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
  18 #include <semaphore.h>
  19 #include <pthread.h>
  20 #include <math.h>
  21
  22 static char                     const *input_name = "perf.data";
  23
  24 static u64                      sample_type;
  25
  26 static char                     default_sort_order[] = "avg, max, switch, runtime";
  27 static char                     *sort_order = default_sort_order;
  28
  29 static int                      profile_cpu = -1;
  30
  31 #define PR_SET_NAME             15               /* Set process name */
  32 #define MAX_CPUS                4096
  33
  34 static u64                      run_measurement_overhead;
  35 static u64                      sleep_measurement_overhead;
  36
  37 #define COMM_LEN                20
  38 #define SYM_LEN                 129
  39
  40 #define MAX_PID                 65536
  41
  42 static unsigned long            nr_tasks;
  43
  44 struct sched_atom;
  45
  46 struct task_desc {
  47         unsigned long           nr;
  48         unsigned long           pid;
  49         char                    comm[COMM_LEN];
  50
  51         unsigned long           nr_events;
  52         unsigned long           curr_event;
  53         struct sched_atom       **atoms;
  54
  55         pthread_t               thread;
  56         sem_t                   sleep_sem;
  57
  58         sem_t                   ready_for_work;
  59         sem_t                   work_done_sem;
  60
  61         u64                     cpu_usage;
  62 };
  63
  64 enum sched_event_type {
  65         SCHED_EVENT_RUN,
  66         SCHED_EVENT_SLEEP,
  67         SCHED_EVENT_WAKEUP,
  68         SCHED_EVENT_MIGRATION,
  69 };
  70
  71 struct sched_atom {
  72         enum sched_event_type   type;
  73         u64                     timestamp;
  74         u64                     duration;
  75         unsigned long           nr;
  76         int                     specific_wait;
  77         sem_t                   *wait_sem;
  78         struct task_desc        *wakee;
  79 };
  80
  81 static struct task_desc         *pid_to_task[MAX_PID];
  82
  83 static struct task_desc         **tasks;
  84
  85 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  86 static u64                      start_time;
  87
  88 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  89
  90 static unsigned long            nr_run_events;
  91 static unsigned long            nr_sleep_events;
  92 static unsigned long            nr_wakeup_events;
  93
  94 static unsigned long            nr_sleep_corrections;
  95 static unsigned long            nr_run_events_optimized;
  96
  97 static unsigned long            targetless_wakeups;
  98 static unsigned long            multitarget_wakeups;
  99
 100 static u64                      cpu_usage;
 101 static u64                      runavg_cpu_usage;
 102 static u64                      parent_cpu_usage;
 103 static u64                      runavg_parent_cpu_usage;
 104
 105 static unsigned long            nr_runs;
 106 static u64                      sum_runtime;
 107 static u64                      sum_fluct;
 108 static u64                      run_avg;
 109
 110 static unsigned long            replay_repeat = 10;
 111 static unsigned long            nr_timestamps;
 112 static unsigned long            nr_unordered_timestamps;
 113 static unsigned long            nr_state_machine_bugs;
 114 static unsigned long            nr_context_switch_bugs;
 115 static unsigned long            nr_events;
 116 static unsigned long            nr_lost_chunks;
 117 static unsigned long            nr_lost_events;
 118
 119 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 120
 121 enum thread_state {
 122         THREAD_SLEEPING = 0,
 123         THREAD_WAIT_CPU,
 124         THREAD_SCHED_IN,
 125         THREAD_IGNORE
 126 };
 127
 128 struct work_atom {
 129         struct list_head        list;
 130         enum thread_state       state;
 131         u64                     sched_out_time;
 132         u64                     wake_up_time;
 133         u64                     sched_in_time;
 134         u64                     runtime;
 135 };
 136
 137 struct work_atoms {
 138         struct list_head        work_list;
 139         struct thread           *thread;
 140         struct rb_node          node;
 141         u64                     max_lat;
 142         u64                     max_lat_at;
 143         u64                     total_lat;
 144         u64                     nb_atoms;
 145         u64                     total_runtime;
 146 };
 147
 148 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 149
 150 static struct rb_root           atom_root, sorted_atom_root;
 151
 152 static u64                      all_runtime;
 153 static u64                      all_count;
 154
 155
 156 static u64 get_nsecs(void)
