +1000530 }
531
Jiri Olsa6a4bb042012-08-08 12:22:36 +0200532 if (group)
Arnaldo Carvalho de Melo63dab222012-08-14 16:35:48 -0300533 perf_evlist__set_leader(evsel_list);
Jiri Olsa6a4bb042012-08-08 12:22:36 +0200534
Andi Kleen4804e012019-11-20 16:15:19 -0800535 if (affinity__setup(&affinity) < 0)
536 return -1;
Andi Kleen5a5dfe42017-08-31 12:40:26 -0700537
Andi Kleen4804e012019-11-20 16:15:19 -0800538 evlist__for_each_cpu (evsel_list, i, cpu) {
539 affinity__set(&affinity, cpu);
Andi Kleen5a5dfe42017-08-31 12:40:26 -0700540
Andi Kleen4804e012019-11-20 16:15:19 -0800541 evlist__for_each_entry(evsel_list, counter) {
542 if (evsel__cpu_iter_skip(counter, cpu))
Andi Kleene0e6a6c2019-11-20 16:15:18 -0800543 continue;
Andi Kleen4804e012019-11-20 16:15:19 -0800544 if (counter->reset_group || counter->errored)
545 continue;
546try_again:
547 if (create_perf_stat_counter(counter, &stat_config, &target,
548 counter->cpu_iter - 1) < 0) {
549
550 /*
551 * Weak group failed. We cannot just undo this here
552 * because earlier CPUs might be in group mode, and the kernel
553 * doesn't support mixing group and non group reads. Defer
554 * it to later.
555 * Don't close here because we're in the wrong affinity.
556 */
557 if ((errno == EINVAL || errno == EBADF) &&
558 counter->leader != counter &&
559 counter->weak_group) {
560 perf_evlist__reset_weak_group(evsel_list, counter, false);
561 assert(counter->reset_group);
562 second_pass = true;
563 continue;
564 }
565
566 switch (stat_handle_error(counter)) {
567 case COUNTER_FATAL:
568 return -1;
569 case COUNTER_RETRY:
570 goto try_again;
571 case COUNTER_SKIP:
572 continue;
573 default:
574 break;
575 }
576
577 }
578 counter->supported = true;
579 }
580 }
581
582 if (second_pass) {
583 /*
584 * Now redo all the weak group after closing them,
585 * and also close errored counters.
586 */
587
588 evlist__for_each_cpu(evsel_list, i, cpu) {
589 affinity__set(&affinity, cpu);
590 /* First close errored or weak retry */
591 evlist__for_each_entry(evsel_list, counter) {
592 if (!counter->reset_group && !counter->errored)
593 continue;
594 if (evsel__cpu_iter_skip_no_inc(counter, cpu))
595 continue;
596 perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
597 }
598 /* Now reopen weak */
599 evlist__for_each_entry(evsel_list, counter) {
600 if (!counter->reset_group && !counter->errored)
601 continue;
602 if (evsel__cpu_iter_skip(counter, cpu))
603 continue;
604 if (!counter->reset_group)
605 continue;
606try_again_reset:
607 pr_debug2("reopening weak %s\n", perf_evsel__name(counter));
608 if (create_perf_stat_counter(counter, &stat_config, &target,
609 counter->cpu_iter - 1) < 0) {
610
611 switch (stat_handle_error(counter)) {
612 case COUNTER_FATAL:
613 return -1;
614 case COUNTER_RETRY:
615 goto try_again_reset;
616 case COUNTER_SKIP:
617 continue;
618 default:
619 break;
620 }
621 }
622 counter->supported = true;
Jin Yaoab6c79b2018-01-16 23:43:08 +0800623 }
Arnaldo Carvalho de Melo48290602011-01-03 17:48:12 -0200624 }
Andi Kleen4804e012019-11-20 16:15:19 -0800625 }
626 affinity__cleanup(&affinity);
627
628 evlist__for_each_entry(evsel_list, counter) {
629 if (!counter->supported) {
630 perf_evsel__free_fd(&counter->core);
631 continue;
632 }
Stephane Eranian410136f2013-11-12 17:58:49 +0100633
634 l = strlen(counter->unit);
Jiri Olsadf4f7b42018-08-30 08:32:32 +0200635 if (l > stat_config.unit_width)
636 stat_config.unit_width = l;
Jiri Olsa2af46462015-11-05 15:40:49 +0100637
Jiri Olsa82bf3112017-07-26 14:02:06 +0200638 if (perf_evsel__should_store_id(counter) &&
Jiri Olsa650d6222018-08-30 08:32:16 +0200639 perf_evsel__store_ids(counter, evsel_list))
Jiri Olsa2af46462015-11-05 15:40:49 +0100640 return -1;
Arnaldo Carvalho de Melo084ab9f2010-03-22 13:10:28 -0300641 }
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200642
Arnaldo Carvalho de Melo23d4aad2015-03-24 19:23:47 -0300643 if (perf_evlist__apply_filters(evsel_list, &counter)) {
Arnaldo Carvalho de Melo62d94b02017-06-27 11:22:31 -0300644 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
Arnaldo Carvalho de Melo23d4aad2015-03-24 19:23:47 -0300645 counter->filter, perf_evsel__name(counter), errno,
Arnaldo Carvalho de Meloc8b5f2c2016-07-06 11:56:20 -0300646 str_error_r(errno, msg, sizeof(msg)));
Frederic Weisbeckercfd748a2011-03-14 16:40:30 +0100647 return -1;
648 }
649
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100650 if (STAT_RECORD) {
Jiri Olsa8ceb41d2017-01-23 22:07:59 +0100651 int err, fd = perf_data__fd(&perf_stat.data);
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100652
Jiri Olsa664c98d2015-11-05 15:40:50 +0100653 if (is_pipe) {
Jiri Olsa8ceb41d2017-01-23 22:07:59 +0100654 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
Jiri Olsa664c98d2015-11-05 15:40:50 +0100655 } else {
656 err = perf_session__write_header(perf_stat.session, evsel_list,
657 fd, false);
658 }
659
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100660 if (err < 0)
661 return err;
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100662
Arnaldo Carvalho de Melob2518922019-09-10 17:17:33 +0100663 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
664 process_synthesized_event, is_pipe);
Jiri Olsa8b99b1a2015-11-05 15:40:48 +0100665 if (err < 0)
666 return err;
Jiri Olsa4979d0c2015-11-05 15:40:46 +0100667 }
668
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200669 /*
670 * Enable counters and exec the command:
671 */
672 t0 = rdclock();
Stephane Eranian13370a92013-01-29 12:47:44 +0100673 clock_gettime(CLOCK_MONOTONIC, &ref_time);
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200674
Liming Wang60666c62009-12-31 16:05:50 +0800675 if (forks) {
Namhyung Kimacf28922013-03-11 16:43:18 +0900676 perf_evlist__start_workload(evsel_list);
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100677 enable_counters();
Namhyung Kimacf28922013-03-11 16:43:18 +0900678
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100679 if (interval || timeout) {
Stephane Eranian13370a92013-01-29 12:47:44 +0100680 while (!waitpid(child_pid, &status, WNOHANG)) {
681 nanosleep(&ts, NULL);
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100682 if (timeout)
683 break;
Jiri Olsaba411a92015-06-26 11:29:24 +0200684 process_interval();
yuzhoujiandb06a262018-01-29 10:25:22 +0100685 if (interval_count && !(--times))
686 break;
Stephane Eranian13370a92013-01-29 12:47:44 +0100687 }
688 }
Jin Yao8a992552019-01-03 15:40:45 +0800689 if (child_pid != -1)
690 wait4(child_pid, &status, 0, &stat_config.ru_data);
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300691
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300692 if (workload_exec_errno) {
Arnaldo Carvalho de Meloc8b5f2c2016-07-06 11:56:20 -0300693 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300694 pr_err("Workload failed: %s\n", emsg);
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300695 return -1;
Arnaldo Carvalho de Melof33cbe72014-01-02 15:11:25 -0300696 }
Arnaldo Carvalho de Melo6af206f2013-12-28 15:45:08 -0300697
Andi Kleen33e49ea2011-09-15 14:31:40 -0700698 if (WIFSIGNALED(status))
699 psignal(WTERMSIG(status), argv[0]);
Liming Wang60666c62009-12-31 16:05:50 +0800700 } else {
Jiri Olsa67ccdec2015-12-03 10:06:44 +0100701 enable_counters();
Stephane Eranian13370a92013-01-29 12:47:44 +0100702 while (!done) {
703 nanosleep(&ts, NULL);
Jiri Olsa03617c22019-07-21 13:24:42 +0200704 if (!is_target_alive(&target, evsel_list->core.threads))
Jiri Olsacbb5df7e2018-10-22 11:30:15 +0200705 break;
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100706 if (timeout)
707 break;
yuzhoujiandb06a262018-01-29 10:25:22 +0100708 if (interval) {
Jiri Olsaba411a92015-06-26 11:29:24 +0200709 process_interval();
yuzhoujiandb06a262018-01-29 10:25:22 +0100710 if (interval_count && !(--times))
711 break;
712 }
Stephane Eranian13370a92013-01-29 12:47:44 +0100713 }
Liming Wang60666c62009-12-31 16:05:50 +0800714 }
Ingo Molnar44db76c2009-06-03 19:36:07 +0200715
Mark Rutland3df33ef2016-08-09 14:04:29 +0100716 disable_counters();
717
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200718 t1 = rdclock();
719
Jiri Olsa54ac0b12018-08-30 08:32:50 +0200720 if (stat_config.walltime_run_table)
721 stat_config.walltime_run[run_idx] = t1 - t0;
Jiri Olsae55c14a2018-04-23 11:08:21 +0200722
Peter Zijlstra9e9772c2009-09-04 15:36:08 +0200723 update_stats(&walltime_nsecs_stats, t1 - t0);
Ingo Molnar42202dd2009-06-13 14:57:28 +0200724
Mark Rutland3df33ef2016-08-09 14:04:29 +0100725 /*
726 * Closing a group leader splits the group, and as we only disable
727 * group leaders, results in remaining events becoming enabled. To
728 * avoid arbitrary skew, we must read all counters before closing any
729 * group leaders.
