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 #include "util/cpumap.h" 49 50 #include <sys/prctl.h> 51 #include <math.h> 52 53 static struct perf_event_attr default_attrs[] = { 54 55 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 56 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 57 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 58 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 59 60 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 61 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 62 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 63 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 64 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES }, 65 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, 66 67 }; 68 69 static int system_wide = 0; 70 static unsigned int nr_cpus = 0; 71 static int run_idx = 0; 72 73 static int run_count = 1; 74 static int inherit = 1; 75 static int scale = 1; 76 static pid_t target_pid = -1; 77 static pid_t child_pid = -1; 78 static int null_run = 0; 79 80 static int fd[MAX_NR_CPUS][MAX_COUNTERS]; 81 82 static int event_scaled[MAX_COUNTERS]; 83 84 static volatile int done = 0; 85 86 struct stats 87 { 88 double n, mean, M2; 89 }; 90 91 static void update_stats(struct stats *stats, u64 val) 92 { 93 double delta; 94 95 stats->n++; 96 delta = val - stats->mean; 97 stats->mean += delta / stats->n; 98 stats->M2 += delta*(val - stats->mean); 99 } 100 101 static double avg_stats(struct stats *stats) 102 { 103 return stats->mean; 104 } 105 106 /* 107 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 108 * 109 * (\Sum n_i^2) - ((\Sum n_i)^2)/n 110 * s^2 = ------------------------------- 111 * n - 1 112 * 113 * http://en.wikipedia.org/wiki/Stddev 114 * 115 * The std dev of the mean is related to the std dev by: 116 * 117 * s 118 * s_mean = ------- 119 * sqrt(n) 120 * 121 */ 122 static double stddev_stats(struct stats *stats) 123 { 124 double variance = stats->M2 / (stats->n - 1); 125 double variance_mean = variance / stats->n; 126 127 return sqrt(variance_mean); 128 } 129 130 struct stats event_res_stats[MAX_COUNTERS][3]; 131 struct stats runtime_nsecs_stats; 132 struct stats walltime_nsecs_stats; 133 struct stats runtime_cycles_stats; 134 struct stats runtime_branches_stats; 135 136 #define MATCH_EVENT(t, c, counter) \ 137 (attrs[counter].type == PERF_TYPE_##t && \ 138 attrs[counter].config == PERF_COUNT_##c) 139 140 #define ERR_PERF_OPEN \ 141 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n" 142 143 static void create_perf_stat_counter(int counter, int pid) 144 { 145 struct perf_event_attr *attr = attrs + counter; 146 147 if (scale) 148 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 149 PERF_FORMAT_TOTAL_TIME_RUNNING; 150 151 if (system_wide) { 152 unsigned int cpu; 153 154 for (cpu = 0; cpu < nr_cpus; cpu++) { 155 fd[cpu][counter] = sys_perf_event_open(attr, -1, cpumap[cpu], -1, 0); 156 if (fd[cpu][counter] < 0 && verbose) 157 fprintf(stderr, ERR_PERF_OPEN, counter, 158 fd[cpu][counter], strerror(errno)); 159 } 160 } else { 161 attr->inherit = inherit; 162 attr->disabled = 1; 163 attr->enable_on_exec = 1; 164 165 fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0); 166 if (fd[0][counter] < 0 && verbose) 167 fprintf(stderr, ERR_PERF_OPEN, counter, 168 fd[0][counter], strerror(errno)); 169 } 170 } 171 172 /* 173 * Does the counter have nsecs as a unit? 174 */ 175 static inline int nsec_counter(int counter) 176 { 177 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) || 178 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 179 return 1; 180 181 return 0; 182 } 183 184 /* 185 * Read out the results of a single counter: 186 */ 187 static void read_counter(int counter) 188 { 189 u64 count[3], single_count[3]; 190 unsigned int cpu; 191 size_t res, nv; 192 int scaled; 193 int i; 194 195 count[0] = count[1] = count[2] = 0; 196 197 nv = scale ? 