]>
Commit | Line | Data |
---|---|---|
91007045 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
ddcacfa0 | 2 | /* |
bf9e1876 IM |
3 | * 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: | |
ddcacfa0 | 9 | |
2cba3ffb | 10 | $ perf stat ./hackbench 10 |
ddcacfa0 | 11 | |
2cba3ffb | 12 | Time: 0.118 |
ddcacfa0 | 13 | |
2cba3ffb | 14 | Performance counter stats for './hackbench 10': |
ddcacfa0 | 15 | |
2cba3ffb IM |
16 | 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 | |
ddcacfa0 | 29 | |
5242519b | 30 | * |
2cba3ffb | 31 | * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
5242519b IM |
32 | * |
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> | |
6e750a8f | 40 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
ddcacfa0 IM |
41 | */ |
42 | ||
16f762a2 | 43 | #include "builtin.h" |
c1a604df | 44 | #include "perf.h" |
f14d5707 | 45 | #include "util/cgroup.h" |
4b6ab94e | 46 | #include <subcmd/parse-options.h> |
5242519b | 47 | #include "util/parse-events.h" |
4cabc3d1 | 48 | #include "util/pmu.h" |
8f28827a | 49 | #include "util/event.h" |
361c99a6 | 50 | #include "util/evlist.h" |
69aad6f1 | 51 | #include "util/evsel.h" |
8f28827a | 52 | #include "util/debug.h" |
a5d243d0 | 53 | #include "util/color.h" |
0007ecea | 54 | #include "util/stat.h" |
60666c63 | 55 | #include "util/header.h" |
a12b51c4 | 56 | #include "util/cpumap.h" |
fd78260b | 57 | #include "util/thread_map.h" |
d809560b | 58 | #include "util/counts.h" |
687986bb | 59 | #include "util/topdown.h" |
4979d0c7 | 60 | #include "util/session.h" |
ba6039b6 | 61 | #include "util/tool.h" |
a067558e | 62 | #include "util/string2.h" |
b18f3e36 | 63 | #include "util/metricgroup.h" |
ea49e01c | 64 | #include "util/synthetic-events.h" |
aeb00b1a | 65 | #include "util/target.h" |
f3711020 | 66 | #include "util/time-utils.h" |
9660e08e | 67 | #include "util/top.h" |
4804e011 | 68 | #include "util/affinity.h" |
70943490 | 69 | #include "util/pfm.h" |
ba6039b6 | 70 | #include "asm/bug.h" |
ddcacfa0 | 71 | |
bd48c63e | 72 | #include <linux/time64.h> |
7f7c536f | 73 | #include <linux/zalloc.h> |
44b1e60a | 74 | #include <api/fs/fs.h> |
a43783ae | 75 | #include <errno.h> |
9607ad3a | 76 | #include <signal.h> |
1f16c575 | 77 | #include <stdlib.h> |
ddcacfa0 | 78 | #include <sys/prctl.h> |
fd20e811 | 79 | #include <inttypes.h> |
5af52b51 | 80 | #include <locale.h> |
e3b03b6c | 81 | #include <math.h> |
7a8ef4c4 ACM |
82 | #include <sys/types.h> |
83 | #include <sys/stat.h> | |
4208735d | 84 | #include <sys/wait.h> |
7a8ef4c4 | 85 | #include <unistd.h> |
0ce2da14 JO |
86 | #include <sys/time.h> |
87 | #include <sys/resource.h> | |
6ef81c55 | 88 | #include <linux/err.h> |
16c8a109 | 89 | |
3052ba56 | 90 | #include <linux/ctype.h> |
453fa030 | 91 | #include <perf/evlist.h> |
3d689ed6 | 92 | |
d7470b6a | 93 | #define DEFAULT_SEPARATOR " " |
daefd0bc | 94 | #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi" |
d7470b6a | 95 | |
d4f63a47 | 96 | static void print_counters(struct timespec *ts, int argc, const char **argv); |
13370a9b | 97 | |
4cabc3d1 | 98 | /* Default events used for perf stat -T */ |
a454742c JO |
99 | static const char *transaction_attrs = { |
100 | "task-clock," | |
4cabc3d1 AK |
101 | "{" |
102 | "instructions," | |
103 | "cycles," | |
104 | "cpu/cycles-t/," | |
105 | "cpu/tx-start/," | |
106 | "cpu/el-start/," | |
107 | "cpu/cycles-ct/" | |
108 | "}" | |
109 | }; | |
110 | ||
111 | /* More limited version when the CPU does not have all events. */ | |
a454742c JO |
112 | static const char * transaction_limited_attrs = { |
113 | "task-clock," | |
4cabc3d1 AK |
114 | "{" |
115 | "instructions," | |
116 | "cycles," | |
117 | "cpu/cycles-t/," | |
118 | "cpu/tx-start/" | |
119 | "}" | |
120 | }; | |
121 | ||
44b1e60a AK |
122 | static const char * topdown_attrs[] = { |
123 | "topdown-total-slots", | |
124 | "topdown-slots-retired", | |
125 | "topdown-recovery-bubbles", | |
126 | "topdown-fetch-bubbles", | |
127 | "topdown-slots-issued", | |
128 | NULL, | |
129 | }; | |
130 | ||
55c36a9f AK |
131 | static const char *topdown_metric_attrs[] = { |
132 | "slots", | |
133 | "topdown-retiring", | |
134 | "topdown-bad-spec", | |
135 | "topdown-fe-bound", | |
136 | "topdown-be-bound", | |
137 | NULL, | |
138 | }; | |
139 | ||
daefd0bc KL |
140 | static const char *smi_cost_attrs = { |
141 | "{" | |
142 | "msr/aperf/," | |
143 | "msr/smi/," | |
144 | "cycles" | |
145 | "}" | |
146 | }; | |
147 | ||
63503dba | 148 | static struct evlist *evsel_list; |
361c99a6 | 149 | |
602ad878 | 150 | static struct target target = { |
77a6f014 NK |
151 | .uid = UINT_MAX, |
152 | }; | |
ddcacfa0 | 153 | |
c1a1f5d9 JO |
154 | #define METRIC_ONLY_LEN 20 |
155 | ||
d07f0b12 | 156 | static volatile pid_t child_pid = -1; |
2cba3ffb | 157 | static int detailed_run = 0; |
4cabc3d1 | 158 | static bool transaction_run; |
44b1e60a | 159 | static bool topdown_run = false; |
daefd0bc KL |
160 | static bool smi_cost = false; |
161 | static bool smi_reset = false; | |
d7470b6a | 162 | static int big_num_opt = -1; |
43bece79 | 163 | static bool group = false; |
1f16c575 PZ |
164 | static const char *pre_cmd = NULL; |
165 | static const char *post_cmd = NULL; | |
166 | static bool sync_run = false; | |
a7e191c3 | 167 | static bool forever = false; |
44b1e60a | 168 | static bool force_metric_only = false; |
13370a9b | 169 | static struct timespec ref_time; |
e0547311 | 170 | static bool append_file; |
db06a269 | 171 | static bool interval_count; |
e0547311 JO |
172 | static const char *output_name; |
173 | static int output_fd; | |
5af52b51 | 174 | |
4979d0c7 JO |
175 | struct perf_stat { |
176 | bool record; | |
8ceb41d7 | 177 | struct perf_data data; |
4979d0c7 JO |
178 | struct perf_session *session; |
179 | u64 bytes_written; | |
ba6039b6 | 180 | struct perf_tool tool; |
1975d36e | 181 | bool maps_allocated; |
f854839b | 182 | struct perf_cpu_map *cpus; |
9749b90e | 183 | struct perf_thread_map *threads; |
89af4e05 | 184 | enum aggr_mode aggr_mode; |
4979d0c7 JO |
185 | }; |
186 | ||
187 | static struct perf_stat perf_stat; | |
188 | #define STAT_RECORD perf_stat.record | |
189 | ||
60666c63 LW |
190 | static volatile int done = 0; |
191 | ||
421a50f3 | 192 | static struct perf_stat_config stat_config = { |
26893a60 JO |
193 | .aggr_mode = AGGR_GLOBAL, |
194 | .scale = true, | |
195 | .unit_width = 4, /* strlen("unit") */ | |
196 | .run_count = 1, | |
197 | .metric_only_len = METRIC_ONLY_LEN, | |
198 | .walltime_nsecs_stats = &walltime_nsecs_stats, | |
34ff0866 | 199 | .big_num = true, |
27e9769a AB |
200 | .ctl_fd = -1, |
201 | .ctl_fd_ack = -1 | |
421a50f3 JO |
202 | }; |
203 | ||
a9a17902 JO |
204 | static bool cpus_map_matched(struct evsel *a, struct evsel *b) |
205 | { | |
206 | if (!a->core.cpus && !b->core.cpus) | |
207 | return true; | |
208 | ||
209 | if (!a->core.cpus || !b->core.cpus) | |
210 | return false; | |
211 | ||
212 | if (a->core.cpus->nr != b->core.cpus->nr) | |
213 | return false; | |
214 | ||
215 | for (int i = 0; i < a->core.cpus->nr; i++) { | |
216 | if (a->core.cpus->map[i] != b->core.cpus->map[i]) | |
217 | return false; | |
218 | } | |
219 | ||
220 | return true; | |
221 | } | |
222 | ||
223 | static void evlist__check_cpu_maps(struct evlist *evlist) | |
224 | { | |
225 | struct evsel *evsel, *pos, *leader; | |
226 | char buf[1024]; | |
227 | ||
228 | evlist__for_each_entry(evlist, evsel) { | |
229 | leader = evsel->leader; | |
230 | ||
231 | /* Check that leader matches cpus with each member. */ | |
232 | if (leader == evsel) | |
233 | continue; | |
234 | if (cpus_map_matched(leader, evsel)) | |
235 | continue; | |
236 | ||
237 | /* If there's mismatch disable the group and warn user. */ | |
238 | WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n"); | |
239 | evsel__group_desc(leader, buf, sizeof(buf)); | |
240 | pr_warning(" %s\n", buf); | |
241 | ||
242 | if (verbose) { | |
243 | cpu_map__snprint(leader->core.cpus, buf, sizeof(buf)); | |
244 | pr_warning(" %s: %s\n", leader->name, buf); | |
245 | cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf)); | |
246 | pr_warning(" %s: %s\n", evsel->name, buf); | |
247 | } | |
248 | ||
249 | for_each_group_evsel(pos, leader) { | |
250 | pos->leader = pos; | |
251 | pos->core.nr_members = 0; | |
252 | } | |
253 | evsel->leader->core.nr_members = 0; | |
254 | } | |
255 | } | |
256 | ||
13370a9b SE |
257 | static inline void diff_timespec(struct timespec *r, struct timespec *a, |
258 | struct timespec *b) | |
259 | { | |
260 | r->tv_sec = a->tv_sec - b->tv_sec; | |
261 | if (a->tv_nsec < b->tv_nsec) { | |
310ebb93 | 262 | r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; |
13370a9b SE |
263 | r->tv_sec--; |
264 | } else { | |
265 | r->tv_nsec = a->tv_nsec - b->tv_nsec ; | |
266 | } | |
267 | } | |
268 | ||
254ecbc7 JO |
269 | static void perf_stat__reset_stats(void) |
270 | { | |
56739444 JY |
271 | int i; |
272 | ||
254ecbc7 | 273 | perf_evlist__reset_stats(evsel_list); |
f87027b9 | 274 | perf_stat__reset_shadow_stats(); |
56739444 JY |
275 | |
276 | for (i = 0; i < stat_config.stats_num; i++) | |
277 | perf_stat__reset_shadow_per_stat(&stat_config.stats[i]); | |
1eda3b21 JO |
278 | } |
279 | ||
8b99b1a4 JO |
280 | static int process_synthesized_event(struct perf_tool *tool __maybe_unused, |
281 | union perf_event *event, | |
282 | struct perf_sample *sample __maybe_unused, | |
283 | struct machine *machine __maybe_unused) | |
4979d0c7 | 284 | { |
8ceb41d7 | 285 | if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { |
4979d0c7 JO |
286 | pr_err("failed to write perf data, error: %m\n"); |
287 | return -1; | |
288 | } | |
289 | ||
8b99b1a4 | 290 | perf_stat.bytes_written += event->header.size; |
4979d0c7 JO |
291 | return 0; |
292 | } | |
293 | ||
1975d36e | 294 | static int write_stat_round_event(u64 tm, u64 type) |
7aad0c32 | 295 | { |
1975d36e | 296 | return perf_event__synthesize_stat_round(NULL, tm, type, |
7aad0c32 JO |
297 | process_synthesized_event, |
298 | NULL); | |
299 | } | |
300 | ||
301 | #define WRITE_STAT_ROUND_EVENT(time, interval) \ | |
302 | write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) | |
303 | ||
8cd36f3e | 304 | #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) |
5a6ea81b | 305 | |
ddc6999e ACM |
306 | static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread, |
307 | struct perf_counts_values *count) | |
5a6ea81b JO |
308 | { |
309 | struct perf_sample_id *sid = SID(counter, cpu, thread); | |
310 | ||
311 | return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, | |
312 | process_synthesized_event, NULL); | |
313 | } | |
314 | ||
32dcd021 | 315 | static int read_single_counter(struct evsel *counter, int cpu, |
f0fbb114 AK |
316 | int thread, struct timespec *rs) |
317 | { | |
318 | if (counter->tool_event == PERF_TOOL_DURATION_TIME) { | |
319 | u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL; | |
320 | struct perf_counts_values *count = | |
321 | perf_counts(counter->counts, cpu, thread); | |
322 | count->ena = count->run = val; | |
323 | count->val = val; | |
324 | return 0; | |
325 | } | |
ea089692 | 326 | return evsel__read_counter(counter, cpu, thread); |
f0fbb114 AK |
327 | } |
328 | ||
f5b4a9c3 SE |
329 | /* |
330 | * Read out the results of a single counter: | |
331 | * do not aggregate counts across CPUs in system-wide mode | |
332 | */ | |
4b49ab70 | 333 | static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu) |
f5b4a9c3 | 334 | { |
a2f354e3 | 335 | int nthreads = perf_thread_map__nr(evsel_list->core.threads); |
4b49ab70 | 336 | int thread; |
f5b4a9c3 | 337 | |
3b4331d9 SP |
338 | if (!counter->supported) |