730 */
Andi Kleenf0fbb112019-03-26 15:18:21 -0700731 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
Jiri Olsa08ef3af2019-07-15 16:21:21 +0200732
733 /*
734 * We need to keep evsel_list alive, because it's processed
735 * later the evsel_list will be closed after.
736 */
737 if (!STAT_RECORD)
Jiri Olsa750b4ed2019-07-21 13:24:07 +0200738 evlist__close(evsel_list);
Arnaldo Carvalho de Meloc52b12e2011-01-03 17:45:52 -0200739
Ingo Molnar42202dd2009-06-13 14:57:28 +0200740 return WEXITSTATUS(status);
741}
742
Jiri Olsae55c14a2018-04-23 11:08:21 +0200743static int run_perf_stat(int argc, const char **argv, int run_idx)
Peter Zijlstra1f16c572012-10-23 13:40:14 +0200744{
745 int ret;
746
747 if (pre_cmd) {
748 ret = system(pre_cmd);
749 if (ret)
750 return ret;
751 }
752
753 if (sync_run)
754 sync();
755
Jiri Olsae55c14a2018-04-23 11:08:21 +0200756 ret = __run_perf_stat(argc, argv, run_idx);
Peter Zijlstra1f16c572012-10-23 13:40:14 +0200757 if (ret)
758 return ret;
759
760 if (post_cmd) {
761 ret = system(post_cmd);
762 if (ret)
763 return ret;
764 }
765
766 return ret;
767}
768
Jiri Olsaa5a9eac2018-08-30 08:32:24 +0200769static void print_counters(struct timespec *ts, int argc, const char **argv)
770{
Jiri Olsa01748202018-08-30 08:32:25 +0200771 /* Do not print anything if we record to the pipe. */
772 if (STAT_RECORD && perf_stat.data.is_pipe)
773 return;
774
Jiri Olsac512e0e2018-08-30 08:32:33 +0200775 perf_evlist__print_counters(evsel_list, &stat_config, &target,
Jiri Olsab64df7f2018-08-30 08:32:26 +0200776 ts, argc, argv);
Jiri Olsaa5a9eac2018-08-30 08:32:24 +0200777}
778
Peter Zijlstraf7b7c262009-06-10 15:55:59 +0200779static volatile int signr = -1;
780
Ingo Molnar52425192009-05-26 09:17:18 +0200781static void skip_signal(int signo)
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200782{
Jiri Olsaec0d3d12015-07-21 14:31:25 +0200783 if ((child_pid == -1) || stat_config.interval)
Liming Wang60666c62009-12-31 16:05:50 +0800784 done = 1;
785
Peter Zijlstraf7b7c262009-06-10 15:55:59 +0200786 signr = signo;
Stephane Eraniand07f0b12013-06-04 17:44:26 +0200787 /*
788 * render child_pid harmless
789 * won't send SIGTERM to a random
790 * process in case of race condition
791 * and fast PID recycling
792 */
793 child_pid = -1;
Peter Zijlstraf7b7c262009-06-10 15:55:59 +0200794}
795
796static void sig_atexit(void)
797{
Stephane Eraniand07f0b12013-06-04 17:44:26 +0200798 sigset_t set, oset;
799
800 /*
801 * avoid race condition with SIGCHLD handler
802 * in skip_signal() which is modifying child_pid
803 * goal is to avoid send SIGTERM to a random
804 * process
805 */
806 sigemptyset(&set);
807 sigaddset(&set, SIGCHLD);
808 sigprocmask(SIG_BLOCK, &set, &oset);
809
Chris Wilson933da832009-10-04 01:35:01 +0100810 if (child_pid != -1)
811 kill(child_pid, SIGTERM);
812
Stephane Eraniand07f0b12013-06-04 17:44:26 +0200813 sigprocmask(SIG_SETMASK, &oset, NULL);
814
Peter Zijlstraf7b7c262009-06-10 15:55:59 +0200815 if (signr == -1)
816 return;
817
818 signal(signr, SIG_DFL);
819 kill(getpid(), signr);
Ingo Molnar52425192009-05-26 09:17:18 +0200820}
Ingo Molnarddcacfa2009-04-20 15:37:32 +0200821
Irina Tirdea1d037ca2012-09-11 01:15:03 +0300822static int stat__set_big_num(const struct option *opt __maybe_unused,
823 const char *s __maybe_unused, int unset)
Stephane Eraniand7470b62010-12-01 18:49:05 +0200824{
825 big_num_opt = unset ? 0 : 1;
826 return 0;
827}
828
Andi Kleen44b1e602016-05-30 12:49:42 -0300829static int enable_metric_only(const struct option *opt __maybe_unused,
830 const char *s __maybe_unused, int unset)
831{
832 force_metric_only = true;
Jiri Olsa0ce5aa02018-08-30 08:32:31 +0200833 stat_config.metric_only = !unset;
Andi Kleen44b1e602016-05-30 12:49:42 -0300834 return 0;
835}
836
Andi Kleenb18f3e32017-08-31 12:40:31 -0700837static int parse_metric_groups(const struct option *opt,
838 const char *str,
839 int unset __maybe_unused)
840{
Jiri Olsad0192fd2018-08-30 08:32:51 +0200841 return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
Andi Kleenb18f3e32017-08-31 12:40:31 -0700842}
843
Michael Petlan51433ea2018-12-10 11:00:04 -0500844static struct option stat_options[] = {
Jiri Olsae0547312015-11-05 15:40:45 +0100845 OPT_BOOLEAN('T', "transaction", &transaction_run,
846 "hardware transaction statistics"),
847 OPT_CALLBACK('e', "event", &evsel_list, "event",
848 "event selector. use 'perf list' to list available events",
849 parse_events_option),
850 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
851 "event filter", parse_filter),
Jiri Olsa5698f262018-08-30 08:32:12 +0200852 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
Jiri Olsae0547312015-11-05 15:40:45 +0100853 "child tasks do not inherit counters"),
854 OPT_STRING('p', "pid", &target.pid, "pid",
855 "stat events on existing process id"),
856 OPT_STRING('t', "tid", &target.tid, "tid",
857 "stat events on existing thread id"),
858 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
859 "system-wide collection from all CPUs"),
860 OPT_BOOLEAN('g', "group", &group,
861 "put the counters into a counter group"),
Andi Kleen75998bb2019-03-14 15:50:01 -0700862 OPT_BOOLEAN(0, "scale", &stat_config.scale,
863 "Use --no-scale to disable counter scaling for multiplexing"),
Jiri Olsae0547312015-11-05 15:40:45 +0100864 OPT_INCR('v', "verbose", &verbose,
865 "be more verbose (show counter open errors, etc)"),
Jiri Olsad97ae042018-08-30 08:32:36 +0200866 OPT_INTEGER('r', "repeat", &stat_config.run_count,
Jiri Olsae0547312015-11-05 15:40:45 +0100867 "repeat command and print average + stddev (max: 100, forever: 0)"),
Jiri Olsa54ac0b12018-08-30 08:32:50 +0200868 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
Jiri Olsae55c14a2018-04-23 11:08:21 +0200869 "display details about each run (only with -r option)"),
Jiri Olsaaea0dca2018-08-30 08:32:41 +0200870 OPT_BOOLEAN('n', "null", &stat_config.null_run,
Jiri Olsae0547312015-11-05 15:40:45 +0100871 "null run - dont start any counters"),
872 OPT_INCR('d', "detailed", &detailed_run,
873 "detailed run - start a lot of events"),
874 OPT_BOOLEAN('S', "sync", &sync_run,
875 "call sync() before starting a run"),
876 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
877 "print large numbers with thousands\' separators",
878 stat__set_big_num),
879 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
880 "list of cpus to monitor in system-wide"),
881 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
882 "disable CPU count aggregation", AGGR_NONE),
Jiri Olsafdee3352018-08-30 08:32:48 +0200883 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
Jiri Olsafa7070a2018-08-30 08:32:29 +0200884 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
Jiri Olsae0547312015-11-05 15:40:45 +0100885 "print counts with custom separator"),
886 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
887 "monitor event in cgroup name only", parse_cgroups),
888 OPT_STRING('o', "output", &output_name, "file", "output file name"),
889 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
890 OPT_INTEGER(0, "log-fd", &output_fd,
891 "log output to fd, instead of stderr"),
892 OPT_STRING(0, "pre", &pre_cmd, "command",
893 "command to run prior to the measured command"),
894 OPT_STRING(0, "post", &post_cmd, "command",
895 "command to run after to the measured command"),
896 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
Alexey Budankov9dc9a952018-04-03 21:18:33 +0300897 "print counts at regular interval in ms "
898 "(overhead is possible for values <= 100ms)"),
yuzhoujiandb06a262018-01-29 10:25:22 +0100899 OPT_INTEGER(0, "interval-count", &stat_config.times,
900 "print counts for fixed number of times"),
Jiri Olsa132c6ba2018-08-30 08:32:30 +0200901 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
Jiri Olsa9660e082018-06-07 00:15:06 +0200902 "clear screen in between new interval"),
yuzhoujianf1f8ad52018-01-29 10:25:23 +0100903 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
904 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
Jiri Olsae0547312015-11-05 15:40:45 +0100905 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
