1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * builtin-stat.c 4 * 5 * Builtin stat command: Give a precise performance counters summary 6 * overview about any workload, CPU or specific PID. 7 * 8 * Sample output: 9 10 $ perf stat ./hackbench 10 11 12 Time: 0.118 13 14 Performance counter stats for './hackbench 10': 15 16 1708.761321 task-clock # 11.037 CPUs utilized 17 41,190 context-switches # 0.024 M/sec 18 6,735 CPU-migrations # 0.004 M/sec 19 17,318 page-faults # 0.010 M/sec 20 5,205,202,243 cycles # 3.046 GHz 21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle 22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle 23 2,603,501,247 instructions # 0.50 insns per cycle 24 # 1.48 stalled cycles per insn 25 484,357,498 branches # 283.455 M/sec 26 6,388,934 branch-misses # 1.32% of all branches 27 28 0.154822978 seconds time elapsed 29 30 * 31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 32 * 33 * Improvements and fixes by: 34 * 35 * Arjan van de Ven <arjan@linux.intel.com> 36 * Yanmin Zhang <yanmin.zhang@intel.com> 37 * Wu Fengguang <fengguang.wu@intel.com> 38 * Mike Galbraith <efault@gmx.de> 39 * Paul Mackerras <paulus@samba.org> 40 * Jaswinder Singh Rajput <jaswinder@kernel.org> 41 */ 42 43 #include "perf.h" 44 #include "builtin.h" 45 #include "util/cgroup.h" 46 #include <subcmd/parse-options.h> 47 #include "util/parse-events.h" 48 #include "util/pmu.h" 49 #include "util/event.h" 50 #include "util/evlist.h" 51 #include "util/evsel.h" 52 #include "util/debug.h" 53 #include "util/color.h" 54 #include "util/stat.h" 55 #include "util/header.h" 56 #include "util/cpumap.h" 57 #include "util/thread.h" 58 #include "util/thread_map.h" 59 #include "util/counts.h" 60 #include "util/group.h" 61 #include "util/session.h" 62 #include "util/tool.h" 63 #include "util/string2.h" 64 #include "util/metricgroup.h" 65 #include "util/top.h" 66 #include "asm/bug.h" 67 68 #include <linux/time64.h> 69 #include <linux/zalloc.h> 70 #include <api/fs/fs.h> 71 #include <errno.h> 72 #include <signal.h> 73 #include <stdlib.h> 74 #include <sys/prctl.h> 75 #include <inttypes.h> 76 #include <locale.h> 77 #include <math.h> 78 #include <sys/types.h> 79 #include <sys/stat.h> 80 #include <sys/wait.h> 81 #include <unistd.h> 82 #include <sys/time.h> 83 #include <sys/resource.h> 84 85 #include <linux/ctype.h> 86 #include <perf/evlist.h> 87 88 #define DEFAULT_SEPARATOR " " 89 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi" 90 91 static void print_counters(struct timespec *ts, int argc, const char **argv); 92 93 /* Default events used for perf stat -T */ 94 static const char *transaction_attrs = { 95 "task-clock," 96 "{" 97 "instructions," 98 "cycles," 99 "cpu/cycles-t/," 100 "cpu/tx-start/," 101 "cpu/el-start/," 102 "cpu/cycles-ct/" 103 "}" 104 }; 105 106 /* More limited version when the CPU does not have all events. */ 107 static const char * transaction_limited_attrs = { 108 "task-clock," 109 "{" 110 "instructions," 111 "cycles," 112 "cpu/cycles-t/," 113 "cpu/tx-start/" 114 "}" 115 }; 116 117 static const char * topdown_attrs[] = { 118 "topdown-total-slots", 119 "topdown-slots-retired", 120 "topdown-recovery-bubbles", 121 "topdown-fetch-bubbles", 122 "topdown-slots-issued", 123 NULL, 124 }; 125 126 static const char *smi_cost_attrs = { 127 "{" 128 "msr/aperf/," 129 "msr/smi/," 130 "cycles" 131 "}" 132 }; 133 134 static struct evlist *evsel_list; 135 136 static struct target target = { 137 .uid = UINT_MAX, 138 }; 139 140 #define METRIC_ONLY_LEN 20 141 142 static volatile pid_t child_pid = -1; 143 static int detailed_run = 0; 144 static bool transaction_run; 145 static bool topdown_run = false; 146 static bool smi_cost = false; 147 static bool smi_reset = false; 148 static int big_num_opt = -1; 149 static bool group = false; 150 static const char *pre_cmd = NULL; 151 static const char *post_cmd = NULL; 152 static bool sync_run = false; 153 static bool forever = false; 154 static bool force_metric_only = false; 155 static struct timespec ref_time; 156 static bool append_file; 157 static bool interval_count; 158 static const char *output_name; 159 static int output_fd; 160 161 struct perf_stat { 162 bool record; 163 struct perf_data data; 164 struct perf_session *session; 165 u64 bytes_written; 166 struct perf_tool tool; 167 bool maps_allocated; 168 struct perf_cpu_map *cpus; 169 struct perf_thread_map *threads; 170 enum aggr_mode aggr_mode; 171 }; 172 173 static struct perf_stat perf_stat; 174 #define STAT_RECORD perf_stat.record 175 176 static volatile int done = 0; 177 178 static struct perf_stat_config stat_config = { 179 .aggr_mode = AGGR_GLOBAL, 180 .scale = true, 181 .unit_width = 4, /* strlen("unit") */ 182 .run_count = 1, 183 .metric_only_len = METRIC_ONLY_LEN, 184 .walltime_nsecs_stats = &walltime_nsecs_stats, 185 .big_num = true, 186 }; 187 188 static inline void diff_timespec(struct timespec *r, struct timespec *a, 189 struct timespec *b) 190 { 191 r->tv_sec = a->tv_sec - b->tv_sec; 192 if (a->tv_nsec < b->tv_nsec) { 193 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec; 194 r->tv_sec--; 195 } else { 196 r->tv_nsec = a->tv_nsec - b->tv_nsec ; 197 } 198 } 199 200 static void perf_stat__reset_stats(void) 201 { 202 int i; 203 204 perf_evlist__reset_stats(evsel_list); 205 perf_stat__reset_shadow_stats(); 206 207 for (i = 0; i < stat_config.stats_num; i++) 208 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]); 209 } 210 211 static int process_synthesized_event(struct perf_tool *tool __maybe_unused, 212 union perf_event *event, 213 struct perf_sample *sample __maybe_unused, 214 struct machine *machine __maybe_unused) 215 { 216 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) { 217 pr_err("failed to write perf data, error: %m\n"); 218 return -1; 219 } 220 221 perf_stat.bytes_written += event->header.size; 222 return 0; 223 } 224 225 static int write_stat_round_event(u64 tm, u64 type) 226 { 227 return perf_event__synthesize_stat_round(NULL, tm, type, 228 process_synthesized_event, 229 NULL); 230 } 231 232 #define WRITE_STAT_ROUND_EVENT(time, interval) \ 233 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval) 234 235 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) 236 237 static int 238 perf_evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread, 239 struct perf_counts_values *count) 240 { 241 struct perf_sample_id *sid = SID(counter, cpu, thread); 242 243 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count, 244 process_synthesized_event, NULL); 245 } 246 247 static int read_single_counter(struct evsel *counter, int cpu, 248 int thread, struct timespec *rs) 249 { 250 if (counter->tool_event == PERF_TOOL_DURATION_TIME) { 251 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL; 252 struct perf_counts_values *count = 253 perf_counts(counter->counts, cpu, thread); 254 count->ena = count->run = val; 255 count->val = val; 256 return 0; 257 } 258 return perf_evsel__read_counter(counter, cpu, thread); 259 } 260 261 /* 262 * Read out the results of a single counter: 263 * do not aggregate counts across CPUs in system-wide mode 264 */ 265 static int read_counter(struct evsel *counter, struct timespec *rs) 266 { 267 int nthreads = thread_map__nr(evsel_list->core.threads); 268 int ncpus, cpu, thread; 269 270 if (target__has_cpu(&target) && !target__has_per_thread(&target)) 271 ncpus = perf_evsel__nr_cpus(counter); 272 else 273 ncpus = 1; 274 275 if (!counter->supported) 276 return -ENOENT; 277 278 if (counter->system_wide) 279 nthreads = 1; 280 281 for (thread = 0; thread < nthreads; thread++) { 282 for (cpu = 0; cpu < ncpus; cpu++) { 283 struct perf_counts_values *count; 284 285 count = perf_counts(counter->counts, cpu, thread); 286 287 /* 288 * The leader's group read loads data into its group members 289 * (via perf_evsel__read_counter) and sets threir count->loaded. 