339 | return -ENOENT; | |
340 | ||
648b5af3 | 341 | if (counter->core.system_wide) |
9bf1a529 JO |
342 | nthreads = 1; |
343 | ||
344 | for (thread = 0; thread < nthreads; thread++) { | |
4b49ab70 AK |
345 | struct perf_counts_values *count; |
346 | ||
347 | count = perf_counts(counter->counts, cpu, thread); | |
348 | ||
349 | /* | |
350 | * The leader's group read loads data into its group members | |
ea089692 | 351 | * (via evsel__read_counter()) and sets their count->loaded. |
4b49ab70 AK |
352 | */ |
353 | if (!perf_counts__is_loaded(counter->counts, cpu, thread) && | |
354 | read_single_counter(counter, cpu, thread, rs)) { | |
355 | counter->counts->scaled = -1; | |
356 | perf_counts(counter->counts, cpu, thread)->ena = 0; | |
357 | perf_counts(counter->counts, cpu, thread)->run = 0; | |
358 | return -1; | |
359 | } | |
5a6ea81b | 360 | |
4b49ab70 | 361 | perf_counts__set_loaded(counter->counts, cpu, thread, false); |
82bf311e | 362 | |
4b49ab70 | 363 | if (STAT_RECORD) { |
ddc6999e | 364 | if (evsel__write_stat_event(counter, cpu, thread, count)) { |
4b49ab70 AK |
365 | pr_err("failed to write stat event\n"); |
366 | return -1; | |
5a6ea81b | 367 | } |
4b49ab70 | 368 | } |
0b1abbf4 | 369 | |
4b49ab70 AK |
370 | if (verbose > 1) { |
371 | fprintf(stat_config.output, | |
372 | "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", | |
8ab2e96d | 373 | evsel__name(counter), |
4b49ab70 AK |
374 | cpu, |
375 | count->val, count->ena, count->run); | |
9bf1a529 | 376 | } |
f5b4a9c3 | 377 | } |
c52b12ed ACM |
378 | |
379 | return 0; | |
2996f5dd IM |
380 | } |
381 | ||
c7e5b328 | 382 | static int read_affinity_counters(struct timespec *rs) |
13370a9b | 383 | { |
32dcd021 | 384 | struct evsel *counter; |
4b49ab70 AK |
385 | struct affinity affinity; |
386 | int i, ncpus, cpu; | |
387 | ||
388 | if (affinity__setup(&affinity) < 0) | |
c7e5b328 | 389 | return -1; |
4b49ab70 AK |
390 | |
391 | ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus); | |
392 | if (!target__has_cpu(&target) || target__has_per_thread(&target)) | |
393 | ncpus = 1; | |
394 | evlist__for_each_cpu(evsel_list, i, cpu) { | |
395 | if (i >= ncpus) | |
396 | break; | |
397 | affinity__set(&affinity, cpu); | |
398 | ||
399 | evlist__for_each_entry(evsel_list, counter) { | |
400 | if (evsel__cpu_iter_skip(counter, cpu)) | |
401 | continue; | |
402 | if (!counter->err) { | |
403 | counter->err = read_counter_cpu(counter, rs, | |
404 | counter->cpu_iter - 1); | |
405 | } | |
406 | } | |
407 | } | |
408 | affinity__cleanup(&affinity); | |
c7e5b328 JY |
409 | return 0; |
410 | } | |
411 | ||
412 | static void read_counters(struct timespec *rs) | |
413 | { | |
414 | struct evsel *counter; | |
415 | ||
ee6a9614 | 416 | if (!stat_config.stop_read_counter && (read_affinity_counters(rs) < 0)) |
c7e5b328 | 417 | return; |
13370a9b | 418 | |
e5cadb93 | 419 | evlist__for_each_entry(evsel_list, counter) { |
4b49ab70 | 420 | if (counter->err) |
245bad8e | 421 | pr_debug("failed to read counter %s\n", counter->name); |
4b49ab70 | 422 | if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter)) |
3b3eb044 | 423 | pr_warning("failed to process counter %s\n", counter->name); |
4b49ab70 | 424 | counter->err = 0; |
13370a9b | 425 | } |
106a94a0 JO |
426 | } |
427 | ||
72f02a94 JY |
428 | static int runtime_stat_new(struct perf_stat_config *config, int nthreads) |
429 | { | |
430 | int i; | |
431 | ||
432 | config->stats = calloc(nthreads, sizeof(struct runtime_stat)); | |
433 | if (!config->stats) | |
434 | return -1; | |
435 | ||
436 | config->stats_num = nthreads; | |
437 | ||
438 | for (i = 0; i < nthreads; i++) | |
439 | runtime_stat__init(&config->stats[i]); | |
440 | ||
441 | return 0; | |
442 | } | |
443 | ||
444 | static void runtime_stat_delete(struct perf_stat_config *config) | |
445 | { | |
446 | int i; | |
447 | ||
448 | if (!config->stats) | |
449 | return; | |
450 | ||
451 | for (i = 0; i < config->stats_num; i++) | |
452 | runtime_stat__exit(&config->stats[i]); | |
453 | ||
454 | zfree(&config->stats); | |
455 | } | |
456 | ||
457 | static void runtime_stat_reset(struct perf_stat_config *config) | |
458 | { | |
459 | int i; | |
460 | ||
461 | if (!config->stats) | |
462 | return; | |
463 | ||
464 | for (i = 0; i < config->stats_num; i++) | |
465 | perf_stat__reset_shadow_per_stat(&config->stats[i]); | |
466 | } | |
467 | ||
ba411a95 | 468 | static void process_interval(void) |
106a94a0 | 469 | { |
106a94a0 | 470 | struct timespec ts, rs; |
106a94a0 | 471 | |
13370a9b SE |
472 | clock_gettime(CLOCK_MONOTONIC, &ts); |
473 | diff_timespec(&rs, &ts, &ref_time); | |
13370a9b | 474 | |
197ba86f | 475 | perf_stat__reset_shadow_per_stat(&rt_stat); |
72f02a94 | 476 | runtime_stat_reset(&stat_config); |
f0fbb114 AK |
477 | read_counters(&rs); |
478 | ||
7aad0c32 | 479 | if (STAT_RECORD) { |
bd48c63e | 480 | if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) |
7aad0c32 JO |
481 | pr_err("failed to write stat round event\n"); |
482 | } | |
483 | ||
b90f1333 | 484 | init_stats(&walltime_nsecs_stats); |
ea9eb1f4 | 485 | update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL); |
d4f63a47 | 486 | print_counters(&rs, 0, NULL); |
13370a9b SE |
487 | } |
488 | ||
dece3a4d AB |
489 | static bool handle_interval(unsigned int interval, int *times) |
490 | { | |
491 | if (interval) { | |
492 | process_interval(); | |
493 | if (interval_count && !(--(*times))) | |
494 | return true; | |
495 | } | |
496 | return false; | |
497 | } | |
498 | ||
67ccdecd | 499 | static void enable_counters(void) |
41191688 | 500 | { |
2162b9c6 AB |
501 | if (stat_config.initial_delay < 0) { |
502 | pr_info(EVLIST_DISABLED_MSG); | |
503 | return; | |
504 | } | |
505 | ||
506 | if (stat_config.initial_delay > 0) { | |
507 | pr_info(EVLIST_DISABLED_MSG); | |
728c0ee0 | 508 | usleep(stat_config.initial_delay * USEC_PER_MSEC); |
2162b9c6 | 509 | } |
67ccdecd JO |
510 | |
511 | /* | |
512 | * We need to enable counters only if: | |
513 | * - we don't have tracee (attaching to task or cpu) | |
514 | * - we have initial delay configured | |
515 | */ | |
2162b9c6 | 516 | if (!target__none(&target) || stat_config.initial_delay) { |
1c87f165 | 517 | evlist__enable(evsel_list); |
2162b9c6 AB |
518 | if (stat_config.initial_delay > 0) |
519 | pr_info(EVLIST_ENABLED_MSG); | |
520 | } | |
41191688 AK |
521 | } |
522 | ||
3df33eff MR |
523 | static void disable_counters(void) |
524 | { | |
525 | /* | |
526 | * If we don't have tracee (attaching to task or cpu), counters may | |
527 | * still be running. To get accurate group ratios, we must stop groups | |
528 | * from counting before reading their constituent counters. | |
529 | */ | |
530 | if (!target__none(&target)) | |
e74676de | 531 | evlist__disable(evsel_list); |
3df33eff MR |
532 | } |
533 | ||
f33cbe72 | 534 | static volatile int workload_exec_errno; |
6af206fd ACM |
535 | |
536 | /* | |
537 | * perf_evlist__prepare_workload will send a SIGUSR1 | |
538 | * if the fork fails, since we asked by setting its | |
539 | * want_signal to true. | |
540 | */ | |
f33cbe72 ACM |
541 | static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, |
542 | void *ucontext __maybe_unused) | |
6af206fd | 543 | { |
f33cbe72 | 544 | workload_exec_errno = info->si_value.sival_int; |
6af206fd ACM |
545 | } |
546 | ||
ddc6999e | 547 | static bool evsel__should_store_id(struct evsel *counter) |
82bf311e | 548 | { |
1fc632ce | 549 | return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; |
82bf311e JO |
550 | } |
551 | ||
cbb5df7e | 552 | static bool is_target_alive(struct target *_target, |
9749b90e | 553 | struct perf_thread_map *threads) |
cbb5df7e JO |
554 | { |
555 | struct stat st; | |
556 | int i; | |
557 | ||
558 | if (!target__has_task(_target)) | |
559 | return true; | |
560 | ||
561 | for (i = 0; i < threads->nr; i++) { | |
562 | char path[PATH_MAX]; | |
563 | ||
564 | scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), | |
565 | threads->map[i].pid); | |
566 | ||
567 | if (!stat(path, &st)) | |
568 | return true; | |
569 | } | |
570 | ||
571 | return false; | |
572 | } | |
573 | ||
bee328cb AB |
574 | static void process_evlist(struct evlist *evlist, unsigned int interval) |
575 | { | |
576 | enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED; | |
577 | ||
578 | if (evlist__ctlfd_process(evlist, &cmd) > 0) { | |
579 | switch (cmd) { | |
580 | case EVLIST_CTL_CMD_ENABLE: | |
581 | pr_info(EVLIST_ENABLED_MSG); | |
582 | if (interval) | |
583 | process_interval(); | |
584 | break; | |
585 | case EVLIST_CTL_CMD_DISABLE: | |
586 | if (interval) | |
587 | process_interval(); | |
588 | pr_info(EVLIST_DISABLED_MSG); | |
589 | break; | |
d20aff15 | 590 | case EVLIST_CTL_CMD_SNAPSHOT: |
bee328cb AB |
591 | case EVLIST_CTL_CMD_ACK: |
592 | case EVLIST_CTL_CMD_UNSUPPORTED: | |
593 | default: | |
594 | break; | |
595 | } | |
596 | } | |
597 | } | |
598 | ||
599 | static void compute_tts(struct timespec *time_start, struct timespec *time_stop, | |
600 | int *time_to_sleep) | |
601 | { | |
602 | int tts = *time_to_sleep; | |
603 | struct timespec time_diff; | |
604 | ||
605 | diff_timespec(&time_diff, time_stop, time_start); | |
606 | ||
607 | tts -= time_diff.tv_sec * MSEC_PER_SEC + | |
608 | time_diff.tv_nsec / NSEC_PER_MSEC; | |
609 | ||
610 | if (tts < 0) | |
611 | tts = 0; | |
612 | ||
613 | *time_to_sleep = tts; | |
614 | } | |
615 | ||
616 | static int dispatch_events(bool forks, int timeout, int interval, int *times) | |
987b8238 AB |
617 | { |
618 | int child_exited = 0, status = 0; | |
bee328cb AB |
619 | int time_to_sleep, sleep_time; |
620 | struct timespec time_start, time_stop; | |
621 | ||
622 | if (interval) | |
623 | sleep_time = interval; | |
624 | else if (timeout) | |
625 | sleep_time = timeout; | |
626 | else | |
627 | sleep_time = 1000; | |
628 | ||
629 | time_to_sleep = sleep_time; | |
987b8238 AB |
630 | |
631 | while (!done) { | |
632 | if (forks) | |
633 | child_exited = waitpid(child_pid, &status, WNOHANG); | |
634 | else | |
635 | child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0; | |
636 | ||
637 | if (child_exited) | |
638 | break; | |
639 | ||
bee328cb AB |
640 | clock_gettime(CLOCK_MONOTONIC, &time_start); |
641 | if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */ | |
642 | if (timeout || handle_interval(interval, times)) | |
643 | break; | |
644 | time_to_sleep = sleep_time; | |
645 | } else { /* fd revent */ | |
646 | process_evlist(evsel_list, interval); | |
647 | clock_gettime(CLOCK_MONOTONIC, &time_stop); | |
648 | compute_tts(&time_start, &time_stop, &time_to_sleep); | |
649 | } | |
987b8238 AB |
650 | } |
651 | ||
652 | return status; | |
653 | } | |
654 | ||
e0e6a6ca AK |
655 | enum counter_recovery { |
656 | COUNTER_SKIP, | |
657 | COUNTER_RETRY, | |
658 | COUNTER_FATAL, | |
659 | }; | |
660 | ||
661 | static enum counter_recovery stat_handle_error(struct evsel *counter) | |
662 | { | |
663 | char msg[BUFSIZ]; | |
664 | /* | |
665 | * PPC returns ENXIO for HW counters until 2.6.37 | |
666 | * (behavior changed with commit b0a873e). | |
667 | */ | |
668 | if (errno == EINVAL || errno == ENOSYS || | |
669 | errno == ENOENT || errno == EOPNOTSUPP || | |
670 | errno == ENXIO) { | |
671 | if (verbose > 0) | |
672 | ui__warning("%s event is not supported by the kernel.\n", | |
8ab2e96d | 673 | evsel__name(counter)); |
e0e6a6ca | 674 | counter->supported = false; |
4804e011 AK |
675 | /* |
676 | * errored is a sticky flag that means one of the counter's | |
677 | * cpu event had a problem and needs to be reexamined. | |
678 | */ | |
679 | counter->errored = true; | |
e0e6a6ca AK |
