1 // SPDX-License-Identifier: GPL-2.0 2 #include "builtin.h" 3 #include "perf.h" 4 #include "perf-sys.h" 5 6 #include "util/cpumap.h" 7 #include "util/evlist.h" 8 #include "util/evsel.h" 9 #include "util/evsel_fprintf.h" 10 #include "util/mutex.h" 11 #include "util/symbol.h" 12 #include "util/thread.h" 13 #include "util/header.h" 14 #include "util/session.h" 15 #include "util/tool.h" 16 #include "util/cloexec.h" 17 #include "util/thread_map.h" 18 #include "util/color.h" 19 #include "util/stat.h" 20 #include "util/string2.h" 21 #include "util/callchain.h" 22 #include "util/time-utils.h" 23 24 #include <subcmd/pager.h> 25 #include <subcmd/parse-options.h> 26 #include "util/trace-event.h" 27 28 #include "util/debug.h" 29 #include "util/event.h" 30 #include "util/util.h" 31 32 #include <linux/kernel.h> 33 #include <linux/log2.h> 34 #include <linux/zalloc.h> 35 #include <sys/prctl.h> 36 #include <sys/resource.h> 37 #include <inttypes.h> 38 39 #include <errno.h> 40 #include <semaphore.h> 41 #include <pthread.h> 42 #include <math.h> 43 #include <api/fs/fs.h> 44 #include <perf/cpumap.h> 45 #include <linux/time64.h> 46 #include <linux/err.h> 47 48 #include <linux/ctype.h> 49 50 #define PR_SET_NAME 15 /* Set process name */ 51 #define MAX_CPUS 4096 52 #define COMM_LEN 20 53 #define SYM_LEN 129 54 #define MAX_PID 1024000 55 #define MAX_PRIO 140 56 57 static const char *cpu_list; 58 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); 59 60 struct sched_atom; 61 62 struct task_desc { 63 unsigned long nr; 64 unsigned long pid; 65 char comm[COMM_LEN]; 66 67 unsigned long nr_events; 68 unsigned long curr_event; 69 struct sched_atom **atoms; 70 71 pthread_t thread; 72 73 sem_t ready_for_work; 74 sem_t work_done_sem; 75 76 u64 cpu_usage; 77 }; 78 79 enum sched_event_type { 80 SCHED_EVENT_RUN, 81 SCHED_EVENT_SLEEP, 82 SCHED_EVENT_WAKEUP, 83 }; 84 85 struct sched_atom { 86 enum sched_event_type type; 87 u64 timestamp; 88 u64 duration; 89 unsigned long nr; 90 sem_t *wait_sem; 91 struct task_desc *wakee; 92 }; 93 94 enum thread_state { 95 THREAD_SLEEPING = 0, 96 THREAD_WAIT_CPU, 97 THREAD_SCHED_IN, 98 THREAD_IGNORE 99 }; 100 101 struct work_atom { 102 struct list_head list; 103 enum thread_state state; 104 u64 sched_out_time; 105 u64 wake_up_time; 106 u64 sched_in_time; 107 u64 runtime; 108 }; 109 110 struct work_atoms { 111 struct list_head work_list; 112 struct thread *thread; 113 struct rb_node node; 114 u64 max_lat; 115 u64 max_lat_start; 116 u64 max_lat_end; 117 u64 total_lat; 118 u64 nb_atoms; 119 u64 total_runtime; 120 int num_merged; 121 }; 122 123 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *); 124 125 struct perf_sched; 126 127 struct trace_sched_handler { 128 int (*switch_event)(struct perf_sched *sched, struct evsel *evsel, 129 struct perf_sample *sample, struct machine *machine); 130 131 int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel, 132 struct perf_sample *sample, struct machine *machine); 133 134 int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel, 135 struct perf_sample *sample, struct machine *machine); 136 137 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */ 138 int (*fork_event)(struct perf_sched *sched, union perf_event *event, 139 struct machine *machine); 140 141 int (*migrate_task_event)(struct perf_sched *sched, 142 struct evsel *evsel, 143 struct perf_sample *sample, 144 struct machine *machine); 145 }; 146 147 #define COLOR_PIDS PERF_COLOR_BLUE 148 #define COLOR_CPUS PERF_COLOR_BG_RED 149 150 struct perf_sched_map { 151 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS); 152 struct perf_cpu *comp_cpus; 153 bool comp; 154 struct perf_thread_map *color_pids; 155 const char *color_pids_str; 156 struct perf_cpu_map *color_cpus; 157 const char *color_cpus_str; 158 const char *task_name; 159 struct strlist *task_names; 160 bool fuzzy; 161 struct perf_cpu_map *cpus; 162 const char *cpus_str; 163 }; 164 165 struct perf_sched { 166 struct perf_tool tool; 167 const char *sort_order; 168 unsigned long nr_tasks; 169 struct task_desc **pid_to_task; 170 struct task_desc **tasks; 171 const struct trace_sched_handler *tp_handler; 172 struct mutex start_work_mutex; 173 struct mutex work_done_wait_mutex; 174 int profile_cpu; 175 /* 176 * Track the current task - that way we can know whether there's any 177 * weird events, such as a task being switched away that is not current. 178 */ 179 struct perf_cpu max_cpu; 180 u32 *curr_pid; 181 struct thread **curr_thread; 182 struct thread **curr_out_thread; 183 char next_shortname1; 184 char next_shortname2; 185 unsigned int replay_repeat; 186 unsigned long nr_run_events; 187 unsigned long nr_sleep_events; 188 unsigned long nr_wakeup_events; 189 unsigned long nr_sleep_corrections; 190 unsigned long nr_run_events_optimized; 191 unsigned long targetless_wakeups; 192 unsigned long multitarget_wakeups; 193 unsigned long nr_runs; 194 unsigned long nr_timestamps; 195 unsigned long nr_unordered_timestamps; 196 unsigned long nr_context_switch_bugs; 197 unsigned long nr_events; 198 unsigned long nr_lost_chunks; 199 unsigned long nr_lost_events; 200 u64 run_measurement_overhead; 201 u64 sleep_measurement_overhead; 202 u64 start_time; 203 u64 cpu_usage; 204 u64 runavg_cpu_usage; 205 u64 parent_cpu_usage; 206 u64 runavg_parent_cpu_usage; 207 u64 sum_runtime; 208 u64 sum_fluct; 209 u64 run_avg; 210 u64 all_runtime; 211 u64 all_count; 212 u64 *cpu_last_switched; 213 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root; 214 struct list_head sort_list, cmp_pid; 215 bool force; 216 bool skip_merge; 217 struct perf_sched_map map; 218 219 /* options for timehist command */ 220 bool summary; 221 bool summary_only; 222 bool idle_hist; 223 bool show_callchain; 224 unsigned int max_stack; 225 bool show_cpu_visual; 226 bool show_wakeups; 227 bool show_next; 228 bool show_migrations; 229 bool pre_migrations; 230 bool show_state; 231 bool show_prio; 232 u64 skipped_samples; 233 const char *time_str; 234 struct perf_time_interval ptime; 235 struct perf_time_interval hist_time; 236 volatile bool thread_funcs_exit; 237 const char *prio_str; 238 DECLARE_BITMAP(prio_bitmap, MAX_PRIO); 239 }; 240 241 /* per thread run time data */ 242 struct thread_runtime { 243 u64 last_time; /* time of previous sched in/out event */ 244 u64 dt_run; /* run time */ 245 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */ 246 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */ 247 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */ 248 u64 dt_delay; /* time between wakeup and sched-in */ 249 u64 dt_pre_mig; /* time between migration and wakeup */ 250 u64 ready_to_run; /* time of wakeup */ 251 u64 migrated; /* time when a thread is migrated */ 252 253 struct stats run_stats; 254 u64 total_run_time; 255 u64 total_sleep_time; 256 u64 total_iowait_time; 257 u64 total_preempt_time; 258 u64 total_delay_time; 259 u64 total_pre_mig_time; 260 261 char last_state; 262 263 char shortname[3]; 264 bool comm_changed; 265 266 u64 migrations; 267 268 int prio; 269 }; 270 271 /* per event run time data */ 272 struct evsel_runtime { 273 u64 *last_time; /* time this event was last seen per cpu */ 274 u32 ncpu; /* highest cpu slot allocated */ 275 }; 276 277 /* per cpu idle time data */ 278 struct idle_thread_runtime { 279 struct thread_runtime tr; 280 struct thread *last_thread; 281 struct rb_root_cached sorted_root; 282 struct callchain_root callchain; 283 struct callchain_cursor cursor; 284 }; 285 286 /* track idle times per cpu */ 287 static struct thread **idle_threads; 288 static int idle_max_cpu; 289 static char idle_comm[] = "<idle>"; 290 291 static u64 get_nsecs(void) 292 { 293 struct timespec ts; 294 295 clock_gettime(CLOCK_MONOTONIC, &ts); 296 297 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec; 298 } 299 300 static void burn_nsecs(struct perf_sched *sched, u64 nsecs) 301 { 302 u64 T0 = get_nsecs(), T1; 303 304 do { 305 T1 = get_nsecs(); 306 } while (T1 + sched->run_measurement_overhead < T0 + nsecs); 307 } 308 309 static void sleep_nsecs(u64 nsecs) 310 { 311 struct timespec ts; 312 313 ts.tv_nsec = nsecs % 999999999; 314 ts.tv_sec = nsecs / 999999999; 315 316 nanosleep(&ts, NULL); 317 } 318 319 static void calibrate_run_measurement_overhead(struct perf_sched *sched) 320 { 321 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 322 int i; 323 324 for (i = 0; i < 10; i++) { 325 T0 = get_nsecs(); 326 burn_nsecs(sched, 0); 327 T1 = get_nsecs(); 328 delta = T1-T0; 329 min_delta = min(min_delta, delta); 330 } 331 sched->run_measurement_overhead = min_delta; 332 333 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta); 334 } 335 336 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched) 337 { 338 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 339 int i; 340 341 for (i = 0; i < 10; i++) { 342 T0 = get_nsecs(); 343 sleep_nsecs(10000); 344 T1 = get_nsecs(); 345 delta = T1-T0; 346 min_delta = min(min_delta, delta); 347 } 348 min_delta -= 10000; 349 sched->sleep_measurement_overhead = min_delta; 350 351 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta); 352 } 353 354 static struct sched_atom * 355 get_new_event(struct task_desc *task, u64 timestamp) 356 { 357 struct sched_atom *event = zalloc(sizeof(*event)); 358 unsigned long idx = task->nr_events; 359 size_t size; 360 361 event->timestamp = timestamp; 362 event->nr = idx; 363 364 task->nr_events++; 365 size = sizeof(struct sched_atom *) * task->nr_events; 366 task->atoms = realloc(task->atoms, size); 367 BUG_ON(!task->atoms); 368 369 task->atoms[idx] = event; 370 371 return event; 372 } 373 374 static struct sched_atom *last_event(struct task_desc *task) 375 { 376 if (!task->nr_events) 377 return NULL; 378 379 return task->atoms[task->nr_events - 1]; 380 } 381 382 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task, 383 u64 timestamp, u64 duration) 384 { 385 struct sched_atom *event, *curr_event = last_event(task); 386 387 /* 388 * optimize an existing RUN event by merging this one 389 * to it: 390 */ 391 if (curr_event && curr_event->type == SCHED_EVENT_RUN) { 392 sched->nr_run_events_optimized++; 393 curr_event->duration += duration; 394 return; 395 } 396 397 event = get_new_event(task, timestamp); 398 399 event->type = SCHED_EVENT_RUN; 400 event->duration = duration; 401 402 sched->nr_run_events++; 403 } 404 405 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task, 406 u64 timestamp, struct task_desc *wakee) 407 { 408 struct sched_atom *event, *wakee_event; 409 410 event = get_new_event(task, timestamp); 411 event->type = SCHED_EVENT_WAKEUP; 412 event->wakee = wakee; 413 414 wakee_event = last_event(wakee); 415 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { 416 sched->targetless_wakeups++; 417 return; 418 } 419 if (wakee_event->wait_sem) { 420 sched->multitarget_wakeups++; 421 return; 422 } 423 424 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem)); 425 sem_init(wakee_event->wait_sem, 0, 0); 426 event->wait_sem = wakee_event->wait_sem; 427 428 sched->nr_wakeup_events++; 429 } 430 431 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task, 432 u64 timestamp) 433 { 434 struct sched_atom *event = get_new_event(task, timestamp); 435 436 event->type = SCHED_EVENT_SLEEP; 437 438 sched->nr_sleep_events++; 439 } 440 441 static struct task_desc *register_pid(struct perf_sched *sched, 442 unsigned long pid, const char *comm) 443 { 444 struct task_desc *task; 445 static int pid_max; 446 447 if (sched->pid_to_task == NULL) { 448 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0) 449 pid_max = MAX_PID; 450 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL); 451 } 452 if (pid >= (unsigned long)pid_max) { 453 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) * 454 sizeof(struct task_desc *))) == NULL); 455 while (pid >= (unsigned long)pid_max) 456 sched->pid_to_task[pid_max++] = NULL; 457 } 458 459 task = sched->pid_to_task[pid]; 460 461 if (task) 462 return task; 463 464 task = zalloc(sizeof(*task)); 465 task->pid = pid; 466 task->nr = sched->nr_tasks; 467 strcpy(task->comm, comm); 468 /* 469 * every task starts in sleeping state - this gets ignored 470 * if there's no wakeup pointing to this sleep state: 471 */ 472 add_sched_event_sleep(sched, task, 0); 473 474 sched->pid_to_task[pid] = task; 475 sched->nr_tasks++; 476 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *)); 477 BUG_ON(!sched->tasks); 478 sched->tasks[task->nr] = task; 479 480 if (verbose > 0) 481 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm); 482 483 return task; 484 } 485 486 487 static void print_task_traces(struct perf_sched *sched) 488 { 489 struct task_desc *task; 490 unsigned long i; 491 492 for (i = 0; i < sched->nr_tasks; i++) { 493 task = sched->tasks[i]; 494 printf("task %6ld (%20s:%10ld), nr_events: %ld\n", 495 task->nr, task->comm, task->pid, task->nr_events); 496 } 497 } 498 499 static void add_cross_task_wakeups(struct perf_sched *sched) 500 { 501 struct task_desc *task1, *task2; 502 unsigned long i, j; 503 504 for (i = 0; i < sched->nr_tasks; i++) { 505 task1 = sched->tasks[i]; 506 j = i + 1; 507 if (j == sched->nr_tasks) 508 j = 0; 509 task2 = sched->tasks[j]; 510 add_sched_event_wakeup(sched, task1, 0, task2); 511 } 512 } 513 514 static void perf_sched__process_event(struct perf_sched *sched, 515 struct sched_atom *atom) 516 { 517 int ret = 0; 518 519 switch (atom->type) { 520 case SCHED_EVENT_RUN: 521 burn_nsecs(sched, atom->duration); 522 break; 523 case SCHED_EVENT_SLEEP: 524 if (atom->wait_sem) 525 ret = sem_wait(atom->wait_sem); 526 BUG_ON(ret); 527 break; 528 case SCHED_EVENT_WAKEUP: 529 if (atom->wait_sem) 530 ret = sem_post(atom->wait_sem); 531 BUG_ON(ret); 532 break; 533 default: 534 BUG_ON(1); 535 } 536 } 537 538 static u64 get_cpu_usage_nsec_parent(void) 539 { 540 struct rusage ru; 541 u64 sum; 542 int err; 543 544 err = getrusage(RUSAGE_SELF, &ru); 545 BUG_ON(err); 546 547 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC; 548 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC; 549 550 return sum; 551 } 552 553 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task) 554 { 555 struct perf_event_attr attr; 556 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE]; 557 int