290 */ 291 if (!perf_counts__is_loaded(counter->counts, cpu, thread) && 292 read_single_counter(counter, cpu, thread, rs)) { 293 counter->counts->scaled = -1; 294 perf_counts(counter->counts, cpu, thread)->ena = 0; 295 perf_counts(counter->counts, cpu, thread)->run = 0; 296 return -1; 297 } 298 299 perf_counts__set_loaded(counter->counts, cpu, thread, false); 300 301 if (STAT_RECORD) { 302 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) { 303 pr_err("failed to write stat event\n"); 304 return -1; 305 } 306 } 307 308 if (verbose > 1) { 309 fprintf(stat_config.output, 310 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", 311 perf_evsel__name(counter), 312 cpu, 313 count->val, count->ena, count->run); 314 } 315 } 316 } 317 318 return 0; 319 } 320 321 static void read_counters(struct timespec *rs) 322 { 323 struct evsel *counter; 324 int ret; 325 326 evlist__for_each_entry(evsel_list, counter) { 327 ret = read_counter(counter, rs); 328 if (ret) 329 pr_debug("failed to read counter %s\n", counter->name); 330 331 if (ret == 0 && perf_stat_process_counter(&stat_config, counter)) 332 pr_warning("failed to process counter %s\n", counter->name); 333 } 334 } 335 336 static void process_interval(void) 337 { 338 struct timespec ts, rs; 339 340 clock_gettime(CLOCK_MONOTONIC, &ts); 341 diff_timespec(&rs, &ts, &ref_time); 342 343 read_counters(&rs); 344 345 if (STAT_RECORD) { 346 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL)) 347 pr_err("failed to write stat round event\n"); 348 } 349 350 init_stats(&walltime_nsecs_stats); 351 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000); 352 print_counters(&rs, 0, NULL); 353 } 354 355 static void enable_counters(void) 356 { 357 if (stat_config.initial_delay) 358 usleep(stat_config.initial_delay * USEC_PER_MSEC); 359 360 /* 361 * We need to enable counters only if: 362 * - we don't have tracee (attaching to task or cpu) 363 * - we have initial delay configured 364 */ 365 if (!target__none(&target) || stat_config.initial_delay) 366 evlist__enable(evsel_list); 367 } 368 369 static void disable_counters(void) 370 { 371 /* 372 * If we don't have tracee (attaching to task or cpu), counters may 373 * still be running. To get accurate group ratios, we must stop groups 374 * from counting before reading their constituent counters. 375 */ 376 if (!target__none(&target)) 377 evlist__disable(evsel_list); 378 } 379 380 static volatile int workload_exec_errno; 381 382 /* 383 * perf_evlist__prepare_workload will send a SIGUSR1 384 * if the fork fails, since we asked by setting its 385 * want_signal to true. 386 */ 387 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info, 388 void *ucontext __maybe_unused) 389 { 390 workload_exec_errno = info->si_value.sival_int; 391 } 392 393 static bool perf_evsel__should_store_id(struct evsel *counter) 394 { 395 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID; 396 } 397 398 static bool is_target_alive(struct target *_target, 399 struct perf_thread_map *threads) 400 { 401 struct stat st; 402 int i; 403 404 if (!target__has_task(_target)) 405 return true; 406 407 for (i = 0; i < threads->nr; i++) { 408 char path[PATH_MAX]; 409 410 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(), 411 threads->map[i].pid); 412 413 if (!stat(path, &st)) 414 return true; 415 } 416 417 return false; 418 } 419 420 static int __run_perf_stat(int argc, const char **argv, int run_idx) 421 { 422 int interval = stat_config.interval; 423 int times = stat_config.times; 424 int timeout = stat_config.timeout; 425 char msg[BUFSIZ]; 426 unsigned long long t0, t1; 427 struct evsel *counter; 428 struct timespec ts; 429 size_t l; 430 int status = 0; 431 const bool forks = (argc > 0); 432 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false; 433 434 if (interval) { 435 ts.tv_sec = interval / USEC_PER_MSEC; 436 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC; 437 } else if (timeout) { 438 ts.tv_sec = timeout / USEC_PER_MSEC; 439 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC; 440 } else { 441 ts.tv_sec = 1; 442 ts.tv_nsec = 0; 443 } 444 445 if (forks) { 446 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe, 447 workload_exec_failed_signal) < 0) { 448 perror("failed to prepare workload"); 449 return -1; 450 } 451 child_pid = evsel_list->workload.pid; 452 } 453 454 if (group) 455 perf_evlist__set_leader(evsel_list); 456 457 evlist__for_each_entry(evsel_list, counter) { 458 try_again: 459 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) { 460 461 /* Weak group failed. Reset the group. */ 462 if ((errno == EINVAL || errno == EBADF) && 463 counter->leader != counter && 464 counter->weak_group) { 465 counter = perf_evlist__reset_weak_group(evsel_list, counter); 466 goto try_again; 467 } 468 469 /* 470 * PPC returns ENXIO for HW counters until 2.6.37 471 * (behavior changed with commit b0a873e). 472 */ 473 if (errno == EINVAL || errno == ENOSYS || 474 errno == ENOENT || errno == EOPNOTSUPP || 475 errno == ENXIO) { 476 if (verbose > 0) 477 ui__warning("%s event is not supported by the kernel.\n", 478 perf_evsel__name(counter)); 479 counter->supported = false; 480 481 if ((counter->leader != counter) || 482 !(counter->leader->core.nr_members > 1)) 483 continue; 484 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) { 485 if (verbose > 0) 486 ui__warning("%s\n", msg); 487 goto try_again; 488 } else if (target__has_per_thread(&target) && 489 evsel_list->core.threads && 490 evsel_list->core.threads->err_thread != -1) { 491 /* 492 * For global --per-thread case, skip current 493 * error thread. 494 */ 495 if (!thread_map__remove(evsel_list->core.threads, 496 evsel_list->core.threads->err_thread)) { 497 evsel_list->core.threads->err_thread = -1; 498 goto try_again; 499 } 500 } 501 502 perf_evsel__open_strerror(counter, &target, 503 errno, msg, sizeof(msg)); 504 ui__error("%s\n", msg); 505 506 if (child_pid != -1) 507 kill(child_pid, SIGTERM); 508 509 return -1; 510 } 511 counter->supported = true; 512 513 l = strlen(counter->unit); 514 if (l > stat_config.unit_width) 515 stat_config.unit_width = l; 516 517 if (perf_evsel__should_store_id(counter) && 518 perf_evsel__store_ids(counter, evsel_list)) 519 return -1; 520 } 521 522 if (perf_evlist__apply_filters(evsel_list, &counter)) { 523 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n", 524 counter->filter, perf_evsel__name(counter), errno, 525 str_error_r(errno, msg, sizeof(msg))); 526 return -1; 527 } 528 529 if (STAT_RECORD) { 530 int err, fd = perf_data__fd(&perf_stat.data); 531 532 if (is_pipe) { 533 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data)); 534 } else { 535 err = perf_session__write_header(perf_stat.session, evsel_list, 536 fd, false); 537 } 538 539 if (err < 0) 540 return err; 541 542 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list, 543 process_synthesized_event, is_pipe); 544 if (err < 0) 545 return err; 546 } 547 548 /* 549 * Enable counters and exec the command: 550 */ 551 t0 = rdclock(); 552 clock_gettime(CLOCK_MONOTONIC, &ref_time); 553 554 if (forks) { 555 perf_evlist__start_workload(evsel_list); 556 enable_counters(); 557 558 if (interval || timeout) { 559 while (!waitpid(child_pid, &status, WNOHANG)) { 560 nanosleep(&ts, NULL); 561 if (timeout) 562 break; 563 process_interval(); 564 if (interval_count && !