680 | |
681 | if ((counter->leader != counter) || | |
682 | !(counter->leader->core.nr_members > 1)) | |
683 | return COUNTER_SKIP; | |
ae430892 | 684 | } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) { |
e0e6a6ca AK |
685 | if (verbose > 0) |
686 | ui__warning("%s\n", msg); | |
687 | return COUNTER_RETRY; | |
688 | } else if (target__has_per_thread(&target) && | |
689 | evsel_list->core.threads && | |
690 | evsel_list->core.threads->err_thread != -1) { | |
691 | /* | |
692 | * For global --per-thread case, skip current | |
693 | * error thread. | |
694 | */ | |
695 | if (!thread_map__remove(evsel_list->core.threads, | |
696 | evsel_list->core.threads->err_thread)) { | |
697 | evsel_list->core.threads->err_thread = -1; | |
698 | return COUNTER_RETRY; | |
699 | } | |
700 | } | |
701 | ||
2bb72dbb | 702 | evsel__open_strerror(counter, &target, errno, msg, sizeof(msg)); |
e0e6a6ca AK |
703 | ui__error("%s\n", msg); |
704 | ||
705 | if (child_pid != -1) | |
706 | kill(child_pid, SIGTERM); | |
707 | return COUNTER_FATAL; | |
708 | } | |
709 | ||
e55c14af | 710 | static int __run_perf_stat(int argc, const char **argv, int run_idx) |
42202dd5 | 711 | { |
ec0d3d1f | 712 | int interval = stat_config.interval; |
db06a269 | 713 | int times = stat_config.times; |
f1f8ad52 | 714 | int timeout = stat_config.timeout; |
d6195a6a | 715 | char msg[BUFSIZ]; |
42202dd5 | 716 | unsigned long long t0, t1; |
32dcd021 | 717 | struct evsel *counter; |
410136f5 | 718 | size_t l; |
42202dd5 | 719 | int status = 0; |
6be2850e | 720 | const bool forks = (argc > 0); |
8ceb41d7 | 721 | bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; |
4804e011 AK |
722 | struct affinity affinity; |
723 | int i, cpu; | |
724 | bool second_pass = false; | |
42202dd5 | 725 | |
60666c63 | 726 | if (forks) { |
664c98d4 | 727 | if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, |
735f7e0b | 728 | workload_exec_failed_signal) < 0) { |
acf28922 NK |
729 | perror("failed to prepare workload"); |
730 | return -1; | |
60666c63 | 731 | } |
d20a47e7 | 732 | child_pid = evsel_list->workload.pid; |
051ae7f7 PM |
733 | } |
734 | ||
6a4bb04c | 735 | if (group) |
63dab225 | 736 | perf_evlist__set_leader(evsel_list); |
6a4bb04c | 737 | |
4804e011 AK |
738 | if (affinity__setup(&affinity) < 0) |
739 | return -1; | |
740 | ||
741 | evlist__for_each_cpu (evsel_list, i, cpu) { | |
742 | affinity__set(&affinity, cpu); | |
743 | ||
744 | evlist__for_each_entry(evsel_list, counter) { | |
745 | if (evsel__cpu_iter_skip(counter, cpu)) | |
746 | continue; | |
747 | if (counter->reset_group || counter->errored) | |
748 | continue; | |
42ef8a78 | 749 | try_again: |
4804e011 AK |
750 | if (create_perf_stat_counter(counter, &stat_config, &target, |
751 | counter->cpu_iter - 1) < 0) { | |
752 | ||
753 | /* | |
754 | * Weak group failed. We cannot just undo this here | |
755 | * because earlier CPUs might be in group mode, and the kernel | |
756 | * doesn't support mixing group and non group reads. Defer | |
757 | * it to later. | |
758 | * Don't close here because we're in the wrong affinity. | |
759 | */ | |
760 | if ((errno == EINVAL || errno == EBADF) && | |
761 | counter->leader != counter && | |
762 | counter->weak_group) { | |
763 | perf_evlist__reset_weak_group(evsel_list, counter, false); | |
764 | assert(counter->reset_group); | |
765 | second_pass = true; | |
766 | continue; | |
767 | } | |
768 | ||
769 | switch (stat_handle_error(counter)) { | |
770 | case COUNTER_FATAL: | |
771 | return -1; | |
772 | case COUNTER_RETRY: | |
773 | goto try_again; | |
774 | case COUNTER_SKIP: | |
775 | continue; | |
776 | default: | |
777 | break; | |
778 | } | |
779 | ||
5a5dfe4b | 780 | } |
4804e011 AK |
781 | counter->supported = true; |
782 | } | |
783 | } | |
5a5dfe4b | 784 | |
4804e011 AK |
785 | if (second_pass) { |
786 | /* | |
787 | * Now redo all the weak group after closing them, | |
788 | * and also close errored counters. | |
789 | */ | |
790 | ||
791 | evlist__for_each_cpu(evsel_list, i, cpu) { | |
792 | affinity__set(&affinity, cpu); | |
793 | /* First close errored or weak retry */ | |
794 | evlist__for_each_entry(evsel_list, counter) { | |
795 | if (!counter->reset_group && !counter->errored) | |
796 | continue; | |
797 | if (evsel__cpu_iter_skip_no_inc(counter, cpu)) | |
798 | continue; | |
799 | perf_evsel__close_cpu(&counter->core, counter->cpu_iter); | |
ab6c79b8 | 800 | } |
4804e011 AK |
801 | /* Now reopen weak */ |
802 | evlist__for_each_entry(evsel_list, counter) { | |
803 | if (!counter->reset_group && !counter->errored) | |
804 | continue; | |
805 | if (evsel__cpu_iter_skip(counter, cpu)) | |
806 | continue; | |
807 | if (!counter->reset_group) | |
808 | continue; | |
809 | try_again_reset: | |
8ab2e96d | 810 | pr_debug2("reopening weak %s\n", evsel__name(counter)); |
4804e011 AK |
811 | if (create_perf_stat_counter(counter, &stat_config, &target, |
812 | counter->cpu_iter - 1) < 0) { | |
813 | ||
814 | switch (stat_handle_error(counter)) { | |
815 | case COUNTER_FATAL: | |
816 | return -1; | |
817 | case COUNTER_RETRY: | |
818 | goto try_again_reset; | |
819 | case COUNTER_SKIP: | |
820 | continue; | |
821 | default: | |
822 | break; | |
823 | } | |
824 | } | |
825 | counter->supported = true; | |
826 | } | |
827 | } | |
828 | } | |
829 | affinity__cleanup(&affinity); | |
830 | ||
831 | evlist__for_each_entry(evsel_list, counter) { | |
832 | if (!counter->supported) { | |
833 | perf_evsel__free_fd(&counter->core); | |
834 | continue; | |
48290609 | 835 | } |
410136f5 SE |
836 | |
837 | l = strlen(counter->unit); | |
df4f7b4d JO |
838 | if (l > stat_config.unit_width) |
839 | stat_config.unit_width = l; | |
2af4646d | 840 | |
ddc6999e | 841 | if (evsel__should_store_id(counter) && |
34397753 | 842 | evsel__store_ids(counter, evsel_list)) |
2af4646d | 843 | return -1; |
084ab9f8 | 844 | } |
42202dd5 | 845 | |
23d4aad4 | 846 | if (perf_evlist__apply_filters(evsel_list, &counter)) { |
62d94b00 | 847 | pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", |
8ab2e96d | 848 | counter->filter, evsel__name(counter), errno, |
c8b5f2c9 | 849 | str_error_r(errno, msg, sizeof(msg))); |
cfd748ae FW |
850 | return -1; |
851 | } | |
852 | ||
4979d0c7 | 853 | if (STAT_RECORD) { |
8ceb41d7 | 854 | int err, fd = perf_data__fd(&perf_stat.data); |
4979d0c7 | 855 | |
664c98d4 | 856 | if (is_pipe) { |
8ceb41d7 | 857 | err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); |
664c98d4 JO |
858 | } else { |
859 | err = perf_session__write_header(perf_stat.session, evsel_list, | |
860 | fd, false); | |
861 | } | |
862 | ||
4979d0c7 JO |
863 | if (err < 0) |
864 | return err; | |
8b99b1a4 | 865 | |
b251892d ACM |
866 | err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list, |
867 | process_synthesized_event, is_pipe); | |
8b99b1a4 JO |
868 | if (err < 0) |
869 | return err; | |
4979d0c7 JO |
870 | } |
871 | ||
42202dd5 IM |
872 | /* |
873 | * Enable counters and exec the command: | |
874 | */ | |
875 | t0 = rdclock(); | |
13370a9b | 876 | clock_gettime(CLOCK_MONOTONIC, &ref_time); |
42202dd5 | 877 | |
60666c63 | 878 | if (forks) { |
acf28922 | 879 | perf_evlist__start_workload(evsel_list); |
67ccdecd | 880 | enable_counters(); |
acf28922 | 881 | |
27e9769a | 882 | if (interval || timeout || evlist__ctlfd_initialized(evsel_list)) |
bee328cb | 883 | status = dispatch_events(forks, timeout, interval, ×); |
cfbd41b7 ACM |
884 | if (child_pid != -1) { |
885 | if (timeout) | |
886 | kill(child_pid, SIGTERM); | |
8a99255a | 887 | wait4(child_pid, &status, 0, &stat_config.ru_data); |
cfbd41b7 | 888 | } |
6af206fd | 889 | |
f33cbe72 | 890 | if (workload_exec_errno) { |
c8b5f2c9 | 891 | const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); |
f33cbe72 | 892 | pr_err("Workload failed: %s\n", emsg); |
6af206fd | 893 | return -1; |
f33cbe72 | 894 | } |
6af206fd | 895 | |
33e49ea7 AK |
896 | if (WIFSIGNALED(status)) |
897 | psignal(WTERMSIG(status), argv[0]); | |
60666c63 | 898 | } else { |
67ccdecd | 899 | enable_counters(); |
bee328cb | 900 | status = dispatch_events(forks, timeout, interval, ×); |
60666c63 | 901 | } |
42202dd5 | 902 | |
3df33eff MR |
903 | disable_counters(); |
904 | ||
42202dd5 IM |
905 | t1 = rdclock(); |
906 | ||
54ac0b1b JO |
907 | if (stat_config.walltime_run_table) |
908 | stat_config.walltime_run[run_idx] = t1 - t0; | |
e55c14af | 909 | |
ee6a9614 | 910 | if (interval && stat_config.summary) { |
c7e5b328 | 911 | stat_config.interval = 0; |
ee6a9614 | 912 | stat_config.stop_read_counter = true; |
c7e5b328 JY |
913 | init_stats(&walltime_nsecs_stats); |
914 | update_stats(&walltime_nsecs_stats, t1 - t0); | |
915 | ||
916 | if (stat_config.aggr_mode == AGGR_GLOBAL) | |
917 | perf_evlist__save_aggr_prev_raw_counts(evsel_list); | |
918 | ||
919 | perf_evlist__copy_prev_raw_counts(evsel_list); | |
920 | perf_evlist__reset_prev_raw_counts(evsel_list); | |
921 | runtime_stat_reset(&stat_config); | |
922 | perf_stat__reset_shadow_per_stat(&rt_stat); | |
923 | } else | |
924 | update_stats(&walltime_nsecs_stats, t1 - t0); | |
42202dd5 | 925 | |
3df33eff MR |
926 | /* |
927 | * Closing a group leader splits the group, and as we only disable | |
928 | * group leaders, results in remaining events becoming enabled. To | |
929 | * avoid arbitrary skew, we must read all counters before closing any | |
930 | * group leaders. | |
931 | */ | |
f0fbb114 | 932 | read_counters(&(struct timespec) { .tv_nsec = t1-t0 }); |
08ef3af1 JO |
933 | |
934 | /* | |
935 | * We need to keep evsel_list alive, because it's processed | |
936 | * later the evsel_list will be closed after. | |
937 | */ | |
938 | if (!STAT_RECORD) | |
750b4ede | 939 | evlist__close(evsel_list); |
c52b12ed | 940 | |
42202dd5 IM |
941 | return WEXITSTATUS(status); |
942 | } | |
943 | ||
e55c14af | 944 | static int run_perf_stat(int argc, const char **argv, int run_idx) |
1f16c575 PZ |
945 | { |
946 | int ret; | |
947 | ||
948 | if (pre_cmd) { | |
949 | ret = system(pre_cmd); | |
950 | if (ret) | |
951 | return ret; | |
952 | } | |
953 | ||
954 | if (sync_run) | |
955 | sync(); | |
956 | ||
e55c14af | 957 | ret = __run_perf_stat(argc, argv, run_idx); |
1f16c575 PZ |
958 | if (ret) |
959 | return ret; | |
960 | ||
961 | if (post_cmd) { | |
962 | ret = system(post_cmd); | |
963 | if (ret) | |
964 | return ret; | |
965 | } | |
966 | ||
967 | return ret; | |
968 | } | |
969 | ||
a5a9eac1 JO |
970 | static void print_counters(struct timespec *ts, int argc, const char **argv) |
971 | { | |
0174820a JO |
972 | /* Do not print anything if we record to the pipe. */ |
973 | if (STAT_RECORD && perf_stat.data.is_pipe) | |
974 | return; | |
975 | ||
c512e0ea | 976 | perf_evlist__print_counters(evsel_list, &stat_config, &target, |
b64df7f3 | 977 | ts, argc, argv); |
a5a9eac1 JO |
978 | } |
979 | ||
f7b7c26e PZ |
980 | static volatile int signr = -1; |
981 | ||
5242519b | 982 | static void skip_signal(int signo) |
ddcacfa0 | 983 | { |
ec0d3d1f | 984 | if ((child_pid == -1) || stat_config.interval) |
60666c63 LW |
985 | done = 1; |
986 | ||
f7b7c26e | 987 | signr = signo; |
d07f0b12 SE |
988 | /* |
989 | * render child_pid harmless | |
990 | * won't send SIGTERM to a random | |
991 | * process in case of race condition | |
992 | * and fast PID recycling | |
993 | */ | |
994 | child_pid = -1; | |
f7b7c26e PZ |
995 | } |
996 | ||
997 | static void sig_atexit(void) | |
998 | { | |
d07f0b12 SE |
999 | sigset_t set, oset; |
1000 | ||
1001 | /* | |
1002 | * avoid race condition with SIGCHLD handler | |
1003 | * in skip_signal() which is modifying child_pid | |
1004 | * goal is to avoid send SIGTERM to a random | |