fd; 558 struct rlimit limit; 559 bool need_privilege = false; 560 561 memset(&attr, 0, sizeof(attr)); 562 563 attr.type = PERF_TYPE_SOFTWARE; 564 attr.config = PERF_COUNT_SW_TASK_CLOCK; 565 566 force_again: 567 fd = sys_perf_event_open(&attr, 0, -1, -1, 568 perf_event_open_cloexec_flag()); 569 570 if (fd < 0) { 571 if (errno == EMFILE) { 572 if (sched->force) { 573 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1); 574 limit.rlim_cur += sched->nr_tasks - cur_task; 575 if (limit.rlim_cur > limit.rlim_max) { 576 limit.rlim_max = limit.rlim_cur; 577 need_privilege = true; 578 } 579 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) { 580 if (need_privilege && errno == EPERM) 581 strcpy(info, "Need privilege\n"); 582 } else 583 goto force_again; 584 } else 585 strcpy(info, "Have a try with -f option\n"); 586 } 587 pr_err("Error: sys_perf_event_open() syscall returned " 588 "with %d (%s)\n%s", fd, 589 str_error_r(errno, sbuf, sizeof(sbuf)), info); 590 exit(EXIT_FAILURE); 591 } 592 return fd; 593 } 594 595 static u64 get_cpu_usage_nsec_self(int fd) 596 { 597 u64 runtime; 598 int ret; 599 600 ret = read(fd, &runtime, sizeof(runtime)); 601 BUG_ON(ret != sizeof(runtime)); 602 603 return runtime; 604 } 605 606 struct sched_thread_parms { 607 struct task_desc *task; 608 struct perf_sched *sched; 609 int fd; 610 }; 611 612 static void *thread_func(void *ctx) 613 { 614 struct sched_thread_parms *parms = ctx; 615 struct task_desc *this_task = parms->task; 616 struct perf_sched *sched = parms->sched; 617 u64 cpu_usage_0, cpu_usage_1; 618 unsigned long i, ret; 619 char comm2[22]; 620 int fd = parms->fd; 621 622 zfree(&parms); 623 624 sprintf(comm2, ":%s", this_task->comm); 625 prctl(PR_SET_NAME, comm2); 626 if (fd < 0) 627 return NULL; 628 629 while (!sched->thread_funcs_exit) { 630 ret = sem_post(&this_task->ready_for_work); 631 BUG_ON(ret); 632 mutex_lock(&sched->start_work_mutex); 633 mutex_unlock(&sched->start_work_mutex); 634 635 cpu_usage_0 = get_cpu_usage_nsec_self(fd); 636 637 for (i = 0; i < this_task->nr_events; i++) { 638 this_task->curr_event = i; 639 perf_sched__process_event(sched, this_task->atoms[i]); 640 } 641 642 cpu_usage_1 = get_cpu_usage_nsec_self(fd); 643 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; 644 ret = sem_post(&this_task->work_done_sem); 645 BUG_ON(ret); 646 647 mutex_lock(&sched->work_done_wait_mutex); 648 mutex_unlock(&sched->work_done_wait_mutex); 649 } 650 return NULL; 651 } 652 653 static void create_tasks(struct perf_sched *sched) 654 EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex) 655 EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex) 656 { 657 struct task_desc *task; 658 pthread_attr_t attr; 659 unsigned long i; 660 int err; 661 662 err = pthread_attr_init(&attr); 663 BUG_ON(err); 664 err = pthread_attr_setstacksize(&attr, 665 (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN)); 666 BUG_ON(err); 667 mutex_lock(&sched->start_work_mutex); 668 mutex_lock(&sched->work_done_wait_mutex); 669 for (i = 0; i < sched->nr_tasks; i++) { 670 struct sched_thread_parms *parms = malloc(sizeof(*parms)); 671 BUG_ON(parms == NULL); 672 parms->task = task = sched->tasks[i]; 673 parms->sched = sched; 674 parms->fd = self_open_counters(sched, i); 675 sem_init(&task->ready_for_work, 0, 0); 676 sem_init(&task->work_done_sem, 0, 0); 677 task->curr_event = 0; 678 err = pthread_create(&task->thread, &attr, thread_func, parms); 679 BUG_ON(err); 680 } 681 } 682 683 static void destroy_tasks(struct perf_sched *sched) 684 UNLOCK_FUNCTION(sched->start_work_mutex) 685 UNLOCK_FUNCTION(sched->work_done_wait_mutex) 686 { 687 struct task_desc *task; 688 unsigned long i; 689 int err; 690 691 mutex_unlock(&sched->start_work_mutex); 692 mutex_unlock(&sched->work_done_wait_mutex); 693 /* Get rid of threads so they won't be upset by mutex destrunction */ 694 for (i = 0; i < sched->nr_tasks; i++) { 695 task = sched->tasks[i]; 696 err = pthread_join(task->thread, NULL); 697 BUG_ON(err); 698 sem_destroy(&task->ready_for_work); 699 sem_destroy(&task->work_done_sem); 700 } 701 } 702 703 static void wait_for_tasks(struct perf_sched *sched) 704 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex) 705 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex) 706 { 707 u64 cpu_usage_0, cpu_usage_1; 708 struct task_desc *task; 709 unsigned long i, ret; 710 711 sched->start_time = get_nsecs(); 712 sched->cpu_usage = 0; 713 mutex_unlock(&sched->work_done_wait_mutex); 714 715 for (i = 0; i < sched->nr_tasks; i++) { 716 task = sched->tasks[i]; 717 ret = sem_wait(&task->ready_for_work); 718 BUG_ON(ret); 719 sem_init(&task->ready_for_work, 0, 0); 720 } 721 mutex_lock(&sched->work_done_wait_mutex); 722 723 cpu_usage_0 = get_cpu_usage_nsec_parent(); 724 725 mutex_unlock(&sched->start_work_mutex); 726 727 for (i = 0; i < sched->nr_tasks; i++) { 728 task = sched->tasks[i]; 729 ret = sem_wait(&task->work_done_sem); 730 BUG_ON(ret); 731 sem_init(&task->work_done_sem, 0, 0); 732 sched->cpu_usage += task->cpu_usage; 733 task->cpu_usage = 0; 734 } 735 736 cpu_usage_1 = get_cpu_usage_nsec_parent(); 737 if (!sched->runavg_cpu_usage) 738 sched->runavg_cpu_usage = sched->cpu_usage; 739 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat; 740 741 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0; 742 if (!sched->runavg_parent_cpu_usage) 743 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage; 744 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) + 745 sched->parent_cpu_usage)/sched->replay_repeat; 746 747 mutex_lock(&sched->start_work_mutex); 748 749 for (i = 0; i < sched->nr_tasks; i++) { 750 task = sched->tasks[i]; 751 task->curr_event = 0; 752 } 753 } 754 755 static void run_one_test(struct perf_sched *sched) 756 EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex) 757 EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex) 758 { 759 u64 T0, T1, delta, avg_delta, fluct; 760 761 T0 = get_nsecs(); 762 wait_for_tasks(sched); 763 T1 = get_nsecs(); 764 765 delta = T1 - T0; 766 sched->sum_runtime += delta; 767 sched->nr_runs++; 768 769 avg_delta = sched->sum_runtime / sched->nr_runs; 770 if (delta < avg_delta) 771 fluct = avg_delta - delta; 772 else 773 fluct = delta - avg_delta; 774 sched->sum_fluct += fluct; 775 if (!sched->run_avg) 776 sched->run_avg = delta; 777 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat; 778 779 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC); 780 781 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC); 782 783 printf("cpu: %0.2f / %0.2f", 784 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC); 785 786 #if 0 787 /* 788 * rusage statistics done by the parent, these are less 789 * accurate than the sched->sum_exec_runtime based statistics: 790 */ 791 printf(" [%0.2f / %0.2f]", 792 (double)sched->parent_cpu_usage / NSEC_PER_MSEC, 793 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC); 794 #endif 795 796 printf("\n"); 797 798 if (sched->nr_sleep_corrections) 799 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections); 800 sched->nr_sleep_corrections = 0; 801 } 802 803 static void test_calibrations(struct perf_sched *sched) 804 { 805 u64 T0, T1; 806 807 T0 = get_nsecs(); 808 burn_nsecs(sched, NSEC_PER_MSEC); 809 T1 = get_nsecs(); 810 811 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0); 812 813 T0 = get_nsecs(); 814 sleep_nsecs(NSEC_PER_MSEC); 815 T1 = get_nsecs(); 816 817 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0); 818 } 819 820 static int 821 replay_wakeup_event(struct perf_sched *sched, 822 struct evsel *evsel, struct perf_sample *sample, 823 struct machine *machine __maybe_unused) 824 { 825 const char *comm = evsel__strval(evsel, sample, "comm"); 826 const u32 pid = evsel__intval(evsel, sample, "pid"); 827 struct task_desc *waker, *wakee; 828 829 if (verbose > 0) { 830 printf("sched_wakeup event %p\n", evsel); 831 832 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid); 833 } 834 835 waker = register_pid(sched, sample->tid, "<unknown>"); 836 wakee = register_pid(sched, pid, comm); 837 838 add_sched_event_wakeup(sched, waker, sample->time, wakee); 839 return 0; 840 } 841 842 static int replay_switch_event(struct perf_sched *sched, 843 struct evsel *evsel, 844 struct perf_sample *sample, 845 struct machine *machine __maybe_unused) 846 { 847 const char *prev_comm = evsel__strval(evsel, sample, "prev_comm"), 848 *next_comm = evsel__strval(evsel, sample, "next_comm"); 849 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"), 850 next_pid = evsel__intval(evsel, sample, "next_pid"); 851 struct task_desc *prev, __maybe_unused *next; 852 u64 timestamp0, timestamp = sample->time; 853 int cpu = sample->cpu; 854 s64 delta; 855 856 if (verbose > 0) 857 printf("sched_switch event %p\n", evsel); 858 859 if (cpu >= MAX_CPUS || cpu < 0) 860 return 0; 861 862 timestamp0 = sched->cpu_last_switched[cpu]; 863 if (timestamp0) 864 delta = timestamp - timestamp0; 865 else 866 delta = 0; 867 868 if (delta < 0) { 869 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 870 return -1; 871 } 872 873 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n", 874 prev_comm, prev_pid, next_comm, next_pid, delta); 875 876 prev = register_pid(sched, prev_pid, prev_comm); 877 next = register_pid(sched, next_pid, next_comm); 878 879 sched->cpu_last_switched[cpu] = timestamp; 880 881 add_sched_event_run(sched, prev, timestamp, delta); 882 add_sched_event_sleep(sched, prev, timestamp); 883 884 return 0; 885 } 886 887 static int replay_fork_event(struct perf_sched *sched, 888 union perf_event *event, 889 struct machine *machine) 890 { 891 struct thread *child, *parent; 892 893 child = machine__findnew_thread(machine, event->fork.pid, 894 event->fork.tid); 895 parent = machine__findnew_thread(machine, event->fork.ppid, 896 event->fork.ptid); 897 898 if (child == NULL || parent == NULL) { 899 pr_debug("thread does not exist on fork event: child %p, parent %p\n", 900 child, parent); 901 goto out_put; 902 } 903 904 if (verbose > 0) { 905 printf("fork event\n"); 906 printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent)); 907 printf("... child: %s/%d\n", thread__comm_str(child), thread__tid(child)); 908 } 909 910 register_pid(sched, thread__tid(parent), thread__comm_str(parent)); 911 register_pid(sched, thread__tid(child), thread__comm_str(child)); 912 out_put: 913 thread__put(child); 914 thread__put(parent); 915 return 0; 916 } 917 918 struct sort_dimension { 919 const char *name; 920 sort_fn_t cmp; 921 struct list_head list; 922 }; 923 924 static inline void init_prio(struct thread_runtime *r) 925 { 926 r->prio = -1; 927 } 928 929 /* 930 * handle runtime stats saved per thread 931 */ 932 static struct thread_runtime *thread__init_runtime(struct thread *thread) 933 { 934 struct thread_runtime *r; 935 936 r = zalloc(sizeof(struct thread_runtime)); 937 if (!r) 938 return NULL; 939 940 init_stats(&r->run_stats); 941 init_prio(r); 942 thread__set_priv(thread, r); 943 944 return r; 945 } 946 947 static struct thread_runtime *thread__get_runtime(struct thread *thread) 948 { 949 struct thread_runtime *tr; 950 951 tr = thread__priv(thread); 952 if (tr == NULL) { 953 tr = thread__init_runtime(thread); 954 if (tr == NULL) 955 pr_debug("Failed to malloc memory for runtime data.\n"); 956 } 957 958 return tr; 959 } 960 961 static int 962 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r) 963 { 964 struct sort_dimension *sort; 965 int ret = 0; 966 967 BUG_ON(list_empty(list)); 968 969 list_for_each_entry(sort, list, list) { 970 ret = sort->cmp(l, r); 971 if (ret) 972 return ret; 973 } 974 975 return ret; 976 } 977 978 static struct work_atoms * 979 thread_atoms_search(struct rb_root_cached *root, struct thread *thread, 980 struct list_head *sort_list) 981 { 982 struct rb_node *node = root->rb_root.rb_node; 983 struct work_atoms key = { .thread = thread }; 984 985 while (node) { 986 struct work_atoms *atoms; 987 int cmp; 988 989 atoms = container_of(node, struct work_atoms, node); 990 991 cmp = thread_lat_cmp(sort_list, &key, atoms); 992 if (cmp > 0) 993 node = node->rb_left; 994 else if (cmp < 0) 995 node = node->rb_right; 996 else { 997 BUG_ON(!RC_CHK_EQUAL(thread, atoms->thread)); 998 return atoms; 999 } 1000 } 1001 return NULL; 1002 } 1003 1004 static void 1005 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data, 1006 struct list_head *sort_list) 1007 { 1008 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; 1009 bool leftmost = true; 1010 1011 while (*new) { 1012 struct work_atoms *this; 1013 int cmp; 1014 1015 this = container_of(*new, struct work_atoms, node); 1016 parent = *new; 1017 1018 cmp = thread_lat_cmp(sort_list, data, this); 1019 1020 if (cmp > 0) 1021 new = &((*new)->rb_left); 1022 else { 1023 new = &((*new)->rb_right); 1024 leftmost = false; 1025 } 1026 } 1027 1028 rb_link_node(&data->node, parent, new); 1029 rb_insert_color_cached(&data->node, root, leftmost); 1030 } 1031 1032 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread) 1033 { 1034 struct work_atoms *atoms = zalloc(sizeof(*atoms)); 1035 if (!atoms) { 1036 pr_err("No memory at %s\n", __func__); 1037 return -1; 1038 } 1039 1040 atoms->thread = thread__get(thread); 1041 INIT_LIST_HEAD(&atoms->work_list); 1042 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid); 1043 return 0; 1044 } 1045 1046 static int 1047 add_sched_out_event(struct work_atoms *atoms, 1048 char run_state, 1049 u64 timestamp) 1050 { 1051 struct work_atom *atom = zalloc(sizeof(*atom)); 1052 if (!atom) { 1053 pr_err("Non memory at %s", __func__); 1054 return -1; 1055 } 1056 1057 atom->sched_out_time = timestamp; 1058 1059 if (run_state == 'R') { 1060 atom->state = THREAD_WAIT_CPU; 1061 atom->wake_up_time = atom->sched_out_time; 1062 } 1063 1064 list_add_tail(&atom->list, &atoms->work_list); 1065 return 0; 1066 } 1067 1068 static void 1069 add_runtime_event(struct work_atoms *atoms, u64 delta, 1070 u64 timestamp __maybe_unused) 1071 { 1072 struct work_atom *atom; 1073 1074 BUG_ON(list_empty(&atoms->work_list)); 1075 1076 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1077 1078 atom->runtime += delta; 1079 atoms->total_runtime += delta; 1080 } 1081 1082 static void 1083 add_sched_in_event(struct