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 #include "util/cloexec.h"
  14
  15 #include "util/parse-options.h"
  16 #include "util/trace-event.h"
  17
  18 #include "util/debug.h"
  19
  20 #include <sys/prctl.h>
  21 #include <sys/resource.h>
  22
  23 #include <semaphore.h>
  24 #include <pthread.h>
  25 #include <math.h>
  26 #include <api/fs/fs.h>
  27
  28 #define PR_SET_NAME             15               /* Set process name */
  29 #define MAX_CPUS                4096
  30 #define COMM_LEN                20
  31 #define SYM_LEN                 129
  32 #define MAX_PID                 1024000
  33
  34 struct sched_atom;
  35
  36 struct task_desc {
  37         unsigned long           nr;
  38         unsigned long           pid;
  39         char                    comm[COMM_LEN];
  40
  41         unsigned long           nr_events;
  42         unsigned long           curr_event;
  43         struct sched_atom       **atoms;
  44
  45         pthread_t               thread;
  46         sem_t                   sleep_sem;
  47
  48         sem_t                   ready_for_work;
  49         sem_t                   work_done_sem;
  50
  51         u64                     cpu_usage;
  52 };
  53
  54 enum sched_event_type {
  55         SCHED_EVENT_RUN,
  56         SCHED_EVENT_SLEEP,
  57         SCHED_EVENT_WAKEUP,
  58         SCHED_EVENT_MIGRATION,
  59 };
  60
  61 struct sched_atom {
  62         enum sched_event_type   type;
  63         int                     specific_wait;
  64         u64                     timestamp;
  65         u64                     duration;
  66         unsigned long           nr;
  67         sem_t                   *wait_sem;
  68         struct task_desc        *wakee;
  69 };
  70
  71 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  72
  73 enum thread_state {
  74         THREAD_SLEEPING = 0,
  75         THREAD_WAIT_CPU,
  76         THREAD_SCHED_IN,
  77         THREAD_IGNORE
  78 };
  79
  80 struct work_atom {
  81         struct list_head        list;
  82         enum thread_state       state;
  83         u64                     sched_out_time;
  84         u64                     wake_up_time;
  85         u64                     sched_in_time;
  86         u64                     runtime;
  87 };
  88
  89 struct work_atoms {
  90         struct list_head        work_list;
  91         struct thread           *thread;
  92         struct rb_node          node;
  93         u64                     max_lat;
  94         u64                     max_lat_at;
  95         u64                     total_lat;
  96         u64                     nb_atoms;
  97         u64                     total_runtime;
  98 };
  99
 100 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 101
 102 struct perf_sched;
 103
 104 struct trace_sched_handler {
 105         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 106                             struct perf_sample *sample, struct machine *machine);
 107
 108         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 109                              struct perf_sample *sample, struct machine *machine);
 110
 111         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 112                             struct perf_sample *sample, struct machine *machine);
 113
 114         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 115         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 116                           struct machine *machine);
 117
 118         int (*migrate_task_event)(struct perf_sched *sched,
 119                                   struct perf_evsel *evsel,
 120                                   struct perf_sample *sample,
 121                                   struct machine *machine);
 122 };
 123
 124 struct perf_sched {
 125         struct perf_tool tool;
 126         const char       *sort_order;
 127         unsigned long    nr_tasks;
 128         struct task_desc **pid_to_task;
 129         struct task_desc **tasks;
 130         const struct trace_sched_handler *tp_handler;
 131         pthread_mutex_t  start_work_mutex;
 132         pthread_mutex_t  work_done_wait_mutex;
 133         int              profile_cpu;
 134 /*
 135  * Track the current task - that way we can know whether there's any
 136  * weird events, such as a task being switched away that is not current.
 137  */
 138         int              max_cpu;
 139         u32              curr_pid[MAX_CPUS];
 140         struct thread    *curr_thread[MAX_CPUS];
 141         char             next_shortname1;
 142         char             next_shortname2;
 143         unsigned int     replay_repeat;
 144         unsigned long    nr_run_events;
 145         unsigned long    nr_sleep_events;
 146         unsigned long    nr_wakeup_events;
 147         unsigned long    nr_sleep_corrections;
 148         unsigned long    nr_run_events_optimized;
 149         unsigned long    targetless_wakeups;
 150         unsigned long    multitarget_wakeups;
 151         unsigned long    nr_runs;
 152         unsigned long    nr_timestamps;
 153         unsigned long    nr_unordered_timestamps;
 154         unsigned long    nr_context_switch_bugs;
 155         unsigned long    nr_events;
 156         unsigned long    nr_lost_chunks;
 157         unsigned long    nr_lost_events;
 158         u64              run_measurement_overhead;
 159         u64              sleep_measurement_overhead;
 160         u64              start_time;
 161         u64              cpu_usage;
 162         u64              runavg_cpu_usage;
 163         u64              parent_cpu_usage;
 164         u64              runavg_parent_cpu_usage;
 165         u64              sum_runtime;
 166         u64              sum_fluct;
 167         u64              run_avg;
 168         u64              all_runtime;
 169         u64              all_count;
 170         u64              cpu_last_switched[MAX_CPUS];
 171         struct rb_root   atom_root, sorted_atom_root;
 172         struct list_head sort_list, cmp_pid;
 173 };
 174
 175 static u64 get_nsecs(void)
 176 {
 177         struct timespec ts;
 178
 179         clock_gettime(CLOCK_MONOTONIC, &ts);
 180
 181         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 182 }
 183
 184 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 185 {
 186         u64 T0 = get_nsecs(), T1;
 187
 188         do {
 189                 T1 = get_nsecs();
 190         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 191 }
 192
 193 static void sleep_nsecs(u64 nsecs)
 194 {
 195         struct timespec ts;
 196
 197         ts.tv_nsec = nsecs % 999999999;
 198         ts.tv_sec = nsecs / 999999999;
 199
 200         nanosleep(&ts, NULL);
 201 }
 202
 203 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 204 {
 205         u64 T0, T1, delta, min_delta = 1000000000ULL;
 206         int i;
 207
 208         for (i = 0; i < 10; i++) {
 209                 T0 = get_nsecs();
 210                 burn_nsecs(sched, 0);
 211                 T1 = get_nsecs();
 212                 delta = T1-T0;
 213                 min_delta = min(min_delta, delta);
 214         }
 215         sched->run_measurement_overhead = min_delta;
 216
 217         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 218 }
 219
 220 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 221 {
 222         u64 T0, T1, delta, min_delta = 1000000000ULL;
 223         int i;
 224
 225         for (i = 0; i < 10; i++) {
 226                 T0 = get_nsecs();
 227                 sleep_nsecs(10000);
 228                 T1 = get_nsecs();
 229                 delta = T1-T0;
