1 // SPDX-License-Identifier: GPL-2.0 2 #include "builtin.h" 3 #include "perf.h" 4 5 #include "util/util.h" 6 #include "util/evlist.h" 7 #include "util/cache.h" 8 #include "util/evsel.h" 9 #include "util/symbol.h" 10 #include "util/thread.h" 11 #include "util/header.h" 12 #include "util/session.h" 13 #include "util/tool.h" 14 #include "util/cloexec.h" 15 #include "util/thread_map.h" 16 #include "util/color.h" 17 #include "util/stat.h" 18 #include "util/callchain.h" 19 #include "util/time-utils.h" 20 21 #include <subcmd/parse-options.h> 22 #include "util/trace-event.h" 23 24 #include "util/debug.h" 25 26 #include <linux/kernel.h> 27 #include <linux/log2.h> 28 #include <sys/prctl.h> 29 #include <sys/resource.h> 30 #include <inttypes.h> 31 32 #include <errno.h> 33 #include <semaphore.h> 34 #include <pthread.h> 35 #include <math.h> 36 #include <api/fs/fs.h> 37 #include <linux/time64.h> 38 39 #include "sane_ctype.h" 40 41 #define PR_SET_NAME 15 /* Set process name */ 42 #define MAX_CPUS 4096 43 #define COMM_LEN 20 44 #define SYM_LEN 129 45 #define MAX_PID 1024000 46 47 struct sched_atom; 48 49 struct task_desc { 50 unsigned long nr; 51 unsigned long pid; 52 char comm[COMM_LEN]; 53 54 unsigned long nr_events; 55 unsigned long curr_event; 56 struct sched_atom **atoms; 57 58 pthread_t thread; 59 sem_t sleep_sem; 60 61 sem_t ready_for_work; 62 sem_t work_done_sem; 63 64 u64 cpu_usage; 65 }; 66 67 enum sched_event_type { 68 SCHED_EVENT_RUN, 69 SCHED_EVENT_SLEEP, 70 SCHED_EVENT_WAKEUP, 71 SCHED_EVENT_MIGRATION, 72 }; 73 74 struct sched_atom { 75 enum sched_event_type type; 76 int specific_wait; 77 u64 timestamp; 78 u64 duration; 79 unsigned long nr; 80 sem_t *wait_sem; 81 struct task_desc *wakee; 82 }; 83 84 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP" 85 86 /* task state bitmask, copied from include/linux/sched.h */ 87 #define TASK_RUNNING 0 88 #define TASK_INTERRUPTIBLE 1 89 #define TASK_UNINTERRUPTIBLE 2 90 #define __TASK_STOPPED 4 91 #define __TASK_TRACED 8 92 /* in tsk->exit_state */ 93 #define EXIT_DEAD 16 94 #define EXIT_ZOMBIE 32 95 #define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD) 96 /* in tsk->state again */ 97 #define TASK_DEAD 64 98 #define TASK_WAKEKILL 128 99 #define TASK_WAKING 256 100 #define TASK_PARKED 512 101 102 enum thread_state { 103 THREAD_SLEEPING = 0, 104 THREAD_WAIT_CPU, 105 THREAD_SCHED_IN, 106 THREAD_IGNORE 107 }; 108 109 struct work_atom { 110 struct list_head list; 111 enum thread_state state; 112 u64 sched_out_time; 113 u64 wake_up_time; 114 u64 sched_in_time; 115 u64 runtime; 116 }; 117 118 struct work_atoms { 119 struct list_head work_list; 120 struct thread *thread; 121 struct rb_node node; 122 u64 max_lat; 123 u64 max_lat_at; 124 u64 total_lat; 125 u64 nb_atoms; 126 u64 total_runtime; 127 int num_merged; 128 }; 129 130 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *); 131 132 struct perf_sched; 133 134 struct trace_sched_handler { 135 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel, 136 struct perf_sample *sample, struct machine *machine); 137 138 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel, 139 struct perf_sample *sample, struct machine *machine); 140 141 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel, 142 struct perf_sample *sample, struct machine *machine); 143 144 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */ 145 int (*fork_event)(struct perf_sched *sched, union perf_event *event, 146 struct machine *machine); 147 148 int (*migrate_task_event)(struct perf_sched *sched, 149 struct perf_evsel *evsel, 150 struct perf_sample *sample, 151 struct machine *machine); 152 }; 153 154 #define COLOR_PIDS PERF_COLOR_BLUE 155 #define COLOR_CPUS PERF_COLOR_BG_RED 156 157 struct perf_sched_map { 158 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS); 159 int *comp_cpus; 160 bool comp; 161 struct thread_map *color_pids; 162 const char *color_pids_str; 163 struct cpu_map *color_cpus; 164 const char *color_cpus_str; 165 struct cpu_map *cpus; 166 const char *cpus_str; 167 }; 168 169 struct perf_sched { 170 struct perf_tool tool; 171 const char *sort_order; 172 unsigned long nr_tasks; 173 struct task_desc **pid_to_task; 174 struct task_desc **tasks; 175 const struct trace_sched_handler *tp_handler; 176 pthread_mutex_t start_work_mutex; 177 pthread_mutex_t work_done_wait_mutex; 178 int profile_cpu; 179 /* 180 * Track the current task - that way we can know whether there's any 181 * weird events, such as a task being switched away that is not current. 182 */ 183 int max_cpu; 184 u32 curr_pid[MAX_CPUS]; 185 struct thread *curr_thread[MAX_CPUS]; 186 char next_shortname1; 187 char next_shortname2; 188 unsigned int replay_repeat; 189 unsigned long nr_run_events; 190 unsigned long nr_sleep_events; 191 unsigned long nr_wakeup_events; 192 unsigned long nr_sleep_corrections; 193 unsigned long nr_run_events_optimized; 194 unsigned long targetless_wakeups; 195 unsigned long multitarget_wakeups; 196 unsigned long nr_runs; 197 unsigned long nr_timestamps; 198 unsigned long nr_unordered_timestamps; 199 unsigned long nr_context_switch_bugs; 200 unsigned long nr_events; 201 unsigned long nr_lost_chunks; 202 unsigned long nr_lost_events; 203 u64 run_measurement_overhead; 204 u64 sleep_measurement_overhead; 205 u64 start_time; 206 u64 cpu_usage; 207 u64 runavg_cpu_usage; 208 u64 parent_cpu_usage; 209 u64 runavg_parent_cpu_usage; 210 u64 sum_runtime; 211 u64 sum_fluct; 212 u64 run_avg; 213 u64 all_runtime; 214 u64 all_count; 215 u64 cpu_last_switched[MAX_CPUS]; 216 struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root; 217 struct list_head sort_list, cmp_pid; 218 bool force; 219 bool skip_merge; 220 struct perf_sched_map map; 221 222 /* options for timehist command */ 223 bool summary; 224 bool summary_only; 225 bool idle_hist; 226 bool show_callchain; 227 unsigned int max_stack; 228 bool show_cpu_visual; 229 bool show_wakeups; 230 bool show_next; 231 bool show_migrations; 232 bool show_state; 233 u64 skipped_samples; 234 const char *time_str; 235 struct perf_time_interval ptime; 236 struct perf_time_interval hist_time; 237 }; 238 239 /* per thread run time data */ 240 struct thread_runtime { 241 u64 last_time; /* time of previous sched in/out event */ 242 u64 dt_run; /* run time */ 243 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */ 244 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */ 245 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */ 246 u64 dt_delay; /* time between wakeup and sched-in */ 247 u64 ready_to_run; /* time of wakeup */ 248 249 struct stats run_stats; 250 u64 total_run_time; 251 u64 total_sleep_time; 252 u64 total_iowait_time; 253 u64 total_preempt_time; 254 u64 total_delay_time; 255 256 int last_state; 257 258 char shortname[3]; 259 bool comm_changed; 260 261 u64 migrations; 262 }; 263 264 /* per event run time data */ 265 struct evsel_runtime { 266 u64 *last_time; /* time this event was last seen per cpu */ 267 u32 ncpu; /* highest cpu slot allocated */ 268 }; 269 270 /* per cpu idle time data */ 271 struct idle_thread_runtime { 272 struct thread_runtime tr; 273 struct thread *last_thread; 274 struct rb_root_cached sorted_root; 275 struct callchain_root callchain; 276 struct callchain_cursor cursor; 277 }; 278 279 /* track idle times per cpu */ 280 static struct thread **idle_threads; 281 static int idle_max_cpu; 282 static char idle_comm[] = "<idle>"; 283 284 static u64 get_nsecs(void) 285 { 286 struct timespec ts; 287 288 clock_gettime(CLOCK_MONOTONIC, &ts); 289 290 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec; 291 } 292 293 static void burn_nsecs(struct perf_sched *sched, u64 nsecs) 294 { 295 u64 T0 = get_nsecs(), T1; 296 297 do { 298 T1 = get_nsecs(); 299 } while (T1 + sched->run_measurement_overhead < T0 + nsecs); 300 } 301 302 static void sleep_nsecs(u64 nsecs) 303 { 304 struct timespec ts; 305 306 ts.tv_nsec = nsecs % 999999999; 307 ts.tv_sec = nsecs / 999999999; 308 309 nanosleep(&ts, NULL); 310 } 311 312 static void calibrate_run_measurement_overhead(struct perf_sched *sched) 313 { 314 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 315 int i; 316 317 for (i = 0; i < 10; i++) { 318 T0 = get_nsecs(); 319 burn_nsecs(sched, 0); 320 T1 = get_nsecs(); 321 delta = T1-T0; 322 min_delta = min(min_delta, delta); 323 } 324 sched->run_measurement_overhead = min_delta; 325 326 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta); 327 } 328 329 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched) 330 { 331 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 332 int i; 333 334 for (i = 0; i < 10; i++) { 335 T0 = get_nsecs(); 336 sleep_nsecs(10000); 337 T1 = get_nsecs(); 338 delta = T1-T0; 339 min_delta = min(min_delta, delta); 340 } 341 min_delta -= 10000; 342 sched->sleep_measurement_overhead = min_delta; 343 344 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta); 345 } 346 347 static struct sched_atom * 348 get_new_event(struct task_desc *task, u64 timestamp) 349 { 350 struct sched_atom *event = zalloc(sizeof(*event)); 351 unsigned long idx = task->nr_events; 352 size_t size; 353 354 event->timestamp = timestamp; 355 event->nr = idx; 356 357 task->nr_events++; 358 size = sizeof(struct sched_atom *) * task->nr_events; 359 task->atoms = realloc(task->atoms, size); 360 BUG_ON(!task->atoms); 361 362 task->atoms[idx] = event; 363 364 return event; 365 } 366 367 static struct sched_atom *last_event(struct task_desc *task) 368 { 369 if (!task->nr_events) 370 return NULL; 371 372 return task->atoms[task->nr_events - 1]; 373 } 374 375 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task, 376 u64 timestamp, u64 duration) 377 { 378 struct sched_atom *event, *curr_event = last_event(task); 379 380 /* 381 * optimize an existing RUN event by merging this one 382 * to it: 383 */ 384 if (curr_event && curr_event->type == SCHED_EVENT_RUN) { 385 sched->nr_run_events_optimized++; 386 curr_event->duration += duration; 387 return; 388 } 389 390 event = get_new_event(task, timestamp); 391 392 event->type = SCHED_EVENT_RUN; 393 event->duration = duration; 394 395 sched->nr_run_events++; 396 } 397 398 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task, 399 u64 timestamp, struct task_desc *wakee) 400 { 401 struct sched_atom *event, *wakee_event; 402 403 event = get_new_event(task, timestamp); 404 event->type = SCHED_EVENT_WAKEUP; 405 event->wakee = wakee; 406 407 wakee_event = last_event(wakee); 408 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { 409 sched->targetless_wakeups++; 410 return; 411 } 412 if (wakee_event->wait_sem) { 413 sched->multitarget_wakeups++; 414 return; 415 } 416 417 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem)); 418 sem_init(wakee_event->wait_sem, 0, 0); 419 wakee_event->specific_wait = 1; 420 event->wait_sem = wakee_event->wait_sem; 421 422 sched->nr_wakeup_events++; 423 } 424 425 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task, 426 u64 timestamp, u64 task_state __maybe_unused) 427 { 428 struct sched_atom *event = get_new_event(task, timestamp); 429 430 event->type = SCHED_EVENT_SLEEP; 431 432 sched->nr_sleep_events++; 433 } 434 435 static struct task_desc *register_pid(struct perf_sched *sched, 436 unsigned long pid, const char *comm) 437 { 438 struct task_desc *task; 439 static int pid_max; 440 441 if (sched->pid_to_task == NULL) { 442 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0) 443 pid_max = MAX_PID; 444 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL); 445 } 446 if (pid >= (unsigned long)pid_max) { 447 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) * 448 sizeof(struct task_desc *))) == NULL); 449 while (pid >= (unsigned long)pid_max) 450 sched->pid_to_task[pid_max++] = NULL; 451 } 452 453 task = sched->pid_to_task[pid]; 454 455 if (task) 456 return task; 457 458 task = zalloc(sizeof(*task)); 459 task->pid = pid; 460 task->nr = sched->nr_tasks; 461 strcpy(task->comm, comm); 462 /* 463 * every task starts in sleeping state - this gets ignored 464 * if there's no wakeup pointing to this sleep state: 465 */ 466 add_sched_event_sleep(sched, task, 0, 0); 467 468 sched->pid_to_task[pid] = task; 469 sched->nr_tasks++; 470 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *)); 