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 * 36 * Released under the GPL v2. (and only v2, not any later version) 37 */ 38 39 #include "perf.h" 40 #include "builtin.h" 41 #include "util/util.h" 42 #include "util/parse-options.h" 43 #include "util/parse-events.h" 44 45 #include <sys/prctl.h> 46 #include <math.h> 47 48 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = { 49 50 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 51 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES}, 52 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 53 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 54 55 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 56 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 57 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES}, 58 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES }, 59 60 }; 61 62 static int system_wide = 0; 63 static int inherit = 1; 64 static int verbose = 0; 65 66 static int fd[MAX_NR_CPUS][MAX_COUNTERS]; 67 68 static int target_pid = -1; 69 static int nr_cpus = 0; 70 static unsigned int page_size; 71 72 static int scale = 1; 73 74 static const unsigned int default_count[] = { 75 1000000, 76 1000000, 77 10000, 78 10000, 79 1000000, 80 10000, 81 }; 82 83 #define MAX_RUN 100 84 85 static int run_count = 1; 86 static int run_idx = 0; 87 88 static u64 event_res[MAX_RUN][MAX_COUNTERS][3]; 89 static u64 event_scaled[MAX_RUN][MAX_COUNTERS]; 90 91 //static u64 event_hist[MAX_RUN][MAX_COUNTERS][3]; 92 93 94 static u64 runtime_nsecs[MAX_RUN]; 95 static u64 walltime_nsecs[MAX_RUN]; 96 static u64 runtime_cycles[MAX_RUN]; 97 98 static u64 event_res_avg[MAX_COUNTERS][3]; 99 static u64 event_res_noise[MAX_COUNTERS][3]; 100 101 static u64 event_scaled_avg[MAX_COUNTERS]; 102 103 static u64 runtime_nsecs_avg; 104 static u64 runtime_nsecs_noise; 105 106 static u64 walltime_nsecs_avg; 107 static u64 walltime_nsecs_noise; 108 109 static u64 runtime_cycles_avg; 110 static u64 runtime_cycles_noise; 111 112 static void create_perf_stat_counter(int counter) 113 { 114 struct perf_counter_attr *attr = attrs + counter; 115 116 if (scale) 117 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 118 PERF_FORMAT_TOTAL_TIME_RUNNING; 119 120 if (system_wide) { 121 int cpu; 122 for (cpu = 0; cpu < nr_cpus; cpu ++) { 123 fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0); 124 if (fd[cpu][counter] < 0 && verbose) { 125 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno)); 126 } 127 } 128 } else { 129 attr->inherit = inherit; 130 attr->disabled = 1; 131 132 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0); 133 if (fd[0][counter] < 0 && verbose) { 134 printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno)); 135 } 136 } 137 } 138 139 /* 140 * Does the counter have nsecs as a unit? 141 */ 142 static inline int nsec_counter(int counter) 143 { 144 if (attrs[counter].type != PERF_TYPE_SOFTWARE) 145 return 0; 146 147 if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK) 148 return 1; 149 150 if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) 151 return 1; 152 153 return 0; 154 } 155 156 /* 157 * Read out the results of a single counter: 158 */ 159 static void read_counter(int counter) 160 { 161 u64 *count, single_count[3]; 162 ssize_t res; 163 int cpu, nv; 164 int scaled; 165 166 count = event_res[run_idx][counter]; 167 168 count[0] = count[1] = count[2] = 0; 169 170 nv = scale ? 3 : 1; 171 for (cpu = 0; cpu < nr_cpus; cpu ++) { 172 if (fd[cpu][counter] < 0) 173 continue; 174 175 res = read(fd[cpu][counter], single_count, nv * sizeof(u64)); 176 assert(res == nv * sizeof(u64)); 177 close(fd[cpu][counter]); 178 fd[cpu][counter] = -1; 179 180 count[0] += single_count[0]; 181 if (scale) { 182 count[1] += single_count[1]; 183 count[2] += single_count[2]; 184 } 185 } 186 187 scaled = 0; 188 if (scale) { 189 if (count[2] == 0) { 190 event_scaled[run_idx][counter] = -1; 191 count[0] = 0; 192 return; 193 } 194 195 if (count[2] < count[1]) { 196 event_scaled[run_idx][counter] = 1; 197 count[0] = (unsigned long long) 198 ((double)count[0] * count[1] / count[2] + 0.5); 199 } 200 } 201 /* 202 * Save the full runtime - to allow normalization during printout: 203 */ 204 if (attrs[counter].type == PERF_TYPE_SOFTWARE && 205 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) 206 runtime_nsecs[run_idx] = count[0]; 207 if (attrs[counter].type == PERF_TYPE_HARDWARE && 208 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES) 209 runtime_cycles[run_idx] = count[0]; 210 } 211 212 static int run_perf_stat(int argc, const char **argv) 213 { 214 unsigned long long t0, t1; 215 int status = 0; 216 int counter; 217 int pid; 218 219 if (!system_wide) 220 nr_cpus = 1; 221 222 for (counter = 0; counter < nr_counters; counter++) 223 create_perf_stat_counter(counter); 224 225 /* 226 * Enable counters and exec the command: 227 */ 228 t0 = rdclock(); 229 prctl(PR_TASK_PERF_COUNTERS_ENABLE); 230 231 if ((pid = fork()) < 0) 232 perror("failed to fork"); 233 234 if (!pid) { 235 if (execvp(argv[0], (char **)argv)) { 236 perror(argv[0]); 237 exit(-1); 238 } 239 } 240 241 wait(&status); 242 243 prctl(PR_TASK_PERF_COUNTERS_DISABLE); 244 t1 = rdclock(); 245 246 walltime_nsecs[run_idx] = t1 - t0; 247 248 for (counter = 0; counter < nr_counters; counter++) 249 read_counter(counter); 250 251 return WEXITSTATUS(status); 252 } 253 254 static void print_noise(u64 *count, u64 *noise) 255 { 256 if (run_count > 1) 257 fprintf(stderr, " ( +- %7.3f%% )", 258 (double)noise[0]/(count[0]+1)*100.0); 259 } 260 261 static void nsec_printout(int counter, u64 *count, u64 *noise) 262 { 263 double msecs = (double)count[0] / 1000000; 264 265 fprintf(stderr, " %14.6f %-20s", msecs, event_name(counter)); 266 267 if (attrs[counter].type == PERF_TYPE_SOFTWARE && 268 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) { 269 270 if (walltime_nsecs_avg) 271 fprintf(stderr, " # %10.3f CPUs ", 272 (double)count[0] / (double)walltime_nsecs_avg); 273 } 274 print_noise(count, noise); 275 } 276 277 static void abs_printout(int counter, u64 *count, u64 *noise) 278 { 279 fprintf(stderr, " %14Ld %-20s", count[0], event_name(counter)); 280 281 if (runtime_cycles_avg && 282 attrs[counter].type == PERF_TYPE_HARDWARE && 283 attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) { 284 285 fprintf(stderr, " # %10.3f IPC ", 286 (double)count[0] / (double)runtime_cycles_avg); 287 } else { 288 if (runtime_nsecs_avg) { 289 fprintf(stderr, " # %10.3f M/sec", 290 (double)count[0]/runtime_nsecs_avg*1000.0); 291 } 292 } 293 print_noise(count, noise); 294 } 295 296 /* 297 * Print out the results of a single counter: 298 */ 299 static void print_counter(int counter) 300 { 301 u64 *count, *noise; 302 int scaled; 303 304 count = event_res_avg[counter]; 305 noise = event_res_noise[counter]; 306 scaled = event_scaled_avg[counter]; 307 308 if (scaled == -1) { 309 fprintf(stderr, " %14s %-20s\n", 310 "<not counted>", event_name(counter)); 311 return; 312 } 313 314 if (nsec_counter(counter)) 315 nsec_printout(counter, count, noise); 316 else 317 abs_printout(counter, count, noise); 318 319 if (scaled) 320 fprintf(stderr, " (scaled from %.2f%%)", 321 (double) count[2] / count[1] * 100); 322 323 fprintf(stderr, "\n"); 324 } 325 326 /* 327 * normalize_noise noise values down to stddev: 328 */ 329 static void normalize_noise(u64 *val) 330 { 331 double res; 332 333 res = (double)*val / (run_count * sqrt((double)run_count)); 334 335 *val = (u64)res; 336 } 337 338 static void update_avg(const char *name, int idx, u64 *avg, u64 *val) 339 { 340 *avg += *val; 341 342 if (verbose > 1) 343 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val); 344 } 345 /* 346 * Calculate the averages and noises: 347 */ 348 static void calc_avg(void) 349 { 350 int i, j; 351 352 if (verbose > 1) 353 fprintf(stderr, "\n"); 354 355 for (i = 0; i < run_count; i++) { 356 update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i); 357 update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i); 358 update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i); 359 360 for (j = 0; j < nr_counters; j++) { 361 update_avg("counter/0", j, 362 event_res_avg[j]+0, event_res[i][j]+0); 363 update_avg("counter/1", j, 364 event_res_avg[j]+1, event_res[i][j]+1); 365 update_avg("counter/2", j, 366 event_res_avg[j]+2, event_res[i][j]+2); 367 update_avg("scaled", j, 368 event_scaled_avg + j, event_scaled[i]+j); 369 } 370 } 371 