906 "aggregate counts per processor socket", AGGR_SOCKET),
Kan Liangdb5742b2019-06-04 15:50:42 -0700907 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
908 "aggregate counts per processor die", AGGR_DIE),
Jiri Olsae0547312015-11-05 15:40:45 +0100909 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
910 "aggregate counts per physical processor core", AGGR_CORE),
911 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
912 "aggregate counts per thread", AGGR_THREAD),
Jiri Olsa86895b42019-08-28 10:17:43 +0200913 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
914 "aggregate counts per numa node", AGGR_NODE),
Jiri Olsa728c0ee2018-08-30 08:32:11 +0200915 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
Jiri Olsae0547312015-11-05 15:40:45 +0100916 "ms to wait before starting measurement after program start"),
Jiri Olsa0ce5aa02018-08-30 08:32:31 +0200917 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
Andi Kleen44b1e602016-05-30 12:49:42 -0300918 "Only print computed metrics. No raw values", enable_metric_only),
919 OPT_BOOLEAN(0, "topdown", &topdown_run,
920 "measure topdown level 1 statistics"),
Kan Liangdaefd0b2017-05-26 12:05:38 -0700921 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
922 "measure SMI cost"),
Andi Kleenb18f3e32017-08-31 12:40:31 -0700923 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
924 "monitor specified metrics or metric groups (separated by ,)",
925 parse_metric_groups),
Jin Yaodd071022019-10-11 13:05:45 +0800926 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
927 "Configure all used events to run in kernel space.",
928 PARSE_OPT_EXCLUSIVE),
929 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
930 "Configure all used events to run in user space.",
931 PARSE_OPT_EXCLUSIVE),
Jin Yao1af62ce2020-02-14 16:04:52 +0800932 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
933 "Use with 'percore' event qualifier to show the event "
934 "counts of one hardware thread by sum up total hardware "
935 "threads of same physical core"),
Jiri Olsae0547312015-11-05 15:40:45 +0100936 OPT_END()
937};
938
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200939static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +0200940 struct perf_cpu_map *map, int cpu)
Jiri Olsa1fe7a302015-10-16 12:41:15 +0200941{
942 return cpu_map__get_socket(map, cpu, NULL);
943}
944
Kan Liangdb5742b2019-06-04 15:50:42 -0700945static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +0200946 struct perf_cpu_map *map, int cpu)
Kan Liangdb5742b2019-06-04 15:50:42 -0700947{
948 return cpu_map__get_die(map, cpu, NULL);
949}
950
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200951static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +0200952 struct perf_cpu_map *map, int cpu)
Jiri Olsa1fe7a302015-10-16 12:41:15 +0200953{
954 return cpu_map__get_core(map, cpu, NULL);
955}
956
Jiri Olsa86895b42019-08-28 10:17:43 +0200957static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
958 struct perf_cpu_map *map, int cpu)
959{
960 return cpu_map__get_node(map, cpu, NULL);
961}
962
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200963static int perf_stat__get_aggr(struct perf_stat_config *config,
Jiri Olsaf8548392019-07-21 13:23:49 +0200964 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100965{
966 int cpu;
967
968 if (idx >= map->nr)
969 return -1;
970
971 cpu = map->map[idx];
972
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200973 if (config->cpus_aggr_map->map[cpu] == -1)
974 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100975
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200976 return config->cpus_aggr_map->map[cpu];
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100977}
978
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200979static int perf_stat__get_socket_cached(struct perf_stat_config *config,
Jiri Olsaf8548392019-07-21 13:23:49 +0200980 struct perf_cpu_map *map, int idx)
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100981{
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200982 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100983}
984
Kan Liangdb5742b2019-06-04 15:50:42 -0700985static int perf_stat__get_die_cached(struct perf_stat_config *config,
Jiri Olsaf8548392019-07-21 13:23:49 +0200986 struct perf_cpu_map *map, int idx)
Kan Liangdb5742b2019-06-04 15:50:42 -0700987{
988 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
989}
990
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200991static int perf_stat__get_core_cached(struct perf_stat_config *config,
Jiri Olsaf8548392019-07-21 13:23:49 +0200992 struct perf_cpu_map *map, int idx)
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100993{
Jiri Olsa6f6b6592018-08-30 08:32:45 +0200994 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
Jiri Olsa1e5a2932015-10-25 15:51:18 +0100995}
996
Jiri Olsa86895b42019-08-28 10:17:43 +0200997static int perf_stat__get_node_cached(struct perf_stat_config *config,
998 struct perf_cpu_map *map, int idx)
999{
1000 return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1001}
1002
Jin Yao4fc4d8d2019-04-12 21:59:49 +08001003static bool term_percore_set(void)
1004{
Jiri Olsa32dcd022019-07-21 13:23:51 +02001005 struct evsel *counter;
Jin Yao4fc4d8d2019-04-12 21:59:49 +08001006
1007 evlist__for_each_entry(evsel_list, counter) {
1008 if (counter->percore)
1009 return true;
1010 }
1011
1012 return false;
1013}
1014
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001015static int perf_stat_init_aggr_mode(void)
1016{
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001017 int nr;
1018
Jiri Olsa421a50f2015-07-21 14:31:22 +02001019 switch (stat_config.aggr_mode) {
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001020 case AGGR_SOCKET:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001021 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001022 perror("cannot build socket map");
1023 return -1;
1024 }
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001025 stat_config.aggr_get_id = perf_stat__get_socket_cached;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001026 break;
Kan Liangdb5742b2019-06-04 15:50:42 -07001027 case AGGR_DIE:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001028 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
Kan Liangdb5742b2019-06-04 15:50:42 -07001029 perror("cannot build die map");
1030 return -1;
1031 }
1032 stat_config.aggr_get_id = perf_stat__get_die_cached;
1033 break;
Stephane Eranian12c08a92013-02-14 13:57:29 +01001034 case AGGR_CORE:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001035 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
Stephane Eranian12c08a92013-02-14 13:57:29 +01001036 perror("cannot build core map");
1037 return -1;
1038 }
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001039 stat_config.aggr_get_id = perf_stat__get_core_cached;
Stephane Eranian12c08a92013-02-14 13:57:29 +01001040 break;
Jiri Olsa86895b42019-08-28 10:17:43 +02001041 case AGGR_NODE:
1042 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1043 perror("cannot build core map");
1044 return -1;
1045 }
1046 stat_config.aggr_get_id = perf_stat__get_node_cached;
1047 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001048 case AGGR_NONE:
Jin Yao4fc4d8d2019-04-12 21:59:49 +08001049 if (term_percore_set()) {
Jiri Olsaf72f9012019-07-21 13:24:41 +02001050 if (cpu_map__build_core_map(evsel_list->core.cpus,
Jin Yao4fc4d8d2019-04-12 21:59:49 +08001051 &stat_config.aggr_map)) {
1052 perror("cannot build core map");
1053 return -1;
1054 }
1055 stat_config.aggr_get_id = perf_stat__get_core_cached;
1056 }
1057 break;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001058 case AGGR_GLOBAL:
Jiri Olsa32b8af82015-06-26 11:29:27 +02001059 case AGGR_THREAD:
Jiri Olsa208df992015-10-16 12:41:04 +02001060 case AGGR_UNSET:
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001061 default:
1062 break;
1063 }
Jiri Olsa1e5a2932015-10-25 15:51:18 +01001064
1065 /*
1066 * The evsel_list->cpus is the base we operate on,
1067 * taking the highest cpu number to be the size of
1068 * the aggregation translate cpumap.
