| 1 | /* |
| 2 | * builtin-stat.c |
| 3 | * |
| 4 | * Builtin stat command: Give a precise performance counters summary |
| 5 | * overview about any workload, CPU or specific PID. |
| 6 | * |
| 7 | * Sample output: |
| 8 | |
| 9 | $ perf stat ~/hackbench 10 |
| 10 | Time: 0.104 |
| 11 | |
| 12 | Performance counter stats for '/home/mingo/hackbench': |
| 13 | |
| 14 | 1255.538611 task clock ticks # 10.143 CPU utilization factor |
| 15 | 54011 context switches # 0.043 M/sec |
| 16 | 385 CPU migrations # 0.000 M/sec |
| 17 | 17755 pagefaults # 0.014 M/sec |
| 18 | 3808323185 CPU cycles # 3033.219 M/sec |
| 19 | 1575111190 instructions # 1254.530 M/sec |
| 20 | 17367895 cache references # 13.833 M/sec |
| 21 | 7674421 cache misses # 6.112 M/sec |
| 22 | |
| 23 | Wall-clock time elapsed: 123.786620 msecs |
| 24 | |
| 25 | * |
| 26 | * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> |
| 27 | * |
| 28 | * Improvements and fixes by: |
| 29 | * |
| 30 | * Arjan van de Ven <arjan@linux.intel.com> |
| 31 | * Yanmin Zhang <yanmin.zhang@intel.com> |
| 32 | * Wu Fengguang <fengguang.wu@intel.com> |
| 33 | * Mike Galbraith <efault@gmx.de> |
| 34 | * Paul Mackerras <paulus@samba.org> |
| 35 | * Jaswinder Singh Rajput <jaswinder@kernel.org> |
| 36 | * |
| 37 | * Released under the GPL v2. (and only v2, not any later version) |
| 38 | */ |
| 39 | |
| 40 | #include "perf.h" |
| 41 | #include "builtin.h" |
| 42 | #include "util/util.h" |
| 43 | #include "util/parse-options.h" |
| 44 | #include "util/parse-events.h" |
| 45 | #include "util/event.h" |
| 46 | #include "util/evlist.h" |
| 47 | #include "util/evsel.h" |
| 48 | #include "util/debug.h" |
| 49 | #include "util/header.h" |
| 50 | #include "util/cpumap.h" |
| 51 | #include "util/thread.h" |
| 52 | #include "util/thread_map.h" |
| 53 | |
| 54 | #include <sys/prctl.h> |
| 55 | #include <math.h> |
| 56 | #include <locale.h> |
| 57 | |
| 58 | #define DEFAULT_SEPARATOR " " |
| 59 | |
| 60 | static struct perf_event_attr default_attrs[] = { |
| 61 | |
| 62 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, |
| 63 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, |
| 64 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, |
| 65 | { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, |
| 66 | |
| 67 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, |
| 68 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, |
| 69 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, |
| 70 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, |
| 71 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES }, |
| 72 | { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, |
| 73 | |
| 74 | }; |
| 75 | |
| 76 | struct perf_evlist *evsel_list; |
| 77 | |
| 78 | static bool system_wide = false; |
| 79 | static int run_idx = 0; |
| 80 | |
| 81 | static int run_count = 1; |
| 82 | static bool no_inherit = false; |
| 83 | static bool scale = true; |
| 84 | static bool no_aggr = false; |
| 85 | static pid_t target_pid = -1; |
| 86 | static pid_t target_tid = -1; |
| 87 | static pid_t child_pid = -1; |
| 88 | static bool null_run = false; |
| 89 | static bool big_num = true; |
| 90 | static int big_num_opt = -1; |
| 91 | static const char *cpu_list; |
| 92 | static const char *csv_sep = NULL; |
| 93 | static bool csv_output = false; |
| 94 | |