1005 | * process | |
1006 | */ | |
1007 | sigemptyset(&set); | |
1008 | sigaddset(&set, SIGCHLD); | |
1009 | sigprocmask(SIG_BLOCK, &set, &oset); | |
1010 | ||
933da83a CW |
1011 | if (child_pid != -1) |
1012 | kill(child_pid, SIGTERM); | |
1013 | ||
d07f0b12 SE |
1014 | sigprocmask(SIG_SETMASK, &oset, NULL); |
1015 | ||
f7b7c26e PZ |
1016 | if (signr == -1) |
1017 | return; | |
1018 | ||
1019 | signal(signr, SIG_DFL); | |
1020 | kill(getpid(), signr); | |
5242519b IM |
1021 | } |
1022 | ||
d778a778 PC |
1023 | void perf_stat__set_big_num(int set) |
1024 | { | |
1025 | stat_config.big_num = (set != 0); | |
1026 | } | |
1027 | ||
1d037ca1 IT |
1028 | static int stat__set_big_num(const struct option *opt __maybe_unused, |
1029 | const char *s __maybe_unused, int unset) | |
d7470b6a SE |
1030 | { |
1031 | big_num_opt = unset ? 0 : 1; | |
d778a778 | 1032 | perf_stat__set_big_num(!unset); |
d7470b6a SE |
1033 | return 0; |
1034 | } | |
1035 | ||
44b1e60a AK |
1036 | static int enable_metric_only(const struct option *opt __maybe_unused, |
1037 | const char *s __maybe_unused, int unset) | |
1038 | { | |
1039 | force_metric_only = true; | |
0ce5aa02 | 1040 | stat_config.metric_only = !unset; |
44b1e60a AK |
1041 | return 0; |
1042 | } | |
1043 | ||
b18f3e36 AK |
1044 | static int parse_metric_groups(const struct option *opt, |
1045 | const char *str, | |
1046 | int unset __maybe_unused) | |
1047 | { | |
05530a79 IR |
1048 | return metricgroup__parse_groups(opt, str, |
1049 | stat_config.metric_no_group, | |
1050 | stat_config.metric_no_merge, | |
1051 | &stat_config.metric_events); | |
b18f3e36 AK |
1052 | } |
1053 | ||
27e9769a AB |
1054 | static int parse_control_option(const struct option *opt, |
1055 | const char *str, | |
1056 | int unset __maybe_unused) | |
1057 | { | |
9864a66d | 1058 | struct perf_stat_config *config = opt->value; |
27e9769a | 1059 | |
a8fcbd26 AH |
1060 | return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close); |
1061 | } | |
1062 | ||
d1c5a0e8 NK |
1063 | static int parse_stat_cgroups(const struct option *opt, |
1064 | const char *str, int unset) | |
1065 | { | |
1066 | if (stat_config.cgroup_list) { | |
1067 | pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); | |
1068 | return -1; | |
1069 | } | |
1070 | ||
1071 | return parse_cgroups(opt, str, unset); | |
1072 | } | |
1073 | ||
51433ead | 1074 | static struct option stat_options[] = { |
e0547311 JO |
1075 | OPT_BOOLEAN('T', "transaction", &transaction_run, |
1076 | "hardware transaction statistics"), | |
1077 | OPT_CALLBACK('e', "event", &evsel_list, "event", | |
1078 | "event selector. use 'perf list' to list available events", | |
1079 | parse_events_option), | |
1080 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", | |
1081 | "event filter", parse_filter), | |
5698f26b | 1082 | OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, |
e0547311 JO |
1083 | "child tasks do not inherit counters"), |
1084 | OPT_STRING('p', "pid", &target.pid, "pid", | |
1085 | "stat events on existing process id"), | |
1086 | OPT_STRING('t', "tid", &target.tid, "tid", | |
1087 | "stat events on existing thread id"), | |
1088 | OPT_BOOLEAN('a', "all-cpus", &target.system_wide, | |
1089 | "system-wide collection from all CPUs"), | |
1090 | OPT_BOOLEAN('g', "group", &group, | |
1091 | "put the counters into a counter group"), | |
75998bb2 AK |
1092 | OPT_BOOLEAN(0, "scale", &stat_config.scale, |
1093 | "Use --no-scale to disable counter scaling for multiplexing"), | |
e0547311 JO |
1094 | OPT_INCR('v', "verbose", &verbose, |
1095 | "be more verbose (show counter open errors, etc)"), | |
d97ae04b | 1096 | OPT_INTEGER('r', "repeat", &stat_config.run_count, |
e0547311 | 1097 | "repeat command and print average + stddev (max: 100, forever: 0)"), |
54ac0b1b | 1098 | OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, |
e55c14af | 1099 | "display details about each run (only with -r option)"), |
aea0dca1 | 1100 | OPT_BOOLEAN('n', "null", &stat_config.null_run, |
e0547311 JO |
1101 | "null run - dont start any counters"), |
1102 | OPT_INCR('d', "detailed", &detailed_run, | |
1103 | "detailed run - start a lot of events"), | |
1104 | OPT_BOOLEAN('S', "sync", &sync_run, | |
1105 | "call sync() before starting a run"), | |
1106 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, | |
1107 | "print large numbers with thousands\' separators", | |
1108 | stat__set_big_num), | |
1109 | OPT_STRING('C', "cpu", &target.cpu_list, "cpu", | |
1110 | "list of cpus to monitor in system-wide"), | |
1111 | OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, | |
1112 | "disable CPU count aggregation", AGGR_NONE), | |
fdee335b | 1113 | OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"), |
fa7070a3 | 1114 | OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", |
e0547311 JO |
1115 | "print counts with custom separator"), |
1116 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", | |
d1c5a0e8 NK |
1117 | "monitor event in cgroup name only", parse_stat_cgroups), |
1118 | OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name", | |
1119 | "expand events for each cgroup"), | |
e0547311 JO |
1120 | OPT_STRING('o', "output", &output_name, "file", "output file name"), |
1121 | OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), | |
1122 | OPT_INTEGER(0, "log-fd", &output_fd, | |
1123 | "log output to fd, instead of stderr"), | |
1124 | OPT_STRING(0, "pre", &pre_cmd, "command", | |
1125 | "command to run prior to the measured command"), | |
1126 | OPT_STRING(0, "post", &post_cmd, "command", | |
1127 | "command to run after to the measured command"), | |
1128 | OPT_UINTEGER('I', "interval-print", &stat_config.interval, | |
9dc9a95f AB |
1129 | "print counts at regular interval in ms " |
1130 | "(overhead is possible for values <= 100ms)"), | |
db06a269 | 1131 | OPT_INTEGER(0, "interval-count", &stat_config.times, |
1132 | "print counts for fixed number of times"), | |
132c6ba3 | 1133 | OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, |
9660e08e | 1134 | "clear screen in between new interval"), |
f1f8ad52 | 1135 | OPT_UINTEGER(0, "timeout", &stat_config.timeout, |
1136 | "stop workload and print counts after a timeout period in ms (>= 10ms)"), | |
e0547311 JO |
1137 | OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, |
1138 | "aggregate counts per processor socket", AGGR_SOCKET), | |
db5742b6 KL |
1139 | OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode, |
1140 | "aggregate counts per processor die", AGGR_DIE), | |
e0547311 JO |
1141 | OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, |
1142 | "aggregate counts per physical processor core", AGGR_CORE), | |
1143 | OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, | |
1144 | "aggregate counts per thread", AGGR_THREAD), | |
86895b48 JO |
1145 | OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode, |
1146 | "aggregate counts per numa node", AGGR_NODE), | |
2162b9c6 AB |
1147 | OPT_INTEGER('D', "delay", &stat_config.initial_delay, |
1148 | "ms to wait before starting measurement after program start (-1: start with events disabled)"), | |
0ce5aa02 | 1149 | OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, |
44b1e60a | 1150 | "Only print computed metrics. No raw values", enable_metric_only), |
05530a79 IR |
1151 | OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group, |
1152 | "don't group metric events, impacts multiplexing"), | |
1153 | OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge, | |
1154 | "don't try to share events between metrics in a group"), | |
44b1e60a AK |
1155 | OPT_BOOLEAN(0, "topdown", &topdown_run, |
1156 | "measure topdown level 1 statistics"), | |
daefd0bc KL |
1157 | OPT_BOOLEAN(0, "smi-cost", &smi_cost, |
1158 | "measure SMI cost"), | |
b18f3e36 AK |
1159 | OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", |
1160 | "monitor specified metrics or metric groups (separated by ,)", | |
1161 | parse_metric_groups), | |
dd071024 JY |
1162 | OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel, |
1163 | "Configure all used events to run in kernel space.", | |
1164 | PARSE_OPT_EXCLUSIVE), | |
1165 | OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user, | |
1166 | "Configure all used events to run in user space.", | |
1167 | PARSE_OPT_EXCLUSIVE), | |
1af62ce6 JY |
1168 | OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread, |
1169 | "Use with 'percore' event qualifier to show the event " | |
1170 | "counts of one hardware thread by sum up total hardware " | |
1171 | "threads of same physical core"), | |
ee6a9614 JY |
1172 | OPT_BOOLEAN(0, "summary", &stat_config.summary, |
1173 | "print summary for interval mode"), | |
70943490 SE |
1174 | #ifdef HAVE_LIBPFM |
1175 | OPT_CALLBACK(0, "pfm-events", &evsel_list, "event", | |
1176 | "libpfm4 event selector. use 'perf list' to list available events", | |
1177 | parse_libpfm_events_option), | |
1178 | #endif | |
a8fcbd26 | 1179 | OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]", |
27e9769a | 1180 | "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n" |
a8fcbd26 AH |
1181 | "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n" |
1182 | "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.", | |
27e9769a | 1183 | parse_control_option), |
e0547311 JO |
1184 | OPT_END() |
1185 | }; | |
1186 | ||
6f6b6594 | 1187 | static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, |
f854839b | 1188 | struct perf_cpu_map *map, int cpu) |
1fe7a300 JO |
1189 | { |
1190 | return cpu_map__get_socket(map, cpu, NULL); | |
1191 | } | |
1192 | ||
db5742b6 | 1193 | static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused, |
f854839b | 1194 | struct perf_cpu_map *map, int cpu) |
db5742b6 KL |
1195 | { |
1196 | return cpu_map__get_die(map, cpu, NULL); | |
1197 | } | |
1198 | ||
6f6b6594 | 1199 | static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused, |
f854839b | 1200 | struct perf_cpu_map *map, int cpu) |
1fe7a300 JO |
1201 | { |
1202 | return cpu_map__get_core(map, cpu, NULL); | |
1203 | } | |
1204 | ||
86895b48 JO |
1205 | static int perf_stat__get_node(struct perf_stat_config *config __maybe_unused, |
1206 | struct perf_cpu_map *map, int cpu) | |
1207 | { | |
1208 | return cpu_map__get_node(map, cpu, NULL); | |
1209 | } | |
1210 | ||
6f6b6594 | 1211 | static int perf_stat__get_aggr(struct perf_stat_config *config, |
f854839b | 1212 | aggr_get_id_t get_id, struct perf_cpu_map *map, int idx) |
1e5a2931 JO |
1213 | { |
1214 | int cpu; | |
1215 | ||
1216 | if (idx >= map->nr) | |
1217 | return -1; | |
1218 | ||
1219 | cpu = map->map[idx]; | |
1220 | ||
6f6b6594 JO |
1221 | if (config->cpus_aggr_map->map[cpu] == -1) |
1222 | config->cpus_aggr_map->map[cpu] = get_id(config, map, idx); | |
1e5a2931 | 1223 | |
6f6b6594 | 1224 | return config->cpus_aggr_map->map[cpu]; |
1e5a2931 JO |
1225 | } |
1226 | ||
6f6b6594 | 1227 | static int perf_stat__get_socket_cached(struct perf_stat_config *config, |
f854839b | 1228 | struct perf_cpu_map *map, int idx) |
1e5a2931 | 1229 | { |
6f6b6594 | 1230 | return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx); |
1e5a2931 JO |
1231 | } |
1232 | ||
db5742b6 | 1233 | static int perf_stat__get_die_cached(struct perf_stat_config *config, |
f854839b | 1234 | struct perf_cpu_map *map, int idx) |
db5742b6 KL |
1235 | { |
1236 | return perf_stat__get_aggr(config, perf_stat__get_die, map, idx); | |
1237 | } | |
1238 | ||
6f6b6594 | 1239 | static int perf_stat__get_core_cached(struct perf_stat_config *config, |
f854839b | 1240 | struct perf_cpu_map *map, int idx) |
1e5a2931 | 1241 | { |