work_atoms *atoms, u64 timestamp) 1084 { 1085 struct work_atom *atom; 1086 u64 delta; 1087 1088 if (list_empty(&atoms->work_list)) 1089 return; 1090 1091 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1092 1093 if (atom->state != THREAD_WAIT_CPU) 1094 return; 1095 1096 if (timestamp < atom->wake_up_time) { 1097 atom->state = THREAD_IGNORE; 1098 return; 1099 } 1100 1101 atom->state = THREAD_SCHED_IN; 1102 atom->sched_in_time = timestamp; 1103 1104 delta = atom->sched_in_time - atom->wake_up_time; 1105 atoms->total_lat += delta; 1106 if (delta > atoms->max_lat) { 1107 atoms->max_lat = delta; 1108 atoms->max_lat_start = atom->wake_up_time; 1109 atoms->max_lat_end = timestamp; 1110 } 1111 atoms->nb_atoms++; 1112 } 1113 1114 static void free_work_atoms(struct work_atoms *atoms) 1115 { 1116 struct work_atom *atom, *tmp; 1117 1118 if (atoms == NULL) 1119 return; 1120 1121 list_for_each_entry_safe(atom, tmp, &atoms->work_list, list) { 1122 list_del(&atom->list); 1123 free(atom); 1124 } 1125 thread__zput(atoms->thread); 1126 free(atoms); 1127 } 1128 1129 static int latency_switch_event(struct perf_sched *sched, 1130 struct evsel *evsel, 1131 struct perf_sample *sample, 1132 struct machine *machine) 1133 { 1134 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"), 1135 next_pid = evsel__intval(evsel, sample, "next_pid"); 1136 const char prev_state = evsel__taskstate(evsel, sample, "prev_state"); 1137 struct work_atoms *out_events, *in_events; 1138 struct thread *sched_out, *sched_in; 1139 u64 timestamp0, timestamp = sample->time; 1140 int cpu = sample->cpu, err = -1; 1141 s64 delta; 1142 1143 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1144 1145 timestamp0 = sched->cpu_last_switched[cpu]; 1146 sched->cpu_last_switched[cpu] = timestamp; 1147 if (timestamp0) 1148 delta = timestamp - timestamp0; 1149 else 1150 delta = 0; 1151 1152 if (delta < 0) { 1153 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1154 return -1; 1155 } 1156 1157 sched_out = machine__findnew_thread(machine, -1, prev_pid); 1158 sched_in = machine__findnew_thread(machine, -1, next_pid); 1159 if (sched_out == NULL || sched_in == NULL) 1160 goto out_put; 1161 1162 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1163 if (!out_events) { 1164 if (thread_atoms_insert(sched, sched_out)) 1165 goto out_put; 1166 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1167 if (!out_events) { 1168 pr_err("out-event: Internal tree error"); 1169 goto out_put; 1170 } 1171 } 1172 if (add_sched_out_event(out_events, prev_state, timestamp)) 1173 return -1; 1174 1175 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1176 if (!in_events) { 1177 if (thread_atoms_insert(sched, sched_in)) 1178 goto out_put; 1179 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1180 if (!in_events) { 1181 pr_err("in-event: Internal tree error"); 1182 goto out_put; 1183 } 1184 /* 1185 * Take came in we have not heard about yet, 1186 * add in an initial atom in runnable state: 1187 */ 1188 if (add_sched_out_event(in_events, 'R', timestamp)) 1189 goto out_put; 1190 } 1191 add_sched_in_event(in_events, timestamp); 1192 err = 0; 1193 out_put: 1194 thread__put(sched_out); 1195 thread__put(sched_in); 1196 return err; 1197 } 1198 1199 static int latency_runtime_event(struct perf_sched *sched, 1200 struct evsel *evsel, 1201 struct perf_sample *sample, 1202 struct machine *machine) 1203 { 1204 const u32 pid = evsel__intval(evsel, sample, "pid"); 1205 const u64 runtime = evsel__intval(evsel, sample, "runtime"); 1206 struct thread *thread = machine__findnew_thread(machine, -1, pid); 1207 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1208 u64 timestamp = sample->time; 1209 int cpu = sample->cpu, err = -1; 1210 1211 if (thread == NULL) 1212 return -1; 1213 1214 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1215 if (!atoms) { 1216 if (thread_atoms_insert(sched, thread)) 1217 goto out_put; 1218 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1219 if (!atoms) { 1220 pr_err("in-event: Internal tree error"); 1221 goto out_put; 1222 } 1223 if (add_sched_out_event(atoms, 'R', timestamp)) 1224 goto out_put; 1225 } 1226 1227 add_runtime_event(atoms, runtime, timestamp); 1228 err = 0; 1229 out_put: 1230 thread__put(thread); 1231 return err; 1232 } 1233 1234 static int latency_wakeup_event(struct perf_sched *sched, 1235 struct evsel *evsel, 1236 struct perf_sample *sample, 1237 struct machine *machine) 1238 { 1239 const u32 pid = evsel__intval(evsel, sample, "pid"); 1240 struct work_atoms *atoms; 1241 struct work_atom *atom; 1242 struct thread *wakee; 1243 u64 timestamp = sample->time; 1244 int err = -1; 1245 1246 wakee = machine__findnew_thread(machine, -1, pid); 1247 if (wakee == NULL) 1248 return -1; 1249 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1250 if (!atoms) { 1251 if (thread_atoms_insert(sched, wakee)) 1252 goto out_put; 1253 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1254 if (!atoms) { 1255 pr_err("wakeup-event: Internal tree error"); 1256 goto out_put; 1257 } 1258 if (add_sched_out_event(atoms, 'S', timestamp)) 1259 goto out_put; 1260 } 1261 1262 BUG_ON(list_empty(&atoms->work_list)); 1263 1264 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1265 1266 /* 1267 * As we do not guarantee the wakeup event happens when 1268 * task is out of run queue, also may happen when task is 1269 * on run queue and wakeup only change ->state to TASK_RUNNING, 1270 * then we should not set the ->wake_up_time when wake up a 1271 * task which is on run queue. 1272 * 1273 * You WILL be missing events if you've recorded only 1274 * one CPU, or are only looking at only one, so don't 1275 * skip in this case. 1276 */ 1277 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING) 1278 goto out_ok; 1279 1280 sched->nr_timestamps++; 1281 if (atom->sched_out_time > timestamp) { 1282 sched->nr_unordered_timestamps++; 1283 goto out_ok; 1284 } 1285 1286 atom->state = THREAD_WAIT_CPU; 1287 atom->wake_up_time = timestamp; 1288 out_ok: 1289 err = 0; 1290 out_put: 1291 thread__put(wakee); 1292 return err; 1293 } 1294 1295 static int latency_migrate_task_event(struct perf_sched *sched, 1296 struct evsel *evsel, 1297 struct perf_sample *sample, 1298 struct machine *machine) 1299 { 1300 const u32 pid = evsel__intval(evsel, sample, "pid"); 1301 u64 timestamp = sample->time; 1302 struct work_atoms *atoms; 1303 struct work_atom *atom; 1304 struct thread *migrant; 1305 int err = -1; 1306 1307 /* 1308 * Only need to worry about migration when profiling one CPU. 1309 */ 1310 if (sched->profile_cpu == -1) 1311 return 0; 1312 1313 migrant = machine__findnew_thread(machine, -1, pid); 1314 if (migrant == NULL) 1315 return -1; 1316 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1317 if (!atoms) { 1318 if (thread_atoms_insert(sched, migrant)) 1319 goto out_put; 1320 register_pid(sched, thread__tid(migrant), thread__comm_str(migrant)); 1321 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1322 if (!atoms) { 1323 pr_err("migration-event: Internal tree error"); 1324 goto out_put; 1325 } 1326 if (add_sched_out_event(atoms, 'R', timestamp)) 1327 goto out_put; 1328 } 1329 1330 BUG_ON(list_empty(&atoms->work_list)); 1331 1332 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1333 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp; 1334 1335 sched->nr_timestamps++; 1336 1337 if (atom->sched_out_time > timestamp) 1338 sched->nr_unordered_timestamps++; 1339 err = 0; 1340 out_put: 1341 thread__put(migrant); 1342 return err; 1343 } 1344 1345 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list) 1346 { 1347 int i; 1348 int ret; 1349 u64 avg; 1350 char max_lat_start[32], max_lat_end[32]; 1351 1352 if (!work_list->nb_atoms) 1353 return; 1354 /* 1355 * Ignore idle threads: 1356 */ 1357 if (!strcmp(thread__comm_str(work_list->thread), "swapper")) 1358 return; 1359 1360 sched->all_runtime += work_list->total_runtime; 1361 sched->all_count += work_list->nb_atoms; 1362 1363 if (work_list->num_merged > 1) { 1364 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), 1365 work_list->num_merged); 1366 } else { 1367 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), 1368 thread__tid(work_list->thread)); 1369 } 1370 1371 for (i = 0; i < 24 - ret; i++) 1372 printf(" "); 1373 1374 avg = work_list->total_lat / work_list->nb_atoms; 1375 timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start)); 1376 timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end)); 1377 1378 printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n", 1379 (double)work_list->total_runtime / NSEC_PER_MSEC, 1380 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC, 1381 (double)work_list->max_lat / NSEC_PER_MSEC, 1382 max_lat_start, max_lat_end); 1383 } 1384 1385 static int pid_cmp(struct work_atoms *l, struct work_atoms *r) 1386 { 1387 pid_t l_tid, r_tid; 1388 1389 if (RC_CHK_EQUAL(l->thread, r->thread)) 1390 return 0; 1391 l_tid = thread__tid(l->thread); 1392 r_tid = thread__tid(r->thread); 1393 if (l_tid < r_tid) 1394 return -1; 1395 if (l_tid > r_tid) 1396 return 1; 1397 return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread)); 1398 } 1399 1400 static int avg_cmp(struct work_atoms *l, struct work_atoms *r) 1401 { 1402 u64 avgl, avgr; 1403 1404 if (!l->nb_atoms) 1405 return -1; 1406 1407 if (!r->nb_atoms) 1408 return 1; 1409 1410 avgl = l->total_lat / l->nb_atoms; 1411 avgr = r->total_lat / r->nb_atoms; 1412 1413 if (avgl < avgr) 1414 return -1; 1415 if (avgl > avgr) 1416 return 1; 1417 1418 return 0; 1419 } 1420 1421 static int max_cmp(struct work_atoms *l, struct work_atoms *r) 1422 { 1423 if (l->max_lat < r->max_lat) 1424 return -1; 1425 if (l->max_lat > r->max_lat) 1426 return 1; 1427 1428 return 0; 1429 } 1430 1431 static int switch_cmp(struct work_atoms *l, struct work_atoms *r) 1432 { 1433 if (l->nb_atoms < r->nb_atoms) 1434 return -1; 1435 if (l->nb_atoms > r->nb_atoms) 1436 return 1; 1437 1438 return 0; 1439 } 1440 1441 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r) 1442 { 1443 if (l->total_runtime < r->total_runtime) 1444 return -1; 1445 if (l->total_runtime > r->total_runtime) 1446 return 1; 1447 1448 return 0; 1449 } 1450 1451 static int sort_dimension__add(const char *tok, struct list_head *list) 1452 { 1453 size_t i; 1454 static struct sort_dimension avg_sort_dimension = { 1455 .name = "avg", 1456 .cmp = avg_cmp, 1457 }; 1458 static struct sort_dimension max_sort_dimension = { 1459 .name = "max", 1460 .cmp = max_cmp, 1461 }; 1462 static struct sort_dimension pid_sort_dimension = { 1463 .name = "pid", 1464 .cmp = pid_cmp, 1465 }; 1466 static struct sort_dimension runtime_sort_dimension = { 1467 .name = "runtime", 1468 .cmp = runtime_cmp, 1469 }; 1470 static struct sort_dimension switch_sort_dimension = { 1471 .name = "switch", 1472 .cmp = switch_cmp, 1473 }; 1474 struct sort_dimension *available_sorts[] = { 1475 &pid_sort_dimension, 1476 &avg_sort_dimension, 1477 &max_sort_dimension, 1478 &switch_sort_dimension, 1479 &runtime_sort_dimension, 1480 }; 1481 1482 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { 1483 if (!strcmp(available_sorts[i]->name, tok)) { 1484 list_add_tail(&available_sorts[i]->list, list); 1485 1486 return 0; 1487 } 1488 } 1489 1490 return -1; 1491 } 1492 1493 static void perf_sched__sort_lat(struct perf_sched *sched) 1494 { 1495 struct rb_node *node; 1496 struct rb_root_cached *root = &sched->atom_root; 1497 again: 1498 for (;;) { 1499 struct work_atoms *data; 1500 node = rb_first_cached(root); 1501 if (!node) 1502 break; 1503 1504 rb_erase_cached(node, root); 1505 data = rb_entry(node, struct work_atoms, node); 1506 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list); 1507 } 1508 if (root == &sched->atom_root) { 1509 root = &sched->merged_atom_root; 1510 goto again; 1511 } 1512 } 1513 1514 static int process_sched_wakeup_event(const struct perf_tool *tool, 1515 struct evsel *evsel, 1516 struct perf_sample *sample, 1517 struct machine *machine) 1518 { 1519 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1520 1521 if (sched->tp_handler->wakeup_event) 1522 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine); 1523 1524 return 0; 1525 } 1526 1527 static int process_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused, 1528 struct evsel *evsel __maybe_unused, 1529 struct perf_sample *sample __maybe_unused, 1530 struct machine *machine __maybe_unused) 1531 { 1532 return 0; 1533 } 1534 1535 union map_priv { 1536 void *ptr; 1537 bool color; 1538 }; 1539 1540 static bool thread__has_color(struct thread *thread) 1541 { 1542 union map_priv priv = { 1543 .ptr = thread__priv(thread), 1544 }; 1545 1546 return priv.color; 1547 } 1548 1549 static struct thread* 1550 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid) 1551 { 1552 struct thread *thread = machine__findnew_thread(machine, pid, tid); 1553 union map_priv priv = { 1554 .color = false, 1555 }; 1556 1557 if (!sched->map.color_pids || !thread || thread__priv(thread)) 1558 return thread; 1559 1560 if (thread_map__has(sched->map.color_pids, tid)) 1561 priv.color = true; 1562 1563 thread__set_priv(thread, priv.ptr); 1564 return thread; 1565 } 1566 1567 static bool sched_match_task(struct perf_sched *sched, const char *comm_str) 1568 { 1569 bool fuzzy_match = sched->map.fuzzy; 1570 struct strlist *task_names = sched->map.task_names; 1571 struct str_node *node; 1572 1573 strlist__for_each_entry(node, task_names) { 1574 bool match_found = fuzzy_match ? !!strstr(comm_str, node->s) : 1575 !strcmp(comm_str, node->s); 1576 if (match_found) 1577 return true; 1578 } 1579 1580 return false; 1581 } 1582 1583 static void print_sched_map(struct perf_sched *sched, struct perf_cpu this_cpu, int cpus_nr, 1584 const char *color, bool sched_out) 1585 { 1586 for (int i = 0; i < cpus_nr; i++) { 1587 struct perf_cpu cpu = { 1588 .