tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/cache.h"
   6 #include "util/symbol.h"
   7 #include "util/thread.h"
   8 #include "util/header.h"
   9 #include "util/session.h"
  10
  11 #include "util/parse-options.h"
  12 #include "util/trace-event.h"
  13
  14 #include "util/debug.h"
  15
  16 #include <sys/prctl.h>
  17 #include <sys/resource.h>
  18
  19 #include <semaphore.h>
  20 #include <pthread.h>
  21 #include <math.h>
  22
  23 static char                     const *input_name = "perf.data";
  24
  25 static char                     default_sort_order[] = "avg, max, switch, runtime";
  26 static const char               *sort_order = default_sort_order;
  27
  28 static int                      profile_cpu = -1;
  29
  30 #define PR_SET_NAME             15               /* Set process name */
  31 #define MAX_CPUS                4096
  32
  33 static u64                      run_measurement_overhead;
  34 static u64                      sleep_measurement_overhead;
  35
  36 #define COMM_LEN                20
  37 #define SYM_LEN                 129
  38
  39 #define MAX_PID                 65536
  40
  41 static unsigned long            nr_tasks;
  42
  43 struct sched_atom;
  44
  45 struct task_desc {
  46         unsigned long           nr;
  47         unsigned long           pid;
  48         char                    comm[COMM_LEN];
  49
  50         unsigned long           nr_events;
  51         unsigned long           curr_event;
  52         struct sched_atom       **atoms;
  53
  54         pthread_t               thread;
  55         sem_t                   sleep_sem;
  56
  57         sem_t                   ready_for_work;
  58         sem_t                   work_done_sem;
  59
  60         u64                     cpu_usage;
  61 };
  62
  63 enum sched_event_type {
  64         SCHED_EVENT_RUN,
  65         SCHED_EVENT_SLEEP,
  66         SCHED_EVENT_WAKEUP,
  67         SCHED_EVENT_MIGRATION,
  68 };
  69
  70 struct sched_atom {
  71         enum sched_event_type   type;
  72         int                     specific_wait;
  73         u64                     timestamp;
  74         u64                     duration;
  75         unsigned long           nr;
  76         sem_t                   *wait_sem;
  77         struct task_desc        *wakee;
  78 };
  79
  80 static struct task_desc         *pid_to_task[MAX_PID];
  81
  82 static struct task_desc         **tasks;
  83
  84 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  85 static u64                      start_time;
  86
  87 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  88
  89 static unsigned long            nr_run_events;
  90 static unsigned long            nr_sleep_events;
  91 static unsigned long            nr_wakeup_events;
  92
  93 static unsigned long            nr_sleep_corrections;
  94 static unsigned long            nr_run_events_optimized;
  95
  96 static unsigned long            targetless_wakeups;
  97 static unsigned long            multitarget_wakeups;
  98
  99 static u64                      cpu_usage;
 100 static u64                      runavg_cpu_usage;
 101 static u64                      parent_cpu_usage;
 102 static u64                      runavg_parent_cpu_usage;
 103
 104 static unsigned long            nr_runs;
 105 static u64                      sum_runtime;
 106 static u64                      sum_fluct;
 107 static u64                      run_avg;
 108
 109 static unsigned int             replay_repeat = 10;
 110 static unsigned long            nr_timestamps;
 111 static unsigned long            nr_unordered_timestamps;
 112 static unsigned long            nr_state_machine_bugs;
 113 static unsigned long            nr_context_switch_bugs;
 114 static unsigned long            nr_events;
 115 static unsigned long            nr_lost_chunks;
 116 static unsigned long            nr_lost_events;
 117
 118 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 119
 120 enum thread_state {
 121         THREAD_SLEEPING = 0,
 122         THREAD_WAIT_CPU,
 123         THREAD_SCHED_IN,
 124         THREAD_IGNORE
 125 };
 126
 127 struct work_atom {
 128         struct list_head        list;
 129         enum thread_state       state;
 130         u64                     sched_out_time;
 131         u64                     wake_up_time;
 132         u64                     sched_in_time;
 133         u64                     runtime;
 134 };
 135
 136 struct work_atoms {
 137         struct list_head        work_list;
 138         struct thread           *thread;
 139         struct rb_node          node;
 140         u64                     max_lat;
 141         u64                     max_lat_at;
 142         u64                     total_lat;
 143         u64                     nb_atoms;
 144         u64                     total_runtime;
 145 };
 146
 147 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 148
 149 static struct rb_root           atom_root, sorted_atom_root;
 150
 151 static u64                      all_runtime;
 152 static u64                      all_count;
 153
 154
 155 static u64 get_nsecs(void)
 156 {
 157         struct timespec ts;
 158
 159         clock_gettime(CLOCK_MONOTONIC, &ts);
 160
 161         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 162 }
 163
 164 static void burn_nsecs(u64 nsecs)
 165 {
 166         u64 T0 = get_nsecs(), T1;
 167
 168         do {
 169                 T1 = get_nsecs();
 170         } while (T1 + run_measurement_overhead < T0 + nsecs);
 171 }
 172
 173 static void sleep_nsecs(u64 nsecs)
 174 {
 175         struct timespec ts;
 176
 177         ts.tv_nsec = nsecs % 999999999;
 178         ts.tv_sec = nsecs / 999999999;
 179
 180         nanosleep(&ts, NULL);
 181 }
 182
 183 static void calibrate_run_measurement_overhead(void)
 184 {
 185         u64 T0, T1, delta, min_delta = 1000000000ULL;
 186         int i;
 187
 188         for (i = 0; i < 10; i++) {
 189                 T0 = get_nsecs();
 190                 burn_nsecs(0);
 191                 T1 = get_nsecs();
 192                 delta = T1-T0;
 193                 min_delta = min(min_delta, delta);
 194         }
 195         run_measurement_overhead = min_delta;
 196
 197         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 198 }
 199
 200 static void calibrate_sleep_measurement_overhead(void)
 201 {
 202         u64 T0, T1, delta, min_delta = 1000000000ULL;
 203         int i;
 204
 205         for (i = 0; i < 10; i++) {
 206                 T0 = get_nsecs();
 207                 sleep_nsecs(10000);
 208                 T1 = get_nsecs();
 209                 delta = T1-T0;
 210                 min_delta = min(min_delta, delta);
 211         }
 212         min_delta -= 10000;
 213         sleep_measurement_overhead = min_delta;
 214
 215         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 216 }
 217
 218 static struct sched_atom *
 219 get_new_event(struct task_desc *task, u64 timestamp)
 220 {
 221         struct sched_atom *event = zalloc(sizeof(*event));
 222         unsigned long idx = task->nr_events;
 223         size_t size;
 224
 225         event->timestamp = timestamp;
 226         event->nr = idx;
 227
 228         task->nr_events++;
 229         size = sizeof(struct sched_atom *) * task->nr_events;
 230         task->atoms = realloc(task->atoms, size);
 231         BUG_ON(!task->atoms);
 232
 233         task->atoms[idx] = event;
 234
 235         return event;
 236 }
 237
 238 static struct sched_atom *last_event(struct task_desc *task)
 239 {
 240         if (!task->nr_events)
 241                 return NULL;
 242
 243         return task->atoms[task->nr_events - 1];
 244 }
 245
 246 static void
 247 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 248 {
 249         struct sched_atom *event, *curr_event = last_event(task);
 250
 251         /*
 252          * optimize an existing RUN event by merging this one
 253          * to it:
 254          */
 255         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 256                 nr_run_events_optimized++;
 257                 curr_event->duration += duration;
 258                 return;
 259         }
 260
 261         event = get_new_event(task, timestamp);
 262