6f6b6594 | 1242 | return perf_stat__get_aggr(config, perf_stat__get_core, map, idx); |
1e5a2931 JO |
1243 | } |
1244 | ||
86895b48 JO |
1245 | static int perf_stat__get_node_cached(struct perf_stat_config *config, |
1246 | struct perf_cpu_map *map, int idx) | |
1247 | { | |
1248 | return perf_stat__get_aggr(config, perf_stat__get_node, map, idx); | |
1249 | } | |
1250 | ||
4fc4d8df JY |
1251 | static bool term_percore_set(void) |
1252 | { | |
32dcd021 | 1253 | struct evsel *counter; |
4fc4d8df JY |
1254 | |
1255 | evlist__for_each_entry(evsel_list, counter) { | |
1256 | if (counter->percore) | |
1257 | return true; | |
1258 | } | |
1259 | ||
1260 | return false; | |
1261 | } | |
1262 | ||
86ee6e18 SE |
1263 | static int perf_stat_init_aggr_mode(void) |
1264 | { | |
1e5a2931 JO |
1265 | int nr; |
1266 | ||
421a50f3 | 1267 | switch (stat_config.aggr_mode) { |
86ee6e18 | 1268 | case AGGR_SOCKET: |
f72f901d | 1269 | if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) { |
86ee6e18 SE |
1270 | perror("cannot build socket map"); |
1271 | return -1; | |
1272 | } | |
6f6b6594 | 1273 | stat_config.aggr_get_id = perf_stat__get_socket_cached; |
86ee6e18 | 1274 | break; |
db5742b6 | 1275 | case AGGR_DIE: |
f72f901d | 1276 | if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) { |
db5742b6 KL |
1277 | perror("cannot build die map"); |
1278 | return -1; | |
1279 | } | |
1280 | stat_config.aggr_get_id = perf_stat__get_die_cached; | |
1281 | break; | |
12c08a9f | 1282 | case AGGR_CORE: |
f72f901d | 1283 | if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) { |
12c08a9f SE |
1284 | perror("cannot build core map"); |
1285 | return -1; | |
1286 | } | |
6f6b6594 | 1287 | stat_config.aggr_get_id = perf_stat__get_core_cached; |
12c08a9f | 1288 | break; |
86895b48 JO |
1289 | case AGGR_NODE: |
1290 | if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) { | |
1291 | perror("cannot build core map"); | |
1292 | return -1; | |
1293 | } | |
1294 | stat_config.aggr_get_id = perf_stat__get_node_cached; | |
1295 | break; | |
86ee6e18 | 1296 | case AGGR_NONE: |
4fc4d8df | 1297 | if (term_percore_set()) { |
f72f901d | 1298 | if (cpu_map__build_core_map(evsel_list->core.cpus, |
4fc4d8df JY |
1299 | &stat_config.aggr_map)) { |
1300 | perror("cannot build core map"); | |
1301 | return -1; | |
1302 | } | |
1303 | stat_config.aggr_get_id = perf_stat__get_core_cached; | |
1304 | } | |
1305 | break; | |
86ee6e18 | 1306 | case AGGR_GLOBAL: |
32b8af82 | 1307 | case AGGR_THREAD: |
208df99e | 1308 | case AGGR_UNSET: |
86ee6e18 SE |
1309 | default: |
1310 | break; | |
1311 | } | |
1e5a2931 JO |
1312 | |
1313 | /* | |
1314 | * The evsel_list->cpus is the base we operate on, | |
1315 | * taking the highest cpu number to be the size of | |
1316 | * the aggregation translate cpumap. | |
1317 | */ | |
4256d434 | 1318 | nr = perf_cpu_map__max(evsel_list->core.cpus); |
315c0a1f | 1319 | stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1); |
6f6b6594 | 1320 | return stat_config.cpus_aggr_map ? 0 : -ENOMEM; |
86ee6e18 SE |
1321 | } |
1322 | ||
544c2ae7 MH |
1323 | static void perf_stat__exit_aggr_mode(void) |
1324 | { | |
38f01d8d JO |
1325 | perf_cpu_map__put(stat_config.aggr_map); |
1326 | perf_cpu_map__put(stat_config.cpus_aggr_map); | |
6f6b6594 JO |
1327 | stat_config.aggr_map = NULL; |
1328 | stat_config.cpus_aggr_map = NULL; | |
544c2ae7 MH |
1329 | } |
1330 | ||
f854839b | 1331 | static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx) |
68d702f7 JO |
1332 | { |
1333 | int cpu; | |
1334 | ||
1335 | if (idx > map->nr) | |
1336 | return -1; | |
1337 | ||
1338 | cpu = map->map[idx]; | |
1339 | ||
da8a58b5 | 1340 | if (cpu >= env->nr_cpus_avail) |
68d702f7 JO |
1341 | return -1; |
1342 | ||
1343 | return cpu; | |
1344 | } | |
1345 | ||
f854839b | 1346 | static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data) |
68d702f7 JO |
1347 | { |
1348 | struct perf_env *env = data; | |
1349 | int cpu = perf_env__get_cpu(env, map, idx); | |
1350 | ||
1351 | return cpu == -1 ? -1 : env->cpu[cpu].socket_id; | |
1352 | } | |
1353 | ||
f854839b | 1354 | static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data) |
db5742b6 KL |
1355 | { |
1356 | struct perf_env *env = data; | |
1357 | int die_id = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1358 | ||
1359 | if (cpu != -1) { | |
1360 | /* | |
1361 | * Encode socket in bit range 15:8 | |
1362 | * die_id is relative to socket, | |
1363 | * we need a global id. So we combine | |
1364 | * socket + die id | |
1365 | */ | |
1366 | if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) | |
1367 | return -1; | |
1368 | ||
1369 | if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) | |
1370 | return -1; | |
1371 | ||
1372 | die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff); | |
1373 | } | |
1374 | ||
1375 | return die_id; | |
1376 | } | |
1377 | ||
f854839b | 1378 | static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data) |
68d702f7 JO |
1379 | { |
1380 | struct perf_env *env = data; | |
1381 | int core = -1, cpu = perf_env__get_cpu(env, map, idx); | |
1382 | ||
1383 | if (cpu != -1) { | |
68d702f7 | 1384 | /* |
db5742b6 KL |
1385 | * Encode socket in bit range 31:24 |
1386 | * encode die id in bit range 23:16 | |
1387 | * core_id is relative to socket and die, | |
68d702f7 | 1388 | * we need a global id. So we combine |
db5742b6 | 1389 | * socket + die id + core id |
68d702f7 | 1390 | */ |
db5742b6 KL |
1391 | if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) |
1392 | return -1; | |
1393 | ||
1394 | if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) | |
1395 | return -1; | |
1396 | ||
1397 | if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n")) | |
1398 | return -1; | |
1399 | ||
1400 | core = (env->cpu[cpu].socket_id << 24) | | |
1401 | (env->cpu[cpu].die_id << 16) | | |
1402 | (env->cpu[cpu].core_id & 0xffff); | |
68d702f7 JO |
1403 | } |
1404 | ||
1405 | return core; | |
1406 | } | |
1407 | ||
86895b48 JO |
1408 | static int perf_env__get_node(struct perf_cpu_map *map, int idx, void *data) |
1409 | { | |
1410 | int cpu = perf_env__get_cpu(data, map, idx); | |
1411 | ||
1412 | return perf_env__numa_node(data, cpu); | |
1413 | } | |
1414 | ||
f854839b JO |
1415 | static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus, |
1416 | struct perf_cpu_map **sockp) | |
68d702f7 JO |
1417 | { |
1418 | return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); | |
1419 | } | |
1420 | ||
f854839b JO |
1421 | static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus, |
1422 | struct perf_cpu_map **diep) | |
db5742b6 KL |
1423 | { |
1424 | return cpu_map__build_map(cpus, diep, perf_env__get_die, env); | |
1425 | } | |
1426 | ||
f854839b JO |
1427 | static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus, |
1428 | struct perf_cpu_map **corep) | |
68d702f7 JO |
1429 | { |
1430 | return cpu_map__build_map(cpus, corep, perf_env__get_core, env); | |
1431 | } | |
1432 | ||
86895b48 JO |
1433 | static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus, |
1434 | struct perf_cpu_map **nodep) | |
1435 | { | |
1436 | return cpu_map__build_map(cpus, nodep, perf_env__get_node, env); | |
1437 | } | |
1438 | ||
6f6b6594 | 1439 | static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, |
f854839b | 1440 | struct perf_cpu_map *map, int idx) |
68d702f7 JO |
1441 | { |
1442 | return perf_env__get_socket(map, idx, &perf_stat.session->header.env); | |
1443 | } | |
db5742b6 | 1444 | static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, |
f854839b | 1445 | struct perf_cpu_map *map, int idx) |
db5742b6 KL |
1446 | { |
1447 | return perf_env__get_die(map, idx, &perf_stat.session->header.env); | |
1448 | } | |
68d702f7 | 1449 | |
6f6b6594 | 1450 | static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, |
f854839b | 1451 | struct perf_cpu_map *map, int idx) |
68d702f7 JO |
1452 | { |
1453 | return perf_env__get_core(map, idx, &perf_stat.session->header.env); | |
1454 | } | |
1455 | ||
86895b48 JO |
1456 | static int perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused, |
1457 | struct perf_cpu_map *map, int idx) | |
1458 | { | |
1459 | return perf_env__get_node(map, idx, &perf_stat.session->header.env); | |
1460 | } | |
1461 | ||
68d702f7 JO |
1462 | static int perf_stat_init_aggr_mode_file(struct perf_stat *st) |
1463 | { | |
1464 | struct perf_env *env = &st->session->header.env; | |
1465 | ||
1466 | switch (stat_config.aggr_mode) { | |
1467 | case AGGR_SOCKET: | |
f72f901d | 1468 | if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { |
68d702f7 JO |
1469 | perror("cannot build socket map"); |
1470 | return -1; | |
1471 | } | |
6f6b6594 | 1472 | stat_config.aggr_get_id = perf_stat__get_socket_file; |
68d702f7 | 1473 | break; |
db5742b6 | 1474 | case AGGR_DIE: |
f72f901d | 1475 | if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { |
db5742b6 KL |
1476 | perror("cannot build die map"); |
1477 | return -1; | |
1478 | } | |
1479 | stat_config.aggr_get_id = perf_stat__get_die_file; | |
1480 | break; | |
68d702f7 | 1481 | case AGGR_CORE: |
f72f901d | 1482 | if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { |
68d702f7 JO |
1483 | perror("cannot build core map"); |
1484 | return -1; | |
1485 | } | |
6f6b6594 | 1486 | stat_config.aggr_get_id = perf_stat__get_core_file; |
68d702f7 | 1487 | break; |
86895b48 JO |
1488 | case AGGR_NODE: |
1489 | if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { | |
1490 | perror("cannot build core map"); | |
1491 | return -1; | |
1492 | } | |
1493 | stat_config.aggr_get_id = perf_stat__get_node_file; | |
1494 | break; | |
68d702f7 JO |
1495 | case AGGR_NONE: |
1496 | case AGGR_GLOBAL: | |
1497 | case AGGR_THREAD: | |
1498 | case AGGR_UNSET: | |
1499 | default: | |
1500 | break; | |
1501 | } | |
1502 | ||
1503 | return 0; | |
1504 | } | |
1505 | ||
2cba3ffb IM |
1506 | /* |
1507 | * Add default attributes, if there were no attributes specified or | |
1508 | * if -d/--detailed, -d -d or -d -d -d is used: | |
1509 | */ | |
1510 | static int add_default_attributes(void) | |
1511 | { | |
44b1e60a | 1512 | int err; |
9dec4473 | 1513 | struct perf_event_attr default_attrs0[] = { |
b070a547 ACM |
1514 | |
1515 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, | |
1516 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, | |
1517 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, | |
1518 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, | |
1519 | ||
1520 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, | |
9dec4473 AK |
1521 | }; |
1522 | struct perf_event_attr frontend_attrs[] = { | |
b070a547 | 1523 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, |
9dec4473 AK |
1524 | }; |
1525 | struct perf_event_attr backend_attrs[] = { | |
b070a547 | 1526 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, |
9dec4473 AK |
1527 | }; |
1528 | struct perf_event_attr default_attrs1[] = { | |
b070a547 ACM |
1529 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
1530 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, | |
1531 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, | |
1532 | ||
1533 | }; | |
1534 | ||
1535 | /* | |
1536 | * Detailed stats (-d), covering the L1 and last level data caches: | |
1537 | */ | |
1538 | struct perf_event_attr detailed_attrs[] = { | |
1539 | ||
1540 | { .type = PERF_TYPE_HW_CACHE, | |
1541 | .config = | |
1542 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1543 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1544 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1545 | ||
1546 | { .type = PERF_TYPE_HW_CACHE, | |
1547 | .config = | |