471 BUG_ON(!sched->tasks); 472 sched->tasks[task->nr] = task; 473 474 if (verbose > 0) 475 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm); 476 477 return task; 478 } 479 480 481 static void print_task_traces(struct perf_sched *sched) 482 { 483 struct task_desc *task; 484 unsigned long i; 485 486 for (i = 0; i < sched->nr_tasks; i++) { 487 task = sched->tasks[i]; 488 printf("task %6ld (%20s:%10ld), nr_events: %ld\n", 489 task->nr, task->comm, task->pid, task->nr_events); 490 } 491 } 492 493 static void add_cross_task_wakeups(struct perf_sched *sched) 494 { 495 struct task_desc *task1, *task2; 496 unsigned long i, j; 497 498 for (i = 0; i < sched->nr_tasks; i++) { 499 task1 = sched->tasks[i]; 500 j = i + 1; 501 if (j == sched->nr_tasks) 502 j = 0; 503 task2 = sched->tasks[j]; 504 add_sched_event_wakeup(sched, task1, 0, task2); 505 } 506 } 507 508 static void perf_sched__process_event(struct perf_sched *sched, 509 struct sched_atom *atom) 510 { 511 int ret = 0; 512 513 switch (atom->type) { 514 case SCHED_EVENT_RUN: 515 burn_nsecs(sched, atom->duration); 516 break; 517 case SCHED_EVENT_SLEEP: 518 if (atom->wait_sem) 519 ret = sem_wait(atom->wait_sem); 520 BUG_ON(ret); 521 break; 522 case SCHED_EVENT_WAKEUP: 523 if (atom->wait_sem) 524 ret = sem_post(atom->wait_sem); 525 BUG_ON(ret); 526 break; 527 case SCHED_EVENT_MIGRATION: 528 break; 529 default: 530 BUG_ON(1); 531 } 532 } 533 534 static u64 get_cpu_usage_nsec_parent(void) 535 { 536 struct rusage ru; 537 u64 sum; 538 int err; 539 540 err = getrusage(RUSAGE_SELF, &ru); 541 BUG_ON(err); 542 543 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC; 544 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC; 545 546 return sum; 547 } 548 549 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task) 550 { 551 struct perf_event_attr attr; 552 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE]; 553 int fd; 554 struct rlimit limit; 555 bool need_privilege = false; 556 557 memset(&attr, 0, sizeof(attr)); 558 559 attr.type = PERF_TYPE_SOFTWARE; 560 attr.config = PERF_COUNT_SW_TASK_CLOCK; 561 562 force_again: 563 fd = sys_perf_event_open(&attr, 0, -1, -1, 564 perf_event_open_cloexec_flag()); 565 566 if (fd < 0) { 567 if (errno == EMFILE) { 568 if (sched->force) { 569 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1); 570 limit.rlim_cur += sched->nr_tasks - cur_task; 571 if (limit.rlim_cur > limit.rlim_max) { 572 limit.rlim_max = limit.rlim_cur; 573 need_privilege = true; 574 } 575 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) { 576 if (need_privilege && errno == EPERM) 577 strcpy(info, "Need privilege\n"); 578 } else 579 goto force_again; 580 } else 581 strcpy(info, "Have a try with -f option\n"); 582 } 583 pr_err("Error: sys_perf_event_open() syscall returned " 584 "with %d (%s)\n%s", fd, 585 str_error_r(errno, sbuf, sizeof(sbuf)), info); 586 exit(EXIT_FAILURE); 587 } 588 return fd; 589 } 590 591 static u64 get_cpu_usage_nsec_self(int fd) 592 { 593 u64 runtime; 594 int ret; 595 596 ret = read(fd, &runtime, sizeof(runtime)); 597 BUG_ON(ret != sizeof(runtime)); 598 599 return runtime; 600 } 601 602 struct sched_thread_parms { 603 struct task_desc *task; 604 struct perf_sched *sched; 605 int fd; 606 }; 607 608 static void *thread_func(void *ctx) 609 { 610 struct sched_thread_parms *parms = ctx; 611 struct task_desc *this_task = parms->task; 612 struct perf_sched *sched = parms->sched; 613 u64 cpu_usage_0, cpu_usage_1; 614 unsigned long i, ret; 615 char comm2[22]; 616 int fd = parms->fd; 617 618 zfree(&parms); 619 620 sprintf(comm2, ":%s", this_task->comm); 621 prctl(PR_SET_NAME, comm2); 622 if (fd < 0) 623 return NULL; 624 again: 625 ret = sem_post(&this_task->ready_for_work); 626 BUG_ON(ret); 627 ret = pthread_mutex_lock(&sched->start_work_mutex); 628 BUG_ON(ret); 629 ret = pthread_mutex_unlock(&sched->start_work_mutex); 630 BUG_ON(ret); 631 632 cpu_usage_0 = get_cpu_usage_nsec_self(fd); 633 634 for (i = 0; i < this_task->nr_events; i++) { 635 this_task->curr_event = i; 636 perf_sched__process_event(sched, this_task->atoms[i]); 637 } 638 639 cpu_usage_1 = get_cpu_usage_nsec_self(fd); 640 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; 641 ret = sem_post(&this_task->work_done_sem); 642 BUG_ON(ret); 643 644 ret = pthread_mutex_lock(&sched->work_done_wait_mutex); 645 BUG_ON(ret); 646 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex); 647 BUG_ON(ret); 648 649 goto again; 650 } 651 652 static void create_tasks(struct perf_sched *sched) 653 { 654 struct task_desc *task; 655 pthread_attr_t attr; 656 unsigned long i; 657 int err; 658 659 err = pthread_attr_init(&attr); 660 BUG_ON(err); 661 err = pthread_attr_setstacksize(&attr, 662 (size_t) max(16 * 1024, PTHREAD_STACK_MIN)); 663 BUG_ON(err); 664 err = pthread_mutex_lock(&sched->start_work_mutex); 665 BUG_ON(err); 666 err = pthread_mutex_lock(&sched->work_done_wait_mutex); 667 BUG_ON(err); 668 for (i = 0; i < sched->nr_tasks; i++) { 669 struct sched_thread_parms *parms = malloc(sizeof(*parms)); 670 BUG_ON(parms == NULL); 671 parms->task = task = sched->tasks[i]; 672 parms->sched = sched; 673 parms->fd = self_open_counters(sched, i); 674 sem_init(&task->sleep_sem, 0, 0); 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 wait_for_tasks(struct perf_sched *sched) 684 { 685 u64 cpu_usage_0, cpu_usage_1; 686 struct task_desc *task; 687 unsigned long i, ret; 688 689 sched->start_time = get_nsecs(); 690 sched->cpu_usage = 0; 691 pthread_mutex_unlock(&sched->work_done_wait_mutex); 692 693 for (i = 0; i < sched->nr_tasks; i++) { 694 task = sched->tasks[i]; 695 ret = sem_wait(&task->ready_for_work); 696 BUG_ON(ret); 697 sem_init(&task->ready_for_work, 0, 0); 698 } 699 ret = pthread_mutex_lock(&sched->work_done_wait_mutex); 700 BUG_ON(ret); 701 702 cpu_usage_0 = get_cpu_usage_nsec_parent(); 703 704 pthread_mutex_unlock(&sched->start_work_mutex); 705 706 for (i = 0; i < sched->nr_tasks; i++) { 707 task = sched->tasks[i]; 708 ret = sem_wait(&task->work_done_sem); 709 BUG_ON(ret); 710 sem_init(&task->work_done_sem, 0, 0); 711 sched->cpu_usage += task->cpu_usage; 712 task->cpu_usage = 0; 713 } 714 715 cpu_usage_1 = get_cpu_usage_nsec_parent(); 716 if (!sched->runavg_cpu_usage) 717 sched->runavg_cpu_usage = sched->cpu_usage; 718 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat; 719 720 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0; 721 if (!sched->runavg_parent_cpu_usage) 722 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage; 723 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) + 724 sched->parent_cpu_usage)/sched->replay_repeat; 725 726 ret = pthread_mutex_lock(&sched->start_work_mutex); 727 BUG_ON(ret); 728 729 for (i = 0; i < sched->nr_tasks; i++) { 730 task = sched->tasks[i]; 731 sem_init(&task->sleep_sem, 0, 0); 732 task->curr_event = 0; 733 } 734 } 735 736 static void run_one_test(struct perf_sched *sched) 737 { 738 u64 T0, T1, delta, avg_delta, fluct; 739 740 T0 = get_nsecs(); 741 wait_for_tasks(sched); 742 T1 = get_nsecs(); 743 744 delta = T1 - T0; 745 sched->sum_runtime += delta; 746 sched->nr_runs++; 747 748 avg_delta = sched->sum_runtime / sched->nr_runs; 749 if (delta < avg_delta) 750 fluct = avg_delta - delta; 751 else 752 fluct = delta - avg_delta; 753 sched->sum_fluct += fluct; 754 if (!sched->run_avg) 755 sched->run_avg = delta; 756 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat; 757 758 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC); 759 760 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC); 761 762 printf("cpu: %0.2f / %0.2f", 763 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC); 764 765 #if 0 766 /* 767 * rusage statistics done by the parent, these are less 768 * accurate than the sched->sum_exec_runtime based statistics: 769 */ 770 printf(" [%0.2f / %0.2f]", 771 (double)sched->parent_cpu_usage / NSEC_PER_MSEC, 772 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC); 773 #endif 774 775 printf("\n"); 776 777 if (sched->nr_sleep_corrections) 778 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections); 779 sched->nr_sleep_corrections = 0; 780 } 781 782 static void test_calibrations(struct perf_sched *sched) 783 { 784 u64 T0, T1; 785 786 T0 = get_nsecs(); 787 burn_nsecs(sched, NSEC_PER_MSEC); 788 T1 = get_nsecs(); 789 790 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0); 791 792 T0 = get_nsecs(); 793 sleep_nsecs(NSEC_PER_MSEC); 794 T1 = get_nsecs(); 795 796 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0); 797 } 798 799 static int 800 replay_wakeup_event(struct perf_sched *sched, 801 struct perf_evsel *evsel, struct perf_sample *sample, 802 struct machine *machine __maybe_unused) 803 { 804 const char *comm = perf_evsel__strval(evsel, sample, "comm"); 805 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 806 struct task_desc *waker, *wakee; 807 808 if (verbose > 0) { 809 printf("sched_wakeup event %p\n", evsel); 810 811 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid); 812 } 813 814 waker = register_pid(sched, sample->tid, "<unknown>"); 815 wakee = register_pid(sched, pid, comm); 816 817 add_sched_event_wakeup(sched, waker, sample->time, wakee); 818 return 0; 819 } 820 821 static int replay_switch_event(struct perf_sched *sched, 822 struct perf_evsel *evsel, 823 struct perf_sample *sample, 824 struct machine *machine __maybe_unused) 825 { 826 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"), 827 *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); 828 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 829 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 830 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); 831 struct task_desc *prev, __maybe_unused *next; 832 u64 timestamp0, timestamp = sample->time; 833 int cpu = sample->cpu; 834 s64 delta; 835 836 if (verbose > 0) 837 printf("sched_switch event %p\n", evsel); 838 839 if (cpu >= MAX_CPUS || cpu < 0) 840 return 0; 841 842 timestamp0 = sched->cpu_last_switched[cpu]; 843 if (timestamp0) 844 delta = timestamp - timestamp0; 845 else 846 delta = 0; 847 848 if (delta < 0) { 849 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 850 return -1; 851 } 852 853 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n", 854 prev_comm, prev_pid, next_comm, next_pid, delta); 855 856 prev = register_pid(sched, prev_pid, prev_comm); 857 next = register_pid(sched, next_pid, next_comm); 858 859 sched->cpu_last_switched[cpu] = timestamp; 860 861 add_sched_event_run(sched, prev, timestamp, delta); 862 add_sched_event_sleep(sched, prev, timestamp, prev_state); 863 864 return 0; 865 } 866 867 static int replay_fork_event(struct perf_sched *sched, 868 union perf_event *event, 869 struct machine *machine) 870 { 871 struct thread *child, *parent; 872 873 child = machine__findnew_thread(machine, event->fork.pid, 874 event->fork.tid); 875 parent = machine__findnew_thread(machine, event->fork.ppid, 876 event->fork.ptid); 877 878 if (child == NULL || parent == NULL) { 879 pr_debug("thread does not exist on fork event: child %p, parent %p\n", 880 child, parent); 881 goto out_put; 882 } 883 884 if (verbose > 0) { 885 printf("fork event\n"); 886 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid); 887 printf("... child: %s/%d\n", thread__comm_str(child), child->tid); 888 } 889 890 register_pid(sched, parent->tid, thread__comm_str(parent)); 891 register_pid(sched, child->tid, thread__comm_str(child)); 892 out_put: 893 thread__put(child); 894 thread__put(parent); 895 return 0; 896 } 897 898 struct sort_dimension { 899 const char *name; 900 sort_fn_t cmp; 901 struct list_head list; 902 }; 903 904 /* 905 * handle runtime stats saved per thread 906 */ 907 static struct thread_runtime *thread__init_runtime(struct thread *thread) 908 { 909 struct thread_runtime *r; 910 911 r = zalloc(sizeof(struct thread_runtime)); 912 if (!r) 913 return NULL; 914 915 init_stats(&r->run_stats); 916 thread__set_priv(thread, r); 917 918 return r; 919 } 920 921 static struct thread_runtime *thread__get_runtime(struct thread *thread) 922 { 923 struct thread_runtime *tr; 924 925 tr = thread__priv(thread); 926 if (tr == NULL) { 927 tr = thread__init_runtime(thread); 928 if (tr == NULL) 929 pr_debug("Failed to malloc memory for runtime data.