| 95 | static volatile int done = 0; |
| 96 | |
| 97 | struct stats |
| 98 | { |
| 99 | double n, mean, M2; |
| 100 | }; |
| 101 | |
| 102 | struct perf_stat { |
| 103 | struct stats res_stats[3]; |
| 104 | }; |
| 105 | |
| 106 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
| 107 | { |
| 108 | evsel->priv = zalloc(sizeof(struct perf_stat)); |
| 109 | return evsel->priv == NULL ? -ENOMEM : 0; |
| 110 | } |
| 111 | |
| 112 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) |
| 113 | { |
| 114 | free(evsel->priv); |
| 115 | evsel->priv = NULL; |
| 116 | } |
| 117 | |
| 118 | static void update_stats(struct stats *stats, u64 val) |
| 119 | { |
| 120 | double delta; |
| 121 | |
| 122 | stats->n++; |
| 123 | delta = val - stats->mean; |
| 124 | stats->mean += delta / stats->n; |
| 125 | stats->M2 += delta*(val - stats->mean); |
| 126 | } |
| 127 | |
| 128 | static double avg_stats(struct stats *stats) |
| 129 | { |
| 130 | return stats->mean; |
| 131 | } |
| 132 | |
| 133 | /* |
| 134 | * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance |
| 135 | * |
| 136 | * (\Sum n_i^2) - ((\Sum n_i)^2)/n |
| 137 | * s^2 = ------------------------------- |
| 138 | * n - 1 |
| 139 | * |
| 140 | * http://en.wikipedia.org/wiki/Stddev |
| 141 | * |
| 142 | * The std dev of the mean is related to the std dev by: |
| 143 | * |
| 144 | * s |
| 145 | * s_mean = ------- |
| 146 | * sqrt(n) |
| 147 | * |
| 148 | */ |
| 149 | static double stddev_stats(struct stats *stats) |
| 150 | { |
| 151 | double variance = stats->M2 / (stats->n - 1); |
| 152 | double variance_mean = variance / stats->n; |
| 153 | |
| 154 | return sqrt(variance_mean); |
| 155 | } |
| 156 | |
| 157 | struct stats runtime_nsecs_stats[MAX_NR_CPUS]; |
| 158 | struct stats runtime_cycles_stats[MAX_NR_CPUS]; |
| 159 | struct stats runtime_branches_stats[MAX_NR_CPUS]; |
| 160 | struct stats runtime_cacherefs_stats[MAX_NR_CPUS]; |
| 161 | struct stats walltime_nsecs_stats; |
| 162 | |
| 163 | static int create_perf_stat_counter(struct perf_evsel *evsel) |
| 164 | { |
| 165 | struct perf_event_attr *attr = &evsel->attr; |
| 166 | |
| 167 | if (scale) |
| 168 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | |
| 169 | PERF_FORMAT_TOTAL_TIME_RUNNING; |
| 170 | |
| 171 | attr->inherit = !no_inherit; |
| 172 | |
| 173 | if (system_wide) |
| 174 | return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false); |
| 175 | |
| 176 | if (target_pid == -1 && target_tid == -1) { |
| 177 | attr->disabled = 1; |
| 178 | attr->enable_on_exec = 1; |
| 179 | } |
| 180 | |
| 181 | return perf_evsel__open_per_thread(evsel, evsel_list->threads, false); |
| 182 | } |
| 183 | |
| 184 | /* |
| 185 | * Does the counter have nsecs as a unit? |
| 186 | */ |
| 187 | static inline int nsec_counter(struct perf_evsel *evsel) |
| 188 | { |
| 189 | if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) || |
| 190 | perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
| 191 | return 1; |
| 192 | |
| 193 | return 0; |
| 194 | } |
| 195 | |
| 196 | /* |
| 197 | * Update various tracking values we maintain to print |
| 198 | * more semantic information such as miss/hit ratios, |
| 199 | * instruction rates, etc: |
| 200 | */ |
| 201 | static void update_shadow_stats(struct perf_evsel *counter, u64 *count) |
| 202 | { |
| 203 | if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) |