(--times)) 565 break; 566 } 567 } 568 if (child_pid != -1) 569 wait4(child_pid, &status, 0, &stat_config.ru_data); 570 571 if (workload_exec_errno) { 572 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg)); 573 pr_err("Workload failed: %s\n", emsg); 574 return -1; 575 } 576 577 if (WIFSIGNALED(status)) 578 psignal(WTERMSIG(status), argv[0]); 579 } else { 580 enable_counters(); 581 while (!done) { 582 nanosleep(&ts, NULL); 583 if (!is_target_alive(&target, evsel_list->core.threads)) 584 break; 585 if (timeout) 586 break; 587 if (interval) { 588 process_interval(); 589 if (interval_count && !(--times)) 590 break; 591 } 592 } 593 } 594 595 disable_counters(); 596 597 t1 = rdclock(); 598 599 if (stat_config.walltime_run_table) 600 stat_config.walltime_run[run_idx] = t1 - t0; 601 602 update_stats(&walltime_nsecs_stats, t1 - t0); 603 604 /* 605 * Closing a group leader splits the group, and as we only disable 606 * group leaders, results in remaining events becoming enabled. To 607 * avoid arbitrary skew, we must read all counters before closing any 608 * group leaders. 609 */ 610 read_counters(&(struct timespec) { .tv_nsec = t1-t0 }); 611 612 /* 613 * We need to keep evsel_list alive, because it's processed 614 * later the evsel_list will be closed after. 615 */ 616 if (!STAT_RECORD) 617 evlist__close(evsel_list); 618 619 return WEXITSTATUS(status); 620 } 621 622 static int run_perf_stat(int argc, const char **argv, int run_idx) 623 { 624 int ret; 625 626 if (pre_cmd) { 627 ret = system(pre_cmd); 628 if (ret) 629 return ret; 630 } 631 632 if (sync_run) 633 sync(); 634 635 ret = __run_perf_stat(argc, argv, run_idx); 636 if (ret) 637 return ret; 638 639 if (post_cmd) { 640 ret = system(post_cmd); 641 if (ret) 642 return ret; 643 } 644 645 return ret; 646 } 647 648 static void print_counters(struct timespec *ts, int argc, const char **argv) 649 { 650 /* Do not print anything if we record to the pipe. */ 651 if (STAT_RECORD && perf_stat.data.is_pipe) 652 return; 653 654 perf_evlist__print_counters(evsel_list, &stat_config, &target, 655 ts, argc, argv); 656 } 657 658 static volatile int signr = -1; 659 660 static void skip_signal(int signo) 661 { 662 if ((child_pid == -1) || stat_config.interval) 663 done = 1; 664 665 signr = signo; 666 /* 667 * render child_pid harmless 668 * won't send SIGTERM to a random 669 * process in case of race condition 670 * and fast PID recycling 671 */ 672 child_pid = -1; 673 } 674 675 static void sig_atexit(void) 676 { 677 sigset_t set, oset; 678 679 /* 680 * avoid race condition with SIGCHLD handler 681 * in skip_signal() which is modifying child_pid 682 * goal is to avoid send SIGTERM to a random 683 * process 684 */ 685 sigemptyset(&set); 686 sigaddset(&set, SIGCHLD); 687 sigprocmask(SIG_BLOCK, &set, &oset); 688 689 if (child_pid != -1) 690 kill(child_pid, SIGTERM); 691 692 sigprocmask(SIG_SETMASK, &oset, NULL); 693 694 if (signr == -1) 695 return; 696 697 signal(signr, SIG_DFL); 698 kill(getpid(), signr); 699 } 700 701 static int stat__set_big_num(const struct option *opt __maybe_unused, 702 const char *s __maybe_unused, int unset) 703 { 704 big_num_opt = unset ? 0 : 1; 705 return 0; 706 } 707 708 static int enable_metric_only(const struct option *opt __maybe_unused, 709 const char *s __maybe_unused, int unset) 710 { 711 force_metric_only = true; 712 stat_config.metric_only = !unset; 713 return 0; 714 } 715 716 static int parse_metric_groups(const struct option *opt, 717 const char *str, 718 int unset __maybe_unused) 719 { 720 return metricgroup__parse_groups(opt, str, &stat_config.metric_events); 721 } 722 723 static struct option stat_options[] = { 724 OPT_BOOLEAN('T', "transaction", &transaction_run, 725 "hardware transaction statistics"), 726 OPT_CALLBACK('e', "event", &evsel_list, "event", 727 "event selector. use 'perf list' to list available events", 728 parse_events_option), 729 OPT_CALLBACK(0, "filter", &evsel_list, "filter", 730 "event filter", parse_filter), 731 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit, 732 "child tasks do not inherit counters"), 733 OPT_STRING('p', "pid", &target.pid, "pid", 734 "stat events on existing process id"), 735 OPT_STRING('t', "tid", &target.tid, "tid", 736 "stat events on existing thread id"), 737 OPT_BOOLEAN('a', "all-cpus", &target.system_wide, 738 "system-wide collection from all CPUs"), 739 OPT_BOOLEAN('g', "group", &group, 740 "put the counters into a counter group"), 741 OPT_BOOLEAN(0, "scale", &stat_config.scale, 742 "Use --no-scale to disable counter scaling for multiplexing"), 743 OPT_INCR('v', "verbose", &verbose, 744 "be more verbose (show counter open errors, etc)"), 745 OPT_INTEGER('r', "repeat", &stat_config.run_count, 746 "repeat command and print average + stddev (max: 100, forever: 0)"), 747 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table, 748 "display details about each run (only with -r option)"), 749 OPT_BOOLEAN('n', "null", &stat_config.null_run, 750 "null run - dont start any counters"), 751 OPT_INCR('d', "detailed", &detailed_run, 752 "detailed run - start a lot of events"), 753 OPT_BOOLEAN('S', "sync", &sync_run, 754 "call sync() before starting a run"), 755 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 756 "print large numbers with thousands\' separators", 757 stat__set_big_num), 758 OPT_STRING('C', "cpu", &target.cpu_list, "cpu", 759 "list of cpus to monitor in system-wide"), 760 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode, 761 "disable CPU count aggregation", AGGR_NONE), 762 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"), 763 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator", 764 "print counts with custom separator"), 765 OPT_CALLBACK('G', "cgroup", &evsel_list, "name", 766 "monitor event in cgroup name only", parse_cgroups), 767 OPT_STRING('o', "output", &output_name, "file", "output file name"), 768 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"), 769 OPT_INTEGER(0, "log-fd", &output_fd, 770 "log output to fd, instead of stderr"), 771 OPT_STRING(0, "pre", &pre_cmd, "command", 772 "command to run prior to the measured command"), 773 OPT_STRING(0, "post", &post_cmd, "command", 774 "command to run after to the measured command"), 775 OPT_UINTEGER('I', "interval-print", &stat_config.interval, 776 "print counts at regular interval in ms " 777 "(overhead is possible for values <= 100ms)"), 778 OPT_INTEGER(0, "interval-count", &stat_config.times, 779 "print counts for fixed number of times"), 780 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear, 781 "clear screen in between new interval"), 782 OPT_UINTEGER(0, "timeout", &stat_config.timeout, 783 "stop workload and print counts after a timeout period in ms (>= 10ms)"), 784 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode, 