tools/perf/builtin-sched.c

   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
  14 #include "util/parse-options.h"
  15 #include "util/trace-event.h"
  16
  17 #include "util/debug.h"
  18
  19 #include <sys/prctl.h>
  20 #include <sys/resource.h>
  21
  22 #include <semaphore.h>
  23 #include <pthread.h>
  24 #include <math.h>
  25
  26 #define PR_SET_NAME             15               /* Set process name */
  27 #define MAX_CPUS                4096
  28 #define COMM_LEN                20
  29 #define SYM_LEN                 129
  30 #define MAX_PID                 65536
  31
  32 struct sched_atom;
  33
  34 struct task_desc {
  35         unsigned long           nr;
  36         unsigned long           pid;
  37         char                    comm[COMM_LEN];
  38
  39         unsigned long           nr_events;
  40         unsigned long           curr_event;
  41         struct sched_atom       **atoms;
  42
  43         pthread_t               thread;
  44         sem_t                   sleep_sem;
  45
  46         sem_t                   ready_for_work;
  47         sem_t                   work_done_sem;
  48
  49         u64                     cpu_usage;
  50 };
  51
  52 enum sched_event_type {
  53         SCHED_EVENT_RUN,
  54         SCHED_EVENT_SLEEP,
  55         SCHED_EVENT_WAKEUP,
  56         SCHED_EVENT_MIGRATION,
  57 };
  58
  59 struct sched_atom {
  60         enum sched_event_type   type;
  61         int                     specific_wait;
  62         u64                     timestamp;
  63         u64                     duration;
  64         unsigned long           nr;
  65         sem_t                   *wait_sem;
  66         struct task_desc        *wakee;
  67 };
  68
  69 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  70
  71 enum thread_state {
  72         THREAD_SLEEPING = 0,
  73         THREAD_WAIT_CPU,
  74         THREAD_SCHED_IN,
  75         THREAD_IGNORE
  76 };
  77
  78 struct work_atom {
  79         struct list_head        list;
  80         enum thread_state       state;
  81         u64                     sched_out_time;
  82         u64                     wake_up_time;
  83         u64                     sched_in_time;
  84         u64                     runtime;
  85 };
  86
  87 struct work_atoms {
  88         struct list_head        work_list;
  89         struct thread           *thread;
  90         struct rb_node          node;
  91         u64                     max_lat;
  92         u64                     max_lat_at;
  93         u64                     total_lat;
  94         u64                     nb_atoms;
  95         u64                     total_runtime;
  96 };
  97
  98 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
  99
 100 struct perf_sched;
 101
 102 struct trace_sched_handler {
 103         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 104                             struct perf_sample *sample, struct machine *machine);
 105
 106         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 107                              struct perf_sample *sample, struct machine *machine);
 108
 109         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 110                             struct perf_sample *sample, struct machine *machine);
 111
 112         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 113         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 114                           struct machine *machine);
 115
 116         int (*migrate_task_event)(struct perf_sched *sched,
 117                                   struct perf_evsel *evsel,
 118                                   struct perf_sample *sample,
 119                                   struct machine *machine);
 120 };
 121
 122 struct perf_sched {
 123         struct perf_tool tool;
 124         const char       *sort_order;
 125         unsigned long    nr_tasks;
 126         struct task_desc *pid_to_task[MAX_PID];
 127         struct task_desc **tasks;
 128         const struct trace_sched_handler *tp_handler;
 129         pthread_mutex_t  start_work_mutex;
 130         pthread_mutex_t  work_done_wait_mutex;
 131         int              profile_cpu;
 132 /*
 133  * Track the current task - that way we can know whether there's any
 134  * weird events, such as a task being switched away that is not current.
 135  */
 136         int              max_cpu;
 137         u32              curr_pid[MAX_CPUS];
 138         struct thread    *curr_thread[MAX_CPUS];
 139         char             next_shortname1;
 140         char             next_shortname2;
 141         unsigned int     replay_repeat;
 142         unsigned long    nr_run_events;
 143         unsigned long    nr_sleep_events;
 144         unsigned long    nr_wakeup_events;
 145         unsigned long    nr_sleep_corrections;
 146         unsigned long    nr_run_events_optimized;
 147         unsigned long    targetless_wakeups;
 148         unsigned long    multitarget_wakeups;
 149         unsigned long    nr_runs;
 150         unsigned long    nr_timestamps;
 151         unsigned long    nr_unordered_timestamps;
 152         unsigned long    nr_context_switch_bugs;
 153         unsigned long    nr_events;
 154         unsigned long    nr_lost_chunks;
 155         unsigned long    nr_lost_events;
 156         u64              run_measurement_overhead;
 157         u64              sleep_measurement_overhead;
 158         u64              start_time;
 159         u64              cpu_usage;
 160         u64              runavg_cpu_usage;
 161         u64              parent_cpu_usage;
 162         u64              runavg_parent_cpu_usage;
 163         u64              sum_runtime;
 164         u64              sum_fluct;
 165         u64              run_avg;
 166         u64              all_runtime;
 167         u64              all_count;
 168         u64              cpu_last_switched[MAX_CPUS];
 169         struct rb_root   atom_root, sorted_atom_root;
 170         struct list_head sort_list, cmp_pid;
 171 };
 172
 173 static u64 get_nsecs(void)
 174 {
 175         struct timespec ts;
 176
 177         clock_gettime(CLOCK_MONOTONIC, &ts);
 178
 179         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 180 }
 181
 182 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 183 {
 184         u64 T0 = get_nsecs(), T1;
 185
 186         do {
 187                 T1 = get_nsecs();
 188         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 189 }
 190
 191 static void sleep_nsecs(u64 nsecs)
 192 {
 193         struct timespec ts;
 194
 195         ts.tv_nsec = nsecs % 999999999;
 196         ts.tv_sec = nsecs / 999999999;
 197
 198         nanosleep(&ts, NULL);
 199 }
 200
 201 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 202 {
 203         u64 T0, T1, delta, min_delta = 1000000000ULL;
 204         int i;
 205
 206         for (i = 0; i < 10; i++) {
 207                 T0 = get_nsecs();
 208                 burn_nsecs(sched, 0);
 209                 T1 = get_nsecs();
 210                 delta = T1-T0;
 211                 min_delta = min(min_delta, delta);
 212         }
 213         sched->run_measurement_overhead = min_delta;
 214
 215         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 216 }
 217
 218 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 219 {
 220         u64 T0, T1, delta, min_delta = 1000000000ULL;
 221         int i;
 222
 223         for (i = 0; i < 10; i++) {
 224                 T0 = get_nsecs();
