1 #include "builtin.h" 2 #include "perf.h" 3 4 #include "util/util.h" 5 #include "util/evlist.h" 6 #include "util/cache.h" 7 #include "util/evsel.h" 8 #include "util/symbol.h" 9 #include "util/thread.h" 10 #include "util/header.h" 11 #include "util/session.h" 12 #include "util/tool.h" 13 #include "util/cloexec.h" 14 #include "util/thread_map.h" 15 #include "util/color.h" 16 #include "util/stat.h" 17 #include "util/callchain.h" 18 #include "util/time-utils.h" 19 20 #include <subcmd/parse-options.h> 21 #include "util/trace-event.h" 22 23 #include "util/debug.h" 24 25 #include <linux/kernel.h> 26 #include <linux/log2.h> 27 #include <sys/prctl.h> 28 #include <sys/resource.h> 29 #include <inttypes.h> 30 31 #include <errno.h> 32 #include <semaphore.h> 33 #include <pthread.h> 34 #include <math.h> 35 #include <api/fs/fs.h> 36 #include <linux/time64.h> 37 38 #include "sane_ctype.h" 39 40 #define PR_SET_NAME 15 /* Set process name */ 41 #define MAX_CPUS 4096 42 #define COMM_LEN 20 43 #define SYM_LEN 129 44 #define MAX_PID 1024000 45 46 struct sched_atom; 47 48 struct task_desc { 49 unsigned long nr; 50 unsigned long pid; 51 char comm[COMM_LEN]; 52 53 unsigned long nr_events; 54 unsigned long curr_event; 55 struct sched_atom **atoms; 56 57 pthread_t thread; 58 sem_t sleep_sem; 59 60 sem_t ready_for_work; 61 sem_t work_done_sem; 62 63 u64 cpu_usage; 64 }; 65 66 enum sched_event_type { 67 SCHED_EVENT_RUN, 68 SCHED_EVENT_SLEEP, 69 SCHED_EVENT_WAKEUP, 70 SCHED_EVENT_MIGRATION, 71 }; 72 73 struct sched_atom { 74 enum sched_event_type type; 75 int specific_wait; 76 u64 timestamp; 77 u64 duration; 78 unsigned long nr; 79 sem_t *wait_sem; 80 struct task_desc *wakee; 81 }; 82 83 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP" 84 85 /* task state bitmask, copied from include/linux/sched.h */ 86 #define TASK_RUNNING 0 87 #define TASK_INTERRUPTIBLE 1 88 #define TASK_UNINTERRUPTIBLE 2 89 #define __TASK_STOPPED 4 90 #define __TASK_TRACED 8 91 /* in tsk->exit_state */ 92 #define EXIT_DEAD 16 93 #define EXIT_ZOMBIE 32 94 #define EXIT_TRACE (EXIT_ZOMBIE | EXIT_DEAD) 95 /* in tsk->state again */ 96 #define TASK_DEAD 64 97 #define TASK_WAKEKILL 128 98 #define TASK_WAKING 256 99 #define TASK_PARKED 512 100 101 enum thread_state { 102 THREAD_SLEEPING = 0, 103 THREAD_WAIT_CPU, 104 THREAD_SCHED_IN, 105 THREAD_IGNORE 106 }; 107 108 struct work_atom { 109 struct list_head list; 110 enum thread_state state; 111 u64 sched_out_time; 112 u64 wake_up_time; 113 u64 sched_in_time; 114 u64 runtime; 115 }; 116 117 struct work_atoms { 118 struct list_head work_list; 119 struct thread *thread; 120 struct rb_node node; 121 u64 max_lat; 122 u64 max_lat_at; 123 u64 total_lat; 124 u64 nb_atoms; 125 u64 total_runtime; 126 int num_merged; 127 }; 128 129 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *); 130 131 struct perf_sched; 132 133 struct trace_sched_handler { 134 int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel, 135 struct perf_sample *sample, struct machine *machine); 136 137 int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel, 138 struct perf_sample *sample, struct machine *machine); 139 140 int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel, 141 struct perf_sample *sample, struct machine *machine); 142 143 /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */ 144 int (*fork_event)(struct perf_sched *sched, union perf_event *event, 145 struct machine *machine); 146 147 int (*migrate_task_event)(struct perf_sched *sched, 148 struct perf_evsel *evsel, 149 struct perf_sample *sample, 150 struct machine *machine); 151 }; 152 153 #define COLOR_PIDS PERF_COLOR_BLUE 154 #define COLOR_CPUS PERF_COLOR_BG_RED 155 156 struct perf_sched_map { 157 DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS); 158 int *comp_cpus; 159 bool comp; 160 struct thread_map *color_pids; 161 const char *color_pids_str; 162 struct cpu_map *color_cpus; 163 const char *color_cpus_str; 164 struct cpu_map *cpus; 165 const char *cpus_str; 166 }; 167 168 struct perf_sched { 169 struct perf_tool tool; 170 const char *sort_order; 171 unsigned long nr_tasks; 172 struct task_desc **pid_to_task; 173 struct task_desc **tasks; 174 const struct trace_sched_handler *tp_handler; 175 pthread_mutex_t start_work_mutex; 176 pthread_mutex_t work_done_wait_mutex; 177 int profile_cpu; 178 /* 179 * Track the current task - that way we can know whether there's any 180 * weird events, such as a task being switched away that is not current. 181 */ 182 int max_cpu; 183 u32 curr_pid[MAX_CPUS]; 184 struct thread *curr_thread[MAX_CPUS]; 185 char next_shortname1; 186 char next_shortname2; 187 unsigned int replay_repeat; 188 unsigned long nr_run_events; 189 unsigned long nr_sleep_events; 190 unsigned long nr_wakeup_events; 191 unsigned long nr_sleep_corrections; 192 unsigned long nr_run_events_optimized; 193 unsigned long targetless_wakeups; 194 unsigned long multitarget_wakeups; 195 unsigned long nr_runs; 196 unsigned long nr_timestamps; 197 unsigned long nr_unordered_timestamps; 198 unsigned long nr_context_switch_bugs; 199 unsigned long nr_events; 200 unsigned long nr_lost_chunks; 201 unsigned long nr_lost_events; 202 u64 run_measurement_overhead; 203 u64 sleep_measurement_overhead; 204 u64 start_time; 205 u64 cpu_usage; 206 u64 runavg_cpu_usage; 207 u64 parent_cpu_usage; 208 u64 runavg_parent_cpu_usage; 209 u64 sum_runtime; 210 u64 sum_fluct; 211 u64 run_avg; 212 u64 all_runtime; 213 u64 all_count; 214 u64 cpu_last_switched[MAX_CPUS]; 215 struct rb_root atom_root, sorted_atom_root, merged_atom_root; 216 struct list_head sort_list, cmp_pid; 217 bool force; 218 bool skip_merge; 219 struct perf_sched_map map; 220 221 /* options for timehist command */ 222 bool summary; 223 bool summary_only; 224 bool idle_hist; 225 bool show_callchain; 226 unsigned int max_stack; 227 bool show_cpu_visual; 228 bool show_wakeups; 229 bool show_next; 230 bool show_migrations; 231 bool show_state; 232 u64 skipped_samples; 233 const char *time_str; 234 struct perf_time_interval ptime; 235 struct perf_time_interval hist_time; 236 }; 237 238 /* per thread run time data */ 239 struct thread_runtime { 240 u64 last_time; /* time of previous sched in/out event */ 241 u64 dt_run; /* run time */ 242 u64 dt_sleep; /* time between CPU access by sleep (off cpu) */ 243 u64 dt_iowait; /* time between CPU access by iowait (off cpu) */ 244 u64 dt_preempt; /* time between CPU access by preempt (off cpu) */ 245 u64 dt_delay; /* time between wakeup and sched-in */ 246 u64 ready_to_run; /* time of wakeup */ 247 248 struct stats run_stats; 249 u64 total_run_time; 250 u64 total_sleep_time; 251 u64 total_iowait_time; 252 u64 total_preempt_time; 253 u64 total_delay_time; 254 255 int last_state; 256 u64 migrations; 257 }; 258 259 /* per event run time data */ 260 struct evsel_runtime { 261 u64 *last_time; /* time this event was last seen per cpu */ 262 u32 ncpu; /* highest cpu slot allocated */ 263 }; 264 265 /* per cpu idle time data */ 266 struct idle_thread_runtime { 267 struct thread_runtime tr; 268 struct thread *last_thread; 269 struct rb_root sorted_root; 270 struct callchain_root callchain; 271 struct callchain_cursor cursor; 272 }; 273 274 /* track idle times per cpu */ 275 static struct thread **idle_threads; 276 static int idle_max_cpu; 277 static char idle_comm[] = "<idle>"; 278 279 static u64 get_nsecs(void) 280 { 281 struct timespec ts; 282 283 clock_gettime(CLOCK_MONOTONIC, &ts); 284 285 return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec; 286 } 287 288 static void burn_nsecs(struct perf_sched *sched, u64 nsecs) 289 { 290 u64 T0 = get_nsecs(), T1; 291 292 do { 293 T1 = get_nsecs(); 294 } while (T1 + sched->run_measurement_overhead < T0 + nsecs); 295 } 296 297 static void sleep_nsecs(u64 nsecs) 298 { 299 struct timespec ts; 300 301 ts.tv_nsec = nsecs % 999999999; 302 ts.tv_sec = nsecs / 999999999; 303 304 nanosleep(&ts, NULL); 305 } 306 307 static void calibrate_run_measurement_overhead(struct perf_sched *sched) 308 { 309 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 310 int i; 311 312 for (i = 0; i < 10; i++) { 313 T0 = get_nsecs(); 314 burn_nsecs(sched, 0); 315 T1 = get_nsecs(); 316 delta = T1-T0; 317 min_delta = min(min_delta, delta); 318 } 319 sched->run_measurement_overhead = min_delta; 320 321 printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta); 322 } 323 324 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched) 325 { 326 u64 T0, T1, delta, min_delta = NSEC_PER_SEC; 327 int i; 328 329 for (i = 0; i < 10; i++) { 330 T0 = get_nsecs(); 331 sleep_nsecs(10000); 332 T1 = get_nsecs(); 333 delta = T1-T0; 334 min_delta = min(min_delta, delta); 335 } 336 min_delta -= 10000; 337 sched->sleep_measurement_overhead = min_delta; 338 339 printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta); 340 } 341 342 static struct sched_atom * 343 get_new_event(struct task_desc *task, u64 timestamp) 344 { 345 struct sched_atom *event = zalloc(sizeof(*event)); 346 unsigned long idx = task->nr_events; 347 size_t size; 348 349 event->timestamp = timestamp; 350 event->nr = idx; 351 352 task->nr_events++; 353 size = sizeof(struct sched_atom *) * task->nr_events; 354 task->atoms = realloc(task->atoms, size); 355 BUG_ON(!task->atoms); 356 357 task->atoms[idx] = event; 358 359 return event; 360 } 361 362 static struct sched_atom *last_event(struct task_desc *task) 363 { 364 if (!task->nr_events) 365 return NULL; 366 367 return task->atoms[task->nr_events - 1]; 368 } 369 370 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task, 371 u64 timestamp, u64 duration) 372 { 373 struct sched_atom *event, *curr_event = last_event(task); 374 375 /* 376 * optimize an existing RUN event by merging this one 377 * to it: 378 */ 379 if (curr_event && curr_event->type == SCHED_EVENT_RUN) { 380 sched->nr_run_events_optimized++; 381 curr_event->duration += duration; 382 return; 383 } 384 385 event = get_new_event(task, timestamp); 386 387 event->type = SCHED_EVENT_RUN; 388 event->duration = duration; 389 390 sched->nr_run_events++; 391 } 392 393 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task, 394 u64 timestamp, struct task_desc *wakee) 395 { 396 struct sched_atom *event, *wakee_event; 397 398 event = get_new_event(task, timestamp); 399 event->type = SCHED_EVENT_WAKEUP; 400 event->wakee = wakee; 401 402 wakee_event = last_event(wakee); 403 if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) { 404 sched->targetless_wakeups++; 405 return; 406 } 407 if (wakee_event->wait_sem) { 408 sched->multitarget_wakeups++; 409 return; 410 } 411 412 wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem)); 413 sem_init(wakee_event->wait_sem, 0, 0); 414 wakee_event->specific_wait = 1; 415 event->wait_sem = wakee_event->wait_sem; 416 417 sched->nr_wakeup_events++; 418 } 419 420 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task, 421 u64 timestamp, u64 task_state __maybe_unused) 422 { 423 struct sched_atom *event = get_new_event(task, timestamp); 424 425 event->type = SCHED_EVENT_SLEEP; 426 427 sched->nr_sleep_events++; 428 } 429 430 static struct task_desc *register_pid(struct perf_sched *sched, 431 unsigned long pid, const char *comm) 432 { 433 struct task_desc *task; 434 static int pid_max; 435 436 if (sched->pid_to_task == NULL) { 437 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0) 438 pid_max = MAX_PID; 439 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL); 440 } 441 if (pid >= (unsigned long)pid_max) { 442 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) * 443 sizeof(struct task_desc *))) == NULL); 444 while (pid >= (unsigned long)pid_max) 445 sched->pid_to_task[pid_max++] = NULL; 446 } 447 448 task = sched->pid_to_task[pid]; 449 450 if (task) 451 return task; 452 453 task = zalloc(sizeof(*task)); 454 task->pid = pid; 455 task->nr = sched->nr_tasks; 456 strcpy(task->comm, comm); 457 /* 458 * every task starts in sleeping state - this gets ignored 459 * if there's no wakeup pointing to this sleep state: 460 */ 461 add_sched_event_sleep(sched, task, 0, 0); 462 463 sched->pid_to_task[pid] = task; 464 