 230                 min_delta = min(min_delta, delta);
 231         }
 232         min_delta -= 10000;
 233         sched->sleep_measurement_overhead = min_delta;
 234
 235         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 236 }
 237
 238 static struct sched_atom *
 239 get_new_event(struct task_desc *task, u64 timestamp)
 240 {
 241         struct sched_atom *event = zalloc(sizeof(*event));
 242         unsigned long idx = task->nr_events;
 243         size_t size;
 244
 245         event->timestamp = timestamp;
 246         event->nr = idx;
 247
 248         task->nr_events++;
 249         size = sizeof(struct sched_atom *) * task->nr_events;
 250         task->atoms = realloc(task->atoms, size);
 251         BUG_ON(!task->atoms);
 252
 253         task->atoms[idx] = event;
 254
 255         return event;
 256 }
 257
 258 static struct sched_atom *last_event(struct task_desc *task)
 259 {
 260         if (!task->nr_events)
 261                 return NULL;
 262
 263         return task->atoms[task->nr_events - 1];
 264 }
 265
 266 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 267                                 u64 timestamp, u64 duration)
 268 {
 269         struct sched_atom *event, *curr_event = last_event(task);
 270
 271         /*
 272          * optimize an existing RUN event by merging this one
 273          * to it:
 274          */
 275         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 276                 sched->nr_run_events_optimized++;
 277                 curr_event->duration += duration;
 278                 return;
 279         }
 280
 281         event = get_new_event(task, timestamp);
 282
 283         event->type = SCHED_EVENT_RUN;
 284         event->duration = duration;
 285
 286         sched->nr_run_events++;
 287 }
 288
 289 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 290                                    u64 timestamp, struct task_desc *wakee)
 291 {
 292         struct sched_atom *event, *wakee_event;
 293
 294         event = get_new_event(task, timestamp);
 295         event->type = SCHED_EVENT_WAKEUP;
 296         event->wakee = wakee;
 297
 298         wakee_event = last_event(wakee);
 299         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 300                 sched->targetless_wakeups++;
 301                 return;
 302         }
 303         if (wakee_event->wait_sem) {
 304                 sched->multitarget_wakeups++;
 305                 return;
 306         }
 307
 308         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 309         sem_init(wakee_event->wait_sem, 0, 0);
 310         wakee_event->specific_wait = 1;
 311         event->wait_sem = wakee_event->wait_sem;
 312
 313         sched->nr_wakeup_events++;
 314 }
 315
 316 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 317                                   u64 timestamp, u64 task_state __maybe_unused)
 318 {
 319         struct sched_atom *event = get_new_event(task, timestamp);
 320
 321         event->type = SCHED_EVENT_SLEEP;
 322
 323         sched->nr_sleep_events++;
 324 }
 325
 326 static struct task_desc *register_pid(struct perf_sched *sched,
 327                                       unsigned long pid, const char *comm)
 328 {
 329         struct task_desc *task;
 330         static int pid_max;
 331
 332         if (sched->pid_to_task == NULL) {
 333                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 334                         pid_max = MAX_PID;
 335                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 336         }
 337         if (pid >= (unsigned long)pid_max) {
 338                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 339                         sizeof(struct task_desc *))) == NULL);
 340                 while (pid >= (unsigned long)pid_max)
 341                         sched->pid_to_task[pid_max++] = NULL;
 342         }
 343
 344         task = sched->pid_to_task[pid];
 345
 346         if (task)
 347                 return task;
 348
 349         task = zalloc(sizeof(*task));
 350         task->pid = pid;
 351         task->nr = sched->nr_tasks;
 352         strcpy(task->comm, comm);
 353         /*
 354          * every task starts in sleeping state - this gets ignored
 355          * if there's no wakeup pointing to this sleep state:
 356          */
 357         add_sched_event_sleep(sched, task, 0, 0);
 358
 359         sched->pid_to_task[pid] = task;
 360         sched->nr_tasks++;
 361         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 362         BUG_ON(!sched->tasks);
 363         sched->tasks[task->nr] = task;
 364
 365         if (verbose)
 366                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 367
 368         return task;
 369 }
 370
 371
 372 static void print_task_traces(struct perf_sched *sched)
 373 {
 374         struct task_desc *task;
 375         unsigned long i;
 376
 377         for (i = 0; i < sched->nr_tasks; i++) {
 378                 task = sched->tasks[i];
 379                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 380                         task->nr, task->comm, task->pid, task->nr_events);
 381         }
 382 }
 383
 384 static void add_cross_task_wakeups(struct perf_sched *sched)
 385 {
 386         struct task_desc *task1, *task2;
 387         unsigned long i, j;
 388
 389         for (i = 0; i < sched->nr_tasks; i++) {
 390                 task1 = sched->tasks[i];
 391                 j = i + 1;
 392                 if (j == sched->nr_tasks)
 393                         j = 0;
 394                 task2 = sched->tasks[j];
 395                 add_sched_event_wakeup(sched, task1, 0, task2);
 396         }
 397 }
 398
 399 static void perf_sched__process_event(struct perf_sched *sched,
 400                                       struct sched_atom *atom)
 401 {
 402         int ret = 0;
 403
 404         switch (atom->type) {
 405                 case SCHED_EVENT_RUN:
 406                         burn_nsecs(sched, atom->duration);
 407                         break;
 408                 case SCHED_EVENT_SLEEP:
 409                         if (atom->wait_sem)
 410                                 ret = sem_wait(atom->wait_sem);
 411                         BUG_ON(ret);
 412                         break;
 413                 case SCHED_EVENT_WAKEUP:
 414                         if (atom->wait_sem)
 415                                 ret = sem_post(atom->wait_sem);
 416                         BUG_ON(ret);
 417                         break;
 418                 case SCHED_EVENT_MIGRATION:
 419                         break;
 420                 default:
 421                         BUG_ON(1);
 422         }
 423 }
 424
 425 static u64 get_cpu_usage_nsec_parent(void)
 426 {
 427         struct rusage ru;
 428         u64 sum;
 429         int err;
 430
 431         err = getrusage(RUSAGE_SELF, &ru);
 432         BUG_ON(err);
 433
 434         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 435         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 436
 437         return sum;
 438 }
 439
 440 static int self_open_counters(void)
 441 {
 442         struct perf_event_attr attr;
 443         char sbuf[STRERR_BUFSIZE];
 444         int fd;
 445
 446         memset(&attr, 0, sizeof(attr));
 447
 448         attr.type = PERF_TYPE_SOFTWARE;
 449         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 450
 451         fd = sys_perf_event_open(&attr, 0, -1, -1,
 452                                  perf_event_open_cloexec_flag());
 453
 454         if (fd < 0)
 455                 pr_err("Error: sys_perf_event_open() syscall returned "
