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/debug.h" 47 #include "util/header.h" 48 49 #include <sys/prctl.h> 50 #include <math.h> 51 52 static struct perf_event_attr default_attrs[] = { 53 54 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 55 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 56 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 57 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 58 59 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 60 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 61 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 62 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 63 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES }, 64 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, 65 66 }; 67 68 static int system_wide = 0; 69 static unsigned int nr_cpus = 0; 70 static int run_idx = 0; 71 72 static int run_count = 1; 73 static int inherit = 1; 74 static int scale = 1; 75 static pid_t target_pid = -1; 76 static pid_t child_pid = -1; 77 static int null_run = 0; 78 79 static int fd[MAX_NR_CPUS][MAX_COUNTERS]; 80 81 static int event_scaled[MAX_COUNTERS]; 82 83 static volatile int done = 0; 84 85 struct stats 86 { 87 double n, mean, M2; 88 }; 89 90 static void update_stats(struct stats *stats, u64 val) 91 { 92 double delta; 93 94 stats->n++; 95 delta = val - stats->mean; 96 stats->mean += delta / stats->n; 97 stats->M2 += delta*(val - stats->mean); 98 } 99 100 static double avg_stats(struct stats *stats) 101 { 102 return stats->mean; 103 } 104 105 /* 106 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 107 * 108 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 109 * s^2 = ------------------------------- 110 * n - 1 111 * 112 * http://en.wikipedia.org/wiki/Stddev 113 * 114 * The std dev of the mean is related to the std dev by: 115 * 116 * s 117 * s_mean = ------- 118 * sqrt(n) 119 * 120 */ 121 static double stddev_stats(struct stats *stats) 122 { 123 double variance = stats->M2 / (stats->n - 1); 124 double variance_mean = variance / stats->n; 125 126 return sqrt(variance_mean); 127 } 128 129 struct stats event_res_stats[MAX_COUNTERS][3]; 130 struct stats runtime_nsecs_stats; 131 struct stats walltime_nsecs_stats; 132 struct stats runtime_cycles_stats; 133 struct stats runtime_branches_stats; 134 135 #define MATCH_EVENT(t, c, counter) \ 136 (attrs[counter].type == PERF_TYPE_##t && \ 137 attrs[counter].config == PERF_COUNT_##c) 138 139 #define ERR_PERF_OPEN \ 140 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n" 141 142 static void create_perf_stat_counter(int counter, int pid) 143 { 144 struct perf_event_attr *attr = attrs + counter; 145 146 if (scale) 147 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 148 PERF_FORMAT_TOTAL_TIME_RUNNING; 149 150 if (system_wide) { 151 unsigned int cpu; 152 153 for (cpu = 0; cpu < nr_cpus; cpu++) { 154 fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0); 155 if (fd[cpu][counter] < 0 && verbose) 156 fprintf(stderr, ERR_PERF_OPEN, counter, 157 fd[cpu][counter], strerror(errno)); 158 } 159 } else { 160 attr->inherit = inherit; 161 attr->disabled = 1; 162 attr->enable_on_exec = 1; 163 164 fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0); 165 if (fd[0][counter] < 0 && verbose) 166 fprintf(stderr, ERR_PERF_OPEN, counter, 167 fd[0][counter], strerror(errno)); 168 } 169 } 170 171 /* 172 * Does the counter have nsecs as a unit? 173 */ 174 static inline int nsec_counter(int counter) 175 { 176 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) || 177 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 178 return 1; 179 180 return 0; 181 } 182 183 /* 184 * Read out the results of a single counter: 185 */ 186 static void read_counter(int counter) 187 { 188 u64 count[3], single_count[3]; 189 unsigned int cpu; 190 size_t res, nv; 191 int scaled; 192 int i; 193 194 count[0] = count[1] = count[2] = 0; 195 196 nv = scale ? 