sched->nr_tasks++; 465 sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *)); 466 BUG_ON(!sched->tasks); 467 sched->tasks[task->nr] = task; 468 469 if (verbose > 0) 470 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm); 471 472 return task; 473 } 474 475 476 static void print_task_traces(struct perf_sched *sched) 477 { 478 struct task_desc *task; 479 unsigned long i; 480 481 for (i = 0; i < sched->nr_tasks; i++) { 482 task = sched->tasks[i]; 483 printf("task %6ld (%20s:%10ld), nr_events: %ld\n", 484 task->nr, task->comm, task->pid, task->nr_events); 485 } 486 } 487 488 static void add_cross_task_wakeups(struct perf_sched *sched) 489 { 490 struct task_desc *task1, *task2; 491 unsigned long i, j; 492 493 for (i = 0; i < sched->nr_tasks; i++) { 494 task1 = sched->tasks[i]; 495 j = i + 1; 496 if (j == sched->nr_tasks) 497 j = 0; 498 task2 = sched->tasks[j]; 499 add_sched_event_wakeup(sched, task1, 0, task2); 500 } 501 } 502 503 static void perf_sched__process_event(struct perf_sched *sched, 504 struct sched_atom *atom) 505 { 506 int ret = 0; 507 508 switch (atom->type) { 509 case SCHED_EVENT_RUN: 510 burn_nsecs(sched, atom->duration); 511 break; 512 case SCHED_EVENT_SLEEP: 513 if (atom->wait_sem) 514 ret = sem_wait(atom->wait_sem); 515 BUG_ON(ret); 516 break; 517 case SCHED_EVENT_WAKEUP: 518 if (atom->wait_sem) 519 ret = sem_post(atom->wait_sem); 520 BUG_ON(ret); 521 break; 522 case SCHED_EVENT_MIGRATION: 523 break; 524 default: 525 BUG_ON(1); 526 } 527 } 528 529 static u64 get_cpu_usage_nsec_parent(void) 530 { 531 struct rusage ru; 532 u64 sum; 533 int err; 534 535 err = getrusage(RUSAGE_SELF, &ru); 536 BUG_ON(err); 537 538 sum = ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC; 539 sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC; 540 541 return sum; 542 } 543 544 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task) 545 { 546 struct perf_event_attr attr; 547 char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE]; 548 int fd; 549 struct rlimit limit; 550 bool need_privilege = false; 551 552 memset(&attr, 0, sizeof(attr)); 553 554 attr.type = PERF_TYPE_SOFTWARE; 555 attr.config = PERF_COUNT_SW_TASK_CLOCK; 556 557 force_again: 558 fd = sys_perf_event_open(&attr, 0, -1, -1, 559 perf_event_open_cloexec_flag()); 560 561 if (fd < 0) { 562 if (errno == EMFILE) { 563 if (sched->force) { 564 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1); 565 limit.rlim_cur += sched->nr_tasks - cur_task; 566 if (limit.rlim_cur > limit.rlim_max) { 567 limit.rlim_max = limit.rlim_cur; 568 need_privilege = true; 569 } 570 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) { 571 if (need_privilege && errno == EPERM) 572 strcpy(info, "Need privilege\n"); 573 } else 574 goto force_again; 575 } else 576 strcpy(info, "Have a try with -f option\n"); 577 } 578 pr_err("Error: sys_perf_event_open() syscall returned " 579 "with %d (%s)\n%s", fd, 580 str_error_r(errno, sbuf, sizeof(sbuf)), info); 581 exit(EXIT_FAILURE); 582 } 583 return fd; 584 } 585 586 static u64 get_cpu_usage_nsec_self(int fd) 587 { 588 u64 runtime; 589 int ret; 590 591 ret = read(fd, &runtime, sizeof(runtime)); 592 BUG_ON(ret != sizeof(runtime)); 593 594 return runtime; 595 } 596 597 struct sched_thread_parms { 598 struct task_desc *task; 599 struct perf_sched *sched; 600 int fd; 601 }; 602 603 static void *thread_func(void *ctx) 604 { 605 struct sched_thread_parms *parms = ctx; 606 struct task_desc *this_task = parms->task; 607 struct perf_sched *sched = parms->sched; 608 u64 cpu_usage_0, cpu_usage_1; 609 unsigned long i, ret; 610 char comm2[22]; 611 int fd = parms->fd; 612 613 zfree(&parms); 614 615 sprintf(comm2, ":%s", this_task->comm); 616 prctl(PR_SET_NAME, comm2); 617 if (fd < 0) 618 return NULL; 619 again: 620 ret = sem_post(&this_task->ready_for_work); 621 BUG_ON(ret); 622 ret = pthread_mutex_lock(&sched->start_work_mutex); 623 BUG_ON(ret); 624 ret = pthread_mutex_unlock(&sched->start_work_mutex); 625 BUG_ON(ret); 626 627 cpu_usage_0 = get_cpu_usage_nsec_self(fd); 628 629 for (i = 0; i < this_task->nr_events; i++) { 630 this_task->curr_event = i; 631 perf_sched__process_event(sched, this_task->atoms[i]); 632 } 633 634 cpu_usage_1 = get_cpu_usage_nsec_self(fd); 635 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0; 636 ret = sem_post(&this_task->work_done_sem); 637 BUG_ON(ret); 638 639 ret = pthread_mutex_lock(&sched->work_done_wait_mutex); 640 BUG_ON(ret); 641 ret = pthread_mutex_unlock(&sched->work_done_wait_mutex); 642 BUG_ON(ret); 643 644 goto again; 645 } 646 647 static void create_tasks(struct perf_sched *sched) 648 { 649 struct task_desc *task; 650 pthread_attr_t attr; 651 unsigned long i; 652 int err; 653 654 err = pthread_attr_init(&attr); 655 BUG_ON(err); 656 err = pthread_attr_setstacksize(&attr, 657 (size_t) max(16 * 1024, PTHREAD_STACK_MIN)); 658 BUG_ON(err); 659 err = pthread_mutex_lock(&sched->start_work_mutex); 660 BUG_ON(err); 661 err = pthread_mutex_lock(&sched->work_done_wait_mutex); 662 BUG_ON(err); 663 for (i = 0; i < sched->nr_tasks; i++) { 664 struct sched_thread_parms *parms = malloc(sizeof(*parms)); 665 BUG_ON(parms == NULL); 666 parms->task = task = sched->tasks[i]; 667 parms->sched = sched; 668 parms->fd = self_open_counters(sched, i); 669 sem_init(&task->sleep_sem, 0, 0); 670 sem_init(&task->ready_for_work, 0, 0); 671 sem_init(&task->work_done_sem, 0, 0); 672 task->curr_event = 0; 673 err = pthread_create(&task->thread, &attr, thread_func, parms); 674 BUG_ON(err); 675 } 676 } 677 678 static void wait_for_tasks(struct perf_sched *sched) 679 { 680 u64 cpu_usage_0, cpu_usage_1; 681 struct task_desc *task; 682 unsigned long i, ret; 683 684 sched->start_time = get_nsecs(); 685 sched->cpu_usage = 0; 686 pthread_mutex_unlock(&sched->work_done_wait_mutex); 687 688 for (i = 0; i < sched->nr_tasks; i++) { 689 task = sched->tasks[i]; 690 ret = sem_wait(&task->ready_for_work); 691 BUG_ON(ret); 692 sem_init(&task->ready_for_work, 0, 0); 693 } 694 ret = pthread_mutex_lock(&sched->work_done_wait_mutex); 695 BUG_ON(ret); 696 697 cpu_usage_0 = get_cpu_usage_nsec_parent(); 698 699 pthread_mutex_unlock(&sched->start_work_mutex); 700 701 for (i = 0; i < sched->nr_tasks; i++) { 702 task = sched->tasks[i]; 703 ret = sem_wait(&task->work_done_sem); 704 BUG_ON(ret); 705 sem_init(&task->work_done_sem, 0, 0); 706 sched->cpu_usage += task->cpu_usage; 707 task->cpu_usage = 0; 708 } 709 710 cpu_usage_1 = get_cpu_usage_nsec_parent(); 711 if (!sched->runavg_cpu_usage) 712 sched->runavg_cpu_usage = sched->cpu_usage; 713 sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat; 714 715 sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0; 716 if (!sched->runavg_parent_cpu_usage) 717 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage; 718 sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) + 719 sched->parent_cpu_usage)/sched->replay_repeat; 720 721 ret = pthread_mutex_lock(&sched->start_work_mutex); 722 BUG_ON(ret); 723 724 for (i = 0; i < sched->nr_tasks; i++) { 725 task = sched->tasks[i]; 726 sem_init(&task->sleep_sem, 0, 0); 727 task->curr_event = 0; 728 } 729 } 730 731 static void run_one_test(struct perf_sched *sched) 732 { 733 u64 T0, T1, delta, avg_delta, fluct; 734 735 T0 = get_nsecs(); 736 wait_for_tasks(sched); 737 T1 = get_nsecs(); 738 739 delta = T1 - T0; 740 sched->sum_runtime += delta; 741 sched->nr_runs++; 742 743 avg_delta = sched->sum_runtime / sched->nr_runs; 744 if (delta < avg_delta) 745 fluct = avg_delta - delta; 746 else 747 fluct = delta - avg_delta; 748 sched->sum_fluct += fluct; 749 if (!sched->run_avg) 750 sched->run_avg = delta; 751 sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat; 752 753 printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC); 754 755 printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC); 756 757 printf("cpu: %0.2f / %0.2f", 758 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC); 759 760 #if 0 761 /* 762 * rusage statistics done by the parent, these are less 763 * accurate than the sched->sum_exec_runtime based statistics: 764 */ 765 printf(" [%0.2f / %0.2f]", 766 (double)sched->parent_cpu_usage / NSEC_PER_MSEC, 767 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC); 768 #endif 769 770 printf("\n"); 771 772 if (sched->nr_sleep_corrections) 773 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections); 774 sched->nr_sleep_corrections = 0; 775 } 776 777 static void test_calibrations(struct perf_sched *sched) 778 { 779 u64 T0, T1; 780 781 T0 = get_nsecs(); 782 burn_nsecs(sched, NSEC_PER_MSEC); 783 T1 = get_nsecs(); 784 785 printf("the run test took %" PRIu64 " nsecs\n", T1 - T0); 786 787 T0 = get_nsecs(); 788 sleep_nsecs(NSEC_PER_MSEC); 789 T1 = get_nsecs(); 790 791 printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0); 792 } 793 794 static int 795 replay_wakeup_event(struct perf_sched *sched, 796 struct perf_evsel *evsel, struct perf_sample *sample, 797 struct machine *machine __maybe_unused) 798 { 799 const char *comm = perf_evsel__strval(evsel, sample, "comm"); 800 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 801 struct task_desc *waker, *wakee; 802 803 if (verbose > 0) { 804 printf("sched_wakeup event %p\n", evsel); 805 806 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid); 807 } 808 809 waker = register_pid(sched, sample->tid, "<unknown>"); 810 wakee = register_pid(sched, pid, comm); 811 812 add_sched_event_wakeup(sched, waker, sample->time, wakee); 813 return 0; 814 } 815 816 static int replay_switch_event(struct perf_sched *sched, 817 struct perf_evsel *evsel, 818 struct perf_sample *sample, 819 struct machine *machine __maybe_unused) 820 { 821 const char *prev_comm = perf_evsel__strval(evsel, sample, "prev_comm"), 822 *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); 823 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 824 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 825 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); 826 struct task_desc *prev, __maybe_unused *next; 827 u64 timestamp0, timestamp = sample->time; 828 int cpu = sample->cpu; 829 s64 delta; 830 831 if (verbose > 0) 832 printf("sched_switch event %p\n", evsel); 833 834 if (cpu >= MAX_CPUS || cpu < 0) 835 return 0; 836 837 timestamp0 = sched->cpu_last_switched[cpu]; 838 if (timestamp0) 839 delta = timestamp - timestamp0; 840 else 841 delta = 0; 842 843 if (delta < 0) { 844 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 845 return -1; 846 } 847 848 pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n", 849 prev_comm, prev_pid, next_comm, next_pid, delta); 850 851 prev = register_pid(sched, prev_pid, prev_comm); 852 next = register_pid(sched, next_pid, next_comm); 853 854 sched->cpu_last_switched[cpu] = timestamp; 855 856 add_sched_event_run(sched, prev, timestamp, delta); 857 add_sched_event_sleep(sched, prev, timestamp, prev_state); 858 859 return 0; 860 } 861 862 static int replay_fork_event(struct perf_sched *sched, 863 union perf_event *event, 864 struct machine *machine) 865 { 866 struct thread *child, *parent; 867 868 child = machine__findnew_thread(machine, event->fork.pid, 869 event->fork.tid); 870 parent = machine__findnew_thread(machine, event->fork.ppid, 871 event->fork.ptid); 872 873 if (child == NULL || parent == NULL) { 874 pr_debug("thread does not exist on fork event: child %p, parent %p\n", 875 child, parent); 876 goto out_put; 877 } 878 879 if (verbose > 0) { 880 printf("fork event\n"); 881 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid); 882 printf("... child: %s/%d\n", thread__comm_str(child), child->tid); 883 } 884 885 register_pid(sched, parent->tid, thread__comm_str(parent)); 886 register_pid(sched, child->tid, thread__comm_str(child)); 887 out_put: 888 thread__put(child); 889 thread__put(parent); 890 return 0; 891 } 892 893 struct sort_dimension { 894 const char *name; 895 sort_fn_t cmp; 896 struct list_head list; 897 }; 898 899 static int 900 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r) 901 { 902 struct sort_dimension *sort; 903 int ret = 0; 904 905 BUG_ON(list_empty(list)); 906 907 list_for_each_entry(sort, list, list) { 