1069 */
Jiri Olsa4256d432019-09-02 14:12:53 +02001070 nr = perf_cpu_map__max(evsel_list->core.cpus);
Jiri Olsa315c0a12019-08-22 13:11:39 +02001071 stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001072 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01001073}
1074
Masami Hiramatsu544c2ae2015-12-09 11:11:27 +09001075static void perf_stat__exit_aggr_mode(void)
1076{
Jiri Olsa38f01d82019-07-21 13:24:17 +02001077 perf_cpu_map__put(stat_config.aggr_map);
1078 perf_cpu_map__put(stat_config.cpus_aggr_map);
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001079 stat_config.aggr_map = NULL;
1080 stat_config.cpus_aggr_map = NULL;
Masami Hiramatsu544c2ae2015-12-09 11:11:27 +09001081}
1082
Jiri Olsaf8548392019-07-21 13:23:49 +02001083static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001084{
1085 int cpu;
1086
1087 if (idx > map->nr)
1088 return -1;
1089
1090 cpu = map->map[idx];
1091
Jan Stancekda8a58b2017-02-17 12:10:26 +01001092 if (cpu >= env->nr_cpus_avail)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001093 return -1;
1094
1095 return cpu;
1096}
1097
Jiri Olsaf8548392019-07-21 13:23:49 +02001098static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001099{
1100 struct perf_env *env = data;
1101 int cpu = perf_env__get_cpu(env, map, idx);
1102
1103 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1104}
1105
Jiri Olsaf8548392019-07-21 13:23:49 +02001106static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
Kan Liangdb5742b2019-06-04 15:50:42 -07001107{
1108 struct perf_env *env = data;
1109 int die_id = -1, cpu = perf_env__get_cpu(env, map, idx);
1110
1111 if (cpu != -1) {
1112 /*
1113 * Encode socket in bit range 15:8
1114 * die_id is relative to socket,
1115 * we need a global id. So we combine
1116 * socket + die id
1117 */
1118 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1119 return -1;
1120
1121 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1122 return -1;
1123
1124 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff);
1125 }
1126
1127 return die_id;
1128}
1129
Jiri Olsaf8548392019-07-21 13:23:49 +02001130static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001131{
1132 struct perf_env *env = data;
1133 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1134
1135 if (cpu != -1) {
Jiri Olsa68d702f2015-11-05 15:40:58 +01001136 /*
Kan Liangdb5742b2019-06-04 15:50:42 -07001137 * Encode socket in bit range 31:24
1138 * encode die id in bit range 23:16
1139 * core_id is relative to socket and die,
Jiri Olsa68d702f2015-11-05 15:40:58 +01001140 * we need a global id. So we combine
Kan Liangdb5742b2019-06-04 15:50:42 -07001141 * socket + die id + core id
Jiri Olsa68d702f2015-11-05 15:40:58 +01001142 */
Kan Liangdb5742b2019-06-04 15:50:42 -07001143 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n"))
1144 return -1;
1145
1146 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n"))
1147 return -1;
1148
1149 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n"))
1150 return -1;
1151
1152 core = (env->cpu[cpu].socket_id << 24) |
1153 (env->cpu[cpu].die_id << 16) |
1154 (env->cpu[cpu].core_id & 0xffff);
Jiri Olsa68d702f2015-11-05 15:40:58 +01001155 }
1156
1157 return core;
1158}
1159
Jiri Olsa86895b42019-08-28 10:17:43 +02001160static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1161{
1162 int cpu = perf_env__get_cpu(data, map, idx);
1163
1164 return perf_env__numa_node(data, cpu);
1165}
1166
Jiri Olsaf8548392019-07-21 13:23:49 +02001167static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1168 struct perf_cpu_map **sockp)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001169{
1170 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1171}
1172
Jiri Olsaf8548392019-07-21 13:23:49 +02001173static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1174 struct perf_cpu_map **diep)
Kan Liangdb5742b2019-06-04 15:50:42 -07001175{
1176 return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1177}
1178
Jiri Olsaf8548392019-07-21 13:23:49 +02001179static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1180 struct perf_cpu_map **corep)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001181{
1182 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1183}
1184
Jiri Olsa86895b42019-08-28 10:17:43 +02001185static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1186 struct perf_cpu_map **nodep)
1187{
1188 return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1189}
1190
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001191static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +02001192 struct perf_cpu_map *map, int idx)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001193{
1194 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1195}
Kan Liangdb5742b2019-06-04 15:50:42 -07001196static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +02001197 struct perf_cpu_map *map, int idx)
Kan Liangdb5742b2019-06-04 15:50:42 -07001198{
1199 return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1200}
Jiri Olsa68d702f2015-11-05 15:40:58 +01001201
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001202static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
Jiri Olsaf8548392019-07-21 13:23:49 +02001203 struct perf_cpu_map *map, int idx)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001204{
1205 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1206}
1207
Jiri Olsa86895b42019-08-28 10:17:43 +02001208static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1209 struct perf_cpu_map *map, int idx)
1210{
1211 return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1212}
1213
Jiri Olsa68d702f2015-11-05 15:40:58 +01001214static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1215{
1216 struct perf_env *env = &st->session->header.env;
1217
1218 switch (stat_config.aggr_mode) {
1219 case AGGR_SOCKET:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001220 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
Jiri Olsa68d702f2015-11-05 15:40:58 +01001221 perror("cannot build socket map");
1222 return -1;
1223 }
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001224 stat_config.aggr_get_id = perf_stat__get_socket_file;
Jiri Olsa68d702f2015-11-05 15:40:58 +01001225 break;
Kan Liangdb5742b2019-06-04 15:50:42 -07001226 case AGGR_DIE:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001227 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
Kan Liangdb5742b2019-06-04 15:50:42 -07001228 perror("cannot build die map");
1229 return -1;
1230 }
1231 stat_config.aggr_get_id = perf_stat__get_die_file;
1232 break;
Jiri Olsa68d702f2015-11-05 15:40:58 +01001233 case AGGR_CORE:
Jiri Olsaf72f9012019-07-21 13:24:41 +02001234 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
Jiri Olsa68d702f2015-11-05 15:40:58 +01001235 perror("cannot build core map");
1236 return -1;
1237 }
Jiri Olsa6f6b6592018-08-30 08:32:45 +02001238 stat_config.aggr_get_id = perf_stat__get_core_file;
Jiri Olsa68d702f2015-11-05 15:40:58 +01001239 break;
Jiri Olsa86895b42019-08-28 10:17:43 +02001240 case AGGR_NODE:
1241 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1242 perror("cannot build core map");
1243 return -1;
1244 }
1245 stat_config.aggr_get_id = perf_stat__get_node_file;
1246 break;
Jiri Olsa68d702f2015-11-05 15:40:58 +01001247 case AGGR_NONE:
1248 case AGGR_GLOBAL:
1249 case AGGR_THREAD:
1250 case AGGR_UNSET:
1251 default:
1252 break;
1253 }
1254
1255 return 0;
1256}
1257
Andi Kleen44b1e602016-05-30 12:49:42 -03001258static int topdown_filter_events(const char **attr, char **str, bool use_group)
1259{
1260 int off = 0;
1261 int i;
1262 int len = 0;
1263 char *s;
1264
1265 for (i = 0; attr[i]; i++) {
1266 if (pmu_have_event("cpu", attr[i])) {
1267 len += strlen(attr[i]) + 1;
1268 attr[i - off] = attr[i];
1269 } else
1270 off++;
1271 }
1272 attr[i - off] = NULL;
1273
1274 *str = malloc(len + 1 + 2);
1275 if (!*str)
1276 return -1;
1277 s = *str;
1278 if (i - off == 0) {
1279 *s = 0;
1280 return 0;
1281 }
1282 if (use_group)
1283 *s++ = '{';
1284 for (i = 0; attr[i]; i++) {
1285 strcpy(s, attr[i]);
1286 s += strlen(s);
1287 *s++ = ',';
1288 }
1289 if (use_group) {
1290 s[-1] = '}';
1291 *s = 0;
1292 } else
1293 s[-1] = 0;
1294 return 0;
1295}
1296
1297__weak bool arch_topdown_check_group(bool *warn)
1298{
1299 *warn = false;
1300 return false;
1301}
1302
1303__weak void arch_topdown_group_warn(void)
1304{
1305}
1306
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001307/*
1308 * Add default attributes, if there were no attributes specified or
1309 * if -d/--detailed, -d -d or -d -d -d is used:
1310 */
1311static int add_default_attributes(void)
1312{
Andi Kleen44b1e602016-05-30 12:49:42 -03001313 int err;
Andi Kleen9dec4472016-02-26 16:27:56 -08001314 struct perf_event_attr default_attrs0[] = {
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001315
1316 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1317 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1318 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1319 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1320
1321 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
Andi Kleen9dec4472016-02-26 16:27:56 -08001322};
1323 struct perf_event_attr frontend_attrs[] = {
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001324 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
Andi Kleen9dec4472016-02-26 16:27:56 -08001325};
1326 struct perf_event_attr backend_attrs[] = {
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001327 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
Andi Kleen9dec4472016-02-26 16:27:56 -08001328};
1329 struct perf_event_attr default_attrs1[] = {
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001330 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1331 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1332 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1333
1334};
1335
1336/*
1337 * Detailed stats (-d), covering the L1 and last level data caches:
1338 */
1339 struct perf_event_attr detailed_attrs[] = {
1340
1341 { .type = PERF_TYPE_HW_CACHE,
1342 .config =
1343 PERF_COUNT_HW_CACHE_L1D << 0 |
1344 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1345 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1346
1347 { .type = PERF_TYPE_HW_CACHE,
1348 .config =
1349 PERF_COUNT_HW_CACHE_L1D << 0 |
1350 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1351 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1352
1353 { .type = PERF_TYPE_HW_CACHE,
1354 .config =
1355 PERF_COUNT_HW_CACHE_LL << 0 |
1356 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1357 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1358
1359 { .type = PERF_TYPE_HW_CACHE,
1360 .config =
1361 PERF_COUNT_HW_CACHE_LL << 0 |
1362 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1363 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1364};
1365
1366/*
1367 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1368 */
1369 struct perf_event_attr very_detailed_attrs[] = {
1370
1371 { .type = PERF_TYPE_HW_CACHE,
1372 .config =
1373 PERF_COUNT_HW_CACHE_L1I << 0 |
1374 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1375 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1376
1377 { .type = PERF_TYPE_HW_CACHE,
1378 .config =
1379 PERF_COUNT_HW_CACHE_L1I << 0 |
1380 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1381 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1382
1383 { .type = PERF_TYPE_HW_CACHE,
1384 .config =
1385 PERF_COUNT_HW_CACHE_DTLB << 0 |
1386 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1387 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1388
1389 { .type = PERF_TYPE_HW_CACHE,
1390 .config =
1391 PERF_COUNT_HW_CACHE_DTLB << 0 |
1392 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1393 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1394
1395 { .type = PERF_TYPE_HW_CACHE,
1396 .config =
1397 PERF_COUNT_HW_CACHE_ITLB << 0 |
1398 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1399 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1400
1401 { .type = PERF_TYPE_HW_CACHE,
1402 .config =
1403 PERF_COUNT_HW_CACHE_ITLB << 0 |
1404 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1405 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1406
1407};
1408
1409/*
1410 * Very, very detailed stats (-d -d -d), adding prefetch events:
1411 */
1412 struct perf_event_attr very_very_detailed_attrs[] = {
1413
1414 { .type = PERF_TYPE_HW_CACHE,
1415 .config =
1416 PERF_COUNT_HW_CACHE_L1D << 0 |
1417 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1418 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1419
1420 { .type = PERF_TYPE_HW_CACHE,
1421 .config =
1422 PERF_COUNT_HW_CACHE_L1D << 0 |
1423 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1424 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1425};
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001426 struct parse_events_error errinfo;
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001427
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001428 /* Set attrs if no event is selected and !null_run: */
Jiri Olsaaea0dca2018-08-30 08:32:41 +02001429 if (stat_config.null_run)
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001430 return 0;
1431
Ian Rogersa910e462019-11-15 23:46:52 -08001432 bzero(&errinfo, sizeof(errinfo));
Andi Kleen4cabc3d2013-08-21 16:47:26 -07001433 if (transaction_run) {
Thomas Richter742d92f2018-06-26 09:17:01 +02001434 /* Handle -T as -M transaction. Once platform specific metrics
1435 * support has been added to the json files, all archictures
1436 * will use this approach. To determine transaction support
1437 * on an architecture test for such a metric name.