 456                        "with %d (%s)\n", fd,
 457                        strerror_r(errno, sbuf, sizeof(sbuf)));
 458         return fd;
 459 }
 460
 461 static u64 get_cpu_usage_nsec_self(int fd)
 462 {
 463         u64 runtime;
 464         int ret;
 465
 466         ret = read(fd, &runtime, sizeof(runtime));
 467         BUG_ON(ret != sizeof(runtime));
 468
 469         return runtime;
 470 }
 471
 472 struct sched_thread_parms {
 473         struct task_desc  *task;
 474         struct perf_sched *sched;
 475         int fd;
 476 };
 477
 478 static void *thread_func(void *ctx)
 479 {
 480         struct sched_thread_parms *parms = ctx;
 481         struct task_desc *this_task = parms->task;
 482         struct perf_sched *sched = parms->sched;
 483         u64 cpu_usage_0, cpu_usage_1;
 484         unsigned long i, ret;
 485         char comm2[22];
 486         int fd = parms->fd;
 487
 488         zfree(&parms);
 489
 490         sprintf(comm2, ":%s", this_task->comm);
 491         prctl(PR_SET_NAME, comm2);
 492         if (fd < 0)
 493                 return NULL;
 494 again:
 495         ret = sem_post(&this_task->ready_for_work);
 496         BUG_ON(ret);
 497         ret = pthread_mutex_lock(&sched->start_work_mutex);
 498         BUG_ON(ret);
 499         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 500         BUG_ON(ret);
 501
 502         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 503
 504         for (i = 0; i < this_task->nr_events; i++) {
 505                 this_task->curr_event = i;
 506                 perf_sched__process_event(sched, this_task->atoms[i]);
 507         }
 508
 509         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 510         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 511         ret = sem_post(&this_task->work_done_sem);
 512         BUG_ON(ret);
 513
 514         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 515         BUG_ON(ret);
 516         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 517         BUG_ON(ret);
 518
 519         goto again;
 520 }
 521
 522 static void create_tasks(struct perf_sched *sched)
 523 {
 524         struct task_desc *task;
 525         pthread_attr_t attr;
 526         unsigned long i;
 527         int err;
 528
 529         err = pthread_attr_init(&attr);
 530         BUG_ON(err);
 531         err = pthread_attr_setstacksize(&attr,
 532                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 533         BUG_ON(err);
 534         err = pthread_mutex_lock(&sched->start_work_mutex);
 535         BUG_ON(err);
 536         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 537         BUG_ON(err);
 538         for (i = 0; i < sched->nr_tasks; i++) {
 539                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 540                 BUG_ON(parms == NULL);
 541                 parms->task = task = sched->tasks[i];
 542                 parms->sched = sched;
 543                 parms->fd = self_open_counters();
 544                 sem_init(&task->sleep_sem, 0, 0);
 545                 sem_init(&task->ready_for_work, 0, 0);
 546                 sem_init(&task->work_done_sem, 0, 0);
 547                 task->curr_event = 0;
 548                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 549                 BUG_ON(err);
 550         }
 551 }
 552
 553 static void wait_for_tasks(struct perf_sched *sched)
 554 {
 555         u64 cpu_usage_0, cpu_usage_1;
 556         struct task_desc *task;
 557         unsigned long i, ret;
 558
 559         sched->start_time = get_nsecs();
 560         sched->cpu_usage = 0;
 561         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 562
 563         for (i = 0; i < sched->nr_tasks; i++) {
 564                 task = sched->tasks[i];
 565                 ret = sem_wait(&task->ready_for_work);
 566                 BUG_ON(ret);
 567                 sem_init(&task->ready_for_work, 0, 0);
 568         }
 569         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 570         BUG_ON(ret);
 571
 572         cpu_usage_0 = get_cpu_usage_nsec_parent();
 573
 574         pthread_mutex_unlock(&sched->start_work_mutex);
 575
 576         for (i = 0; i < sched->nr_tasks; i++) {
 577                 task = sched->tasks[i];
 578                 ret = sem_wait(&task->work_done_sem);
 579                 BUG_ON(ret);
 580                 sem_init(&task->work_done_sem, 0, 0);
 581                 sched->cpu_usage += task->cpu_usage;
 582                 task->cpu_usage = 0;
 583         }
 584
 585         cpu_usage_1 = get_cpu_usage_nsec_parent();
 586         if (!sched->runavg_cpu_usage)
 587                 sched->runavg_cpu_usage = sched->cpu_usage;
 588         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
 589
 590         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 591         if (!sched->runavg_parent_cpu_usage)
 592                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 593         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
 594                                          sched->parent_cpu_usage)/10;
 595
 596         ret = pthread_mutex_lock(&sched->start_work_mutex);
 597         BUG_ON(ret);
 598
 599         for (i = 0; i < sched->nr_tasks; i++) {
 600                 task = sched->tasks[i];
 601                 sem_init(&task->sleep_sem, 0, 0);
 602                 task->curr_event = 0;
 603         }
 604 }
 605
 606 static void run_one_test(struct perf_sched *sched)
 607 {
 608         u64 T0, T1, delta, avg_delta, fluct;
 609
 610         T0 = get_nsecs();
 611         wait_for_tasks(sched);
 612         T1 = get_nsecs();
 613
 614         delta = T1 - T0;
 615         sched->sum_runtime += delta;
 616         sched->nr_runs++;
 617
 618         avg_delta = sched->sum_runtime / sched->nr_runs;
 619         if (delta < avg_delta)
 620                 fluct = avg_delta - delta;
 621         else
 622                 fluct = delta - avg_delta;
 623         sched->sum_fluct += fluct;
 624         if (!sched->run_avg)
 625                 sched->run_avg = delta;
 626         sched->run_avg = (sched->run_avg * 9 + delta) / 10;
 627
 628         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 629
 630         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 631
 632         printf("cpu: %0.2f / %0.2f",
 633                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 634
 635 #if 0
 636         /*
 637          * rusage statistics done by the parent, these are less
 638          * accurate than the sched->sum_exec_runtime based statistics:
 639          */
 640         printf(" [%0.2f / %0.2f]",
 641                 (double)sched->parent_cpu_usage/1e6,
 642                 (double)sched->runavg_parent_cpu_usage/1e6);
 643 #endif
 644
 645         printf("\n");
 646
 647         if (sched->nr_sleep_corrections)
 648                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 649         sched->nr_sleep_corrections = 0;
 650 }
 651
 652 static void test_calibrations(struct perf_sched *sched)
 653 {
 654         u64 T0, T1;
 655
 656         T0 = get_nsecs();
 657         burn_nsecs(sched, 1e6);
 658         T1 = get_nsecs();
 659
 660         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 661
 662         T0 = get_nsecs();
 663         sleep_nsecs(1e6);
 664         T1 = get_nsecs();
 665
 666         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 667 }
 668
 669 static int
 670 replay_wakeup_event(struct perf_sched *sched,
 671                     struct perf_evsel *evsel, struct perf_sample *sample,
 672                     struct machine *machine __maybe_unused)
 673 {