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i, 1589 }; 1590 struct thread *curr_thread = sched->curr_thread[cpu.cpu]; 1591 struct thread *curr_out_thread = sched->curr_out_thread[cpu.cpu]; 1592 struct thread_runtime *curr_tr; 1593 const char *pid_color = color; 1594 const char *cpu_color = color; 1595 char symbol = ' '; 1596 struct thread *thread_to_check = sched_out ? curr_out_thread : curr_thread; 1597 1598 if (thread_to_check && thread__has_color(thread_to_check)) 1599 pid_color = COLOR_PIDS; 1600 1601 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu)) 1602 cpu_color = COLOR_CPUS; 1603 1604 if (cpu.cpu == this_cpu.cpu) 1605 symbol = '*'; 1606 1607 color_fprintf(stdout, cpu.cpu != this_cpu.cpu ? color : cpu_color, "%c", symbol); 1608 1609 thread_to_check = sched_out ? sched->curr_out_thread[cpu.cpu] : 1610 sched->curr_thread[cpu.cpu]; 1611 1612 if (thread_to_check) { 1613 curr_tr = thread__get_runtime(thread_to_check); 1614 if (curr_tr == NULL) 1615 return; 1616 1617 if (sched_out) { 1618 if (cpu.cpu == this_cpu.cpu) 1619 color_fprintf(stdout, color, "- "); 1620 else { 1621 curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]); 1622 if (curr_tr != NULL) 1623 color_fprintf(stdout, pid_color, "%2s ", 1624 curr_tr->shortname); 1625 } 1626 } else 1627 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname); 1628 } else 1629 color_fprintf(stdout, color, " "); 1630 } 1631 } 1632 1633 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel, 1634 struct perf_sample *sample, struct machine *machine) 1635 { 1636 const u32 next_pid = evsel__intval(evsel, sample, "next_pid"); 1637 const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"); 1638 struct thread *sched_in, *sched_out; 1639 struct thread_runtime *tr; 1640 int new_shortname; 1641 u64 timestamp0, timestamp = sample->time; 1642 s64 delta; 1643 struct perf_cpu this_cpu = { 1644 .cpu = sample->cpu, 1645 }; 1646 int cpus_nr; 1647 int proceed; 1648 bool new_cpu = false; 1649 const char *color = PERF_COLOR_NORMAL; 1650 char stimestamp[32]; 1651 const char *str; 1652 int ret = -1; 1653 1654 BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0); 1655 1656 if (this_cpu.cpu > sched->max_cpu.cpu) 1657 sched->max_cpu = this_cpu; 1658 1659 if (sched->map.comp) { 1660 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS); 1661 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) { 1662 sched->map.comp_cpus[cpus_nr++] = this_cpu; 1663 new_cpu = true; 1664 } 1665 } else 1666 cpus_nr = sched->max_cpu.cpu; 1667 1668 timestamp0 = sched->cpu_last_switched[this_cpu.cpu]; 1669 sched->cpu_last_switched[this_cpu.cpu] = timestamp; 1670 if (timestamp0) 1671 delta = timestamp - timestamp0; 1672 else 1673 delta = 0; 1674 1675 if (delta < 0) { 1676 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1677 return -1; 1678 } 1679 1680 sched_in = map__findnew_thread(sched, machine, -1, next_pid); 1681 sched_out = map__findnew_thread(sched, machine, -1, prev_pid); 1682 if (sched_in == NULL || sched_out == NULL) 1683 goto out; 1684 1685 tr = thread__get_runtime(sched_in); 1686 if (tr == NULL) 1687 goto out; 1688 1689 thread__put(sched->curr_thread[this_cpu.cpu]); 1690 thread__put(sched->curr_out_thread[this_cpu.cpu]); 1691 1692 sched->curr_thread[this_cpu.cpu] = thread__get(sched_in); 1693 sched->curr_out_thread[this_cpu.cpu] = thread__get(sched_out); 1694 1695 ret = 0; 1696 1697 str = thread__comm_str(sched_in); 1698 new_shortname = 0; 1699 if (!tr->shortname[0]) { 1700 if (!strcmp(thread__comm_str(sched_in), "swapper")) { 1701 /* 1702 * Don't allocate a letter-number for swapper:0 1703 * as a shortname. Instead, we use '.' for it. 1704 */ 1705 tr->shortname[0] = '.'; 1706 tr->shortname[1] = ' '; 1707 } else if (!sched->map.task_name || sched_match_task(sched, str)) { 1708 tr->shortname[0] = sched->next_shortname1; 1709 tr->shortname[1] = sched->next_shortname2; 1710 1711 if (sched->next_shortname1 < 'Z') { 1712 sched->next_shortname1++; 1713 } else { 1714 sched->next_shortname1 = 'A'; 1715 if (sched->next_shortname2 < '9') 1716 sched->next_shortname2++; 1717 else 1718 sched->next_shortname2 = '0'; 1719 } 1720 } else { 1721 tr->shortname[0] = '-'; 1722 tr->shortname[1] = ' '; 1723 } 1724 new_shortname = 1; 1725 } 1726 1727 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu)) 1728 goto out; 1729 1730 proceed = 0; 1731 str = thread__comm_str(sched_in); 1732 /* 1733 * Check which of sched_in and sched_out matches the passed --task-name 1734 * arguments and call the corresponding print_sched_map. 1735 */ 1736 if (sched->map.task_name && !sched_match_task(sched, str)) { 1737 if (!sched_match_task(sched, thread__comm_str(sched_out))) 1738 goto out; 1739 else 1740 goto sched_out; 1741 1742 } else { 1743 str = thread__comm_str(sched_out); 1744 if (!(sched->map.task_name && !sched_match_task(sched, str))) 1745 proceed = 1; 1746 } 1747 1748 printf(" "); 1749 1750 print_sched_map(sched, this_cpu, cpus_nr, color, false); 1751 1752 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp)); 1753 color_fprintf(stdout, color, " %12s secs ", stimestamp); 1754 if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) { 1755 const char *pid_color = color; 1756 1757 if (thread__has_color(sched_in)) 1758 pid_color = COLOR_PIDS; 1759 1760 color_fprintf(stdout, pid_color, "%s => %s:%d", 1761 tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in)); 1762 tr->comm_changed = false; 1763 } 1764 1765 if (sched->map.comp && new_cpu) 1766 color_fprintf(stdout, color, " (CPU %d)", this_cpu.cpu); 1767 1768 if (proceed != 1) { 1769 color_fprintf(stdout, color, "\n"); 1770 goto out; 1771 } 1772 1773 sched_out: 1774 if (sched->map.task_name) { 1775 tr = thread__get_runtime(sched->curr_out_thread[this_cpu.cpu]); 1776 if (strcmp(tr->shortname, "") == 0) 1777 goto out; 1778 1779 if (proceed == 1) 1780 color_fprintf(stdout, color, "\n"); 1781 1782 printf(" "); 1783 print_sched_map(sched, this_cpu, cpus_nr, color, true); 1784 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp)); 1785 color_fprintf(stdout, color, " %12s secs ", stimestamp); 1786 } 1787 1788 color_fprintf(stdout, color, "\n"); 1789 1790 out: 1791 thread__put(sched_out); 1792 thread__put(sched_in); 1793 1794 return ret; 1795 } 1796 1797 static int process_sched_switch_event(const struct perf_tool *tool, 1798 struct evsel *evsel, 1799 struct perf_sample *sample, 1800 struct machine *machine) 1801 { 1802 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1803 int this_cpu = sample->cpu, err = 0; 1804 u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"), 1805 next_pid = evsel__intval(evsel, sample, "next_pid"); 1806 1807 if (sched->curr_pid[this_cpu] != (u32)-1) { 1808 /* 1809 * Are we trying to switch away a PID that is 1810 * not current? 1811 */ 1812 if (sched->curr_pid[this_cpu] != prev_pid) 1813 sched->nr_context_switch_bugs++; 1814 } 1815 1816 if (sched->tp_handler->switch_event) 1817 err = sched->tp_handler->switch_event(sched, evsel, sample, machine); 1818 1819 sched->curr_pid[this_cpu] = next_pid; 1820 return err; 1821 } 1822 1823 static int process_sched_runtime_event(const struct perf_tool *tool, 1824 struct evsel *evsel, 1825 struct perf_sample *sample, 1826 struct machine *machine) 1827 { 1828 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1829 1830 if (sched->tp_handler->runtime_event) 1831 return sched->tp_handler->runtime_event(sched, evsel, sample, machine); 1832 1833 return 0; 1834 } 1835 1836 static int perf_sched__process_fork_event(const struct perf_tool *tool, 1837 union perf_event *event, 1838 struct perf_sample *sample, 1839 struct machine *machine) 1840 { 1841 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1842 1843 /* run the fork event through the perf machinery */ 1844 perf_event__process_fork(tool, event, sample, machine); 1845 1846 /* and then run additional processing needed for this command */ 1847 if (sched->tp_handler->fork_event) 1848 return sched->tp_handler->fork_event(sched, event, machine); 1849 1850 return 0; 1851 } 1852 1853 static int process_sched_migrate_task_event(const struct perf_tool *tool, 1854 struct evsel *evsel, 1855 struct perf_sample *sample, 1856 struct machine *machine) 1857 { 1858 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1859 1860 if (sched->tp_handler->migrate_task_event) 1861 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine); 1862 1863 return 0; 1864 } 1865 1866 typedef int (*tracepoint_handler)(const struct perf_tool *tool, 1867 struct evsel *evsel, 1868 struct perf_sample *sample, 1869 struct machine *machine); 1870 1871 static int perf_sched__process_tracepoint_sample(const struct perf_tool *tool __maybe_unused, 1872 union perf_event *event __maybe_unused, 1873 struct perf_sample *sample, 1874 struct evsel *evsel, 1875 struct machine *machine) 1876 { 1877 int err = 0; 1878 1879 if (evsel->handler != NULL) { 1880 tracepoint_handler f = evsel->handler; 1881 err = f(tool, evsel, sample, machine); 1882 } 1883 1884 return err; 1885 } 1886 1887 static int perf_sched__process_comm(const struct perf_tool *tool __maybe_unused, 1888 union perf_event *event, 1889 struct perf_sample *sample, 1890 struct machine *machine) 1891 { 1892 struct thread *thread; 1893 struct thread_runtime *tr; 1894 int err; 1895 1896 err = perf_event__process_comm(tool, event, sample, machine); 1897 if (err) 1898 return err; 1899 1900 thread = machine__find_thread(machine, sample->pid, sample->tid); 1901 if (!thread) { 1902 pr_err("Internal error: can't find thread\n"); 1903 return -1; 1904 } 1905 1906 tr = thread__get_runtime(thread); 1907 if (tr == NULL) { 1908 thread__put(thread); 1909 return -1; 1910 } 1911 1912 tr->comm_changed = true; 1913 thread__put(thread); 1914 1915 return 0; 1916 } 1917 1918 static int perf_sched__read_events(struct perf_sched *sched) 1919 { 1920 struct evsel_str_handler handlers[] = { 1921 { "sched:sched_switch", process_sched_switch_event, }, 1922 { "sched:sched_stat_runtime", process_sched_runtime_event, }, 1923 { "sched:sched_wakeup", process_sched_wakeup_event, }, 1924 { "sched:sched_waking", process_sched_wakeup_event, }, 1925 { "sched:sched_wakeup_new", process_sched_wakeup_event, }, 1926 { "sched:sched_migrate_task", process_sched_migrate_task_event, }, 1927 }; 1928 struct perf_session *session; 1929 struct perf_data data = { 1930 .path = input_name, 1931 .mode = PERF_DATA_MODE_READ, 1932 .force = sched->force, 1933 }; 1934 int rc = -1; 1935 1936 session = perf_session__new(&data, &sched->tool); 1937 if (IS_ERR(session)) { 1938 pr_debug("Error creating perf session"); 1939 return PTR_ERR(session); 1940 } 1941 1942 symbol__init(&session->header.env); 1943 1944 /* prefer sched_waking if it is captured */ 1945 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking")) 1946 handlers[2].handler = process_sched_wakeup_ignore; 1947 1948 if (perf_session__set_tracepoints_handlers(session, handlers)) 1949 goto out_delete; 1950 1951 if (perf_session__has_traces(session, "record -R")) { 1952 int err = perf_session__process_events(session); 1953 if (err) { 1954 pr_err("Failed to process events, error %d", err); 1955 goto out_delete; 1956 } 1957 1958 sched->nr_events = session->evlist->stats.nr_events[0]; 1959 sched->nr_lost_events = session->evlist->stats.total_lost; 1960 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST]; 1961 } 1962 1963 rc = 0; 1964 out_delete: 1965 perf_session__delete(session); 1966 return rc; 1967 } 1968 1969 /* 1970 * scheduling times are printed as msec.usec 1971 */ 1972 static inline void print_sched_time(unsigned long long nsecs, int width) 1973 { 1974 unsigned long msecs; 1975 unsigned long usecs; 1976 1977 msecs = nsecs / NSEC_PER_MSEC; 1978 nsecs -= msecs * NSEC_PER_MSEC; 1979 usecs = nsecs / NSEC_PER_USEC; 1980 printf("%*lu.%03lu ", width, msecs, usecs); 1981 } 1982 1983 /* 1984 * returns runtime data for event, allocating memory for it the 1985 * first time it is used. 1986 */ 1987 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel) 1988 { 1989 struct evsel_runtime *r = evsel->priv; 1990 1991 if (r == NULL) { 1992 r = zalloc(sizeof(struct evsel_runtime)); 1993 evsel->priv = r; 1994 } 1995 1996 return r; 1997 } 1998 1999 /* 2000 * save last time event was seen per cpu 2001 */ 2002 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu) 2003 { 2004 struct evsel_runtime *r = evsel__get_runtime(evsel); 2005 2006 if (r == NULL) 2007 return; 2008 2009 if ((cpu >= r->ncpu) || (r->last_time == NULL)) { 2010 int i, n = __roundup_pow_of_two(cpu+1); 2011 void *p = r->last_time; 2012 2013 p = realloc(r->last_time, n * sizeof(u64)); 2014 if (!p) 2015 return; 2016 2017 r->last_time = p; 2018 for (i = r->ncpu; i < n; ++i) 2019 r->last_time[i] = (u64) 0; 2020 2021 r->ncpu = n; 2022 } 2023 2024 r->last_time[cpu] = timestamp; 2025 } 2026 2027 /* returns last time this event was seen on the given cpu */ 2028 static u64 evsel__get_time(struct evsel *evsel, u32 cpu) 2029 { 2030 struct evsel_runtime *r = evsel__get_runtime(evsel); 2031 2032 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu)) 2033 return 0; 2034 2035 return r->last_time[cpu]; 2036 } 2037 2038 static void timehist__evsel_priv_destructor(void *priv) 2039 { 2040 struct evsel_runtime *r = priv; 2041 2042 if (r) { 2043 free(r->last_time); 2044 free(r); 2045 } 2046 } 2047 2048 static int comm_width = 30; 2049 2050 static char *timehist_get_commstr(struct thread *thread) 2051 { 2052 static char str[32]; 2053 const char *comm = thread__comm_str(thread); 2054 pid_t tid = thread__tid(thread); 2055 pid_t pid = thread__pid(thread); 2056 int n; 2057 2058 if (pid == 0) 2059 n = scnprintf(str, sizeof(str), "%s", comm); 2060 2061 else if (tid != pid) 2062 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid); 2063 2064 else 2065 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid); 2066 2067 if (n > comm_width) 2068 comm_width = n; 2069 2070 return str; 2071 } 2072 2073 /* prio field format: xxx or xxx->yyy */ 2074 #define MAX_PRIO_STR_LEN 8 2075 static char *timehist_get_priostr(struct evsel *evsel, 2076 struct thread *thread, 2077 struct perf_sample *sample) 2078 { 2079 static char prio_str[16]; 2080 int prev_prio = (int)evsel__intval(evsel, sample, "prev_prio"); 2081 struct thread_runtime *tr = thread__priv(thread); 2082 2083 if (tr->prio != prev_prio && tr->prio != -1) 2084 scnprintf(prio_str, sizeof(prio_str), "%d->%d", tr->prio, prev_prio); 2085 else 2086 scnprintf(prio_str, sizeof(prio_str), "%d", prev_prio); 2087 2088 return prio_str; 2089 } 2090 2091 static void timehist_header(struct perf_sched *sched) 2092 { 2093 u32 ncpus = sched->max_cpu.cpu + 1; 2094 u32 i, j; 2095 2096 printf("%15s %6s ", "time", "cpu"); 2097 2098 if (sched->show_cpu_visual) { 2099 printf(" "); 2100 for (i = 0, j = 0; i < ncpus; ++i) { 2101 printf("%x", j++); 2102 if (j > 15) 2103 j = 0; 2104 } 2105 printf(" "); 2106 } 2107 2108 printf(" %-*s", comm_width, "task name"); 2109 2110 if (sched->show_prio) 2111 printf(" %-*s", MAX_PRIO_STR_LEN, "prio"); 2112 2113 printf(" %9s %9s %9s", "wait time", "sch delay", "run time"); 2114 2115 if (sched->pre_migrations) 2116 printf(" %9s", "pre-mig time"); 2117 2118 if (sched->show_state) 2119 printf(" %s", "state"); 2120 2121 printf("\n"); 2122 2123 /* 2124 * units row 2125 */ 2126 printf("%15s %-6s ", "", ""); 2127 2128 if (sched->show_cpu_visual) 2129 printf(" %*s ", ncpus, ""); 2130 2131 printf(" %-*s", comm_width, "[tid/pid]"); 2132 2133 if (sched->show_prio) 2134 printf(" %-*s", MAX_PRIO_STR_LEN, ""); 2135 2136 printf(" %9s %9s %9s", "(msec)", "(msec)", "(msec)"); 2137 2138 if (sched->pre_migrations) 2139 printf(" %9s", "(msec)"); 2140 2141 printf("\n"); 2142 2143 /* 2144 * separator 2145 */ 2146 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line); 2147 2148 if (sched->show_cpu_visual) 2149 printf(" %.