1438 */
1439 if (metricgroup__has_metric("transaction")) {
1440 struct option opt = { .value = &evsel_list };
1441
1442 return metricgroup__parse_groups(&opt, "transaction",
Jiri Olsad0192fd2018-08-30 08:32:51 +02001443 &stat_config.metric_events);
Thomas Richter742d92f2018-06-26 09:17:01 +02001444 }
1445
Andi Kleen4cabc3d2013-08-21 16:47:26 -07001446 if (pmu_have_event("cpu", "cycles-ct") &&
1447 pmu_have_event("cpu", "el-start"))
Thomas Richterfca323402018-03-08 15:57:35 +01001448 err = parse_events(evsel_list, transaction_attrs,
1449 &errinfo);
Andi Kleen4cabc3d2013-08-21 16:47:26 -07001450 else
Thomas Richterfca323402018-03-08 15:57:35 +01001451 err = parse_events(evsel_list,
1452 transaction_limited_attrs,
1453 &errinfo);
Jiri Olsaa4547422015-06-03 16:25:53 +02001454 if (err) {
Andi Kleen4cabc3d2013-08-21 16:47:26 -07001455 fprintf(stderr, "Cannot set up transaction events\n");
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001456 parse_events_print_error(&errinfo, transaction_attrs);
Andi Kleen4cabc3d2013-08-21 16:47:26 -07001457 return -1;
1458 }
1459 return 0;
1460 }
1461
Kan Liangdaefd0b2017-05-26 12:05:38 -07001462 if (smi_cost) {
1463 int smi;
1464
1465 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1466 fprintf(stderr, "freeze_on_smi is not supported.\n");
1467 return -1;
1468 }
1469
1470 if (!smi) {
1471 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1472 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1473 return -1;
1474 }
1475 smi_reset = true;
1476 }
1477
1478 if (pmu_have_event("msr", "aperf") &&
1479 pmu_have_event("msr", "smi")) {
1480 if (!force_metric_only)
Jiri Olsa0ce5aa02018-08-30 08:32:31 +02001481 stat_config.metric_only = true;
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001482 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
Kan Liangdaefd0b2017-05-26 12:05:38 -07001483 } else {
1484 fprintf(stderr, "To measure SMI cost, it needs "
1485 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001486 parse_events_print_error(&errinfo, smi_cost_attrs);
Kan Liangdaefd0b2017-05-26 12:05:38 -07001487 return -1;
1488 }
1489 if (err) {
Ian Rogersa910e462019-11-15 23:46:52 -08001490 parse_events_print_error(&errinfo, smi_cost_attrs);
Kan Liangdaefd0b2017-05-26 12:05:38 -07001491 fprintf(stderr, "Cannot set up SMI cost events\n");
1492 return -1;
1493 }
1494 return 0;
1495 }
1496
Andi Kleen44b1e602016-05-30 12:49:42 -03001497 if (topdown_run) {
1498 char *str = NULL;
1499 bool warn = false;
1500
1501 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1502 stat_config.aggr_mode != AGGR_CORE) {
1503 pr_err("top down event configuration requires --per-core mode\n");
1504 return -1;
1505 }
1506 stat_config.aggr_mode = AGGR_CORE;
1507 if (nr_cgroups || !target__has_cpu(&target)) {
1508 pr_err("top down event configuration requires system-wide mode (-a)\n");
1509 return -1;
1510 }
1511
1512 if (!force_metric_only)
Jiri Olsa0ce5aa02018-08-30 08:32:31 +02001513 stat_config.metric_only = true;
Andi Kleen44b1e602016-05-30 12:49:42 -03001514 if (topdown_filter_events(topdown_attrs, &str,
1515 arch_topdown_check_group(&warn)) < 0) {
1516 pr_err("Out of memory\n");
1517 return -1;
1518 }
1519 if (topdown_attrs[0] && str) {
1520 if (warn)
1521 arch_topdown_group_warn();
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001522 err = parse_events(evsel_list, str, &errinfo);
Andi Kleen44b1e602016-05-30 12:49:42 -03001523 if (err) {
1524 fprintf(stderr,
1525 "Cannot set up top down events %s: %d\n",
1526 str, err);
Jiri Olsaa5cfa622018-06-07 00:15:10 +02001527 parse_events_print_error(&errinfo, str);
Leo Yanc74b0502019-07-02 18:34:11 +08001528 free(str);
Andi Kleen44b1e602016-05-30 12:49:42 -03001529 return -1;
1530 }
1531 } else {
1532 fprintf(stderr, "System does not support topdown\n");
1533 return -1;
1534 }
1535 free(str);
1536 }
1537
Jiri Olsa6484d2f2019-07-21 13:24:28 +02001538 if (!evsel_list->core.nr_entries) {
Namhyung Kima1f3d562016-05-13 15:01:03 +09001539 if (target__has_cpu(&target))
1540 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1541
Andi Kleen9dec4472016-02-26 16:27:56 -08001542 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1543 return -1;
1544 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1545 if (perf_evlist__add_default_attrs(evsel_list,
1546 frontend_attrs) < 0)
1547 return -1;
1548 }
1549 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1550 if (perf_evlist__add_default_attrs(evsel_list,
1551 backend_attrs) < 0)
1552 return -1;
1553 }
1554 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02001555 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001556 }
1557
1558 /* Detailed events get appended to the event list: */
1559
1560 if (detailed_run < 1)
1561 return 0;
1562
1563 /* Append detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03001564 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02001565 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001566
1567 if (detailed_run < 2)
1568 return 0;
1569
1570 /* Append very detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03001571 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
Arnaldo Carvalho de Melo50d08e42011-11-04 09:10:59 -02001572 return -1;
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001573
1574 if (detailed_run < 3)
1575 return 0;
1576
1577 /* Append very, very detailed run extra attributes: */
Arnaldo Carvalho de Melo79695e12012-05-30 13:53:54 -03001578 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02001579}
1580
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01001581static const char * const stat_record_usage[] = {
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001582 "perf stat record [<options>]",
1583 NULL,
1584};
1585
Jiri Olsa3ba78bd2015-11-05 15:40:47 +01001586static void init_features(struct perf_session *session)
1587{
1588 int feat;
1589
1590 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1591 perf_header__set_feat(&session->header, feat);
1592
Jiri Olsa8002a632019-04-09 12:01:56 +02001593 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
Jiri Olsa3ba78bd2015-11-05 15:40:47 +01001594 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1595 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1596 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1597 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1598}
1599
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001600static int __cmd_record(int argc, const char **argv)
1601{
1602 struct perf_session *session;
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01001603 struct perf_data *data = &perf_stat.data;
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001604
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01001605 argc = parse_options(argc, argv, stat_options, stat_record_usage,
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001606 PARSE_OPT_STOP_AT_NON_OPTION);
1607
1608 if (output_name)
Jiri Olsa2d4f2792019-02-21 10:41:30 +01001609 data->path = output_name;
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001610
Jiri Olsad97ae042018-08-30 08:32:36 +02001611 if (stat_config.run_count != 1 || forever) {
Jiri Olsae9d6db8e82015-11-05 15:40:53 +01001612 pr_err("Cannot use -r option with perf stat record.\n");
1613 return -1;
1614 }
1615
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01001616 session = perf_session__new(data, false, NULL);
Mamatha Inamdar6ef81c52019-08-22 12:50:49 +05301617 if (IS_ERR(session)) {
1618 pr_err("Perf session creation failed\n");
1619 return PTR_ERR(session);
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001620 }
1621
Jiri Olsa3ba78bd2015-11-05 15:40:47 +01001622 init_features(session);
1623
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001624 session->evlist = evsel_list;
1625 perf_stat.session = session;
1626 perf_stat.record = true;
1627 return argc;
1628}
1629
Jiri Olsa89f16882018-09-13 14:54:03 +02001630static int process_stat_round_event(struct perf_session *session,
1631 union perf_event *event)
Jiri Olsaa56f9392015-11-05 15:40:59 +01001632{
Jiri Olsa72932372019-08-28 15:57:16 +02001633 struct perf_record_stat_round *stat_round = &event->stat_round;
Jiri Olsa32dcd022019-07-21 13:23:51 +02001634 struct evsel *counter;
Jiri Olsaa56f9392015-11-05 15:40:59 +01001635 struct timespec tsh, *ts = NULL;
1636 const char **argv = session->header.env.cmdline_argv;
1637 int argc = session->header.env.nr_cmdline;
1638
Arnaldo Carvalho de Meloe5cadb92016-06-23 11:26:15 -03001639 evlist__for_each_entry(evsel_list, counter)
Jiri Olsaa56f9392015-11-05 15:40:59 +01001640 perf_stat_process_counter(&stat_config, counter);
1641
Andi Kleene3b03b62016-05-05 16:04:03 -07001642 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1643 update_stats(&walltime_nsecs_stats, stat_round->time);
Jiri Olsaa56f9392015-11-05 15:40:59 +01001644
Andi Kleene3b03b62016-05-05 16:04:03 -07001645 if (stat_config.interval && stat_round->time) {
Arnaldo Carvalho de Melobd48c632016-08-05 15:40:30 -03001646 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
1647 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
Jiri Olsaa56f9392015-11-05 15:40:59 +01001648 ts = &tsh;
1649 }
1650
1651 print_counters(ts, argc, argv);
1652 return 0;
1653}
1654
Jiri Olsa62ba18ba2015-11-05 15:40:57 +01001655static
Jiri Olsa89f16882018-09-13 14:54:03 +02001656int process_stat_config_event(struct perf_session *session,
1657 union perf_event *event)
Jiri Olsa62ba18ba2015-11-05 15:40:57 +01001658{
Jiri Olsa89f16882018-09-13 14:54:03 +02001659 struct perf_tool *tool = session->tool;
Jiri Olsa68d702f2015-11-05 15:40:58 +01001660 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1661
Jiri Olsa62ba18ba2015-11-05 15:40:57 +01001662 perf_event__read_stat_config(&stat_config, &event->stat_config);
Jiri Olsa68d702f2015-11-05 15:40:58 +01001663
Jiri Olsa315c0a12019-08-22 13:11:39 +02001664 if (perf_cpu_map__empty(st->cpus)) {
Jiri Olsa89af4e02015-11-05 15:41:02 +01001665 if (st->aggr_mode != AGGR_UNSET)