 263         event->type = SCHED_EVENT_RUN;
 264         event->duration = duration;
 265
 266         nr_run_events++;
 267 }
 268
 269 static void
 270 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 271                        struct task_desc *wakee)
 272 {
 273         struct sched_atom *event, *wakee_event;
 274
 275         event = get_new_event(task, timestamp);
 276         event->type = SCHED_EVENT_WAKEUP;
 277         event->wakee = wakee;
 278
 279         wakee_event = last_event(wakee);
 280         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 281                 targetless_wakeups++;
 282                 return;
 283         }
 284         if (wakee_event->wait_sem) {
 285                 multitarget_wakeups++;
 286                 return;
 287         }
 288
 289         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 290         sem_init(wakee_event->wait_sem, 0, 0);
 291         wakee_event->specific_wait = 1;
 292         event->wait_sem = wakee_event->wait_sem;
 293
 294         nr_wakeup_events++;
 295 }
 296
 297 static void
 298 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 299                       u64 task_state __used)
 300 {
 301         struct sched_atom *event = get_new_event(task, timestamp);
 302
 303         event->type = SCHED_EVENT_SLEEP;
 304
 305         nr_sleep_events++;
 306 }
 307
 308 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 309 {
 310         struct task_desc *task;
 311
 312         BUG_ON(pid >= MAX_PID);
 313
 314         task = pid_to_task[pid];
 315
 316         if (task)
 317                 return task;
 318
 319         task = zalloc(sizeof(*task));
 320         task->pid = pid;
 321         task->nr = nr_tasks;
 322         strcpy(task->comm, comm);
 323         /*
 324          * every task starts in sleeping state - this gets ignored
 325          * if there's no wakeup pointing to this sleep state:
 326          */
 327         add_sched_event_sleep(task, 0, 0);
 328
 329         pid_to_task[pid] = task;
 330         nr_tasks++;
 331         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 332         BUG_ON(!tasks);
 333         tasks[task->nr] = task;
 334
 335         if (verbose)
 336                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 337
 338         return task;
 339 }
 340
 341
 342 static void print_task_traces(void)
 343 {
 344         struct task_desc *task;
 345         unsigned long i;
 346
 347         for (i = 0; i < nr_tasks; i++) {
 348                 task = tasks[i];
 349                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 350                         task->nr, task->comm, task->pid, task->nr_events);
 351         }
 352 }
 353
 354 static void add_cross_task_wakeups(void)
 355 {
 356         struct task_desc *task1, *task2;
 357         unsigned long i, j;
 358
 359         for (i = 0; i < nr_tasks; i++) {
 360                 task1 = tasks[i];
 361                 j = i + 1;
 362                 if (j == nr_tasks)
 363                         j = 0;
 364                 task2 = tasks[j];
 365                 add_sched_event_wakeup(task1, 0, task2);
 366         }
 367 }
 368
 369 static void
 370 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 371 {
 372         int ret = 0;
 373
 374         switch (atom->type) {
 375                 case SCHED_EVENT_RUN:
 376                         burn_nsecs(atom->duration);
 377                         break;
 378                 case SCHED_EVENT_SLEEP:
 379                         if (atom->wait_sem)
 380                                 ret = sem_wait(atom->wait_sem);
 381                         BUG_ON(ret);
 382                         break;
 383                 case SCHED_EVENT_WAKEUP:
 384                         if (atom->wait_sem)
 385                                 ret = sem_post(atom->wait_sem);
 386                         BUG_ON(ret);
 387                         break;
 388                 case SCHED_EVENT_MIGRATION:
 389                         break;
 390                 default:
 391                         BUG_ON(1);
 392         }
 393 }
 394
 395 static u64 get_cpu_usage_nsec_parent(void)
 396 {
 397         struct rusage ru;
 398         u64 sum;
 399         int err;
 400
 401         err = getrusage(RUSAGE_SELF, &ru);
 402         BUG_ON(err);
 403
 404         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 405         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 406
 407         return sum;
 408 }
 409
 410 static int self_open_counters(void)
 411 {
 412         struct perf_event_attr attr;
 413         int fd;
 414
 415         memset(&attr, 0, sizeof(attr));
 416
 417         attr.type = PERF_TYPE_SOFTWARE;
 418         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 419
 420         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 421
 422         if (fd < 0)
 423                 die("Error: sys_perf_event_open() syscall returned"
 424                     "with %d (%s)\n", fd, strerror(errno));
 425         return fd;
 426 }
 427
 428 static u64 get_cpu_usage_nsec_self(int fd)
 429 {
 430         u64 runtime;
 431         int ret;
 432
 433         ret = read(fd, &runtime, sizeof(runtime));
 434         BUG_ON(ret != sizeof(runtime));
 435
 436         return runtime;
 437 }
 438
 439 static void *thread_func(void *ctx)
 440 {
 441         struct task_desc *this_task = ctx;
 442         u64 cpu_usage_0, cpu_usage_1;
 443         unsigned long i, ret;
 444         char comm2[22];
 445         int fd;
 446
 447         sprintf(comm2, ":%s", this_task->comm);
 448         prctl(PR_SET_NAME, comm2);
 449         fd = self_open_counters();
 450
 451 again:
 452         ret = sem_post(&this_task->ready_for_work);
 453         BUG_ON(ret);
 454         ret = pthread_mutex_lock(&start_work_mutex);
 455         BUG_ON(ret);
 456         ret = pthread_mutex_unlock(&start_work_mutex);
 457         BUG_ON(ret);
 458
 459         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 460
 461         for (i = 0; i < this_task->nr_events; i++) {
 462                 this_task->curr_event = i;
 463                 process_sched_event(this_task, this_task->atoms[i]);
 464         }
 465
 466         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 467         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 468         ret = sem_post(&this_task->work_done_sem);
 469         BUG_ON(ret);
 470
 471         ret = pthread_mutex_lock(&work_done_wait_mutex);
 472         BUG_ON(ret);
 473         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 474         BUG_ON(ret);
 475
 476         goto again;
 477 }
 478
 479 static void create_tasks(void)
 480 {
 481         struct task_desc *task;
 482         pthread_attr_t attr;
 483         unsigned long i;
 484         int err;
 485
 486         err = pthread_attr_init(&attr);
 487         BUG_ON(err);
 488         err = pthread_attr_setstacksize(&attr,
 489                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 490         BUG_ON(err);
 491         err = pthread_mutex_lock(&start_work_mutex);
 492         BUG_ON(err);
 493         err = pthread_mutex_lock(&work_done_wait_mutex);
 494         BUG_ON(err);
 495         for (i = 0; i < nr_tasks; i++) {
 496                 task = tasks[i];
 497                 sem_init(&task->sleep_sem, 0, 0);
 498                 sem_init(&task->ready_for_work, 0, 0);
 499                 sem_init(&task->work_done_sem, 0, 0);
 500                 task->curr_event = 0;
 501                 err = pthread_create(&task->thread, &attr, thread_func, task);
 502                 BUG_ON(err);
 503         }
 504 }
 505
 506 static void wait_for_tasks(void)
 507 {
 508         u64 cpu_usage_0, cpu_usage_1;
 509         struct task_desc *task;
 510         unsigned long i, ret;
 511
 512         start_time = get_nsecs();
 513         cpu_usage = 0;
 514         pthread_mutex_unlock(&work_done_wait_mutex);
 515
 516         for (i = 0; i < nr_tasks; i++) {
 517                 task = tasks[i];
 518                 ret = sem_wait(&task->ready_for_work);
 519                 BUG_ON(ret);
 520                 sem_init(&task->ready_for_work, 0, 0);
 521         }
 522         ret = pthread_mutex_lock(&work_done_wait_mutex);