| 204 | update_stats(&runtime_nsecs_stats[0], count[0]); |
| 205 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) |
| 206 | update_stats(&runtime_cycles_stats[0], count[0]); |
| 207 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
| 208 | update_stats(&runtime_branches_stats[0], count[0]); |
| 209 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) |
| 210 | update_stats(&runtime_cacherefs_stats[0], count[0]); |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * Read out the results of a single counter: |
| 215 | * aggregate counts across CPUs in system-wide mode |
| 216 | */ |
| 217 | static int read_counter_aggr(struct perf_evsel *counter) |
| 218 | { |
| 219 | struct perf_stat *ps = counter->priv; |
| 220 | u64 *count = counter->counts->aggr.values; |
| 221 | int i; |
| 222 | |
| 223 | if (__perf_evsel__read(counter, evsel_list->cpus->nr, |
| 224 | evsel_list->threads->nr, scale) < 0) |
| 225 | return -1; |
| 226 | |
| 227 | for (i = 0; i < 3; i++) |
| 228 | update_stats(&ps->res_stats[i], count[i]); |
| 229 | |
| 230 | if (verbose) { |
| 231 | fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
| 232 | event_name(counter), count[0], count[1], count[2]); |
| 233 | } |
| 234 | |
| 235 | /* |
| 236 | * Save the full runtime - to allow normalization during printout: |
| 237 | */ |
| 238 | update_shadow_stats(counter, count); |
| 239 | |
| 240 | return 0; |
| 241 | } |
| 242 | |
| 243 | /* |
| 244 | * Read out the results of a single counter: |
| 245 | * do not aggregate counts across CPUs in system-wide mode |
| 246 | */ |
| 247 | static int read_counter(struct perf_evsel *counter) |
| 248 | { |
| 249 | u64 *count; |
| 250 | int cpu; |
| 251 | |
| 252 | for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { |
| 253 | if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0) |
| 254 | return -1; |
| 255 | |
| 256 | count = counter->counts->cpu[cpu].values; |
| 257 | |
| 258 | update_shadow_stats(counter, count); |
| 259 | } |
| 260 | |
| 261 | return 0; |
| 262 | } |
| 263 | |
| 264 | static int run_perf_stat(int argc __used, const char **argv) |
| 265 | { |
| 266 | unsigned long long t0, t1; |
| 267 | struct perf_evsel *counter; |
| 268 | int status = 0; |
| 269 | int child_ready_pipe[2], go_pipe[2]; |
| 270 | const bool forks = (argc > 0); |
| 271 | char buf; |
| 272 | |
| 273 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
| 274 | perror("failed to create pipes"); |
| 275 | exit(1); |
| 276 | } |
| 277 | |
| 278 | if (forks) { |
| 279 | if ((child_pid = fork()) < 0) |
| 280 | perror("failed to fork"); |
| 281 | |
| 282 | if (!child_pid) { |
| 283 | close(child_ready_pipe[0]); |
| 284 | close(go_pipe[1]); |
| 285 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); |
| 286 | |
| 287 | /* |
| 288 | * Do a dummy execvp to get the PLT entry resolved, |
| 289 | * so we avoid the resolver overhead on the real |
| 290 | * execvp call. |
| 291 | */ |
| 292 | execvp("", (char **)argv); |
| 293 | |
| 294 | /* |
| 295 | * Tell the parent we're ready to go |
| 296 | */ |
| 297 | close(child_ready_pipe[1]); |
| 298 | |
| 299 | /* |
| 300 | * Wait until the parent tells us to go. |
| 301 | */ |
| 302 | if (read(go_pipe[0], &buf, 1) == -1) |
| 303 | perror("unable to read pipe"); |
| 304 | |
| 305 | execvp(argv[0], (char **)argv); |
| 306 | |
| 307 | perror(argv[0]); |