1666 pr_warning("warning: processing task data, aggregation mode not set\n");
1667 return 0;
1668 }
1669
1670 if (st->aggr_mode != AGGR_UNSET)
1671 stat_config.aggr_mode = st->aggr_mode;
1672
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01001673 if (perf_stat.data.is_pipe)
Jiri Olsa68d702f2015-11-05 15:40:58 +01001674 perf_stat_init_aggr_mode();
1675 else
1676 perf_stat_init_aggr_mode_file(st);
1677
Jiri Olsa62ba18ba2015-11-05 15:40:57 +01001678 return 0;
1679}
1680
Jiri Olsa1975d362015-11-05 15:40:56 +01001681static int set_maps(struct perf_stat *st)
1682{
1683 if (!st->cpus || !st->threads)
1684 return 0;
1685
1686 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1687 return -EINVAL;
1688
Jiri Olsa453fa032019-07-21 13:24:43 +02001689 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
Jiri Olsa1975d362015-11-05 15:40:56 +01001690
1691 if (perf_evlist__alloc_stats(evsel_list, true))
1692 return -ENOMEM;
1693
1694 st->maps_allocated = true;
1695 return 0;
1696}
1697
1698static
Jiri Olsa89f16882018-09-13 14:54:03 +02001699int process_thread_map_event(struct perf_session *session,
1700 union perf_event *event)
Jiri Olsa1975d362015-11-05 15:40:56 +01001701{
Jiri Olsa89f16882018-09-13 14:54:03 +02001702 struct perf_tool *tool = session->tool;
Jiri Olsa1975d362015-11-05 15:40:56 +01001703 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1704
1705 if (st->threads) {
1706 pr_warning("Extra thread map event, ignoring.\n");
1707 return 0;
1708 }
1709
1710 st->threads = thread_map__new_event(&event->thread_map);
1711 if (!st->threads)
1712 return -ENOMEM;
1713
1714 return set_maps(st);
1715}
1716
1717static
Jiri Olsa89f16882018-09-13 14:54:03 +02001718int process_cpu_map_event(struct perf_session *session,
1719 union perf_event *event)
Jiri Olsa1975d362015-11-05 15:40:56 +01001720{
Jiri Olsa89f16882018-09-13 14:54:03 +02001721 struct perf_tool *tool = session->tool;
Jiri Olsa1975d362015-11-05 15:40:56 +01001722 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
Jiri Olsaf8548392019-07-21 13:23:49 +02001723 struct perf_cpu_map *cpus;
Jiri Olsa1975d362015-11-05 15:40:56 +01001724
1725 if (st->cpus) {
1726 pr_warning("Extra cpu map event, ignoring.\n");
1727 return 0;
1728 }
1729
1730 cpus = cpu_map__new_data(&event->cpu_map.data);
1731 if (!cpus)
1732 return -ENOMEM;
1733
1734 st->cpus = cpus;
1735 return set_maps(st);
1736}
1737
Jin Yao56739442017-12-05 22:03:07 +08001738static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1739{
1740 int i;
1741
1742 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1743 if (!config->stats)
1744 return -1;
1745
1746 config->stats_num = nthreads;
1747
1748 for (i = 0; i < nthreads; i++)
1749 runtime_stat__init(&config->stats[i]);
1750
1751 return 0;
1752}
1753
1754static void runtime_stat_delete(struct perf_stat_config *config)
1755{
1756 int i;
1757
1758 if (!config->stats)
1759 return;
1760
1761 for (i = 0; i < config->stats_num; i++)
1762 runtime_stat__exit(&config->stats[i]);
1763
Arnaldo Carvalho de Melod8f9da22019-07-04 12:06:20 -03001764 zfree(&config->stats);
Jin Yao56739442017-12-05 22:03:07 +08001765}
1766
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01001767static const char * const stat_report_usage[] = {
Jiri Olsaba6039b62015-11-05 15:40:55 +01001768 "perf stat report [<options>]",
1769 NULL,
1770};
1771
1772static struct perf_stat perf_stat = {
1773 .tool = {
1774 .attr = perf_event__process_attr,
Jiri Olsafa6ea782015-11-05 15:41:00 +01001775 .event_update = perf_event__process_event_update,
Jiri Olsa1975d362015-11-05 15:40:56 +01001776 .thread_map = process_thread_map_event,
1777 .cpu_map = process_cpu_map_event,
Jiri Olsa62ba18ba2015-11-05 15:40:57 +01001778 .stat_config = process_stat_config_event,
Jiri Olsaa56f9392015-11-05 15:40:59 +01001779 .stat = perf_event__process_stat_event,
1780 .stat_round = process_stat_round_event,
Jiri Olsaba6039b62015-11-05 15:40:55 +01001781 },
Jiri Olsa89af4e02015-11-05 15:41:02 +01001782 .aggr_mode = AGGR_UNSET,
Jiri Olsaba6039b62015-11-05 15:40:55 +01001783};
1784
1785static int __cmd_report(int argc, const char **argv)
1786{
1787 struct perf_session *session;
1788 const struct option options[] = {
1789 OPT_STRING('i', "input", &input_name, "file", "input file name"),
Jiri Olsa89af4e02015-11-05 15:41:02 +01001790 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1791 "aggregate counts per processor socket", AGGR_SOCKET),
Kan Liangdb5742b2019-06-04 15:50:42 -07001792 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
1793 "aggregate counts per processor die", AGGR_DIE),
Jiri Olsa89af4e02015-11-05 15:41:02 +01001794 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1795 "aggregate counts per physical processor core", AGGR_CORE),
Jiri Olsa86895b42019-08-28 10:17:43 +02001796 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
1797 "aggregate counts per numa node", AGGR_NODE),
Jiri Olsa89af4e02015-11-05 15:41:02 +01001798 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1799 "disable CPU count aggregation", AGGR_NONE),
Jiri Olsaba6039b62015-11-05 15:40:55 +01001800 OPT_END()
1801 };
1802 struct stat st;
1803 int ret;
1804
Jiri Olsa8a59f3c2016-01-12 10:35:29 +01001805 argc = parse_options(argc, argv, options, stat_report_usage, 0);
Jiri Olsaba6039b62015-11-05 15:40:55 +01001806
1807 if (!input_name || !strlen(input_name)) {
1808 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1809 input_name = "-";
1810 else
1811 input_name = "perf.data";
1812 }
1813
Jiri Olsa2d4f2792019-02-21 10:41:30 +01001814 perf_stat.data.path = input_name;
1815 perf_stat.data.mode = PERF_DATA_MODE_READ;
Jiri Olsaba6039b62015-11-05 15:40:55 +01001816
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01001817 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
Mamatha Inamdar6ef81c52019-08-22 12:50:49 +05301818 if (IS_ERR(session))
1819 return PTR_ERR(session);
Jiri Olsaba6039b62015-11-05 15:40:55 +01001820
1821 perf_stat.session = session;
1822 stat_config.output = stderr;
1823 evsel_list = session->evlist;
1824
1825 ret = perf_session__process_events(session);
1826 if (ret)
1827 return ret;
1828
1829 perf_session__delete(session);
1830 return 0;
1831}
1832
Jiri Olsae3ba76d2017-02-27 10:48:18 +01001833static void setup_system_wide(int forks)
1834{
1835 /*
1836 * Make system wide (-a) the default target if
1837 * no target was specified and one of following
1838 * conditions is met:
1839 *
1840 * - there's no workload specified
1841 * - there is workload specified but all requested
1842 * events are system wide events
1843 */
1844 if (!target__none(&target))
1845 return;
1846
1847 if (!forks)
1848 target.system_wide = true;
1849 else {
Jiri Olsa32dcd022019-07-21 13:23:51 +02001850 struct evsel *counter;
Jiri Olsae3ba76d2017-02-27 10:48:18 +01001851
1852 evlist__for_each_entry(evsel_list, counter) {
Jiri Olsa648b5af2019-08-06 11:35:19 +02001853 if (!counter->core.system_wide)
Jiri Olsae3ba76d2017-02-27 10:48:18 +01001854 return;
1855 }
1856
Jiri Olsa6484d2f2019-07-21 13:24:28 +02001857 if (evsel_list->core.nr_entries)
Jiri Olsae3ba76d2017-02-27 10:48:18 +01001858 target.system_wide = true;
1859 }
1860}
1861
Arnaldo Carvalho de Melob0ad8ea2017-03-27 11:47:20 -03001862int cmd_stat(int argc, const char **argv)
Ingo Molnar52425192009-05-26 09:17:18 +02001863{
Arnaldo Carvalho de Melob070a5472012-10-01 15:20:58 -03001864 const char * const stat_usage[] = {
1865 "perf stat [<options>] [<command>]",
1866 NULL
1867 };
Namhyung Kimcc03c542013-11-01 16:33:15 +09001868 int status = -EINVAL, run_idx;
Stephane Eranian4aa90152011-08-15 22:22:33 +02001869 const char *mode;
Jiri Olsa58215222015-07-21 14:31:24 +02001870 FILE *output = stderr;
yuzhoujianf1f8ad52018-01-29 10:25:23 +01001871 unsigned int interval, timeout;
Jiri Olsaba6039b62015-11-05 15:40:55 +01001872 const char * const stat_subcommands[] = { "record", "report" };
Ingo Molnar42202dd2009-06-13 14:57:28 +02001873
Stephane Eranian5af52b52010-05-18 15:00:01 +02001874 setlocale(LC_ALL, "");
1875
Jiri Olsa0f98b112019-07-21 13:23:55 +02001876 evsel_list = evlist__new();
Arnaldo Carvalho de Melo361c99a2011-01-11 20:56:53 -02001877 if (evsel_list == NULL)
1878 return -ENOMEM;
1879
Wang Nan1669e502016-02-19 11:43:58 +00001880 parse_events__shrink_config_terms();
Michael Petlan51433ea2018-12-10 11:00:04 -05001881
1882 /* String-parsing callback-based options would segfault when negated */
1883 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1884 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1885 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1886
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001887 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1888 (const char **) stat_usage,
1889 PARSE_OPT_STOP_AT_NON_OPTION);
Andi Kleen37932c12017-03-20 13:17:08 -07001890 perf_stat__collect_metric_expr(evsel_list);
Andi Kleenfb4605b2016-03-01 10:57:52 -08001891 perf_stat__init_shadow_stats();
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001892
Jiri Olsafa7070a2018-08-30 08:32:29 +02001893 if (stat_config.csv_sep) {
1894 stat_config.csv_output = true;
1895 if (!strcmp(stat_config.csv_sep, "\\t"))
1896 stat_config.csv_sep = "\t";
Jiri Olsa6edb78a2015-11-05 15:41:01 +01001897 } else
Jiri Olsafa7070a2018-08-30 08:32:29 +02001898 stat_config.csv_sep = DEFAULT_SEPARATOR;
Jiri Olsa6edb78a2015-11-05 15:41:01 +01001899
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001900 if (argc && !strncmp(argv[0], "rec", 3)) {
1901 argc = __cmd_record(argc, argv);
1902 if (argc < 0)
1903 return -1;
Jiri Olsaba6039b62015-11-05 15:40:55 +01001904 } else if (argc && !strncmp(argv[0], "rep", 3))
1905 return __cmd_report(argc, argv);
Stephane Eraniand7470b62010-12-01 18:49:05 +02001906
Jiri Olsaec0d3d12015-07-21 14:31:25 +02001907 interval = stat_config.interval;
yuzhoujianf1f8ad52018-01-29 10:25:23 +01001908 timeout = stat_config.timeout;
Jiri Olsaec0d3d12015-07-21 14:31:25 +02001909
Jiri Olsa4979d0c2015-11-05 15:40:46 +01001910 /*
1911 * For record command the -o is already taken care of.