runtime_nsecs_avg /= run_count; 372 walltime_nsecs_avg /= run_count; 373 runtime_cycles_avg /= run_count; 374 375 for (j = 0; j < nr_counters; j++) { 376 event_res_avg[j][0] /= run_count; 377 event_res_avg[j][1] /= run_count; 378 event_res_avg[j][2] /= run_count; 379 } 380 381 for (i = 0; i < run_count; i++) { 382 runtime_nsecs_noise += 383 abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg)); 384 walltime_nsecs_noise += 385 abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg)); 386 runtime_cycles_noise += 387 abs((s64)(runtime_cycles[i] - runtime_cycles_avg)); 388 389 for (j = 0; j < nr_counters; j++) { 390 event_res_noise[j][0] += 391 abs((s64)(event_res[i][j][0] - event_res_avg[j][0])); 392 event_res_noise[j][1] += 393 abs((s64)(event_res[i][j][1] - event_res_avg[j][1])); 394 event_res_noise[j][2] += 395 abs((s64)(event_res[i][j][2] - event_res_avg[j][2])); 396 } 397 } 398 399 normalize_noise(&runtime_nsecs_noise); 400 normalize_noise(&walltime_nsecs_noise); 401 normalize_noise(&runtime_cycles_noise); 402 403 for (j = 0; j < nr_counters; j++) { 404 normalize_noise(&event_res_noise[j][0]); 405 normalize_noise(&event_res_noise[j][1]); 406 normalize_noise(&event_res_noise[j][2]); 407 } 408 } 409 410 static void print_stat(int argc, const char **argv) 411 { 412 int i, counter; 413 414 calc_avg(); 415 416 fflush(stdout); 417 418 fprintf(stderr, "\n"); 419 fprintf(stderr, " Performance counter stats for \'%s", argv[0]); 420 421 for (i = 1; i < argc; i++) 422 fprintf(stderr, " %s", argv[i]); 423 424 fprintf(stderr, "\'"); 425 if (run_count > 1) 426 fprintf(stderr, " (%d runs)", run_count); 427 fprintf(stderr, ":\n\n"); 428 429 for (counter = 0; counter < nr_counters; counter++) 430 print_counter(counter); 431 432 433 fprintf(stderr, "\n"); 434 fprintf(stderr, " %14.9f seconds time elapsed.\n", 435 (double)walltime_nsecs_avg/1e9); 436 fprintf(stderr, "\n"); 437 } 438 439 static volatile int signr = -1; 440 441 static void skip_signal(int signo) 442 { 443 signr = signo; 444 } 445 446 static void sig_atexit(void) 447 { 448 if (signr == -1) 449 return; 450 451 signal(signr, SIG_DFL); 452 kill(getpid(), signr); 453 } 454 455 static const char * const stat_usage[] = { 456 "perf stat [<options>] <command>", 457 NULL 458 }; 459 460 static const struct option options[] = { 461 OPT_CALLBACK('e', "event", NULL, "event", 462 "event selector. use 'perf list' to list available events", 463 parse_events), 464 OPT_BOOLEAN('i', "inherit", &inherit, 465 "child tasks inherit counters"), 466 OPT_INTEGER('p', "pid", &target_pid, 467 "stat events on existing pid"), 468 OPT_BOOLEAN('a', "all-cpus", &system_wide, 469 "system-wide collection from all CPUs"), 470 OPT_BOOLEAN('S', "scale", &scale, 471 "scale/normalize counters"), 472 OPT_BOOLEAN('v', "verbose", &verbose, 473 "be more verbose (show counter open errors, etc)"), 474 OPT_INTEGER('r', "repeat", &run_count, 475 "repeat command and print average + stddev (max: 100)"), 476 OPT_END() 477 }; 478 479 int cmd_stat(int argc, const char **argv, const char *prefix) 480 { 481 int status; 482 483 page_size = sysconf(_SC_PAGE_SIZE); 484 485 memcpy(attrs, default_attrs, sizeof(attrs)); 486 487 argc = parse_options(argc, argv, options, stat_usage, 0); 488 if (!argc) 489 usage_with_options(stat_usage, options); 490 if (run_count <= 0 || run_count > MAX_RUN) 491 usage_with_options(stat_usage, options); 492 493 if (!nr_counters) 494 nr_counters = 8; 495 496 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN); 497 assert(nr_cpus <= MAX_NR_CPUS); 498 assert(nr_cpus >= 0); 499 500 /* 501 * We dont want to block the signals - that would cause 502 * child tasks to inherit that and Ctrl-C would not work. 503 * What we want is for Ctrl-C to work in the exec()-ed 504 * task, but being ignored by perf stat itself: 505 */ 506 atexit(sig_atexit); 507 signal(SIGINT, skip_signal); 508 signal(SIGALRM, skip_signal); 509 signal(SIGABRT, skip_signal); 510 511 status = 0; 512 for (run_idx = 0; run_idx < run_count; run_idx++) { 513 if (run_count != 1 && verbose) 514 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx+1); 515 status = run_perf_stat(argc, argv); 516 } 517 518 print_stat(argc, argv); 519 520 return status; 521 } 522