 225                 sleep_nsecs(10000);
 226                 T1 = get_nsecs();
 227                 delta = T1-T0;
 228                 min_delta = min(min_delta, delta);
 229         }
 230         min_delta -= 10000;
 231         sched->sleep_measurement_overhead = min_delta;
 232
 233         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 234 }
 235
 236 static struct sched_atom *
 237 get_new_event(struct task_desc *task, u64 timestamp)
 238 {
 239         struct sched_atom *event = zalloc(sizeof(*event));
 240         unsigned long idx = task->nr_events;
 241         size_t size;
 242
 243         event->timestamp = timestamp;
 244         event->nr = idx;
 245
 246         task->nr_events++;
 247         size = sizeof(struct sched_atom *) * task->nr_events;
 248         task->atoms = realloc(task->atoms, size);
 249         BUG_ON(!task->atoms);
 250
 251         task->atoms[idx] = event;
 252
 253         return event;
 254 }
 255
 256 static struct sched_atom *last_event(struct task_desc *task)
 257 {
 258         if (!task->nr_events)
 259                 return NULL;
 260
 261         return task->atoms[task->nr_events - 1];
 262 }
 263
 264 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 265                                 u64 timestamp, u64 duration)
 266 {
 267         struct sched_atom *event, *curr_event = last_event(task);
 268
 269         /*
 270          * optimize an existing RUN event by merging this one
 271          * to it:
 272          */
 273         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 274                 sched->nr_run_events_optimized++;
 275                 curr_event->duration += duration;
 276                 return;
 277         }
 278
 279         event = get_new_event(task, timestamp);
 280
 281         event->type = SCHED_EVENT_RUN;
 282         event->duration = duration;
 283
 284         sched->nr_run_events++;
 285 }
 286
 287 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 288                                    u64 timestamp, struct task_desc *wakee)
 289 {
 290         struct sched_atom *event, *wakee_event;
 291
 292         event = get_new_event(task, timestamp);
 293         event->type = SCHED_EVENT_WAKEUP;
 294         event->wakee = wakee;
 295
 296         wakee_event = last_event(wakee);
 297         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 298                 sched->targetless_wakeups++;
 299                 return;
 300         }
 301         if (wakee_event->wait_sem) {
 302                 sched->multitarget_wakeups++;
 303                 return;
 304         }
 305
 306         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 307         sem_init(wakee_event->wait_sem, 0, 0);
 308         wakee_event->specific_wait = 1;
 309         event->wait_sem = wakee_event->wait_sem;
 310
 311         sched->nr_wakeup_events++;
 312 }
 313
 314 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 315                                   u64 timestamp, u64 task_state __maybe_unused)
 316 {
 317         struct sched_atom *event = get_new_event(task, timestamp);
 318
 319         event->type = SCHED_EVENT_SLEEP;
 320
 321         sched->nr_sleep_events++;
 322 }
 323
 324 static struct task_desc *register_pid(struct perf_sched *sched,
 325                                       unsigned long pid, const char *comm)
 326 {
 327         struct task_desc *task;
 328
 329         BUG_ON(pid >= MAX_PID);
 330
 331         task = sched->pid_to_task[pid];
 332
 333         if (task)
 334                 return task;
 335
 336         task = zalloc(sizeof(*task));
 337         task->pid = pid;
 338         task->nr = sched->nr_tasks;
 339         strcpy(task->comm, comm);
 340         /*
 341          * every task starts in sleeping state - this gets ignored
 342          * if there's no wakeup pointing to this sleep state:
 343          */
 344         add_sched_event_sleep(sched, task, 0, 0);
 345
 346         sched->pid_to_task[pid] = task;
 347         sched->nr_tasks++;
 348         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_task *));
 349         BUG_ON(!sched->tasks);
 350         sched->tasks[task->nr] = task;
 351
 352         if (verbose)
 353                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 354
 355         return task;
 356 }
 357
 358
 359 static void print_task_traces(struct perf_sched *sched)
 360 {
 361         struct task_desc *task;
 362         unsigned long i;
 363
 364         for (i = 0; i < sched->nr_tasks; i++) {
 365                 task = sched->tasks[i];
 366                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 367                         task->nr, task->comm, task->pid, task->nr_events);
 368         }
 369 }
 370
 371 static void add_cross_task_wakeups(struct perf_sched *sched)
 372 {
 373         struct task_desc *task1, *task2;
 374         unsigned long i, j;
 375
 376         for (i = 0; i < sched->nr_tasks; i++) {
 377                 task1 = sched->tasks[i];
 378                 j = i + 1;
 379                 if (j == sched->nr_tasks)
 380                         j = 0;
 381                 task2 = sched->tasks[j];
 382                 add_sched_event_wakeup(sched, task1, 0, task2);
 383         }
 384 }
 385
 386 static void perf_sched__process_event(struct perf_sched *sched,
 387                                       struct sched_atom *atom)
 388 {
 389         int ret = 0;
 390
 391         switch (atom->type) {
 392                 case SCHED_EVENT_RUN:
 393                         burn_nsecs(sched, atom->duration);
 394                         break;
 395                 case SCHED_EVENT_SLEEP:
 396                         if (atom->wait_sem)
 397                                 ret = sem_wait(atom->wait_sem);
 398                         BUG_ON(ret);
 399                         break;
 400                 case SCHED_EVENT_WAKEUP:
 401                         if (atom->wait_sem)
 402                                 ret = sem_post(atom->wait_sem);
 403                         BUG_ON(ret);
 404                         break;
 405                 case SCHED_EVENT_MIGRATION:
 406                         break;
 407                 default:
 408                         BUG_ON(1);
 409         }
 410 }
 411
 412 static u64 get_cpu_usage_nsec_parent(void)
 413 {
 414         struct rusage ru;
 415         u64 sum;
 416         int err;
 417
 418         err = getrusage(RUSAGE_SELF, &ru);
 419         BUG_ON(err);
 420
 421         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 422         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 423
 424         return sum;
 425 }
 426
 427 static int self_open_counters(void)
 428 {
 429         struct perf_event_attr attr;
 430         int fd;
 431
 432         memset(&attr, 0, sizeof(attr));
 433
 434         attr.type = PERF_TYPE_SOFTWARE;
 435         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 436
 437         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 438
 439         if (fd < 0)
 440                 pr_err("Error: sys_perf_event_open() syscall returned "
 441                        "with %d (%s)\n", fd, strerror(errno));
 442         return fd;
 443 }
 444
 445 static u64 get_cpu_usage_nsec_self(int fd)
 446 {
 447         u64 runtime;
 448         int ret;
 449
 450         ret = read(fd, &runtime, sizeof(runtime));
 451         BUG_ON(ret != sizeof(runtime));
 452
 453         return runtime;
 454 }
 455
 456 struct sched_thread_parms {
