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