\n"); 930 } 931 932 return tr; 933 } 934 935 static int 936 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r) 937 { 938 struct sort_dimension *sort; 939 int ret = 0; 940 941 BUG_ON(list_empty(list)); 942 943 list_for_each_entry(sort, list, list) { 944 ret = sort->cmp(l, r); 945 if (ret) 946 return ret; 947 } 948 949 return ret; 950 } 951 952 static struct work_atoms * 953 thread_atoms_search(struct rb_root_cached *root, struct thread *thread, 954 struct list_head *sort_list) 955 { 956 struct rb_node *node = root->rb_root.rb_node; 957 struct work_atoms key = { .thread = thread }; 958 959 while (node) { 960 struct work_atoms *atoms; 961 int cmp; 962 963 atoms = container_of(node, struct work_atoms, node); 964 965 cmp = thread_lat_cmp(sort_list, &key, atoms); 966 if (cmp > 0) 967 node = node->rb_left; 968 else if (cmp < 0) 969 node = node->rb_right; 970 else { 971 BUG_ON(thread != atoms->thread); 972 return atoms; 973 } 974 } 975 return NULL; 976 } 977 978 static void 979 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data, 980 struct list_head *sort_list) 981 { 982 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; 983 bool leftmost = true; 984 985 while (*new) { 986 struct work_atoms *this; 987 int cmp; 988 989 this = container_of(*new, struct work_atoms, node); 990 parent = *new; 991 992 cmp = thread_lat_cmp(sort_list, data, this); 993 994 if (cmp > 0) 995 new = &((*new)->rb_left); 996 else { 997 new = &((*new)->rb_right); 998 leftmost = false; 999 } 1000 } 1001 1002 rb_link_node(&data->node, parent, new); 1003 rb_insert_color_cached(&data->node, root, leftmost); 1004 } 1005 1006 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread) 1007 { 1008 struct work_atoms *atoms = zalloc(sizeof(*atoms)); 1009 if (!atoms) { 1010 pr_err("No memory at %s\n", __func__); 1011 return -1; 1012 } 1013 1014 atoms->thread = thread__get(thread); 1015 INIT_LIST_HEAD(&atoms->work_list); 1016 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid); 1017 return 0; 1018 } 1019 1020 static char sched_out_state(u64 prev_state) 1021 { 1022 const char *str = TASK_STATE_TO_CHAR_STR; 1023 1024 return str[prev_state]; 1025 } 1026 1027 static int 1028 add_sched_out_event(struct work_atoms *atoms, 1029 char run_state, 1030 u64 timestamp) 1031 { 1032 struct work_atom *atom = zalloc(sizeof(*atom)); 1033 if (!atom) { 1034 pr_err("Non memory at %s", __func__); 1035 return -1; 1036 } 1037 1038 atom->sched_out_time = timestamp; 1039 1040 if (run_state == 'R') { 1041 atom->state = THREAD_WAIT_CPU; 1042 atom->wake_up_time = atom->sched_out_time; 1043 } 1044 1045 list_add_tail(&atom->list, &atoms->work_list); 1046 return 0; 1047 } 1048 1049 static void 1050 add_runtime_event(struct work_atoms *atoms, u64 delta, 1051 u64 timestamp __maybe_unused) 1052 { 1053 struct work_atom *atom; 1054 1055 BUG_ON(list_empty(&atoms->work_list)); 1056 1057 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1058 1059 atom->runtime += delta; 1060 atoms->total_runtime += delta; 1061 } 1062 1063 static void 1064 add_sched_in_event(struct work_atoms *atoms, u64 timestamp) 1065 { 1066 struct work_atom *atom; 1067 u64 delta; 1068 1069 if (list_empty(&atoms->work_list)) 1070 return; 1071 1072 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1073 1074 if (atom->state != THREAD_WAIT_CPU) 1075 return; 1076 1077 if (timestamp < atom->wake_up_time) { 1078 atom->state = THREAD_IGNORE; 1079 return; 1080 } 1081 1082 atom->state = THREAD_SCHED_IN; 1083 atom->sched_in_time = timestamp; 1084 1085 delta = atom->sched_in_time - atom->wake_up_time; 1086 atoms->total_lat += delta; 1087 if (delta > atoms->max_lat) { 1088 atoms->max_lat = delta; 1089 atoms->max_lat_at = timestamp; 1090 } 1091 atoms->nb_atoms++; 1092 } 1093 1094 static int latency_switch_event(struct perf_sched *sched, 1095 struct perf_evsel *evsel, 1096 struct perf_sample *sample, 1097 struct machine *machine) 1098 { 1099 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 1100 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1101 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); 1102 struct work_atoms *out_events, *in_events; 1103 struct thread *sched_out, *sched_in; 1104 u64 timestamp0, timestamp = sample->time; 1105 int cpu = sample->cpu, err = -1; 1106 s64 delta; 1107 1108 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1109 1110 timestamp0 = sched->cpu_last_switched[cpu]; 1111 sched->cpu_last_switched[cpu] = timestamp; 1112 if (timestamp0) 1113 delta = timestamp - timestamp0; 1114 else 1115 delta = 0; 1116 1117 if (delta < 0) { 1118 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1119 return -1; 1120 } 1121 1122 sched_out = machine__findnew_thread(machine, -1, prev_pid); 1123 sched_in = machine__findnew_thread(machine, -1, next_pid); 1124 if (sched_out == NULL || sched_in == NULL) 1125 goto out_put; 1126 1127 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1128 if (!out_events) { 1129 if (thread_atoms_insert(sched, sched_out)) 1130 goto out_put; 1131 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1132 if (!out_events) { 1133 pr_err("out-event: Internal tree error"); 1134 goto out_put; 1135 } 1136 } 1137 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp)) 1138 return -1; 1139 1140 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1141 if (!in_events) { 1142 if (thread_atoms_insert(sched, sched_in)) 1143 goto out_put; 1144 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1145 if (!in_events) { 1146 pr_err("in-event: Internal tree error"); 1147 goto out_put; 1148 } 1149 /* 1150 * Take came in we have not heard about yet, 1151 * add in an initial atom in runnable state: 1152 */ 1153 if (add_sched_out_event(in_events, 'R', timestamp)) 1154 goto out_put; 1155 } 1156 add_sched_in_event(in_events, timestamp); 1157 err = 0; 1158 out_put: 1159 thread__put(sched_out); 1160 thread__put(sched_in); 1161 return err; 1162 } 1163 1164 static int latency_runtime_event(struct perf_sched *sched, 1165 struct perf_evsel *evsel, 1166 struct perf_sample *sample, 1167 struct machine *machine) 1168 { 1169 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1170 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime"); 1171 struct thread *thread = machine__findnew_thread(machine, -1, pid); 1172 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1173 u64 timestamp = sample->time; 1174 int cpu = sample->cpu, err = -1; 1175 1176 if (thread == NULL) 1177 return -1; 1178 1179 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1180 if (!atoms) { 1181 if (thread_atoms_insert(sched, thread)) 1182 goto out_put; 1183 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1184 if (!atoms) { 1185 pr_err("in-event: Internal tree error"); 1186 goto out_put; 1187 } 1188 if (add_sched_out_event(atoms, 'R', timestamp)) 1189 goto out_put; 1190 } 1191 1192 add_runtime_event(atoms, runtime, timestamp); 1193 err = 0; 1194 out_put: 1195 thread__put(thread); 1196 return err; 1197 } 1198 1199 static int latency_wakeup_event(struct perf_sched *sched, 1200 struct perf_evsel *evsel, 1201 struct perf_sample *sample, 1202 struct machine *machine) 1203 { 1204 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1205 struct work_atoms *atoms; 1206 struct work_atom *atom; 1207 struct thread *wakee; 1208 u64 timestamp = sample->time; 1209 int err = -1; 1210 1211 wakee = machine__findnew_thread(machine, -1, pid); 1212 if (wakee == NULL) 1213 return -1; 1214 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1215 if (!atoms) { 1216 if (thread_atoms_insert(sched, wakee)) 1217 goto out_put; 1218 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1219 if (!atoms) { 1220 pr_err("wakeup-event: Internal tree error"); 1221 goto out_put; 1222 } 1223 if (add_sched_out_event(atoms, 'S', timestamp)) 1224 goto out_put; 1225 } 1226 1227 BUG_ON(list_empty(&atoms->work_list)); 1228 1229 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1230 1231 /* 1232 * As we do not guarantee the wakeup event happens when 1233 * task is out of run queue, also may happen when task is 1234 * on run queue and wakeup only change ->state to TASK_RUNNING, 1235 * then we should not set the ->wake_up_time when wake up a 1236 * task which is on run queue. 1237 * 1238 * You WILL be missing events if you've recorded only 1239 * one CPU, or are only looking at only one, so don't 1240 * skip in this case. 1241 */ 1242 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING) 1243 goto out_ok; 1244 1245 sched->nr_timestamps++; 1246 if (atom->sched_out_time > timestamp) { 1247 sched->nr_unordered_timestamps++; 1248 goto out_ok; 1249 } 1250 1251 atom->state = THREAD_WAIT_CPU; 1252 atom->wake_up_time = timestamp; 1253 out_ok: 1254 err = 0; 1255 out_put: 1256 thread__put(wakee); 1257 return err; 1258 } 1259 1260 static int latency_migrate_task_event(struct perf_sched *sched, 1261 struct perf_evsel *evsel, 1262 struct perf_sample *sample, 1263 struct machine *machine) 1264 { 1265 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1266 u64 timestamp = sample->time; 1267 struct work_atoms *atoms; 1268 struct work_atom *atom; 1269 struct thread *migrant; 1270 int err = -1; 1271 1272 /* 1273 * Only need to worry about migration when profiling one CPU. 1274 */ 1275 if (sched->profile_cpu == -1) 1276 return 0; 1277 1278 migrant = machine__findnew_thread(machine, -1, pid); 1279 if (migrant == NULL) 1280 return -1; 1281 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1282 if (!atoms) { 1283 if (thread_atoms_insert(sched, migrant)) 1284 goto out_put; 1285 register_pid(sched, migrant->tid, thread__comm_str(migrant)); 1286 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1287 if (!atoms) { 1288 pr_err("migration-event: Internal tree error"); 1289 goto out_put; 1290 } 1291 if (add_sched_out_event(atoms, 'R', timestamp)) 1292 goto out_put; 1293 } 1294 1295 BUG_ON(list_empty(&atoms->work_list)); 1296 1297 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1298 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp; 1299 1300 sched->nr_timestamps++; 1301 1302 if (atom->sched_out_time > timestamp) 1303 sched->nr_unordered_timestamps++; 1304 err = 0; 1305 out_put: 1306 thread__put(migrant); 1307 return err; 1308 } 1309 1310 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list) 1311 { 1312 int i; 1313 int ret; 1314 u64 avg; 1315 char max_lat_at[32]; 1316 1317 if (!work_list->nb_atoms) 1318 return; 1319 /* 1320 * Ignore idle threads: 1321 */ 1322 if (!strcmp(thread__comm_str(work_list->thread), "swapper")) 1323 return; 1324 1325 sched->all_runtime += work_list->total_runtime; 1326 sched->all_count += work_list->nb_atoms; 1327 1328 if (work_list->num_merged > 1) 1329 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged); 1330 else 1331 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid); 1332 1333 for (i = 0; i < 24 - ret; i++) 1334 printf(" "); 1335 1336 avg = work_list->total_lat / work_list->nb_atoms; 1337 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at)); 1338 1339 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n", 1340 (double)work_list->total_runtime / NSEC_PER_MSEC, 1341 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC, 1342 (double)work_list->max_lat / NSEC_PER_MSEC, 1343 max_lat_at); 1344 } 1345 1346 static int pid_cmp(struct work_atoms *l, struct work_atoms *r) 1347 { 1348 if (l->thread == r->thread) 1349 return 0; 1350 if (l->thread->tid < r->thread->tid) 1351 return -1; 1352 if (l->thread->tid > r->thread->tid) 1353 return 1; 1354 return (int)(l->thread - r->thread); 1355 } 1356 1357 static int avg_cmp(struct work_atoms *l, struct work_atoms *r) 1358 { 1359 u64 avgl, avgr; 1360 1361 if (!l->nb_atoms) 1362 return -1; 1363 1364 if (!r->nb_atoms) 1365 return 1; 1366 1367 avgl = l->total_lat / l->nb_atoms; 1368 avgr = r->total_lat / r->nb_atoms; 1369 1370 if (avgl < avgr) 1371 return -1; 1372 if (avgl > avgr) 1373 return 1; 1374 1375 return 0; 1376 } 1377 1378 static int max_cmp(struct work_atoms *l, struct work_atoms *r) 1379 { 1380 if (l->max_lat < r->max_lat) 1381 return -1; 1382 if (l->max_lat > r->max_lat) 1383 return 1; 1384 1385 return 0; 1386 } 1387 1388 static int switch_cmp(struct work_atoms *l, struct work_atoms *r) 1389 { 1390 if (l->nb_atoms < r->nb_atoms) 