 457         struct task_desc  *task;
 458         struct perf_sched *sched;
 459 };
 460
 461 static void *thread_func(void *ctx)
 462 {
 463         struct sched_thread_parms *parms = ctx;
 464         struct task_desc *this_task = parms->task;
 465         struct perf_sched *sched = parms->sched;
 466         u64 cpu_usage_0, cpu_usage_1;
 467         unsigned long i, ret;
 468         char comm2[22];
 469         int fd;
 470
 471         zfree(&parms);
 472
 473         sprintf(comm2, ":%s", this_task->comm);
 474         prctl(PR_SET_NAME, comm2);
 475         fd = self_open_counters();
 476         if (fd < 0)
 477                 return NULL;
 478 again:
 479         ret = sem_post(&this_task->ready_for_work);
 480         BUG_ON(ret);
 481         ret = pthread_mutex_lock(&sched->start_work_mutex);
 482         BUG_ON(ret);
 483         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 484         BUG_ON(ret);
 485
 486         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 487
 488         for (i = 0; i < this_task->nr_events; i++) {
 489                 this_task->curr_event = i;
 490                 perf_sched__process_event(sched, this_task->atoms[i]);
 491         }
 492
 493         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 494         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 495         ret = sem_post(&this_task->work_done_sem);
 496         BUG_ON(ret);
 497
 498         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 499         BUG_ON(ret);
 500         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 501         BUG_ON(ret);
 502
 503         goto again;
 504 }
 505
 506 static void create_tasks(struct perf_sched *sched)
 507 {
 508         struct task_desc *task;
 509         pthread_attr_t attr;
 510         unsigned long i;
 511         int err;
 512
 513         err = pthread_attr_init(&attr);
 514         BUG_ON(err);
 515         err = pthread_attr_setstacksize(&attr,
 516                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 517         BUG_ON(err);
 518         err = pthread_mutex_lock(&sched->start_work_mutex);
 519         BUG_ON(err);
 520         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 521         BUG_ON(err);
 522         for (i = 0; i < sched->nr_tasks; i++) {
 523                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 524                 BUG_ON(parms == NULL);
 525                 parms->task = task = sched->tasks[i];
 526                 parms->sched = sched;
 527                 sem_init(&task->sleep_sem, 0, 0);
 528                 sem_init(&task->ready_for_work, 0, 0);
 529                 sem_init(&task->work_done_sem, 0, 0);
 530                 task->curr_event = 0;
 531                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 532                 BUG_ON(err);
 533         }
 534 }
 535
 536 static void wait_for_tasks(struct perf_sched *sched)
 537 {
 538         u64 cpu_usage_0, cpu_usage_1;
 539         struct task_desc *task;
 540         unsigned long i, ret;
 541
 542         sched->start_time = get_nsecs();
 543         sched->cpu_usage = 0;
 544         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 545
 546         for (i = 0; i < sched->nr_tasks; i++) {
 547                 task = sched->tasks[i];
 548                 ret = sem_wait(&task->ready_for_work);
 549                 BUG_ON(ret);
 550                 sem_init(&task->ready_for_work, 0, 0);
 551         }
 552         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 553         BUG_ON(ret);
 554
 555         cpu_usage_0 = get_cpu_usage_nsec_parent();
 556
 557         pthread_mutex_unlock(&sched->start_work_mutex);
 558
 559         for (i = 0; i < sched->nr_tasks; i++) {
 560                 task = sched->tasks[i];
 561                 ret = sem_wait(&task->work_done_sem);
 562                 BUG_ON(ret);
 563                 sem_init(&task->work_done_sem, 0, 0);
 564                 sched->cpu_usage += task->cpu_usage;
 565                 task->cpu_usage = 0;
 566         }
 567
 568         cpu_usage_1 = get_cpu_usage_nsec_parent();
 569         if (!sched->runavg_cpu_usage)
 570                 sched->runavg_cpu_usage = sched->cpu_usage;
 571         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
 572
 573         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 574         if (!sched->runavg_parent_cpu_usage)
 575                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 576         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
 577                                          sched->parent_cpu_usage)/10;
 578
 579         ret = pthread_mutex_lock(&sched->start_work_mutex);
 580         BUG_ON(ret);
 581
 582         for (i = 0; i < sched->nr_tasks; i++) {
 583                 task = sched->tasks[i];
 584                 sem_init(&task->sleep_sem, 0, 0);
 585                 task->curr_event = 0;
 586         }
 587 }
 588
 589 static void run_one_test(struct perf_sched *sched)
 590 {
 591         u64 T0, T1, delta, avg_delta, fluct;
 592
 593         T0 = get_nsecs();
 594         wait_for_tasks(sched);
 595         T1 = get_nsecs();
 596
 597         delta = T1 - T0;
 598         sched->sum_runtime += delta;
 599         sched->nr_runs++;
 600
 601         avg_delta = sched->sum_runtime / sched->nr_runs;
 602         if (delta < avg_delta)
 603                 fluct = avg_delta - delta;
 604         else
 605                 fluct = delta - avg_delta;
 606         sched->sum_fluct += fluct;
 607         if (!sched->run_avg)
 608                 sched->run_avg = delta;
 609         sched->run_avg = (sched->run_avg * 9 + delta) / 10;
 610
 611         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 612
 613         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 614
 615         printf("cpu: %0.2f / %0.2f",
 616                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 617
 618 #if 0
 619         /*
 620          * rusage statistics done by the parent, these are less
 621          * accurate than the sched->sum_exec_runtime based statistics:
 622          */
 623         printf(" [%0.2f / %0.2f]",
 624                 (double)sched->parent_cpu_usage/1e6,
 625                 (double)sched->runavg_parent_cpu_usage/1e6);
 626 #endif
 627
 628         printf("\n");
 629
 630         if (sched->nr_sleep_corrections)
 631                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 632         sched->nr_sleep_corrections = 0;
 633 }
 634
 635 static void test_calibrations(struct perf_sched *sched)
 636 {
 637         u64 T0, T1;
 638
 639         T0 = get_nsecs();
 640         burn_nsecs(sched, 1e6);
 641         T1 = get_nsecs();
 642
 643         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 644
 645         T0 = get_nsecs();
 646         sleep_nsecs(1e6);
 647         T1 = get_nsecs();
 648
 649         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 650 }
 651
 652 static int
 653 replay_wakeup_event(struct perf_sched *sched,
 654                     struct perf_evsel *evsel, struct perf_sample *sample,
 655                     struct machine *machine __maybe_unused)
 656 {
 657         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 658         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 659         struct task_desc *waker, *wakee;
 660
 661         if (verbose) {
 662                 printf("sched_wakeup event %p\n", evsel);
 663