 674         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 675         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 676         struct task_desc *waker, *wakee;
 677
 678         if (verbose) {
 679                 printf("sched_wakeup event %p\n", evsel);
 680
 681                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 682         }
 683
 684         waker = register_pid(sched, sample->tid, "<unknown>");
 685         wakee = register_pid(sched, pid, comm);
 686
 687         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 688         return 0;
 689 }
 690
 691 static int replay_switch_event(struct perf_sched *sched,
 692                                struct perf_evsel *evsel,
 693                                struct perf_sample *sample,
 694                                struct machine *machine __maybe_unused)
 695 {
 696         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 697                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 698         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 699                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 700         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 701         struct task_desc *prev, __maybe_unused *next;
 702         u64 timestamp0, timestamp = sample->time;
 703         int cpu = sample->cpu;
 704         s64 delta;
 705
 706         if (verbose)
 707                 printf("sched_switch event %p\n", evsel);
 708
 709         if (cpu >= MAX_CPUS || cpu < 0)
 710                 return 0;
 711
 712         timestamp0 = sched->cpu_last_switched[cpu];
 713         if (timestamp0)
 714                 delta = timestamp - timestamp0;
 715         else
 716                 delta = 0;
 717
 718         if (delta < 0) {
 719                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 720                 return -1;
 721         }
 722
 723         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 724                  prev_comm, prev_pid, next_comm, next_pid, delta);
 725
 726         prev = register_pid(sched, prev_pid, prev_comm);
 727         next = register_pid(sched, next_pid, next_comm);
 728
 729         sched->cpu_last_switched[cpu] = timestamp;
 730
 731         add_sched_event_run(sched, prev, timestamp, delta);
 732         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 733
 734         return 0;
 735 }
 736
 737 static int replay_fork_event(struct perf_sched *sched,
 738                              union perf_event *event,
 739                              struct machine *machine)
 740 {
 741         struct thread *child, *parent;
 742
 743         child = machine__findnew_thread(machine, event->fork.pid,
 744                                         event->fork.tid);
 745         parent = machine__findnew_thread(machine, event->fork.ppid,
 746                                          event->fork.ptid);
 747
 748         if (child == NULL || parent == NULL) {
 749                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 750                                  child, parent);
 751                 return 0;
 752         }
 753
 754         if (verbose) {
 755                 printf("fork event\n");
 756                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 757                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 758         }
 759
 760         register_pid(sched, parent->tid, thread__comm_str(parent));
 761         register_pid(sched, child->tid, thread__comm_str(child));
 762         return 0;
 763 }
 764
 765 struct sort_dimension {
 766         const char              *name;
 767         sort_fn_t               cmp;
 768         struct list_head        list;
 769 };
 770
 771 static int
 772 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 773 {
 774         struct sort_dimension *sort;
 775         int ret = 0;
 776
 777         BUG_ON(list_empty(list));
 778
 779         list_for_each_entry(sort, list, list) {
 780                 ret = sort->cmp(l, r);
 781                 if (ret)
 782                         return ret;
 783         }
 784
 785         return ret;
 786 }
 787
 788 static struct work_atoms *
 789 thread_atoms_search(struct rb_root *root, struct thread *thread,
 790                          struct list_head *sort_list)
 791 {
 792         struct rb_node *node = root->rb_node;
 793         struct work_atoms key = { .thread = thread };
 794
 795         while (node) {
 796                 struct work_atoms *atoms;
 797                 int cmp;
 798
 799                 atoms = container_of(node, struct work_atoms, node);
 800
 801                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 802                 if (cmp > 0)
 803                         node = node->rb_left;
 804                 else if (cmp < 0)
 805                         node = node->rb_right;
 806                 else {
 807                         BUG_ON(thread != atoms->thread);
 808                         return atoms;
 809                 }
 810         }
 811         return NULL;
 812 }
 813
 814 static void
 815 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 816                          struct list_head *sort_list)
 817 {
 818         struct rb_node **new = &(root->rb_node), *parent = NULL;
 819
 820         while (*new) {
 821                 struct work_atoms *this;
 822                 int cmp;
 823
 824                 this = container_of(*new, struct work_atoms, node);
 825                 parent = *new;
 826
 827                 cmp = thread_lat_cmp(sort_list, data, this);
 828
 829                 if (cmp > 0)
 830                         new = &((*new)->rb_left);
 831                 else
 832                         new = &((*new)->rb_right);
 833         }
 834
 835         rb_link_node(&data->node, parent, new);
 836         rb_insert_color(&data->node, root);
 837 }
 838
 839 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 840 {
 841         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 842         if (!atoms) {
 843                 pr_err("No memory at %s\n", __func__);
 844                 return -1;
 845         }
 846
 847         atoms->thread = thread__get(thread);
 848         INIT_LIST_HEAD(&atoms->work_list);
 849         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 850         return 0;
 851 }
 852
 853 static char sched_out_state(u64 prev_state)
 854 {
 855         const char *str = TASK_STATE_TO_CHAR_STR;
 856
 857         return str[prev_state];
 858 }
 859
 860 static int
 861 add_sched_out_event(struct work_atoms *atoms,
 862                     char run_state,
 863                     u64 timestamp)
 864 {
 865         struct work_atom *atom = zalloc(sizeof(*atom));
 866         if (!atom) {
 867                 pr_err("Non memory at %s", __func__);
 868                 return -1;
 869         }
 870
 871         atom->sched_out_time = timestamp;
 872
 873         if (run_state == 'R') {
 874                 atom->state = THREAD_WAIT_CPU;
 875                 atom->wake_up_time = atom->sched_out_time;
 876         }
 877
 878         list_add_tail(&atom->list, &atoms->work_list);
 879         return 0;
 880 }
 881
 882 static void
 883 add_runtime_event(struct work_atoms *atoms, u64 delta,
 884                   u64 timestamp __maybe_unused)
 885 {
 886         struct work_atom *atom;
 887
 888         BUG_ON(list_empty(&atoms->work_list));
 889
 890         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 891
 892         atom->runtime += delta;
 893         atoms->total_runtime += delta;
 894 }
 895
 896 static void
 897 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 898 {