3 : 1; 198 for (cpu = 0; cpu < nr_cpus; cpu++) { 199 if (fd[cpu][counter] < 0) 200 continue; 201 202 res = read(fd[cpu][counter], single_count, nv * sizeof(u64)); 203 assert(res == nv * sizeof(u64)); 204 205 close(fd[cpu][counter]); 206 fd[cpu][counter] = -1; 207 208 count[0] += single_count[0]; 209 if (scale) { 210 count[1] += single_count[1]; 211 count[2] += single_count[2]; 212 } 213 } 214 215 scaled = 0; 216 if (scale) { 217 if (count[2] == 0) { 218 event_scaled[counter] = -1; 219 count[0] = 0; 220 return; 221 } 222 223 if (count[2] < count[1]) { 224 event_scaled[counter] = 1; 225 count[0] = (unsigned long long) 226 ((double)count[0] * count[1] / count[2] + 0.5); 227 } 228 } 229 230 for (i = 0; i < 3; i++) 231 update_stats(&event_res_stats[counter][i], count[i]); 232 233 if (verbose) { 234 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter), 235 count[0], count[1], count[2]); 236 } 237 238 /* 239 * Save the full runtime - to allow normalization during printout: 240 */ 241 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) 242 update_stats(&runtime_nsecs_stats, count[0]); 243 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter)) 244 update_stats(&runtime_cycles_stats, count[0]); 245 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter)) 246 update_stats(&runtime_branches_stats, count[0]); 247 } 248 249 static int run_perf_stat(int argc __used, const char **argv) 250 { 251 unsigned long long t0, t1; 252 int status = 0; 253 int counter; 254 int pid = target_pid; 255 int child_ready_pipe[2], go_pipe[2]; 256 const bool forks = (target_pid == -1 && argc > 0); 257 char buf; 258 259 if (!system_wide) 260 nr_cpus = 1; 261 262 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { 263 perror("failed to create pipes"); 264 exit(1); 265 } 266 267 if (forks) { 268 if ((pid = fork()) < 0) 269 perror("failed to fork"); 270 271 if (!pid) { 272 close(child_ready_pipe[0]); 273 close(go_pipe[1]); 274 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 275 276 /* 277 * Do a dummy execvp to get the PLT entry resolved, 278 * so we avoid the resolver overhead on the real 279 * execvp call. 280 */ 281 execvp("", (char **)argv); 282 283 /* 284 * Tell the parent we're ready to go 285 */ 286 close(child_ready_pipe[1]); 287 288 /* 289 * Wait until the parent tells us to go. 290 */ 291 if (read(go_pipe[0], &buf, 1) == -1) 292 perror("unable to read pipe"); 293 294 execvp(argv[0], (char **)argv); 295 296 perror(argv[0]); 297 exit(-1); 298 } 299 300 child_pid = pid; 301 302 /* 303 * Wait for the child to be ready to exec. 304 */ 305 close(child_ready_pipe[1]); 306 close(go_pipe[0]); 307 if (read(child_ready_pipe[0], &buf, 1) == -1) 308 perror("unable to read pipe"); 309 close(child_ready_pipe[0]); 310 } 311 312 for (counter = 0; counter < nr_counters; counter++) 313 create_perf_stat_counter(counter, pid); 314 315 /* 316 * Enable counters and exec the command: 317 */ 318 t0 = rdclock(); 319 320 if (forks) { 321 close(go_pipe[1]); 322 wait(&status); 323 } else { 324 while(!done); 325 } 326 327 t1 = rdclock(); 328 329 update_stats(&walltime_nsecs_stats, t1 - t0); 330 331 for (counter = 0; counter < nr_counters; counter++) 332 read_counter(counter); 333 334 return WEXITSTATUS(status); 335 } 336 337 static void print_noise(int counter, double avg) 338 { 339 if (run_count == 1) 340 return; 341 342 fprintf(stderr, " ( +- %7.3f%% )", 343 100 * stddev_stats(&event_res_stats[counter][0]) / avg); 344 } 345 346 static void nsec_printout(int counter, double avg) 347 { 348 double msecs = avg / 1e6; 349 350 fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter)); 351 352 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) { 353 fprintf(stderr, " # %10.3f CPUs ", 354 avg / avg_stats(&walltime_nsecs_stats)); 355 } 356 } 357 358 static void abs_printout(int counter, double avg) 359 { 360 double total, ratio = 0.0; 361 362 fprintf(stderr, " %14.0f %-24s", avg, event_name(counter)); 363 364 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) { 365 total = avg_stats(&runtime_cycles_stats); 366 367 if (total) 368 ratio = avg / total; 369 370 fprintf(stderr, " # %10.3f IPC ", ratio); 371 } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) && 372 runtime_branches_stats.n != 0) { 373 total = avg_stats(&runtime_branches_stats); 374 375 if (total) 376 ratio = avg * 100 / total; 377 378 fprintf(stderr, " # %10.3f %% ", ratio); 379 380 } else if (runtime_nsecs_stats.n != 0) { 381 total = avg_stats(&runtime_nsecs_stats); 382 383 if (total) 384 ratio = 1000.0 * avg / total; 385 386 fprintf(stderr, " # %10.3f M/sec", ratio); 387 } 388 } 389 390 /* 391 * Print