 664                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 665         }
 666
 667         waker = register_pid(sched, sample->tid, "<unknown>");
 668         wakee = register_pid(sched, pid, comm);
 669
 670         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 671         return 0;
 672 }
 673
 674 static int replay_switch_event(struct perf_sched *sched,
 675                                struct perf_evsel *evsel,
 676                                struct perf_sample *sample,
 677                                struct machine *machine __maybe_unused)
 678 {
 679         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 680                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 681         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 682                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 683         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 684         struct task_desc *prev, __maybe_unused *next;
 685         u64 timestamp0, timestamp = sample->time;
 686         int cpu = sample->cpu;
 687         s64 delta;
 688
 689         if (verbose)
 690                 printf("sched_switch event %p\n", evsel);
 691
 692         if (cpu >= MAX_CPUS || cpu < 0)
 693                 return 0;
 694
 695         timestamp0 = sched->cpu_last_switched[cpu];
 696         if (timestamp0)
 697                 delta = timestamp - timestamp0;
 698         else
 699                 delta = 0;
 700
 701         if (delta < 0) {
 702                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 703                 return -1;
 704         }
 705
 706         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 707                  prev_comm, prev_pid, next_comm, next_pid, delta);
 708
 709         prev = register_pid(sched, prev_pid, prev_comm);
 710         next = register_pid(sched, next_pid, next_comm);
 711
 712         sched->cpu_last_switched[cpu] = timestamp;
 713
 714         add_sched_event_run(sched, prev, timestamp, delta);
 715         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 716
 717         return 0;
 718 }
 719
 720 static int replay_fork_event(struct perf_sched *sched,
 721                              union perf_event *event,
 722                              struct machine *machine)
 723 {
 724         struct thread *child, *parent;
 725
 726         child = machine__findnew_thread(machine, event->fork.pid,
 727                                         event->fork.tid);
 728         parent = machine__findnew_thread(machine, event->fork.ppid,
 729                                          event->fork.ptid);
 730
 731         if (child == NULL || parent == NULL) {
 732                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 733                                  child, parent);
 734                 return 0;
 735         }
 736
 737         if (verbose) {
 738                 printf("fork event\n");
 739                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 740                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 741         }
 742
 743         register_pid(sched, parent->tid, thread__comm_str(parent));
 744         register_pid(sched, child->tid, thread__comm_str(child));
 745         return 0;
 746 }
 747
 748 struct sort_dimension {
 749         const char              *name;
 750         sort_fn_t               cmp;
 751         struct list_head        list;
 752 };
 753
 754 static int
 755 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 756 {
 757         struct sort_dimension *sort;
 758         int ret = 0;
 759
 760         BUG_ON(list_empty(list));
 761
 762         list_for_each_entry(sort, list, list) {
 763                 ret = sort->cmp(l, r);
 764                 if (ret)
 765                         return ret;
 766         }
 767
 768         return ret;
 769 }
 770
 771 static struct work_atoms *
 772 thread_atoms_search(struct rb_root *root, struct thread *thread,
 773                          struct list_head *sort_list)
 774 {
 775         struct rb_node *node = root->rb_node;
 776         struct work_atoms key = { .thread = thread };
 777
 778         while (node) {
 779                 struct work_atoms *atoms;
 780                 int cmp;
 781
 782                 atoms = container_of(node, struct work_atoms, node);
 783
 784                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 785                 if (cmp > 0)
 786                         node = node->rb_left;
 787                 else if (cmp < 0)
 788                         node = node->rb_right;
 789                 else {
 790                         BUG_ON(thread != atoms->thread);
 791                         return atoms;
 792                 }
 793         }
 794         return NULL;
 795 }
 796
 797 static void
 798 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 799                          struct list_head *sort_list)
 800 {
 801         struct rb_node **new = &(root->rb_node), *parent = NULL;
 802
 803         while (*new) {
 804                 struct work_atoms *this;
 805                 int cmp;
 806
 807                 this = container_of(*new, struct work_atoms, node);
 808                 parent = *new;
 809
 810                 cmp = thread_lat_cmp(sort_list, data, this);
 811
 812                 if (cmp > 0)
 813                         new = &((*new)->rb_left);
 814                 else
 815                         new = &((*new)->rb_right);
 816         }
 817
 818         rb_link_node(&data->node, parent, new);
 819         rb_insert_color(&data->node, root);
 820 }
 821
 822 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 823 {
 824         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 825         if (!atoms) {
 826                 pr_err("No memory at %s\n", __func__);
 827                 return -1;
 828         }
 829
 830         atoms->thread = thread;
 831         INIT_LIST_HEAD(&atoms->work_list);
 832         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 833         return 0;
 834 }
 835
 836 static char sched_out_state(u64 prev_state)
 837 {
 838         const char *str = TASK_STATE_TO_CHAR_STR;
 839
 840         return str[prev_state];
 841 }
 842
 843 static int
 844 add_sched_out_event(struct work_atoms *atoms,
 845                     char run_state,
 846                     u64 timestamp)
 847 {
 848         struct work_atom *atom = zalloc(sizeof(*atom));
 849         if (!atom) {
 850                 pr_err("Non memory at %s", __func__);
 851                 return -1;
 852         }
 853
 854         atom->sched_out_time = timestamp;
 855
 856         if (run_state == 'R') {
 857                 atom->state = THREAD_WAIT_CPU;
 858                 atom->wake_up_time = atom->sched_out_time;
 859         }
 860
 861         list_add_tail(&atom->list, &atoms->work_list);
 862         return 0;
 863 }
 864
 865 static void
 866 add_runtime_event(struct work_atoms *atoms, u64 delta,
 867                   u64 timestamp __maybe_unused)
 868 {
 869         struct work_atom *atom;
 870
 871         BUG_ON(list_empty(&atoms->work_list));
 872
 873         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 874
 875         atom->runtime += delta;
 876         atoms->total_runtime += delta;
 877 }
 878
 879 static void
 880 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 881 {
 882         struct work_atom *atom;
 883         u64 delta;
 884
 885         if (list_empty(&atoms->work_list))
 886                 return;
 887