1548 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1549 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1550 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1551 | ||
1552 | { .type = PERF_TYPE_HW_CACHE, | |
1553 | .config = | |
1554 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1555 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1556 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1557 | ||
1558 | { .type = PERF_TYPE_HW_CACHE, | |
1559 | .config = | |
1560 | PERF_COUNT_HW_CACHE_LL << 0 | | |
1561 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1562 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1563 | }; | |
1564 | ||
1565 | /* | |
1566 | * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: | |
1567 | */ | |
1568 | struct perf_event_attr very_detailed_attrs[] = { | |
1569 | ||
1570 | { .type = PERF_TYPE_HW_CACHE, | |
1571 | .config = | |
1572 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1573 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1574 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1575 | ||
1576 | { .type = PERF_TYPE_HW_CACHE, | |
1577 | .config = | |
1578 | PERF_COUNT_HW_CACHE_L1I << 0 | | |
1579 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1580 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1581 | ||
1582 | { .type = PERF_TYPE_HW_CACHE, | |
1583 | .config = | |
1584 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1585 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1586 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1587 | ||
1588 | { .type = PERF_TYPE_HW_CACHE, | |
1589 | .config = | |
1590 | PERF_COUNT_HW_CACHE_DTLB << 0 | | |
1591 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1592 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1593 | ||
1594 | { .type = PERF_TYPE_HW_CACHE, | |
1595 | .config = | |
1596 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1597 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1598 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1599 | ||
1600 | { .type = PERF_TYPE_HW_CACHE, | |
1601 | .config = | |
1602 | PERF_COUNT_HW_CACHE_ITLB << 0 | | |
1603 | (PERF_COUNT_HW_CACHE_OP_READ << 8) | | |
1604 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1605 | ||
1606 | }; | |
1607 | ||
1608 | /* | |
1609 | * Very, very detailed stats (-d -d -d), adding prefetch events: | |
1610 | */ | |
1611 | struct perf_event_attr very_very_detailed_attrs[] = { | |
1612 | ||
1613 | { .type = PERF_TYPE_HW_CACHE, | |
1614 | .config = | |
1615 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1616 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1617 | (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, | |
1618 | ||
1619 | { .type = PERF_TYPE_HW_CACHE, | |
1620 | .config = | |
1621 | PERF_COUNT_HW_CACHE_L1D << 0 | | |
1622 | (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | | |
1623 | (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, | |
1624 | }; | |
a5cfa621 | 1625 | struct parse_events_error errinfo; |
b070a547 | 1626 | |
2cba3ffb | 1627 | /* Set attrs if no event is selected and !null_run: */ |
aea0dca1 | 1628 | if (stat_config.null_run) |
2cba3ffb IM |
1629 | return 0; |
1630 | ||
a910e466 | 1631 | bzero(&errinfo, sizeof(errinfo)); |
4cabc3d1 | 1632 | if (transaction_run) { |
742d92ff TR |
1633 | /* Handle -T as -M transaction. Once platform specific metrics |
1634 | * support has been added to the json files, all archictures | |
1635 | * will use this approach. To determine transaction support | |
1636 | * on an architecture test for such a metric name. | |
1637 | */ | |
1638 | if (metricgroup__has_metric("transaction")) { | |
1639 | struct option opt = { .value = &evsel_list }; | |
1640 | ||
1641 | return metricgroup__parse_groups(&opt, "transaction", | |
05530a79 IR |
1642 | stat_config.metric_no_group, |
1643 | stat_config.metric_no_merge, | |
d0192fdb | 1644 | &stat_config.metric_events); |
742d92ff TR |
1645 | } |
1646 | ||
4cabc3d1 AK |
1647 | if (pmu_have_event("cpu", "cycles-ct") && |
1648 | pmu_have_event("cpu", "el-start")) | |
fca32340 TR |
1649 | err = parse_events(evsel_list, transaction_attrs, |
1650 | &errinfo); | |
4cabc3d1 | 1651 | else |
fca32340 TR |
1652 | err = parse_events(evsel_list, |
1653 | transaction_limited_attrs, | |
1654 | &errinfo); | |
a454742c | 1655 | if (err) { |
4cabc3d1 | 1656 | fprintf(stderr, "Cannot set up transaction events\n"); |
a5cfa621 | 1657 | parse_events_print_error(&errinfo, transaction_attrs); |
4cabc3d1 AK |
1658 | return -1; |
1659 | } | |
1660 | return 0; | |
1661 | } | |
1662 | ||
daefd0bc KL |
1663 | if (smi_cost) { |
1664 | int smi; | |
1665 | ||
1666 | if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { | |
1667 | fprintf(stderr, "freeze_on_smi is not supported.\n"); | |
1668 | return -1; | |
1669 | } | |
1670 | ||
1671 | if (!smi) { | |
1672 | if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { | |
1673 | fprintf(stderr, "Failed to set freeze_on_smi.\n"); | |
1674 | return -1; | |
1675 | } | |
1676 | smi_reset = true; | |
1677 | } | |
1678 | ||
1679 | if (pmu_have_event("msr", "aperf") && | |
1680 | pmu_have_event("msr", "smi")) { | |
1681 | if (!force_metric_only) | |
0ce5aa02 | 1682 | stat_config.metric_only = true; |
a5cfa621 | 1683 | err = parse_events(evsel_list, smi_cost_attrs, &errinfo); |
daefd0bc KL |
1684 | } else { |
1685 | fprintf(stderr, "To measure SMI cost, it needs " | |
1686 | "msr/aperf/, msr/smi/ and cpu/cycles/ support\n"); | |
a5cfa621 | 1687 | parse_events_print_error(&errinfo, smi_cost_attrs); |
daefd0bc KL |
1688 | return -1; |
1689 | } | |
1690 | if (err) { | |
a910e466 | 1691 | parse_events_print_error(&errinfo, smi_cost_attrs); |
daefd0bc KL |
1692 | fprintf(stderr, "Cannot set up SMI cost events\n"); |
1693 | return -1; | |
1694 | } | |
1695 | return 0; | |
1696 | } | |
1697 | ||
44b1e60a AK |
1698 | if (topdown_run) { |
1699 | char *str = NULL; | |
1700 | bool warn = false; | |
1701 | ||
55c36a9f AK |
1702 | if (!force_metric_only) |
1703 | stat_config.metric_only = true; | |
1704 | ||
1705 | if (topdown_filter_events(topdown_metric_attrs, &str, 1) < 0) { | |
1706 | pr_err("Out of memory\n"); | |
1707 | return -1; | |
1708 | } | |
1709 | if (topdown_metric_attrs[0] && str) { | |
1710 | if (!stat_config.interval && !stat_config.metric_only) { | |
1711 | fprintf(stat_config.output, | |
1712 | "Topdown accuracy may decrease when measuring long periods.\n" | |
1713 | "Please print the result regularly, e.g. -I1000\n"); | |
1714 | } | |
1715 | goto setup_metrics; | |
1716 | } | |
1717 | ||
1718 | zfree(&str); | |
1719 | ||
44b1e60a AK |
1720 | if (stat_config.aggr_mode != AGGR_GLOBAL && |
1721 | stat_config.aggr_mode != AGGR_CORE) { | |
1722 | pr_err("top down event configuration requires --per-core mode\n"); | |
1723 | return -1; | |
1724 | } | |
1725 | stat_config.aggr_mode = AGGR_CORE; | |
1726 | if (nr_cgroups || !target__has_cpu(&target)) { | |
1727 | pr_err("top down event configuration requires system-wide mode (-a)\n"); | |
1728 | return -1; | |
1729 | } | |
1730 | ||
44b1e60a AK |
1731 | if (topdown_filter_events(topdown_attrs, &str, |
1732 | arch_topdown_check_group(&warn)) < 0) { | |
1733 | pr_err("Out of memory\n"); | |
1734 | return -1; | |
1735 | } | |
1736 | if (topdown_attrs[0] && str) { | |
1737 | if (warn) | |
1738 | arch_topdown_group_warn(); | |
55c36a9f | 1739 | setup_metrics: |
a5cfa621 | 1740 | err = parse_events(evsel_list, str, &errinfo); |
44b1e60a AK |
1741 | if (err) { |
1742 | fprintf(stderr, | |
1743 | "Cannot set up top down events %s: %d\n", | |
1744 | str, err); | |
a5cfa621 | 1745 | parse_events_print_error(&errinfo, str); |
c74b0503 | 1746 | free(str); |
44b1e60a AK |
1747 | return -1; |
1748 | } | |
1749 | } else { | |
1750 | fprintf(stderr, "System does not support topdown\n"); | |
1751 | return -1; | |
1752 | } | |
1753 | free(str); | |
1754 | } | |
1755 | ||
6484d2f9 | 1756 | if (!evsel_list->core.nr_entries) { |
a1f3d567 NK |
1757 | if (target__has_cpu(&target)) |
1758 | default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; | |
1759 | ||
e251abee | 1760 | if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0) |
9dec4473 AK |
1761 | return -1; |
1762 | if (pmu_have_event("cpu", "stalled-cycles-frontend")) { | |
e251abee | 1763 | if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0) |
9dec4473 AK |
1764 | return -1; |
1765 | } | |
1766 | if (pmu_have_event("cpu", "stalled-cycles-backend")) { | |
e251abee | 1767 | if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0) |
9dec4473 AK |
1768 | return -1; |
1769 | } | |
e251abee | 1770 | if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0) |
50d08e47 | 1771 | return -1; |
2cba3ffb IM |
1772 | } |
1773 | ||
1774 | /* Detailed events get appended to the event list: */ | |
1775 | ||
1776 | if (detailed_run < 1) | |
1777 | return 0; | |
1778 | ||
1779 | /* Append detailed run extra attributes: */ | |
e251abee | 1780 | if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) |
50d08e47 | 1781 | return -1; |
2cba3ffb IM |
1782 | |
1783 | if (detailed_run < 2) | |
1784 | return 0; | |
1785 | ||
1786 | /* Append very detailed run extra attributes: */ | |
e251abee | 1787 | if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) |
50d08e47 | 1788 | return -1; |
2cba3ffb IM |
1789 | |
1790 | if (detailed_run < 3) | |
1791 | return 0; | |
1792 | ||
1793 | /* Append very, very detailed run extra attributes: */ | |
e251abee | 1794 | return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); |
2cba3ffb IM |
1795 | } |
1796 | ||
8a59f3cc | 1797 | static const char * const stat_record_usage[] = { |
4979d0c7 JO |
1798 | "perf stat record [<options>]", |
1799 | NULL, | |
1800 | }; | |
1801 | ||
3ba78bd0 JO |
1802 | static void init_features(struct perf_session *session) |
1803 | { | |
1804 | int feat; | |
1805 | ||
1806 | for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) | |
1807 | perf_header__set_feat(&session->header, feat); | |
1808 | ||
8002a63f | 1809 | perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); |
3ba78bd0 JO |
1810 | perf_header__clear_feat(&session->header, HEADER_BUILD_ID); |
1811 | perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); | |
1812 | perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); | |
1813 | perf_header__clear_feat(&session->header, HEADER_AUXTRACE); | |
1814 | } | |
1815 | ||
4979d0c7 JO |
1816 | static int __cmd_record(int argc, const char **argv) |
1817 | { | |
1818 | struct perf_session *session; | |
8ceb41d7 | 1819 | struct perf_data *data = &perf_stat.data; |
4979d0c7 | 1820 | |
8a59f3cc | 1821 | argc = parse_options(argc, argv, stat_options, stat_record_usage, |
4979d0c7 JO |
1822 | PARSE_OPT_STOP_AT_NON_OPTION); |
1823 | ||
1824 | if (output_name) | |
2d4f2799 | 1825 | data->path = output_name; |
4979d0c7 | 1826 | |
d97ae04b | 1827 | if (stat_config.run_count != 1 || forever) { |
e9d6db8e JO |
1828 | pr_err("Cannot use -r option with perf stat record.\n"); |
1829 | return -1; | |
1830 | } | |
1831 | ||
8ceb41d7 | 1832 | session = perf_session__new(data, false, NULL); |
6ef81c55 MI |
1833 | if (IS_ERR(session)) { |
1834 | pr_err("Perf session creation failed\n"); | |
1835 | return PTR_ERR(session); | |
4979d0c7 JO |
1836 | } |
1837 | ||
3ba78bd0 JO |
1838 | init_features(session); |
1839 | ||
4979d0c7 JO |
1840 | session->evlist = evsel_list; |
1841 | perf_stat.session = session; | |
1842 | perf_stat.record = true; | |
1843 | return argc; | |
1844 | } | |
1845 | ||
89f1688a JO |
1846 | static int process_stat_round_event(struct perf_session *session, |
1847 | union perf_event *event) | |
a56f9390 | 1848 | { |