| 308 | exit(-1); |
| 309 | } |
| 310 | |
| 311 | if (target_tid == -1 && target_pid == -1 && !system_wide) |
| 312 | evsel_list->threads->map[0] = child_pid; |
| 313 | |
| 314 | /* |
| 315 | * Wait for the child to be ready to exec. |
| 316 | */ |
| 317 | close(child_ready_pipe[1]); |
| 318 | close(go_pipe[0]); |
| 319 | if (read(child_ready_pipe[0], &buf, 1) == -1) |
| 320 | perror("unable to read pipe"); |
| 321 | close(child_ready_pipe[0]); |
| 322 | } |
| 323 | |
| 324 | list_for_each_entry(counter, &evsel_list->entries, node) { |
| 325 | if (create_perf_stat_counter(counter) < 0) { |
| 326 | if (errno == -EPERM || errno == -EACCES) { |
| 327 | error("You may not have permission to collect %sstats.\n" |
| 328 | "\t Consider tweaking" |
| 329 | " /proc/sys/kernel/perf_event_paranoid or running as root.", |
| 330 | system_wide ? "system-wide " : ""); |
| 331 | } else if (errno == ENOENT) { |
| 332 | error("%s event is not supported. ", event_name(counter)); |
| 333 | } else { |
| 334 | error("open_counter returned with %d (%s). " |
| 335 | "/bin/dmesg may provide additional information.\n", |
| 336 | errno, strerror(errno)); |
| 337 | } |
| 338 | if (child_pid != -1) |
| 339 | kill(child_pid, SIGTERM); |
| 340 | die("Not all events could be opened.\n"); |
| 341 | return -1; |
| 342 | } |
| 343 | } |
| 344 | |
| 345 | if (perf_evlist__set_filters(evsel_list)) { |
| 346 | error("failed to set filter with %d (%s)\n", errno, |
| 347 | strerror(errno)); |
| 348 | return -1; |
| 349 | } |
| 350 | |
| 351 | /* |
| 352 | * Enable counters and exec the command: |
| 353 | */ |
| 354 | t0 = rdclock(); |
| 355 | |
| 356 | if (forks) { |
| 357 | close(go_pipe[1]); |
| 358 | wait(&status); |
| 359 | } else { |
| 360 | while(!done) sleep(1); |
| 361 | } |
| 362 | |
| 363 | t1 = rdclock(); |
| 364 | |
| 365 | update_stats(&walltime_nsecs_stats, t1 - t0); |
| 366 | |
| 367 | if (no_aggr) { |
| 368 | list_for_each_entry(counter, &evsel_list->entries, node) { |
| 369 | read_counter(counter); |
| 370 | perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1); |
| 371 | } |
| 372 | } else { |
| 373 | list_for_each_entry(counter, &evsel_list->entries, node) { |
| 374 | read_counter_aggr(counter); |
| 375 | perf_evsel__close_fd(counter, evsel_list->cpus->nr, |
| 376 | evsel_list->threads->nr); |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | return WEXITSTATUS(status); |
| 381 | } |
| 382 | |
| 383 | static void print_noise(struct perf_evsel *evsel, double avg) |
| 384 | { |
| 385 | struct perf_stat *ps; |
| 386 | |
| 387 | if (run_count == 1) |
| 388 | return; |
| 389 | |
| 390 | ps = evsel->priv; |
| 391 | fprintf(stderr, " ( +- %7.3f%% )", |
| 392 | 100 * stddev_stats(&ps->res_stats[0]) / avg); |
| 393 | } |
| 394 | |
| 395 | static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) |
| 396 | { |
| 397 | double msecs = avg / 1e6; |
| 398 | char cpustr[16] = { '\0', }; |
| 399 | const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s"; |
| 400 | |
| 401 | if (no_aggr) |
| 402 | sprintf(cpustr, "CPU%*d%s", |
| 403 | csv_output ? 0 : -4, |
| 404 | evsel_list->cpus->map[cpu], csv_sep); |
| 405 | |
| 406 | fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel)); |
| 407 | |
| 408 | if (evsel->cgrp) |
| 409 | fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name); |
| 410 | |