 888         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 889
 890         if (atom->state != THREAD_WAIT_CPU)
 891                 return;
 892
 893         if (timestamp < atom->wake_up_time) {
 894                 atom->state = THREAD_IGNORE;
 895                 return;
 896         }
 897
 898         atom->state = THREAD_SCHED_IN;
 899         atom->sched_in_time = timestamp;
 900
 901         delta = atom->sched_in_time - atom->wake_up_time;
 902         atoms->total_lat += delta;
 903         if (delta > atoms->max_lat) {
 904                 atoms->max_lat = delta;
 905                 atoms->max_lat_at = timestamp;
 906         }
 907         atoms->nb_atoms++;
 908 }
 909
 910 static int latency_switch_event(struct perf_sched *sched,
 911                                 struct perf_evsel *evsel,
 912                                 struct perf_sample *sample,
 913                                 struct machine *machine)
 914 {
 915         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 916                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 917         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 918         struct work_atoms *out_events, *in_events;
 919         struct thread *sched_out, *sched_in;
 920         u64 timestamp0, timestamp = sample->time;
 921         int cpu = sample->cpu;
 922         s64 delta;
 923
 924         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 925
 926         timestamp0 = sched->cpu_last_switched[cpu];
 927         sched->cpu_last_switched[cpu] = timestamp;
 928         if (timestamp0)
 929                 delta = timestamp - timestamp0;
 930         else
 931                 delta = 0;
 932
 933         if (delta < 0) {
 934                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 935                 return -1;
 936         }
 937
 938         sched_out = machine__findnew_thread(machine, 0, prev_pid);
 939         sched_in = machine__findnew_thread(machine, 0, next_pid);
 940
 941         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 942         if (!out_events) {
 943                 if (thread_atoms_insert(sched, sched_out))
 944                         return -1;
 945                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 946                 if (!out_events) {
 947                         pr_err("out-event: Internal tree error");
 948                         return -1;
 949                 }
 950         }
 951         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 952                 return -1;
 953
 954         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 955         if (!in_events) {
 956                 if (thread_atoms_insert(sched, sched_in))
 957                         return -1;
 958                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 959                 if (!in_events) {
 960                         pr_err("in-event: Internal tree error");
 961                         return -1;
 962                 }
 963                 /*
 964                  * Take came in we have not heard about yet,
 965                  * add in an initial atom in runnable state:
 966                  */
 967                 if (add_sched_out_event(in_events, 'R', timestamp))
 968                         return -1;
 969         }
 970         add_sched_in_event(in_events, timestamp);
 971
 972         return 0;
 973 }
 974
 975 static int latency_runtime_event(struct perf_sched *sched,
 976                                  struct perf_evsel *evsel,
 977                                  struct perf_sample *sample,
 978                                  struct machine *machine)
 979 {
 980         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
 981         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
 982         struct thread *thread = machine__findnew_thread(machine, 0, pid);
 983         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 984         u64 timestamp = sample->time;
 985         int cpu = sample->cpu;
 986
 987         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 988         if (!atoms) {
 989                 if (thread_atoms_insert(sched, thread))
 990                         return -1;
 991                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 992                 if (!atoms) {
 993                         pr_err("in-event: Internal tree error");
 994                         return -1;
 995                 }
 996                 if (add_sched_out_event(atoms, 'R', timestamp))
 997                         return -1;
 998         }
 999
1000         add_runtime_event(atoms, runtime, timestamp);
1001         return 0;
1002 }
1003
1004 static int latency_wakeup_event(struct perf_sched *sched,
1005                                 struct perf_evsel *evsel,
1006                                 struct perf_sample *sample,
1007                                 struct machine *machine)
1008 {
1009         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1010         struct work_atoms *atoms;
1011         struct work_atom *atom;
1012         struct thread *wakee;
1013         u64 timestamp = sample->time;
1014
1015         wakee = machine__findnew_thread(machine, 0, pid);
1016         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1017         if (!atoms) {
1018                 if (thread_atoms_insert(sched, wakee))
1019                         return -1;
1020                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1021                 if (!atoms) {
1022                         pr_err("wakeup-event: Internal tree error");
1023                         return -1;
1024                 }
1025                 if (add_sched_out_event(atoms, 'S', timestamp))
1026                         return -1;
1027         }
1028
1029         BUG_ON(list_empty(&atoms->work_list));
1030
1031         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1032
1033         /*
1034          * As we do not guarantee the wakeup event happens when
1035          * task is out of run queue, also may happen when task is
1036          * on run queue and wakeup only change ->state to TASK_RUNNING,
1037          * then we should not set the ->wake_up_time when wake up a
1038          * task which is on run queue.
1039          *
1040          * You WILL be missing events if you've recorded only
1041          * one CPU, or are only looking at only one, so don't
1042          * skip in this case.
1043          */
1044         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1045                 return 0;
1046
1047         sched->nr_timestamps++;
1048         if (atom->sched_out_time > timestamp) {
1049                 sched->nr_unordered_timestamps++;
1050                 return 0;
1051         }
1052
1053         atom->state = THREAD_WAIT_CPU;
1054         atom->wake_up_time = timestamp;
1055         return 0;
1056 }
1057
1058 static int latency_migrate_task_event(struct perf_sched *sched,
1059                                       struct perf_evsel *evsel,
1060                                       struct perf_sample *sample,
1061                                       struct machine *machine)
1062 {
1063         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1064         u64 timestamp = sample->time;
1065         struct work_atoms *atoms;
1066         struct work_atom *atom;
1067         struct thread *migrant;
1068
1069         /*
1070          * Only need to worry about migration when profiling one CPU.