908 ret = sort->cmp(l, r); 909 if (ret) 910 return ret; 911 } 912 913 return ret; 914 } 915 916 static struct work_atoms * 917 thread_atoms_search(struct rb_root *root, struct thread *thread, 918 struct list_head *sort_list) 919 { 920 struct rb_node *node = root->rb_node; 921 struct work_atoms key = { .thread = thread }; 922 923 while (node) { 924 struct work_atoms *atoms; 925 int cmp; 926 927 atoms = container_of(node, struct work_atoms, node); 928 929 cmp = thread_lat_cmp(sort_list, &key, atoms); 930 if (cmp > 0) 931 node = node->rb_left; 932 else if (cmp < 0) 933 node = node->rb_right; 934 else { 935 BUG_ON(thread != atoms->thread); 936 return atoms; 937 } 938 } 939 return NULL; 940 } 941 942 static void 943 __thread_latency_insert(struct rb_root *root, struct work_atoms *data, 944 struct list_head *sort_list) 945 { 946 struct rb_node **new = &(root->rb_node), *parent = NULL; 947 948 while (*new) { 949 struct work_atoms *this; 950 int cmp; 951 952 this = container_of(*new, struct work_atoms, node); 953 parent = *new; 954 955 cmp = thread_lat_cmp(sort_list, data, this); 956 957 if (cmp > 0) 958 new = &((*new)->rb_left); 959 else 960 new = &((*new)->rb_right); 961 } 962 963 rb_link_node(&data->node, parent, new); 964 rb_insert_color(&data->node, root); 965 } 966 967 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread) 968 { 969 struct work_atoms *atoms = zalloc(sizeof(*atoms)); 970 if (!atoms) { 971 pr_err("No memory at %s\n", __func__); 972 return -1; 973 } 974 975 atoms->thread = thread__get(thread); 976 INIT_LIST_HEAD(&atoms->work_list); 977 __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid); 978 return 0; 979 } 980 981 static char sched_out_state(u64 prev_state) 982 { 983 const char *str = TASK_STATE_TO_CHAR_STR; 984 985 return str[prev_state]; 986 } 987 988 static int 989 add_sched_out_event(struct work_atoms *atoms, 990 char run_state, 991 u64 timestamp) 992 { 993 struct work_atom *atom = zalloc(sizeof(*atom)); 994 if (!atom) { 995 pr_err("Non memory at %s", __func__); 996 return -1; 997 } 998 999 atom->sched_out_time = timestamp; 1000 1001 if (run_state == 'R') { 1002 atom->state = THREAD_WAIT_CPU; 1003 atom->wake_up_time = atom->sched_out_time; 1004 } 1005 1006 list_add_tail(&atom->list, &atoms->work_list); 1007 return 0; 1008 } 1009 1010 static void 1011 add_runtime_event(struct work_atoms *atoms, u64 delta, 1012 u64 timestamp __maybe_unused) 1013 { 1014 struct work_atom *atom; 1015 1016 BUG_ON(list_empty(&atoms->work_list)); 1017 1018 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1019 1020 atom->runtime += delta; 1021 atoms->total_runtime += delta; 1022 } 1023 1024 static void 1025 add_sched_in_event(struct work_atoms *atoms, u64 timestamp) 1026 { 1027 struct work_atom *atom; 1028 u64 delta; 1029 1030 if (list_empty(&atoms->work_list)) 1031 return; 1032 1033 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1034 1035 if (atom->state != THREAD_WAIT_CPU) 1036 return; 1037 1038 if (timestamp < atom->wake_up_time) { 1039 atom->state = THREAD_IGNORE; 1040 return; 1041 } 1042 1043 atom->state = THREAD_SCHED_IN; 1044 atom->sched_in_time = timestamp; 1045 1046 delta = atom->sched_in_time - atom->wake_up_time; 1047 atoms->total_lat += delta; 1048 if (delta > atoms->max_lat) { 1049 atoms->max_lat = delta; 1050 atoms->max_lat_at = timestamp; 1051 } 1052 atoms->nb_atoms++; 1053 } 1054 1055 static int latency_switch_event(struct perf_sched *sched, 1056 struct perf_evsel *evsel, 1057 struct perf_sample *sample, 1058 struct machine *machine) 1059 { 1060 const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 1061 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1062 const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state"); 1063 struct work_atoms *out_events, *in_events; 1064 struct thread *sched_out, *sched_in; 1065 u64 timestamp0, timestamp = sample->time; 1066 int cpu = sample->cpu, err = -1; 1067 s64 delta; 1068 1069 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1070 1071 timestamp0 = sched->cpu_last_switched[cpu]; 1072 sched->cpu_last_switched[cpu] = timestamp; 1073 if (timestamp0) 1074 delta = timestamp - timestamp0; 1075 else 1076 delta = 0; 1077 1078 if (delta < 0) { 1079 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1080 return -1; 1081 } 1082 1083 sched_out = machine__findnew_thread(machine, -1, prev_pid); 1084 sched_in = machine__findnew_thread(machine, -1, next_pid); 1085 if (sched_out == NULL || sched_in == NULL) 1086 goto out_put; 1087 1088 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1089 if (!out_events) { 1090 if (thread_atoms_insert(sched, sched_out)) 1091 goto out_put; 1092 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid); 1093 if (!out_events) { 1094 pr_err("out-event: Internal tree error"); 1095 goto out_put; 1096 } 1097 } 1098 if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp)) 1099 return -1; 1100 1101 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1102 if (!in_events) { 1103 if (thread_atoms_insert(sched, sched_in)) 1104 goto out_put; 1105 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid); 1106 if (!in_events) { 1107 pr_err("in-event: Internal tree error"); 1108 goto out_put; 1109 } 1110 /* 1111 * Take came in we have not heard about yet, 1112 * add in an initial atom in runnable state: 1113 */ 1114 if (add_sched_out_event(in_events, 'R', timestamp)) 1115 goto out_put; 1116 } 1117 add_sched_in_event(in_events, timestamp); 1118 err = 0; 1119 out_put: 1120 thread__put(sched_out); 1121 thread__put(sched_in); 1122 return err; 1123 } 1124 1125 static int latency_runtime_event(struct perf_sched *sched, 1126 struct perf_evsel *evsel, 1127 struct perf_sample *sample, 1128 struct machine *machine) 1129 { 1130 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1131 const u64 runtime = perf_evsel__intval(evsel, sample, "runtime"); 1132 struct thread *thread = machine__findnew_thread(machine, -1, pid); 1133 struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1134 u64 timestamp = sample->time; 1135 int cpu = sample->cpu, err = -1; 1136 1137 if (thread == NULL) 1138 return -1; 1139 1140 BUG_ON(cpu >= MAX_CPUS || cpu < 0); 1141 if (!atoms) { 1142 if (thread_atoms_insert(sched, thread)) 1143 goto out_put; 1144 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid); 1145 if (!atoms) { 1146 pr_err("in-event: Internal tree error"); 1147 goto out_put; 1148 } 1149 if (add_sched_out_event(atoms, 'R', timestamp)) 1150 goto out_put; 1151 } 1152 1153 add_runtime_event(atoms, runtime, timestamp); 1154 err = 0; 1155 out_put: 1156 thread__put(thread); 1157 return err; 1158 } 1159 1160 static int latency_wakeup_event(struct perf_sched *sched, 1161 struct perf_evsel *evsel, 1162 struct perf_sample *sample, 1163 struct machine *machine) 1164 { 1165 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1166 struct work_atoms *atoms; 1167 struct work_atom *atom; 1168 struct thread *wakee; 1169 u64 timestamp = sample->time; 1170 int err = -1; 1171 1172 wakee = machine__findnew_thread(machine, -1, pid); 1173 if (wakee == NULL) 1174 return -1; 1175 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1176 if (!atoms) { 1177 if (thread_atoms_insert(sched, wakee)) 1178 goto out_put; 1179 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid); 1180 if (!atoms) { 1181 pr_err("wakeup-event: Internal tree error"); 1182 goto out_put; 1183 } 1184 if (add_sched_out_event(atoms, 'S', timestamp)) 1185 goto out_put; 1186 } 1187 1188 BUG_ON(list_empty(&atoms->work_list)); 1189 1190 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1191 1192 /* 1193 * As we do not guarantee the wakeup event happens when 1194 * task is out of run queue, also may happen when task is 1195 * on run queue and wakeup only change ->state to TASK_RUNNING, 1196 * then we should not set the ->wake_up_time when wake up a 1197 * task which is on run queue. 1198 * 1199 * You WILL be missing events if you've recorded only 1200 * one CPU, or are only looking at only one, so don't 1201 * skip in this case. 1202 */ 1203 if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING) 1204 goto out_ok; 1205 1206 sched->nr_timestamps++; 1207 if (atom->sched_out_time > timestamp) { 1208 sched->nr_unordered_timestamps++; 1209 goto out_ok; 1210 } 1211 1212 atom->state = THREAD_WAIT_CPU; 1213 atom->wake_up_time = timestamp; 1214 out_ok: 1215 err = 0; 1216 out_put: 1217 thread__put(wakee); 1218 return err; 1219 } 1220 1221 static int latency_migrate_task_event(struct perf_sched *sched, 1222 struct perf_evsel *evsel, 1223 struct perf_sample *sample, 1224 struct machine *machine) 1225 { 1226 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 1227 u64 timestamp = sample->time; 1228 struct work_atoms *atoms; 1229 struct work_atom *atom; 1230 struct thread *migrant; 1231 int err = -1; 1232 1233 /* 1234 * Only need to worry about migration when profiling one CPU. 1235 */ 1236 if (sched->profile_cpu == -1) 1237 return 0; 1238 1239 migrant = machine__findnew_thread(machine, -1, pid); 1240 if (migrant == NULL) 1241 return -1; 1242 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1243 if (!atoms) { 1244 if (thread_atoms_insert(sched, migrant)) 1245 goto out_put; 1246 register_pid(sched, migrant->tid, thread__comm_str(migrant)); 1247 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid); 1248 if (!atoms) { 1249 pr_err("migration-event: Internal tree error"); 1250 goto out_put; 1251 } 1252 if (add_sched_out_event(atoms, 'R', timestamp)) 1253 goto out_put; 1254 } 1255 1256 BUG_ON(list_empty(&atoms->work_list)); 1257 1258 atom = list_entry(atoms->work_list.prev, struct work_atom, list); 1259 atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp; 1260 1261 sched->nr_timestamps++; 1262 1263 if (atom->sched_out_time > timestamp) 1264 sched->nr_unordered_timestamps++; 1265 err = 0; 1266 out_put: 1267 thread__put(migrant); 1268 return err; 1269 } 1270 1271 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list) 1272 { 1273 int i; 1274 int ret; 1275 u64 avg; 1276 char max_lat_at[32]; 1277 1278 if (!work_list->nb_atoms) 1279 return; 1280 /* 1281 * Ignore idle threads: 1282 */ 1283 if (!strcmp(thread__comm_str(work_list->thread), "swapper")) 1284 return; 1285 1286 sched->all_runtime += work_list->total_runtime; 1287 sched->all_count += work_list->nb_atoms; 1288 1289 if (work_list->num_merged > 1) 1290 ret = printf(" %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged); 1291 else 1292 ret = printf(" %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid); 1293 1294 for (i = 0; i < 24 - ret; i++) 1295 printf(" "); 1296 1297 avg = work_list->total_lat / work_list->nb_atoms; 1298 timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at)); 1299 1300 printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n", 1301 (double)work_list->total_runtime / NSEC_PER_MSEC, 1302 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC, 1303 (double)work_list->max_lat / NSEC_PER_MSEC, 1304 max_lat_at); 1305 } 1306 1307 static int pid_cmp(struct work_atoms *l, struct work_atoms *r) 1308 { 1309 if (l->thread == r->thread) 1310 return 0; 1311 if (l->thread->tid < r->thread->tid) 1312 return -1; 1313 if (l->thread->tid > r->thread->tid) 1314 return 1; 1315 return (int)(l->thread - r->thread); 1316 } 1317 1318 static int avg_cmp(struct work_atoms *l, struct work_atoms *r) 1319 { 1320 u64 avgl, avgr; 1321 1322 if (!l->nb_atoms) 1323 return -1; 1324 1325 if (!r->nb_atoms) 1326 return 1; 1327 1328 avgl = l->total_lat / l->nb_atoms; 1329 avgr = r->total_lat / r->nb_atoms; 1330 1331 if (avgl < avgr) 1332 return -1; 1333 if (avgl > avgr) 1334 return 1; 1335 1336 return 0; 1337 } 1338 1339 static int max_cmp(struct work_atoms *l, struct work_atoms *r) 1340 { 1341 if (l->max_lat < r->max_lat) 1342 return -1; 1343 if (l->max_lat > r->max_lat) 1344 return 1; 1345 1346 return 0; 1347 } 1348 1349 static int switch_cmp(struct work_atoms *l, struct work_atoms *r) 1350 { 1351 if (l->nb_atoms < r->nb_atoms) 1352 return -1; 1353 if (l->nb_atoms > r->nb_atoms) 1354 return 1; 1355 1356 return 0; 1357 } 1358 1359 