 523         BUG_ON(ret);
 524
 525         cpu_usage_0 = get_cpu_usage_nsec_parent();
 526
 527         pthread_mutex_unlock(&start_work_mutex);
 528
 529         for (i = 0; i < nr_tasks; i++) {
 530                 task = tasks[i];
 531                 ret = sem_wait(&task->work_done_sem);
 532                 BUG_ON(ret);
 533                 sem_init(&task->work_done_sem, 0, 0);
 534                 cpu_usage += task->cpu_usage;
 535                 task->cpu_usage = 0;
 536         }
 537
 538         cpu_usage_1 = get_cpu_usage_nsec_parent();
 539         if (!runavg_cpu_usage)
 540                 runavg_cpu_usage = cpu_usage;
 541         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 542
 543         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 544         if (!runavg_parent_cpu_usage)
 545                 runavg_parent_cpu_usage = parent_cpu_usage;
 546         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 547                                    parent_cpu_usage)/10;
 548
 549         ret = pthread_mutex_lock(&start_work_mutex);
 550         BUG_ON(ret);
 551
 552         for (i = 0; i < nr_tasks; i++) {
 553                 task = tasks[i];
 554                 sem_init(&task->sleep_sem, 0, 0);
 555                 task->curr_event = 0;
 556         }
 557 }
 558
 559 static void run_one_test(void)
 560 {
 561         u64 T0, T1, delta, avg_delta, fluct;
 562
 563         T0 = get_nsecs();
 564         wait_for_tasks();
 565         T1 = get_nsecs();
 566
 567         delta = T1 - T0;
 568         sum_runtime += delta;
 569         nr_runs++;
 570
 571         avg_delta = sum_runtime / nr_runs;
 572         if (delta < avg_delta)
 573                 fluct = avg_delta - delta;
 574         else
 575                 fluct = delta - avg_delta;
 576         sum_fluct += fluct;
 577         if (!run_avg)
 578                 run_avg = delta;
 579         run_avg = (run_avg*9 + delta)/10;
 580
 581         printf("#%-3ld: %0.3f, ",
 582                 nr_runs, (double)delta/1000000.0);
 583
 584         printf("ravg: %0.2f, ",
 585                 (double)run_avg/1e6);
 586
 587         printf("cpu: %0.2f / %0.2f",
 588                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 589
 590 #if 0
 591         /*
 592          * rusage statistics done by the parent, these are less
 593          * accurate than the sum_exec_runtime based statistics:
 594          */
 595         printf(" [%0.2f / %0.2f]",
 596                 (double)parent_cpu_usage/1e6,
 597                 (double)runavg_parent_cpu_usage/1e6);
 598 #endif
 599
 600         printf("\n");
 601
 602         if (nr_sleep_corrections)
 603                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 604         nr_sleep_corrections = 0;
 605 }
 606
 607 static void test_calibrations(void)
 608 {
 609         u64 T0, T1;
 610
 611         T0 = get_nsecs();
 612         burn_nsecs(1e6);
 613         T1 = get_nsecs();
 614
 615         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 616
 617         T0 = get_nsecs();
 618         sleep_nsecs(1e6);
 619         T1 = get_nsecs();
 620
 621         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 622 }
 623
 624 #define FILL_FIELD(ptr, field, event, data)     \
 625         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 626
 627 #define FILL_ARRAY(ptr, array, event, data)                     \
 628 do {                                                            \
 629         void *__array = raw_field_ptr(event, #array, data);     \
 630         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 631 } while(0)
 632
 633 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 634 do {                                                            \
 635         FILL_FIELD(ptr, common_type, event, data);              \
 636         FILL_FIELD(ptr, common_flags, event, data);             \
 637         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 638         FILL_FIELD(ptr, common_pid, event, data);               \
 639         FILL_FIELD(ptr, common_tgid, event, data);              \
 640 } while (0)
 641
 642
 643
 644 struct trace_switch_event {
 645         u32 size;
 646
 647         u16 common_type;
 648         u8 common_flags;
 649         u8 common_preempt_count;
 650         u32 common_pid;
 651         u32 common_tgid;
 652
 653         char prev_comm[16];
 654         u32 prev_pid;
 655         u32 prev_prio;
 656         u64 prev_state;
 657         char next_comm[16];
 658         u32 next_pid;
 659         u32 next_prio;
 660 };
 661
 662 struct trace_runtime_event {
 663         u32 size;
 664
 665         u16 common_type;
 666         u8 common_flags;
 667         u8 common_preempt_count;
 668         u32 common_pid;
 669         u32 common_tgid;
 670
 671         char comm[16];
 672         u32 pid;
 673         u64 runtime;
 674         u64 vruntime;
 675 };
 676
 677 struct trace_wakeup_event {
 678         u32 size;
 679
 680         u16 common_type;
 681         u8 common_flags;
 682         u8 common_preempt_count;
 683         u32 common_pid;
 684         u32 common_tgid;
 685
 686         char comm[16];
 687         u32 pid;
 688
 689         u32 prio;
 690         u32 success;
 691         u32 cpu;
 692 };
 693
 694 struct trace_fork_event {
 695         u32 size;
 696
 697         u16 common_type;
 698         u8 common_flags;
 699         u8 common_preempt_count;
 700         u32 common_pid;
 701         u32 common_tgid;
 702
 703         char parent_comm[16];
 704         u32 parent_pid;
 705         char child_comm[16];
 706         u32 child_pid;
 707 };
 708
 709 struct trace_migrate_task_event {
 710         u32 size;
 711
 712         u16 common_type;
 713         u8 common_flags;
 714         u8 common_preempt_count;
 715         u32 common_pid;
 716         u32 common_tgid;
 717
 718         char comm[16];
 719         u32 pid;
 720
 721         u32 prio;
 722         u32 cpu;
 723 };
 724
 725 struct trace_sched_handler {
 726         void (*switch_event)(struct trace_switch_event *,
 727                              struct perf_session *,
 728                              struct event *,
 729                              int cpu,
 730                              u64 timestamp,
 731                              struct thread *thread);
 732
 733         void (*runtime_event)(struct trace_runtime_event *,
 734                               struct perf_session *,
 735                               struct event *,
 736                               int cpu,
 737                               u64 timestamp,
 738                               struct thread *thread);
 739
 740         void (*wakeup_event)(struct trace_wakeup_event *,
 741                              struct perf_session *,
 742                              struct event *,
 743                              int cpu,
 744                              u64 timestamp,
 745                              struct thread *thread);
 746
 747         void (*fork_event)(struct trace_fork_event *,
 748                            struct event *,
 749                            int cpu,
 750                            u64 timestamp,
 751                            struct thread *thread);
 752
 753         void (*migrate_task_event)(struct trace_migrate_task_event *,
 754                            struct perf_session *session,
 755                            struct event *,
 756                            int cpu,
 757                            u64 timestamp,
 758                            struct thread *thread);
 759 };
 760
 761
 762 static void
 763 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 764                     struct perf_session *session __used,
 765                     struct event *event,
 766                     int cpu __used,
 767                     u64 timestamp __used,
 768                     struct thread *thread __used)
 769 {
 770         struct task_desc *waker, *wakee;
 771
 772         if (verbose) {
 773                 printf("sched_wakeup event %p\n", event);
 774
 775                 printf(" ... pid %d woke up %s/%d\n",
 776                         wakeup_event->common_pid,
 777                         wakeup_event->comm,