1071          */
1072         if (sched->profile_cpu == -1)
1073                 return 0;
1074
1075         migrant = machine__findnew_thread(machine, 0, pid);
1076         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1077         if (!atoms) {
1078                 if (thread_atoms_insert(sched, migrant))
1079                         return -1;
1080                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1081                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1082                 if (!atoms) {
1083                         pr_err("migration-event: Internal tree error");
1084                         return -1;
1085                 }
1086                 if (add_sched_out_event(atoms, 'R', timestamp))
1087                         return -1;
1088         }
1089
1090         BUG_ON(list_empty(&atoms->work_list));
1091
1092         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1093         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1094
1095         sched->nr_timestamps++;
1096
1097         if (atom->sched_out_time > timestamp)
1098                 sched->nr_unordered_timestamps++;
1099
1100         return 0;
1101 }
1102
1103 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1104 {
1105         int i;
1106         int ret;
1107         u64 avg;
1108
1109         if (!work_list->nb_atoms)
1110                 return;
1111         /*
1112          * Ignore idle threads:
1113          */
1114         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1115                 return;
1116
1117         sched->all_runtime += work_list->total_runtime;
1118         sched->all_count   += work_list->nb_atoms;
1119
1120         ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1121
1122         for (i = 0; i < 24 - ret; i++)
1123                 printf(" ");
1124
1125         avg = work_list->total_lat / work_list->nb_atoms;
1126
1127         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1128               (double)work_list->total_runtime / 1e6,
1129                  work_list->nb_atoms, (double)avg / 1e6,
1130                  (double)work_list->max_lat / 1e6,
1131                  (double)work_list->max_lat_at / 1e9);
1132 }
1133
1134 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1135 {
1136         if (l->thread->tid < r->thread->tid)
1137                 return -1;
1138         if (l->thread->tid > r->thread->tid)
1139                 return 1;
1140
1141         return 0;
1142 }
1143
1144 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1145 {
1146         u64 avgl, avgr;
1147
1148         if (!l->nb_atoms)
1149                 return -1;
1150
1151         if (!r->nb_atoms)
1152                 return 1;
1153
1154         avgl = l->total_lat / l->nb_atoms;
1155         avgr = r->total_lat / r->nb_atoms;
1156
1157         if (avgl < avgr)
1158                 return -1;
1159         if (avgl > avgr)
1160                 return 1;
1161
1162         return 0;
1163 }
1164
1165 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1166 {
1167         if (l->max_lat < r->max_lat)
1168                 return -1;
1169         if (l->max_lat > r->max_lat)
1170                 return 1;
1171
1172         return 0;
1173 }
1174
1175 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1176 {
1177         if (l->nb_atoms < r->nb_atoms)
1178                 return -1;
1179         if (l->nb_atoms > r->nb_atoms)
1180                 return 1;
1181
1182         return 0;
1183 }
1184
1185 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1186 {
1187         if (l->total_runtime < r->total_runtime)
1188                 return -1;
1189         if (l->total_runtime > r->total_runtime)
1190                 return 1;
1191
1192         return 0;
1193 }
1194
1195 static int sort_dimension__add(const char *tok, struct list_head *list)
1196 {
1197         size_t i;
1198         static struct sort_dimension avg_sort_dimension = {
1199                 .name = "avg",
1200                 .cmp  = avg_cmp,
1201         };
1202         static struct sort_dimension max_sort_dimension = {
1203                 .name = "max",
1204                 .cmp  = max_cmp,
1205         };
1206         static struct sort_dimension pid_sort_dimension = {
1207                 .name = "pid",
1208                 .cmp  = pid_cmp,
1209         };
1210         static struct sort_dimension runtime_sort_dimension = {
1211                 .name = "runtime",
1212                 .cmp  = runtime_cmp,
1213         };
1214         static struct sort_dimension switch_sort_dimension = {
1215                 .name = "switch",
1216                 .cmp  = switch_cmp,
1217         };
1218         struct sort_dimension *available_sorts[] = {
1219                 &pid_sort_dimension,
1220                 &avg_sort_dimension,
1221                 &max_sort_dimension,
1222                 &switch_sort_dimension,
1223                 &runtime_sort_dimension,
1224         };
1225
1226         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1227                 if (!strcmp(available_sorts[i]->name, tok)) {
1228                         list_add_tail(&available_sorts[i]->list, list);
1229
1230                         return 0;
1231                 }
1232         }
1233
1234         return -1;
1235 }
1236
1237 static void perf_sched__sort_lat(struct perf_sched *sched)
1238 {
1239         struct rb_node *node;
1240
1241         for (;;) {
1242                 struct work_atoms *data;
1243                 node = rb_first(&sched->atom_root);
1244                 if (!node)
1245                         break;
1246
1247                 rb_erase(node, &sched->atom_root);
1248                 data = rb_entry(node, struct work_atoms, node);
1249                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1250         }
1251 }
1252
1253 static int process_sched_wakeup_event(struct perf_tool *tool,
1254                                       struct perf_evsel *evsel,
1255                                       struct perf_sample *sample,
1256                                       struct machine *machine)
1257 {
1258         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1259
1260         if (sched->tp_handler->wakeup_event)
1261                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1262
1263         return 0;
1264 }
1265
1266 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1267                             struct perf_sample *sample, struct machine *machine)
1268 {
1269         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1270         struct thread *sched_in;
1271         int new_shortname;
1272         u64 timestamp0, timestamp = sample->time;
1273         s64 delta;
1274         int cpu, this_cpu = sample->cpu;
1275
1276         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1277
1278         if (this_cpu > sched->max_cpu)
1279                 sched->max_cpu = this_cpu;
1280
1281         timestamp0 = sched->cpu_last_switched[this_cpu];
1282         sched->cpu_last_switched[this_cpu] = timestamp;
1283         if (timestamp0)
1284                 delta = timestamp - timestamp0;
1285         else
1286                 delta = 0;
1287
1288         if (delta < 0) {
1289                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1290                 return -1;
1291         }
1292
1293         sched_in = machine__findnew_thread(machine, 0, next_pid);
1294
1295         sched->curr_thread[this_cpu] = sched_in;
1296
1297         printf("  ");
1298
1299         new_shortname = 0;
1300         if (!sched_in->shortname[0]) {
1301                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1302                         /*
1303                          * Don't allocate a letter-number for swapper:0
1304                          * as a shortname. Instead, we use '.' for it.
1305                          */
1306                         sched_in->shortname[0] = '.';
1307                         sched_in->shortname[1] = ' ';
1308                 } else {
1309                         sched_in->shortname[0] = sched->next_shortname1;
1310                         sched_in->shortname[1] = sched->next_shortname2;
1311
1312                         if (sched->next_shortname1 < 'Z') {
1313                                 sched->next_shortname1++;
1314                         } else {
1315                                 sched->next_shortname1 = 'A';
1316                                 if (sched->next_shortname2 < '9')
1317                                         sched->next_shortname2++;
1318                                 else
1319                                         sched->next_shortname2 = '0';
1320                         }
1321                 }
1322                 new_shortname = 1;
1323         }
1324
1325         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1326                 if (cpu != this_cpu)
1327                         printf(" ");
1328                 else
1329                         printf("*");
1330
1331                 if (sched->curr_thread[cpu])
1332                         printf("%2s ", sched->curr_thread[cpu]->shortname);
1333                 else
1334                         printf("   ");
1335         }
1336
1337         printf("  %12.6f secs ", (double)timestamp/1e9);
1338         if (new_shortname) {
1339                 printf("%s => %s:%d\n",
1340                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1341         } else {
1342                 printf("\n");
1343         }
1344
1345         return 0;
1346 }
1347
1348 static int process_sched_switch_event(struct perf_tool *tool,
1349                                       struct perf_evsel *evsel,
1350                                       struct perf_sample *sample,
1351                                       struct machine *machine)
1352 {
1353         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1354         int this_cpu = sample->cpu, err = 0;
1355         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1356             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1357
1358         if (sched->curr_pid[this_cpu] != (u32)-1) {
1359                 /*
1360                  * Are we trying to switch away a PID that is
1361                  * not current?