1912 */
1913 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
Stephane Eranian4aa90152011-08-15 22:22:33 +02001914 output = NULL;
1915
Jim Cromie56f3bae2011-09-07 17:14:00 -06001916 if (output_name && output_fd) {
1917 fprintf(stderr, "cannot use both --output and --log-fd\n");
Jiri Olsae0547312015-11-05 15:40:45 +01001918 parse_options_usage(stat_usage, stat_options, "o", 1);
1919 parse_options_usage(NULL, stat_options, "log-fd", 0);
Namhyung Kimcc03c542013-11-01 16:33:15 +09001920 goto out;
Jim Cromie56f3bae2011-09-07 17:14:00 -06001921 }
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02001922
Jiri Olsa0ce5aa02018-08-30 08:32:31 +02001923 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
Andi Kleen54b50912016-03-03 15:57:36 -08001924 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1925 goto out;
1926 }
1927
Jiri Olsad97ae042018-08-30 08:32:36 +02001928 if (stat_config.metric_only && stat_config.run_count > 1) {
Andi Kleen54b50912016-03-03 15:57:36 -08001929 fprintf(stderr, "--metric-only is not supported with -r\n");
1930 goto out;
1931 }
1932
Jiri Olsa54ac0b12018-08-30 08:32:50 +02001933 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
Jiri Olsae55c14a2018-04-23 11:08:21 +02001934 fprintf(stderr, "--table is only supported with -r\n");
1935 parse_options_usage(stat_usage, stat_options, "r", 1);
1936 parse_options_usage(NULL, stat_options, "table", 0);
1937 goto out;
1938 }
1939
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02001940 if (output_fd < 0) {
1941 fprintf(stderr, "argument to --log-fd must be a > 0\n");
Jiri Olsae0547312015-11-05 15:40:45 +01001942 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
Namhyung Kimcc03c542013-11-01 16:33:15 +09001943 goto out;
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02001944 }
1945
Stephane Eranian4aa90152011-08-15 22:22:33 +02001946 if (!output) {
1947 struct timespec tm;
1948 mode = append_file ? "a" : "w";
1949
1950 output = fopen(output_name, mode);
1951 if (!output) {
1952 perror("failed to create output file");
David Ahernfceda7f2012-08-26 12:24:44 -06001953 return -1;
Stephane Eranian4aa90152011-08-15 22:22:33 +02001954 }
1955 clock_gettime(CLOCK_REALTIME, &tm);
1956 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
Stephane Eranianfc3e4d02012-05-15 13:11:11 +02001957 } else if (output_fd > 0) {
Jim Cromie56f3bae2011-09-07 17:14:00 -06001958 mode = append_file ? "a" : "w";
1959 output = fdopen(output_fd, mode);
1960 if (!output) {
1961 perror("Failed opening logfd");
1962 return -errno;
1963 }
Stephane Eranian4aa90152011-08-15 22:22:33 +02001964 }
1965
Jiri Olsa58215222015-07-21 14:31:24 +02001966 stat_config.output = output;
1967
Stephane Eraniand7470b62010-12-01 18:49:05 +02001968 /*
1969 * let the spreadsheet do the pretty-printing
1970 */
Jiri Olsafa7070a2018-08-30 08:32:29 +02001971 if (stat_config.csv_output) {
Jim Cromie61a9f322011-09-07 17:14:04 -06001972 /* User explicitly passed -B? */
Stephane Eraniand7470b62010-12-01 18:49:05 +02001973 if (big_num_opt == 1) {
1974 fprintf(stderr, "-B option not supported with -x\n");
Jiri Olsae0547312015-11-05 15:40:45 +01001975 parse_options_usage(stat_usage, stat_options, "B", 1);
1976 parse_options_usage(NULL, stat_options, "x", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09001977 goto out;
Stephane Eraniand7470b62010-12-01 18:49:05 +02001978 } else /* Nope, so disable big number formatting */
Jiri Olsa34ff0862018-08-30 08:32:47 +02001979 stat_config.big_num = false;
Stephane Eraniand7470b62010-12-01 18:49:05 +02001980 } else if (big_num_opt == 0) /* User passed --no-big-num */
Jiri Olsa34ff0862018-08-30 08:32:47 +02001981 stat_config.big_num = false;
Stephane Eraniand7470b62010-12-01 18:49:05 +02001982
Jiri Olsae3ba76d2017-02-27 10:48:18 +01001983 setup_system_wide(argc);
David Ahernac3063b2013-09-30 07:37:37 -06001984
Jiri Olsa0ce2da12018-06-05 14:13:13 +02001985 /*
1986 * Display user/system times only for single
1987 * run and when there's specified tracee.
1988 */
Jiri Olsad97ae042018-08-30 08:32:36 +02001989 if ((stat_config.run_count == 1) && target__none(&target))
Jiri Olsa8897a892018-08-30 08:32:44 +02001990 stat_config.ru_display = true;
Jiri Olsa0ce2da12018-06-05 14:13:13 +02001991
Jiri Olsad97ae042018-08-30 08:32:36 +02001992 if (stat_config.run_count < 0) {
Namhyung Kimcc03c542013-11-01 16:33:15 +09001993 pr_err("Run count must be a positive number\n");
Jiri Olsae0547312015-11-05 15:40:45 +01001994 parse_options_usage(stat_usage, stat_options, "r", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09001995 goto out;
Jiri Olsad97ae042018-08-30 08:32:36 +02001996 } else if (stat_config.run_count == 0) {
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05001997 forever = true;
Jiri Olsad97ae042018-08-30 08:32:36 +02001998 stat_config.run_count = 1;
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05001999 }
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002000
Jiri Olsa54ac0b12018-08-30 08:32:50 +02002001 if (stat_config.walltime_run_table) {
2002 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2003 if (!stat_config.walltime_run) {
Jiri Olsae55c14a2018-04-23 11:08:21 +02002004 pr_err("failed to setup -r option");
2005 goto out;
2006 }
2007 }
2008
Jin Yao1d9f8d12017-12-05 22:03:10 +08002009 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2010 !target__has_task(&target)) {
2011 if (!target.system_wide || target.cpu_list) {
2012 fprintf(stderr, "The --per-thread option is only "
2013 "available when monitoring via -p -t -a "
2014 "options or only --per-thread.\n");
2015 parse_options_usage(NULL, stat_options, "p", 1);
2016 parse_options_usage(NULL, stat_options, "t", 1);
2017 goto out;
2018 }
Jiri Olsa32b8af82015-06-26 11:29:27 +02002019 }
2020
2021 /*
2022 * no_aggr, cgroup are for system-wide only
2023 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2024 */
Jiri Olsa421a50f2015-07-21 14:31:22 +02002025 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2026 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002027 !target__has_cpu(&target)) {
Stephane Eranian023695d2011-02-14 11:20:01 +02002028 fprintf(stderr, "both cgroup and no-aggregation "
2029 "modes only available in system-wide mode\n");
2030
Jiri Olsae0547312015-11-05 15:40:45 +01002031 parse_options_usage(stat_usage, stat_options, "G", 1);
2032 parse_options_usage(NULL, stat_options, "A", 1);
2033 parse_options_usage(NULL, stat_options, "a", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002034 goto out;
Stephane Eraniand7e7a452013-02-06 15:46:02 +01002035 }
2036
Ingo Molnar2cba3ff2011-05-19 13:30:56 +02002037 if (add_default_attributes())
2038 goto out;
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002039
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002040 target__validate(&target);
Arnaldo Carvalho de Melo5c98d4662011-01-03 17:53:33 -02002041
Jin Yao1d9f8d12017-12-05 22:03:10 +08002042 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2043 target.per_thread = true;
2044
Namhyung Kim77a6f012012-05-07 14:09:04 +09002045 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002046 if (target__has_task(&target)) {
Namhyung Kim77a6f012012-05-07 14:09:04 +09002047 pr_err("Problems finding threads of monitor\n");
Jiri Olsae0547312015-11-05 15:40:45 +01002048 parse_options_usage(stat_usage, stat_options, "p", 1);
2049 parse_options_usage(NULL, stat_options, "t", 1);
Arnaldo Carvalho de Melo602ad872013-11-12 16:46:16 -03002050 } else if (target__has_cpu(&target)) {
Namhyung Kim77a6f012012-05-07 14:09:04 +09002051 perror("failed to parse CPUs map");
Jiri Olsae0547312015-11-05 15:40:45 +01002052 parse_options_usage(stat_usage, stat_options, "C", 1);
2053 parse_options_usage(NULL, stat_options, "a", 1);
Namhyung Kimcc03c542013-11-01 16:33:15 +09002054 }
2055 goto out;
Arnaldo Carvalho de Melo60d567e2011-01-03 17:49:48 -02002056 }
Jiri Olsa32b8af82015-06-26 11:29:27 +02002057
2058 /*
2059 * Initialize thread_map with comm names,
2060 * so we could print it out on output.