| 411 | if (csv_output) |
| 412 | return; |
| 413 | |
| 414 | if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) |
| 415 | fprintf(stderr, " # %10.3f CPUs", |
| 416 | avg / avg_stats(&walltime_nsecs_stats)); |
| 417 | } |
| 418 | |
| 419 | static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) |
| 420 | { |
| 421 | double total, ratio = 0.0; |
| 422 | char cpustr[16] = { '\0', }; |
| 423 | const char *fmt; |
| 424 | |
| 425 | if (csv_output) |
| 426 | fmt = "%s%.0f%s%s"; |
| 427 | else if (big_num) |
| 428 | fmt = "%s%'18.0f%s%-24s"; |
| 429 | else |
| 430 | fmt = "%s%18.0f%s%-24s"; |
| 431 | |
| 432 | if (no_aggr) |
| 433 | sprintf(cpustr, "CPU%*d%s", |
| 434 | csv_output ? 0 : -4, |
| 435 | evsel_list->cpus->map[cpu], csv_sep); |
| 436 | else |
| 437 | cpu = 0; |
| 438 | |
| 439 | fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel)); |
| 440 | |
| 441 | if (evsel->cgrp) |
| 442 | fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name); |
| 443 | |
| 444 | if (csv_output) |
| 445 | return; |
| 446 | |
| 447 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { |
| 448 | total = avg_stats(&runtime_cycles_stats[cpu]); |
| 449 | |
| 450 | if (total) |
| 451 | ratio = avg / total; |
| 452 | |
| 453 | fprintf(stderr, " # ( %4.2f instructions per cycle )", ratio); |
| 454 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && |
| 455 | runtime_branches_stats[cpu].n != 0) { |
| 456 | total = avg_stats(&runtime_branches_stats[cpu]); |
| 457 | |
| 458 | if (total) |
| 459 | ratio = avg * 100 / total; |
| 460 | |
| 461 | fprintf(stderr, " # %10.3f %%", ratio); |
| 462 | |
| 463 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) && |
| 464 | runtime_cacherefs_stats[cpu].n != 0) { |
| 465 | total = avg_stats(&runtime_cacherefs_stats[cpu]); |
| 466 | |
| 467 | if (total) |
| 468 | ratio = avg * 100 / total; |
| 469 | |
| 470 | fprintf(stderr, " # %10.3f %%", ratio); |
| 471 | |
| 472 | } else if (runtime_nsecs_stats[cpu].n != 0) { |
| 473 | total = avg_stats(&runtime_nsecs_stats[cpu]); |
| 474 | |
| 475 | if (total) |
| 476 | ratio = 1000.0 * avg / total; |
| 477 | |
| 478 | fprintf(stderr, " # %10.3f M/sec", ratio); |
| 479 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES)) { |
| 480 | total = avg_stats(&runtime_cycles_stats[cpu]); |
| 481 | |
| 482 | if (total) |
| 483 | ratio = avg / total * 100.0; |
| 484 | |
| 485 | fprintf(stderr, " # (%5.2f%% of all cycles )", ratio); |
| 486 | } |
| 487 | } |
| 488 | |
| 489 | /* |
| 490 | * Print out the results of a single counter: |
| 491 | * aggregated counts in system-wide mode |
| 492 | */ |
| 493 | static void print_counter_aggr(struct perf_evsel *counter) |
| 494 | { |
| 495 | struct perf_stat *ps = counter->priv; |
| 496 | double avg = avg_stats(&ps->res_stats[0]); |
| 497 | int scaled = counter->counts->scaled; |
| 498 | |
| 499 | if (scaled == -1) { |
| 500 | fprintf(stderr, "%*s%s%*s", |
| 501 | csv_output ? 0 : 18, |
| 502 | "<not counted>", |
| 503 | csv_sep, |
| 504 | csv_output ? 0 : -24, |
| 505 | event_name(counter)); |
| 506 | |
| 507 | if (counter->cgrp) |
| 508 | fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name); |
| 509 | |
| 510 | fputc('\n', stderr); |
| 511 | return; |
| 512 | } |
| 513 | |
| 514 | if (nsec_counter(counter)) |
| 515 | nsec_printout(-1, counter, avg); |
| 516 | else |