785 "aggregate counts per processor socket", AGGR_SOCKET), 786 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode, 787 "aggregate counts per processor die", AGGR_DIE), 788 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode, 789 "aggregate counts per physical processor core", AGGR_CORE), 790 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode, 791 "aggregate counts per thread", AGGR_THREAD), 792 OPT_UINTEGER('D', "delay", &stat_config.initial_delay, 793 "ms to wait before starting measurement after program start"), 794 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL, 795 "Only print computed metrics. No raw values", enable_metric_only), 796 OPT_BOOLEAN(0, "topdown", &topdown_run, 797 "measure topdown level 1 statistics"), 798 OPT_BOOLEAN(0, "smi-cost", &smi_cost, 799 "measure SMI cost"), 800 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list", 801 "monitor specified metrics or metric groups (separated by ,)", 802 parse_metric_groups), 803 OPT_END() 804 }; 805 806 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused, 807 struct perf_cpu_map *map, int cpu) 808 { 809 return cpu_map__get_socket(map, cpu, NULL); 810 } 811 812 static int perf_stat__get_die(struct perf_stat_config *config __maybe_unused, 813 struct perf_cpu_map *map, int cpu) 814 { 815 return cpu_map__get_die(map, cpu, NULL); 816 } 817 818 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused, 819 struct perf_cpu_map *map, int cpu) 820 { 821 return cpu_map__get_core(map, cpu, NULL); 822 } 823 824 static int cpu_map__get_max(struct perf_cpu_map *map) 825 { 826 int i, max = -1; 827 828 for (i = 0; i < map->nr; i++) { 829 if (map->map[i] > max) 830 max = map->map[i]; 831 } 832 833 return max; 834 } 835 836 static int perf_stat__get_aggr(struct perf_stat_config *config, 837 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx) 838 { 839 int cpu; 840 841 if (idx >= map->nr) 842 return -1; 843 844 cpu = map->map[idx]; 845 846 if (config->cpus_aggr_map->map[cpu] == -1) 847 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx); 848 849 return config->cpus_aggr_map->map[cpu]; 850 } 851 852 static int perf_stat__get_socket_cached(struct perf_stat_config *config, 853 struct perf_cpu_map *map, int idx) 854 { 855 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx); 856 } 857 858 static int perf_stat__get_die_cached(struct perf_stat_config *config, 859 struct perf_cpu_map *map, int idx) 860 { 861 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx); 862 } 863 864 static int perf_stat__get_core_cached(struct perf_stat_config *config, 865 struct perf_cpu_map *map, int idx) 866 { 867 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx); 868 } 869 870 static bool term_percore_set(void) 871 { 872 struct evsel *counter; 873 874 evlist__for_each_entry(evsel_list, counter) { 875 if (counter->percore) 876 return true; 877 } 878 879 return false; 880 } 881 882 static int perf_stat_init_aggr_mode(void) 883 { 884 int nr; 885 886 switch (stat_config.aggr_mode) { 887 case AGGR_SOCKET: 888 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 889 perror("cannot build socket map"); 890 return -1; 891 } 892 stat_config.aggr_get_id = perf_stat__get_socket_cached; 893 break; 894 case AGGR_DIE: 895 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 896 perror("cannot build die map"); 897 return -1; 898 } 899 stat_config.aggr_get_id = perf_stat__get_die_cached; 900 break; 901 case AGGR_CORE: 902 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) { 903 perror("cannot build core map"); 904 return -1; 905 } 906 stat_config.aggr_get_id = perf_stat__get_core_cached; 907 break; 908 case AGGR_NONE: 909 if (term_percore_set()) { 910 if (cpu_map__build_core_map(evsel_list->core.cpus, 911 &stat_config.aggr_map)) { 912 perror("cannot build core map"); 913 return -1; 914 } 915 stat_config.aggr_get_id = perf_stat__get_core_cached; 916 } 917 break; 918 case AGGR_GLOBAL: 919 case AGGR_THREAD: 920 case AGGR_UNSET: 921 default: 922 break; 923 } 924 925 /* 926 * The evsel_list->cpus is the base we operate on, 927 * taking the highest cpu number to be the size of 928 * the aggregation translate cpumap. 929 */ 930 nr = cpu_map__get_max(evsel_list->core.cpus); 931 stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1); 932 return stat_config.cpus_aggr_map ? 0 : -ENOMEM; 933 } 934 935 static void perf_stat__exit_aggr_mode(void) 936 { 937 perf_cpu_map__put(stat_config.aggr_map); 938 perf_cpu_map__put(stat_config.cpus_aggr_map); 939 stat_config.aggr_map = NULL; 940 stat_config.cpus_aggr_map = NULL; 941 } 942 943 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx) 944 { 945 int cpu; 946 947 if (idx > map->nr) 948 return -1; 949 950 cpu = map->map[idx]; 951 952 if (cpu >= env->nr_cpus_avail) 953 return -1; 954 955 return cpu; 956 } 957 958 static int perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data) 959 { 960 struct perf_env *env = data; 961 int cpu = perf_env__get_cpu(env, map, idx); 962 963 return cpu == -1 ? -1 : env->cpu[cpu].socket_id; 964 } 965 966 static int perf_env__get_die(struct perf_cpu_map *map, int idx, void *data) 967 { 968 struct perf_env *env = data; 969 int die_id = -1, cpu = perf_env__get_cpu(env, map, idx); 970 971 if (cpu != -1) { 972 /* 973 * Encode socket in bit range 15:8 974 * die_id is relative to socket, 975 * we need a global id. So we combine 976 * socket + die id 977 */ 978 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) 979 return -1; 980 981 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) 982 return -1; 983 984 die_id = (env->cpu[cpu].socket_id << 8) | (env->cpu[cpu].die_id & 0xff); 985 } 986 987 return die_id; 988 } 989 990 static int perf_env__get_core(struct perf_cpu_map *map, int idx, void *data) 991 { 992 struct perf_env *env = data; 993 int core = -1, cpu = perf_env__get_cpu(env, map, idx); 994 995 if (cpu != -1) { 996 /* 997 * Encode socket in bit range 31:24 998 * encode die id in bit range 23:16 999 * core_id is relative to socket and die, 1000 * we need a global id. So we combine 1001 * socket + die id + core id 1002 */ 1003 if (WARN_ONCE(env->cpu[cpu].socket_id >> 8, "The socket id number is too big.\n")) 1004 return -1; 1005 1006 if (WARN_ONCE(env->cpu[cpu].die_id >> 8, "The die id number is too big.\n")) 1007 return -1; 1008 1009 if (WARN_ONCE(env->cpu[cpu].core_id >> 16, "The core id number is too big.