*s ", ncpus, graph_dotted_line); 2150 2151 printf(" %.*s", comm_width, graph_dotted_line); 2152 2153 if (sched->show_prio) 2154 printf(" %.*s", MAX_PRIO_STR_LEN, graph_dotted_line); 2155 2156 printf(" %.9s %.9s %.9s", graph_dotted_line, graph_dotted_line, graph_dotted_line); 2157 2158 if (sched->pre_migrations) 2159 printf(" %.9s", graph_dotted_line); 2160 2161 if (sched->show_state) 2162 printf(" %.5s", graph_dotted_line); 2163 2164 printf("\n"); 2165 } 2166 2167 static void timehist_print_sample(struct perf_sched *sched, 2168 struct evsel *evsel, 2169 struct perf_sample *sample, 2170 struct addr_location *al, 2171 struct thread *thread, 2172 u64 t, const char state) 2173 { 2174 struct thread_runtime *tr = thread__priv(thread); 2175 const char *next_comm = evsel__strval(evsel, sample, "next_comm"); 2176 const u32 next_pid = evsel__intval(evsel, sample, "next_pid"); 2177 u32 max_cpus = sched->max_cpu.cpu + 1; 2178 char tstr[64]; 2179 char nstr[30]; 2180 u64 wait_time; 2181 2182 if (cpu_list && !test_bit(sample->cpu, cpu_bitmap)) 2183 return; 2184 2185 timestamp__scnprintf_usec(t, tstr, sizeof(tstr)); 2186 printf("%15s [%04d] ", tstr, sample->cpu); 2187 2188 if (sched->show_cpu_visual) { 2189 u32 i; 2190 char c; 2191 2192 printf(" "); 2193 for (i = 0; i < max_cpus; ++i) { 2194 /* flag idle times with 'i'; others are sched events */ 2195 if (i == sample->cpu) 2196 c = (thread__tid(thread) == 0) ? 'i' : 's'; 2197 else 2198 c = ' '; 2199 printf("%c", c); 2200 } 2201 printf(" "); 2202 } 2203 2204 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2205 2206 if (sched->show_prio) 2207 printf(" %-*s ", MAX_PRIO_STR_LEN, timehist_get_priostr(evsel, thread, sample)); 2208 2209 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt; 2210 print_sched_time(wait_time, 6); 2211 2212 print_sched_time(tr->dt_delay, 6); 2213 print_sched_time(tr->dt_run, 6); 2214 if (sched->pre_migrations) 2215 print_sched_time(tr->dt_pre_mig, 6); 2216 2217 if (sched->show_state) 2218 printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state); 2219 2220 if (sched->show_next) { 2221 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid); 2222 printf(" %-*s", comm_width, nstr); 2223 } 2224 2225 if (sched->show_wakeups && !sched->show_next) 2226 printf(" %-*s", comm_width, ""); 2227 2228 if (thread__tid(thread) == 0) 2229 goto out; 2230 2231 if (sched->show_callchain) 2232 printf(" "); 2233 2234 sample__fprintf_sym(sample, al, 0, 2235 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE | 2236 EVSEL__PRINT_CALLCHAIN_ARROW | 2237 EVSEL__PRINT_SKIP_IGNORED, 2238 get_tls_callchain_cursor(), symbol_conf.bt_stop_list, stdout); 2239 2240 out: 2241 printf("\n"); 2242 } 2243 2244 /* 2245 * Explanation of delta-time stats: 2246 * 2247 * t = time of current schedule out event 2248 * tprev = time of previous sched out event 2249 * also time of schedule-in event for current task 2250 * last_time = time of last sched change event for current task 2251 * (i.e, time process was last scheduled out) 2252 * ready_to_run = time of wakeup for current task 2253 * migrated = time of task migration to another CPU 2254 * 2255 * -----|-------------|-------------|-------------|-------------|----- 2256 * last ready migrated tprev t 2257 * time to run 2258 * 2259 * |---------------- dt_wait ----------------| 2260 * |--------- dt_delay ---------|-- dt_run --| 2261 * |- dt_pre_mig -| 2262 * 2263 * dt_run = run time of current task 2264 * dt_wait = time between last schedule out event for task and tprev 2265 * represents time spent off the cpu 2266 * dt_delay = time between wakeup and schedule-in of task 2267 * dt_pre_mig = time between wakeup and migration to another CPU 2268 */ 2269 2270 static void timehist_update_runtime_stats(struct thread_runtime *r, 2271 u64 t, u64 tprev) 2272 { 2273 r->dt_delay = 0; 2274 r->dt_sleep = 0; 2275 r->dt_iowait = 0; 2276 r->dt_preempt = 0; 2277 r->dt_run = 0; 2278 r->dt_pre_mig = 0; 2279 2280 if (tprev) { 2281 r->dt_run = t - tprev; 2282 if (r->ready_to_run) { 2283 if (r->ready_to_run > tprev) 2284 pr_debug("time travel: wakeup time for task > previous sched_switch event\n"); 2285 else 2286 r->dt_delay = tprev - r->ready_to_run; 2287 2288 if ((r->migrated > r->ready_to_run) && (r->migrated < tprev)) 2289 r->dt_pre_mig = r->migrated - r->ready_to_run; 2290 } 2291 2292 if (r->last_time > tprev) 2293 pr_debug("time travel: last sched out time for task > previous sched_switch event\n"); 2294 else if (r->last_time) { 2295 u64 dt_wait = tprev - r->last_time; 2296 2297 if (r->last_state == 'R') 2298 r->dt_preempt = dt_wait; 2299 else if (r->last_state == 'D') 2300 r->dt_iowait = dt_wait; 2301 else 2302 r->dt_sleep = dt_wait; 2303 } 2304 } 2305 2306 update_stats(&r->run_stats, r->dt_run); 2307 2308 r->total_run_time += r->dt_run; 2309 r->total_delay_time += r->dt_delay; 2310 r->total_sleep_time += r->dt_sleep; 2311 r->total_iowait_time += r->dt_iowait; 2312 r->total_preempt_time += r->dt_preempt; 2313 r->total_pre_mig_time += r->dt_pre_mig; 2314 } 2315 2316 static bool is_idle_sample(struct perf_sample *sample, 2317 struct evsel *evsel) 2318 { 2319 /* pid 0 == swapper == idle task */ 2320 if (evsel__name_is(evsel, "sched:sched_switch")) 2321 return evsel__intval(evsel, sample, "prev_pid") == 0; 2322 2323 return sample->pid == 0; 2324 } 2325 2326 static void save_task_callchain(struct perf_sched *sched, 2327 struct perf_sample *sample, 2328 struct evsel *evsel, 2329 struct machine *machine) 2330 { 2331 struct callchain_cursor *cursor; 2332 struct thread *thread; 2333 2334 /* want main thread for process - has maps */ 2335 thread = machine__findnew_thread(machine, sample->pid, sample->pid); 2336 if (thread == NULL) { 2337 pr_debug("Failed to get thread for pid %d.\n", sample->pid); 2338 return; 2339 } 2340 2341 if (!sched->show_callchain || sample->callchain == NULL) { 2342 thread__put(thread); 2343 return; 2344 } 2345 2346 cursor = get_tls_callchain_cursor(); 2347 2348 if (thread__resolve_callchain(thread, cursor, evsel, sample, 2349 NULL, NULL, sched->max_stack + 2) != 0) { 2350 if (verbose > 0) 2351 pr_err("Failed to resolve callchain. Skipping\n"); 2352 2353 thread__put(thread); 2354 return; 2355 } 2356 2357 callchain_cursor_commit(cursor); 2358 thread__put(thread); 2359 2360 while (true) { 2361 struct callchain_cursor_node *node; 2362 struct symbol *sym; 2363 2364 node = callchain_cursor_current(cursor); 2365 if (node == NULL) 2366 break; 2367 2368 sym = node->ms.sym; 2369 if (sym) { 2370 if (!strcmp(sym->name, "schedule") || 2371 !strcmp(sym->name, "__schedule") || 2372 !strcmp(sym->name, "preempt_schedule")) 2373 sym->ignore = 1; 2374 } 2375 2376 callchain_cursor_advance(cursor); 2377 } 2378 } 2379 2380 static int init_idle_thread(struct thread *thread) 2381 { 2382 struct idle_thread_runtime *itr; 2383 2384 thread__set_comm(thread, idle_comm, 0); 2385 2386 itr = zalloc(sizeof(*itr)); 2387 if (itr == NULL) 2388 return -ENOMEM; 2389 2390 init_prio(&itr->tr); 2391 init_stats(&itr->tr.run_stats); 2392 callchain_init(&itr->callchain); 2393 callchain_cursor_reset(&itr->cursor); 2394 thread__set_priv(thread, itr); 2395 2396 return 0; 2397 } 2398 2399 /* 2400 * Track idle stats per cpu by maintaining a local thread 2401 * struct for the idle task on each cpu. 2402 */ 2403 static int init_idle_threads(int ncpu) 2404 { 2405 int i, ret; 2406 2407 idle_threads = zalloc(ncpu * sizeof(struct thread *)); 2408 if (!idle_threads) 2409 return -ENOMEM; 2410 2411 idle_max_cpu = ncpu; 2412 2413 /* allocate the actual thread struct if needed */ 2414 for (i = 0; i < ncpu; ++i) { 2415 idle_threads[i] = thread__new(0, 0); 2416 if (idle_threads[i] == NULL) 2417 return -ENOMEM; 2418 2419 ret = init_idle_thread(idle_threads[i]); 2420 if (ret < 0) 2421 return ret; 2422 } 2423 2424 return 0; 2425 } 2426 2427 static void free_idle_threads(void) 2428 { 2429 int i; 2430 2431 if (idle_threads == NULL) 2432 return; 2433 2434 for (i = 0; i < idle_max_cpu; ++i) { 2435 struct thread *idle = idle_threads[i]; 2436 2437 if (idle) { 2438 struct idle_thread_runtime *itr; 2439 2440 itr = thread__priv(idle); 2441 if (itr) 2442 thread__put(itr->last_thread); 2443 2444 thread__delete(idle); 2445 } 2446 } 2447 2448 free(idle_threads); 2449 } 2450 2451 static struct thread *get_idle_thread(int cpu) 2452 { 2453 /* 2454 * expand/allocate array of pointers to local thread 2455 * structs if needed 2456 */ 2457 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) { 2458 int i, j = __roundup_pow_of_two(cpu+1); 2459 void *p; 2460 2461 p = realloc(idle_threads, j * sizeof(struct thread *)); 2462 if (!p) 2463 return NULL; 2464 2465 idle_threads = (struct thread **) p; 2466 for (i = idle_max_cpu; i < j; ++i) 2467 idle_threads[i] = NULL; 2468 2469 idle_max_cpu = j; 2470 } 2471 2472 /* allocate a new thread struct if needed */ 2473 if (idle_threads[cpu] == NULL) { 2474 idle_threads[cpu] = thread__new(0, 0); 2475 2476 if (idle_threads[cpu]) { 2477 if (init_idle_thread(idle_threads[cpu]) < 0) 2478 return NULL; 2479 } 2480 } 2481 2482 return thread__get(idle_threads[cpu]); 2483 } 2484 2485 static void save_idle_callchain(struct perf_sched *sched, 2486 struct idle_thread_runtime *itr, 2487 struct perf_sample *sample) 2488 { 2489 struct callchain_cursor *cursor; 2490 2491 if (!sched->show_callchain || sample->callchain == NULL) 2492 return; 2493 2494 cursor = get_tls_callchain_cursor(); 2495 if (cursor == NULL) 2496 return; 2497 2498 callchain_cursor__copy(&itr->cursor, cursor); 2499 } 2500 2501 static struct thread *timehist_get_thread(struct perf_sched *sched, 2502 struct perf_sample *sample, 2503 struct machine *machine, 2504 struct evsel *evsel) 2505 { 2506 struct thread *thread; 2507 2508 if (is_idle_sample(sample, evsel)) { 2509 thread = get_idle_thread(sample->cpu); 2510 if (thread == NULL) 2511 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2512 2513 } else { 2514 /* there were samples with tid 0 but non-zero pid */ 2515 thread = machine__findnew_thread(machine, sample->pid, 2516 sample->tid ?: sample->pid); 2517 if (thread == NULL) { 2518 pr_debug("Failed to get thread for tid %d. skipping sample.\n", 2519 sample->tid); 2520 } 2521 2522 save_task_callchain(sched, sample, evsel, machine); 2523 if (sched->idle_hist) { 2524 struct thread *idle; 2525 struct idle_thread_runtime *itr; 2526 2527 idle = get_idle_thread(sample->cpu); 2528 if (idle == NULL) { 2529 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2530 return NULL; 2531 } 2532 2533 itr = thread__priv(idle); 2534 if (itr == NULL) 2535 return NULL; 2536 2537 thread__put(itr->last_thread); 2538 itr->last_thread = thread__get(thread); 2539 2540 /* copy task callchain when entering to idle */ 2541 if (evsel__intval(evsel, sample, "next_pid") == 0) 2542 save_idle_callchain(sched, itr, sample); 2543 } 2544 } 2545 2546 return thread; 2547 } 2548 2549 static bool timehist_skip_sample(struct perf_sched *sched, 2550 struct thread *thread, 2551 struct evsel *evsel, 2552 struct perf_sample *sample) 2553 { 2554 bool rc = false; 2555 int prio = -1; 2556 struct thread_runtime *tr = NULL; 2557 2558 if (thread__is_filtered(thread)) { 2559 rc = true; 2560 sched->skipped_samples++; 2561 } 2562 2563 if (sched->prio_str) { 2564 /* 2565 * Because priority may be changed during task execution, 2566 * first read priority from prev sched_in event for current task. 2567 * If prev sched_in event is not saved, then read priority from 2568 * current task sched_out event. 