| 517 | abs_printout(-1, counter, avg); |
| 518 | |
| 519 | if (csv_output) { |
| 520 | fputc('\n', stderr); |
| 521 | return; |
| 522 | } |
| 523 | |
| 524 | print_noise(counter, avg); |
| 525 | |
| 526 | if (scaled) { |
| 527 | double avg_enabled, avg_running; |
| 528 | |
| 529 | avg_enabled = avg_stats(&ps->res_stats[1]); |
| 530 | avg_running = avg_stats(&ps->res_stats[2]); |
| 531 | |
| 532 | fprintf(stderr, " (scaled from %.2f%%)", |
| 533 | 100 * avg_running / avg_enabled); |
| 534 | } |
| 535 | fprintf(stderr, "\n"); |
| 536 | } |
| 537 | |
| 538 | /* |
| 539 | * Print out the results of a single counter: |
| 540 | * does not use aggregated count in system-wide |
| 541 | */ |
| 542 | static void print_counter(struct perf_evsel *counter) |
| 543 | { |
| 544 | u64 ena, run, val; |
| 545 | int cpu; |
| 546 | |
| 547 | for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) { |
| 548 | val = counter->counts->cpu[cpu].val; |
| 549 | ena = counter->counts->cpu[cpu].ena; |
| 550 | run = counter->counts->cpu[cpu].run; |
| 551 | if (run == 0 || ena == 0) { |
| 552 | fprintf(stderr, "CPU%*d%s%*s%s%*s", |
| 553 | csv_output ? 0 : -4, |
| 554 | evsel_list->cpus->map[cpu], csv_sep, |
| 555 | csv_output ? 0 : 18, |
| 556 | "<not counted>", csv_sep, |
| 557 | csv_output ? 0 : -24, |
| 558 | event_name(counter)); |
| 559 | |
| 560 | if (counter->cgrp) |
| 561 | fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name); |
| 562 | |
| 563 | fputc('\n', stderr); |
| 564 | continue; |
| 565 | } |
| 566 | |
| 567 | if (nsec_counter(counter)) |
| 568 | nsec_printout(cpu, counter, val); |
| 569 | else |
| 570 | abs_printout(cpu, counter, val); |
| 571 | |
| 572 | if (!csv_output) { |
| 573 | print_noise(counter, 1.0); |
| 574 | |
| 575 | if (run != ena) { |
| 576 | fprintf(stderr, " (scaled from %.2f%%)", |
| 577 | 100.0 * run / ena); |
| 578 | } |
| 579 | } |
| 580 | fputc('\n', stderr); |
| 581 | } |
| 582 | } |
| 583 | |
| 584 | static void print_stat(int argc, const char **argv) |
| 585 | { |
| 586 | struct perf_evsel *counter; |
| 587 | int i; |
| 588 | |
| 589 | fflush(stdout); |
| 590 | |
| 591 | if (!csv_output) { |
| 592 | fprintf(stderr, "\n"); |
| 593 | fprintf(stderr, " Performance counter stats for "); |
| 594 | if(target_pid == -1 && target_tid == -1) { |
| 595 | fprintf(stderr, "\'%s", argv[0]); |
| 596 | for (i = 1; i < argc; i++) |
| 597 | fprintf(stderr, " %s", argv[i]); |
| 598 | } else if (target_pid != -1) |
| 599 | fprintf(stderr, "process id \'%d", target_pid); |
| 600 | else |
| 601 | fprintf(stderr, "thread id \'%d", target_tid); |
| 602 | |
| 603 | fprintf(stderr, "\'"); |
| 604 | if (run_count > 1) |
| 605 | fprintf(stderr, " (%d runs)", run_count); |
| 606 | fprintf(stderr, ":\n\n"); |
| 607 | } |
| 608 | |
| 609 | if (no_aggr) { |
| 610 | list_for_each_entry(counter, &evsel_list->entries, node) |
| 611 | print_counter(counter); |
| 612 | } else { |
| 613 | list_for_each_entry(counter, &evsel_list->entries, node) |
| 614 | print_counter_aggr(counter); |
| 615 | } |
| 616 | |
| 617 | if (!csv_output) { |
| 618 | fprintf(stderr, "\n"); |
| 619 | fprintf(stderr, " %18.9f seconds time elapsed", |
| 620 | avg_stats(&walltime_nsecs_stats)/1e9); |
| 621 | if (run_count > 1) { |
| 622 | fprintf(stderr, " ( +- %7.3f%% )", |
| 623 | 100*stddev_stats(&walltime_nsecs_stats) / |