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r) 1360 { 1361 if (l->total_runtime < r->total_runtime) 1362 return -1; 1363 if (l->total_runtime > r->total_runtime) 1364 return 1; 1365 1366 return 0; 1367 } 1368 1369 static int sort_dimension__add(const char *tok, struct list_head *list) 1370 { 1371 size_t i; 1372 static struct sort_dimension avg_sort_dimension = { 1373 .name = "avg", 1374 .cmp = avg_cmp, 1375 }; 1376 static struct sort_dimension max_sort_dimension = { 1377 .name = "max", 1378 .cmp = max_cmp, 1379 }; 1380 static struct sort_dimension pid_sort_dimension = { 1381 .name = "pid", 1382 .cmp = pid_cmp, 1383 }; 1384 static struct sort_dimension runtime_sort_dimension = { 1385 .name = "runtime", 1386 .cmp = runtime_cmp, 1387 }; 1388 static struct sort_dimension switch_sort_dimension = { 1389 .name = "switch", 1390 .cmp = switch_cmp, 1391 }; 1392 struct sort_dimension *available_sorts[] = { 1393 &pid_sort_dimension, 1394 &avg_sort_dimension, 1395 &max_sort_dimension, 1396 &switch_sort_dimension, 1397 &runtime_sort_dimension, 1398 }; 1399 1400 for (i = 0; i < ARRAY_SIZE(available_sorts); i++) { 1401 if (!strcmp(available_sorts[i]->name, tok)) { 1402 list_add_tail(&available_sorts[i]->list, list); 1403 1404 return 0; 1405 } 1406 } 1407 1408 return -1; 1409 } 1410 1411 static void perf_sched__sort_lat(struct perf_sched *sched) 1412 { 1413 struct rb_node *node; 1414 struct rb_root *root = &sched->atom_root; 1415 again: 1416 for (;;) { 1417 struct work_atoms *data; 1418 node = rb_first(root); 1419 if (!node) 1420 break; 1421 1422 rb_erase(node, root); 1423 data = rb_entry(node, struct work_atoms, node); 1424 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list); 1425 } 1426 if (root == &sched->atom_root) { 1427 root = &sched->merged_atom_root; 1428 goto again; 1429 } 1430 } 1431 1432 static int process_sched_wakeup_event(struct perf_tool *tool, 1433 struct perf_evsel *evsel, 1434 struct perf_sample *sample, 1435 struct machine *machine) 1436 { 1437 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1438 1439 if (sched->tp_handler->wakeup_event) 1440 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine); 1441 1442 return 0; 1443 } 1444 1445 union map_priv { 1446 void *ptr; 1447 bool color; 1448 }; 1449 1450 static bool thread__has_color(struct thread *thread) 1451 { 1452 union map_priv priv = { 1453 .ptr = thread__priv(thread), 1454 }; 1455 1456 return priv.color; 1457 } 1458 1459 static struct thread* 1460 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid) 1461 { 1462 struct thread *thread = machine__findnew_thread(machine, pid, tid); 1463 union map_priv priv = { 1464 .color = false, 1465 }; 1466 1467 if (!sched->map.color_pids || !thread || thread__priv(thread)) 1468 return thread; 1469 1470 if (thread_map__has(sched->map.color_pids, tid)) 1471 priv.color = true; 1472 1473 thread__set_priv(thread, priv.ptr); 1474 return thread; 1475 } 1476 1477 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel, 1478 struct perf_sample *sample, struct machine *machine) 1479 { 1480 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1481 struct thread *sched_in; 1482 int new_shortname; 1483 u64 timestamp0, timestamp = sample->time; 1484 s64 delta; 1485 int i, this_cpu = sample->cpu; 1486 int cpus_nr; 1487 bool new_cpu = false; 1488 const char *color = PERF_COLOR_NORMAL; 1489 char stimestamp[32]; 1490 1491 BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0); 1492 1493 if (this_cpu > sched->max_cpu) 1494 sched->max_cpu = this_cpu; 1495 1496 if (sched->map.comp) { 1497 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS); 1498 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) { 1499 sched->map.comp_cpus[cpus_nr++] = this_cpu; 1500 new_cpu = true; 1501 } 1502 } else 1503 cpus_nr = sched->max_cpu; 1504 1505 timestamp0 = sched->cpu_last_switched[this_cpu]; 1506 sched->cpu_last_switched[this_cpu] = timestamp; 1507 if (timestamp0) 1508 delta = timestamp - timestamp0; 1509 else 1510 delta = 0; 1511 1512 if (delta < 0) { 1513 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta); 1514 return -1; 1515 } 1516 1517 sched_in = map__findnew_thread(sched, machine, -1, next_pid); 1518 if (sched_in == NULL) 1519 return -1; 1520 1521 sched->curr_thread[this_cpu] = thread__get(sched_in); 1522 1523 printf(" "); 1524 1525 new_shortname = 0; 1526 if (!sched_in->shortname[0]) { 1527 if (!strcmp(thread__comm_str(sched_in), "swapper")) { 1528 /* 1529 * Don't allocate a letter-number for swapper:0 1530 * as a shortname. Instead, we use '.' for it. 1531 */ 1532 sched_in->shortname[0] = '.'; 1533 sched_in->shortname[1] = ' '; 1534 } else { 1535 sched_in->shortname[0] = sched->next_shortname1; 1536 sched_in->shortname[1] = sched->next_shortname2; 1537 1538 if (sched->next_shortname1 < 'Z') { 1539 sched->next_shortname1++; 1540 } else { 1541 sched->next_shortname1 = 'A'; 1542 if (sched->next_shortname2 < '9') 1543 sched->next_shortname2++; 1544 else 1545 sched->next_shortname2 = '0'; 1546 } 1547 } 1548 new_shortname = 1; 1549 } 1550 1551 for (i = 0; i < cpus_nr; i++) { 1552 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i; 1553 struct thread *curr_thread = sched->curr_thread[cpu]; 1554 const char *pid_color = color; 1555 const char *cpu_color = color; 1556 1557 if (curr_thread && thread__has_color(curr_thread)) 1558 pid_color = COLOR_PIDS; 1559 1560 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu)) 1561 continue; 1562 1563 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu)) 1564 cpu_color = COLOR_CPUS; 1565 1566 if (cpu != this_cpu) 1567 color_fprintf(stdout, color, " "); 1568 else 1569 color_fprintf(stdout, cpu_color, "*"); 1570 1571 if (sched->curr_thread[cpu]) 1572 color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname); 1573 else 1574 color_fprintf(stdout, color, " "); 1575 } 1576 1577 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu)) 1578 goto out; 1579 1580 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp)); 1581 color_fprintf(stdout, color, " %12s secs ", stimestamp); 1582 if (new_shortname || (verbose > 0 && sched_in->tid)) { 1583 const char *pid_color = color; 1584 1585 if (thread__has_color(sched_in)) 1586 pid_color = COLOR_PIDS; 1587 1588 color_fprintf(stdout, pid_color, "%s => %s:%d", 1589 sched_in->shortname, thread__comm_str(sched_in), sched_in->tid); 1590 } 1591 1592 if (sched->map.comp && new_cpu) 1593 color_fprintf(stdout, color, " (CPU %d)", this_cpu); 1594 1595 out: 1596 color_fprintf(stdout, color, "\n"); 1597 1598 thread__put(sched_in); 1599 1600 return 0; 1601 } 1602 1603 static int process_sched_switch_event(struct perf_tool *tool, 1604 struct perf_evsel *evsel, 1605 struct perf_sample *sample, 1606 struct machine *machine) 1607 { 1608 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1609 int this_cpu = sample->cpu, err = 0; 1610 u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"), 1611 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1612 1613 if (sched->curr_pid[this_cpu] != (u32)-1) { 1614 /* 1615 * Are we trying to switch away a PID that is 1616 * not current? 1617 */ 1618 if (sched->curr_pid[this_cpu] != prev_pid) 1619 sched->nr_context_switch_bugs++; 1620 } 1621 1622 if (sched->tp_handler->switch_event) 1623 err = sched->tp_handler->switch_event(sched, evsel, sample, machine); 1624 1625 sched->curr_pid[this_cpu] = next_pid; 1626 return err; 1627 } 1628 1629 static int process_sched_runtime_event(struct perf_tool *tool, 1630 struct perf_evsel *evsel, 1631 struct perf_sample *sample, 1632 struct machine *machine) 1633 { 1634 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1635 1636 if (sched->tp_handler->runtime_event) 1637 return sched->tp_handler->runtime_event(sched, evsel, sample, machine); 1638 1639 return 0; 1640 } 1641 1642 static int perf_sched__process_fork_event(struct perf_tool *tool, 1643 union perf_event *event, 1644 struct perf_sample *sample, 1645 struct machine *machine) 1646 { 1647 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1648 1649 /* run the fork event through the perf machineruy */ 1650 perf_event__process_fork(tool, event, sample, machine); 1651 1652 /* and then run additional processing needed for this command */ 1653 if (sched->tp_handler->fork_event) 1654 return sched->tp_handler->fork_event(sched, event, machine); 1655 1656 return 0; 1657 } 1658 1659 static int process_sched_migrate_task_event(struct perf_tool *tool, 1660 struct perf_evsel *evsel, 1661 struct perf_sample *sample, 1662 struct machine *machine) 1663 { 1664 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 1665 1666 if (sched->tp_handler->migrate_task_event) 1667 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine); 1668 1669 return 0; 1670 } 1671 1672 typedef int (*tracepoint_handler)(struct perf_tool *tool, 1673 struct perf_evsel *evsel, 1674 struct perf_sample *sample, 1675 struct machine *machine); 1676 1677 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused, 1678 union perf_event *event __maybe_unused, 1679 struct perf_sample *sample, 1680 struct perf_evsel *evsel, 1681 struct machine *machine) 1682 { 1683 int err = 0; 1684 1685 if (evsel->handler != NULL) { 1686 tracepoint_handler f = evsel->handler; 1687 err = f(tool, evsel, sample, machine); 1688 } 1689 1690 return err; 1691 } 1692 1693 static int perf_sched__read_events(struct perf_sched *sched) 1694 { 1695 const struct perf_evsel_str_handler handlers[] = { 1696 { "sched:sched_switch", process_sched_switch_event, }, 1697 { "sched:sched_stat_runtime", process_sched_runtime_event, }, 1698 { "sched:sched_wakeup", process_sched_wakeup_event, }, 1699 { "sched:sched_wakeup_new", process_sched_wakeup_event, }, 1700 { "sched:sched_migrate_task", process_sched_migrate_task_event, }, 1701 }; 1702 struct perf_session *session; 1703 struct perf_data_file file = { 1704 .path = input_name, 1705 .mode = PERF_DATA_MODE_READ, 1706 .force = sched->force, 1707 }; 1708 int rc = -1; 1709 1710 session = perf_session__new(&file, false, &sched->tool); 1711 if (session == NULL) { 1712 pr_debug("No Memory for session\n"); 1713 return -1; 1714 } 1715 1716 symbol__init(&session->header.env); 1717 1718 if (perf_session__set_tracepoints_handlers(session, handlers)) 1719 goto out_delete; 1720 1721 if (perf_session__has_traces(session, "record -R")) { 1722 int err = perf_session__process_events(session); 1723 if (err) { 1724 pr_err("Failed to process events, error %d", err); 1725 goto out_delete; 1726 } 1727 1728 sched->nr_events = session->evlist->stats.nr_events[0]; 1729 sched->nr_lost_events = session->evlist->stats.total_lost; 1730 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST]; 1731 } 1732 1733 rc = 0; 1734 out_delete: 1735 perf_session__delete(session); 1736 return rc; 1737 } 1738 1739 /* 1740 * scheduling times are printed as msec.usec 1741 */ 1742 static inline void print_sched_time(unsigned long long nsecs, int width) 1743 { 1744 unsigned long msecs; 1745 unsigned long usecs; 1746 1747 msecs = nsecs / NSEC_PER_MSEC; 1748 nsecs -= msecs * NSEC_PER_MSEC; 1749 usecs = nsecs / NSEC_PER_USEC; 1750 printf("%*lu.