out the results of a single counter: 392 */ 393 static void print_counter(int counter) 394 { 395 double avg = avg_stats(&event_res_stats[counter][0]); 396 int scaled = event_scaled[counter]; 397 398 if (scaled == -1) { 399 fprintf(stderr, " %14s %-24s\n", 400 "<not counted>", event_name(counter)); 401 return; 402 } 403 404 if (nsec_counter(counter)) 405 nsec_printout(counter, avg); 406 else 407 abs_printout(counter, avg); 408 409 print_noise(counter, avg); 410 411 if (scaled) { 412 double avg_enabled, avg_running; 413 414 avg_enabled = avg_stats(&event_res_stats[counter][1]); 415 avg_running = avg_stats(&event_res_stats[counter][2]); 416 417 fprintf(stderr, " (scaled from %.2f%%)", 418 100 * avg_running / avg_enabled); 419 } 420 421 fprintf(stderr, "\n"); 422 } 423 424 static void print_stat(int argc, const char **argv) 425 { 426 int i, counter; 427 428 fflush(stdout); 429 430 fprintf(stderr, "\n"); 431 fprintf(stderr, " Performance counter stats for "); 432 if(target_pid == -1) { 433 fprintf(stderr, "\'%s", argv[0]); 434 for (i = 1; i < argc; i++) 435 fprintf(stderr, " %s", argv[i]); 436 }else 437 fprintf(stderr, "task pid \'%d", target_pid); 438 439 fprintf(stderr, "\'"); 440 if (run_count > 1) 441 fprintf(stderr, " (%d runs)", run_count); 442 fprintf(stderr, ":\n\n"); 443 444 for (counter = 0; counter < nr_counters; counter++) 445 print_counter(counter); 446 447 fprintf(stderr, "\n"); 448 fprintf(stderr, " %14.9f seconds time elapsed", 449 avg_stats(&walltime_nsecs_stats)/1e9); 450 if (run_count > 1) { 451 fprintf(stderr, " ( +- %7.3f%% )", 452 100*stddev_stats(&walltime_nsecs_stats) / 453 avg_stats(&walltime_nsecs_stats)); 454 } 455 fprintf(stderr, "\n\n"); 456 } 457 458 static volatile int signr = -1; 459 460 static void skip_signal(int signo) 461 { 462 if(target_pid != -1) 463 done = 1; 464 465 signr = signo; 466 } 467 468 static void sig_atexit(void) 469 { 470 if (child_pid != -1) 471 kill(child_pid, SIGTERM); 472 473 if (signr == -1) 474 return; 475 476 signal(signr, SIG_DFL); 477 kill(getpid(), signr); 478 } 479 480 static const char * const stat_usage[] = { 481 "perf stat [<options>] [<command>]", 482 NULL 483 }; 484 485 static const struct option options[] = { 486 OPT_CALLBACK('e', "event", NULL, "event", 487 "event selector. use 'perf list' to list available events", 488 parse_events), 489 OPT_BOOLEAN('i', "inherit", &inherit, 490 "child tasks inherit counters"), 491 OPT_INTEGER('p', "pid", &target_pid, 492 "stat events on existing pid"), 493 OPT_BOOLEAN('a', "all-cpus", &system_wide, 494 "system-wide collection from all CPUs"), 495 OPT_BOOLEAN('c', "scale", &scale, 496 "scale/normalize counters"), 497 OPT_BOOLEAN('v', "verbose", &verbose, 498 "be more verbose (show counter open errors, etc)"), 499 OPT_INTEGER('r', "repeat", &run_count, 500 "repeat command and print average + stddev (max: 100)"), 501 OPT_BOOLEAN('n', "null", &null_run, 502 "null run - dont start any counters"), 503 OPT_END() 504 }; 505 506 int cmd_stat(int argc, const char **argv, const char *prefix __used) 507 { 508 int status; 509 510 argc = parse_options(argc, argv, options, stat_usage, 511 PARSE_OPT_STOP_AT_NON_OPTION); 512 if (!argc && target_pid == -1) 513 usage_with_options(stat_usage, options); 514 if (run_count <= 0) 515 usage_with_options(stat_usage, options); 516 517 /* Set attrs and nr_counters if no event is selected and !null_run */ 518 if (!null_run && !nr_counters) { 519 memcpy(attrs, default_attrs, sizeof(default_attrs)); 520 nr_counters = ARRAY_SIZE(default_attrs); 521 } 522 523 if (system_wide) 524 nr_cpus = read_cpu_map(); 525 else 526 nr_cpus = 1; 527 528 /* 529 * We dont want to block the signals - that would cause 530 * child tasks to inherit that and Ctrl-C would not work. 531 * What we want is for Ctrl-C to work in the exec()-ed 532 * task, but being ignored by perf stat itself: 533 */ 534 atexit(sig_atexit); 535 signal(SIGINT, skip_signal); 536 signal(SIGALRM, skip_signal); 537 signal(SIGABRT, skip_signal); 538 539 status = 0; 540 for (run_idx = 0; run_idx < run_count; run_idx++) { 541 if (run_count != 1 && verbose) 542 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 543 status = run_perf_stat(argc, argv); 544 } 545 546 print_stat(argc, argv); 547 548 return status; 549 } 550