| 624 | avg_stats(&walltime_nsecs_stats)); |
| 625 | } |
| 626 | fprintf(stderr, "\n\n"); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | static volatile int signr = -1; |
| 631 | |
| 632 | static void skip_signal(int signo) |
| 633 | { |
| 634 | if(child_pid == -1) |
| 635 | done = 1; |
| 636 | |
| 637 | signr = signo; |
| 638 | } |
| 639 | |
| 640 | static void sig_atexit(void) |
| 641 | { |
| 642 | if (child_pid != -1) |
| 643 | kill(child_pid, SIGTERM); |
| 644 | |
| 645 | if (signr == -1) |
| 646 | return; |
| 647 | |
| 648 | signal(signr, SIG_DFL); |
| 649 | kill(getpid(), signr); |
| 650 | } |
| 651 | |
| 652 | static const char * const stat_usage[] = { |
| 653 | "perf stat [<options>] [<command>]", |
| 654 | NULL |
| 655 | }; |
| 656 | |
| 657 | static int stat__set_big_num(const struct option *opt __used, |
| 658 | const char *s __used, int unset) |
| 659 | { |
| 660 | big_num_opt = unset ? 0 : 1; |
| 661 | return 0; |
| 662 | } |
| 663 | |
| 664 | static const struct option options[] = { |
| 665 | OPT_CALLBACK('e', "event", &evsel_list, "event", |
| 666 | "event selector. use 'perf list' to list available events", |
| 667 | parse_events), |
| 668 | OPT_CALLBACK(0, "filter", &evsel_list, "filter", |
| 669 | "event filter", parse_filter), |
| 670 | OPT_BOOLEAN('i', "no-inherit", &no_inherit, |
| 671 | "child tasks do not inherit counters"), |
| 672 | OPT_INTEGER('p', "pid", &target_pid, |
| 673 | "stat events on existing process id"), |
| 674 | OPT_INTEGER('t', "tid", &target_tid, |
| 675 | "stat events on existing thread id"), |
| 676 | OPT_BOOLEAN('a', "all-cpus", &system_wide, |
| 677 | "system-wide collection from all CPUs"), |
| 678 | OPT_BOOLEAN('c', "scale", &scale, |
| 679 | "scale/normalize counters"), |
| 680 | OPT_INCR('v', "verbose", &verbose, |
| 681 | "be more verbose (show counter open errors, etc)"), |
| 682 | OPT_INTEGER('r', "repeat", &run_count, |
| 683 | "repeat command and print average + stddev (max: 100)"), |
| 684 | OPT_BOOLEAN('n', "null", &null_run, |
| 685 | "null run - dont start any counters"), |
| 686 | OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, |
| 687 | "print large numbers with thousands\' separators", |
| 688 | stat__set_big_num), |
| 689 | OPT_STRING('C', "cpu", &cpu_list, "cpu", |
| 690 | "list of cpus to monitor in system-wide"), |
| 691 | OPT_BOOLEAN('A', "no-aggr", &no_aggr, |
| 692 | "disable CPU count aggregation"), |
| 693 | OPT_STRING('x', "field-separator", &csv_sep, "separator", |
| 694 | "print counts with custom separator"), |
| 695 | OPT_CALLBACK('G', "cgroup", &evsel_list, "name", |
| 696 | "monitor event in cgroup name only", |
| 697 | parse_cgroups), |
| 698 | OPT_END() |
| 699 | }; |
| 700 | |
| 701 | int cmd_stat(int argc, const char **argv, const char *prefix __used) |
| 702 | { |
| 703 | struct perf_evsel *pos; |
| 704 | int status = -ENOMEM; |
| 705 | |
| 706 | setlocale(LC_ALL, ""); |
| 707 | |
| 708 | evsel_list = perf_evlist__new(NULL, NULL); |
| 709 | if (evsel_list == NULL) |
| 710 | return -ENOMEM; |
| 711 | |
| 712 | argc = parse_options(argc, argv, options, stat_usage, |
| 713 | PARSE_OPT_STOP_AT_NON_OPTION); |
| 714 | |
| 715 | if (csv_sep) |
| 716 | csv_output = true; |
| 717 | else |
| 718 | csv_sep = DEFAULT_SEPARATOR; |
| 719 | |
| 720 | /* |
| 721 | * let the spreadsheet do the pretty-printing |
| 722 | */ |