 899         struct work_atom *atom;
 900         u64 delta;
 901
 902         if (list_empty(&atoms->work_list))
 903                 return;
 904
 905         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 906
 907         if (atom->state != THREAD_WAIT_CPU)
 908                 return;
 909
 910         if (timestamp < atom->wake_up_time) {
 911                 atom->state = THREAD_IGNORE;
 912                 return;
 913         }
 914
 915         atom->state = THREAD_SCHED_IN;
 916         atom->sched_in_time = timestamp;
 917
 918         delta = atom->sched_in_time - atom->wake_up_time;
 919         atoms->total_lat += delta;
 920         if (delta > atoms->max_lat) {
 921                 atoms->max_lat = delta;
 922                 atoms->max_lat_at = timestamp;
 923         }
 924         atoms->nb_atoms++;
 925 }
 926
 927 static int latency_switch_event(struct perf_sched *sched,
 928                                 struct perf_evsel *evsel,
 929                                 struct perf_sample *sample,
 930                                 struct machine *machine)
 931 {
 932         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 933                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 934         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 935         struct work_atoms *out_events, *in_events;
 936         struct thread *sched_out, *sched_in;
 937         u64 timestamp0, timestamp = sample->time;
 938         int cpu = sample->cpu;
 939         s64 delta;
 940
 941         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 942
 943         timestamp0 = sched->cpu_last_switched[cpu];
 944         sched->cpu_last_switched[cpu] = timestamp;
 945         if (timestamp0)
 946                 delta = timestamp - timestamp0;
 947         else
 948                 delta = 0;
 949
 950         if (delta < 0) {
 951                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 952                 return -1;
 953         }
 954
 955         sched_out = machine__findnew_thread(machine, -1, prev_pid);
 956         sched_in = machine__findnew_thread(machine, -1, next_pid);
 957
 958         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 959         if (!out_events) {
 960                 if (thread_atoms_insert(sched, sched_out))
 961                         return -1;
 962                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 963                 if (!out_events) {
 964                         pr_err("out-event: Internal tree error");
 965                         return -1;
 966                 }
 967         }
 968         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 969                 return -1;
 970
 971         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 972         if (!in_events) {
 973                 if (thread_atoms_insert(sched, sched_in))
 974                         return -1;
 975                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 976                 if (!in_events) {
 977                         pr_err("in-event: Internal tree error");
 978                         return -1;
 979                 }
 980                 /*
 981                  * Take came in we have not heard about yet,
 982                  * add in an initial atom in runnable state:
 983                  */
 984                 if (add_sched_out_event(in_events, 'R', timestamp))
 985                         return -1;
 986         }
 987         add_sched_in_event(in_events, timestamp);
 988
 989         return 0;
 990 }
 991
 992 static int latency_runtime_event(struct perf_sched *sched,
 993                                  struct perf_evsel *evsel,
 994                                  struct perf_sample *sample,
 995                                  struct machine *machine)
 996 {
 997         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
 998         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
 999         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1000         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1001         u64 timestamp = sample->time;
1002         int cpu = sample->cpu;
1003
1004         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1005         if (!atoms) {
1006                 if (thread_atoms_insert(sched, thread))
1007                         return -1;
1008                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1009                 if (!atoms) {
1010                         pr_err("in-event: Internal tree error");
1011                         return -1;
1012                 }
1013                 if (add_sched_out_event(atoms, 'R', timestamp))
1014                         return -1;
1015         }
1016
1017         add_runtime_event(atoms, runtime, timestamp);
1018         return 0;
1019 }
1020
1021 static int latency_wakeup_event(struct perf_sched *sched,
1022                                 struct perf_evsel *evsel,
1023                                 struct perf_sample *sample,
1024                                 struct machine *machine)
1025 {
1026         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1027         struct work_atoms *atoms;
1028         struct work_atom *atom;
1029         struct thread *wakee;
1030         u64 timestamp = sample->time;
1031
1032         wakee = machine__findnew_thread(machine, -1, pid);
1033         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1034         if (!atoms) {
1035                 if (thread_atoms_insert(sched, wakee))
1036                         return -1;
1037                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1038                 if (!atoms) {
1039                         pr_err("wakeup-event: Internal tree error");
1040                         return -1;
1041                 }
1042                 if (add_sched_out_event(atoms, 'S', timestamp))
1043                         return -1;
1044         }
1045
1046         BUG_ON(list_empty(&atoms->work_list));
1047
1048         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1049
1050         /*
1051          * As we do not guarantee the wakeup event happens when
1052          * task is out of run queue, also may happen when task is
1053          * on run queue and wakeup only change ->state to TASK_RUNNING,
1054          * then we should not set the ->wake_up_time when wake up a
1055          * task which is on run queue.
1056          *
1057          * You WILL be missing events if you've recorded only
1058          * one CPU, or are only looking at only one, so don't
1059          * skip in this case.
1060          */
1061         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1062                 return 0;
1063
1064         sched->nr_timestamps++;
1065         if (atom->sched_out_time > timestamp) {
1066                 sched->nr_unordered_timestamps++;
1067                 return 0;
1068         }
1069
1070         atom->state = THREAD_WAIT_CPU;
1071         atom->wake_up_time = timestamp;
1072         return 0;
1073 }
1074
1075 static int latency_migrate_task_event(struct perf_sched *sched,
1076                                       struct perf_evsel *evsel,
1077                                       struct perf_sample *sample,
1078                                       struct machine *machine)
1079 {
1080         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1081         u64 timestamp = sample->time;
1082         struct work_atoms *atoms;
1083         struct work_atom *atom;
1084         struct thread *migrant;
1085
1086         /*
1087          * Only need to worry about migration when profiling one CPU.