72932371 | 1849 | struct perf_record_stat_round *stat_round = &event->stat_round; |
32dcd021 | 1850 | struct evsel *counter; |
a56f9390 JO |
1851 | struct timespec tsh, *ts = NULL; |
1852 | const char **argv = session->header.env.cmdline_argv; | |
1853 | int argc = session->header.env.nr_cmdline; | |
1854 | ||
e5cadb93 | 1855 | evlist__for_each_entry(evsel_list, counter) |
a56f9390 JO |
1856 | perf_stat_process_counter(&stat_config, counter); |
1857 | ||
e3b03b6c AK |
1858 | if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) |
1859 | update_stats(&walltime_nsecs_stats, stat_round->time); | |
a56f9390 | 1860 | |
e3b03b6c | 1861 | if (stat_config.interval && stat_round->time) { |
bd48c63e ACM |
1862 | tsh.tv_sec = stat_round->time / NSEC_PER_SEC; |
1863 | tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; | |
a56f9390 JO |
1864 | ts = &tsh; |
1865 | } | |
1866 | ||
1867 | print_counters(ts, argc, argv); | |
1868 | return 0; | |
1869 | } | |
1870 | ||
62ba18ba | 1871 | static |
89f1688a JO |
1872 | int process_stat_config_event(struct perf_session *session, |
1873 | union perf_event *event) | |
62ba18ba | 1874 | { |
89f1688a | 1875 | struct perf_tool *tool = session->tool; |
68d702f7 JO |
1876 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
1877 | ||
62ba18ba | 1878 | perf_event__read_stat_config(&stat_config, &event->stat_config); |
68d702f7 | 1879 | |
315c0a1f | 1880 | if (perf_cpu_map__empty(st->cpus)) { |
89af4e05 JO |
1881 | if (st->aggr_mode != AGGR_UNSET) |
1882 | pr_warning("warning: processing task data, aggregation mode not set\n"); | |
1883 | return 0; | |
1884 | } | |
1885 | ||
1886 | if (st->aggr_mode != AGGR_UNSET) | |
1887 | stat_config.aggr_mode = st->aggr_mode; | |
1888 | ||
8ceb41d7 | 1889 | if (perf_stat.data.is_pipe) |
68d702f7 JO |
1890 | perf_stat_init_aggr_mode(); |
1891 | else | |
1892 | perf_stat_init_aggr_mode_file(st); | |
1893 | ||
62ba18ba JO |
1894 | return 0; |
1895 | } | |
1896 | ||
1975d36e JO |
1897 | static int set_maps(struct perf_stat *st) |
1898 | { | |
1899 | if (!st->cpus || !st->threads) | |
1900 | return 0; | |
1901 | ||
1902 | if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) | |
1903 | return -EINVAL; | |
1904 | ||
453fa030 | 1905 | perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); |
1975d36e JO |
1906 | |
1907 | if (perf_evlist__alloc_stats(evsel_list, true)) | |
1908 | return -ENOMEM; | |
1909 | ||
1910 | st->maps_allocated = true; | |
1911 | return 0; | |
1912 | } | |
1913 | ||
1914 | static | |
89f1688a JO |
1915 | int process_thread_map_event(struct perf_session *session, |
1916 | union perf_event *event) | |
1975d36e | 1917 | { |
89f1688a | 1918 | struct perf_tool *tool = session->tool; |
1975d36e JO |
1919 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
1920 | ||
1921 | if (st->threads) { | |
1922 | pr_warning("Extra thread map event, ignoring.\n"); | |
1923 | return 0; | |
1924 | } | |
1925 | ||
1926 | st->threads = thread_map__new_event(&event->thread_map); | |
1927 | if (!st->threads) | |
1928 | return -ENOMEM; | |
1929 | ||
1930 | return set_maps(st); | |
1931 | } | |
1932 | ||
1933 | static | |
89f1688a JO |
1934 | int process_cpu_map_event(struct perf_session *session, |
1935 | union perf_event *event) | |
1975d36e | 1936 | { |
89f1688a | 1937 | struct perf_tool *tool = session->tool; |
1975d36e | 1938 | struct perf_stat *st = container_of(tool, struct perf_stat, tool); |
f854839b | 1939 | struct perf_cpu_map *cpus; |
1975d36e JO |
1940 | |
1941 | if (st->cpus) { | |
1942 | pr_warning("Extra cpu map event, ignoring.\n"); | |
1943 | return 0; | |
1944 | } | |
1945 | ||
1946 | cpus = cpu_map__new_data(&event->cpu_map.data); | |
1947 | if (!cpus) | |
1948 | return -ENOMEM; | |
1949 | ||
1950 | st->cpus = cpus; | |
1951 | return set_maps(st); | |
1952 | } | |
1953 | ||
8a59f3cc | 1954 | static const char * const stat_report_usage[] = { |
ba6039b6 JO |
1955 | "perf stat report [<options>]", |
1956 | NULL, | |
1957 | }; | |
1958 | ||
1959 | static struct perf_stat perf_stat = { | |
1960 | .tool = { | |
1961 | .attr = perf_event__process_attr, | |
fa6ea781 | 1962 | .event_update = perf_event__process_event_update, |
1975d36e JO |
1963 | .thread_map = process_thread_map_event, |
1964 | .cpu_map = process_cpu_map_event, | |
62ba18ba | 1965 | .stat_config = process_stat_config_event, |
a56f9390 JO |
1966 | .stat = perf_event__process_stat_event, |
1967 | .stat_round = process_stat_round_event, | |
ba6039b6 | 1968 | }, |
89af4e05 | 1969 | .aggr_mode = AGGR_UNSET, |
ba6039b6 JO |
1970 | }; |
1971 | ||
1972 | static int __cmd_report(int argc, const char **argv) | |
1973 | { | |
1974 | struct perf_session *session; | |
1975 | const struct option options[] = { | |
1976 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
89af4e05 JO |
1977 | OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, |
1978 | "aggregate counts per processor socket", AGGR_SOCKET), | |
db5742b6 KL |
1979 | OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, |
1980 | "aggregate counts per processor die", AGGR_DIE), | |
89af4e05 JO |
1981 | OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, |
1982 | "aggregate counts per physical processor core", AGGR_CORE), | |
86895b48 JO |
1983 | OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode, |
1984 | "aggregate counts per numa node", AGGR_NODE), | |
89af4e05 JO |
1985 | OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, |
1986 | "disable CPU count aggregation", AGGR_NONE), | |
ba6039b6 JO |
1987 | OPT_END() |
1988 | }; | |
1989 | struct stat st; | |
1990 | int ret; | |
1991 | ||
8a59f3cc | 1992 | argc = parse_options(argc, argv, options, stat_report_usage, 0); |
ba6039b6 JO |
1993 | |
1994 | if (!input_name || !strlen(input_name)) { | |
1995 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) | |
1996 | input_name = "-"; | |
1997 | else | |
1998 | input_name = "perf.data"; | |
1999 | } | |
2000 | ||
2d4f2799 JO |
2001 | perf_stat.data.path = input_name; |
2002 | perf_stat.data.mode = PERF_DATA_MODE_READ; | |
ba6039b6 | 2003 | |
8ceb41d7 | 2004 | session = perf_session__new(&perf_stat.data, false, &perf_stat.tool); |
6ef81c55 MI |
2005 | if (IS_ERR(session)) |
2006 | return PTR_ERR(session); | |
ba6039b6 JO |
2007 | |
2008 | perf_stat.session = session; | |
2009 | stat_config.output = stderr; | |
2010 | evsel_list = session->evlist; | |
2011 | ||
2012 | ret = perf_session__process_events(session); | |
2013 | if (ret) | |
2014 | return ret; | |
2015 | ||
2016 | perf_session__delete(session); | |
2017 | return 0; | |
2018 | } | |
2019 | ||
e3ba76de JO |
2020 | static void setup_system_wide(int forks) |
2021 | { | |
2022 | /* | |
2023 | * Make system wide (-a) the default target if | |
2024 | * no target was specified and one of following | |
2025 | * conditions is met: | |
2026 | * | |
2027 | * - there's no workload specified | |
2028 | * - there is workload specified but all requested | |
2029 | * events are system wide events | |
2030 | */ | |
2031 | if (!target__none(&target)) | |
2032 | return; | |
2033 | ||
2034 | if (!forks) | |
2035 | target.system_wide = true; | |
2036 | else { | |
32dcd021 | 2037 | struct evsel *counter; |
e3ba76de JO |
2038 | |
2039 | evlist__for_each_entry(evsel_list, counter) { | |
002a3d69 JY |
2040 | if (!counter->core.system_wide && |
2041 | strcmp(counter->name, "duration_time")) { | |
e3ba76de | 2042 | return; |
002a3d69 | 2043 | } |
e3ba76de JO |
2044 | } |
2045 | ||
6484d2f9 | 2046 | if (evsel_list->core.nr_entries) |
e3ba76de JO |
2047 | target.system_wide = true; |
2048 | } | |
2049 | } | |
2050 | ||
b0ad8ea6 | 2051 | int cmd_stat(int argc, const char **argv) |
5242519b | 2052 | { |
b070a547 ACM |
2053 | const char * const stat_usage[] = { |
2054 | "perf stat [<options>] [<command>]", | |
2055 | NULL | |
2056 | }; | |
cc03c542 | 2057 | int status = -EINVAL, run_idx; |
4aa9015f | 2058 | const char *mode; |
5821522e | 2059 | FILE *output = stderr; |
f1f8ad52 | 2060 | unsigned int interval, timeout; |
ba6039b6 | 2061 | const char * const stat_subcommands[] = { "record", "report" }; |
42202dd5 | 2062 | |
5af52b51 SE |
2063 | setlocale(LC_ALL, ""); |
2064 | ||
0f98b11c | 2065 | evsel_list = evlist__new(); |
361c99a6 ACM |
2066 | if (evsel_list == NULL) |
2067 | return -ENOMEM; | |
2068 | ||
1669e509 | 2069 | parse_events__shrink_config_terms(); |
51433ead MP |
2070 | |
2071 | /* String-parsing callback-based options would segfault when negated */ | |
2072 | set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); | |
2073 | set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); | |
2074 | set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); | |
2075 | ||
4979d0c7 JO |
2076 | argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, |
2077 | (const char **) stat_usage, | |
2078 | PARSE_OPT_STOP_AT_NON_OPTION); | |
37932c18 | 2079 | perf_stat__collect_metric_expr(evsel_list); |
fb4605ba | 2080 | perf_stat__init_shadow_stats(); |
4979d0c7 | 2081 | |
fa7070a3 JO |
2082 | if (stat_config.csv_sep) { |
2083 | stat_config.csv_output = true; | |
2084 | if (!strcmp(stat_config.csv_sep, "\\t")) | |
2085 | stat_config.csv_sep = "\t"; | |
6edb78a2 | 2086 | } else |
fa7070a3 | 2087 | stat_config.csv_sep = DEFAULT_SEPARATOR; |
6edb78a2 | 2088 | |
4979d0c7 JO |
2089 | if (argc && !strncmp(argv[0], "rec", 3)) { |
2090 | argc = __cmd_record(argc, argv); | |
2091 | if (argc < 0) | |
2092 | return -1; | |
ba6039b6 JO |
2093 | } else if (argc && !strncmp(argv[0], "rep", 3)) |
2094 | return __cmd_report(argc, argv); | |
d7470b6a | 2095 | |
ec0d3d1f | 2096 | interval = stat_config.interval; |
f1f8ad52 | 2097 | timeout = stat_config.timeout; |
ec0d3d1f | 2098 | |
4979d0c7 JO |
2099 | /* |
2100 | * For record command the -o is already taken care of. | |
2101 | */ | |
2102 | if (!STAT_RECORD && output_name && strcmp(output_name, "-")) | |
4aa9015f SE |
2103 | output = NULL; |
2104 | ||
56f3bae7 JC |
2105 | if (output_name && output_fd) { |
2106 | fprintf(stderr, "cannot use both --output and --log-fd\n"); | |
e0547311 JO |
2107 | parse_options_usage(stat_usage, stat_options, "o", 1); |
2108 | parse_options_usage(NULL, stat_options, "log-fd", 0); | |
cc03c542 | 2109 | goto out; |
56f3bae7 | 2110 | } |
fc3e4d07 | 2111 | |
0ce5aa02 | 2112 | if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { |
54b50916 AK |
2113 | fprintf(stderr, "--metric-only is not supported with --per-thread\n"); |
2114 | goto out; | |
2115 | } | |
2116 | ||
d97ae04b | 2117 | if (stat_config.metric_only && stat_config.run_count > 1) { |
54b50916 AK |
2118 | fprintf(stderr, "--metric-only is not supported with -r\n"); |
2119 | goto out; | |
2120 | } | |
2121 | ||
54ac0b1b | 2122 | if (stat_config.walltime_run_table && stat_config.run_count <= 1) { |
e55c14af JO |
2123 | fprintf(stderr, "--table is only supported with -r\n"); |
2124 | parse_options_usage(stat_usage, stat_options, "r", 1); | |
2125 | parse_options_usage(NULL, stat_options, "table", 0); | |
2126 | goto out; | |
2127 | } | |
2128 | ||
fc3e4d07 SE |
2129 | if (output_fd < 0) { |
2130 | fprintf(stderr, "argument to --log-fd must be a > 0\n"); | |
e0547311 | 2131 | parse_options_usage(stat_usage, stat_options, "log-fd", 0); |
cc03c542 | 2132 | goto out; |
fc3e4d07 SE |
2133 | } |
2134 | ||
4aa9015f SE |
2135 | if (!output) { |
2136 | struct timespec tm; | |
2137 | mode = append_file ? "a" : "w"; | |
2138 | ||
2139 | output = fopen(output_name, mode); | |
2140 | if (!output) { | |
2141 | perror("failed to create output file"); | |
fceda7fe | 2142 | return -1; |
4aa9015f SE |
2143 | } |
2144 | clock_gettime(CLOCK_REALTIME, &tm); | |