 778                         wakeup_event->pid);
 779         }
 780
 781         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 782         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 783
 784         add_sched_event_wakeup(waker, timestamp, wakee);
 785 }
 786
 787 static u64 cpu_last_switched[MAX_CPUS];
 788
 789 static void
 790 replay_switch_event(struct trace_switch_event *switch_event,
 791                     struct perf_session *session __used,
 792                     struct event *event,
 793                     int cpu,
 794                     u64 timestamp,
 795                     struct thread *thread __used)
 796 {
 797         struct task_desc *prev, __used *next;
 798         u64 timestamp0;
 799         s64 delta;
 800
 801         if (verbose)
 802                 printf("sched_switch event %p\n", event);
 803
 804         if (cpu >= MAX_CPUS || cpu < 0)
 805                 return;
 806
 807         timestamp0 = cpu_last_switched[cpu];
 808         if (timestamp0)
 809                 delta = timestamp - timestamp0;
 810         else
 811                 delta = 0;
 812
 813         if (delta < 0)
 814                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 815
 816         if (verbose) {
 817                 printf(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 818                         switch_event->prev_comm, switch_event->prev_pid,
 819                         switch_event->next_comm, switch_event->next_pid,
 820                         delta);
 821         }
 822
 823         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 824         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 825
 826         cpu_last_switched[cpu] = timestamp;
 827
 828         add_sched_event_run(prev, timestamp, delta);
 829         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 830 }
 831
 832
 833 static void
 834 replay_fork_event(struct trace_fork_event *fork_event,
 835                   struct event *event,
 836                   int cpu __used,
 837                   u64 timestamp __used,
 838                   struct thread *thread __used)
 839 {
 840         if (verbose) {
 841                 printf("sched_fork event %p\n", event);
 842                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 843                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 844         }
 845         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 846         register_pid(fork_event->child_pid, fork_event->child_comm);
 847 }
 848
 849 static struct trace_sched_handler replay_ops  = {
 850         .wakeup_event           = replay_wakeup_event,
 851         .switch_event           = replay_switch_event,
 852         .fork_event             = replay_fork_event,
 853 };
 854
 855 struct sort_dimension {
 856         const char              *name;
 857         sort_fn_t               cmp;
 858         struct list_head        list;
 859 };
 860
 861 static LIST_HEAD(cmp_pid);
 862
 863 static int
 864 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 865 {
 866         struct sort_dimension *sort;
 867         int ret = 0;
 868
 869         BUG_ON(list_empty(list));
 870
 871         list_for_each_entry(sort, list, list) {
 872                 ret = sort->cmp(l, r);
 873                 if (ret)
 874                         return ret;
 875         }
 876
 877         return ret;
 878 }
 879
 880 static struct work_atoms *
 881 thread_atoms_search(struct rb_root *root, struct thread *thread,
 882                          struct list_head *sort_list)
 883 {
 884         struct rb_node *node = root->rb_node;
 885         struct work_atoms key = { .thread = thread };
 886
 887         while (node) {
 888                 struct work_atoms *atoms;
 889                 int cmp;
 890
 891                 atoms = container_of(node, struct work_atoms, node);
 892
 893                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 894                 if (cmp > 0)
 895                         node = node->rb_left;
 896                 else if (cmp < 0)
 897                         node = node->rb_right;
 898                 else {
 899                         BUG_ON(thread != atoms->thread);
 900                         return atoms;
 901                 }
 902         }
 903         return NULL;
 904 }
 905
 906 static void
 907 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 908                          struct list_head *sort_list)
 909 {
 910         struct rb_node **new = &(root->rb_node), *parent = NULL;
 911
 912         while (*new) {
 913                 struct work_atoms *this;
 914                 int cmp;
 915
 916                 this = container_of(*new, struct work_atoms, node);
 917                 parent = *new;
 918
 919                 cmp = thread_lat_cmp(sort_list, data, this);
 920
 921                 if (cmp > 0)
 922                         new = &((*new)->rb_left);
 923                 else
 924                         new = &((*new)->rb_right);
 925         }
 926
 927         rb_link_node(&data->node, parent, new);
 928         rb_insert_color(&data->node, root);
 929 }
 930
 931 static void thread_atoms_insert(struct thread *thread)
 932 {
 933         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 934         if (!atoms)
 935                 die("No memory");
 936
 937         atoms->thread = thread;
 938         INIT_LIST_HEAD(&atoms->work_list);
 939         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 940 }
 941
 942 static void
 943 latency_fork_event(struct trace_fork_event *fork_event __used,
 944                    struct event *event __used,
 945                    int cpu __used,
 946                    u64 timestamp __used,
 947                    struct thread *thread __used)
 948 {
 949         /* should insert the newcomer */
 950 }
 951
 952 __used
 953 static char sched_out_state(struct trace_switch_event *switch_event)
 954 {
 955         const char *str = TASK_STATE_TO_CHAR_STR;
 956
 957         return str[switch_event->prev_state];
 958 }
 959
 960 static void
 961 add_sched_out_event(struct work_atoms *atoms,
 962                     char run_state,
 963                     u64 timestamp)
 964 {
 965         struct work_atom *atom = zalloc(sizeof(*atom));
 966         if (!atom)
 967                 die("Non memory");
 968
 969         atom->sched_out_time = timestamp;
 970
 971         if (run_state == 'R') {
 972                 atom->state = THREAD_WAIT_CPU;
 973                 atom->wake_up_time = atom->sched_out_time;
 974         }
 975
 976         list_add_tail(&atom->list, &atoms->work_list);
 977 }
 978
 979 static void
 980 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
 981 {
 982         struct work_atom *atom;
 983
 984         BUG_ON(list_empty(&atoms->work_list));
 985
 986         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 987
 988         atom->runtime += delta;
 989         atoms->total_runtime += delta;
 990 }
 991
 992 static void
 993 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 994 {
 995         struct work_atom *atom;
 996         u64 delta;
 997
 998         if (list_empty(&atoms->work_list))
 999                 return;
1000
1001         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1002
1003         if (atom->state != THREAD_WAIT_CPU)
1004                 return;
1005
1006         if (timestamp < atom->wake_up_time) {
1007                 atom->state = THREAD_IGNORE;
1008                 return;
1009         }
1010
1011         atom->state = THREAD_SCHED_IN;
1012         atom->sched_in_time = timestamp;
1013
1014         delta = atom->sched_in_time - atom->wake_up_time;
1015         atoms->total_lat += delta;
1016         if (delta > atoms->max_lat) {
1017                 atoms->max_lat = delta;
1018                 atoms->max_lat_at = timestamp;
1019         }
1020         atoms->nb_atoms++;
1021 }
1022
1023 static void
1024 latency_switch_event(struct trace_switch_event *switch_event,
1025                      struct perf_session *session,
1026                      struct event *event __used,
1027                      int cpu,
1028                      u64 timestamp,
1029                      struct thread *thread __used)