1088          */
1089         if (sched->profile_cpu == -1)
1090                 return 0;
1091
1092         migrant = machine__findnew_thread(machine, -1, pid);
1093         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1094         if (!atoms) {
1095                 if (thread_atoms_insert(sched, migrant))
1096                         return -1;
1097                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1098                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1099                 if (!atoms) {
1100                         pr_err("migration-event: Internal tree error");
1101                         return -1;
1102                 }
1103                 if (add_sched_out_event(atoms, 'R', timestamp))
1104                         return -1;
1105         }
1106
1107         BUG_ON(list_empty(&atoms->work_list));
1108
1109         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1110         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1111
1112         sched->nr_timestamps++;
1113
1114         if (atom->sched_out_time > timestamp)
1115                 sched->nr_unordered_timestamps++;
1116
1117         return 0;
1118 }
1119
1120 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1121 {
1122         int i;
1123         int ret;
1124         u64 avg;
1125
1126         if (!work_list->nb_atoms)
1127                 return;
1128         /*
1129          * Ignore idle threads:
1130          */
1131         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1132                 return;
1133
1134         sched->all_runtime += work_list->total_runtime;
1135         sched->all_count   += work_list->nb_atoms;
1136
1137         ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1138
1139         for (i = 0; i < 24 - ret; i++)
1140                 printf(" ");
1141
1142         avg = work_list->total_lat / work_list->nb_atoms;
1143
1144         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1145               (double)work_list->total_runtime / 1e6,
1146                  work_list->nb_atoms, (double)avg / 1e6,
1147                  (double)work_list->max_lat / 1e6,
1148                  (double)work_list->max_lat_at / 1e9);
1149 }
1150
1151 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1152 {
1153         if (l->thread->tid < r->thread->tid)
1154                 return -1;
1155         if (l->thread->tid > r->thread->tid)
1156                 return 1;
1157
1158         return 0;
1159 }
1160
1161 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1162 {
1163         u64 avgl, avgr;
1164
1165         if (!l->nb_atoms)
1166                 return -1;
1167
1168         if (!r->nb_atoms)
1169                 return 1;
1170
1171         avgl = l->total_lat / l->nb_atoms;
1172         avgr = r->total_lat / r->nb_atoms;
1173
1174         if (avgl < avgr)
1175                 return -1;
1176         if (avgl > avgr)
1177                 return 1;
1178
1179         return 0;
1180 }
1181
1182 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1183 {
1184         if (l->max_lat < r->max_lat)
1185                 return -1;
1186         if (l->max_lat > r->max_lat)
1187                 return 1;
1188
1189         return 0;
1190 }
1191
1192 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1193 {
1194         if (l->nb_atoms < r->nb_atoms)
1195                 return -1;
1196         if (l->nb_atoms > r->nb_atoms)
1197                 return 1;
1198
1199         return 0;
1200 }
1201
1202 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1203 {
1204         if (l->total_runtime < r->total_runtime)
1205                 return -1;
1206         if (l->total_runtime > r->total_runtime)
1207                 return 1;
1208
1209         return 0;
1210 }
1211
1212 static int sort_dimension__add(const char *tok, struct list_head *list)
1213 {
1214         size_t i;
1215         static struct sort_dimension avg_sort_dimension = {
1216                 .name = "avg",
1217                 .cmp  = avg_cmp,
1218         };
1219         static struct sort_dimension max_sort_dimension = {
1220                 .name = "max",
1221                 .cmp  = max_cmp,
1222         };
1223         static struct sort_dimension pid_sort_dimension = {
1224                 .name = "pid",
1225                 .cmp  = pid_cmp,
1226         };
1227         static struct sort_dimension runtime_sort_dimension = {
1228                 .name = "runtime",
1229                 .cmp  = runtime_cmp,
1230         };
1231         static struct sort_dimension switch_sort_dimension = {
1232                 .name = "switch",
1233                 .cmp  = switch_cmp,
1234         };
1235         struct sort_dimension *available_sorts[] = {
1236                 &pid_sort_dimension,
1237                 &avg_sort_dimension,
1238                 &max_sort_dimension,
1239                 &switch_sort_dimension,
1240                 &runtime_sort_dimension,
1241         };
1242
1243         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1244                 if (!strcmp(available_sorts[i]->name, tok)) {
1245                         list_add_tail(&available_sorts[i]->list, list);
1246
1247                         return 0;
1248                 }
1249         }
1250
1251         return -1;
1252 }
1253
1254 static void perf_sched__sort_lat(struct perf_sched *sched)
1255 {
1256         struct rb_node *node;
1257
1258         for (;;) {
1259                 struct work_atoms *data;
1260                 node = rb_first(&sched->atom_root);
1261                 if (!node)
1262                         break;
1263
1264                 rb_erase(node, &sched->atom_root);
1265                 data = rb_entry(node, struct work_atoms, node);
1266                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1267         }
1268 }
1269
1270 static int process_sched_wakeup_event(struct perf_tool *tool,
1271                                       struct perf_evsel *evsel,
1272                                       struct perf_sample *sample,
1273                                       struct machine *machine)
1274 {
1275         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1276
1277         if (sched->tp_handler->wakeup_event)
1278                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1279
1280         return 0;
1281 }
1282
1283 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1284                             struct perf_sample *sample, struct machine *machine)
1285 {
1286         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1287         struct thread *sched_in;
1288         int new_shortname;
1289         u64 timestamp0, timestamp = sample->time;
1290         s64 delta;
1291         int cpu, this_cpu = sample->cpu;
1292
1293         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1294
1295         if (this_cpu > sched->max_cpu)
1296                 sched->max_cpu = this_cpu;
1297
1298         timestamp0 = sched->cpu_last_switched[this_cpu];
1299         sched->cpu_last_switched[this_cpu] = timestamp;
1300         if (timestamp0)
1301                 delta = timestamp - timestamp0;
1302         else
1303                 delta = 0;
1304
1305         if (delta < 0) {
1306                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1307                 return -1;
1308         }
1309
1310         sched_in = machine__findnew_thread(machine, -1, next_pid);
1311
1312         sched->curr_thread[this_cpu] = sched_in;
1313
1314         printf("  ");
1315
1316         new_shortname = 0;
1317         if (!sched_in->shortname[0]) {
1318                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1319                         /*
1320                          * Don't allocate a letter-number for swapper:0
1321                          * as a shortname. Instead, we use '.' for it.