2569 */ 2570 tr = thread__get_runtime(thread); 2571 if (tr && tr->prio != -1) 2572 prio = tr->prio; 2573 else if (evsel__name_is(evsel, "sched:sched_switch")) 2574 prio = evsel__intval(evsel, sample, "prev_prio"); 2575 2576 if (prio != -1 && !test_bit(prio, sched->prio_bitmap)) { 2577 rc = true; 2578 sched->skipped_samples++; 2579 } 2580 } 2581 2582 if (sched->idle_hist) { 2583 if (!evsel__name_is(evsel, "sched:sched_switch")) 2584 rc = true; 2585 else if (evsel__intval(evsel, sample, "prev_pid") != 0 && 2586 evsel__intval(evsel, sample, "next_pid") != 0) 2587 rc = true; 2588 } 2589 2590 return rc; 2591 } 2592 2593 static void timehist_print_wakeup_event(struct perf_sched *sched, 2594 struct evsel *evsel, 2595 struct perf_sample *sample, 2596 struct machine *machine, 2597 struct thread *awakened) 2598 { 2599 struct thread *thread; 2600 char tstr[64]; 2601 2602 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2603 if (thread == NULL) 2604 return; 2605 2606 /* show wakeup unless both awakee and awaker are filtered */ 2607 if (timehist_skip_sample(sched, thread, evsel, sample) && 2608 timehist_skip_sample(sched, awakened, evsel, sample)) { 2609 thread__put(thread); 2610 return; 2611 } 2612 2613 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2614 printf("%15s [%04d] ", tstr, sample->cpu); 2615 if (sched->show_cpu_visual) 2616 printf(" %*s ", sched->max_cpu.cpu + 1, ""); 2617 2618 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2619 2620 /* dt spacer */ 2621 printf(" %9s %9s %9s ", "", "", ""); 2622 2623 printf("awakened: %s", timehist_get_commstr(awakened)); 2624 2625 printf("\n"); 2626 2627 thread__put(thread); 2628 } 2629 2630 static int timehist_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused, 2631 union perf_event *event __maybe_unused, 2632 struct evsel *evsel __maybe_unused, 2633 struct perf_sample *sample __maybe_unused, 2634 struct machine *machine __maybe_unused) 2635 { 2636 return 0; 2637 } 2638 2639 static int timehist_sched_wakeup_event(const struct perf_tool *tool, 2640 union perf_event *event __maybe_unused, 2641 struct evsel *evsel, 2642 struct perf_sample *sample, 2643 struct machine *machine) 2644 { 2645 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2646 struct thread *thread; 2647 struct thread_runtime *tr = NULL; 2648 /* want pid of awakened task not pid in sample */ 2649 const u32 pid = evsel__intval(evsel, sample, "pid"); 2650 2651 thread = machine__findnew_thread(machine, 0, pid); 2652 if (thread == NULL) 2653 return -1; 2654 2655 tr = thread__get_runtime(thread); 2656 if (tr == NULL) { 2657 thread__put(thread); 2658 return -1; 2659 } 2660 2661 if (tr->ready_to_run == 0) 2662 tr->ready_to_run = sample->time; 2663 2664 /* show wakeups if requested */ 2665 if (sched->show_wakeups && 2666 !perf_time__skip_sample(&sched->ptime, sample->time)) 2667 timehist_print_wakeup_event(sched, evsel, sample, machine, thread); 2668 2669 thread__put(thread); 2670 return 0; 2671 } 2672 2673 static void timehist_print_migration_event(struct perf_sched *sched, 2674 struct evsel *evsel, 2675 struct perf_sample *sample, 2676 struct machine *machine, 2677 struct thread *migrated) 2678 { 2679 struct thread *thread; 2680 char tstr[64]; 2681 u32 max_cpus; 2682 u32 ocpu, dcpu; 2683 2684 if (sched->summary_only) 2685 return; 2686 2687 max_cpus = sched->max_cpu.cpu + 1; 2688 ocpu = evsel__intval(evsel, sample, "orig_cpu"); 2689 dcpu = evsel__intval(evsel, sample, "dest_cpu"); 2690 2691 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2692 if (thread == NULL) 2693 return; 2694 2695 if (timehist_skip_sample(sched, thread, evsel, sample) && 2696 timehist_skip_sample(sched, migrated, evsel, sample)) { 2697 thread__put(thread); 2698 return; 2699 } 2700 2701 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2702 printf("%15s [%04d] ", tstr, sample->cpu); 2703 2704 if (sched->show_cpu_visual) { 2705 u32 i; 2706 char c; 2707 2708 printf(" "); 2709 for (i = 0; i < max_cpus; ++i) { 2710 c = (i == sample->cpu) ? 'm' : ' '; 2711 printf("%c", c); 2712 } 2713 printf(" "); 2714 } 2715 2716 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2717 2718 /* dt spacer */ 2719 printf(" %9s %9s %9s ", "", "", ""); 2720 2721 printf("migrated: %s", timehist_get_commstr(migrated)); 2722 printf(" cpu %d => %d", ocpu, dcpu); 2723 2724 printf("\n"); 2725 thread__put(thread); 2726 } 2727 2728 static int timehist_migrate_task_event(const struct perf_tool *tool, 2729 union perf_event *event __maybe_unused, 2730 struct evsel *evsel, 2731 struct perf_sample *sample, 2732 struct machine *machine) 2733 { 2734 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2735 struct thread *thread; 2736 struct thread_runtime *tr = NULL; 2737 /* want pid of migrated task not pid in sample */ 2738 const u32 pid = evsel__intval(evsel, sample, "pid"); 2739 2740 thread = machine__findnew_thread(machine, 0, pid); 2741 if (thread == NULL) 2742 return -1; 2743 2744 tr = thread__get_runtime(thread); 2745 if (tr == NULL) { 2746 thread__put(thread); 2747 return -1; 2748 } 2749 2750 tr->migrations++; 2751 tr->migrated = sample->time; 2752 2753 /* show migrations if requested */ 2754 if (sched->show_migrations) { 2755 timehist_print_migration_event(sched, evsel, sample, 2756 machine, thread); 2757 } 2758 thread__put(thread); 2759 2760 return 0; 2761 } 2762 2763 static void timehist_update_task_prio(struct evsel *evsel, 2764 struct perf_sample *sample, 2765 struct machine *machine) 2766 { 2767 struct thread *thread; 2768 struct thread_runtime *tr = NULL; 2769 const u32 next_pid = evsel__intval(evsel, sample, "next_pid"); 2770 const u32 next_prio = evsel__intval(evsel, sample, "next_prio"); 2771 2772 if (next_pid == 0) 2773 thread = get_idle_thread(sample->cpu); 2774 else 2775 thread = machine__findnew_thread(machine, -1, next_pid); 2776 2777 if (thread == NULL) 2778 return; 2779 2780 tr = thread__get_runtime(thread); 2781 if (tr != NULL) 2782 tr->prio = next_prio; 2783 2784 thread__put(thread); 2785 } 2786 2787 static int timehist_sched_change_event(const struct perf_tool *tool, 2788 union perf_event *event, 2789 struct evsel *evsel, 2790 struct perf_sample *sample, 2791 struct machine *machine) 2792 { 2793 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2794 struct perf_time_interval *ptime = &sched->ptime; 2795 struct addr_location al; 2796 struct thread *thread = NULL; 2797 struct thread_runtime *tr = NULL; 2798 u64 tprev, t = sample->time; 2799 int rc = 0; 2800 const char state = evsel__taskstate(evsel, sample, "prev_state"); 2801 2802 addr_location__init(&al); 2803 if (machine__resolve(machine, &al, sample) < 0) { 2804 pr_err("problem processing %d event. skipping it\n", 2805 event->header.type); 2806 rc = -1; 2807 goto out; 2808 } 2809 2810 if (sched->show_prio || sched->prio_str) 2811 timehist_update_task_prio(evsel, sample, machine); 2812 2813 thread = timehist_get_thread(sched, sample, machine, evsel); 2814 if (thread == NULL) { 2815 rc = -1; 2816 goto out; 2817 } 2818 2819 if (timehist_skip_sample(sched, thread, evsel, sample)) 2820 goto out; 2821 2822 tr = thread__get_runtime(thread); 2823 if (tr == NULL) { 2824 rc = -1; 2825 goto out; 2826 } 2827 2828 tprev = evsel__get_time(evsel, sample->cpu); 2829 2830 /* 2831 * If start time given: 2832 * - sample time is under window user cares about - skip sample 2833 * - tprev is under window user cares about - reset to start of window 2834 */ 2835 if (ptime->start && ptime->start > t) 2836 goto out; 2837 2838 if (tprev && ptime->start > tprev) 2839 tprev = ptime->start; 2840 2841 /* 2842 * If end time given: 2843 * - previous sched event is out of window - we are done 2844 * - sample time is beyond window user cares about - reset it 2845 * to close out stats for time window interest 2846 * - If tprev is 0, that is, sched_in event for current task is 2847 * not recorded, cannot determine whether sched_in event is 2848 * within time window interest - ignore it 2849 */ 2850 if (ptime->end) { 2851 if (!tprev || tprev > ptime->end) 2852 goto out; 2853 2854 if (t > ptime->end) 2855 t = ptime->end; 2856 } 2857 2858 if (!sched->idle_hist || thread__tid(thread) == 0) { 2859 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap)) 2860 timehist_update_runtime_stats(tr, t, tprev); 2861 2862 if (sched->idle_hist) { 2863 struct idle_thread_runtime *itr = (void *)tr; 2864 struct thread_runtime *last_tr; 2865 2866 if (itr->last_thread == NULL) 2867 goto out; 2868 2869 /* add current idle time as last thread's runtime */ 2870 last_tr = thread__get_runtime(itr->last_thread); 2871 if (last_tr == NULL) 2872 goto out; 2873 2874 timehist_update_runtime_stats(last_tr, t, tprev); 2875 /* 2876 * remove delta time of last thread as it's not updated 2877 * and otherwise it will show an invalid value next 2878 * time. we only care total run time and run stat. 2879 */ 2880 last_tr->dt_run = 0; 2881 last_tr->dt_delay = 0; 2882 last_tr->dt_sleep = 0; 2883 last_tr->dt_iowait = 0; 2884 last_tr->dt_preempt = 0; 2885 2886 if (itr->cursor.nr) 2887 callchain_append(&itr->callchain, &itr->cursor, t - tprev); 2888 2889 itr->last_thread = NULL; 2890 } 2891 2892 if (!sched->summary_only) 2893 timehist_print_sample(sched, evsel, sample, &al, thread, t, state); 2894 } 2895 2896 out: 2897 if (sched->hist_time.start == 0 && t >= ptime->start) 2898 sched->hist_time.start = t; 2899 if (ptime->end == 0 || t <= ptime->end) 2900 sched->hist_time.end = t; 2901 2902 if (tr) { 2903 /* time of this sched_switch event becomes last time task seen */ 2904 tr->last_time = sample->time; 2905 2906 /* last state is used to determine where to account wait time */ 2907 tr->last_state = state; 2908 2909 /* sched out event for task so reset ready to run time and migrated time */ 2910 if (state == 'R') 2911 tr->ready_to_run = t; 2912 else 2913 tr->ready_to_run = 0; 2914 2915 tr->migrated = 0; 2916 } 2917 2918 evsel__save_time(evsel, sample->time, sample->cpu); 2919 2920 thread__put(thread); 2921 addr_location__exit(&al); 2922 return rc; 2923 } 2924 2925 static int timehist_sched_switch_event(const struct perf_tool *tool, 2926 union perf_event *event, 2927 struct evsel *evsel, 2928 struct perf_sample *sample, 2929 struct machine *machine __maybe_unused) 2930 { 2931 return timehist_sched_change_event(tool, event, evsel, sample, machine); 2932 } 2933 2934 static int process_lost(const struct perf_tool *tool __maybe_unused, 2935 union perf_event *event, 2936 struct perf_sample *sample, 2937 struct machine *machine __maybe_unused) 2938 { 2939 char tstr[64]; 2940 2941 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2942 printf("%15s ", tstr); 2943 printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu); 2944 2945 return 0; 2946 } 2947 2948 2949 static void print_thread_runtime(struct thread *t, 2950 struct thread_runtime *r) 2951 { 2952 double mean = avg_stats(&r->run_stats); 2953 float stddev; 2954 2955 printf("%*s %5d %9" PRIu64 " ", 2956 comm_width, timehist_get_commstr(t), thread__ppid(t), 2957 (u64) r->run_stats.n); 2958 2959 print_sched_time(r->total_run_time, 8); 2960 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean); 2961 print_sched_time(r->run_stats.min, 6); 2962 printf(" "); 2963 print_sched_time((u64) mean, 6); 2964 printf(" "); 2965 print_sched_time(r->run_stats.max, 6); 2966 printf(" "); 2967 printf("%5.2f", stddev); 2968 printf(" %5" PRIu64, r->migrations); 2969 printf("\n"); 2970 } 2971 2972 static void print_thread_waittime(struct thread *t, 2973 struct thread_runtime *r) 2974 { 2975 printf("%*s %5d %9" PRIu64 " ", 2976 comm_width, timehist_get_commstr(t), thread__ppid(t), 2977 (u64) r->run_stats.n); 2978 2979 print_sched_time(r->total_run_time, 8); 2980 print_sched_time(r->total_sleep_time, 6); 2981 printf(" "); 2982 print_sched_time(r->total_iowait_time, 6); 2983 printf(" "); 2984 print_sched_time(r->total_preempt_time, 6); 2985 printf(" "); 2986 print_sched_time(r->total_delay_time, 6); 2987 printf("\n"); 2988 } 2989 2990 struct total_run_stats { 2991 struct perf_sched *sched; 2992 u64 sched_count; 2993 u64 task_count; 2994 u64 total_run_time; 2995 }; 2996 2997 static int show_thread_runtime(struct thread *t, void *priv) 2998 { 2999 struct total_run_stats *stats = priv; 3000 struct thread_runtime *r; 3001 3002 if (thread__is_filtered(t)) 3003 return 0; 3004 3005 r = thread__priv(t); 3006 if (r && r->run_stats.n) { 3007 stats->task_count++; 3008 stats->sched_count += r->run_stats.n; 3009 stats->total_run_time += r->total_run_time; 3010 3011 if (stats->sched->show_state) 3012 print_thread_waittime(t, r); 3013 else 3014 print_thread_runtime(t, r); 3015 } 3016 3017 return 0; 3018 } 3019 3020 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node) 3021 { 3022 const char *sep = " <- "; 3023 struct callchain_list *chain; 3024 size_t ret = 0; 3025 char bf[1024]; 3026 bool first; 3027 3028 if (node == NULL) 3029 return 0; 3030 3031 ret = callchain__fprintf_folded(fp, node->parent); 3032 first = (ret == 0); 3033 3034 list_for_each_entry(chain, &node->val, list) { 3035 if (chain->ip >= PERF_CONTEXT_MAX) 3036 continue; 3037 if (chain->ms.sym && chain->ms.sym->ignore) 3038 continue; 3039 ret += fprintf(fp, "%s%s", first ? "" : sep, 3040 callchain_list__sym_name(chain, bf, sizeof(bf), 3041 false)); 3042 first = false; 3043 } 3044 3045 return ret; 3046 } 3047 3048 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root) 3049 { 3050 size_t ret = 0; 3051 FILE *fp = stdout; 3052 struct callchain_node *chain; 3053 struct rb_node *rb_node = rb_first_cached(root); 3054 3055 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains"); 3056 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line, 3057 graph_dotted_line); 3058 3059 while (rb_node) { 3060 chain = rb_entry(rb_node, struct callchain_node, rb_node); 3061 rb_node = rb_next(rb_node); 3062 3063 ret += fprintf(fp, " "); 3064 print_sched_time(chain->hit, 12); 3065 ret += 16; /* print_sched_time returns 2nd arg + 4 */ 3066 ret += fprintf(fp, " %8d ", chain->count); 3067 ret += callchain__fprintf_folded(fp, chain); 3068 ret += fprintf(fp, "\n"); 3069 } 3070 3071 return ret; 3072 } 3073 3074 static void timehist_print_summary(struct perf_sched *sched, 3075 struct perf_session *session) 3076 { 3077 struct machine *m = &session->machines.host; 3078 struct total_run_stats totals; 3079 u64 task_count; 3080 struct thread *t; 3081 struct thread_runtime *r; 3082 int i; 3083 u64 hist_time = sched->hist_time.end - sched->hist_time.start; 3084 3085 memset(&totals, 0, sizeof(totals)); 3086 totals.sched = sched; 3087 3088 if (sched->idle_hist) { 3089 printf("\nIdle-time summary\n"); 3090 printf("%*s parent sched-out ", comm_width, "comm"); 3091 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n"); 3092 } else if (sched->show_state) { 3093 printf("\nWait-time summary\n"); 3094 printf("%*s parent sched-in ", comm_width, "comm"); 3095 printf(" run-time sleep iowait preempt delay\n"); 3096 } else { 3097 printf("\nRuntime summary\n"); 3098 printf("%*s parent sched-in ", comm_width, "comm"); 3099 printf(" run-time min-run avg-run max-run stddev migrations\n"); 3100 } 3101 printf("%*s (count) ", comm_width, ""); 3102 printf(" (msec) (msec) (msec) (msec) %s\n", 3103 sched->show_state ? "(msec)" : "%"); 3104 printf("%.117s\n", graph_dotted_line); 3105 3106 machine__for_each_thread(m, show_thread_runtime, &totals); 3107 task_count = totals.task_count; 3108 if (!task_count) 3109 printf("<no still running tasks>\n"); 3110 3111 /* CPU idle stats not tracked when samples were skipped */ 3112 if (sched->skipped_samples && !sched->idle_hist) 3113 return; 3114 3115 printf("\nIdle stats:\n"); 3116 for (i = 0; i < idle_max_cpu; ++i) { 3117 if (cpu_list && !test_bit(i, cpu_bitmap)) 3118 continue; 3119 3120 t = idle_threads[i]; 3121 if (!t) 3122 