1362                  */
1363                 if (sched->curr_pid[this_cpu] != prev_pid)
1364                         sched->nr_context_switch_bugs++;
1365         }
1366
1367         if (sched->tp_handler->switch_event)
1368                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1369
1370         sched->curr_pid[this_cpu] = next_pid;
1371         return err;
1372 }
1373
1374 static int process_sched_runtime_event(struct perf_tool *tool,
1375                                        struct perf_evsel *evsel,
1376                                        struct perf_sample *sample,
1377                                        struct machine *machine)
1378 {
1379         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1380
1381         if (sched->tp_handler->runtime_event)
1382                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1383
1384         return 0;
1385 }
1386
1387 static int perf_sched__process_fork_event(struct perf_tool *tool,
1388                                           union perf_event *event,
1389                                           struct perf_sample *sample,
1390                                           struct machine *machine)
1391 {
1392         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1393
1394         /* run the fork event through the perf machineruy */
1395         perf_event__process_fork(tool, event, sample, machine);
1396
1397         /* and then run additional processing needed for this command */
1398         if (sched->tp_handler->fork_event)
1399                 return sched->tp_handler->fork_event(sched, event, machine);
1400
1401         return 0;
1402 }
1403
1404 static int process_sched_migrate_task_event(struct perf_tool *tool,
1405                                             struct perf_evsel *evsel,
1406                                             struct perf_sample *sample,
1407                                             struct machine *machine)
1408 {
1409         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1410
1411         if (sched->tp_handler->migrate_task_event)
1412                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1413
1414         return 0;
1415 }
1416
1417 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1418                                   struct perf_evsel *evsel,
1419                                   struct perf_sample *sample,
1420                                   struct machine *machine);
1421
1422 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1423                                                  union perf_event *event __maybe_unused,
1424                                                  struct perf_sample *sample,
1425                                                  struct perf_evsel *evsel,
1426                                                  struct machine *machine)
1427 {
1428         int err = 0;
1429
1430         evsel->hists.stats.total_period += sample->period;
1431         hists__inc_nr_samples(&evsel->hists, true);
1432
1433         if (evsel->handler != NULL) {
1434                 tracepoint_handler f = evsel->handler;
1435                 err = f(tool, evsel, sample, machine);
1436         }
1437
1438         return err;
1439 }
1440
1441 static int perf_sched__read_events(struct perf_sched *sched,
1442                                    struct perf_session **psession)
1443 {
1444         const struct perf_evsel_str_handler handlers[] = {
1445                 { "sched:sched_switch",       process_sched_switch_event, },
1446                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1447                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1448                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1449                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1450         };
1451         struct perf_session *session;
1452         struct perf_data_file file = {
1453                 .path = input_name,
1454                 .mode = PERF_DATA_MODE_READ,
1455         };
1456
1457         session = perf_session__new(&file, false, &sched->tool);
1458         if (session == NULL) {
1459                 pr_debug("No Memory for session\n");
1460                 return -1;
1461         }
1462
1463         if (perf_session__set_tracepoints_handlers(session, handlers))
1464                 goto out_delete;
1465
1466         if (perf_session__has_traces(session, "record -R")) {
1467                 int err = perf_session__process_events(session, &sched->tool);
1468                 if (err) {
1469                         pr_err("Failed to process events, error %d", err);
1470                         goto out_delete;
1471                 }
1472
1473                 sched->nr_events      = session->stats.nr_events[0];
1474                 sched->nr_lost_events = session->stats.total_lost;
1475                 sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST];
1476         }
1477
1478         if (psession)
1479                 *psession = session;
1480         else
1481                 perf_session__delete(session);
1482
1483         return 0;
1484
1485 out_delete:
1486         perf_session__delete(session);
1487         return -1;
1488 }
1489
1490 static void print_bad_events(struct perf_sched *sched)
1491 {
1492         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1493                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1494                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1495                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1496         }
1497         if (sched->nr_lost_events && sched->nr_events) {
1498                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1499                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1500                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1501         }
1502         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1503                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1504                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1505                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1506                 if (sched->nr_lost_events)
1507                         printf(" (due to lost events?)");
1508                 printf("\n");
1509         }
1510 }
1511
1512 static int perf_sched__lat(struct perf_sched *sched)
1513 {
1514         struct rb_node *next;
1515         struct perf_session *session;
1516
1517         setup_pager();
1518
1519         /* save session -- references to threads are held in work_list */
1520         if (perf_sched__read_events(sched, &session))
1521                 return -1;
1522
1523         perf_sched__sort_lat(sched);
1524
1525         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1526         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1527         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1528
1529         next = rb_first(&sched->sorted_atom_root);
1530
1531         while (next) {
1532                 struct work_atoms *work_list;
1533
1534                 work_list = rb_entry(next, struct work_atoms, node);
1535                 output_lat_thread(sched, work_list);
1536                 next = rb_next(next);
1537         }
1538
1539         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1540         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1541                 (double)sched->all_runtime / 1e6, sched->all_count);
1542
1543         printf(" ---------------------------------------------------\n");
1544
1545         print_bad_events(sched);
1546         printf("\n");
1547
1548         perf_session__delete(session);
1549         return 0;
1550 }
1551
1552 static int perf_sched__map(struct perf_sched *sched)
1553 {
1554         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1555
1556         setup_pager();
1557         if (perf_sched__read_events(sched, NULL))
1558                 return -1;
1559         print_bad_events(sched);
1560         return 0;
1561 }
1562
1563 static int perf_sched__replay(struct perf_sched *sched)
1564 {
1565         unsigned long i;
1566
1567         calibrate_run_measurement_overhead(sched);
1568         calibrate_sleep_measurement_overhead(sched);
1569
1570         test_calibrations(sched);
1571
1572         if (perf_sched__read_events(sched, NULL))
1573                 return -1;
1574
1575         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1576         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1577         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1578
1579         if (sched->targetless_wakeups)