1030 {
1031         struct work_atoms *out_events, *in_events;
1032         struct thread *sched_out, *sched_in;
1033         u64 timestamp0;
1034         s64 delta;
1035
1036         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1037
1038         timestamp0 = cpu_last_switched[cpu];
1039         cpu_last_switched[cpu] = timestamp;
1040         if (timestamp0)
1041                 delta = timestamp - timestamp0;
1042         else
1043                 delta = 0;
1044
1045         if (delta < 0)
1046                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1047
1048
1049         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1050         sched_in = perf_session__findnew(session, switch_event->next_pid);
1051
1052         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1053         if (!out_events) {
1054                 thread_atoms_insert(sched_out);
1055                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1056                 if (!out_events)
1057                         die("out-event: Internal tree error");
1058         }
1059         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1060
1061         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1062         if (!in_events) {
1063                 thread_atoms_insert(sched_in);
1064                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1065                 if (!in_events)
1066                         die("in-event: Internal tree error");
1067                 /*
1068                  * Take came in we have not heard about yet,
1069                  * add in an initial atom in runnable state:
1070                  */
1071                 add_sched_out_event(in_events, 'R', timestamp);
1072         }
1073         add_sched_in_event(in_events, timestamp);
1074 }
1075
1076 static void
1077 latency_runtime_event(struct trace_runtime_event *runtime_event,
1078                      struct perf_session *session,
1079                      struct event *event __used,
1080                      int cpu,
1081                      u64 timestamp,
1082                      struct thread *this_thread __used)
1083 {
1084         struct thread *thread = perf_session__findnew(session, runtime_event->pid);
1085         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1086
1087         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1088         if (!atoms) {
1089                 thread_atoms_insert(thread);
1090                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1091                 if (!atoms)
1092                         die("in-event: Internal tree error");
1093                 add_sched_out_event(atoms, 'R', timestamp);
1094         }
1095
1096         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1097 }
1098
1099 static void
1100 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1101                      struct perf_session *session,
1102                      struct event *__event __used,
1103                      int cpu __used,
1104                      u64 timestamp,
1105                      struct thread *thread __used)
1106 {
1107         struct work_atoms *atoms;
1108         struct work_atom *atom;
1109         struct thread *wakee;
1110
1111         /* Note for later, it may be interesting to observe the failing cases */
1112         if (!wakeup_event->success)
1113                 return;
1114
1115         wakee = perf_session__findnew(session, wakeup_event->pid);
1116         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1117         if (!atoms) {
1118                 thread_atoms_insert(wakee);
1119                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1120                 if (!atoms)
1121                         die("wakeup-event: Internal tree error");
1122                 add_sched_out_event(atoms, 'S', timestamp);
1123         }
1124
1125         BUG_ON(list_empty(&atoms->work_list));
1126
1127         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1128
1129         /*
1130          * You WILL be missing events if you've recorded only
1131          * one CPU, or are only looking at only one, so don't
1132          * make useless noise.
1133          */
1134         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1135                 nr_state_machine_bugs++;
1136
1137         nr_timestamps++;
1138         if (atom->sched_out_time > timestamp) {
1139                 nr_unordered_timestamps++;
1140                 return;
1141         }
1142
1143         atom->state = THREAD_WAIT_CPU;
1144         atom->wake_up_time = timestamp;
1145 }
1146
1147 static void
1148 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1149                      struct perf_session *session,
1150                      struct event *__event __used,
1151                      int cpu __used,
1152                      u64 timestamp,
1153                      struct thread *thread __used)
1154 {
1155         struct work_atoms *atoms;
1156         struct work_atom *atom;
1157         struct thread *migrant;
1158
1159         /*
1160          * Only need to worry about migration when profiling one CPU.
1161          */
1162         if (profile_cpu == -1)
1163                 return;
1164
1165         migrant = perf_session__findnew(session, migrate_task_event->pid);
1166         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1167         if (!atoms) {
1168                 thread_atoms_insert(migrant);
1169                 register_pid(migrant->pid, migrant->comm);
1170                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1171                 if (!atoms)
1172                         die("migration-event: Internal tree error");
1173                 add_sched_out_event(atoms, 'R', timestamp);
1174         }
1175
1176         BUG_ON(list_empty(&atoms->work_list));
1177
1178         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1179         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1180
1181         nr_timestamps++;
1182
1183         if (atom->sched_out_time > timestamp)
1184                 nr_unordered_timestamps++;
1185 }
1186
1187 static struct trace_sched_handler lat_ops  = {
1188         .wakeup_event           = latency_wakeup_event,
1189         .switch_event           = latency_switch_event,
1190         .runtime_event          = latency_runtime_event,
1191         .fork_event             = latency_fork_event,
1192         .migrate_task_event     = latency_migrate_task_event,
1193 };
1194
1195 static void output_lat_thread(struct work_atoms *work_list)
1196 {
1197         int i;
1198         int ret;
1199         u64 avg;
1200
1201         if (!work_list->nb_atoms)
1202                 return;
1203         /*
1204          * Ignore idle threads:
1205          */
1206         if (!strcmp(work_list->thread->comm, "swapper"))
1207                 return;
1208
1209         all_runtime += work_list->total_runtime;
1210         all_count += work_list->nb_atoms;
1211
1212         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1213
1214         for (i = 0; i < 24 - ret; i++)
1215                 printf(" ");
1216
1217         avg = work_list->total_lat / work_list->nb_atoms;
1218
1219         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1220               (double)work_list->total_runtime / 1e6,
1221                  work_list->nb_atoms, (double)avg / 1e6,
1222                  (double)work_list->max_lat / 1e6,
1223                  (double)work_list->max_lat_at / 1e9);
1224 }
1225
1226 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1227 {
1228         if (l->thread->pid < r->thread->pid)
1229                 return -1;
1230         if (l->thread->pid > r->thread->pid)
1231                 return 1;
1232
1233         return 0;
1234 }
1235
1236 static struct sort_dimension pid_sort_dimension = {
1237         .name                   = "pid",
1238         .cmp                    = pid_cmp,
1239 };
1240
1241 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1242 {
1243         u64 avgl, avgr;
1244
1245         if (!l->nb_atoms)
1246                 return -1;
1247
1248         if (!r->nb_atoms)
1249                 return 1;
1250
1251         avgl = l->total_lat / l->nb_atoms;