continue; 3123 3124 r = thread__priv(t); 3125 if (r && r->run_stats.n) { 3126 totals.sched_count += r->run_stats.n; 3127 printf(" CPU %2d idle for ", i); 3128 print_sched_time(r->total_run_time, 6); 3129 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time); 3130 } else 3131 printf(" CPU %2d idle entire time window\n", i); 3132 } 3133 3134 if (sched->idle_hist && sched->show_callchain) { 3135 callchain_param.mode = CHAIN_FOLDED; 3136 callchain_param.value = CCVAL_PERIOD; 3137 3138 callchain_register_param(&callchain_param); 3139 3140 printf("\nIdle stats by callchain:\n"); 3141 for (i = 0; i < idle_max_cpu; ++i) { 3142 struct idle_thread_runtime *itr; 3143 3144 t = idle_threads[i]; 3145 if (!t) 3146 continue; 3147 3148 itr = thread__priv(t); 3149 if (itr == NULL) 3150 continue; 3151 3152 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain, 3153 0, &callchain_param); 3154 3155 printf(" CPU %2d:", i); 3156 print_sched_time(itr->tr.total_run_time, 6); 3157 printf(" msec\n"); 3158 timehist_print_idlehist_callchain(&itr->sorted_root); 3159 printf("\n"); 3160 } 3161 } 3162 3163 printf("\n" 3164 " Total number of unique tasks: %" PRIu64 "\n" 3165 "Total number of context switches: %" PRIu64 "\n", 3166 totals.task_count, totals.sched_count); 3167 3168 printf(" Total run time (msec): "); 3169 print_sched_time(totals.total_run_time, 2); 3170 printf("\n"); 3171 3172 printf(" Total scheduling time (msec): "); 3173 print_sched_time(hist_time, 2); 3174 printf(" (x %d)\n", sched->max_cpu.cpu); 3175 } 3176 3177 typedef int (*sched_handler)(const struct perf_tool *tool, 3178 union perf_event *event, 3179 struct evsel *evsel, 3180 struct perf_sample *sample, 3181 struct machine *machine); 3182 3183 static int perf_timehist__process_sample(const struct perf_tool *tool, 3184 union perf_event *event, 3185 struct perf_sample *sample, 3186 struct evsel *evsel, 3187 struct machine *machine) 3188 { 3189 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 3190 int err = 0; 3191 struct perf_cpu this_cpu = { 3192 .cpu = sample->cpu, 3193 }; 3194 3195 if (this_cpu.cpu > sched->max_cpu.cpu) 3196 sched->max_cpu = this_cpu; 3197 3198 if (evsel->handler != NULL) { 3199 sched_handler f = evsel->handler; 3200 3201 err = f(tool, event, evsel, sample, machine); 3202 } 3203 3204 return err; 3205 } 3206 3207 static int timehist_check_attr(struct perf_sched *sched, 3208 struct evlist *evlist) 3209 { 3210 struct evsel *evsel; 3211 struct evsel_runtime *er; 3212 3213 list_for_each_entry(evsel, &evlist->core.entries, core.node) { 3214 er = evsel__get_runtime(evsel); 3215 if (er == NULL) { 3216 pr_err("Failed to allocate memory for evsel runtime data\n"); 3217 return -1; 3218 } 3219 3220 /* only need to save callchain related to sched_switch event */ 3221 if (sched->show_callchain && 3222 evsel__name_is(evsel, "sched:sched_switch") && 3223 !evsel__has_callchain(evsel)) { 3224 pr_info("Samples of sched_switch event do not have callchains.\n"); 3225 sched->show_callchain = 0; 3226 symbol_conf.use_callchain = 0; 3227 } 3228 } 3229 3230 return 0; 3231 } 3232 3233 static int timehist_parse_prio_str(struct perf_sched *sched) 3234 { 3235 char *p; 3236 unsigned long start_prio, end_prio; 3237 const char *str = sched->prio_str; 3238 3239 if (!str) 3240 return 0; 3241 3242 while (isdigit(*str)) { 3243 p = NULL; 3244 start_prio = strtoul(str, &p, 0); 3245 if (start_prio >= MAX_PRIO || (*p != '\0' && *p != ',' && *p != '-')) 3246 return -1; 3247 3248 if (*p == '-') { 3249 str = ++p; 3250 p = NULL; 3251 end_prio = strtoul(str, &p, 0); 3252 3253 if (end_prio >= MAX_PRIO || (*p != '\0' && *p != ',')) 3254 return -1; 3255 3256 if (end_prio < start_prio) 3257 return -1; 3258 } else { 3259 end_prio = start_prio; 3260 } 3261 3262 for (; start_prio <= end_prio; start_prio++) 3263 __set_bit(start_prio, sched->prio_bitmap); 3264 3265 if (*p) 3266 ++p; 3267 3268 str = p; 3269 } 3270 3271 return 0; 3272 } 3273 3274 static int perf_sched__timehist(struct perf_sched *sched) 3275 { 3276 struct evsel_str_handler handlers[] = { 3277 { "sched:sched_switch", timehist_sched_switch_event, }, 3278 { "sched:sched_wakeup", timehist_sched_wakeup_event, }, 3279 { "sched:sched_waking", timehist_sched_wakeup_event, }, 3280 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, }, 3281 }; 3282 const struct evsel_str_handler migrate_handlers[] = { 3283 { "sched:sched_migrate_task", timehist_migrate_task_event, }, 3284 }; 3285 struct perf_data data = { 3286 .path = input_name, 3287 .mode = PERF_DATA_MODE_READ, 3288 .force = sched->force, 3289 }; 3290 3291 struct perf_session *session; 3292 struct evlist *evlist; 3293 int err = -1; 3294 3295 /* 3296 * event handlers for timehist option 3297 */ 3298 sched->tool.sample = perf_timehist__process_sample; 3299 sched->tool.mmap = perf_event__process_mmap; 3300 sched->tool.comm = perf_event__process_comm; 3301 sched->tool.exit = perf_event__process_exit; 3302 sched->tool.fork = perf_event__process_fork; 3303 sched->tool.lost = process_lost; 3304 sched->tool.attr = perf_event__process_attr; 3305 sched->tool.tracing_data = perf_event__process_tracing_data; 3306 sched->tool.build_id = perf_event__process_build_id; 3307 3308 sched->tool.ordering_requires_timestamps = true; 3309 3310 symbol_conf.use_callchain = sched->show_callchain; 3311 3312 session = perf_session__new(&data, &sched->tool); 3313 if (IS_ERR(session)) 3314 return PTR_ERR(session); 3315 3316 if (cpu_list) { 3317 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap); 3318 if (err < 0) 3319 goto out; 3320 } 3321 3322 evlist = session->evlist; 3323 3324 symbol__init(&session->header.env); 3325 3326 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) { 3327 pr_err("Invalid time string\n"); 3328 err = -EINVAL; 3329 goto out; 3330 } 3331 3332 if (timehist_check_attr(sched, evlist) != 0) 3333 goto out; 3334 3335 if (timehist_parse_prio_str(sched) != 0) { 3336 pr_err("Invalid prio string\n"); 3337 goto out; 3338 } 3339 3340 setup_pager(); 3341 3342 evsel__set_priv_destructor(timehist__evsel_priv_destructor); 3343 3344 /* prefer sched_waking if it is captured */ 3345 if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking")) 3346 handlers[1].handler = timehist_sched_wakeup_ignore; 3347 3348 /* setup per-evsel handlers */ 3349 if (perf_session__set_tracepoints_handlers(session, handlers)) 3350 goto out; 3351 3352 /* sched_switch event at a minimum needs to exist */ 3353 if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) { 3354 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n"); 3355 goto out; 3356 } 3357 3358 if ((sched->show_migrations || sched->pre_migrations) && 3359 perf_session__set_tracepoints_handlers(session, migrate_handlers)) 3360 goto out; 3361 3362 /* pre-allocate struct for per-CPU idle stats */ 3363 sched->max_cpu.cpu = session->header.env.nr_cpus_online; 3364 if (sched->max_cpu.cpu == 0) 3365 sched->max_cpu.cpu = 4; 3366 if (init_idle_threads(sched->max_cpu.cpu)) 3367 goto out; 3368 3369 /* summary_only implies summary option, but don't overwrite summary if set */ 3370 if (sched->summary_only) 3371 sched->summary = sched->summary_only; 3372 3373 if (!sched->summary_only) 3374 timehist_header(sched); 3375 3376 err = perf_session__process_events(session); 3377 if (err) { 3378 pr_err("Failed to process events, error %d", err); 3379 goto out; 3380 } 3381 3382 sched->nr_events = evlist->stats.nr_events[0]; 3383 sched->nr_lost_events = evlist->stats.total_lost; 3384 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST]; 3385 3386 if (sched->summary) 3387 timehist_print_summary(sched, session); 3388 3389 out: 3390 free_idle_threads(); 3391 perf_session__delete(session); 3392 3393 return err; 3394 } 3395 3396 3397 static void print_bad_events(struct perf_sched *sched) 3398 { 3399 if (sched->nr_unordered_timestamps && sched->nr_timestamps) { 3400 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n", 3401 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0, 3402 sched->nr_unordered_timestamps, sched->nr_timestamps); 3403 } 3404 if (sched->nr_lost_events && sched->nr_events) { 3405 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", 3406 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0, 3407 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks); 3408 } 3409 if (sched->nr_context_switch_bugs && sched->nr_timestamps) { 3410 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)", 3411 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0, 3412 sched->nr_context_switch_bugs, sched->nr_timestamps); 3413 if (sched->nr_lost_events) 3414 printf(" (due to lost events?)"); 3415 printf("\n"); 3416 } 3417 } 3418 3419 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data) 3420 { 3421 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; 3422 struct work_atoms *this; 3423 const char *comm = thread__comm_str(data->thread), *this_comm; 3424 bool leftmost = true; 3425 3426 while (*new) { 3427 int cmp; 3428 3429 this = container_of(*new, struct work_atoms, node); 3430 parent = *new; 3431 3432 this_comm = thread__comm_str(this->thread); 3433 cmp = strcmp(comm, this_comm); 3434 if (cmp > 0) { 3435 new = &((*new)->rb_left); 3436 } else if (cmp < 0) { 3437 new = &((*new)->rb_right); 3438 leftmost = false; 3439 } else { 3440 this->num_merged++; 3441 this->total_runtime += data->total_runtime; 3442 this->nb_atoms += data->nb_atoms; 3443 this->total_lat += data->total_lat; 3444 list_splice_init(&data->work_list, &this->work_list); 3445 if (this->max_lat < data->max_lat) { 3446 this->max_lat = data->max_lat; 3447 this->max_lat_start = data->max_lat_start; 3448 this->max_lat_end = data->max_lat_end; 3449 } 3450 free_work_atoms(data); 3451 return; 3452 } 3453 } 3454 3455 data->num_merged++; 3456 rb_link_node(&data->node, parent, new); 3457 rb_insert_color_cached(&data->node, root, leftmost); 3458 } 3459 3460 static void perf_sched__merge_lat(struct perf_sched *sched) 3461 { 3462 struct work_atoms *data; 3463 struct rb_node *node; 3464 3465 if (sched->skip_merge) 3466 return; 3467 3468 while ((node = rb_first_cached(&sched->atom_root))) { 3469 rb_erase_cached(node, &sched->atom_root); 3470 data = rb_entry(node, struct work_atoms, node); 3471 __merge_work_atoms(&sched->merged_atom_root, data); 3472 } 3473 } 3474 3475 static int setup_cpus_switch_event(struct perf_sched *sched) 3476 { 3477 unsigned int i; 3478 3479 sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched))); 3480 if (!sched->cpu_last_switched) 3481 return -1; 3482 3483 sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid))); 3484 if (!sched->curr_pid) { 3485 zfree(&sched->cpu_last_switched); 3486 return -1; 3487 } 3488 3489 for (i = 0; i < MAX_CPUS; i++) 3490 sched->curr_pid[i] = -1; 3491 3492 return 0; 3493 } 3494 3495 static void free_cpus_switch_event(struct perf_sched *sched) 3496 { 3497 zfree(&sched->curr_pid); 3498 zfree(&sched->cpu_last_switched); 3499 } 3500 3501 static int perf_sched__lat(struct perf_sched *sched) 3502 { 3503 int rc = -1; 3504 struct rb_node *next; 3505 3506 setup_pager(); 3507 3508 if (setup_cpus_switch_event(sched)) 3509 return rc; 3510 3511 if (perf_sched__read_events(sched)) 3512 goto out_free_cpus_switch_event; 3513 3514 perf_sched__merge_lat(sched); 3515 perf_sched__sort_lat(sched); 3516 3517 printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n"); 3518 printf(" Task | Runtime ms | Count | Avg delay ms | Max delay ms | Max delay start | Max delay end |\n"); 3519 printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n"); 3520 3521 next = rb_first_cached(&sched->sorted_atom_root); 3522 3523 while (next) { 3524 struct work_atoms *work_list; 3525 3526 work_list = rb_entry(next, struct work_atoms, node); 3527 output_lat_thread(sched, work_list); 3528 next = rb_next(next); 3529 } 3530 3531 printf(" -----------------------------------------------------------------------------------------------------------------\n"); 3532 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n", 3533 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count); 3534 3535 printf(" ---------------------------------------------------\n"); 3536 3537 print_bad_events(sched); 3538 printf("\n"); 3539 3540 rc = 0; 3541 3542 while ((next = rb_first_cached(&sched->sorted_atom_root))) { 3543 struct work_atoms *data; 3544 3545 data = rb_entry(next, struct work_atoms, node); 3546 rb_erase_cached(next, &sched->sorted_atom_root); 3547 free_work_atoms(data); 3548 } 3549 out_free_cpus_switch_event: 3550 free_cpus_switch_event(sched); 3551 return rc; 3552 } 3553 3554 static int setup_map_cpus(struct perf_sched *sched) 3555 { 3556 sched->max_cpu.cpu = sysconf(_SC_NPROCESSORS_CONF); 3557 3558 if (sched->map.comp) { 3559 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int)); 3560 if (!sched->map.comp_cpus) 3561 return -1; 3562 } 3563 3564 if (sched->map.cpus_str) { 3565 sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str); 3566 if (!sched->map.cpus) { 3567 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str); 3568 zfree(&sched->map.comp_cpus); 3569 return -1; 3570 } 3571 } 3572 3573 return 0; 3574 } 3575 3576 static int setup_color_pids(struct perf_sched *sched) 3577 { 3578 struct perf_thread_map *map; 3579 3580 if (!sched->map.color_pids_str) 3581 return 0; 3582 3583 map = thread_map__new_by_tid_str(sched->map.color_pids_str); 3584 if (!map) { 3585 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str); 3586 return -1; 3587 } 3588 3589 sched->map.color_pids = map; 3590 return 0; 3591 } 3592 3593 static int setup_color_cpus(struct perf_sched *sched) 3594 { 3595 struct perf_cpu_map *map; 3596 3597 if (!sched->map.color_cpus_str) 3598 return 0; 3599 3600 map = perf_cpu_map__new(sched->map.color_cpus_str); 3601 if (!map) { 3602 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str); 3603 return -1; 3604 } 3605 3606 sched->map.color_cpus = map; 3607 return 0; 3608 } 3609 3610 static int perf_sched__map(struct perf_sched *sched) 3611 { 3612 int rc = -1; 3613 3614 sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread))); 3615 if (!sched->curr_thread) 3616 return rc; 3617 3618 sched->curr_out_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_out_thread))); 3619 if (!sched->curr_out_thread) 3620 