1322                          */
1323                         sched_in->shortname[0] = '.';
1324                         sched_in->shortname[1] = ' ';
1325                 } else {
1326                         sched_in->shortname[0] = sched->next_shortname1;
1327                         sched_in->shortname[1] = sched->next_shortname2;
1328
1329                         if (sched->next_shortname1 < 'Z') {
1330                                 sched->next_shortname1++;
1331                         } else {
1332                                 sched->next_shortname1 = 'A';
1333                                 if (sched->next_shortname2 < '9')
1334                                         sched->next_shortname2++;
1335                                 else
1336                                         sched->next_shortname2 = '0';
1337                         }
1338                 }
1339                 new_shortname = 1;
1340         }
1341
1342         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1343                 if (cpu != this_cpu)
1344                         printf(" ");
1345                 else
1346                         printf("*");
1347
1348                 if (sched->curr_thread[cpu])
1349                         printf("%2s ", sched->curr_thread[cpu]->shortname);
1350                 else
1351                         printf("   ");
1352         }
1353
1354         printf("  %12.6f secs ", (double)timestamp/1e9);
1355         if (new_shortname) {
1356                 printf("%s => %s:%d\n",
1357                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1358         } else {
1359                 printf("\n");
1360         }
1361
1362         return 0;
1363 }
1364
1365 static int process_sched_switch_event(struct perf_tool *tool,
1366                                       struct perf_evsel *evsel,
1367                                       struct perf_sample *sample,
1368                                       struct machine *machine)
1369 {
1370         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1371         int this_cpu = sample->cpu, err = 0;
1372         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1373             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1374
1375         if (sched->curr_pid[this_cpu] != (u32)-1) {
1376                 /*
1377                  * Are we trying to switch away a PID that is
1378                  * not current?
1379                  */
1380                 if (sched->curr_pid[this_cpu] != prev_pid)
1381                         sched->nr_context_switch_bugs++;
1382         }
1383
1384         if (sched->tp_handler->switch_event)
1385                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1386
1387         sched->curr_pid[this_cpu] = next_pid;
1388         return err;
1389 }
1390
1391 static int process_sched_runtime_event(struct perf_tool *tool,
1392                                        struct perf_evsel *evsel,
1393                                        struct perf_sample *sample,
1394                                        struct machine *machine)
1395 {
1396         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1397
1398         if (sched->tp_handler->runtime_event)
1399                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1400
1401         return 0;
1402 }
1403
1404 static int perf_sched__process_fork_event(struct perf_tool *tool,
1405                                           union perf_event *event,
1406                                           struct perf_sample *sample,
1407                                           struct machine *machine)
1408 {
1409         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1410
1411         /* run the fork event through the perf machineruy */
1412         perf_event__process_fork(tool, event, sample, machine);
1413
1414         /* and then run additional processing needed for this command */
1415         if (sched->tp_handler->fork_event)
1416                 return sched->tp_handler->fork_event(sched, event, machine);
1417
1418         return 0;
1419 }
1420
1421 static int process_sched_migrate_task_event(struct perf_tool *tool,
1422                                             struct perf_evsel *evsel,
1423                                             struct perf_sample *sample,
1424                                             struct machine *machine)
1425 {
1426         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1427
1428         if (sched->tp_handler->migrate_task_event)
1429                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1430
1431         return 0;
1432 }
1433
1434 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1435                                   struct perf_evsel *evsel,
1436                                   struct perf_sample *sample,
1437                                   struct machine *machine);
1438
1439 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1440                                                  union perf_event *event __maybe_unused,
1441                                                  struct perf_sample *sample,
1442                                                  struct perf_evsel *evsel,
1443                                                  struct machine *machine)
1444 {
1445         int err = 0;
1446
1447         if (evsel->handler != NULL) {
1448                 tracepoint_handler f = evsel->handler;
1449                 err = f(tool, evsel, sample, machine);
1450         }
1451
1452         return err;
1453 }
1454
1455 static int perf_sched__read_events(struct perf_sched *sched)
1456 {
1457         const struct perf_evsel_str_handler handlers[] = {
1458                 { "sched:sched_switch",       process_sched_switch_event, },
1459                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1460                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1461                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1462                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1463         };
1464         struct perf_session *session;
1465         struct perf_data_file file = {
1466                 .path = input_name,
1467                 .mode = PERF_DATA_MODE_READ,
1468         };
1469         int rc = -1;
1470
1471         session = perf_session__new(&file, false, &sched->tool);
1472         if (session == NULL) {
1473                 pr_debug("No Memory for session\n");
1474                 return -1;
1475         }
1476
1477         symbol__init(&session->header.env);
1478
1479         if (perf_session__set_tracepoints_handlers(session, handlers))
1480                 goto out_delete;
1481
1482         if (perf_session__has_traces(session, "record -R")) {
1483                 int err = perf_session__process_events(session);
1484                 if (err) {
1485                         pr_err("Failed to process events, error %d", err);
1486                         goto out_delete;
1487                 }
1488
1489                 sched->nr_events      = session->evlist->stats.nr_events[0];
1490                 sched->nr_lost_events = session->evlist->stats.total_lost;
1491                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1492         }
1493
1494         rc = 0;
1495 out_delete:
1496         perf_session__delete(session);
1497         return rc;
1498 }
1499
1500 static void print_bad_events(struct perf_sched *sched)
1501 {
1502         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1503                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1504                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1505                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1506         }
1507         if (sched->nr_lost_events && sched->nr_events) {
1508                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1509                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1510                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1511         }
1512         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1513                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1514                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1515                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1516                 if (sched->nr_lost_events)
1517                         printf(" (due to lost events?)");
1518                 printf("\n");
1519         }
1520 }
1521
1522 static int perf_sched__lat(struct perf_sched *sched)
1523 {
1524         struct rb_node *next;
1525
1526         setup_pager();
1527
1528         if (perf_sched__read_events(sched))
1529                 return -1;
1530
1531         perf_sched__sort_lat(sched);
1532
1533         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1534         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1535         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1536
1537         next = rb_first(&sched->sorted_atom_root);
1538
1539         while (next) {
1540                 struct work_atoms *work_list;
1541
1542                 work_list = rb_entry(next, struct work_atoms, node);
1543                 output_lat_thread(sched, work_list);
1544                 next = rb_next(next);
1545                 thread__zput(work_list->thread);
1546         }
1547
1548         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1549         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1550                 (double)sched->all_runtime / 1e6, sched->all_count);
1551
1552         printf(" ---------------------------------------------------\n");
1553
1554         print_bad_events(sched);