| 723 | if (csv_output) { |
| 724 | /* User explicitely passed -B? */ |
| 725 | if (big_num_opt == 1) { |
| 726 | fprintf(stderr, "-B option not supported with -x\n"); |
| 727 | usage_with_options(stat_usage, options); |
| 728 | } else /* Nope, so disable big number formatting */ |
| 729 | big_num = false; |
| 730 | } else if (big_num_opt == 0) /* User passed --no-big-num */ |
| 731 | big_num = false; |
| 732 | |
| 733 | if (!argc && target_pid == -1 && target_tid == -1) |
| 734 | usage_with_options(stat_usage, options); |
| 735 | if (run_count <= 0) |
| 736 | usage_with_options(stat_usage, options); |
| 737 | |
| 738 | /* no_aggr, cgroup are for system-wide only */ |
| 739 | if ((no_aggr || nr_cgroups) && !system_wide) { |
| 740 | fprintf(stderr, "both cgroup and no-aggregation " |
| 741 | "modes only available in system-wide mode\n"); |
| 742 | |
| 743 | usage_with_options(stat_usage, options); |
| 744 | } |
| 745 | |
| 746 | /* Set attrs and nr_counters if no event is selected and !null_run */ |
| 747 | if (!null_run && !evsel_list->nr_entries) { |
| 748 | size_t c; |
| 749 | |
| 750 | for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) { |
| 751 | pos = perf_evsel__new(&default_attrs[c], c); |
| 752 | if (pos == NULL) |
| 753 | goto out; |
| 754 | perf_evlist__add(evsel_list, pos); |
| 755 | } |
| 756 | } |
| 757 | |
| 758 | if (target_pid != -1) |
| 759 | target_tid = target_pid; |
| 760 | |
| 761 | evsel_list->threads = thread_map__new(target_pid, target_tid); |
| 762 | if (evsel_list->threads == NULL) { |
| 763 | pr_err("Problems finding threads of monitor\n"); |
| 764 | usage_with_options(stat_usage, options); |
| 765 | } |
| 766 | |
| 767 | if (system_wide) |
| 768 | evsel_list->cpus = cpu_map__new(cpu_list); |
| 769 | else |
| 770 | evsel_list->cpus = cpu_map__dummy_new(); |
| 771 | |
| 772 | if (evsel_list->cpus == NULL) { |
| 773 | perror("failed to parse CPUs map"); |
| 774 | usage_with_options(stat_usage, options); |
| 775 | return -1; |
| 776 | } |
| 777 | |
| 778 | list_for_each_entry(pos, &evsel_list->entries, node) { |
| 779 | if (perf_evsel__alloc_stat_priv(pos) < 0 || |
| 780 | perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 || |
| 781 | perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0) |
| 782 | goto out_free_fd; |
| 783 | } |
| 784 | |
| 785 | /* |
| 786 | * We dont want to block the signals - that would cause |
| 787 | * child tasks to inherit that and Ctrl-C would not work. |
| 788 | * What we want is for Ctrl-C to work in the exec()-ed |
| 789 | * task, but being ignored by perf stat itself: |
| 790 | */ |
| 791 | atexit(sig_atexit); |
| 792 | signal(SIGINT, skip_signal); |
| 793 | signal(SIGALRM, skip_signal); |
| 794 | signal(SIGABRT, skip_signal); |
| 795 | |
| 796 | status = 0; |
| 797 | for (run_idx = 0; run_idx < run_count; run_idx++) { |
| 798 | if (run_count != 1 && verbose) |
| 799 | fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); |
| 800 | status = run_perf_stat(argc, argv); |
| 801 | } |
| 802 | |
| 803 | if (status != -1) |
| 804 | print_stat(argc, argv); |
| 805 | out_free_fd: |
| 806 | list_for_each_entry(pos, &evsel_list->entries, node) |
| 807 | perf_evsel__free_stat_priv(pos); |
| 808 | perf_evlist__delete_maps(evsel_list); |
| 809 | out: |
| 810 | perf_evlist__delete(evsel_list); |
| 811 | return status; |
| 812 | } |