1391 return -1; 1392 if (l->nb_atoms > r->nb_atoms) 1393 return 1; 1394 1395 return 0; 1396 } 1397 1398 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r) 1399 { 1400 if (l->total_runtime < r->total_runtime) 1401 return -1; 1402 if (l->total_runtime > r->total_runtime) 1403 return 1; 1404 1405 return 0; 1406 } 1407 1408 static int sort_dimension__add(const char *tok, struct list_head *list) 1409 { 1410 size_t i; 1411 static struct sort_dimension avg_sort_dimension = { 1412 .name = "avg", 1413 .cmp = avg_cmp, 1414 }; 1415 static struct sort_dimension max_sort_dimension = { 1416 .name = "max", 1417 .cmp = max_cmp, 1418 }; 1419 static struct sort_dimension pid_sort_dimension = { 1420 .name = "pid", 1421 .cmp = pid_cmp, 1422 }; 1423 static struct sort_dimension runtime_sort_dimension = { 1424 .name = "runtime", 1425 .cmp = runtime_cmp, 1426 }; 1427 static struct sort_dimension switch_sort_dimension = { 1428 .name = "switch", 1429 .cmp = switch_cmp, 1430 }; 1431 struct sort_dimension *available_sorts[] = { 1432 &pid_sort_dimension, 1433 &avg_sort_dimension, 1434 &max_sort_dimension, 1435 &switch_sort_dimension, 1436 &runtime_sort_dimension, 1437 }; 1438 1439 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { 1440 if (!strcmp(available_sorts[i]->name, tok)) { 1441 list_add_tail(&available_sorts[i]->list, list); 1442 1443 return 0; 1444 } 1445 } 1446 1447 return -1; 1448 } 1449 1450 static void perf_sched__sort_lat(struct perf_sched *sched) 1451 { 1452 struct rb_node *node; 1453 struct rb_root_cached *root = &sched->atom_root; 1454 again: 1455 for (;;) { 1456 struct work_atoms *data; 1457 node = rb_first_cached(root); 1458 if (!node) 1459 break; 1460 1461 rb_erase_cached(node, root); 1462 data = rb_entry(node, struct work_atoms, node); 1463 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list); 1464 } 1465 if (root == &sched->atom_root) { 1466 root = &sched->merged_atom_root; 1467 goto again; 1468 } 1469 } 1470 1471 static int process_sched_wakeup_event(struct perf_tool *tool, 1472 struct perf_evsel *evsel, 1473 struct perf_sample *sample, 1474 struct machine *machine) 1475 { 1476 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1477 1478 if (sched->tp_handler->wakeup_event) 1479 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine); 1480 1481 return 0; 1482 } 1483 1484 union map_priv { 1485 void *ptr; 1486 bool color; 1487 }; 1488 1489 static bool thread__has_color(struct thread *thread) 1490 { 1491 union map_priv priv = { 1492 .ptr = thread__priv(thread), 1493 }; 1494 1495 return priv.color; 1496 } 1497 1498 static struct thread* 1499 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid) 1500 { 1501 struct thread *thread = machine__findnew_thread(machine, pid, tid); 1502 union map_priv priv = { 1503 .color = false, 1504 }; 1505 1506 if (!sched->map.color_pids || !thread || thread__priv(thread)) 1507 return thread; 1508 1509 if (thread_map__has(sched->map.color_pids, tid)) 1510 priv.color = true; 1511 1512 thread__set_priv(thread, priv.ptr); 1513 return thread; 1514 } 1515 1516 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel, 1517 struct perf_sample *sample, struct machine *machine) 1518 { 1519 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1520 struct thread *sched_in; 1521 struct thread_runtime *tr; 1522 int new_shortname; 1523 u64 timestamp0, timestamp = sample->time; 1524 s64 delta; 1525 int i, this_cpu = sample->cpu; 1526 int cpus_nr; 1527 bool new_cpu = false; 1528 const char *color = PERF_COLOR_NORMAL; 1529 char stimestamp[32]; 1530 1531 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0); 1532 1533 if (this_cpu > sched->max_cpu) 1534 sched->max_cpu = this_cpu; 1535 1536 if (sched->map.comp) { 1537 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS); 1538 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) { 1539 sched->map.comp_cpus[cpus_nr++] = this_cpu; 1540 new_cpu = true; 1541 } 1542 } else 1543 cpus_nr = sched->max_cpu; 1544 1545 timestamp0 = sched->cpu_last_switched[this_cpu]; 1546 sched->cpu_last_switched[this_cpu] = timestamp; 1547 if (timestamp0) 1548 delta = timestamp - timestamp0; 1549 else 1550 delta = 0; 1551 1552 if (delta < 0) { 1553 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1554 return -1; 1555 } 1556 1557 sched_in = map__findnew_thread(sched, machine, -1, next_pid); 1558 if (sched_in == NULL) 1559 return -1; 1560 1561 tr = thread__get_runtime(sched_in); 1562 if (tr == NULL) { 1563 thread__put(sched_in); 1564 return -1; 1565 } 1566 1567 sched->curr_thread[this_cpu] = thread__get(sched_in); 1568 1569 printf(" "); 1570 1571 new_shortname = 0; 1572 if (!tr->shortname[0]) { 1573 if (!strcmp(thread__comm_str(sched_in), "swapper")) { 1574 /* 1575 * Don't allocate a letter-number for swapper:0 1576 * as a shortname. Instead, we use '.' for it. 1577 */ 1578 tr->shortname[0] = '.'; 1579 tr->shortname[1] = ' '; 1580 } else { 1581 tr->shortname[0] = sched->next_shortname1; 1582 tr->shortname[1] = sched->next_shortname2; 1583 1584 if (sched->next_shortname1 < 'Z') { 1585 sched->next_shortname1++; 1586 } else { 1587 sched->next_shortname1 = 'A'; 1588 if (sched->next_shortname2 < '9') 1589 sched->next_shortname2++; 1590 else 1591 sched->next_shortname2 = '0'; 1592 } 1593 } 1594 new_shortname = 1; 1595 } 1596 1597 for (i = 0; i < cpus_nr; i++) { 1598 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i; 1599 struct thread *curr_thread = sched->curr_thread[cpu]; 1600 struct thread_runtime *curr_tr; 1601 const char *pid_color = color; 1602 const char *cpu_color = color; 1603 1604 if (curr_thread && thread__has_color(curr_thread)) 1605 pid_color = COLOR_PIDS; 1606 1607 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu)) 1608 continue; 1609 1610 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu)) 1611 cpu_color = COLOR_CPUS; 1612 1613 if (cpu != this_cpu) 1614 color_fprintf(stdout, color, " "); 1615 else 1616 color_fprintf(stdout, cpu_color, "*"); 1617 1618 if (sched->curr_thread[cpu]) { 1619 curr_tr = thread__get_runtime(sched->curr_thread[cpu]); 1620 if (curr_tr == NULL) { 1621 thread__put(sched_in); 1622 return -1; 1623 } 1624 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname); 1625 } else 1626 color_fprintf(stdout, color, " "); 1627 } 1628 1629 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu)) 1630 goto out; 1631 1632 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp)); 1633 color_fprintf(stdout, color, " %12s secs ", stimestamp); 1634 if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) { 1635 const char *pid_color = color; 1636 1637 if (thread__has_color(sched_in)) 1638 pid_color = COLOR_PIDS; 1639 1640 color_fprintf(stdout, pid_color, "%s => %s:%d", 1641 tr->shortname, thread__comm_str(sched_in), sched_in->tid); 1642 tr->comm_changed = false; 1643 } 1644 1645 if (sched->map.comp && new_cpu) 1646 color_fprintf(stdout, color, " (CPU %d)", this_cpu); 1647 1648 out: 1649 color_fprintf(stdout, color, "\n"); 1650 1651 thread__put(sched_in); 1652 1653 return 0; 1654 } 1655 1656 static int process_sched_switch_event(struct perf_tool *tool, 1657 struct perf_evsel *evsel, 1658 struct perf_sample *sample, 1659 struct machine *machine) 1660 { 1661 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1662 int this_cpu = sample->cpu, err = 0; 1663 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 1664 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1665 1666 if (sched->curr_pid[this_cpu] != (u32)-1) { 1667 /* 1668 * Are we trying to switch away a PID that is 1669 * not current? 1670 */ 1671 if (sched->curr_pid[this_cpu] != prev_pid) 1672 sched->nr_context_switch_bugs++; 1673 } 1674 1675 if (sched->tp_handler->switch_event) 1676 err = sched->tp_handler->switch_event(sched, evsel, sample, machine); 1677 1678 sched->curr_pid[this_cpu] = next_pid; 1679 return err; 1680 } 1681 1682 static int process_sched_runtime_event(struct perf_tool *tool, 1683 struct perf_evsel *evsel, 1684 struct perf_sample *sample, 1685 struct machine *machine) 1686 { 1687 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1688 1689 if (sched->tp_handler->runtime_event) 1690 return sched->tp_handler->runtime_event(sched, evsel, sample, machine); 1691 1692 return 0; 1693 } 1694 1695 static int perf_sched__process_fork_event(struct perf_tool *tool, 1696 union perf_event *event, 1697 struct perf_sample *sample, 1698 struct machine *machine) 1699 { 1700 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1701 1702 /* run the fork event through the perf machineruy */ 1703 perf_event__process_fork(tool, event, sample, machine); 1704 1705 /* and then run additional processing needed for this command */ 1706 if (sched->tp_handler->fork_event) 1707 return sched->tp_handler->fork_event(sched, event, machine); 1708 1709 return 0; 1710 } 1711 1712 static int process_sched_migrate_task_event(struct perf_tool *tool, 1713 struct perf_evsel *evsel, 1714 struct perf_sample *sample, 1715 struct machine *machine) 1716 { 1717 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1718 1719 if (sched->tp_handler->migrate_task_event) 1720 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine); 1721 1722 return 0; 1723 } 1724 1725 typedef int (*tracepoint_handler)(struct perf_tool *tool, 1726 struct perf_evsel *evsel, 1727 struct perf_sample *sample, 1728 struct machine *machine); 1729 1730 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused, 1731 union perf_event *event __maybe_unused, 1732 struct perf_sample *sample, 1733 struct perf_evsel *evsel, 1734 struct machine *machine) 1735 { 1736 int err = 0; 1737 1738 if (evsel->handler != NULL) { 1739 tracepoint_handler f = evsel->handler; 1740 err = f(tool, evsel, sample, machine); 1741 } 1742 1743 return err; 1744 } 1745 1746 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused, 1747 union perf_event *event, 1748 struct perf_sample *sample, 1749 struct machine *machine) 1750 { 1751 struct thread *thread; 1752 struct thread_runtime *tr; 1753 int err; 1754 1755 err = perf_event__process_comm(tool, event, sample, machine); 1756 if (err) 1757 return err; 1758 1759 thread = machine__find_thread(machine, sample->pid, sample->tid); 1760 if (!thread) { 1761 pr_err("Internal error: can't find thread\n"); 1762 return -1; 1763 } 1764 1765 tr = thread__get_runtime(thread); 1766 if (tr == NULL) { 1767 thread__put(thread); 1768 return -1; 1769 } 1770 1771 tr->comm_changed = true; 1772 thread__put(thread); 1773 1774 return 0; 1775 } 1776 1777 static int perf_sched__read_events(struct perf_sched *sched) 1778 { 1779 const struct perf_evsel_str_handler handlers[] = { 1780 { "sched:sched_switch", process_sched_switch_event, }, 1781 { "sched:sched_stat_runtime", process_sched_runtime_event, }, 1782 { "sched:sched_wakeup", process_sched_wakeup_event, }, 1783 { "sched:sched_wakeup_new", process_sched_wakeup_event, }, 1784 { "sched:sched_migrate_task", process_sched_migrate_task_event, }, 1785 }; 1786 struct perf_session *session; 1787 struct perf_data data = { 1788 .file = { 1789 .path = input_name, 1790 }, 1791 .mode = PERF_DATA_MODE_READ, 1792 .force = sched->force, 1793 }; 1794 int rc = -1; 1795 1796 session = perf_session__new(&data, false, &sched->tool); 1797 if (session == NULL) { 1798 pr_debug("No Memory for session\n"); 1799 return -1; 1800 } 1801 1802 symbol__init(&session->header.env); 1803 1804 if (perf_session__set_tracepoints_handlers(session, handlers)) 1805 goto out_delete; 1806 1807 if (perf_session__has_traces(session, "record -R")) { 1808 int err = perf_session__process_events(session); 1809 if (err) { 1810 pr_err("Failed to process events, error %d", err); 1811 goto out_delete; 1812 } 1813 1814 sched->nr_events = session->evlist->stats.nr_events[0]; 1815 sched->nr_lost_events = session->evlist->stats.total_lost; 1816 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST]; 1817 } 1818 1819 rc = 0; 1820 out_delete: 1821 perf_session__delete(session); 1822 return rc; 1823 } 1824 1825 /* 1826 * scheduling times are printed as msec.usec 1827 */ 1828 static inline void print_sched_time(unsigned long long nsecs, int width) 1829 { 1830 unsigned long msecs; 1831 unsigned long usecs; 1832 1833 msecs = nsecs / NSEC_PER_MSEC; 1834 nsecs -= msecs * NSEC_PER_MSEC; 1835 usecs = nsecs / NSEC_PER_USEC; 1836 printf("%*lu.