%03lu ", width, msecs, usecs); 1751 } 1752 1753 /* 1754 * returns runtime data for event, allocating memory for it the 1755 * first time it is used. 1756 */ 1757 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel) 1758 { 1759 struct evsel_runtime *r = evsel->priv; 1760 1761 if (r == NULL) { 1762 r = zalloc(sizeof(struct evsel_runtime)); 1763 evsel->priv = r; 1764 } 1765 1766 return r; 1767 } 1768 1769 /* 1770 * save last time event was seen per cpu 1771 */ 1772 static void perf_evsel__save_time(struct perf_evsel *evsel, 1773 u64 timestamp, u32 cpu) 1774 { 1775 struct evsel_runtime *r = perf_evsel__get_runtime(evsel); 1776 1777 if (r == NULL) 1778 return; 1779 1780 if ((cpu >= r->ncpu) || (r->last_time == NULL)) { 1781 int i, n = __roundup_pow_of_two(cpu+1); 1782 void *p = r->last_time; 1783 1784 p = realloc(r->last_time, n * sizeof(u64)); 1785 if (!p) 1786 return; 1787 1788 r->last_time = p; 1789 for (i = r->ncpu; i < n; ++i) 1790 r->last_time[i] = (u64) 0; 1791 1792 r->ncpu = n; 1793 } 1794 1795 r->last_time[cpu] = timestamp; 1796 } 1797 1798 /* returns last time this event was seen on the given cpu */ 1799 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu) 1800 { 1801 struct evsel_runtime *r = perf_evsel__get_runtime(evsel); 1802 1803 if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu)) 1804 return 0; 1805 1806 return r->last_time[cpu]; 1807 } 1808 1809 static int comm_width = 30; 1810 1811 static char *timehist_get_commstr(struct thread *thread) 1812 { 1813 static char str[32]; 1814 const char *comm = thread__comm_str(thread); 1815 pid_t tid = thread->tid; 1816 pid_t pid = thread->pid_; 1817 int n; 1818 1819 if (pid == 0) 1820 n = scnprintf(str, sizeof(str), "%s", comm); 1821 1822 else if (tid != pid) 1823 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid); 1824 1825 else 1826 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid); 1827 1828 if (n > comm_width) 1829 comm_width = n; 1830 1831 return str; 1832 } 1833 1834 static void timehist_header(struct perf_sched *sched) 1835 { 1836 u32 ncpus = sched->max_cpu + 1; 1837 u32 i, j; 1838 1839 printf("%15s %6s ", "time", "cpu"); 1840 1841 if (sched->show_cpu_visual) { 1842 printf(" "); 1843 for (i = 0, j = 0; i < ncpus; ++i) { 1844 printf("%x", j++); 1845 if (j > 15) 1846 j = 0; 1847 } 1848 printf(" "); 1849 } 1850 1851 printf(" %-*s %9s %9s %9s", comm_width, 1852 "task name", "wait time", "sch delay", "run time"); 1853 1854 if (sched->show_state) 1855 printf(" %s", "state"); 1856 1857 printf("\n"); 1858 1859 /* 1860 * units row 1861 */ 1862 printf("%15s %-6s ", "", ""); 1863 1864 if (sched->show_cpu_visual) 1865 printf(" %*s ", ncpus, ""); 1866 1867 printf(" %-*s %9s %9s %9s", comm_width, 1868 "[tid/pid]", "(msec)", "(msec)", "(msec)"); 1869 1870 if (sched->show_state) 1871 printf(" %5s", ""); 1872 1873 printf("\n"); 1874 1875 /* 1876 * separator 1877 */ 1878 printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line); 1879 1880 if (sched->show_cpu_visual) 1881 printf(" %.*s ", ncpus, graph_dotted_line); 1882 1883 printf(" %.*s %.9s %.9s %.9s", comm_width, 1884 graph_dotted_line, graph_dotted_line, graph_dotted_line, 1885 graph_dotted_line); 1886 1887 if (sched->show_state) 1888 printf(" %.5s", graph_dotted_line); 1889 1890 printf("\n"); 1891 } 1892 1893 static char task_state_char(struct thread *thread, int state) 1894 { 1895 static const char state_to_char[] = TASK_STATE_TO_CHAR_STR; 1896 unsigned bit = state ? ffs(state) : 0; 1897 1898 /* 'I' for idle */ 1899 if (thread->tid == 0) 1900 return 'I'; 1901 1902 return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?'; 1903 } 1904 1905 static void timehist_print_sample(struct perf_sched *sched, 1906 struct perf_evsel *evsel, 1907 struct perf_sample *sample, 1908 struct addr_location *al, 1909 struct thread *thread, 1910 u64 t, int state) 1911 { 1912 struct thread_runtime *tr = thread__priv(thread); 1913 const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm"); 1914 const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid"); 1915 u32 max_cpus = sched->max_cpu + 1; 1916 char tstr[64]; 1917 char nstr[30]; 1918 u64 wait_time; 1919 1920 timestamp__scnprintf_usec(t, tstr, sizeof(tstr)); 1921 printf("%15s [%04d] ", tstr, sample->cpu); 1922 1923 if (sched->show_cpu_visual) { 1924 u32 i; 1925 char c; 1926 1927 printf(" "); 1928 for (i = 0; i < max_cpus; ++i) { 1929 /* flag idle times with 'i'; others are sched events */ 1930 if (i == sample->cpu) 1931 c = (thread->tid == 0) ? 'i' : 's'; 1932 else 1933 c = ' '; 1934 printf("%c", c); 1935 } 1936 printf(" "); 1937 } 1938 1939 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 1940 1941 wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt; 1942 print_sched_time(wait_time, 6); 1943 1944 print_sched_time(tr->dt_delay, 6); 1945 print_sched_time(tr->dt_run, 6); 1946 1947 if (sched->show_state) 1948 printf(" %5c ", task_state_char(thread, state)); 1949 1950 if (sched->show_next) { 1951 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid); 1952 printf(" %-*s", comm_width, nstr); 1953 } 1954 1955 if (sched->show_wakeups && !sched->show_next) 1956 printf(" %-*s", comm_width, ""); 1957 1958 if (thread->tid == 0) 1959 goto out; 1960 1961 if (sched->show_callchain) 1962 printf(" "); 1963 1964 sample__fprintf_sym(sample, al, 0, 1965 EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE | 1966 EVSEL__PRINT_CALLCHAIN_ARROW | 1967 EVSEL__PRINT_SKIP_IGNORED, 1968 &callchain_cursor, stdout); 1969 1970 out: 1971 printf("\n"); 1972 } 1973 1974 /* 1975 * Explanation of delta-time stats: 1976 * 1977 * t = time of current schedule out event 1978 * tprev = time of previous sched out event 1979 * also time of schedule-in event for current task 1980 * last_time = time of last sched change event for current task 1981 * (i.e, time process was last scheduled out) 1982 * ready_to_run = time of wakeup for current task 1983 * 1984 * -----|------------|------------|------------|------ 1985 * last ready tprev t 1986 * time to run 1987 * 1988 * |-------- dt_wait --------| 1989 * |- dt_delay -|-- dt_run --| 1990 * 1991 * dt_run = run time of current task 1992 * dt_wait = time between last schedule out event for task and tprev 1993 * represents time spent off the cpu 1994 * dt_delay = time between wakeup and schedule-in of task 1995 */ 1996 1997 static void timehist_update_runtime_stats(struct thread_runtime *r, 1998 u64 t, u64 tprev) 1999 { 2000 r->dt_delay = 0; 2001 r->dt_sleep = 0; 2002 r->dt_iowait = 0; 2003 r->dt_preempt = 0; 2004 r->dt_run = 0; 2005 2006 if (tprev) { 2007 r->dt_run = t - tprev; 2008 if (r->ready_to_run) { 2009 if (r->ready_to_run > tprev) 2010 pr_debug("time travel: wakeup time for task > previous sched_switch event\n"); 2011 else 2012 r->dt_delay = tprev - r->ready_to_run; 2013 } 2014 2015 if (r->last_time > tprev) 2016 pr_debug("time travel: last sched out time for task > previous sched_switch event\n"); 2017 else if (r->last_time) { 2018 u64 dt_wait = tprev - r->last_time; 2019 2020 if (r->last_state == TASK_RUNNING) 2021 r->dt_preempt = dt_wait; 2022 else if (r->last_state == TASK_UNINTERRUPTIBLE) 2023 r->dt_iowait = dt_wait; 2024 else 2025 r->dt_sleep = dt_wait; 2026 } 2027 } 2028 2029 update_stats(&r->run_stats, r->dt_run); 2030 2031 r->total_run_time += r->dt_run; 2032 r->total_delay_time += r->dt_delay; 2033 r->total_sleep_time += r->dt_sleep; 2034 r->total_iowait_time += r->dt_iowait; 2035 r->total_preempt_time += r->dt_preempt; 2036 } 2037 2038 static bool is_idle_sample(struct perf_sample *sample, 2039 struct perf_evsel *evsel) 2040 { 2041 /* pid 0 == swapper == idle task */ 2042 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0) 2043 return perf_evsel__intval(evsel, sample, "prev_pid") == 0; 2044 2045 return sample->pid == 0; 2046 } 2047 2048 static void save_task_callchain(struct perf_sched *sched, 2049 struct perf_sample *sample, 2050 struct perf_evsel *evsel, 2051 struct machine *machine) 2052 { 2053 struct callchain_cursor *cursor = &callchain_cursor; 2054 struct thread *thread; 2055 2056 /* want main thread for process - has maps */ 2057 thread = machine__findnew_thread(machine, sample->pid, sample->pid); 2058 if (thread == NULL) { 2059 pr_debug("Failed to get thread for pid %d.\n", sample->pid); 2060 return; 2061 } 2062 2063 if (!symbol_conf.use_callchain || sample->callchain == NULL) 2064 return; 2065 2066 if (thread__resolve_callchain(thread, cursor, evsel, sample, 2067 NULL, NULL, sched->max_stack + 2) != 0) { 2068 if (verbose > 0) 2069 error("Failed to resolve callchain. Skipping\n"); 2070 2071 return; 2072 } 2073 2074 callchain_cursor_commit(cursor); 2075 2076 while (true) { 2077 struct callchain_cursor_node *node; 2078 struct symbol *sym; 2079 2080 node = callchain_cursor_current(cursor); 2081 if (node == NULL) 2082 break; 2083 2084 sym = node->sym; 2085 if (sym) { 2086 if (!strcmp(sym->name, "schedule") || 2087 !strcmp(sym->name, "__schedule") || 2088 !strcmp(sym->name, "preempt_schedule")) 2089 sym->ignore = 1; 2090 } 2091 2092 callchain_cursor_advance(cursor); 2093 } 2094 } 2095 2096 static int init_idle_thread(struct thread *thread) 2097 { 2098 struct idle_thread_runtime *itr; 2099 2100 thread__set_comm(thread, idle_comm, 0); 2101 2102 itr = zalloc(sizeof(*itr)); 2103 if (itr == NULL) 2104 return -ENOMEM; 2105 2106 init_stats(&itr->tr.run_stats); 2107 callchain_init(&itr->callchain); 2108 callchain_cursor_reset(&itr->cursor); 2109 thread__set_priv(thread, itr); 2110 2111 return 0; 2112 } 2113 2114 /* 2115 * Track idle stats per cpu by maintaining a local thread 2116 * struct for the idle task on each cpu. 2117 */ 2118 static int init_idle_threads(int ncpu) 2119 { 2120 int i, ret; 2121 2122 idle_threads = zalloc(ncpu * sizeof(struct thread *)); 2123 if (!idle_threads) 2124 return -ENOMEM; 2125 2126 idle_max_cpu = ncpu; 2127 2128 /* allocate the actual thread struct if needed */ 2129 for (i = 0; i < ncpu; ++i) { 2130 idle_threads[i] = thread__new(0, 0); 2131 if (idle_threads[i] == NULL) 2132 return -ENOMEM; 2133 2134 ret = init_idle_thread(idle_threads[i]); 2135 if (ret < 0) 2136 return ret; 2137 } 2138 2139 return 0; 2140 } 2141 2142 static void free_idle_threads(void) 2143 { 2144 int i; 2145 2146 if (idle_threads == NULL) 2147 return; 2148 2149 for (i = 0; i < idle_max_cpu; ++i) { 2150 if ((idle_threads[i])) 2151 thread__delete(idle_threads[i]); 2152 } 2153 2154 free(idle_threads); 2155 } 2156 2157 static struct thread *get_idle_thread(int cpu) 2158 { 2159 /* 2160 * expand/allocate array of pointers to local thread 2161 * structs if needed 2162 */ 2163 if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) { 2164 int i, j = __roundup_pow_of_two(cpu+1); 2165 void *p; 2166 2167 p = realloc(idle_threads, j * sizeof(struct thread *)); 2168 if (!p) 2169 return NULL; 2170 2171 idle_threads = (struct thread **) p; 2172 for (i = idle_max_cpu; i < j; ++i) 2173 idle_threads[i] = NULL; 2174 2175 idle_max_cpu = j; 2176 } 2177 2178 /* allocate a new thread struct if needed */ 2179 if (idle_threads[cpu] == NULL) { 2180 idle_threads[cpu] = thread__new(0, 0); 2181 2182 if (idle_threads[cpu]) { 2183 if (init_idle_thread(idle_threads[cpu]) < 0) 2184 return NULL; 2185 } 2186 } 2187 2188 return idle_threads[cpu]; 2189 } 2190 2191 static void save_idle_callchain(struct idle_thread_runtime *itr, 2192 struct perf_sample *sample) 2193 { 2194 if (!symbol_conf.use_callchain || sample->callchain == NULL) 2195 return; 2196 2197 callchain_cursor__copy(&itr->cursor, &callchain_cursor); 2198 } 2199 2200 /* 2201 * handle runtime stats saved per thread 2202 */ 2203 static struct thread_runtime *thread__init_runtime(struct thread *thread) 2204 { 2205 struct thread_runtime *r; 2206 2207 r = zalloc(sizeof(struct thread_runtime)); 2208 if (!r) 2209 return NULL; 2210 2211 init_stats(&r->run_stats); 2212 thread__set_priv(thread, r); 2213 2214 return r; 2215 } 2216 2217 static struct thread_runtime *thread__get_runtime(struct thread *thread) 2218 { 2219 struct thread_runtime *tr; 2220 2221 tr = thread__priv(thread); 2222 if (tr == NULL) { 2223 tr = thread__init_runtime(thread); 2224 if (tr == NULL) 2225 pr_debug("Failed to malloc memory for runtime data.\n"); 2226 } 2227 2228 return tr; 2229 } 2230 2231 static struct thread *timehist_get_thread(struct perf_sched *sched, 2232 struct perf_sample *sample, 2233 struct machine *machine, 2234 struct perf_evsel *evsel) 2235 { 2236 struct thread *thread; 2237 2238 if (is_idle_sample(sample, evsel)) { 2239 thread = get_idle_thread(sample->cpu); 2240 if (thread == NULL) 2241 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2242 2243 } else { 2244 /* there were samples with tid 0 but non-zero pid */ 2245 thread = machine__findnew_thread(machine, sample->pid, 2246 sample->tid ?: sample->pid); 2247 if (thread == NULL) { 2248 pr_debug("Failed to get thread for tid %d. skipping sample.