1252         avgr = r->total_lat / r->nb_atoms;
1253
1254         if (avgl < avgr)
1255                 return -1;
1256         if (avgl > avgr)
1257                 return 1;
1258
1259         return 0;
1260 }
1261
1262 static struct sort_dimension avg_sort_dimension = {
1263         .name                   = "avg",
1264         .cmp                    = avg_cmp,
1265 };
1266
1267 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1268 {
1269         if (l->max_lat < r->max_lat)
1270                 return -1;
1271         if (l->max_lat > r->max_lat)
1272                 return 1;
1273
1274         return 0;
1275 }
1276
1277 static struct sort_dimension max_sort_dimension = {
1278         .name                   = "max",
1279         .cmp                    = max_cmp,
1280 };
1281
1282 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1283 {
1284         if (l->nb_atoms < r->nb_atoms)
1285                 return -1;
1286         if (l->nb_atoms > r->nb_atoms)
1287                 return 1;
1288
1289         return 0;
1290 }
1291
1292 static struct sort_dimension switch_sort_dimension = {
1293         .name                   = "switch",
1294         .cmp                    = switch_cmp,
1295 };
1296
1297 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1298 {
1299         if (l->total_runtime < r->total_runtime)
1300                 return -1;
1301         if (l->total_runtime > r->total_runtime)
1302                 return 1;
1303
1304         return 0;
1305 }
1306
1307 static struct sort_dimension runtime_sort_dimension = {
1308         .name                   = "runtime",
1309         .cmp                    = runtime_cmp,
1310 };
1311
1312 static struct sort_dimension *available_sorts[] = {
1313         &pid_sort_dimension,
1314         &avg_sort_dimension,
1315         &max_sort_dimension,
1316         &switch_sort_dimension,
1317         &runtime_sort_dimension,
1318 };
1319
1320 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1321
1322 static LIST_HEAD(sort_list);
1323
1324 static int sort_dimension__add(const char *tok, struct list_head *list)
1325 {
1326         int i;
1327
1328         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1329                 if (!strcmp(available_sorts[i]->name, tok)) {
1330                         list_add_tail(&available_sorts[i]->list, list);
1331
1332                         return 0;
1333                 }
1334         }
1335
1336         return -1;
1337 }
1338
1339 static void setup_sorting(void);
1340
1341 static void sort_lat(void)
1342 {
1343         struct rb_node *node;
1344
1345         for (;;) {
1346                 struct work_atoms *data;
1347                 node = rb_first(&atom_root);
1348                 if (!node)
1349                         break;
1350
1351                 rb_erase(node, &atom_root);
1352                 data = rb_entry(node, struct work_atoms, node);
1353                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1354         }
1355 }
1356
1357 static struct trace_sched_handler *trace_handler;
1358
1359 static void
1360 process_sched_wakeup_event(void *data, struct perf_session *session,
1361                            struct event *event,
1362                            int cpu __used,
1363                            u64 timestamp __used,
1364                            struct thread *thread __used)
1365 {
1366         struct trace_wakeup_event wakeup_event;
1367
1368         FILL_COMMON_FIELDS(wakeup_event, event, data);
1369
1370         FILL_ARRAY(wakeup_event, comm, event, data);
1371         FILL_FIELD(wakeup_event, pid, event, data);
1372         FILL_FIELD(wakeup_event, prio, event, data);
1373         FILL_FIELD(wakeup_event, success, event, data);
1374         FILL_FIELD(wakeup_event, cpu, event, data);
1375
1376         if (trace_handler->wakeup_event)
1377                 trace_handler->wakeup_event(&wakeup_event, session, event,
1378                                             cpu, timestamp, thread);
1379 }
1380
1381 /*
1382  * Track the current task - that way we can know whether there's any
1383  * weird events, such as a task being switched away that is not current.
1384  */
1385 static int max_cpu;
1386
1387 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1388
1389 static struct thread *curr_thread[MAX_CPUS];
1390
1391 static char next_shortname1 = 'A';
1392 static char next_shortname2 = '0';
1393
1394 static void
1395 map_switch_event(struct trace_switch_event *switch_event,
1396                  struct perf_session *session,
1397                  struct event *event __used,
1398                  int this_cpu,
1399                  u64 timestamp,
1400                  struct thread *thread __used)
1401 {
1402         struct thread *sched_out __used, *sched_in;
1403         int new_shortname;
1404         u64 timestamp0;
1405         s64 delta;
1406         int cpu;
1407
1408         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1409
1410         if (this_cpu > max_cpu)
1411                 max_cpu = this_cpu;
1412
1413         timestamp0 = cpu_last_switched[this_cpu];
1414         cpu_last_switched[this_cpu] = timestamp;
1415         if (timestamp0)
1416                 delta = timestamp - timestamp0;
1417         else
1418                 delta = 0;
1419
1420         if (delta < 0)
1421                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1422
1423
1424         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1425         sched_in = perf_session__findnew(session, switch_event->next_pid);
1426
1427         curr_thread[this_cpu] = sched_in;
1428
1429         printf("  ");
1430
1431         new_shortname = 0;
1432         if (!sched_in->shortname[0]) {
1433                 sched_in->shortname[0] = next_shortname1;
1434                 sched_in->shortname[1] = next_shortname2;
1435
1436                 if (next_shortname1 < 'Z') {
1437                         next_shortname1++;
1438                 } else {
1439                         next_shortname1='A';
1440                         if (next_shortname2 < '9') {
1441                                 next_shortname2++;
1442                         } else {
1443                                 next_shortname2='0';
1444                         }
1445                 }
1446                 new_shortname = 1;
1447         }
1448
1449         for (cpu = 0; cpu <= max_cpu; cpu++) {
1450                 if (cpu != this_cpu)
1451                         printf(" ");
1452                 else
1453                         printf("*");
1454
1455                 if (curr_thread[cpu]) {
1456                         if (curr_thread[cpu]->pid)
1457                                 printf("%2s ", curr_thread[cpu]->shortname);
1458                         else
1459                                 printf(".  ");
1460                 } else
1461                         printf("   ");
1462         }
1463
1464         printf("  %12.6f secs ", (double)timestamp/1e9);
1465         if (new_shortname) {
1466                 printf("%s => %s:%d\n",
1467                         sched_in->shortname, sched_in->comm, sched_in->pid);
1468         } else {
1469                 printf("\n");
1470         }
1471 }
1472
1473
1474 static void
1475 process_sched_switch_event(void *data, struct perf_session *session,
1476                            struct event *event,
1477                            int this_cpu,
1478                            u64 timestamp __used,
1479                            struct thread *thread __used)
1480 {
1481         struct trace_switch_event switch_event;
1482
1483         FILL_COMMON_FIELDS(switch_event, event, data);
1484
1485         FILL_ARRAY(switch_event, prev_comm, event, data);
1486         FILL_FIELD(switch_event, prev_pid, event, data);
1487         FILL_FIELD(switch_event, prev_prio, event, data);
1488         FILL_FIELD(switch_event, prev_state, event, data);
1489         FILL_ARRAY(switch_event, next_comm, event, data);
1490         FILL_FIELD(switch_event, next_pid, event, data);
1491         FILL_FIELD(switch_event, next_prio, event, data);
1492
1493         if (curr_pid[this_cpu] != (u32)-1) {
1494                 /*
1495                  * Are we trying to switch away a PID that is
1496                  * not current?