%03lu ", width, msecs, usecs); 1837 } 1838 1839 /* 1840 * returns runtime data for event, allocating memory for it the 1841 * first time it is used. 1842 */ 1843 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel) 1844 { 1845 struct evsel_runtime *r = evsel->priv; 1846 1847 if (r == NULL) { 1848 r = zalloc(sizeof(struct evsel_runtime)); 1849 evsel->priv = r; 1850 } 1851 1852 return r; 1853 } 1854 1855 /* 1856 * save last time event was seen per cpu 1857 */ 1858 static void perf_evsel__save_time(struct perf_evsel *evsel, 1859 u64 timestamp, u32 cpu) 1860 { 1861 struct evsel_runtime *r = perf_evsel__get_runtime(evsel); 1862 1863 if (r == NULL) 1864 return; 1865 1866 if ((cpu >= r->ncpu) || (r->last_time == NULL)) { 1867 int i, n = __roundup_pow_of_two(cpu+1); 1868 void *p = r->last_time; 1869 1870 p = realloc(r->last_time, n * sizeof(u64)); 1871 if (!p) 1872 return; 1873 1874 r->last_time = p; 1875 for (i = r->ncpu; i < n; ++i) 1876 r->last_time[i] = (u64) 0; 1877 1878 r->ncpu = n; 1879 } 1880 1881 r->last_time[cpu] = timestamp; 1882 } 1883 1884 /* returns last time this event was seen on the given cpu */ 1885 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu) 1886 { 1887 struct evsel_runtime *r = perf_evsel__get_runtime(evsel); 1888 1889 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu)) 1890 return 0; 1891 1892 return r->last_time[cpu]; 1893 } 1894 1895 static int comm_width = 30; 1896 1897 static char *timehist_get_commstr(struct thread *thread) 1898 { 1899 static char str[32]; 1900 const char *comm = thread__comm_str(thread); 1901 pid_t tid = thread->tid; 1902 pid_t pid = thread->pid_; 1903 int n; 1904 1905 if (pid == 0) 1906 n = scnprintf(str, sizeof(str), "%s", comm); 1907 1908 else if (tid != pid) 1909 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid); 1910 1911 else 1912 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid); 1913 1914 if (n > comm_width) 1915 comm_width = n; 1916 1917 return str; 1918 } 1919 1920 static void timehist_header(struct perf_sched *sched) 1921 { 1922 u32 ncpus = sched->max_cpu + 1; 1923 u32 i, j; 1924 1925 printf("%15s %6s ", "time", "cpu"); 1926 1927 if (sched->show_cpu_visual) { 1928 printf(" "); 1929 for (i = 0, j = 0; i < ncpus; ++i) { 1930 printf("%x", j++); 1931 if (j > 15) 1932 j = 0; 1933 } 1934 printf(" "); 1935 } 1936 1937 printf(" %-*s %9s %9s %9s", comm_width, 1938 "task name", "wait time", "sch delay", "run time"); 1939 1940 if (sched->show_state) 1941 printf(" %s", "state"); 1942 1943 printf("\n"); 1944 1945 /* 1946 * units row 1947 */ 1948 printf("%15s %-6s ", "", ""); 1949 1950 if (sched->show_cpu_visual) 1951 printf(" %*s ", ncpus, ""); 1952 1953 printf(" %-*s %9s %9s %9s", comm_width, 1954 "[tid/pid]", "(msec)", "(msec)", "(msec)"); 1955 1956 if (sched->show_state) 1957 printf(" %5s", ""); 1958 1959 printf("\n"); 1960 1961 /* 1962 * separator 1963 */ 1964 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line); 1965 1966 if (sched->show_cpu_visual) 1967 printf(" %.*s ", ncpus, graph_dotted_line); 1968 1969 printf(" %.*s %.9s %.9s %.9s", comm_width, 1970 graph_dotted_line, graph_dotted_line, graph_dotted_line, 1971 graph_dotted_line); 1972 1973 if (sched->show_state) 1974 printf(" %.5s", graph_dotted_line); 1975 1976 printf("\n"); 1977 } 1978 1979 static char task_state_char(struct thread *thread, int state) 1980 { 1981 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR; 1982 unsigned bit = state ? ffs(state) : 0; 1983 1984 /* 'I' for idle */ 1985 if (thread->tid == 0) 1986 return 'I'; 1987 1988 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?'; 1989 } 1990 1991 static void timehist_print_sample(struct perf_sched *sched, 1992 struct perf_evsel *evsel, 1993 struct perf_sample *sample, 1994 struct addr_location *al, 1995 struct thread *thread, 1996 u64 t, int state) 1997 { 1998 struct thread_runtime *tr = thread__priv(thread); 1999 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); 2000 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 2001 u32 max_cpus = sched->max_cpu + 1; 2002 char tstr[64]; 2003 char nstr[30]; 2004 u64 wait_time; 2005 2006 timestamp__scnprintf_usec(t, tstr, sizeof(tstr)); 2007 printf("%15s [%04d] ", tstr, sample->cpu); 2008 2009 if (sched->show_cpu_visual) { 2010 u32 i; 2011 char c; 2012 2013 printf(" "); 2014 for (i = 0; i < max_cpus; ++i) { 2015 /* flag idle times with 'i'; others are sched events */ 2016 if (i == sample->cpu) 2017 c = (thread->tid == 0) ? 'i' : 's'; 2018 else 2019 c = ' '; 2020 printf("%c", c); 2021 } 2022 printf(" "); 2023 } 2024 2025 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2026 2027 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt; 2028 print_sched_time(wait_time, 6); 2029 2030 print_sched_time(tr->dt_delay, 6); 2031 print_sched_time(tr->dt_run, 6); 2032 2033 if (sched->show_state) 2034 printf(" %5c ", task_state_char(thread, state)); 2035 2036 if (sched->show_next) { 2037 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid); 2038 printf(" %-*s", comm_width, nstr); 2039 } 2040 2041 if (sched->show_wakeups && !sched->show_next) 2042 printf(" %-*s", comm_width, ""); 2043 2044 if (thread->tid == 0) 2045 goto out; 2046 2047 if (sched->show_callchain) 2048 printf(" "); 2049 2050 sample__fprintf_sym(sample, al, 0, 2051 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE | 2052 EVSEL__PRINT_CALLCHAIN_ARROW | 2053 EVSEL__PRINT_SKIP_IGNORED, 2054 &callchain_cursor, stdout); 2055 2056 out: 2057 printf("\n"); 2058 } 2059 2060 /* 2061 * Explanation of delta-time stats: 2062 * 2063 * t = time of current schedule out event 2064 * tprev = time of previous sched out event 2065 * also time of schedule-in event for current task 2066 * last_time = time of last sched change event for current task 2067 * (i.e, time process was last scheduled out) 2068 * ready_to_run = time of wakeup for current task 2069 * 2070 * -----|------------|------------|------------|------ 2071 * last ready tprev t 2072 * time to run 2073 * 2074 * |-------- dt_wait --------| 2075 * |- dt_delay -|-- dt_run --| 2076 * 2077 * dt_run = run time of current task 2078 * dt_wait = time between last schedule out event for task and tprev 2079 * represents time spent off the cpu 2080 * dt_delay = time between wakeup and schedule-in of task 2081 */ 2082 2083 static void timehist_update_runtime_stats(struct thread_runtime *r, 2084 u64 t, u64 tprev) 2085 { 2086 r->dt_delay = 0; 2087 r->dt_sleep = 0; 2088 r->dt_iowait = 0; 2089 r->dt_preempt = 0; 2090 r->dt_run = 0; 2091 2092 if (tprev) { 2093 r->dt_run = t - tprev; 2094 if (r->ready_to_run) { 2095 if (r->ready_to_run > tprev) 2096 pr_debug("time travel: wakeup time for task > previous sched_switch event\n"); 2097 else 2098 r->dt_delay = tprev - r->ready_to_run; 2099 } 2100 2101 if (r->last_time > tprev) 2102 pr_debug("time travel: last sched out time for task > previous sched_switch event\n"); 2103 else if (r->last_time) { 2104 u64 dt_wait = tprev - r->last_time; 2105 2106 if (r->last_state == TASK_RUNNING) 2107 r->dt_preempt = dt_wait; 2108 else if (r->last_state == TASK_UNINTERRUPTIBLE) 2109 r->dt_iowait = dt_wait; 2110 else 2111 r->dt_sleep = dt_wait; 2112 } 2113 } 2114 2115 update_stats(&r->run_stats, r->dt_run); 2116 2117 r->total_run_time += r->dt_run; 2118 r->total_delay_time += r->dt_delay; 2119 r->total_sleep_time += r->dt_sleep; 2120 r->total_iowait_time += r->dt_iowait; 2121 r->total_preempt_time += r->dt_preempt; 2122 } 2123 2124 static bool is_idle_sample(struct perf_sample *sample, 2125 struct perf_evsel *evsel) 2126 { 2127 /* pid 0 == swapper == idle task */ 2128 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0) 2129 return perf_evsel__intval(evsel, sample, "prev_pid") == 0; 2130 2131 return sample->pid == 0; 2132 } 2133 2134 static void save_task_callchain(struct perf_sched *sched, 2135 struct perf_sample *sample, 2136 struct perf_evsel *evsel, 2137 struct machine *machine) 2138 { 2139 struct callchain_cursor *cursor = &callchain_cursor; 2140 struct thread *thread; 2141 2142 /* want main thread for process - has maps */ 2143 thread = machine__findnew_thread(machine, sample->pid, sample->pid); 2144 if (thread == NULL) { 2145 pr_debug("Failed to get thread for pid %d.\n", sample->pid); 2146 return; 2147 } 2148 2149 if (!sched->show_callchain || sample->callchain == NULL) 2150 return; 2151 2152 if (thread__resolve_callchain(thread, cursor, evsel, sample, 2153 NULL, NULL, sched->max_stack + 2) != 0) { 2154 if (verbose > 0) 2155 pr_err("Failed to resolve callchain. Skipping\n"); 2156 2157 return; 2158 } 2159 2160 callchain_cursor_commit(cursor); 2161 2162 while (true) { 2163 struct callchain_cursor_node *node; 2164 struct symbol *sym; 2165 2166 node = callchain_cursor_current(cursor); 2167 if (node == NULL) 2168 break; 2169 2170 sym = node->sym; 2171 if (sym) { 2172 if (!strcmp(sym->name, "schedule") || 2173 !strcmp(sym->name, "__schedule") || 2174 !strcmp(sym->name, "preempt_schedule")) 2175 sym->ignore = 1; 2176 } 2177 2178 callchain_cursor_advance(cursor); 2179 } 2180 } 2181 2182 static int init_idle_thread(struct thread *thread) 2183 { 2184 struct idle_thread_runtime *itr; 2185 2186 thread__set_comm(thread, idle_comm, 0); 2187 2188 itr = zalloc(sizeof(*itr)); 2189 if (itr == NULL) 2190 return -ENOMEM; 2191 2192 init_stats(&itr->tr.run_stats); 2193 callchain_init(&itr->callchain); 2194 callchain_cursor_reset(&itr->cursor); 2195 thread__set_priv(thread, itr); 2196 2197 return 0; 2198 } 2199 2200 /* 2201 * Track idle stats per cpu by maintaining a local thread 2202 * struct for the idle task on each cpu. 2203 */ 2204 static int init_idle_threads(int ncpu) 2205 { 2206 int i, ret; 2207 2208 idle_threads = zalloc(ncpu * sizeof(struct thread *)); 2209 if (!idle_threads) 2210 return -ENOMEM; 2211 2212 idle_max_cpu = ncpu; 2213 2214 /* allocate the actual thread struct if needed */ 2215 for (i = 0; i < ncpu; ++i) { 2216 idle_threads[i] = thread__new(0, 0); 2217 if (idle_threads[i] == NULL) 2218 return -ENOMEM; 2219 2220 ret = init_idle_thread(idle_threads[i]); 2221 if (ret < 0) 2222 return ret; 2223 } 2224 2225 return 0; 2226 } 2227 2228 static void free_idle_threads(void) 2229 { 2230 int i; 2231 2232 if (idle_threads == NULL) 2233 return; 2234 2235 for (i = 0; i < idle_max_cpu; ++i) { 2236 if ((idle_threads[i])) 2237 thread__delete(idle_threads[i]); 2238 } 2239 2240 free(idle_threads); 2241 } 2242 2243 static struct thread *get_idle_thread(int cpu) 2244 { 2245 /* 2246 * expand/allocate array of pointers to local thread 2247 * structs if needed 2248 */ 2249 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) { 2250 int i, j = __roundup_pow_of_two(cpu+1); 2251 void *p; 2252 2253 p = realloc(idle_threads, j * sizeof(struct thread *)); 2254 if (!p) 2255 return NULL; 2256 2257 idle_threads = (struct thread **) p; 2258 for (i = idle_max_cpu; i < j; ++i) 2259 idle_threads[i] = NULL; 2260 2261 idle_max_cpu = j; 2262 } 2263 2264 /* allocate a new thread struct if needed */ 2265 if (idle_threads[cpu] == NULL) { 2266 idle_threads[cpu] = thread__new(0, 0); 2267 2268 if (idle_threads[cpu]) { 2269 if (init_idle_thread(idle_threads[cpu]) < 0) 2270 return NULL; 2271 } 2272 } 2273 2274 return idle_threads[cpu]; 2275 } 2276 2277 static void save_idle_callchain(struct perf_sched *sched, 2278 struct idle_thread_runtime *itr, 2279 struct perf_sample *sample) 2280 { 2281 if (!sched->show_callchain || sample->callchain == NULL) 2282 return; 2283 2284 callchain_cursor__copy(&itr->cursor, &callchain_cursor); 2285 } 2286 2287 static struct thread *timehist_get_thread(struct perf_sched *sched, 2288 struct perf_sample *sample, 2289 struct machine *machine, 2290 struct perf_evsel *evsel) 2291 { 2292 struct thread *thread; 2293 2294 if (is_idle_sample(sample, evsel)) { 2295 thread = get_idle_thread(sample->cpu); 2296 if (thread == NULL) 2297 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2298 2299 } else { 2300 /* there were samples with tid 0 but non-zero pid */ 2301 thread = machine__findnew_thread(machine, sample->pid, 2302 sample->tid ?: sample->pid); 2303 if (thread == NULL) { 2304 pr_debug("Failed to get thread for tid %d. skipping sample.