\n", 2249 sample->tid); 2250 } 2251 2252 save_task_callchain(sched, sample, evsel, machine); 2253 if (sched->idle_hist) { 2254 struct thread *idle; 2255 struct idle_thread_runtime *itr; 2256 2257 idle = get_idle_thread(sample->cpu); 2258 if (idle == NULL) { 2259 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu); 2260 return NULL; 2261 } 2262 2263 itr = thread__priv(idle); 2264 if (itr == NULL) 2265 return NULL; 2266 2267 itr->last_thread = thread; 2268 2269 /* copy task callchain when entering to idle */ 2270 if (perf_evsel__intval(evsel, sample, "next_pid") == 0) 2271 save_idle_callchain(itr, sample); 2272 } 2273 } 2274 2275 return thread; 2276 } 2277 2278 static bool timehist_skip_sample(struct perf_sched *sched, 2279 struct thread *thread, 2280 struct perf_evsel *evsel, 2281 struct perf_sample *sample) 2282 { 2283 bool rc = false; 2284 2285 if (thread__is_filtered(thread)) { 2286 rc = true; 2287 sched->skipped_samples++; 2288 } 2289 2290 if (sched->idle_hist) { 2291 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch")) 2292 rc = true; 2293 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 && 2294 perf_evsel__intval(evsel, sample, "next_pid") != 0) 2295 rc = true; 2296 } 2297 2298 return rc; 2299 } 2300 2301 static void timehist_print_wakeup_event(struct perf_sched *sched, 2302 struct perf_evsel *evsel, 2303 struct perf_sample *sample, 2304 struct machine *machine, 2305 struct thread *awakened) 2306 { 2307 struct thread *thread; 2308 char tstr[64]; 2309 2310 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2311 if (thread == NULL) 2312 return; 2313 2314 /* show wakeup unless both awakee and awaker are filtered */ 2315 if (timehist_skip_sample(sched, thread, evsel, sample) && 2316 timehist_skip_sample(sched, awakened, evsel, sample)) { 2317 return; 2318 } 2319 2320 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2321 printf("%15s [%04d] ", tstr, sample->cpu); 2322 if (sched->show_cpu_visual) 2323 printf(" %*s ", sched->max_cpu + 1, ""); 2324 2325 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2326 2327 /* dt spacer */ 2328 printf(" %9s %9s %9s ", "", "", ""); 2329 2330 printf("awakened: %s", timehist_get_commstr(awakened)); 2331 2332 printf("\n"); 2333 } 2334 2335 static int timehist_sched_wakeup_event(struct perf_tool *tool, 2336 union perf_event *event __maybe_unused, 2337 struct perf_evsel *evsel, 2338 struct perf_sample *sample, 2339 struct machine *machine) 2340 { 2341 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2342 struct thread *thread; 2343 struct thread_runtime *tr = NULL; 2344 /* want pid of awakened task not pid in sample */ 2345 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 2346 2347 thread = machine__findnew_thread(machine, 0, pid); 2348 if (thread == NULL) 2349 return -1; 2350 2351 tr = thread__get_runtime(thread); 2352 if (tr == NULL) 2353 return -1; 2354 2355 if (tr->ready_to_run == 0) 2356 tr->ready_to_run = sample->time; 2357 2358 /* show wakeups if requested */ 2359 if (sched->show_wakeups && 2360 !perf_time__skip_sample(&sched->ptime, sample->time)) 2361 timehist_print_wakeup_event(sched, evsel, sample, machine, thread); 2362 2363 return 0; 2364 } 2365 2366 static void timehist_print_migration_event(struct perf_sched *sched, 2367 struct perf_evsel *evsel, 2368 struct perf_sample *sample, 2369 struct machine *machine, 2370 struct thread *migrated) 2371 { 2372 struct thread *thread; 2373 char tstr[64]; 2374 u32 max_cpus = sched->max_cpu + 1; 2375 u32 ocpu, dcpu; 2376 2377 if (sched->summary_only) 2378 return; 2379 2380 max_cpus = sched->max_cpu + 1; 2381 ocpu = perf_evsel__intval(evsel, sample, "orig_cpu"); 2382 dcpu = perf_evsel__intval(evsel, sample, "dest_cpu"); 2383 2384 thread = machine__findnew_thread(machine, sample->pid, sample->tid); 2385 if (thread == NULL) 2386 return; 2387 2388 if (timehist_skip_sample(sched, thread, evsel, sample) && 2389 timehist_skip_sample(sched, migrated, evsel, sample)) { 2390 return; 2391 } 2392 2393 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2394 printf("%15s [%04d] ", tstr, sample->cpu); 2395 2396 if (sched->show_cpu_visual) { 2397 u32 i; 2398 char c; 2399 2400 printf(" "); 2401 for (i = 0; i < max_cpus; ++i) { 2402 c = (i == sample->cpu) ? 'm' : ' '; 2403 printf("%c", c); 2404 } 2405 printf(" "); 2406 } 2407 2408 printf(" %-*s ", comm_width, timehist_get_commstr(thread)); 2409 2410 /* dt spacer */ 2411 printf(" %9s %9s %9s ", "", "", ""); 2412 2413 printf("migrated: %s", timehist_get_commstr(migrated)); 2414 printf(" cpu %d => %d", ocpu, dcpu); 2415 2416 printf("\n"); 2417 } 2418 2419 static int timehist_migrate_task_event(struct perf_tool *tool, 2420 union perf_event *event __maybe_unused, 2421 struct perf_evsel *evsel, 2422 struct perf_sample *sample, 2423 struct machine *machine) 2424 { 2425 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2426 struct thread *thread; 2427 struct thread_runtime *tr = NULL; 2428 /* want pid of migrated task not pid in sample */ 2429 const u32 pid = perf_evsel__intval(evsel, sample, "pid"); 2430 2431 thread = machine__findnew_thread(machine, 0, pid); 2432 if (thread == NULL) 2433 return -1; 2434 2435 tr = thread__get_runtime(thread); 2436 if (tr == NULL) 2437 return -1; 2438 2439 tr->migrations++; 2440 2441 /* show migrations if requested */ 2442 timehist_print_migration_event(sched, evsel, sample, machine, thread); 2443 2444 return 0; 2445 } 2446 2447 static int timehist_sched_change_event(struct perf_tool *tool, 2448 union perf_event *event, 2449 struct perf_evsel *evsel, 2450 struct perf_sample *sample, 2451 struct machine *machine) 2452 { 2453 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2454 struct perf_time_interval *ptime = &sched->ptime; 2455 struct addr_location al; 2456 struct thread *thread; 2457 struct thread_runtime *tr = NULL; 2458 u64 tprev, t = sample->time; 2459 int rc = 0; 2460 int state = perf_evsel__intval(evsel, sample, "prev_state"); 2461 2462 2463 if (machine__resolve(machine, &al, sample) < 0) { 2464 pr_err("problem processing %d event. skipping it\n", 2465 event->header.type); 2466 rc = -1; 2467 goto out; 2468 } 2469 2470 thread = timehist_get_thread(sched, sample, machine, evsel); 2471 if (thread == NULL) { 2472 rc = -1; 2473 goto out; 2474 } 2475 2476 if (timehist_skip_sample(sched, thread, evsel, sample)) 2477 goto out; 2478 2479 tr = thread__get_runtime(thread); 2480 if (tr == NULL) { 2481 rc = -1; 2482 goto out; 2483 } 2484 2485 tprev = perf_evsel__get_time(evsel, sample->cpu); 2486 2487 /* 2488 * If start time given: 2489 * - sample time is under window user cares about - skip sample 2490 * - tprev is under window user cares about - reset to start of window 2491 */ 2492 if (ptime->start && ptime->start > t) 2493 goto out; 2494 2495 if (tprev && ptime->start > tprev) 2496 tprev = ptime->start; 2497 2498 /* 2499 * If end time given: 2500 * - previous sched event is out of window - we are done 2501 * - sample time is beyond window user cares about - reset it 2502 * to close out stats for time window interest 2503 */ 2504 if (ptime->end) { 2505 if (tprev > ptime->end) 2506 goto out; 2507 2508 if (t > ptime->end) 2509 t = ptime->end; 2510 } 2511 2512 if (!sched->idle_hist || thread->tid == 0) { 2513 timehist_update_runtime_stats(tr, t, tprev); 2514 2515 if (sched->idle_hist) { 2516 struct idle_thread_runtime *itr = (void *)tr; 2517 struct thread_runtime *last_tr; 2518 2519 BUG_ON(thread->tid != 0); 2520 2521 if (itr->last_thread == NULL) 2522 goto out; 2523 2524 /* add current idle time as last thread's runtime */ 2525 last_tr = thread__get_runtime(itr->last_thread); 2526 if (last_tr == NULL) 2527 goto out; 2528 2529 timehist_update_runtime_stats(last_tr, t, tprev); 2530 /* 2531 * remove delta time of last thread as it's not updated 2532 * and otherwise it will show an invalid value next 2533 * time. we only care total run time and run stat. 2534 */ 2535 last_tr->dt_run = 0; 2536 last_tr->dt_delay = 0; 2537 last_tr->dt_sleep = 0; 2538 last_tr->dt_iowait = 0; 2539 last_tr->dt_preempt = 0; 2540 2541 if (itr->cursor.nr) 2542 callchain_append(&itr->callchain, &itr->cursor, t - tprev); 2543 2544 itr->last_thread = NULL; 2545 } 2546 } 2547 2548 if (!sched->summary_only) 2549 timehist_print_sample(sched, evsel, sample, &al, thread, t, state); 2550 2551 out: 2552 if (sched->hist_time.start == 0 && t >= ptime->start) 2553 sched->hist_time.start = t; 2554 if (ptime->end == 0 || t <= ptime->end) 2555 sched->hist_time.end = t; 2556 2557 if (tr) { 2558 /* time of this sched_switch event becomes last time task seen */ 2559 tr->last_time = sample->time; 2560 2561 /* last state is used to determine where to account wait time */ 2562 tr->last_state = state; 2563 2564 /* sched out event for task so reset ready to run time */ 2565 tr->ready_to_run = 0; 2566 } 2567 2568 perf_evsel__save_time(evsel, sample->time, sample->cpu); 2569 2570 return rc; 2571 } 2572 2573 static int timehist_sched_switch_event(struct perf_tool *tool, 2574 union perf_event *event, 2575 struct perf_evsel *evsel, 2576 struct perf_sample *sample, 2577 struct machine *machine __maybe_unused) 2578 { 2579 return timehist_sched_change_event(tool, event, evsel, sample, machine); 2580 } 2581 2582 static int process_lost(struct perf_tool *tool __maybe_unused, 2583 union perf_event *event, 2584 struct perf_sample *sample, 2585 struct machine *machine __maybe_unused) 2586 { 2587 char tstr[64]; 2588 2589 timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr)); 2590 printf("%15s ", tstr); 2591 printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu); 2592 2593 return 0; 2594 } 2595 2596 2597 static void print_thread_runtime(struct thread *t, 2598 struct thread_runtime *r) 2599 { 2600 double mean = avg_stats(&r->run_stats); 2601 float stddev; 2602 2603 printf("%*s %5d %9" PRIu64 " ", 2604 comm_width, timehist_get_commstr(t), t->ppid, 2605 (u64) r->run_stats.n); 2606 2607 print_sched_time(r->total_run_time, 8); 2608 stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean); 2609 print_sched_time(r->run_stats.min, 6); 2610 printf(" "); 2611 print_sched_time((u64) mean, 6); 2612 printf(" "); 2613 print_sched_time(r->run_stats.max, 6); 2614 printf(" "); 2615 printf("%5.2f", stddev); 2616 printf(" %5" PRIu64, r->migrations); 2617 printf("\n"); 2618 } 2619 2620 static void print_thread_waittime(struct thread *t, 2621 struct thread_runtime *r) 2622 { 2623 printf("%*s %5d %9" PRIu64 " ", 2624 comm_width, timehist_get_commstr(t), t->ppid, 2625 (u64) r->run_stats.n); 2626 2627 print_sched_time(r->total_run_time, 8); 2628 print_sched_time(r->total_sleep_time, 6); 2629 printf(" "); 2630 print_sched_time(r->total_iowait_time, 6); 2631 printf(" "); 2632 print_sched_time(r->total_preempt_time, 6); 2633 printf(" "); 2634 print_sched_time(r->total_delay_time, 6); 2635 printf("\n"); 2636 } 2637 2638 struct total_run_stats { 2639 struct perf_sched *sched; 2640 u64 sched_count; 2641 u64 task_count; 2642 u64 total_run_time; 2643 }; 2644 2645 static int __show_thread_runtime(struct thread *t, void *priv) 2646 { 2647 struct total_run_stats *stats = priv; 2648 struct thread_runtime *r; 2649 2650 if (thread__is_filtered(t)) 2651 return 0; 2652 2653 r = thread__priv(t); 2654 if (r && r->run_stats.n) { 2655 stats->task_count++; 2656 stats->sched_count += r->run_stats.n; 2657 stats->total_run_time += r->total_run_time; 2658 2659 if (stats->sched->show_state) 2660 print_thread_waittime(t, r); 2661 else 2662 print_thread_runtime(t, r); 2663 } 2664 2665 return 0; 2666 } 2667 2668 static int show_thread_runtime(struct thread *t, void *priv) 2669 { 2670 if (t->dead) 2671 return 0; 2672 2673 return __show_thread_runtime(t, priv); 2674 } 2675 2676 static int show_deadthread_runtime(struct thread *t, void *priv) 2677 { 2678 if (!t->dead) 2679 return 0; 2680 2681 return __show_thread_runtime(t, priv); 2682 } 2683 2684 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node) 2685 { 2686 const char *sep = " <- "; 2687 struct callchain_list *chain; 2688 size_t ret = 0; 2689 char bf[1024]; 2690 bool first; 2691 2692 if (node == NULL) 2693 return 0; 2694 2695 ret = callchain__fprintf_folded(fp, node->parent); 2696 first = (ret == 0); 2697 2698 list_for_each_entry(chain, &node->val, list) { 2699 if (chain->ip >= PERF_CONTEXT_MAX) 2700 continue; 2701 if (chain->ms.sym && chain->ms.sym->ignore) 2702 