1555         printf("\n");
1556
1557         return 0;
1558 }
1559
1560 static int perf_sched__map(struct perf_sched *sched)
1561 {
1562         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1563
1564         setup_pager();
1565         if (perf_sched__read_events(sched))
1566                 return -1;
1567         print_bad_events(sched);
1568         return 0;
1569 }
1570
1571 static int perf_sched__replay(struct perf_sched *sched)
1572 {
1573         unsigned long i;
1574
1575         calibrate_run_measurement_overhead(sched);
1576         calibrate_sleep_measurement_overhead(sched);
1577
1578         test_calibrations(sched);
1579
1580         if (perf_sched__read_events(sched))
1581                 return -1;
1582
1583         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1584         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1585         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1586
1587         if (sched->targetless_wakeups)
1588                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1589         if (sched->multitarget_wakeups)
1590                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1591         if (sched->nr_run_events_optimized)
1592                 printf("run atoms optimized: %ld\n",
1593                         sched->nr_run_events_optimized);
1594
1595         print_task_traces(sched);
1596         add_cross_task_wakeups(sched);
1597
1598         create_tasks(sched);
1599         printf("------------------------------------------------------------\n");
1600         for (i = 0; i < sched->replay_repeat; i++)
1601                 run_one_test(sched);
1602
1603         return 0;
1604 }
1605
1606 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1607                           const char * const usage_msg[])
1608 {
1609         char *tmp, *tok, *str = strdup(sched->sort_order);
1610
1611         for (tok = strtok_r(str, ", ", &tmp);
1612                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1613                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1614                         error("Unknown --sort key: `%s'", tok);
1615                         usage_with_options(usage_msg, options);
1616                 }
1617         }
1618
1619         free(str);
1620
1621         sort_dimension__add("pid", &sched->cmp_pid);
1622 }
1623
1624 static int __cmd_record(int argc, const char **argv)
1625 {
1626         unsigned int rec_argc, i, j;
1627         const char **rec_argv;
1628         const char * const record_args[] = {
1629                 "record",
1630                 "-a",
1631                 "-R",
1632                 "-m", "1024",
1633                 "-c", "1",
1634                 "-e", "sched:sched_switch",
1635                 "-e", "sched:sched_stat_wait",
1636                 "-e", "sched:sched_stat_sleep",
1637                 "-e", "sched:sched_stat_iowait",
1638                 "-e", "sched:sched_stat_runtime",
1639                 "-e", "sched:sched_process_fork",
1640                 "-e", "sched:sched_wakeup",
1641                 "-e", "sched:sched_wakeup_new",
1642                 "-e", "sched:sched_migrate_task",
1643         };
1644
1645         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1646         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1647
1648         if (rec_argv == NULL)
1649                 return -ENOMEM;
1650
1651         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1652                 rec_argv[i] = strdup(record_args[i]);
1653
1654         for (j = 1; j < (unsigned int)argc; j++, i++)
1655                 rec_argv[i] = argv[j];
1656
1657         BUG_ON(i != rec_argc);
1658
1659         return cmd_record(i, rec_argv, NULL);
1660 }
1661
1662 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1663 {
1664         const char default_sort_order[] = "avg, max, switch, runtime";
1665         struct perf_sched sched = {
1666                 .tool = {
1667                         .sample          = perf_sched__process_tracepoint_sample,
1668                         .comm            = perf_event__process_comm,
1669                         .lost            = perf_event__process_lost,
1670                         .fork            = perf_sched__process_fork_event,
1671                         .ordered_events = true,
1672                 },
1673                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1674                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1675                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1676                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1677                 .sort_order           = default_sort_order,
1678                 .replay_repeat        = 10,
1679                 .profile_cpu          = -1,
1680                 .next_shortname1      = 'A',
1681                 .next_shortname2      = '0',
1682         };
1683         const struct option latency_options[] = {
1684         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1685                    "sort by key(s): runtime, switch, avg, max"),
1686         OPT_INCR('v', "verbose", &verbose,
1687                     "be more verbose (show symbol address, etc)"),
1688         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1689                     "CPU to profile on"),
1690         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1691                     "dump raw trace in ASCII"),
1692         OPT_END()
1693         };
1694         const struct option replay_options[] = {
1695         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1696                      "repeat the workload replay N times (-1: infinite)"),
1697         OPT_INCR('v', "verbose", &verbose,
1698                     "be more verbose (show symbol address, etc)"),
1699         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1700                     "dump raw trace in ASCII"),
1701         OPT_END()
1702         };
1703         const struct option sched_options[] = {
1704         OPT_STRING('i', "input", &input_name, "file",
1705                     "input file name"),
1706         OPT_INCR('v', "verbose", &verbose,
1707                     "be more verbose (show symbol address, etc)"),
1708         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1709                     "dump raw trace in ASCII"),
1710         OPT_END()
1711         };
1712         const char * const latency_usage[] = {
1713                 "perf sched latency [<options>]",
1714                 NULL
1715         };
1716         const char * const replay_usage[] = {
1717                 "perf sched replay [<options>]",
1718                 NULL
1719         };
1720         const char *const sched_subcommands[] = { "record", "latency", "map",
1721                                                   "replay", "script", NULL };
1722         const char *sched_usage[] = {
1723                 NULL,
1724                 NULL
1725         };
1726         struct trace_sched_handler lat_ops  = {
1727                 .wakeup_event       = latency_wakeup_event,
1728                 .switch_event       = latency_switch_event,
1729                 .runtime_event      = latency_runtime_event,
1730                 .migrate_task_event = latency_migrate_task_event,
1731         };
1732         struct trace_sched_handler map_ops  = {
1733                 .switch_event       = map_switch_event,
1734         };
1735         struct trace_sched_handler replay_ops  = {
1736                 .wakeup_event       = replay_wakeup_event,
1737                 .switch_event       = replay_switch_event,
1738                 .fork_event         = replay_fork_event,
1739         };
1740         unsigned int i;
1741
1742         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1743                 sched.curr_pid[i] = -1;
1744
1745         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1746                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1747         if (!argc)
1748                 usage_with_options(sched_usage, sched_options);
1749
1750         /*
1751          * Aliased to 'perf script' for now:
1752          */
1753         if (!strcmp(argv[0], "script"))
1754                 return cmd_script(argc, argv, prefix);
1755
1756         if (!strncmp(argv[0], "rec", 3)) {
1757                 return __cmd_record(argc, argv);
1758         } else if (!strncmp(argv[0], "lat", 3)) {
1759                 sched.tp_handler = &lat_ops;
1760                 if (argc > 1) {
1761                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1762                         if (argc)
1763                                 usage_with_options(latency_usage, latency_options);
1764                 }
1765                 setup_sorting(&sched, latency_options, latency_usage);
1766                 return perf_sched__lat(&sched);
1767         } else if (!strcmp(argv[0], "map")) {
1768                 sched.tp_handler = &map_ops;
1769                 setup_sorting(&sched, latency_options, latency_usage);
1770                 return perf_sched__map(&sched);
1771         } else if (!strncmp(argv[0], "rep", 3)) {
1772                 sched.tp_handler = &replay_ops;
1773                 if (argc) {
1774                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1775                         if (argc)
1776                                 usage_with_options(replay_usage, replay_options);
1777                 }
1778                 return perf_sched__replay(&sched);
1779         } else {
1780                 usage_with_options(sched_usage, sched_options);
1781         }
1782
1783         return 0;
1784 }