\n", 2305 sample->tid); 2306 } 2307 2308 save_task_callchain(sched, sample, evsel, machine); 2309 if (sched->idle_hist) { 2310 struct thread *idle; 2311 struct idle_thread_runtime *itr; 2312 2313 idle = get_idle_thread(sample->cpu); 2314 if (idle == NULL) { 2315 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2316 return NULL; 2317 } 2318 2319 itr = thread__priv(idle); 2320 if (itr == NULL) 2321 return NULL; 2322 2323 itr->last_thread = thread; 2324 2325 /* copy task callchain when entering to idle */ 2326 if (perf_evsel__intval(evsel, sample, "next_pid") == 0) 2327 save_idle_callchain(sched, itr, sample); 2328 } 2329 } 2330 2331 return thread; 2332 } 2333 2334 static bool timehist_skip_sample(struct perf_sched *sched, 2335 struct thread *thread, 2336 struct perf_evsel *evsel, 2337 struct perf_sample *sample) 2338 { 2339 bool rc = false; 2340 2341 if (thread__is_filtered(thread)) { 2342 rc = true; 2343 sched->skipped_samples++; 2344 } 2345 2346 if (sched->idle_hist) { 2347 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch")) 2348 rc = true; 2349 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 && 2350 perf_evsel__intval(evsel, sample, "next_pid") != 0) 2351 rc = true; 2352 } 2353 2354 return rc; 2355 } 2356 2357 static void timehist_print_wakeup_event(struct perf_sched *sched, 2358 struct perf_evsel *evsel, 2359 struct perf_sample *sample, 2360 struct machine *machine, 2361 struct thread *awakened) 2362 { 2363 struct thread *thread; 2364 char tstr[64]; 2365 2366 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2367 if (thread == NULL) 2368 return; 2369 2370 /* show wakeup unless both awakee and awaker are filtered */ 2371 if (timehist_skip_sample(sched, thread, evsel, sample) && 2372 timehist_skip_sample(sched, awakened, evsel, sample)) { 2373 return; 2374 } 2375 2376 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2377 printf("%15s [%04d] ", tstr, sample->cpu); 2378 if (sched->show_cpu_visual) 2379 printf(" %*s ", sched->max_cpu + 1, ""); 2380 2381 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2382 2383 /* dt spacer */ 2384 printf(" %9s %9s %9s ", "", "", ""); 2385 2386 printf("awakened: %s", timehist_get_commstr(awakened)); 2387 2388 printf("\n"); 2389 } 2390 2391 static int timehist_sched_wakeup_event(struct perf_tool *tool, 2392 union perf_event *event __maybe_unused, 2393 struct perf_evsel *evsel, 2394 struct perf_sample *sample, 2395 struct machine *machine) 2396 { 2397 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2398 struct thread *thread; 2399 struct thread_runtime *tr = NULL; 2400 /* want pid of awakened task not pid in sample */ 2401 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 2402 2403 thread = machine__findnew_thread(machine, 0, pid); 2404 if (thread == NULL) 2405 return -1; 2406 2407 tr = thread__get_runtime(thread); 2408 if (tr == NULL) 2409 return -1; 2410 2411 if (tr->ready_to_run == 0) 2412 tr->ready_to_run = sample->time; 2413 2414 /* show wakeups if requested */ 2415 if (sched->show_wakeups && 2416 !perf_time__skip_sample(&sched->ptime, sample->time)) 2417 timehist_print_wakeup_event(sched, evsel, sample, machine, thread); 2418 2419 return 0; 2420 } 2421 2422 static void timehist_print_migration_event(struct perf_sched *sched, 2423 struct perf_evsel *evsel, 2424 struct perf_sample *sample, 2425 struct machine *machine, 2426 struct thread *migrated) 2427 { 2428 struct thread *thread; 2429 char tstr[64]; 2430 u32 max_cpus = sched->max_cpu + 1; 2431 u32 ocpu, dcpu; 2432 2433 if (sched->summary_only) 2434 return; 2435 2436 max_cpus = sched->max_cpu + 1; 2437 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu"); 2438 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu"); 2439 2440 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2441 if (thread == NULL) 2442 return; 2443 2444 if (timehist_skip_sample(sched, thread, evsel, sample) && 2445 timehist_skip_sample(sched, migrated, evsel, sample)) { 2446 return; 2447 } 2448 2449 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2450 printf("%15s [%04d] ", tstr, sample->cpu); 2451 2452 if (sched->show_cpu_visual) { 2453 u32 i; 2454 char c; 2455 2456 printf(" "); 2457 for (i = 0; i < max_cpus; ++i) { 2458 c = (i == sample->cpu) ? 'm' : ' '; 2459 printf("%c", c); 2460 } 2461 printf(" "); 2462 } 2463 2464 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2465 2466 /* dt spacer */ 2467 printf(" %9s %9s %9s ", "", "", ""); 2468 2469 printf("migrated: %s", timehist_get_commstr(migrated)); 2470 printf(" cpu %d => %d", ocpu, dcpu); 2471 2472 printf("\n"); 2473 } 2474 2475 static int timehist_migrate_task_event(struct perf_tool *tool, 2476 union perf_event *event __maybe_unused, 2477 struct perf_evsel *evsel, 2478 struct perf_sample *sample, 2479 struct machine *machine) 2480 { 2481 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2482 struct thread *thread; 2483 struct thread_runtime *tr = NULL; 2484 /* want pid of migrated task not pid in sample */ 2485 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 2486 2487 thread = machine__findnew_thread(machine, 0, pid); 2488 if (thread == NULL) 2489 return -1; 2490 2491 tr = thread__get_runtime(thread); 2492 if (tr == NULL) 2493 return -1; 2494 2495 tr->migrations++; 2496 2497 /* show migrations if requested */ 2498 timehist_print_migration_event(sched, evsel, sample, machine, thread); 2499 2500 return 0; 2501 } 2502 2503 static int timehist_sched_change_event(struct perf_tool *tool, 2504 union perf_event *event, 2505 struct perf_evsel *evsel, 2506 struct perf_sample *sample, 2507 struct machine *machine) 2508 { 2509 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2510 struct perf_time_interval *ptime = &sched->ptime; 2511 struct addr_location al; 2512 struct thread *thread; 2513 struct thread_runtime *tr = NULL; 2514 u64 tprev, t = sample->time; 2515 int rc = 0; 2516 int state = perf_evsel__intval(evsel, sample, "prev_state"); 2517 2518 2519 if (machine__resolve(machine, &al, sample) < 0) { 2520 pr_err("problem processing %d event. skipping it\n", 2521 event->header.type); 2522 rc = -1; 2523 goto out; 2524 } 2525 2526 thread = timehist_get_thread(sched, sample, machine, evsel); 2527 if (thread == NULL) { 2528 rc = -1; 2529 goto out; 2530 } 2531 2532 if (timehist_skip_sample(sched, thread, evsel, sample)) 2533 goto out; 2534 2535 tr = thread__get_runtime(thread); 2536 if (tr == NULL) { 2537 rc = -1; 2538 goto out; 2539 } 2540 2541 tprev = perf_evsel__get_time(evsel, sample->cpu); 2542 2543 /* 2544 * If start time given: 2545 * - sample time is under window user cares about - skip sample 2546 * - tprev is under window user cares about - reset to start of window 2547 */ 2548 if (ptime->start && ptime->start > t) 2549 goto out; 2550 2551 if (tprev && ptime->start > tprev) 2552 tprev = ptime->start; 2553 2554 /* 2555 * If end time given: 2556 * - previous sched event is out of window - we are done 2557 * - sample time is beyond window user cares about - reset it 2558 * to close out stats for time window interest 2559 */ 2560 if (ptime->end) { 2561 if (tprev > ptime->end) 2562 goto out; 2563 2564 if (t > ptime->end) 2565 t = ptime->end; 2566 } 2567 2568 if (!sched->idle_hist || thread->tid == 0) { 2569 timehist_update_runtime_stats(tr, t, tprev); 2570 2571 if (sched->idle_hist) { 2572 struct idle_thread_runtime *itr = (void *)tr; 2573 struct thread_runtime *last_tr; 2574 2575 BUG_ON(thread->tid != 0); 2576 2577 if (itr->last_thread == NULL) 2578 goto out; 2579 2580 /* add current idle time as last thread's runtime */ 2581 last_tr = thread__get_runtime(itr->last_thread); 2582 if (last_tr == NULL) 2583 goto out; 2584 2585 timehist_update_runtime_stats(last_tr, t, tprev); 2586 /* 2587 * remove delta time of last thread as it's not updated 2588 * and otherwise it will show an invalid value next 2589 * time. we only care total run time and run stat. 2590 */ 2591 last_tr->dt_run = 0; 2592 last_tr->dt_delay = 0; 2593 last_tr->dt_sleep = 0; 2594 last_tr->dt_iowait = 0; 2595 last_tr->dt_preempt = 0; 2596 2597 if (itr->cursor.nr) 2598 callchain_append(&itr->callchain, &itr->cursor, t - tprev); 2599 2600 itr->last_thread = NULL; 2601 } 2602 } 2603 2604 if (!sched->summary_only) 2605 timehist_print_sample(sched, evsel, sample, &al, thread, t, state); 2606 2607 out: 2608 if (sched->hist_time.start == 0 && t >= ptime->start) 2609 sched->hist_time.start = t; 2610 if (ptime->end == 0 || t <= ptime->end) 2611 sched->hist_time.end = t; 2612 2613 if (tr) { 2614 /* time of this sched_switch event becomes last time task seen */ 2615 tr->last_time = sample->time; 2616 2617 /* last state is used to determine where to account wait time */ 2618 tr->last_state = state; 2619 2620 /* sched out event for task so reset ready to run time */ 2621 tr->ready_to_run = 0; 2622 } 2623 2624 perf_evsel__save_time(evsel, sample->time, sample->cpu); 2625 2626 return rc; 2627 } 2628 2629 static int timehist_sched_switch_event(struct perf_tool *tool, 2630 union perf_event *event, 2631 struct perf_evsel *evsel, 2632 struct perf_sample *sample, 2633 struct machine *machine __maybe_unused) 2634 { 2635 return timehist_sched_change_event(tool, event, evsel, sample, machine); 2636 } 2637 2638 static int process_lost(struct perf_tool *tool __maybe_unused, 2639 union perf_event *event, 2640 struct perf_sample *sample, 2641 struct machine *machine __maybe_unused) 2642 { 2643 char tstr[64]; 2644 2645 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2646 printf("%15s ", tstr); 2647 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu); 2648 2649 return 0; 2650 } 2651 2652 2653 static void print_thread_runtime(struct thread *t, 2654 struct thread_runtime *r) 2655 { 2656 double mean = avg_stats(&r->run_stats); 2657 float stddev; 2658 2659 printf("%*s %5d %9" PRIu64 " ", 2660 comm_width, timehist_get_commstr(t), t->ppid, 2661 (u64) r->run_stats.n); 2662 2663 print_sched_time(r->total_run_time, 8); 2664 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean); 2665 print_sched_time(r->run_stats.min, 6); 2666 printf(" "); 2667 print_sched_time((u64) mean, 6); 2668 printf(" "); 2669 print_sched_time(r->run_stats.max, 6); 2670 printf(" "); 2671 printf("%5.2f", stddev); 2672 printf(" %5" PRIu64, r->migrations); 2673 printf("\n"); 2674 } 2675 2676 static void print_thread_waittime(struct thread *t, 2677 struct thread_runtime *r) 2678 { 2679 printf("%*s %5d %9" PRIu64 " ", 2680 comm_width, timehist_get_commstr(t), t->ppid, 2681 (u64) r->run_stats.n); 2682 2683 print_sched_time(r->total_run_time, 8); 2684 print_sched_time(r->total_sleep_time, 6); 2685 printf(" "); 2686 print_sched_time(r->total_iowait_time, 6); 2687 printf(" "); 2688 print_sched_time(r->total_preempt_time, 6); 2689 printf(" "); 2690 print_sched_time(r->total_delay_time, 6); 2691 printf("\n"); 2692 } 2693 2694 struct total_run_stats { 2695 struct perf_sched *sched; 2696 u64 sched_count; 2697 u64 task_count; 2698 u64 total_run_time; 2699 }; 2700 2701 static int __show_thread_runtime(struct thread *t, void *priv) 2702 { 2703 struct total_run_stats *stats = priv; 2704 struct thread_runtime *r; 2705 2706 if (thread__is_filtered(t)) 2707 return 0; 2708 2709 r = thread__priv(t); 2710 if (r && r->run_stats.n) { 2711 stats->task_count++; 2712 stats->sched_count += r->run_stats.n; 2713 stats->total_run_time += r->total_run_time; 2714 2715 if (stats->sched->show_state) 2716 print_thread_waittime(t, r); 2717 else 2718 print_thread_runtime(t, r); 2719 } 2720 2721 return 0; 2722 } 2723 2724 static int show_thread_runtime(struct thread *t, void *priv) 2725 { 2726 if (t->dead) 2727 return 0; 2728 2729 return __show_thread_runtime(t, priv); 2730 } 2731 2732 static int show_deadthread_runtime(struct thread *t, void *priv) 2733 { 2734 if (!t->dead) 2735 return 0; 2736 2737 return __show_thread_runtime(t, priv); 2738 } 2739 2740 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node) 2741 { 2742 const char *sep = " <- "; 2743 struct callchain_list *chain; 2744 size_t ret = 0; 2745 char bf[1024]; 2746 bool first; 2747 2748 if (node == NULL) 2749 return 0; 2750 2751 ret = callchain__fprintf_folded(fp, node->parent); 2752 first = (ret == 0); 2753 2754 list_for_each_entry(chain, &node->val, list) { 2755 if (chain->ip >= PERF_CONTEXT_MAX) 2756 continue; 2757 if (chain->ms.sym && chain->ms.sym->ignore) 2758 continue; 2759 ret += fprintf(fp, "%s%s", first ? "" : sep, 2760 callchain_list__sym_name(chain, bf, sizeof(bf), 2761 false)); 2762 first = false; 2763 } 2764 2765 return ret; 2766 } 2767 2768 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root) 2769 { 2770 size_t ret = 0; 2771 FILE *fp = stdout; 2772 struct callchain_node *chain; 2773 struct rb_node *rb_node = rb_first_cached(root); 2774 2775 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains"); 2776 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line, 2777 graph_dotted_line); 2778 2779 while (rb_node) { 2780 chain = rb_entry(rb_node, struct callchain_node, rb_node); 2781 rb_node = rb_next(rb_node); 2782 2783 ret += fprintf(fp, " "); 2784 print_sched_time(chain->hit, 12); 2785 ret += 16; /* print_sched_time returns 2nd arg + 4 */ 2786 ret += fprintf(fp, " %8d ", chain->count); 2787 ret += callchain__fprintf_folded(fp, chain); 2788 ret += fprintf(fp, "\n"); 2789 } 2790 2791 return ret; 2792 } 2793 2794 static void timehist_print_summary(struct perf_sched *sched, 2795 struct perf_session *session) 2796 { 2797 struct machine *m = &session->machines.host; 2798 struct total_run_stats totals; 2799 u64 task_count; 2800 struct thread *t; 2801 struct thread_runtime *r; 2802 int i; 2803 u64 hist_time = sched->hist_time.end - sched->hist_time.start; 2804 2805 memset(&totals, 0, sizeof(totals)); 2806 totals.sched = sched; 2807 2808 if (sched->idle_hist) { 2809 printf("\nIdle-time summary\n"); 2810 printf("%*s parent sched-out ", comm_width, "comm"); 2811 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n"); 2812 } else if (sched->show_state) { 2813 printf("\nWait-time summary\n"); 2814 printf("%*s parent sched-in ", comm_width, "comm"); 2815 printf(" run-time sleep iowait preempt delay\n"); 2816 } else { 2817 printf("\nRuntime summary\n"); 2818 printf("%*s parent sched-in ", comm_width, "comm"); 2819 printf(" run-time min-run avg-run max-run stddev migrations\n"); 2820 } 2821 printf("%*s (count) ", comm_width, ""); 2822 printf(" (msec) (msec) (msec) (msec) %s\n", 2823 sched->show_state ? "(msec)" : "%"); 2824 printf("%.117s\n", graph_dotted_line); 2825 2826 machine__for_each_thread(m, show_thread_runtime, &totals); 2827 task_count = totals.task_count; 2828 if (!task_count) 2829 printf("<no still running tasks>\n"); 2830 2831 printf("\nTerminated tasks:\n"); 2832 machine__for_each_thread(m, show_deadthread_runtime, &totals); 2833 if (task_count == totals.task_count) 2834 printf("<no terminated tasks>\n"); 2835 2836 /* CPU idle stats not tracked when samples were skipped */ 2837 if (sched->skipped_samples && !sched->idle_hist) 2838 return; 2839 2840 printf("\nIdle stats:\n"); 2841 for (i = 0; i < idle_max_cpu; ++i) { 2842 t = idle_threads[i]; 2843 if (!t) 2844 continue; 2845 2846 r = thread__priv(t); 2847 if (r && r->run_stats.n) { 2848 totals.sched_count += r->run_stats.n; 2849 printf(" CPU %2d idle for ", i); 2850 print_sched_time(r->total_run_time, 6); 2851 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time); 2852 } else 2853 printf(" CPU %2d idle entire time window\n", i); 2854 } 2855 2856 if (sched->idle_hist && sched->show_callchain) { 2857 callchain_param.mode = CHAIN_FOLDED; 2858 callchain_param.value = CCVAL_PERIOD; 2859 2860 callchain_register_param(&callchain_param); 2861 2862 printf("\nIdle stats by callchain:\n"); 2863 for (i = 0; i < idle_max_cpu; ++i) { 2864 struct idle_thread_runtime *itr; 2865 2866 t = idle_threads[i]; 2867 if (!t) 2868 continue; 2869 2870 itr = thread__priv(t); 2871 if (itr == NULL) 2872 continue; 2873 2874 callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain, 2875 0, &callchain_param); 2876 2877 printf(" CPU %2d:", i); 2878 print_sched_time(itr->tr.total_run_time, 6); 2879 printf(" msec\n"); 2880 timehist_print_idlehist_callchain(&itr->sorted_root); 2881 printf("\n"); 2882 } 2883 } 2884 2885 printf("\n" 2886 " Total number of unique tasks: %" PRIu64 "\n" 2887 "Total number of context switches: %" PRIu64 "\n", 2888 totals.task_count, totals.sched_count); 2889 2890 printf(" Total run time (msec): "); 2891 print_sched_time(totals.total_run_time, 2); 2892 printf("\n"); 2893 2894 printf(" Total scheduling time (msec): "); 2895 print_sched_time(hist_time, 2); 2896 printf(" (x %d)\n", sched->max_cpu); 2897 } 2898 2899 typedef int (*sched_handler)(struct perf_tool *tool, 2900 union perf_event *event, 2901 struct perf_evsel *evsel, 2902 struct perf_sample *sample, 2903 struct machine *machine); 2904 2905 static int perf_timehist__process_sample(struct perf_tool *tool, 2906 union perf_event *event, 2907 struct perf_sample *sample, 2908 struct perf_evsel *evsel, 2909 struct machine *machine) 2910 { 2911 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2912 int err = 0; 2913 int this_cpu = sample->cpu; 2914 2915 if (this_cpu > sched->max_cpu) 2916 sched->max_cpu = this_cpu; 2917 2918 if (evsel->handler != NULL) { 2919 sched_handler f = evsel->handler; 2920 2921 err = f(tool, event, evsel, sample, machine); 2922 } 2923 2924 return err; 2925 } 2926 2927 static int timehist_check_attr(struct perf_sched *sched, 2928 struct perf_evlist *evlist) 2929 { 2930 struct perf_evsel *evsel; 2931 struct evsel_runtime *er; 2932 2933 list_for_each_entry(evsel, &evlist->entries, node) { 2934 er = perf_evsel__get_runtime(evsel); 2935 if (er == NULL) { 2936 pr_err("Failed to allocate memory for evsel runtime data\n"); 2937 return -1; 2938 } 2939 2940 if (sched->show_callchain && !evsel__has_callchain(evsel)) { 2941 pr_info("Samples do not have callchains.\n"); 2942 sched->show_callchain = 0; 2943 symbol_conf.use_callchain = 0; 2944 } 2945 } 2946 2947 return 0; 2948 } 2949 2950 static int perf_sched__timehist(struct perf_sched *sched) 2951 { 2952 const struct perf_evsel_str_handler handlers[] = { 2953 { "sched:sched_switch", timehist_sched_switch_event, }, 2954 { "sched:sched_wakeup", timehist_sched_wakeup_event, }, 2955 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, }, 2956 }; 2957 const struct perf_evsel_str_handler migrate_handlers[] = { 2958 { "sched:sched_migrate_task", timehist_migrate_task_event, }, 2959 }; 2960 struct perf_data data = { 2961 .file = { 2962 .path = input_name, 2963 }, 2964 .mode = PERF_DATA_MODE_READ, 2965 .force = sched->force, 2966 }; 2967 2968 struct perf_session *session; 2969 struct perf_evlist *evlist; 2970 int err = -1; 2971 2972 /* 2973 * event handlers for timehist option 2974 */ 2975 sched->tool.sample = perf_timehist__process_sample; 2976 sched->tool.mmap = perf_event__process_mmap; 2977 sched->tool.comm = perf_event__process_comm; 2978 sched->tool.exit = perf_event__process_exit; 2979 sched->tool.fork = perf_event__process_fork; 2980 sched->tool.lost = process_lost; 2981 sched->tool.attr = perf_event__process_attr; 2982 sched->tool.tracing_data = perf_event__process_tracing_data; 2983 sched->tool.build_id = perf_event__process_build_id; 2984 2985 sched->tool.ordered_events = true; 2986 sched->tool.ordering_requires_timestamps = true; 2987 2988 symbol_conf.use_callchain = sched->show_callchain; 2989 2990 session = perf_session__new(&data, false, &sched->tool); 2991 if (session == NULL) 2992 return -ENOMEM; 2993 2994 evlist = session->evlist; 2995 2996 symbol__init(&session->header.env); 2997 2998 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) { 2999 pr_err("Invalid time string\n"); 3000 return -EINVAL; 3001 } 3002 3003 if (timehist_check_attr(sched, evlist) != 0) 3004 goto out; 3005 3006 setup_pager(); 3007 3008 /* setup per-evsel handlers */ 3009 if (perf_session__set_tracepoints_handlers(session, handlers)) 3010 goto out; 3011 3012 /* sched_switch event at a minimum needs to exist */ 3013 if (!perf_evlist__find_tracepoint_by_name(session->evlist, 3014 "sched:sched_switch")) { 3015 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n"); 3016 goto out; 3017 } 3018 3019 if (sched->show_migrations && 3020 perf_session__set_tracepoints_handlers(session, migrate_handlers)) 3021 goto out; 3022 3023 /* pre-allocate struct for per-CPU idle stats */ 3024 sched->max_cpu = session->header.env.nr_cpus_online; 3025 if (sched->max_cpu == 0) 3026 sched->max_cpu = 4; 3027 if (init_idle_threads(sched->max_cpu)) 3028 goto out; 3029 3030 /* summary_only implies summary option, but don't overwrite summary if set */ 3031 if (sched->summary_only) 3032 sched->summary = sched->summary_only; 3033 3034 if (!sched->summary_only) 3035 timehist_header(sched); 3036 3037 err = perf_session__process_events(session); 3038 if (err) { 3039 pr_err("Failed to process events, error %d", err); 3040 goto out; 3041 } 3042 3043 sched->nr_events = evlist->stats.nr_events[0]; 3044 sched->nr_lost_events = evlist->stats.total_lost; 3045 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST]; 3046 3047 if (sched->summary) 3048 timehist_print_summary(sched, session); 3049 3050 out: 3051 free_idle_threads(); 3052 perf_session__delete(session); 3053 3054 return err; 3055 } 3056 3057 3058 static void print_bad_events(struct perf_sched *sched) 3059 { 3060 if (sched->nr_unordered_timestamps && sched->nr_timestamps) { 3061 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n", 3062 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0, 3063 sched->nr_unordered_timestamps, sched->nr_timestamps); 3064 } 3065 if (sched->nr_lost_events && sched->nr_events) { 3066 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", 3067 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0, 3068 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks); 3069 } 3070 if (sched->nr_context_switch_bugs && sched->nr_timestamps) { 3071 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)", 3072 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0, 3073 sched->nr_context_switch_bugs, sched->nr_timestamps); 3074 if (sched->nr_lost_events) 3075 printf(" (due to lost events?)"); 3076 printf("\n"); 3077 } 3078 } 3079 3080 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data) 3081 { 3082 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL; 3083 struct work_atoms *this; 3084 const char *comm = thread__comm_str(data->thread), *this_comm; 3085 bool leftmost = true; 3086 3087 while (*new) { 3088 int cmp; 3089 3090 this = container_of(*new, struct work_atoms, node); 3091 parent = *new; 3092 3093 this_comm = thread__comm_str(this->thread); 3094 cmp = strcmp(comm, this_comm); 3095 if (cmp > 0) { 3096 new = &((*new)->rb_left); 3097 } else if (cmp < 0) { 3098 new = &((*new)->rb_right); 3099 leftmost = false; 3100 } else { 3101 this->num_merged++; 3102 this->total_runtime += data->total_runtime; 3103 this->nb_atoms += data->nb_atoms; 3104 this->total_lat += data->total_lat; 3105 list_splice(&data->work_list, &this->work_list); 3106 if (this->max_lat < data->max_lat) { 3107 this->max_lat = data->max_lat; 3108 this->max_lat_at = data->max_lat_at; 3109 } 3110 zfree(&data); 3111 return; 3112 } 3113 } 3114 3115 data->num_merged++; 3116 rb_link_node(&data->node, parent, new); 3117 rb_insert_color_cached(&data->node, root, leftmost); 3118 } 3119 3120 static void perf_sched__merge_lat(struct perf_sched *sched) 3121 { 3122 struct work_atoms *data; 3123 struct rb_node *node; 3124 3125 if (sched->skip_merge) 3126 return; 3127 3128 while ((node = rb_first_cached(&sched->atom_root))) { 3129 rb_erase_cached(node, &sched->atom_root); 3130 data = rb_entry(node, struct work_atoms, node); 3131 __merge_work_atoms(&sched->merged_atom_root, data); 3132 } 3133 } 3134 3135 static int perf_sched__lat(struct perf_sched *sched) 3136 { 3137 struct rb_node *next; 3138 3139 setup_pager(); 3140 3141 if (perf_sched__read_events(sched)) 3142 return -1; 3143 3144 perf_sched__merge_lat(sched); 3145 perf_sched__sort_lat(sched); 3146 3147 printf("\n -----------------------------------------------------------------------------------------------------------------\n"); 3148 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n"); 3149 printf(" -----------------------------------------------------------------------------------------------------------------\n"); 3150 3151 next = rb_first_cached(&sched->sorted_atom_root); 3152 3153 while (next) { 3154 struct work_atoms *work_list; 3155 3156 work_list = rb_entry(next, struct