continue; 2703 ret += fprintf(fp, "%s%s", first ? "" : sep, 2704 callchain_list__sym_name(chain, bf, sizeof(bf), 2705 false)); 2706 first = false; 2707 } 2708 2709 return ret; 2710 } 2711 2712 static size_t timehist_print_idlehist_callchain(struct rb_root *root) 2713 { 2714 size_t ret = 0; 2715 FILE *fp = stdout; 2716 struct callchain_node *chain; 2717 struct rb_node *rb_node = rb_first(root); 2718 2719 printf(" %16s %8s %s\n", "Idle time (msec)", "Count", "Callchains"); 2720 printf(" %.16s %.8s %.50s\n", graph_dotted_line, graph_dotted_line, 2721 graph_dotted_line); 2722 2723 while (rb_node) { 2724 chain = rb_entry(rb_node, struct callchain_node, rb_node); 2725 rb_node = rb_next(rb_node); 2726 2727 ret += fprintf(fp, " "); 2728 print_sched_time(chain->hit, 12); 2729 ret += 16; /* print_sched_time returns 2nd arg + 4 */ 2730 ret += fprintf(fp, " %8d ", chain->count); 2731 ret += callchain__fprintf_folded(fp, chain); 2732 ret += fprintf(fp, "\n"); 2733 } 2734 2735 return ret; 2736 } 2737 2738 static void timehist_print_summary(struct perf_sched *sched, 2739 struct perf_session *session) 2740 { 2741 struct machine *m = &session->machines.host; 2742 struct total_run_stats totals; 2743 u64 task_count; 2744 struct thread *t; 2745 struct thread_runtime *r; 2746 int i; 2747 u64 hist_time = sched->hist_time.end - sched->hist_time.start; 2748 2749 memset(&totals, 0, sizeof(totals)); 2750 totals.sched = sched; 2751 2752 if (sched->idle_hist) { 2753 printf("\nIdle-time summary\n"); 2754 printf("%*s parent sched-out ", comm_width, "comm"); 2755 printf(" idle-time min-idle avg-idle max-idle stddev migrations\n"); 2756 } else if (sched->show_state) { 2757 printf("\nWait-time summary\n"); 2758 printf("%*s parent sched-in ", comm_width, "comm"); 2759 printf(" run-time sleep iowait preempt delay\n"); 2760 } else { 2761 printf("\nRuntime summary\n"); 2762 printf("%*s parent sched-in ", comm_width, "comm"); 2763 printf(" run-time min-run avg-run max-run stddev migrations\n"); 2764 } 2765 printf("%*s (count) ", comm_width, ""); 2766 printf(" (msec) (msec) (msec) (msec) %s\n", 2767 sched->show_state ? "(msec)" : "%"); 2768 printf("%.117s\n", graph_dotted_line); 2769 2770 machine__for_each_thread(m, show_thread_runtime, &totals); 2771 task_count = totals.task_count; 2772 if (!task_count) 2773 printf("<no still running tasks>\n"); 2774 2775 printf("\nTerminated tasks:\n"); 2776 machine__for_each_thread(m, show_deadthread_runtime, &totals); 2777 if (task_count == totals.task_count) 2778 printf("<no terminated tasks>\n"); 2779 2780 /* CPU idle stats not tracked when samples were skipped */ 2781 if (sched->skipped_samples && !sched->idle_hist) 2782 return; 2783 2784 printf("\nIdle stats:\n"); 2785 for (i = 0; i < idle_max_cpu; ++i) { 2786 t = idle_threads[i]; 2787 if (!t) 2788 continue; 2789 2790 r = thread__priv(t); 2791 if (r && r->run_stats.n) { 2792 totals.sched_count += r->run_stats.n; 2793 printf(" CPU %2d idle for ", i); 2794 print_sched_time(r->total_run_time, 6); 2795 printf(" msec (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time); 2796 } else 2797 printf(" CPU %2d idle entire time window\n", i); 2798 } 2799 2800 if (sched->idle_hist && symbol_conf.use_callchain) { 2801 callchain_param.mode = CHAIN_FOLDED; 2802 callchain_param.value = CCVAL_PERIOD; 2803 2804 callchain_register_param(&callchain_param); 2805 2806 printf("\nIdle stats by callchain:\n"); 2807 for (i = 0; i < idle_max_cpu; ++i) { 2808 struct idle_thread_runtime *itr; 2809 2810 t = idle_threads[i]; 2811 if (!t) 2812 continue; 2813 2814 itr = thread__priv(t); 2815 if (itr == NULL) 2816 continue; 2817 2818 callchain_param.sort(&itr->sorted_root, &itr->callchain, 2819 0, &callchain_param); 2820 2821 printf(" CPU %2d:", i); 2822 print_sched_time(itr->tr.total_run_time, 6); 2823 printf(" msec\n"); 2824 timehist_print_idlehist_callchain(&itr->sorted_root); 2825 printf("\n"); 2826 } 2827 } 2828 2829 printf("\n" 2830 " Total number of unique tasks: %" PRIu64 "\n" 2831 "Total number of context switches: %" PRIu64 "\n", 2832 totals.task_count, totals.sched_count); 2833 2834 printf(" Total run time (msec): "); 2835 print_sched_time(totals.total_run_time, 2); 2836 printf("\n"); 2837 2838 printf(" Total scheduling time (msec): "); 2839 print_sched_time(hist_time, 2); 2840 printf(" (x %d)\n", sched->max_cpu); 2841 } 2842 2843 typedef int (*sched_handler)(struct perf_tool *tool, 2844 union perf_event *event, 2845 struct perf_evsel *evsel, 2846 struct perf_sample *sample, 2847 struct machine *machine); 2848 2849 static int perf_timehist__process_sample(struct perf_tool *tool, 2850 union perf_event *event, 2851 struct perf_sample *sample, 2852 struct perf_evsel *evsel, 2853 struct machine *machine) 2854 { 2855 struct perf_sched *sched = container_of(tool, struct perf_sched, tool); 2856 int err = 0; 2857 int this_cpu = sample->cpu; 2858 2859 if (this_cpu > sched->max_cpu) 2860 sched->max_cpu = this_cpu; 2861 2862 if (evsel->handler != NULL) { 2863 sched_handler f = evsel->handler; 2864 2865 err = f(tool, event, evsel, sample, machine); 2866 } 2867 2868 return err; 2869 } 2870 2871 static int timehist_check_attr(struct perf_sched *sched, 2872 struct perf_evlist *evlist) 2873 { 2874 struct perf_evsel *evsel; 2875 struct evsel_runtime *er; 2876 2877 list_for_each_entry(evsel, &evlist->entries, node) { 2878 er = perf_evsel__get_runtime(evsel); 2879 if (er == NULL) { 2880 pr_err("Failed to allocate memory for evsel runtime data\n"); 2881 return -1; 2882 } 2883 2884 if (sched->show_callchain && 2885 !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) { 2886 pr_info("Samples do not have callchains.\n"); 2887 sched->show_callchain = 0; 2888 symbol_conf.use_callchain = 0; 2889 } 2890 } 2891 2892 return 0; 2893 } 2894 2895 static int perf_sched__timehist(struct perf_sched *sched) 2896 { 2897 const struct perf_evsel_str_handler handlers[] = { 2898 { "sched:sched_switch", timehist_sched_switch_event, }, 2899 { "sched:sched_wakeup", timehist_sched_wakeup_event, }, 2900 { "sched:sched_wakeup_new", timehist_sched_wakeup_event, }, 2901 }; 2902 const struct perf_evsel_str_handler migrate_handlers[] = { 2903 { "sched:sched_migrate_task", timehist_migrate_task_event, }, 2904 }; 2905 struct perf_data_file file = { 2906 .path = input_name, 2907 .mode = PERF_DATA_MODE_READ, 2908 .force = sched->force, 2909 }; 2910 2911 struct perf_session *session; 2912 struct perf_evlist *evlist; 2913 int err = -1; 2914 2915 /* 2916 * event handlers for timehist option 2917 */ 2918 sched->tool.sample = perf_timehist__process_sample; 2919 sched->tool.mmap = perf_event__process_mmap; 2920 sched->tool.comm = perf_event__process_comm; 2921 sched->tool.exit = perf_event__process_exit; 2922 sched->tool.fork = perf_event__process_fork; 2923 sched->tool.lost = process_lost; 2924 sched->tool.attr = perf_event__process_attr; 2925 sched->tool.tracing_data = perf_event__process_tracing_data; 2926 sched->tool.build_id = perf_event__process_build_id; 2927 2928 sched->tool.ordered_events = true; 2929 sched->tool.ordering_requires_timestamps = true; 2930 2931 symbol_conf.use_callchain = sched->show_callchain; 2932 2933 session = perf_session__new(&file, false, &sched->tool); 2934 if (session == NULL) 2935 return -ENOMEM; 2936 2937 evlist = session->evlist; 2938 2939 symbol__init(&session->header.env); 2940 2941 if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) { 2942 pr_err("Invalid time string\n"); 2943 return -EINVAL; 2944 } 2945 2946 if (timehist_check_attr(sched, evlist) != 0) 2947 goto out; 2948 2949 setup_pager(); 2950 2951 /* setup per-evsel handlers */ 2952 if (perf_session__set_tracepoints_handlers(session, handlers)) 2953 goto out; 2954 2955 /* sched_switch event at a minimum needs to exist */ 2956 if (!perf_evlist__find_tracepoint_by_name(session->evlist, 2957 "sched:sched_switch")) { 2958 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n"); 2959 goto out; 2960 } 2961 2962 if (sched->show_migrations && 2963 perf_session__set_tracepoints_handlers(session, migrate_handlers)) 2964 goto out; 2965 2966 /* pre-allocate struct for per-CPU idle stats */ 2967 sched->max_cpu = session->header.env.nr_cpus_online; 2968 if (sched->max_cpu == 0) 2969 sched->max_cpu = 4; 2970 if (init_idle_threads(sched->max_cpu)) 2971 goto out; 2972 2973 /* summary_only implies summary option, but don't overwrite summary if set */ 2974 if (sched->summary_only) 2975 sched->summary = sched->summary_only; 2976 2977 if (!sched->summary_only) 2978 timehist_header(sched); 2979 2980 err = perf_session__process_events(session); 2981 if (err) { 2982 pr_err("Failed to process events, error %d", err); 2983 goto out; 2984 } 2985 2986 sched->nr_events = evlist->stats.nr_events[0]; 2987 sched->nr_lost_events = evlist->stats.total_lost; 2988 sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST]; 2989 2990 if (sched->summary) 2991 timehist_print_summary(sched, session); 2992 2993 out: 2994 free_idle_threads(); 2995 perf_session__delete(session); 2996 2997 return err; 2998 } 2999 3000 3001 static void print_bad_events(struct perf_sched *sched) 3002 { 3003 if (sched->nr_unordered_timestamps && sched->nr_timestamps) { 3004 printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n", 3005 (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0, 3006 sched->nr_unordered_timestamps, sched->nr_timestamps); 3007 } 3008 if (sched->nr_lost_events && sched->nr_events) { 3009 printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n", 3010 (double)sched->nr_lost_events/(double)sched->nr_events * 100.0, 3011 sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks); 3012 } 3013 if (sched->nr_context_switch_bugs && sched->nr_timestamps) { 3014 printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)", 3015 (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0, 3016 sched->nr_context_switch_bugs, sched->nr_timestamps); 3017 if (sched->nr_lost_events) 3018 printf(" (due to lost events?)"); 3019 printf("\n"); 3020 } 3021 } 3022 3023 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data) 3024 { 3025 struct rb_node **new = &(root->rb_node), *parent = NULL; 3026 struct work_atoms *this; 3027 const char *comm = thread__comm_str(data->thread), *this_comm; 3028 3029 while (*new) { 3030 int cmp; 3031 3032 this = container_of(*new, struct work_atoms, node); 3033 parent = *new; 3034 3035 this_comm = thread__comm_str(this->thread); 3036 cmp = strcmp(comm, this_comm); 3037 if (cmp > 0) { 3038 new = &((*new)->rb_left); 3039 } else if (cmp < 0) { 3040 new = &((*new)->rb_right); 3041 } else { 3042 this->num_merged++; 3043 this->total_runtime += data->total_runtime; 3044 this->nb_atoms += data->nb_atoms; 3045 this->total_lat += data->total_lat; 3046 list_splice(&data->work_list, &this->work_list); 3047 if (this->max_lat < data->max_lat) { 3048 this->max_lat = data->max_lat; 3049 this->max_lat_at = data->max_lat_at; 3050 } 3051 zfree(&data); 3052 return; 3053 } 3054 } 3055 3056 data->num_merged++; 3057 rb_link_node(&data->node, parent, new); 3058 rb_insert_color(&data->node, root); 3059 } 3060 3061 static void perf_sched__merge_lat(struct perf_sched *sched) 3062 { 3063 struct work_atoms *data; 3064 struct rb_node *node; 3065 3066 if (sched->skip_merge) 3067 return; 3068 3069 while ((node = rb_first(&sched->atom_root))) { 3070 rb_erase(node, &sched->atom_root); 3071 data = rb_entry(node, struct work_atoms, node); 3072 __merge_work_atoms(&sched->merged_atom_root, data); 3073 } 3074 } 3075 3076 static int perf_sched__lat(struct perf_sched *sched) 3077 { 3078 struct rb_node *next; 3079 3080 setup_pager(); 3081 3082 if (perf_sched__read_events(sched)) 3083 return -1; 3084 3085 perf_sched__merge_lat(sched); 3086 perf_sched__sort_lat(sched); 3087 3088 printf("\n -----------------------------------------------------------------------------------------------------------------\n"); 3089 printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n"); 3090 printf(" -----------------------------------------------------------------------------------------------------------------\n"); 3091 3092 next = rb_first(&sched->sorted_atom_root); 3093 3094 while (next) { 3095 struct work_atoms *work_list; 3096 3097 work_list = rb_entry(next, struct work_atoms, node); 3098 output_lat_thread(sched, work_list); 3099 next = rb_next(next); 3100 thread__zput(work_list->thread); 3101 } 3102 3103 printf(" -----------------------------------------------------------------------------------------------------------------\n"); 3104 printf(" TOTAL: |%11.3f ms |%9" PRIu64 " |\n", 3105 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count); 3106 3107 printf(" ---------------------------------------------------\n"); 3108 3109 print_bad_events(sched); 3110 printf("\n"); 3111 3112 return 0; 3113 } 3114 3115 static int setup_map_cpus(struct perf_sched *sched) 3116 { 3117 struct cpu_map *map; 3118 3119 sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF); 3120 3121 if (sched->map.comp) { 3122 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int)); 3123 if (!sched->map.comp_cpus) 3124 return -1; 3125 } 3126 3127 if (!sched->map.cpus_str) 3128 return 0; 3129 3130 map = cpu_map__new(sched->map.cpus_str); 3131 if (!map) { 3132 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str); 3133 return -1; 3134 } 3135 3136 sched->map.cpus = map; 3137 return 0; 3138 } 3139 3140 static int setup_color_pids(struct perf_sched *sched) 3141 { 3142 struct thread_map *map; 3143 3144 if (!sched->map.color_pids_str) 3145 return 0; 3146 3147 map = thread_map__new_by_tid_str(sched->map.color_pids_str); 3148 if (!map) { 3149 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str); 3150 return -1; 3151 } 3152 3153 sched->map.color_pids = map; 3154 return 0; 3155 } 3156 3157 static int setup_color_cpus(struct perf_sched *sched) 3158 { 3159 struct cpu_map *map; 3160 3161 if (!sched->map.color_cpus_str) 3162 return 0; 3163 3164 map = cpu_map__new(sched->map.color_cpus_str); 3165 if (!map) { 3166 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str); 3167 return -1; 3168 } 3169 3170 sched->map.color_cpus = map; 3171 return 0; 3172 } 3173 3174 static int perf_sched__map(struct perf_sched *sched) 3175 { 3176 if (setup_map_cpus(sched)) 3177 return -1; 3178 3179 if (setup_color_pids(sched)) 3180 return -1; 3181 3182 if (setup_color_cpus(sched)) 3183 return -1; 3184 3185 setup_pager(); 3186 if (perf_sched__read_events(sched)) 3187 return -1; 3188 print_bad_events(sched); 3189 return 0; 3190 } 3191 3192 static int perf_sched__replay(struct perf_sched *sched) 3193 { 3194 unsigned long i; 3195 3196 calibrate_run_measurement_overhead(sched); 3197 calibrate_sleep_measurement_overhead(sched); 3198 3199 test_calibrations(sched); 3200 3201 if (perf_sched__read_events(sched)) 3202 return -1; 3203 3204 printf("nr_run_events: %ld\n", sched->nr_run_events); 3205 printf("nr_sleep_events: %ld\n", sched->nr_sleep_events); 3206 printf("nr_wakeup_events: %ld\n", sched->nr_wakeup_events); 3207 3208 if (sched->targetless_wakeups) 3209 printf("target-less wakeups: %ld\n", sched->targetless_wakeups); 3210 if (sched->multitarget_wakeups) 3211 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups); 3212 if (sched->nr_run_events_optimized) 3213 printf("run atoms optimized: %ld\n", 3214 sched->nr_run_events_optimized); 3215 3216 print_task_traces(sched); 3217 add_cross_task_wakeups(sched); 3218 3219 create_tasks(sched); 3220 printf("------------------------------------------------------------\n"); 3221 for (i = 0; i < sched->replay_repeat; i++) 3222 run_one_test(sched); 3223 3224 return 0; 3225 } 3226 3227 static void setup_sorting(struct perf_sched *sched, const struct option *options, 3228 const char * const usage_msg[]) 3229 { 3230 char *tmp, *tok, *str = strdup(sched->sort_order); 3231 3232 for (tok = strtok_r(str, ", ", &tmp); 3233 tok; tok = strtok_r(NULL, ", ", &tmp)) { 3234 if (sort_dimension__add(tok, &sched->sort_list) < 0) { 3235 usage_with_options_msg(usage_msg, options, 3236 "Unknown --sort key: `%s'", tok); 3237 } 3238 } 3239 3240 free(str); 3241 3242 sort_dimension__add("pid", &sched->cmp_pid); 3243 } 3244 3245 static int __cmd_record(int argc, const char **argv) 3246 { 3247 unsigned int rec_argc, i, j; 3248 const char **rec_argv; 3249 const char * const record_args[] = { 3250 "record", 3251 "-a", 3252 "-R", 3253 "-m", "1024", 3254 "-c", "1", 3255 "-e", "sched:sched_switch", 3256 "-e", "sched:sched_stat_wait", 3257 "-e", "sched:sched_stat_sleep", 3258 "-e", "sched:sched_stat_iowait", 3259 "-e", "sched:sched_stat_runtime", 3260 "-e", "sched:sched_process_fork", 3261 "-e", "sched:sched_wakeup", 3262 "-e", "sched:sched_wakeup_new", 3263 "-e", "sched:sched_migrate_task", 3264 }; 3265 3266 rec_argc = ARRAY_SIZE(record_args) + argc - 1; 3267 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 3268 3269 if (rec_argv == NULL) 3270 return -ENOMEM; 3271 3272 for (i = 0; i < ARRAY_SIZE(record_args); i++) 3273 rec_argv[i] = strdup(record_args[i]); 3274 3275 for (j = 1; j < (unsigned int)argc; j++, i++) 3276 rec_argv[i] = argv[j]; 3277 3278 BUG_ON(i != rec_argc); 3279 3280 return cmd_record(i, rec_argv); 3281 } 3282 3283 int cmd_sched(int argc, const char **argv) 3284 { 3285 const char default_sort_order[] = "avg, max, switch, runtime"; 3286 struct perf_sched sched = { 3287 .tool = { 3288 .sample = perf_sched__process_tracepoint_sample, 3289 .comm = perf_event__process_comm, 3290 .namespaces = perf_event__process_namespaces, 3291 .lost = perf_event__process_lost, 3292 .fork = perf_sched__process_fork_event, 3293 .ordered_events = true, 3294 }, 3295 .cmp_pid = LIST_HEAD_INIT(sched.cmp_pid), 3296 .sort_list = LIST_HEAD_INIT(sched.sort_list), 3297 .start_work_mutex = PTHREAD_MUTEX_INITIALIZER, 3298 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER, 3299 .sort_order = default_sort_order, 3300 .replay_repeat = 10, 3301 .profile_cpu = -1, 3302 .next_shortname1 = 'A', 3303 .next_shortname2 = '0', 3304 .skip_merge = 0, 3305 .show_callchain = 1, 3306 .max_stack = 5, 3307 }; 3308 const struct option sched_options[] = { 3309 OPT_STRING('i', "input", &input_name, "file", 3310 "input file name"), 3311 OPT_INCR('v', "verbose", &verbose, 3312 "be more verbose (show symbol address, etc)"), 3313 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace, 3314 "dump raw trace in ASCII"), 3315 OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"), 3316 OPT_END() 3317 }; 3318 const struct option latency_options[] = { 3319 OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]", 3320 "sort by key(s): runtime, switch, avg, max"), 3321 OPT_INTEGER('C', "CPU", &sched.profile_cpu, 3322 "CPU to profile on"), 3323 OPT_BOOLEAN('p', "pids", &sched.skip_merge, 3324 "latency stats per pid instead of per comm"), 3325 OPT_PARENT(sched_options) 3326 }; 3327 const struct option replay_options[] = { 3328 OPT_UINTEGER('r', "repeat", &sched.replay_repeat, 3329 "repeat the workload replay N times (-1: infinite)"), 3330 OPT_PARENT(sched_options) 3331 }; 3332 const struct option map_options[] = { 3333 OPT_BOOLEAN(0, "compact", &sched.map.comp, 3334 "map output in compact mode"), 3335 OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids", 3336 "highlight given pids in map"), 3337 OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus", 3338 "highlight given CPUs in map"), 3339 OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus", 3340 "display given CPUs in map"), 3341 OPT_PARENT(sched_options) 3342 }; 3343 const struct option timehist_options[] = { 3344 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name, 3345 "file", "vmlinux pathname"), 3346 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name, 3347 "file", "kallsyms pathname"), 3348 OPT_BOOLEAN('g', "call-graph", &sched.show_callchain, 3349 "Display call chains if present (default on)"), 3350 OPT_UINTEGER(0, "max-stack", &sched.max_stack, 3351 "Maximum number of functions to display backtrace."), 3352 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory", 3353 "Look for files with symbols relative to this directory"), 3354 OPT_BOOLEAN('s', "summary", &sched.summary_only, 3355 "Show only syscall summary with statistics"), 3356 OPT_BOOLEAN('S', "with-summary", &sched.summary, 3357 "Show all syscalls and summary with statistics"), 3358 OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"), 3359 OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"), 3360 OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"), 3361 OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"), 3362 OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"), 3363 OPT_STRING(0, "time", &sched.time_str, "str", 3364 "Time span for analysis (start,stop)"), 3365 OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"), 3366 OPT_PARENT(sched_options) 3367 }; 3368 3369 const char * const latency_usage[] = { 3370 "perf sched latency [<options>]", 3371 NULL 3372 }; 3373 const char * const replay_usage[] = { 3374 "perf sched replay [<options>]", 3375 NULL 3376 }; 3377 const char * const map_usage[] = { 3378 "perf sched map [<options>]", 3379 NULL 3380 }; 3381 const char * const timehist_usage[] = { 3382 "perf sched timehist [<options>]", 3383 NULL 3384 }; 3385 const char *const sched_subcommands[] = { "record", "latency", "map", 3386 "replay", "script", 3387 "timehist", NULL }; 3388 const char *sched_usage[] = { 3389 NULL, 3390 NULL 3391 }; 3392 struct trace_sched_handler lat_ops = { 3393 .wakeup_event = latency_wakeup_event, 3394 .switch_event = latency_switch_event, 3395 .runtime_event = latency_runtime_event, 3396 .migrate_task_event = latency_migrate_task_event, 3397 }; 3398 struct trace_sched_handler map_ops = { 3399 .switch_event = map_switch_event, 3400 }; 3401 struct trace_sched_handler replay_ops = { 3402 .wakeup_event = replay_wakeup_event, 3403 .switch_event = replay_switch_event, 3404 .fork_event = replay_fork_event, 3405 }; 3406 unsigned int i; 3407 3408 for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++) 3409 sched.curr_pid[i] = -1; 3410 3411 argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands, 3412 sched_usage, PARSE_OPT_STOP_AT_NON_OPTION); 3413 if (!argc) 3414 usage_with_options(sched_usage, sched_options); 3415 3416 /* 3417 * Aliased to 'perf script' for now: 3418 */ 3419 if (!strcmp(argv[0], "script")) 3420 return cmd_script(argc, argv); 3421 3422 if (!strncmp(argv[0], "rec", 3)) { 3423 return __cmd_record(argc, argv); 3424 } else if (!strncmp(argv[0], "lat", 3)) { 3425 sched.tp_handler = &lat_ops; 3426 if (argc > 1) { 3427 argc = parse_options(argc, argv, latency_options, latency_usage, 0); 3428 if (argc) 3429 usage_with_options(latency_usage, latency_options); 3430 } 3431 setup_sorting(&sched, latency_options, latency_usage); 3432 return perf_sched__lat(&sched); 3433 } else if (!strcmp(argv[0], "map")) { 3434 if (argc) { 3435 argc = parse_options(argc, argv, map_options, map_usage, 0); 3436 if (argc) 3437 usage_with_options(map_usage, map_options); 3438 } 3439 sched.tp_handler = &map_ops; 3440 setup_sorting(&sched, latency_options, latency_usage); 3441 return perf_sched__map(&sched); 3442 } else if (!strncmp(argv[0], "rep", 3)) { 3443 sched.tp_handler = &replay_ops; 3444 if (argc) { 3445 argc = parse_options(argc, argv, replay_options, replay_usage, 0); 3446 if (argc) 3447 usage_with_options(replay_usage, replay_options); 3448 } 3449 return perf_sched__replay(&sched); 3450 } else if (!strcmp(argv[0], "timehist")) { 3451 if (argc) { 3452 argc = parse_options(argc, argv, timehist_options, 3453 timehist_usage, 0); 3454 if (argc) 3455 usage_with_options(timehist_usage, timehist_options); 3456 } 3457 if ((sched.show_wakeups || sched.show_next) && 3458 sched.summary_only) { 3459 pr_err(" Error: -s and -[n|w] are mutually exclusive.\n"); 3460 parse_options_usage(timehist_usage, timehist_options, "s", true); 3461 if (sched.show_wakeups) 3462 parse_options_usage(NULL, timehist_options, "w", true); 3463 if (sched.show_next) 3464 parse_options_usage(NULL, timehist_options, "n", true); 3465 return -EINVAL; 3466 } 3467 3468 return perf_sched__timehist(&sched); 3469 } else { 3470 usage_with_options(sched_usage, sched_options); 3471 } 3472 3473 return 0; 3474 } 3475