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