3 : 1; 197 for (cpu = 0; cpu < nr_cpus; cpu++) { 198 if (fd[cpu][counter] < 0) 199 continue; 200 201 res = read(fd[cpu][counter], single_count, nv * sizeof(u64)); 202 assert(res == nv * sizeof(u64)); 203 204 close(fd[cpu][counter]); 205 fd[cpu][counter] = -1; 206 207 count[0] += single_count[0]; 208 if (scale) { 209 count[1] += single_count[1]; 210 count[2] += single_count[2]; 211 } 212 } 213 214 scaled = 0; 215 if (scale) { 216 if (count[2] == 0) { 217 event_scaled[counter] = -1; 218 count[0] = 0; 219 return; 220 } 221 222 if (count[2] < count[1]) { 223 event_scaled[counter] = 1; 224 count[0] = (unsigned long long) 225 ((double)count[0] * count[1] / count[2] + 0.5); 226 } 227 } 228 229 for (i = 0; i < 3; i++) 230 update_stats(&event_res_stats[counter][i], count[i]); 231 232 if (verbose) { 233 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter), 234 count[0], count[1], count[2]); 235 } 236 237 /* 238 * Save the full runtime - to allow normalization during printout: 239 */ 240 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 241 update_stats(&runtime_nsecs_stats, count[0]); 242 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter)) 243 update_stats(&runtime_cycles_stats, count[0]); 244 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter)) 245 update_stats(&runtime_branches_stats, count[0]); 246 } 247 248 static int run_perf_stat(int argc __used, const char **argv) 249 { 250 unsigned long long t0, t1; 251 int status = 0; 252 int counter; 253 int pid = target_pid; 254 int child_ready_pipe[2], go_pipe[2]; 255 const bool forks = (target_pid == -1 && argc > 0); 256 char buf; 257 258 if (!system_wide) 259 nr_cpus = 1; 260 261 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 262 perror("failed to create pipes"); 263 exit(1); 264 } 265 266 if (forks) { 267 if ((pid = fork()) < 0) 268 perror("failed to fork"); 269 270 if (!pid) { 271 close(child_ready_pipe[0]); 272 close(go_pipe[1]); 273 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 274 275 /* 276 * Do a dummy execvp to get the PLT entry resolved, 277 * so we avoid the resolver overhead on the real 278 * execvp call. 279 */ 280 execvp("", (char **)argv); 281 282 /* 283 * Tell the parent we're ready to go 284 */ 285 close(child_ready_pipe[1]); 286 287 /* 288 * Wait until the parent tells us to go. 289 */ 290 if (read(go_pipe[0], &buf, 1) == -1) 291 perror("unable to read pipe"); 292 293 execvp(argv[0], (char **)argv); 294 295 perror(argv[0]); 296 exit(-1); 297 } 298 299 child_pid = pid; 300 301 /* 302 * Wait for the child to be ready to exec. 303 */ 304 close(child_ready_pipe[1]); 305 close(go_pipe[0]); 306 if (read(child_ready_pipe[0], &buf, 1) == -1) 307 perror("unable to read pipe"); 308 close(child_ready_pipe[0]); 309 } 310 311 for (counter = 0; counter < nr_counters; counter++) 312 create_perf_stat_counter(counter, pid); 313 314 /* 315 * Enable counters and exec the command: 316 */ 317 t0 = rdclock(); 318 319 if (forks) { 320 close(go_pipe[1]); 321 wait(&status); 322 } else { 323 while(!done); 324 } 325 326 t1 = rdclock(); 327 328 update_stats(&walltime_nsecs_stats, t1 - t0); 329 330 for (counter = 0; counter < nr_counters; counter++) 331 read_counter(counter); 332 333 return WEXITSTATUS(status); 334 } 335 336 static void print_noise(int counter, double avg) 337 { 338 if (run_count == 1) 339 return; 340 341 fprintf(stderr, " ( +- %7.3f%% )", 342 100 * stddev_stats(&event_res_stats[counter][0]) / avg); 343 } 344 345 static void nsec_printout(int counter, double avg) 346 { 347 double msecs = avg / 1e6; 348 349 fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter)); 350 351 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) { 352 fprintf(stderr, " # %10.3f CPUs ", 353 avg / avg_stats(&walltime_nsecs_stats)); 354 } 355 } 356 357 static void abs_printout(int counter, double avg) 358 { 359 double total, ratio = 0.0; 360 361 fprintf(stderr, " %14.0f %-24s", avg, event_name(counter)); 362 363 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) { 364 total = avg_stats(&runtime_cycles_stats); 365 366 if (total) 367 ratio = avg / total; 368 369 fprintf(stderr, " # %10.3f IPC ", ratio); 370 } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) && 371 runtime_branches_stats.n != 0) { 372 total = avg_stats(&runtime_branches_stats); 373 374 if (total) 375 ratio = avg * 100 / total; 376 377 fprintf(stderr, " # %10.3f %% ", ratio); 378 379 } else if (runtime_nsecs_stats.n != 0) { 380 total = avg_stats(&runtime_nsecs_stats); 381 382 if (total) 383 ratio = 1000.0 * avg / total; 384 385 fprintf(stderr, " # %10.3f M/sec", ratio); 386 } 387 } 388 389 /* 390 * Print out the results of a single