2061 */
Jin Yao56739442017-12-05 22:03:07 +08002062 if (stat_config.aggr_mode == AGGR_THREAD) {
Jiri Olsa03617c22019-07-21 13:24:42 +02002063 thread_map__read_comms(evsel_list->core.threads);
Jin Yao56739442017-12-05 22:03:07 +08002064 if (target.system_wide) {
2065 if (runtime_stat_new(&stat_config,
Jiri Olsaa2f354e2019-08-22 13:11:41 +02002066 perf_thread_map__nr(evsel_list->core.threads))) {
Jin Yao56739442017-12-05 22:03:07 +08002067 goto out;
2068 }
2069 }
2070 }
Jiri Olsa32b8af82015-06-26 11:29:27 +02002071
Jiri Olsa86895b42019-08-28 10:17:43 +02002072 if (stat_config.aggr_mode == AGGR_NODE)
2073 cpu__setup_cpunode_map();
2074
yuzhoujiandb06a262018-01-29 10:25:22 +01002075 if (stat_config.times && interval)
2076 interval_count = true;
2077 else if (stat_config.times && !interval) {
2078 pr_err("interval-count option should be used together with "
2079 "interval-print.\n");
2080 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2081 parse_options_usage(stat_usage, stat_options, "I", 1);
2082 goto out;
2083 }
Stephane Eranianc45c6ea2010-05-28 12:00:01 +02002084
yuzhoujianf1f8ad52018-01-29 10:25:23 +01002085 if (timeout && timeout < 100) {
2086 if (timeout < 10) {
2087 pr_err("timeout must be >= 10ms.\n");
2088 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2089 goto out;
2090 } else
2091 pr_warning("timeout < 100ms. "
2092 "The overhead percentage could be high in some cases. "
2093 "Please proceed with caution.\n");
2094 }
2095 if (timeout && interval) {
2096 pr_err("timeout option is not supported with interval-print.\n");
2097 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2098 parse_options_usage(stat_usage, stat_options, "I", 1);
2099 goto out;
2100 }
2101
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002102 if (perf_evlist__alloc_stats(evsel_list, interval))
Arnaldo Carvalho de Melo03ad9742014-01-03 15:56:06 -03002103 goto out;
Zhang, Yanmind6d901c2010-03-18 11:36:05 -03002104
Stephane Eranian86ee6e12013-02-14 13:57:27 +01002105 if (perf_stat_init_aggr_mode())
Arnaldo Carvalho de Melo03ad9742014-01-03 15:56:06 -03002106 goto out;
Stephane Eranian86ee6e12013-02-14 13:57:27 +01002107
Ingo Molnar58d7e992009-05-15 11:03:23 +02002108 /*
Jiri Olsa7d9ad162018-08-30 08:32:14 +02002109 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2110 * while avoiding that older tools show confusing messages.
2111 *
2112 * However for pipe sessions we need to keep it zero,
2113 * because script's perf_evsel__check_attr is triggered
2114 * by attr->sample_type != 0, and we can't run it on
2115 * stat sessions.
2116 */
2117 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2118
2119 /*
Ingo Molnar58d7e992009-05-15 11:03:23 +02002120 * We dont want to block the signals - that would cause
2121 * child tasks to inherit that and Ctrl-C would not work.
2122 * What we want is for Ctrl-C to work in the exec()-ed
2123 * task, but being ignored by perf stat itself:
2124 */
Peter Zijlstraf7b7c262009-06-10 15:55:59 +02002125 atexit(sig_atexit);
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002126 if (!forever)
2127 signal(SIGINT, skip_signal);
Stephane Eranian13370a92013-01-29 12:47:44 +01002128 signal(SIGCHLD, skip_signal);
Ingo Molnar58d7e992009-05-15 11:03:23 +02002129 signal(SIGALRM, skip_signal);
2130 signal(SIGABRT, skip_signal);
2131
Ingo Molnar42202dd2009-06-13 14:57:28 +02002132 status = 0;
Jiri Olsad97ae042018-08-30 08:32:36 +02002133 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2134 if (stat_config.run_count != 1 && verbose > 0)
Stephane Eranian4aa90152011-08-15 22:22:33 +02002135 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2136 run_idx + 1);
Ingo Molnarf9cef0a2011-04-28 18:17:11 +02002137
Srikar Dronamrajub63fd112019-09-04 15:17:37 +05302138 if (run_idx != 0)
2139 perf_evlist__reset_prev_raw_counts(evsel_list);
2140
Jiri Olsae55c14a2018-04-23 11:08:21 +02002141 status = run_perf_stat(argc, argv, run_idx);
Srikar Dronamraju443f2d52019-09-04 15:17:38 +05302142 if (forever && status != -1 && !interval) {
Jiri Olsad4f63a42015-06-26 11:29:26 +02002143 print_counters(NULL, argc, argv);
Jiri Olsa254ecbc72015-06-26 11:29:13 +02002144 perf_stat__reset_stats();
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002145 }
Ingo Molnar42202dd2009-06-13 14:57:28 +02002146 }
2147
Frederik Deweerdta7e191c2013-03-01 13:02:27 -05002148 if (!forever && status != -1 && !interval)
Jiri Olsad4f63a42015-06-26 11:29:26 +02002149 print_counters(NULL, argc, argv);
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002150
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002151 if (STAT_RECORD) {
2152 /*
2153 * We synthesize the kernel mmap record just so that older tools
2154 * don't emit warnings about not being able to resolve symbols
2155 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2156 * a saner message about no samples being in the perf.data file.
2157 *
2158 * This also serves to suppress a warning about f_header.data.size == 0
Jiri Olsa8b99b1a2015-11-05 15:40:48 +01002159 * in header.c at the moment 'perf stat record' gets introduced, which
2160 * is not really needed once we start adding the stat specific PERF_RECORD_
2161 * records, but the need to suppress the kptr_restrict messages in older
2162 * tools remain -acme
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002163 */
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01002164 int fd = perf_data__fd(&perf_stat.data);
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002165 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2166 process_synthesized_event,
2167 &perf_stat.session->machines.host);
2168 if (err) {
2169 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2170 "older tools may produce warnings about this file\n.");
2171 }
2172
Jiri Olsa7aad0c32015-11-05 15:40:52 +01002173 if (!interval) {
2174 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2175 pr_err("failed to write stat round event\n");
2176 }
2177
Jiri Olsa8ceb41d2017-01-23 22:07:59 +01002178 if (!perf_stat.data.is_pipe) {
Jiri Olsa664c98d2015-11-05 15:40:50 +01002179 perf_stat.session->header.data_size += perf_stat.bytes_written;
2180 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2181 }
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002182
Jiri Olsa750b4ed2019-07-21 13:24:07 +02002183 evlist__close(evsel_list);
Jiri Olsa4979d0c2015-11-05 15:40:46 +01002184 perf_session__delete(perf_stat.session);
2185 }
2186
Masami Hiramatsu544c2ae2015-12-09 11:11:27 +09002187 perf_stat__exit_aggr_mode();
Arnaldo Carvalho de Melod134ffb2013-03-18 11:24:21 -03002188 perf_evlist__free_stats(evsel_list);
Arnaldo Carvalho de Melo0015e2e2011-02-01 16:18:10 -02002189out:
Arnaldo Carvalho de Melod8f9da22019-07-04 12:06:20 -03002190 zfree(&stat_config.walltime_run);
Jiri Olsae55c14a2018-04-23 11:08:21 +02002191
Kan Liangdaefd0b2017-05-26 12:05:38 -07002192 if (smi_cost && smi_reset)
2193 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2194
Jiri Olsac12995a2019-07-21 13:23:56 +02002195 evlist__delete(evsel_list);
Jin Yao56739442017-12-05 22:03:07 +08002196
2197 runtime_stat_delete(&stat_config);
2198
Ingo Molnar42202dd2009-06-13 14:57:28 +02002199 return status;
Ingo Molnarddcacfa2009-04-20 15:37:32 +02002200}