\n")) 1010 return -1; 1011 1012 core = (env->cpu[cpu].socket_id << 24) | 1013 (env->cpu[cpu].die_id << 16) | 1014 (env->cpu[cpu].core_id & 0xffff); 1015 } 1016 1017 return core; 1018 } 1019 1020 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus, 1021 struct perf_cpu_map **sockp) 1022 { 1023 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env); 1024 } 1025 1026 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus, 1027 struct perf_cpu_map **diep) 1028 { 1029 return cpu_map__build_map(cpus, diep, perf_env__get_die, env); 1030 } 1031 1032 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus, 1033 struct perf_cpu_map **corep) 1034 { 1035 return cpu_map__build_map(cpus, corep, perf_env__get_core, env); 1036 } 1037 1038 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused, 1039 struct perf_cpu_map *map, int idx) 1040 { 1041 return perf_env__get_socket(map, idx, &perf_stat.session->header.env); 1042 } 1043 static int perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused, 1044 struct perf_cpu_map *map, int idx) 1045 { 1046 return perf_env__get_die(map, idx, &perf_stat.session->header.env); 1047 } 1048 1049 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused, 1050 struct perf_cpu_map *map, int idx) 1051 { 1052 return perf_env__get_core(map, idx, &perf_stat.session->header.env); 1053 } 1054 1055 static int perf_stat_init_aggr_mode_file(struct perf_stat *st) 1056 { 1057 struct perf_env *env = &st->session->header.env; 1058 1059 switch (stat_config.aggr_mode) { 1060 case AGGR_SOCKET: 1061 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1062 perror("cannot build socket map"); 1063 return -1; 1064 } 1065 stat_config.aggr_get_id = perf_stat__get_socket_file; 1066 break; 1067 case AGGR_DIE: 1068 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1069 perror("cannot build die map"); 1070 return -1; 1071 } 1072 stat_config.aggr_get_id = perf_stat__get_die_file; 1073 break; 1074 case AGGR_CORE: 1075 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) { 1076 perror("cannot build core map"); 1077 return -1; 1078 } 1079 stat_config.aggr_get_id = perf_stat__get_core_file; 1080 break; 1081 case AGGR_NONE: 1082 case AGGR_GLOBAL: 1083 case AGGR_THREAD: 1084 case AGGR_UNSET: 1085 default: 1086 break; 1087 } 1088 1089 return 0; 1090 } 1091 1092 static int topdown_filter_events(const char **attr, char **str, bool use_group) 1093 { 1094 int off = 0; 1095 int i; 1096 int len = 0; 1097 char *s; 1098 1099 for (i = 0; attr[i]; i++) { 1100 if (pmu_have_event("cpu", attr[i])) { 1101 len += strlen(attr[i]) + 1; 1102 attr[i - off] = attr[i]; 1103 } else 1104 off++; 1105 } 1106 attr[i - off] = NULL; 1107 1108 *str = malloc(len + 1 + 2); 1109 if (!*str) 1110 return -1; 1111 s = *str; 1112 if (i - off == 0) { 1113 *s = 0; 1114 return 0; 1115 } 1116 if (use_group) 1117 *s++ = '{'; 1118 for (i = 0; attr[i]; i++) { 1119 strcpy(s, attr[i]); 1120 s += strlen(s); 1121 *s++ = ','; 1122 } 1123 if (use_group) { 1124 s[-1] = '}'; 1125 *s = 0; 1126 } else 1127 s[-1] = 0; 1128 return 0; 1129 } 1130 1131 __weak bool arch_topdown_check_group(bool *warn) 1132 { 1133 *warn = false; 1134 return false; 1135 } 1136 1137 __weak void arch_topdown_group_warn(void) 1138 { 1139 } 1140 1141 /* 1142 * Add default attributes, if there were no attributes specified or 1143 * if -d/--detailed, -d -d or -d -d -d is used: 1144 */ 1145 static int add_default_attributes(void) 1146 { 1147 int err; 1148 struct perf_event_attr default_attrs0[] = { 1149 1150 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK }, 1151 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES }, 1152 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS }, 1153 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 1154 1155 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 1156 }; 1157 struct perf_event_attr frontend_attrs[] = { 1158 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND }, 1159 }; 1160 struct perf_event_attr backend_attrs[] = { 1161 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND }, 1162 }; 1163 struct perf_event_attr default_attrs1[] = { 1164 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 1165 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 1166 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 1167 1168 }; 1169 1170 /* 1171 * Detailed stats (-d), covering the L1 and last level data caches: 1172 */ 1173 struct perf_event_attr detailed_attrs[] = { 1174 1175 { .type = PERF_TYPE_HW_CACHE, 1176 .config = 1177 PERF_COUNT_HW_CACHE_L1D << 0 | 1178 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1179 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1180 1181 { .type = PERF_TYPE_HW_CACHE, 1182 .config = 1183 PERF_COUNT_HW_CACHE_L1D << 0 | 1184 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1185 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1186 1187 { .type = PERF_TYPE_HW_CACHE, 1188 .config = 1189 PERF_COUNT_HW_CACHE_LL << 0 | 1190 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1191 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1192 1193 { .type = PERF_TYPE_HW_CACHE, 1194 .config = 1195 PERF_COUNT_HW_CACHE_LL << 0 | 1196 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1197 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1198 }; 1199 1200 /* 1201 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches: 1202 */ 1203 struct perf_event_attr very_detailed_attrs[] = { 1204 1205 { .type = PERF_TYPE_HW_CACHE, 1206 .config = 1207 PERF_COUNT_HW_CACHE_L1I << 0 | 1208 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1209 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1210 1211 { .type = PERF_TYPE_HW_CACHE, 1212 .config = 1213 PERF_COUNT_HW_CACHE_L1I << 0 | 1214 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1215 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1216 1217 { .type = PERF_TYPE_HW_CACHE, 1218 .config = 1219 PERF_COUNT_HW_CACHE_DTLB << 0 | 1220 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1221 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1222 1223 { .type = PERF_TYPE_HW_CACHE, 1224 .config = 1225 PERF_COUNT_HW_CACHE_DTLB << 0 | 1226 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1227 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1228 1229 { .type = PERF_TYPE_HW_CACHE, 1230 .config = 1231 PERF_COUNT_HW_CACHE_ITLB << 0 | 1232 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1233 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1234 1235 { .type = PERF_TYPE_HW_CACHE, 1236 .config = 1237 PERF_COUNT_HW_CACHE_ITLB << 0 | 1238 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 1239 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1240 1241 }; 1242 1243 /* 1244 * Very, very detailed stats (-d -d -d), adding prefetch events: 1245 */ 1246 struct perf_event_attr very_very_detailed_attrs[] = { 1247 1248 { .type = PERF_TYPE_HW_CACHE, 1249 .config = 1250 PERF_COUNT_HW_CACHE_L1D << 0 | 1251 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1252 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) }, 1253 1254 { .type = PERF_TYPE_HW_CACHE, 1255 .config = 1256 PERF_COUNT_HW_CACHE_L1D << 0 | 1257 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) | 1258 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) }, 1259 }; 1260 struct parse_events_error errinfo; 1261 1262 /* Set attrs if no event is selected and !null_run: */ 1263 if (stat_config.null_run) 1264 return 0; 1265 1266 if (transaction_run) { 1267 /* Handle -T as -M transaction. Once platform specific metrics 1268 * support has been added to the json files, all archictures 1269 * will use this approach. To determine transaction support 1270 * on an architecture test for such a metric name. 1271 */ 1272 if (metricgroup__has_metric("transaction")) { 1273 struct option opt = { .value = &evsel_list }; 1274 1275 return metricgroup__parse_groups(&opt, "transaction", 1276 &stat_config.metric_events); 1277 } 1278 1279 if (pmu_have_event("cpu", "cycles-ct") && 1280 pmu_have_event("cpu", "el-start")) 1281 err = parse_events(evsel_list, transaction_attrs, 1282 &errinfo); 1283 else 1284 err = parse_events(evsel_list, 1285 transaction_limited_attrs, 1286 &errinfo); 1287 if (err) { 1288 fprintf(stderr, "Cannot set up transaction events\n"); 1289 parse_events_print_error(&errinfo, transaction_attrs); 1290 return -1; 1291 } 1292 return 0; 1293 } 1294 1295 if (smi_cost) { 1296 int smi; 1297 1298 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) { 1299 fprintf(stderr, "freeze_on_smi is not supported.