counter: 391 */ 392 static void print_counter(int counter) 393 { 394 double avg = avg_stats(&event_res_stats[counter][0]); 395 int scaled = event_scaled[counter]; 396 397 if (scaled == -1) { 398 fprintf(stderr, " %14s %-24s\n", 399 "<not counted>", event_name(counter)); 400 return; 401 } 402 403 if (nsec_counter(counter)) 404 nsec_printout(counter, avg); 405 else 406 abs_printout(counter, avg); 407 408 print_noise(counter, avg); 409 410 if (scaled) { 411 double avg_enabled, avg_running; 412 413 avg_enabled = avg_stats(&event_res_stats[counter][1]); 414 avg_running = avg_stats(&event_res_stats[counter][2]); 415 416 fprintf(stderr, " (scaled from %.2f%%)", 417 100 * avg_running / avg_enabled); 418 } 419 420 fprintf(stderr, "\n"); 421 } 422 423 static void print_stat(int argc, const char **argv) 424 { 425 int i, counter; 426 427 fflush(stdout); 428 429 fprintf(stderr, "\n"); 430 fprintf(stderr, " Performance counter stats for "); 431 if(target_pid == -1) { 432 fprintf(stderr, "\'%s", argv[0]); 433 for (i = 1; i < argc; i++) 434 fprintf(stderr, " %s", argv[i]); 435 }else 436 fprintf(stderr, "task pid \'%d", target_pid); 437 438 fprintf(stderr, "\'"); 439 if (run_count > 1) 440 fprintf(stderr, " (%d runs)", run_count); 441 fprintf(stderr, ":\n\n"); 442 443 for (counter = 0; counter < nr_counters; counter++) 444 print_counter(counter); 445 446 fprintf(stderr, "\n"); 447 fprintf(stderr, " %14.9f seconds time elapsed", 448 avg_stats(&walltime_nsecs_stats)/1e9); 449 if (run_count > 1) { 450 fprintf(stderr, " ( +- %7.3f%% )", 451 100*stddev_stats(&walltime_nsecs_stats) / 452 avg_stats(&walltime_nsecs_stats)); 453 } 454 fprintf(stderr, "\n\n"); 455 } 456 457 static volatile int signr = -1; 458 459 static void skip_signal(int signo) 460 { 461 if(target_pid != -1) 462 done = 1; 463 464 signr = signo; 465 } 466 467 static void sig_atexit(void) 468 { 469 if (child_pid != -1) 470 kill(child_pid, SIGTERM); 471 472 if (signr == -1) 473 return; 474 475 signal(signr, SIG_DFL); 476 kill(getpid(), signr); 477 } 478 479 static const char * const stat_usage[] = { 480 "perf stat [<options>] [<command>]", 481 NULL 482 }; 483 484 static const struct option options[] = { 485 OPT_CALLBACK('e', "event", NULL, "event", 486 "event selector. use 'perf list' to list available events", 487 parse_events), 488 OPT_BOOLEAN('i', "inherit", &inherit, 489 "child tasks inherit counters"), 490 OPT_INTEGER('p', "pid", &target_pid, 491 "stat events on existing pid"), 492 OPT_BOOLEAN('a', "all-cpus", &system_wide, 493 "system-wide collection from all CPUs"), 494 OPT_BOOLEAN('c', "scale", &scale, 495 "scale/normalize counters"), 496 OPT_BOOLEAN('v', "verbose", &verbose, 497 "be more verbose (show counter open errors, etc)"), 498 OPT_INTEGER('r', "repeat", &run_count, 499 "repeat command and print average + stddev (max: 100)"), 500 OPT_BOOLEAN('n', "null", &null_run, 501 "null run - dont start any counters"), 502 OPT_END() 503 }; 504 505 int cmd_stat(int argc, const char **argv, const char *prefix __used) 506 { 507 int status; 508 509 argc = parse_options(argc, argv, options, stat_usage, 510 PARSE_OPT_STOP_AT_NON_OPTION); 511 if (!argc && target_pid == -1) 512 usage_with_options(stat_usage, options); 513 if (run_count <= 0) 514 usage_with_options(stat_usage, options); 515 516 /* Set attrs and nr_counters if no event is selected and !null_run */ 517 if (!null_run && !nr_counters) { 518 memcpy(attrs, default_attrs, sizeof(default_attrs)); 519 nr_counters = ARRAY_SIZE(default_attrs); 520 } 521 522 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); 523 assert(nr_cpus <= MAX_NR_CPUS); 524 assert((int)nr_cpus >= 0); 525 526 /* 527 * We dont want to block the signals - that would cause 528 * child tasks to inherit that and Ctrl-C would not work. 529 * What we want is for Ctrl-C to work in the exec()-ed 530 * task, but being ignored by perf stat itself: 531 */ 532 atexit(sig_atexit); 533 signal(SIGINT, skip_signal); 534 signal(SIGALRM, skip_signal); 535 signal(SIGABRT, skip_signal); 536 537 status = 0; 538 for (run_idx = 0; run_idx < run_count; run_idx++) { 539 if (run_count != 1 && verbose) 540 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 541 status = run_perf_stat(argc, argv); 542 } 543 544 print_stat(argc, argv); 545 546 return status; 547 } 548