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