\n"); 1300 return -1; 1301 } 1302 1303 if (!smi) { 1304 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) { 1305 fprintf(stderr, "Failed to set freeze_on_smi.\n"); 1306 return -1; 1307 } 1308 smi_reset = true; 1309 } 1310 1311 if (pmu_have_event("msr", "aperf") && 1312 pmu_have_event("msr", "smi")) { 1313 if (!force_metric_only) 1314 stat_config.metric_only = true; 1315 err = parse_events(evsel_list, smi_cost_attrs, &errinfo); 1316 } else { 1317 fprintf(stderr, "To measure SMI cost, it needs " 1318 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n"); 1319 parse_events_print_error(&errinfo, smi_cost_attrs); 1320 return -1; 1321 } 1322 if (err) { 1323 fprintf(stderr, "Cannot set up SMI cost events\n"); 1324 return -1; 1325 } 1326 return 0; 1327 } 1328 1329 if (topdown_run) { 1330 char *str = NULL; 1331 bool warn = false; 1332 1333 if (stat_config.aggr_mode != AGGR_GLOBAL && 1334 stat_config.aggr_mode != AGGR_CORE) { 1335 pr_err("top down event configuration requires --per-core mode\n"); 1336 return -1; 1337 } 1338 stat_config.aggr_mode = AGGR_CORE; 1339 if (nr_cgroups || !target__has_cpu(&target)) { 1340 pr_err("top down event configuration requires system-wide mode (-a)\n"); 1341 return -1; 1342 } 1343 1344 if (!force_metric_only) 1345 stat_config.metric_only = true; 1346 if (topdown_filter_events(topdown_attrs, &str, 1347 arch_topdown_check_group(&warn)) < 0) { 1348 pr_err("Out of memory\n"); 1349 return -1; 1350 } 1351 if (topdown_attrs[0] && str) { 1352 if (warn) 1353 arch_topdown_group_warn(); 1354 err = parse_events(evsel_list, str, &errinfo); 1355 if (err) { 1356 fprintf(stderr, 1357 "Cannot set up top down events %s: %d\n", 1358 str, err); 1359 parse_events_print_error(&errinfo, str); 1360 free(str); 1361 return -1; 1362 } 1363 } else { 1364 fprintf(stderr, "System does not support topdown\n"); 1365 return -1; 1366 } 1367 free(str); 1368 } 1369 1370 if (!evsel_list->core.nr_entries) { 1371 if (target__has_cpu(&target)) 1372 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK; 1373 1374 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0) 1375 return -1; 1376 if (pmu_have_event("cpu", "stalled-cycles-frontend")) { 1377 if (perf_evlist__add_default_attrs(evsel_list, 1378 frontend_attrs) < 0) 1379 return -1; 1380 } 1381 if (pmu_have_event("cpu", "stalled-cycles-backend")) { 1382 if (perf_evlist__add_default_attrs(evsel_list, 1383 backend_attrs) < 0) 1384 return -1; 1385 } 1386 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0) 1387 return -1; 1388 } 1389 1390 /* Detailed events get appended to the event list: */ 1391 1392 if (detailed_run < 1) 1393 return 0; 1394 1395 /* Append detailed run extra attributes: */ 1396 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0) 1397 return -1; 1398 1399 if (detailed_run < 2) 1400 return 0; 1401 1402 /* Append very detailed run extra attributes: */ 1403 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0) 1404 return -1; 1405 1406 if (detailed_run < 3) 1407 return 0; 1408 1409 /* Append very, very detailed run extra attributes: */ 1410 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs); 1411 } 1412 1413 static const char * const stat_record_usage[] = { 1414 "perf stat record [<options>]", 1415 NULL, 1416 }; 1417 1418 static void init_features(struct perf_session *session) 1419 { 1420 int feat; 1421 1422 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++) 1423 perf_header__set_feat(&session->header, feat); 1424 1425 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT); 1426 perf_header__clear_feat(&session->header, HEADER_BUILD_ID); 1427 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA); 1428 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK); 1429 perf_header__clear_feat(&session->header, HEADER_AUXTRACE); 1430 } 1431 1432 static int __cmd_record(int argc, const char **argv) 1433 { 1434 struct perf_session *session; 1435 struct perf_data *data = &perf_stat.data; 1436 1437 argc = parse_options(argc, argv, stat_options, stat_record_usage, 1438 PARSE_OPT_STOP_AT_NON_OPTION); 1439 1440 if (output_name) 1441 data->path = output_name; 1442 1443 if (stat_config.run_count != 1 || forever) { 1444 pr_err("Cannot use -r option with perf stat record.\n"); 1445 return -1; 1446 } 1447 1448 session = perf_session__new(data, false, NULL); 1449 if (session == NULL) { 1450 pr_err("Perf session creation failed.\n"); 1451 return -1; 1452 } 1453 1454 init_features(session); 1455 1456 session->evlist = evsel_list; 1457 perf_stat.session = session; 1458 perf_stat.record = true; 1459 return argc; 1460 } 1461 1462 static int process_stat_round_event(struct perf_session *session, 1463 union perf_event *event) 1464 { 1465 struct stat_round_event *stat_round = &event->stat_round; 1466 struct evsel *counter; 1467 struct timespec tsh, *ts = NULL; 1468 const char **argv = session->header.env.cmdline_argv; 1469 int argc = session->header.env.nr_cmdline; 1470 1471 evlist__for_each_entry(evsel_list, counter) 1472 perf_stat_process_counter(&stat_config, counter); 1473 1474 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL) 1475 update_stats(&walltime_nsecs_stats, stat_round->time); 1476 1477 if (stat_config.interval && stat_round->time) { 1478 tsh.tv_sec = stat_round->time / NSEC_PER_SEC; 1479 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC; 1480 ts = &tsh; 1481 } 1482 1483 print_counters(ts, argc, argv); 1484 return 0; 1485 } 1486 1487 static 1488 int process_stat_config_event(struct perf_session *session, 1489 union perf_event *event) 1490 { 1491 struct perf_tool *tool = session->tool; 1492 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1493 1494 perf_event__read_stat_config(&stat_config, &event->stat_config); 1495 1496 if (cpu_map__empty(st->cpus)) { 1497 if (st->aggr_mode != AGGR_UNSET) 1498 pr_warning("warning: processing task data, aggregation mode not set\n"); 1499 return 0; 1500 } 1501 1502 if (st->aggr_mode != AGGR_UNSET) 1503 stat_config.aggr_mode = st->aggr_mode; 1504 1505 if (perf_stat.data.is_pipe) 1506 perf_stat_init_aggr_mode(); 1507 else 1508 perf_stat_init_aggr_mode_file(st); 1509 1510 return 0; 1511 } 1512 1513 static int set_maps(struct perf_stat *st) 1514 { 1515 if (!st->cpus || !st->threads) 1516 return 0; 1517 1518 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n")) 1519 return -EINVAL; 1520 1521 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads); 1522 1523 if (perf_evlist__alloc_stats(evsel_list, true)) 1524 return -ENOMEM; 1525 1526 st->maps_allocated = true; 1527 return 0; 1528 } 1529 1530 static 1531 int process_thread_map_event(struct perf_session *session, 1532 union perf_event *event) 1533 { 1534 struct perf_tool *tool = session->tool; 1535 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1536 1537 if (st->threads) { 1538 pr_warning("Extra thread map event, ignoring.\n"); 1539 return 0; 1540 } 1541 1542 st->threads = thread_map__new_event(&event->thread_map); 1543 if (!st->threads) 1544 return -ENOMEM; 1545 1546 return set_maps(st); 1547 } 1548 1549 static 1550 int process_cpu_map_event(struct perf_session *session, 1551 union perf_event *event) 1552 { 1553 struct perf_tool *tool = session->tool; 1554 struct perf_stat *st = container_of(tool, struct perf_stat, tool); 1555 struct perf_cpu_map *cpus; 1556 1557 if (st->cpus) { 1558 pr_warning("Extra cpu map event, ignoring.