2145 | fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); | |
fc3e4d07 | 2146 | } else if (output_fd > 0) { |
56f3bae7 JC |
2147 | mode = append_file ? "a" : "w"; |
2148 | output = fdopen(output_fd, mode); | |
2149 | if (!output) { | |
2150 | perror("Failed opening logfd"); | |
2151 | return -errno; | |
2152 | } | |
4aa9015f SE |
2153 | } |
2154 | ||
5821522e JO |
2155 | stat_config.output = output; |
2156 | ||
d7470b6a SE |
2157 | /* |
2158 | * let the spreadsheet do the pretty-printing | |
2159 | */ | |
fa7070a3 | 2160 | if (stat_config.csv_output) { |
61a9f324 | 2161 | /* User explicitly passed -B? */ |
d7470b6a SE |
2162 | if (big_num_opt == 1) { |
2163 | fprintf(stderr, "-B option not supported with -x\n"); | |
e0547311 JO |
2164 | parse_options_usage(stat_usage, stat_options, "B", 1); |
2165 | parse_options_usage(NULL, stat_options, "x", 1); | |
cc03c542 | 2166 | goto out; |
d7470b6a | 2167 | } else /* Nope, so disable big number formatting */ |
34ff0866 | 2168 | stat_config.big_num = false; |
d7470b6a | 2169 | } else if (big_num_opt == 0) /* User passed --no-big-num */ |
34ff0866 | 2170 | stat_config.big_num = false; |
d7470b6a | 2171 | |
e3ba76de | 2172 | setup_system_wide(argc); |
ac3063bd | 2173 | |
0ce2da14 JO |
2174 | /* |
2175 | * Display user/system times only for single | |
2176 | * run and when there's specified tracee. | |
2177 | */ | |
d97ae04b | 2178 | if ((stat_config.run_count == 1) && target__none(&target)) |
8897a891 | 2179 | stat_config.ru_display = true; |
0ce2da14 | 2180 | |
d97ae04b | 2181 | if (stat_config.run_count < 0) { |
cc03c542 | 2182 | pr_err("Run count must be a positive number\n"); |
e0547311 | 2183 | parse_options_usage(stat_usage, stat_options, "r", 1); |
cc03c542 | 2184 | goto out; |
d97ae04b | 2185 | } else if (stat_config.run_count == 0) { |
a7e191c3 | 2186 | forever = true; |
d97ae04b | 2187 | stat_config.run_count = 1; |
a7e191c3 | 2188 | } |
ddcacfa0 | 2189 | |
54ac0b1b JO |
2190 | if (stat_config.walltime_run_table) { |
2191 | stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); | |
2192 | if (!stat_config.walltime_run) { | |
e55c14af JO |
2193 | pr_err("failed to setup -r option"); |
2194 | goto out; | |
2195 | } | |
2196 | } | |
2197 | ||
1d9f8d1b JY |
2198 | if ((stat_config.aggr_mode == AGGR_THREAD) && |
2199 | !target__has_task(&target)) { | |
2200 | if (!target.system_wide || target.cpu_list) { | |
2201 | fprintf(stderr, "The --per-thread option is only " | |
2202 | "available when monitoring via -p -t -a " | |
2203 | "options or only --per-thread.\n"); | |
2204 | parse_options_usage(NULL, stat_options, "p", 1); | |
2205 | parse_options_usage(NULL, stat_options, "t", 1); | |
2206 | goto out; | |
2207 | } | |
32b8af82 JO |
2208 | } |
2209 | ||
2210 | /* | |
2211 | * no_aggr, cgroup are for system-wide only | |
2212 | * --per-thread is aggregated per thread, we dont mix it with cpu mode | |
2213 | */ | |
421a50f3 JO |
2214 | if (((stat_config.aggr_mode != AGGR_GLOBAL && |
2215 | stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && | |
602ad878 | 2216 | !target__has_cpu(&target)) { |
023695d9 SE |
2217 | fprintf(stderr, "both cgroup and no-aggregation " |
2218 | "modes only available in system-wide mode\n"); | |
2219 | ||
e0547311 JO |
2220 | parse_options_usage(stat_usage, stat_options, "G", 1); |
2221 | parse_options_usage(NULL, stat_options, "A", 1); | |
2222 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 | 2223 | goto out; |
d7e7a451 SE |
2224 | } |
2225 | ||
2cba3ffb IM |
2226 | if (add_default_attributes()) |
2227 | goto out; | |
ddcacfa0 | 2228 | |
d1c5a0e8 NK |
2229 | if (stat_config.cgroup_list) { |
2230 | if (nr_cgroups > 0) { | |
2231 | pr_err("--cgroup and --for-each-cgroup cannot be used together\n"); | |
2232 | parse_options_usage(stat_usage, stat_options, "G", 1); | |
2233 | parse_options_usage(NULL, stat_options, "for-each-cgroup", 0); | |
2234 | goto out; | |
2235 | } | |
2236 | ||
b214ba8c | 2237 | if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, |
89fb1ca2 | 2238 | &stat_config.metric_events, true) < 0) |
d1c5a0e8 NK |
2239 | goto out; |
2240 | } | |
2241 | ||
602ad878 | 2242 | target__validate(&target); |
5c98d466 | 2243 | |
1d9f8d1b JY |
2244 | if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) |
2245 | target.per_thread = true; | |
2246 | ||
77a6f014 | 2247 | if (perf_evlist__create_maps(evsel_list, &target) < 0) { |
602ad878 | 2248 | if (target__has_task(&target)) { |
77a6f014 | 2249 | pr_err("Problems finding threads of monitor\n"); |
e0547311 JO |
2250 | parse_options_usage(stat_usage, stat_options, "p", 1); |
2251 | parse_options_usage(NULL, stat_options, "t", 1); | |
602ad878 | 2252 | } else if (target__has_cpu(&target)) { |
77a6f014 | 2253 | perror("failed to parse CPUs map"); |
e0547311 JO |
2254 | parse_options_usage(stat_usage, stat_options, "C", 1); |
2255 | parse_options_usage(NULL, stat_options, "a", 1); | |
cc03c542 NK |
2256 | } |
2257 | goto out; | |
60d567e2 | 2258 | } |
32b8af82 | 2259 | |
a9a17902 JO |
2260 | evlist__check_cpu_maps(evsel_list); |
2261 | ||
32b8af82 JO |
2262 | /* |
2263 | * Initialize thread_map with comm names, | |
2264 | * so we could print it out on output. | |
2265 | */ | |
56739444 | 2266 | if (stat_config.aggr_mode == AGGR_THREAD) { |
03617c22 | 2267 | thread_map__read_comms(evsel_list->core.threads); |
56739444 JY |
2268 | if (target.system_wide) { |
2269 | if (runtime_stat_new(&stat_config, | |
a2f354e3 | 2270 | perf_thread_map__nr(evsel_list->core.threads))) { |
56739444 JY |
2271 | goto out; |
2272 | } | |
2273 | } | |
2274 | } | |
32b8af82 | 2275 | |
86895b48 JO |
2276 | if (stat_config.aggr_mode == AGGR_NODE) |
2277 | cpu__setup_cpunode_map(); | |
2278 | ||
db06a269 | 2279 | if (stat_config.times && interval) |
2280 | interval_count = true; | |
2281 | else if (stat_config.times && !interval) { | |
2282 | pr_err("interval-count option should be used together with " | |
2283 | "interval-print.\n"); | |
2284 | parse_options_usage(stat_usage, stat_options, "interval-count", 0); | |
2285 | parse_options_usage(stat_usage, stat_options, "I", 1); | |
2286 | goto out; | |
2287 | } | |
c45c6ea2 | 2288 | |
f1f8ad52 | 2289 | if (timeout && timeout < 100) { |
2290 | if (timeout < 10) { | |
2291 | pr_err("timeout must be >= 10ms.\n"); | |
2292 | parse_options_usage(stat_usage, stat_options, "timeout", 0); | |
2293 | goto out; | |
2294 | } else | |
2295 | pr_warning("timeout < 100ms. " | |
2296 | "The overhead percentage could be high in some cases. " | |
2297 | "Please proceed with caution.\n"); | |
2298 | } | |
2299 | if (timeout && interval) { | |
2300 | pr_err("timeout option is not supported with interval-print.\n"); | |
2301 | parse_options_usage(stat_usage, stat_options, "timeout", 0); | |
2302 | parse_options_usage(stat_usage, stat_options, "I", 1); | |
2303 | goto out; | |
2304 | } | |
2305 | ||
d134ffb9 | 2306 | if (perf_evlist__alloc_stats(evsel_list, interval)) |
03ad9747 | 2307 | goto out; |
d6d901c2 | 2308 | |
86ee6e18 | 2309 | if (perf_stat_init_aggr_mode()) |
03ad9747 | 2310 | goto out; |
86ee6e18 | 2311 | |
7d9ad16a JO |
2312 | /* |
2313 | * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless | |
2314 | * while avoiding that older tools show confusing messages. | |
2315 | * | |
2316 | * However for pipe sessions we need to keep it zero, | |
2317 | * because script's perf_evsel__check_attr is triggered | |
2318 | * by attr->sample_type != 0, and we can't run it on | |
2319 | * stat sessions. | |
2320 | */ | |
2321 | stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); | |
2322 | ||
58d7e993 IM |
2323 | /* |
2324 | * We dont want to block the signals - that would cause | |
2325 | * child tasks to inherit that and Ctrl-C would not work. | |
2326 | * What we want is for Ctrl-C to work in the exec()-ed | |
2327 | * task, but being ignored by perf stat itself: | |
2328 | */ | |
f7b7c26e | 2329 | atexit(sig_atexit); |
a7e191c3 FD |
2330 | if (!forever) |
2331 | signal(SIGINT, skip_signal); | |
13370a9b | 2332 | signal(SIGCHLD, skip_signal); |
58d7e993 IM |
2333 | signal(SIGALRM, skip_signal); |
2334 | signal(SIGABRT, skip_signal); | |
2335 | ||
27e9769a AB |
2336 | if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack)) |
2337 | goto out; | |
2338 | ||
42202dd5 | 2339 | status = 0; |
d97ae04b JO |
2340 | for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { |
2341 | if (stat_config.run_count != 1 && verbose > 0) | |
4aa9015f SE |
2342 | fprintf(output, "[ perf stat: executing run #%d ... ]\n", |
2343 | run_idx + 1); | |
f9cef0a9 | 2344 | |
b63fd11c SD |
2345 | if (run_idx != 0) |
2346 | perf_evlist__reset_prev_raw_counts(evsel_list); | |
2347 | ||
e55c14af | 2348 | status = run_perf_stat(argc, argv, run_idx); |
443f2d5b | 2349 | if (forever && status != -1 && !interval) { |
d4f63a47 | 2350 | print_counters(NULL, argc, argv); |
254ecbc7 | 2351 | perf_stat__reset_stats(); |
a7e191c3 | 2352 | } |
42202dd5 IM |
2353 | } |
2354 | ||
c7e5b328 | 2355 | if (!forever && status != -1 && (!interval || stat_config.summary)) |
d4f63a47 | 2356 | print_counters(NULL, argc, argv); |
d134ffb9 | 2357 | |
27e9769a AB |
2358 | evlist__finalize_ctlfd(evsel_list); |
2359 | ||
4979d0c7 JO |
2360 | if (STAT_RECORD) { |
2361 | /* | |
2362 | * We synthesize the kernel mmap record just so that older tools | |
2363 | * don't emit warnings about not being able to resolve symbols | |
2364 | * due to /proc/sys/kernel/kptr_restrict settings and instear provide | |
2365 | * a saner message about no samples being in the perf.data file. | |
2366 | * | |
2367 | * This also serves to suppress a warning about f_header.data.size == 0 | |
8b99b1a4 JO |
2368 | * in header.c at the moment 'perf stat record' gets introduced, which |
2369 | * is not really needed once we start adding the stat specific PERF_RECORD_ | |
2370 | * records, but the need to suppress the kptr_restrict messages in older | |
2371 | * tools remain -acme | |
4979d0c7 | 2372 | */ |
8ceb41d7 | 2373 | int fd = perf_data__fd(&perf_stat.data); |
4979d0c7 JO |
2374 | int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, |
2375 | process_synthesized_event, | |
2376 | &perf_stat.session->machines.host); | |
2377 | if (err) { | |
2378 | pr_warning("Couldn't synthesize the kernel mmap record, harmless, " | |
2379 | "older tools may produce warnings about this file\n."); | |
2380 | } | |
2381 | ||
7aad0c32 JO |
2382 | if (!interval) { |
2383 | if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) | |
2384 | pr_err("failed to write stat round event\n"); | |
2385 | } | |
2386 | ||
8ceb41d7 | 2387 | if (!perf_stat.data.is_pipe) { |
664c98d4 JO |
2388 | perf_stat.session->header.data_size += perf_stat.bytes_written; |
2389 | perf_session__write_header(perf_stat.session, evsel_list, fd, true); | |
2390 | } | |
4979d0c7 | 2391 | |
750b4ede | 2392 | evlist__close(evsel_list); |
4979d0c7 JO |
2393 | perf_session__delete(perf_stat.session); |
2394 | } | |
2395 | ||
544c2ae7 | 2396 | perf_stat__exit_aggr_mode(); |
d134ffb9 | 2397 | perf_evlist__free_stats(evsel_list); |
0015e2e1 | 2398 | out: |
d8f9da24 | 2399 | zfree(&stat_config.walltime_run); |
e55c14af | 2400 | |
daefd0bc KL |
2401 | if (smi_cost && smi_reset) |
2402 | sysfs__write_int(FREEZE_ON_SMI_PATH, 0); | |
2403 | ||
c12995a5 | 2404 | evlist__delete(evsel_list); |
56739444 | 2405 | |
9afe5658 | 2406 | metricgroup__rblist_exit(&stat_config.metric_events); |
56739444 | 2407 | runtime_stat_delete(&stat_config); |
ee7fe31e | 2408 | evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close); |
56739444 | 2409 | |
42202dd5 | 2410 | return status; |
ddcacfa0 | 2411 | } |