goto out_free_curr_thread; 3621 3622 if (setup_cpus_switch_event(sched)) 3623 goto out_free_curr_out_thread; 3624 3625 if (setup_map_cpus(sched)) 3626 goto out_free_cpus_switch_event; 3627 3628 if (setup_color_pids(sched)) 3629 goto out_put_map_cpus; 3630 3631 if (setup_color_cpus(sched)) 3632 goto out_put_color_pids; 3633 3634 setup_pager(); 3635 if (perf_sched__read_events(sched)) 3636 goto out_put_color_cpus; 3637 3638 rc = 0; 3639 print_bad_events(sched); 3640 3641 out_put_color_cpus: 3642 perf_cpu_map__put(sched->map.color_cpus); 3643 3644 out_put_color_pids: 3645 perf_thread_map__put(sched->map.color_pids); 3646 3647 out_put_map_cpus: 3648 zfree(&sched->map.comp_cpus); 3649 perf_cpu_map__put(sched->map.cpus); 3650 3651 out_free_cpus_switch_event: 3652 free_cpus_switch_event(sched); 3653 3654 out_free_curr_out_thread: 3655 for (int i = 0; i < MAX_CPUS; i++) 3656 thread__put(sched->curr_out_thread[i]); 3657 zfree(&sched->curr_out_thread); 3658 3659 out_free_curr_thread: 3660 for (int i = 0; i < MAX_CPUS; i++) 3661 thread__put(sched->curr_thread[i]); 3662 zfree(&sched->curr_thread); 3663 return rc; 3664 } 3665 3666 static int perf_sched__replay(struct perf_sched *sched) 3667 { 3668 int ret; 3669 unsigned long i; 3670 3671 mutex_init(&sched->start_work_mutex); 3672 mutex_init(&sched->work_done_wait_mutex); 3673 3674 ret = setup_cpus_switch_event(sched); 3675 if (ret) 3676 goto out_mutex_destroy; 3677 3678 calibrate_run_measurement_overhead(sched); 3679 calibrate_sleep_measurement_overhead(sched); 3680 3681 test_calibrations(sched); 3682 3683 ret = perf_sched__read_events(sched); 3684 if (ret) 3685 goto out_free_cpus_switch_event; 3686 3687 printf("nr_run_events: %ld\n", sched->nr_run_events); 3688 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events); 3689 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events); 3690 3691 if (sched->targetless_wakeups) 3692 printf("target-less wakeups: %ld\n", sched->targetless_wakeups); 3693 if (sched->multitarget_wakeups) 3694 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups); 3695 if (sched->nr_run_events_optimized) 3696 printf("run atoms optimized: %ld\n", 3697 sched->nr_run_events_optimized); 3698 3699 print_task_traces(sched); 3700 add_cross_task_wakeups(sched); 3701 3702 sched->thread_funcs_exit = false; 3703 create_tasks(sched); 3704 printf("------------------------------------------------------------\n"); 3705 if (sched->replay_repeat == 0) 3706 sched->replay_repeat = UINT_MAX; 3707 3708 for (i = 0; i < sched->replay_repeat; i++) 3709 run_one_test(sched); 3710 3711 sched->thread_funcs_exit = true; 3712 destroy_tasks(sched); 3713 3714 out_free_cpus_switch_event: 3715 free_cpus_switch_event(sched); 3716 3717 out_mutex_destroy: 3718 mutex_destroy(&sched->start_work_mutex); 3719 mutex_destroy(&sched->work_done_wait_mutex); 3720 return ret; 3721 } 3722 3723 static void setup_sorting(struct perf_sched *sched, const struct option *options, 3724 const char * const usage_msg[]) 3725 { 3726 char *tmp, *tok, *str = strdup(sched->sort_order); 3727 3728 for (tok = strtok_r(str, ", ", &tmp); 3729 tok; tok = strtok_r(NULL, ", ", &tmp)) { 3730 if (sort_dimension__add(tok, &sched->sort_list) < 0) { 3731 usage_with_options_msg(usage_msg, options, 3732 "Unknown --sort key: `%s'", tok); 3733 } 3734 } 3735 3736 free(str); 3737 3738 sort_dimension__add("pid", &sched->cmp_pid); 3739 } 3740 3741 static bool schedstat_events_exposed(void) 3742 { 3743 /* 3744 * Select "sched:sched_stat_wait" event to check 3745 * whether schedstat tracepoints are exposed. 3746 */ 3747 return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ? 3748 false : true; 3749 } 3750 3751 static int __cmd_record(int argc, const char **argv) 3752 { 3753 unsigned int rec_argc, i, j; 3754 char **rec_argv; 3755 const char **rec_argv_copy; 3756 const char * const record_args[] = { 3757 "record", 3758 "-a", 3759 "-R", 3760 "-m", "1024", 3761 "-c", "1", 3762 "-e", "sched:sched_switch", 3763 "-e", "sched:sched_stat_runtime", 3764 "-e", "sched:sched_process_fork", 3765 "-e", "sched:sched_wakeup_new", 3766 "-e", "sched:sched_migrate_task", 3767 }; 3768 3769 /* 3770 * The tracepoints trace_sched_stat_{wait, sleep, iowait} 3771 * are not exposed to user if CONFIG_SCHEDSTATS is not set, 3772 * to prevent "perf sched record" execution failure, determine 3773 * whether to record schedstat events according to actual situation. 3774 */ 3775 const char * const schedstat_args[] = { 3776 "-e", "sched:sched_stat_wait", 3777 "-e", "sched:sched_stat_sleep", 3778 "-e", "sched:sched_stat_iowait", 3779 }; 3780 unsigned int schedstat_argc = schedstat_events_exposed() ? 3781 ARRAY_SIZE(schedstat_args) : 0; 3782 3783 struct tep_event *waking_event; 3784 int ret; 3785 3786 /* 3787 * +2 for either "-e", "sched:sched_wakeup" or 3788 * "-e", "sched:sched_waking" 3789 */ 3790 rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1; 3791 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 3792 if (rec_argv == NULL) 3793 return -ENOMEM; 3794 rec_argv_copy = calloc(rec_argc + 1, sizeof(char *)); 3795 if (rec_argv_copy == NULL) { 3796 free(rec_argv); 3797 return -ENOMEM; 3798 } 3799 3800 for (i = 0; i < ARRAY_SIZE(record_args); i++) 3801 rec_argv[i] = strdup(record_args[i]); 3802 3803 rec_argv[i++] = strdup("-e"); 3804 waking_event = trace_event__tp_format("sched", "sched_waking"); 3805 if (!IS_ERR(waking_event)) 3806 rec_argv[i++] = strdup("sched:sched_waking"); 3807 else 3808 rec_argv[i++] = strdup("sched:sched_wakeup"); 3809 3810 for (j = 0; j < schedstat_argc; j++) 3811 rec_argv[i++] = strdup(schedstat_args[j]); 3812 3813 for (j = 1; j < (unsigned int)argc; j++, i++) 3814 rec_argv[i] = strdup(argv[j]); 3815 3816 BUG_ON(i != rec_argc); 3817 3818 memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc); 3819 ret = cmd_record(rec_argc, rec_argv_copy); 3820 3821 for (i = 0; i < rec_argc; i++) 3822 free(rec_argv[i]); 3823 free(rec_argv); 3824 free(rec_argv_copy); 3825 3826 return ret; 3827 } 3828 3829 int cmd_sched(int argc, const char **argv) 3830 { 3831 static const char default_sort_order[] = "avg, max, switch, runtime"; 3832 struct perf_sched sched = { 3833 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid), 3834 .sort_list = LIST_HEAD_INIT(sched.sort_list), 3835 .sort_order = default_sort_order, 3836 .replay_repeat = 10, 3837 .profile_cpu = -1, 3838 .next_shortname1 = 'A', 3839 .next_shortname2 = '0', 3840 .skip_merge = 0, 3841 .show_callchain = 1, 3842 .max_stack = 5, 3843 }; 3844 const struct option sched_options[] = { 3845 OPT_STRING('i', "input", &input_name, "file", 3846 "input file name"), 3847 OPT_INCR('v', "verbose", &verbose, 3848 "be more verbose (show symbol address, etc)"), 3849 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 3850 "dump raw trace in ASCII"), 3851 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"), 3852 OPT_END() 3853 }; 3854 const struct option latency_options[] = { 3855 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]", 3856 "sort by key(s): runtime, switch, avg, max"), 3857 OPT_INTEGER('C', "CPU", &sched.profile_cpu, 3858 "CPU to profile on"), 3859 OPT_BOOLEAN('p', "pids", &sched.skip_merge, 3860 "latency stats per pid instead of per comm"), 3861 OPT_PARENT(sched_options) 3862 }; 3863 const struct option replay_options[] = { 3864 OPT_UINTEGER('r', "repeat", &sched.replay_repeat, 3865 "repeat the workload replay N times (0: infinite)"), 3866 OPT_PARENT(sched_options) 3867 }; 3868 const struct option map_options[] = { 3869 OPT_BOOLEAN(0, "compact", &sched.map.comp, 3870 "map output in compact mode"), 3871 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids", 3872 "highlight given pids in map"), 3873 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus", 3874 "highlight given CPUs in map"), 3875 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus", 3876 "display given CPUs in map"), 3877 OPT_STRING(0, "task-name", &sched.map.task_name, "task", 3878 "map output only for the given task name(s)."), 3879 OPT_BOOLEAN(0, "fuzzy-name", &sched.map.fuzzy, 3880 "given command name can be partially matched (fuzzy matching)"), 3881 OPT_PARENT(sched_options) 3882 }; 3883 const struct option timehist_options[] = { 3884 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 3885 "file", "vmlinux pathname"), 3886 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 3887 "file", "kallsyms pathname"), 3888 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain, 3889 "Display call chains if present (default on)"), 3890 OPT_UINTEGER(0, "max-stack", &sched.max_stack, 3891 "Maximum number of functions to display backtrace."), 3892 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory", 3893 "Look for files with symbols relative to this directory"), 3894 OPT_BOOLEAN('s', "summary", &sched.summary_only, 3895 "Show only syscall summary with statistics"), 3896 OPT_BOOLEAN('S', "with-summary", &sched.summary, 3897 "Show all syscalls and summary with statistics"), 3898 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"), 3899 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"), 3900 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"), 3901 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"), 3902 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"), 3903 OPT_STRING(0, "time", &sched.time_str, "str", 3904 "Time span for analysis (start,stop)"), 3905 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"), 3906 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]", 3907 "analyze events only for given process id(s)"), 3908 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]", 3909 "analyze events only for given thread id(s)"), 3910 OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"), 3911 OPT_BOOLEAN(0, "show-prio", &sched.show_prio, "Show task priority"), 3912 OPT_STRING(0, "prio", &sched.prio_str, "prio", 3913 "analyze events only for given task priority(ies)"), 3914 OPT_BOOLEAN('P', "pre-migrations", &sched.pre_migrations, "Show pre-migration wait time"), 3915 OPT_PARENT(sched_options) 3916 }; 3917 3918 const char * const latency_usage[] = { 3919 "perf sched latency [<options>]", 3920 NULL 3921 }; 3922 const char * const replay_usage[] = { 3923 "perf sched replay [<options>]", 3924 NULL 3925 }; 3926 const char * const map_usage[] = { 3927 "perf sched map [<options>]", 3928 NULL 3929 }; 3930 const char * const timehist_usage[] = { 3931 "perf sched timehist [<options>]", 3932 NULL 3933 }; 3934 const char *const sched_subcommands[] = { "record", "latency", "map", 3935 "replay", "script", 3936 "timehist", NULL }; 3937 const char *sched_usage[] = { 3938 NULL, 3939 NULL 3940 }; 3941 struct trace_sched_handler lat_ops = { 3942 .wakeup_event = latency_wakeup_event, 3943 .switch_event = latency_switch_event, 3944 .runtime_event = latency_runtime_event, 3945 .migrate_task_event = latency_migrate_task_event, 3946 }; 3947 struct trace_sched_handler map_ops = { 3948 .switch_event = map_switch_event, 3949 }; 3950 struct trace_sched_handler replay_ops = { 3951 .wakeup_event = replay_wakeup_event, 3952 .switch_event = replay_switch_event, 3953 .fork_event = replay_fork_event, 3954 }; 3955 int ret; 3956 3957 perf_tool__init(&sched.tool, /*ordered_events=*/true); 3958 sched.tool.sample = perf_sched__process_tracepoint_sample; 3959 sched.tool.comm = perf_sched__process_comm; 3960 sched.tool.namespaces = perf_event__process_namespaces; 3961 sched.tool.lost = perf_event__process_lost; 3962 sched.tool.fork = perf_sched__process_fork_event; 3963 3964 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands, 3965 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION); 3966 if (!argc) 3967 usage_with_options(sched_usage, sched_options); 3968 3969 thread__set_priv_destructor(free); 3970 3971 /* 3972 * Aliased to 'perf script' for now: 3973 */ 3974 if (!strcmp(argv[0], "script")) { 3975 ret = cmd_script(argc, argv); 3976 } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 3977 ret = __cmd_record(argc, argv); 3978 } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) { 3979 sched.tp_handler = &lat_ops; 3980 if (argc > 1) { 3981 argc = parse_options(argc, argv, latency_options, latency_usage, 0); 3982 if (argc) 3983 usage_with_options(latency_usage, latency_options); 3984 } 3985 setup_sorting(&sched, latency_options, latency_usage); 3986 ret = perf_sched__lat(&sched); 3987 } else if (!strcmp(argv[0], "map")) { 3988 if (argc) { 3989 argc = parse_options(argc, argv, map_options, map_usage, 0); 3990 if (argc) 3991 usage_with_options(map_usage, map_options); 3992 3993 if (sched.map.task_name) { 3994 sched.map.task_names = strlist__new(sched.map.task_name, NULL); 3995 if (sched.map.task_names == NULL) { 3996 fprintf(stderr, "Failed to parse task names\n"); 3997 ret = -1; 3998 goto out; 3999 } 4000 } 4001 } 4002 sched.tp_handler = &map_ops; 4003 setup_sorting(&sched, latency_options, latency_usage); 4004 ret = perf_sched__map(&sched); 4005 } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) { 4006 sched.tp_handler = &replay_ops; 4007 if (argc) { 4008 argc = parse_options(argc, argv, replay_options, replay_usage, 0); 4009 if (argc) 4010 usage_with_options(replay_usage, replay_options); 4011 } 4012 ret = perf_sched__replay(&sched); 4013 } else if (!strcmp(argv[0], "timehist")) { 4014 if (argc) { 4015 argc = parse_options(argc, argv, timehist_options, 4016 timehist_usage, 0); 4017 if (argc) 4018 usage_with_options(timehist_usage, timehist_options); 4019 } 4020 if ((sched.show_wakeups || sched.show_next) && 4021 sched.summary_only) { 4022 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n"); 4023 parse_options_usage(timehist_usage, timehist_options, "s", true); 4024 if (sched.show_wakeups) 4025 parse_options_usage(NULL, timehist_options, "w", true); 4026 if (sched.show_next) 4027 parse_options_usage(NULL, timehist_options, "n", true); 4028 ret = -EINVAL; 4029 goto out; 4030 } 4031 ret = symbol__validate_sym_arguments(); 4032 if (!ret) 4033 ret = perf_sched__timehist(&sched); 4034 } else { 4035 usage_with_options(sched_usage, sched_options); 4036 } 4037 4038 out: 4039 /* free usage string allocated by parse_options_subcommand */ 4040 free((void *)sched_usage[0]); 4041 4042 return ret; 4043 } 4044