\n"); 1559 return 0; 1560 } 1561 1562 cpus = cpu_map__new_data(&event->cpu_map.data); 1563 if (!cpus) 1564 return -ENOMEM; 1565 1566 st->cpus = cpus; 1567 return set_maps(st); 1568 } 1569 1570 static int runtime_stat_new(struct perf_stat_config *config, int nthreads) 1571 { 1572 int i; 1573 1574 config->stats = calloc(nthreads, sizeof(struct runtime_stat)); 1575 if (!config->stats) 1576 return -1; 1577 1578 config->stats_num = nthreads; 1579 1580 for (i = 0; i < nthreads; i++) 1581 runtime_stat__init(&config->stats[i]); 1582 1583 return 0; 1584 } 1585 1586 static void runtime_stat_delete(struct perf_stat_config *config) 1587 { 1588 int i; 1589 1590 if (!config->stats) 1591 return; 1592 1593 for (i = 0; i < config->stats_num; i++) 1594 runtime_stat__exit(&config->stats[i]); 1595 1596 zfree(&config->stats); 1597 } 1598 1599 static const char * const stat_report_usage[] = { 1600 "perf stat report [<options>]", 1601 NULL, 1602 }; 1603 1604 static struct perf_stat perf_stat = { 1605 .tool = { 1606 .attr = perf_event__process_attr, 1607 .event_update = perf_event__process_event_update, 1608 .thread_map = process_thread_map_event, 1609 .cpu_map = process_cpu_map_event, 1610 .stat_config = process_stat_config_event, 1611 .stat = perf_event__process_stat_event, 1612 .stat_round = process_stat_round_event, 1613 }, 1614 .aggr_mode = AGGR_UNSET, 1615 }; 1616 1617 static int __cmd_report(int argc, const char **argv) 1618 { 1619 struct perf_session *session; 1620 const struct option options[] = { 1621 OPT_STRING('i', "input", &input_name, "file", "input file name"), 1622 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode, 1623 "aggregate counts per processor socket", AGGR_SOCKET), 1624 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode, 1625 "aggregate counts per processor die", AGGR_DIE), 1626 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode, 1627 "aggregate counts per physical processor core", AGGR_CORE), 1628 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode, 1629 "disable CPU count aggregation", AGGR_NONE), 1630 OPT_END() 1631 }; 1632 struct stat st; 1633 int ret; 1634 1635 argc = parse_options(argc, argv, options, stat_report_usage, 0); 1636 1637 if (!input_name || !strlen(input_name)) { 1638 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 1639 input_name = "-"; 1640 else 1641 input_name = "perf.data"; 1642 } 1643 1644 perf_stat.data.path = input_name; 1645 perf_stat.data.mode = PERF_DATA_MODE_READ; 1646 1647 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool); 1648 if (session == NULL) 1649 return -1; 1650 1651 perf_stat.session = session; 1652 stat_config.output = stderr; 1653 evsel_list = session->evlist; 1654 1655 ret = perf_session__process_events(session); 1656 if (ret) 1657 return ret; 1658 1659 perf_session__delete(session); 1660 return 0; 1661 } 1662 1663 static void setup_system_wide(int forks) 1664 { 1665 /* 1666 * Make system wide (-a) the default target if 1667 * no target was specified and one of following 1668 * conditions is met: 1669 * 1670 * - there's no workload specified 1671 * - there is workload specified but all requested 1672 * events are system wide events 1673 */ 1674 if (!target__none(&target)) 1675 return; 1676 1677 if (!forks) 1678 target.system_wide = true; 1679 else { 1680 struct evsel *counter; 1681 1682 evlist__for_each_entry(evsel_list, counter) { 1683 if (!counter->system_wide) 1684 return; 1685 } 1686 1687 if (evsel_list->core.nr_entries) 1688 target.system_wide = true; 1689 } 1690 } 1691 1692 int cmd_stat(int argc, const char **argv) 1693 { 1694 const char * const stat_usage[] = { 1695 "perf stat [<options>] [<command>]", 1696 NULL 1697 }; 1698 int status = -EINVAL, run_idx; 1699 const char *mode; 1700 FILE *output = stderr; 1701 unsigned int interval, timeout; 1702 const char * const stat_subcommands[] = { "record", "report" }; 1703 1704 setlocale(LC_ALL, ""); 1705 1706 evsel_list = evlist__new(); 1707 if (evsel_list == NULL) 1708 return -ENOMEM; 1709 1710 parse_events__shrink_config_terms(); 1711 1712 /* String-parsing callback-based options would segfault when negated */ 1713 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG); 1714 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG); 1715 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG); 1716 1717 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands, 1718 (const char **) stat_usage, 1719 PARSE_OPT_STOP_AT_NON_OPTION); 1720 perf_stat__collect_metric_expr(evsel_list); 1721 perf_stat__init_shadow_stats(); 1722 1723 if (stat_config.csv_sep) { 1724 stat_config.csv_output = true; 1725 if (!strcmp(stat_config.csv_sep, "\\t")) 1726 stat_config.csv_sep = "\t"; 1727 } else 1728 stat_config.csv_sep = DEFAULT_SEPARATOR; 1729 1730 if (argc && !strncmp(argv[0], "rec", 3)) { 1731 argc = __cmd_record(argc, argv); 1732 if (argc < 0) 1733 return -1; 1734 } else if (argc && !strncmp(argv[0], "rep", 3)) 1735 return __cmd_report(argc, argv); 1736 1737 interval = stat_config.interval; 1738 timeout = stat_config.timeout; 1739 1740 /* 1741 * For record command the -o is already taken care of. 1742 */ 1743 if (!STAT_RECORD && output_name && strcmp(output_name, "-")) 1744 output = NULL; 1745 1746 if (output_name && output_fd) { 1747 fprintf(stderr, "cannot use both --output and --log-fd\n"); 1748 parse_options_usage(stat_usage, stat_options, "o", 1); 1749 parse_options_usage(NULL, stat_options, "log-fd", 0); 1750 goto out; 1751 } 1752 1753 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) { 1754 fprintf(stderr, "--metric-only is not supported with --per-thread\n"); 1755 goto out; 1756 } 1757 1758 if (stat_config.metric_only && stat_config.run_count > 1) { 1759 fprintf(stderr, "--metric-only is not supported with -r\n"); 1760 goto out; 1761 } 1762 1763 if (stat_config.walltime_run_table && stat_config.run_count <= 1) { 1764 fprintf(stderr, "--table is only supported with -r\n"); 1765 parse_options_usage(stat_usage, stat_options, "r", 1); 1766 parse_options_usage(NULL, stat_options, "table", 0); 1767 goto out; 1768 } 1769 1770 if (output_fd < 0) { 1771 fprintf(stderr, "argument to --log-fd must be a > 0\n"); 1772 parse_options_usage(stat_usage, stat_options, "log-fd", 0); 1773 goto out; 1774 } 1775 1776 if (!output) { 1777 struct timespec tm; 1778 mode = append_file ? "a" : "w"; 1779 1780 output = fopen(output_name, mode); 1781 if (!output) { 1782 perror("failed to create output file"); 1783 return -1; 1784 } 1785 clock_gettime(CLOCK_REALTIME, &tm); 1786 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); 1787 } else if (output_fd > 0) { 1788 mode = append_file ? "a" : "w"; 1789 output = fdopen(output_fd, mode); 1790 if (!output) { 1791 perror("Failed opening logfd"); 1792 return -errno; 1793 } 1794 } 1795 1796 stat_config.output = output; 1797 1798 /* 1799 * let the spreadsheet do the pretty-printing 1800 */ 1801 if (stat_config.csv_output) { 1802 /* User explicitly passed -B? */ 1803 if (big_num_opt == 1) { 1804 fprintf(stderr, "-B option not supported with -x\n"); 1805 parse_options_usage(stat_usage, stat_options, "B", 1); 1806 parse_options_usage(NULL, stat_options, "x", 1); 1807 goto out; 1808 } else /* Nope, so disable big number formatting */ 1809 stat_config.big_num = false; 1810 } else if (big_num_opt == 0) /* User passed --no-big-num */ 1811 stat_config.big_num = false; 1812 1813 setup_system_wide(argc); 1814 1815 /* 1816 * Display user/system times only for single 1817 * run and when there's specified tracee. 