1580                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1581         if (sched->multitarget_wakeups)
1582                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1583         if (sched->nr_run_events_optimized)
1584                 printf("run atoms optimized: %ld\n",
1585                         sched->nr_run_events_optimized);
1586
1587         print_task_traces(sched);
1588         add_cross_task_wakeups(sched);
1589
1590         create_tasks(sched);
1591         printf("------------------------------------------------------------\n");
1592         for (i = 0; i < sched->replay_repeat; i++)
1593                 run_one_test(sched);
1594
1595         return 0;
1596 }
1597
1598 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1599                           const char * const usage_msg[])
1600 {
1601         char *tmp, *tok, *str = strdup(sched->sort_order);
1602
1603         for (tok = strtok_r(str, ", ", &tmp);
1604                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1605                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1606                         error("Unknown --sort key: `%s'", tok);
1607                         usage_with_options(usage_msg, options);
1608                 }
1609         }
1610
1611         free(str);
1612
1613         sort_dimension__add("pid", &sched->cmp_pid);
1614 }
1615
1616 static int __cmd_record(int argc, const char **argv)
1617 {
1618         unsigned int rec_argc, i, j;
1619         const char **rec_argv;
1620         const char * const record_args[] = {
1621                 "record",
1622                 "-a",
1623                 "-R",
1624                 "-m", "1024",
1625                 "-c", "1",
1626                 "-e", "sched:sched_switch",
1627                 "-e", "sched:sched_stat_wait",
1628                 "-e", "sched:sched_stat_sleep",
1629                 "-e", "sched:sched_stat_iowait",
1630                 "-e", "sched:sched_stat_runtime",
1631                 "-e", "sched:sched_process_fork",
1632                 "-e", "sched:sched_wakeup",
1633                 "-e", "sched:sched_wakeup_new",
1634                 "-e", "sched:sched_migrate_task",
1635         };
1636
1637         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1638         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1639
1640         if (rec_argv == NULL)
1641                 return -ENOMEM;
1642
1643         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1644                 rec_argv[i] = strdup(record_args[i]);
1645
1646         for (j = 1; j < (unsigned int)argc; j++, i++)
1647                 rec_argv[i] = argv[j];
1648
1649         BUG_ON(i != rec_argc);
1650
1651         return cmd_record(i, rec_argv, NULL);
1652 }
1653
1654 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1655 {
1656         const char default_sort_order[] = "avg, max, switch, runtime";
1657         struct perf_sched sched = {
1658                 .tool = {
1659                         .sample          = perf_sched__process_tracepoint_sample,
1660                         .comm            = perf_event__process_comm,
1661                         .lost            = perf_event__process_lost,
1662                         .fork            = perf_sched__process_fork_event,
1663                         .ordered_samples = true,
1664                 },
1665                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1666                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1667                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1668                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1669                 .sort_order           = default_sort_order,
1670                 .replay_repeat        = 10,
1671                 .profile_cpu          = -1,
1672                 .next_shortname1      = 'A',
1673                 .next_shortname2      = '0',
1674         };
1675         const struct option latency_options[] = {
1676         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1677                    "sort by key(s): runtime, switch, avg, max"),
1678         OPT_INCR('v', "verbose", &verbose,
1679                     "be more verbose (show symbol address, etc)"),
1680         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1681                     "CPU to profile on"),
1682         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1683                     "dump raw trace in ASCII"),
1684         OPT_END()
1685         };
1686         const struct option replay_options[] = {
1687         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1688                      "repeat the workload replay N times (-1: infinite)"),
1689         OPT_INCR('v', "verbose", &verbose,
1690                     "be more verbose (show symbol address, etc)"),
1691         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1692                     "dump raw trace in ASCII"),
1693         OPT_END()
1694         };
1695         const struct option sched_options[] = {
1696         OPT_STRING('i', "input", &input_name, "file",
1697                     "input file name"),
1698         OPT_INCR('v', "verbose", &verbose,
1699                     "be more verbose (show symbol address, etc)"),
1700         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1701                     "dump raw trace in ASCII"),
1702         OPT_END()
1703         };
1704         const char * const latency_usage[] = {
1705                 "perf sched latency [<options>]",
1706                 NULL
1707         };
1708         const char * const replay_usage[] = {
1709                 "perf sched replay [<options>]",
1710                 NULL
1711         };
1712         const char *const sched_subcommands[] = { "record", "latency", "map",
1713                                                   "replay", "script", NULL };
1714         const char *sched_usage[] = {
1715                 NULL,
1716                 NULL
1717         };
1718         struct trace_sched_handler lat_ops  = {
1719                 .wakeup_event       = latency_wakeup_event,
1720                 .switch_event       = latency_switch_event,
1721                 .runtime_event      = latency_runtime_event,
1722                 .migrate_task_event = latency_migrate_task_event,
1723         };
1724         struct trace_sched_handler map_ops  = {
1725                 .switch_event       = map_switch_event,
1726         };
1727         struct trace_sched_handler replay_ops  = {
1728                 .wakeup_event       = replay_wakeup_event,
1729                 .switch_event       = replay_switch_event,
1730                 .fork_event         = replay_fork_event,
1731         };
1732         unsigned int i;
1733
1734         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1735                 sched.curr_pid[i] = -1;
1736
1737         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1738                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1739         if (!argc)
1740                 usage_with_options(sched_usage, sched_options);
1741
1742         /*
1743          * Aliased to 'perf script' for now:
1744          */
1745         if (!strcmp(argv[0], "script"))
1746                 return cmd_script(argc, argv, prefix);
1747
1748         symbol__init();
1749         if (!strncmp(argv[0], "rec", 3)) {
1750                 return __cmd_record(argc, argv);
1751         } else if (!strncmp(argv[0], "lat", 3)) {
1752                 sched.tp_handler = &lat_ops;
1753                 if (argc > 1) {
1754                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1755                         if (argc)
1756                                 usage_with_options(latency_usage, latency_options);
1757                 }
1758                 setup_sorting(&sched, latency_options, latency_usage);
1759                 return perf_sched__lat(&sched);
1760         } else if (!strcmp(argv[0], "map")) {
1761                 sched.tp_handler = &map_ops;
1762                 setup_sorting(&sched, latency_options, latency_usage);
1763                 return perf_sched__map(&sched);
1764         } else if (!strncmp(argv[0], "rep", 3)) {
1765                 sched.tp_handler = &replay_ops;
1766                 if (argc) {
1767                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1768                         if (argc)
1769                                 usage_with_options(replay_usage, replay_options);
1770                 }
1771                 return perf_sched__replay(&sched);
1772         } else {
1773                 usage_with_options(sched_usage, sched_options);
1774         }
1775
1776         return 0;
1777 }