1497                  */
1498                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1499                         nr_context_switch_bugs++;
1500         }
1501         if (trace_handler->switch_event)
1502                 trace_handler->switch_event(&switch_event, session, event,
1503                                             this_cpu, timestamp, thread);
1504
1505         curr_pid[this_cpu] = switch_event.next_pid;
1506 }
1507
1508 static void
1509 process_sched_runtime_event(void *data, struct perf_session *session,
1510                            struct event *event,
1511                            int cpu __used,
1512                            u64 timestamp __used,
1513                            struct thread *thread __used)
1514 {
1515         struct trace_runtime_event runtime_event;
1516
1517         FILL_ARRAY(runtime_event, comm, event, data);
1518         FILL_FIELD(runtime_event, pid, event, data);
1519         FILL_FIELD(runtime_event, runtime, event, data);
1520         FILL_FIELD(runtime_event, vruntime, event, data);
1521
1522         if (trace_handler->runtime_event)
1523                 trace_handler->runtime_event(&runtime_event, session, event, cpu, timestamp, thread);
1524 }
1525
1526 static void
1527 process_sched_fork_event(void *data,
1528                          struct event *event,
1529                          int cpu __used,
1530                          u64 timestamp __used,
1531                          struct thread *thread __used)
1532 {
1533         struct trace_fork_event fork_event;
1534
1535         FILL_COMMON_FIELDS(fork_event, event, data);
1536
1537         FILL_ARRAY(fork_event, parent_comm, event, data);
1538         FILL_FIELD(fork_event, parent_pid, event, data);
1539         FILL_ARRAY(fork_event, child_comm, event, data);
1540         FILL_FIELD(fork_event, child_pid, event, data);
1541
1542         if (trace_handler->fork_event)
1543                 trace_handler->fork_event(&fork_event, event,
1544                                           cpu, timestamp, thread);
1545 }
1546
1547 static void
1548 process_sched_exit_event(struct event *event,
1549                          int cpu __used,
1550                          u64 timestamp __used,
1551                          struct thread *thread __used)
1552 {
1553         if (verbose)
1554                 printf("sched_exit event %p\n", event);
1555 }
1556
1557 static void
1558 process_sched_migrate_task_event(void *data, struct perf_session *session,
1559                            struct event *event,
1560                            int cpu __used,
1561                            u64 timestamp __used,
1562                            struct thread *thread __used)
1563 {
1564         struct trace_migrate_task_event migrate_task_event;
1565
1566         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1567
1568         FILL_ARRAY(migrate_task_event, comm, event, data);
1569         FILL_FIELD(migrate_task_event, pid, event, data);
1570         FILL_FIELD(migrate_task_event, prio, event, data);
1571         FILL_FIELD(migrate_task_event, cpu, event, data);
1572
1573         if (trace_handler->migrate_task_event)
1574                 trace_handler->migrate_task_event(&migrate_task_event, session,
1575                                                  event, cpu, timestamp, thread);
1576 }
1577
1578 static void process_raw_event(union perf_event *raw_event __used,
1579                               struct perf_session *session, void *data, int cpu,
1580                               u64 timestamp, struct thread *thread)
1581 {
1582         struct event *event;
1583         int type;
1584
1585
1586         type = trace_parse_common_type(data);
1587         event = trace_find_event(type);
1588
1589         if (!strcmp(event->name, "sched_switch"))
1590                 process_sched_switch_event(data, session, event, cpu, timestamp, thread);
1591         if (!strcmp(event->name, "sched_stat_runtime"))
1592                 process_sched_runtime_event(data, session, event, cpu, timestamp, thread);
1593         if (!strcmp(event->name, "sched_wakeup"))
1594                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1595         if (!strcmp(event->name, "sched_wakeup_new"))
1596                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1597         if (!strcmp(event->name, "sched_process_fork"))
1598                 process_sched_fork_event(data, event, cpu, timestamp, thread);
1599         if (!strcmp(event->name, "sched_process_exit"))
1600                 process_sched_exit_event(event, cpu, timestamp, thread);
1601         if (!strcmp(event->name, "sched_migrate_task"))
1602                 process_sched_migrate_task_event(data, session, event, cpu, timestamp, thread);
1603 }
1604
1605 static int process_sample_event(union perf_event *event,
1606                                 struct perf_sample *sample,
1607                                 struct perf_evsel *evsel __used,
1608                                 struct perf_session *session)
1609 {
1610         struct thread *thread;
1611
1612         if (!(session->sample_type & PERF_SAMPLE_RAW))
1613                 return 0;
1614
1615         thread = perf_session__findnew(session, sample->pid);
1616         if (thread == NULL) {
1617                 pr_debug("problem processing %d event, skipping it.\n",
1618                          event->header.type);
1619                 return -1;
1620         }
1621
1622         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1623
1624         if (profile_cpu != -1 && profile_cpu != (int)sample->cpu)
1625                 return 0;
1626
1627         process_raw_event(event, session, sample->raw_data, sample->cpu,
1628                           sample->time, thread);
1629
1630         return 0;
1631 }
1632
1633 static struct perf_event_ops event_ops = {
1634         .sample                 = process_sample_event,
1635         .comm                   = perf_event__process_comm,
1636         .lost                   = perf_event__process_lost,
1637         .fork                   = perf_event__process_task,
1638         .ordered_samples        = true,
1639 };
1640
1641 static void read_events(bool destroy, struct perf_session **psession)
1642 {
1643         int err = -EINVAL;
1644         struct perf_session *session = perf_session__new(input_name, O_RDONLY,
1645                                                          0, false, &event_ops);
1646         if (session == NULL)
1647                 die("No Memory");
1648
1649         if (perf_session__has_traces(session, "record -R")) {
1650                 err = perf_session__process_events(session, &event_ops);
1651                 if (err)
1652                         die("Failed to process events, error %d", err);
1653
1654                 nr_events      = session->hists.stats.nr_events[0];
1655                 nr_lost_events = session->hists.stats.total_lost;
1656                 nr_lost_chunks = session->hists.stats.nr_events[PERF_RECORD_LOST];
1657         }
1658
1659         if (destroy)
1660                 perf_session__delete(session);
1661
1662         if (psession)
1663                 *psession = session;
1664 }
1665
1666 static void print_bad_events(void)
1667 {
1668         if (nr_unordered_timestamps && nr_timestamps) {
1669                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1670                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1671                         nr_unordered_timestamps, nr_timestamps);
1672         }
1673         if (nr_lost_events && nr_events) {
1674                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1675                         (double)nr_lost_events/(double)nr_events*100.0,
1676                         nr_lost_events, nr_events, nr_lost_chunks);
1677         }
1678         if (nr_state_machine_bugs && nr_timestamps) {
1679                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1680                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1681                         nr_state_machine_bugs, nr_timestamps);
1682                 if (nr_lost_events)
1683                         printf(" (due to lost events?)");
1684                 printf("\n");
1685         }
1686         if (nr_context_switch_bugs && nr_timestamps) {
1687                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1688                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1689                         nr_context_switch_bugs, nr_timestamps);