1818 */ 1819 if ((stat_config.run_count == 1) && target__none(&target)) 1820 stat_config.ru_display = true; 1821 1822 if (stat_config.run_count < 0) { 1823 pr_err("Run count must be a positive number\n"); 1824 parse_options_usage(stat_usage, stat_options, "r", 1); 1825 goto out; 1826 } else if (stat_config.run_count == 0) { 1827 forever = true; 1828 stat_config.run_count = 1; 1829 } 1830 1831 if (stat_config.walltime_run_table) { 1832 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0])); 1833 if (!stat_config.walltime_run) { 1834 pr_err("failed to setup -r option"); 1835 goto out; 1836 } 1837 } 1838 1839 if ((stat_config.aggr_mode == AGGR_THREAD) && 1840 !target__has_task(&target)) { 1841 if (!target.system_wide || target.cpu_list) { 1842 fprintf(stderr, "The --per-thread option is only " 1843 "available when monitoring via -p -t -a " 1844 "options or only --per-thread.\n"); 1845 parse_options_usage(NULL, stat_options, "p", 1); 1846 parse_options_usage(NULL, stat_options, "t", 1); 1847 goto out; 1848 } 1849 } 1850 1851 /* 1852 * no_aggr, cgroup are for system-wide only 1853 * --per-thread is aggregated per thread, we dont mix it with cpu mode 1854 */ 1855 if (((stat_config.aggr_mode != AGGR_GLOBAL && 1856 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) && 1857 !target__has_cpu(&target)) { 1858 fprintf(stderr, "both cgroup and no-aggregation " 1859 "modes only available in system-wide mode\n"); 1860 1861 parse_options_usage(stat_usage, stat_options, "G", 1); 1862 parse_options_usage(NULL, stat_options, "A", 1); 1863 parse_options_usage(NULL, stat_options, "a", 1); 1864 goto out; 1865 } 1866 1867 if (add_default_attributes()) 1868 goto out; 1869 1870 target__validate(&target); 1871 1872 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide)) 1873 target.per_thread = true; 1874 1875 if (perf_evlist__create_maps(evsel_list, &target) < 0) { 1876 if (target__has_task(&target)) { 1877 pr_err("Problems finding threads of monitor\n"); 1878 parse_options_usage(stat_usage, stat_options, "p", 1); 1879 parse_options_usage(NULL, stat_options, "t", 1); 1880 } else if (target__has_cpu(&target)) { 1881 perror("failed to parse CPUs map"); 1882 parse_options_usage(stat_usage, stat_options, "C", 1); 1883 parse_options_usage(NULL, stat_options, "a", 1); 1884 } 1885 goto out; 1886 } 1887 1888 /* 1889 * Initialize thread_map with comm names, 1890 * so we could print it out on output. 1891 */ 1892 if (stat_config.aggr_mode == AGGR_THREAD) { 1893 thread_map__read_comms(evsel_list->core.threads); 1894 if (target.system_wide) { 1895 if (runtime_stat_new(&stat_config, 1896 thread_map__nr(evsel_list->core.threads))) { 1897 goto out; 1898 } 1899 } 1900 } 1901 1902 if (stat_config.times && interval) 1903 interval_count = true; 1904 else if (stat_config.times && !interval) { 1905 pr_err("interval-count option should be used together with " 1906 "interval-print.\n"); 1907 parse_options_usage(stat_usage, stat_options, "interval-count", 0); 1908 parse_options_usage(stat_usage, stat_options, "I", 1); 1909 goto out; 1910 } 1911 1912 if (timeout && timeout < 100) { 1913 if (timeout < 10) { 1914 pr_err("timeout must be >= 10ms.\n"); 1915 parse_options_usage(stat_usage, stat_options, "timeout", 0); 1916 goto out; 1917 } else 1918 pr_warning("timeout < 100ms. " 1919 "The overhead percentage could be high in some cases. " 1920 "Please proceed with caution.\n"); 1921 } 1922 if (timeout && interval) { 1923 pr_err("timeout option is not supported with interval-print.\n"); 1924 parse_options_usage(stat_usage, stat_options, "timeout", 0); 1925 parse_options_usage(stat_usage, stat_options, "I", 1); 1926 goto out; 1927 } 1928 1929 if (perf_evlist__alloc_stats(evsel_list, interval)) 1930 goto out; 1931 1932 if (perf_stat_init_aggr_mode()) 1933 goto out; 1934 1935 /* 1936 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless 1937 * while avoiding that older tools show confusing messages. 1938 * 1939 * However for pipe sessions we need to keep it zero, 1940 * because script's perf_evsel__check_attr is triggered 1941 * by attr->sample_type != 0, and we can't run it on 1942 * stat sessions. 1943 */ 1944 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe); 1945 1946 /* 1947 * We dont want to block the signals - that would cause 1948 * child tasks to inherit that and Ctrl-C would not work. 1949 * What we want is for Ctrl-C to work in the exec()-ed 1950 * task, but being ignored by perf stat itself: 1951 */ 1952 atexit(sig_atexit); 1953 if (!forever) 1954 signal(SIGINT, skip_signal); 1955 signal(SIGCHLD, skip_signal); 1956 signal(SIGALRM, skip_signal); 1957 signal(SIGABRT, skip_signal); 1958 1959 status = 0; 1960 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) { 1961 if (stat_config.run_count != 1 && verbose > 0) 1962 fprintf(output, "[ perf stat: executing run #%d ... ]\n", 1963 run_idx + 1); 1964 1965 status = run_perf_stat(argc, argv, run_idx); 1966 if (forever && status != -1) { 1967 print_counters(NULL, argc, argv); 1968 perf_stat__reset_stats(); 1969 } 1970 } 1971 1972 if (!forever && status != -1 && !interval) 1973 print_counters(NULL, argc, argv); 1974 1975 if (STAT_RECORD) { 1976 /* 1977 * We synthesize the kernel mmap record just so that older tools 1978 * don't emit warnings about not being able to resolve symbols 1979 * due to /proc/sys/kernel/kptr_restrict settings and instear provide 1980 * a saner message about no samples being in the perf.data file. 1981 * 1982 * This also serves to suppress a warning about f_header.data.size == 0 1983 * in header.c at the moment 'perf stat record' gets introduced, which 1984 * is not really needed once we start adding the stat specific PERF_RECORD_ 1985 * records, but the need to suppress the kptr_restrict messages in older 1986 * tools remain -acme 1987 */ 1988 int fd = perf_data__fd(&perf_stat.data); 1989 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat, 1990 process_synthesized_event, 1991 &perf_stat.session->machines.host); 1992 if (err) { 1993 pr_warning("Couldn't synthesize the kernel mmap record, harmless, " 1994 "older tools may produce warnings about this file\n."); 1995 } 1996 1997 if (!interval) { 1998 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL)) 1999 pr_err("failed to write stat round event\n"); 2000 } 2001 2002 if (!perf_stat.data.is_pipe) { 2003 perf_stat.session->header.data_size += perf_stat.bytes_written; 2004 perf_session__write_header(perf_stat.session, evsel_list, fd, true); 2005 } 2006 2007 evlist__close(evsel_list); 2008 perf_session__delete(perf_stat.session); 2009 } 2010 2011 perf_stat__exit_aggr_mode(); 2012 perf_evlist__free_stats(evsel_list); 2013 out: 2014 zfree(&stat_config.walltime_run); 2015 2016 if (smi_cost && smi_reset) 2017 sysfs__write_int(FREEZE_ON_SMI_PATH, 0); 2018 2019 evlist__delete(evsel_list); 2020 2021 runtime_stat_delete(&stat_config); 2022 2023 return status; 2024 } 2025