work_atoms, node); 3157 output_lat_thread(sched, work_list); 3158 next = rb_next(next); 3159 thread__zput(work_list->thread); 3160 } 3161 3162 printf(" -----------------------------------------------------------------------------------------------------------------\n"); 3163 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n", 3164 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count); 3165 3166 printf(" ---------------------------------------------------\n"); 3167 3168 print_bad_events(sched); 3169 printf("\n"); 3170 3171 return 0; 3172 } 3173 3174 static int setup_map_cpus(struct perf_sched *sched) 3175 { 3176 struct cpu_map *map; 3177 3178 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF); 3179 3180 if (sched->map.comp) { 3181 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int)); 3182 if (!sched->map.comp_cpus) 3183 return -1; 3184 } 3185 3186 if (!sched->map.cpus_str) 3187 return 0; 3188 3189 map = cpu_map__new(sched->map.cpus_str); 3190 if (!map) { 3191 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str); 3192 return -1; 3193 } 3194 3195 sched->map.cpus = map; 3196 return 0; 3197 } 3198 3199 static int setup_color_pids(struct perf_sched *sched) 3200 { 3201 struct thread_map *map; 3202 3203 if (!sched->map.color_pids_str) 3204 return 0; 3205 3206 map = thread_map__new_by_tid_str(sched->map.color_pids_str); 3207 if (!map) { 3208 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str); 3209 return -1; 3210 } 3211 3212 sched->map.color_pids = map; 3213 return 0; 3214 } 3215 3216 static int setup_color_cpus(struct perf_sched *sched) 3217 { 3218 struct cpu_map *map; 3219 3220 if (!sched->map.color_cpus_str) 3221 return 0; 3222 3223 map = cpu_map__new(sched->map.color_cpus_str); 3224 if (!map) { 3225 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str); 3226 return -1; 3227 } 3228 3229 sched->map.color_cpus = map; 3230 return 0; 3231 } 3232 3233 static int perf_sched__map(struct perf_sched *sched) 3234 { 3235 if (setup_map_cpus(sched)) 3236 return -1; 3237 3238 if (setup_color_pids(sched)) 3239 return -1; 3240 3241 if (setup_color_cpus(sched)) 3242 return -1; 3243 3244 setup_pager(); 3245 if (perf_sched__read_events(sched)) 3246 return -1; 3247 print_bad_events(sched); 3248 return 0; 3249 } 3250 3251 static int perf_sched__replay(struct perf_sched *sched) 3252 { 3253 unsigned long i; 3254 3255 calibrate_run_measurement_overhead(sched); 3256 calibrate_sleep_measurement_overhead(sched); 3257 3258 test_calibrations(sched); 3259 3260 if (perf_sched__read_events(sched)) 3261 return -1; 3262 3263 printf("nr_run_events: %ld\n", sched->nr_run_events); 3264 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events); 3265 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events); 3266 3267 if (sched->targetless_wakeups) 3268 printf("target-less wakeups: %ld\n", sched->targetless_wakeups); 3269 if (sched->multitarget_wakeups) 3270 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups); 3271 if (sched->nr_run_events_optimized) 3272 printf("run atoms optimized: %ld\n", 3273 sched->nr_run_events_optimized); 3274 3275 print_task_traces(sched); 3276 add_cross_task_wakeups(sched); 3277 3278 create_tasks(sched); 3279 printf("------------------------------------------------------------\n"); 3280 for (i = 0; i < sched->replay_repeat; i++) 3281 run_one_test(sched); 3282 3283 return 0; 3284 } 3285 3286 static void setup_sorting(struct perf_sched *sched, const struct option *options, 3287 const char * const usage_msg[]) 3288 { 3289 char *tmp, *tok, *str = strdup(sched->sort_order); 3290 3291 for (tok = strtok_r(str, ", ", &tmp); 3292 tok; tok = strtok_r(NULL, ", ", &tmp)) { 3293 if (sort_dimension__add(tok, &sched->sort_list) < 0) { 3294 usage_with_options_msg(usage_msg, options, 3295 "Unknown --sort key: `%s'", tok); 3296 } 3297 } 3298 3299 free(str); 3300 3301 sort_dimension__add("pid", &sched->cmp_pid); 3302 } 3303 3304 static int __cmd_record(int argc, const char **argv) 3305 { 3306 unsigned int rec_argc, i, j; 3307 const char **rec_argv; 3308 const char * const record_args[] = { 3309 "record", 3310 "-a", 3311 "-R", 3312 "-m", "1024", 3313 "-c", "1", 3314 "-e", "sched:sched_switch", 3315 "-e", "sched:sched_stat_wait", 3316 "-e", "sched:sched_stat_sleep", 3317 "-e", "sched:sched_stat_iowait", 3318 "-e", "sched:sched_stat_runtime", 3319 "-e", "sched:sched_process_fork", 3320 "-e", "sched:sched_wakeup", 3321 "-e", "sched:sched_wakeup_new", 3322 "-e", "sched:sched_migrate_task", 3323 }; 3324 3325 rec_argc = ARRAY_SIZE(record_args) + argc - 1; 3326 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 3327 3328 if (rec_argv == NULL) 3329 return -ENOMEM; 3330 3331 for (i = 0; i < ARRAY_SIZE(record_args); i++) 3332 rec_argv[i] = strdup(record_args[i]); 3333 3334 for (j = 1; j < (unsigned int)argc; j++, i++) 3335 rec_argv[i] = argv[j]; 3336 3337 BUG_ON(i != rec_argc); 3338 3339 return cmd_record(i, rec_argv); 3340 } 3341 3342 int cmd_sched(int argc, const char **argv) 3343 { 3344 static const char default_sort_order[] = "avg, max, switch, runtime"; 3345 struct perf_sched sched = { 3346 .tool = { 3347 .sample = perf_sched__process_tracepoint_sample, 3348 .comm = perf_sched__process_comm, 3349 .namespaces = perf_event__process_namespaces, 3350 .lost = perf_event__process_lost, 3351 .fork = perf_sched__process_fork_event, 3352 .ordered_events = true, 3353 }, 3354 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid), 3355 .sort_list = LIST_HEAD_INIT(sched.sort_list), 3356 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER, 3357 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER, 3358 .sort_order = default_sort_order, 3359 .replay_repeat = 10, 3360 .profile_cpu = -1, 3361 .next_shortname1 = 'A', 3362 .next_shortname2 = '0', 3363 .skip_merge = 0, 3364 .show_callchain = 1, 3365 .max_stack = 5, 3366 }; 3367 const struct option sched_options[] = { 3368 OPT_STRING('i', "input", &input_name, "file", 3369 "input file name"), 3370 OPT_INCR('v', "verbose", &verbose, 3371 "be more verbose (show symbol address, etc)"), 3372 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 3373 "dump raw trace in ASCII"), 3374 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"), 3375 OPT_END() 3376 }; 3377 const struct option latency_options[] = { 3378 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]", 3379 "sort by key(s): runtime, switch, avg, max"), 3380 OPT_INTEGER('C', "CPU", &sched.profile_cpu, 3381 "CPU to profile on"), 3382 OPT_BOOLEAN('p', "pids", &sched.skip_merge, 3383 "latency stats per pid instead of per comm"), 3384 OPT_PARENT(sched_options) 3385 }; 3386 const struct option replay_options[] = { 3387 OPT_UINTEGER('r', "repeat", &sched.replay_repeat, 3388 "repeat the workload replay N times (-1: infinite)"), 3389 OPT_PARENT(sched_options) 3390 }; 3391 const struct option map_options[] = { 3392 OPT_BOOLEAN(0, "compact", &sched.map.comp, 3393 "map output in compact mode"), 3394 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids", 3395 "highlight given pids in map"), 3396 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus", 3397 "highlight given CPUs in map"), 3398 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus", 3399 "display given CPUs in map"), 3400 OPT_PARENT(sched_options) 3401 }; 3402 const struct option timehist_options[] = { 3403 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 3404 "file", "vmlinux pathname"), 3405 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 3406 "file", "kallsyms pathname"), 3407 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain, 3408 "Display call chains if present (default on)"), 3409 OPT_UINTEGER(0, "max-stack", &sched.max_stack, 3410 "Maximum number of functions to display backtrace."), 3411 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory", 3412 "Look for files with symbols relative to this directory"), 3413 OPT_BOOLEAN('s', "summary", &sched.summary_only, 3414 "Show only syscall summary with statistics"), 3415 OPT_BOOLEAN('S', "with-summary", &sched.summary, 3416 "Show all syscalls and summary with statistics"), 3417 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"), 3418 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"), 3419 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"), 3420 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"), 3421 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"), 3422 OPT_STRING(0, "time", &sched.time_str, "str", 3423 "Time span for analysis (start,stop)"), 3424 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"), 3425 OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]", 3426 "analyze events only for given process id(s)"), 3427 OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]", 3428 "analyze events only for given thread id(s)"), 3429 OPT_PARENT(sched_options) 3430 }; 3431 3432 const char * const latency_usage[] = { 3433 "perf sched latency [<options>]", 3434 NULL 3435 }; 3436 const char * const replay_usage[] = { 3437 "perf sched replay [<options>]", 3438 NULL 3439 }; 3440 const char * const map_usage[] = { 3441 "perf sched map [<options>]", 3442 NULL 3443 }; 3444 const char * const timehist_usage[] = { 3445 "perf sched timehist [<options>]", 3446 NULL 3447 }; 3448 const char *const sched_subcommands[] = { "record", "latency", "map", 3449 "replay", "script", 3450 "timehist", NULL }; 3451 const char *sched_usage[] = { 3452 NULL, 3453 NULL 3454 }; 3455 struct trace_sched_handler lat_ops = { 3456 .wakeup_event = latency_wakeup_event, 3457 .switch_event = latency_switch_event, 3458 .runtime_event = latency_runtime_event, 3459 .migrate_task_event = latency_migrate_task_event, 3460 }; 3461 struct trace_sched_handler map_ops = { 3462 .switch_event = map_switch_event, 3463 }; 3464 struct trace_sched_handler replay_ops = { 3465 .wakeup_event = replay_wakeup_event, 3466 .switch_event = replay_switch_event, 3467 .fork_event = replay_fork_event, 3468 }; 3469 unsigned int i; 3470 3471 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++) 3472 sched.curr_pid[i] = -1; 3473 3474 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands, 3475 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION); 3476 if (!argc) 3477 usage_with_options(sched_usage, sched_options); 3478 3479 /* 3480 * Aliased to 'perf script' for now: 3481 */ 3482 if (!strcmp(argv[0], "script")) 3483 return cmd_script(argc, argv); 3484 3485 if (!strncmp(argv[0], "rec", 3)) { 3486 return __cmd_record(argc, argv); 3487 } else if (!strncmp(argv[0], "lat", 3)) { 3488 sched.tp_handler = &lat_ops; 3489 if (argc > 1) { 3490 argc = parse_options(argc, argv, latency_options, latency_usage, 0); 3491 if (argc) 3492 usage_with_options(latency_usage, latency_options); 3493 } 3494 setup_sorting(&sched, latency_options, latency_usage); 3495 return perf_sched__lat(&sched); 3496 } else if (!strcmp(argv[0], "map")) { 3497 if (argc) { 3498 argc = parse_options(argc, argv, map_options, map_usage, 0); 3499 if (argc) 3500 usage_with_options(map_usage, map_options); 3501 } 3502 sched.tp_handler = &map_ops; 3503 setup_sorting(&sched, latency_options, latency_usage); 3504 return perf_sched__map(&sched); 3505 } else if (!strncmp(argv[0], "rep", 3)) { 3506 sched.tp_handler = &replay_ops; 3507 if (argc) { 3508 argc = parse_options(argc, argv, replay_options, replay_usage, 0); 3509 if (argc) 3510 usage_with_options(replay_usage, replay_options); 3511 } 3512 return perf_sched__replay(&sched); 3513 } else if (!strcmp(argv[0], "timehist")) { 3514 if (argc) { 3515 argc = parse_options(argc, argv, timehist_options, 3516 timehist_usage, 0); 3517 if (argc) 3518 usage_with_options(timehist_usage, timehist_options); 3519 } 3520 if ((sched.show_wakeups || sched.show_next) && 3521 sched.summary_only) { 3522 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n"); 3523 parse_options_usage(timehist_usage, timehist_options, "s", true); 3524 if (sched.show_wakeups) 3525 parse_options_usage(NULL, timehist_options, "w", true); 3526 if (sched.show_next) 3527 parse_options_usage(NULL, timehist_options, "n", true); 3528 return -EINVAL; 3529 } 3530 3531 return perf_sched__timehist(&sched); 3532 } else { 3533 usage_with_options(sched_usage, sched_options); 3534 } 3535 3536 return 0; 3537 } 3538