1690                 if (nr_lost_events)
1691                         printf(" (due to lost events?)");
1692                 printf("\n");
1693         }
1694 }
1695
1696 static void __cmd_lat(void)
1697 {
1698         struct rb_node *next;
1699         struct perf_session *session;
1700
1701         setup_pager();
1702         read_events(false, &session);
1703         sort_lat();
1704
1705         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1706         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1707         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1708
1709         next = rb_first(&sorted_atom_root);
1710
1711         while (next) {
1712                 struct work_atoms *work_list;
1713
1714                 work_list = rb_entry(next, struct work_atoms, node);
1715                 output_lat_thread(work_list);
1716                 next = rb_next(next);
1717         }
1718
1719         printf(" -----------------------------------------------------------------------------------------\n");
1720         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1721                 (double)all_runtime/1e6, all_count);
1722
1723         printf(" ---------------------------------------------------\n");
1724
1725         print_bad_events();
1726         printf("\n");
1727
1728         perf_session__delete(session);
1729 }
1730
1731 static struct trace_sched_handler map_ops  = {
1732         .wakeup_event           = NULL,
1733         .switch_event           = map_switch_event,
1734         .runtime_event          = NULL,
1735         .fork_event             = NULL,
1736 };
1737
1738 static void __cmd_map(void)
1739 {
1740         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1741
1742         setup_pager();
1743         read_events(true, NULL);
1744         print_bad_events();
1745 }
1746
1747 static void __cmd_replay(void)
1748 {
1749         unsigned long i;
1750
1751         calibrate_run_measurement_overhead();
1752         calibrate_sleep_measurement_overhead();
1753
1754         test_calibrations();
1755
1756         read_events(true, NULL);
1757
1758         printf("nr_run_events:        %ld\n", nr_run_events);
1759         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1760         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1761
1762         if (targetless_wakeups)
1763                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1764         if (multitarget_wakeups)
1765                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1766         if (nr_run_events_optimized)
1767                 printf("run atoms optimized: %ld\n",
1768                         nr_run_events_optimized);
1769
1770         print_task_traces();
1771         add_cross_task_wakeups();
1772
1773         create_tasks();
1774         printf("------------------------------------------------------------\n");
1775         for (i = 0; i < replay_repeat; i++)
1776                 run_one_test();
1777 }
1778
1779
1780 static const char * const sched_usage[] = {
1781         "perf sched [<options>] {record|latency|map|replay|script}",
1782         NULL
1783 };
1784
1785 static const struct option sched_options[] = {
1786         OPT_STRING('i', "input", &input_name, "file",
1787                     "input file name"),
1788         OPT_INCR('v', "verbose", &verbose,
1789                     "be more verbose (show symbol address, etc)"),
1790         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1791                     "dump raw trace in ASCII"),
1792         OPT_END()
1793 };
1794
1795 static const char * const latency_usage[] = {
1796         "perf sched latency [<options>]",
1797         NULL
1798 };
1799
1800 static const struct option latency_options[] = {
1801         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1802                    "sort by key(s): runtime, switch, avg, max"),
1803         OPT_INCR('v', "verbose", &verbose,
1804                     "be more verbose (show symbol address, etc)"),
1805         OPT_INTEGER('C', "CPU", &profile_cpu,
1806                     "CPU to profile on"),
1807         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1808                     "dump raw trace in ASCII"),
1809         OPT_END()
1810 };
1811
1812 static const char * const replay_usage[] = {
1813         "perf sched replay [<options>]",
1814         NULL
1815 };
1816
1817 static const struct option replay_options[] = {
1818         OPT_UINTEGER('r', "repeat", &replay_repeat,
1819                      "repeat the workload replay N times (-1: infinite)"),
1820         OPT_INCR('v', "verbose", &verbose,
1821                     "be more verbose (show symbol address, etc)"),
1822         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1823                     "dump raw trace in ASCII"),
1824         OPT_END()
1825 };
1826
1827 static void setup_sorting(void)
1828 {
1829         char *tmp, *tok, *str = strdup(sort_order);
1830
1831         for (tok = strtok_r(str, ", ", &tmp);
1832                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1833                 if (sort_dimension__add(tok, &sort_list) < 0) {
1834                         error("Unknown --sort key: `%s'", tok);
1835                         usage_with_options(latency_usage, latency_options);
1836                 }
1837         }
1838
1839         free(str);
1840
1841         sort_dimension__add("pid", &cmp_pid);
1842 }
1843
1844 static const char *record_args[] = {
1845         "record",
1846         "-a",
1847         "-R",
1848         "-f",
1849         "-m", "1024",
1850         "-c", "1",
1851         "-e", "sched:sched_switch",
1852         "-e", "sched:sched_stat_wait",
1853         "-e", "sched:sched_stat_sleep",
1854         "-e", "sched:sched_stat_iowait",
1855         "-e", "sched:sched_stat_runtime",
1856         "-e", "sched:sched_process_exit",
1857         "-e", "sched:sched_process_fork",
1858         "-e", "sched:sched_wakeup",
1859         "-e", "sched:sched_migrate_task",
1860 };
1861
1862 static int __cmd_record(int argc, const char **argv)
1863 {
1864         unsigned int rec_argc, i, j;
1865         const char **rec_argv;
1866
1867         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1868         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1869
1870         if (rec_argv == NULL)
1871                 return -ENOMEM;
1872
1873         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1874                 rec_argv[i] = strdup(record_args[i]);
1875
1876         for (j = 1; j < (unsigned int)argc; j++, i++)
1877                 rec_argv[i] = argv[j];
1878
1879         BUG_ON(i != rec_argc);
1880
1881         return cmd_record(i, rec_argv, NULL);
1882 }
1883
1884 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1885 {
1886         argc = parse_options(argc, argv, sched_options, sched_usage,
1887                              PARSE_OPT_STOP_AT_NON_OPTION);
1888         if (!argc)
1889                 usage_with_options(sched_usage, sched_options);
1890
1891         /*
1892          * Aliased to 'perf script' for now:
1893          */
1894         if (!strcmp(argv[0], "script"))
1895                 return cmd_script(argc, argv, prefix);
1896
1897         symbol__init();
1898         if (!strncmp(argv[0], "rec", 3)) {
1899                 return __cmd_record(argc, argv);
1900         } else if (!strncmp(argv[0], "lat", 3)) {
1901                 trace_handler = &lat_ops;
1902                 if (argc > 1) {
1903                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1904                         if (argc)
1905                                 usage_with_options(latency_usage, latency_options);
1906                 }
1907                 setup_sorting();
1908                 __cmd_lat();
1909         } else if (!strcmp(argv[0], "map")) {
1910                 trace_handler = &map_ops;
1911                 setup_sorting();
1912                 __cmd_map();
1913         } else if (!strncmp(argv[0], "rep", 3)) {
1914                 trace_handler = &replay_ops;
1915                 if (argc) {
1916                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1917                         if (argc)
1918                                 usage_with_options(replay_usage, replay_options);
1919                 }
1920                 __cmd_replay();
1921         } else {
1922                 usage_with_options(sched_usage, sched_options);
1923         }
1924
1925         return 0;
1926 }