 157 {
 158         struct timespec ts;
 159
 160         clock_gettime(CLOCK_MONOTONIC, &ts);
 161
 162         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 163 }
 164
 165 static void burn_nsecs(u64 nsecs)
 166 {
 167         u64 T0 = get_nsecs(), T1;
 168
 169         do {
 170                 T1 = get_nsecs();
 171         } while (T1 + run_measurement_overhead < T0 + nsecs);
 172 }
 173
 174 static void sleep_nsecs(u64 nsecs)
 175 {
 176         struct timespec ts;
 177
 178         ts.tv_nsec = nsecs % 999999999;
 179         ts.tv_sec = nsecs / 999999999;
 180
 181         nanosleep(&ts, NULL);
 182 }
 183
 184 static void calibrate_run_measurement_overhead(void)
 185 {
 186         u64 T0, T1, delta, min_delta = 1000000000ULL;
 187         int i;
 188
 189         for (i = 0; i < 10; i++) {
 190                 T0 = get_nsecs();
 191                 burn_nsecs(0);
 192                 T1 = get_nsecs();
 193                 delta = T1-T0;
 194                 min_delta = min(min_delta, delta);
 195         }
 196         run_measurement_overhead = min_delta;
 197
 198         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 199 }
 200
 201 static void calibrate_sleep_measurement_overhead(void)
 202 {
 203         u64 T0, T1, delta, min_delta = 1000000000ULL;
 204         int i;
 205
 206         for (i = 0; i < 10; i++) {
 207                 T0 = get_nsecs();
 208                 sleep_nsecs(10000);
 209                 T1 = get_nsecs();
 210                 delta = T1-T0;
 211                 min_delta = min(min_delta, delta);
 212         }
 213         min_delta -= 10000;
 214         sleep_measurement_overhead = min_delta;
 215
 216         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 217 }
 218
 219 static struct sched_atom *
 220 get_new_event(struct task_desc *task, u64 timestamp)
 221 {
 222         struct sched_atom *event = zalloc(sizeof(*event));
 223         unsigned long idx = task->nr_events;
 224         size_t size;
 225
 226         event->timestamp = timestamp;
 227         event->nr = idx;
 228
 229         task->nr_events++;
 230         size = sizeof(struct sched_atom *) * task->nr_events;
 231         task->atoms = realloc(task->atoms, size);
 232         BUG_ON(!task->atoms);
 233
 234         task->atoms[idx] = event;
 235
 236         return event;
 237 }
 238
 239 static struct sched_atom *last_event(struct task_desc *task)
 240 {
 241         if (!task->nr_events)
 242                 return NULL;
 243
 244         return task->atoms[task->nr_events - 1];
 245 }
 246
 247 static void
 248 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 249 {
 250         struct sched_atom *event, *curr_event = last_event(task);
 251
 252         /*
 253          * optimize an existing RUN event by merging this one
 254          * to it:
 255          */
 256         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 257                 nr_run_events_optimized++;
 258                 curr_event->duration += duration;
 259                 return;
 260         }
 261
 262         event = get_new_event(task, timestamp);
 263
 264         event->type = SCHED_EVENT_RUN;
 265         event->duration = duration;
 266
 267         nr_run_events++;
 268 }
 269
 270 static void
 271 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 272                        struct task_desc *wakee)
 273 {
 274         struct sched_atom *event, *wakee_event;
 275
 276         event = get_new_event(task, timestamp);
 277         event->type = SCHED_EVENT_WAKEUP;
 278         event->wakee = wakee;
 279
 280         wakee_event = last_event(wakee);
 281         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 282                 targetless_wakeups++;
 283                 return;
 284         }
 285         if (wakee_event->wait_sem) {
 286                 multitarget_wakeups++;
 287                 return;
 288         }
 289
 290         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 291         sem_init(wakee_event->wait_sem, 0, 0);
 292         wakee_event->specific_wait = 1;
 293         event->wait_sem = wakee_event->wait_sem;
 294
 295         nr_wakeup_events++;
 296 }
 297
 298 static void
 299 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 300                       u64 task_state __used)
 301 {
 302         struct sched_atom *event = get_new_event(task, timestamp);
 303
 304         event->type = SCHED_EVENT_SLEEP;
 305
 306         nr_sleep_events++;
 307 }
 308
 309 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 310 {
 311         struct task_desc *task;
 312
 313         BUG_ON(pid >= MAX_PID);
 314
 315         task = pid_to_task[pid];
 316
 317         if (task)
 318                 return task;
 319
 320         task = zalloc(sizeof(*task));
 321         task->pid = pid;
 322         task->nr = nr_tasks;
 323         strcpy(task->comm, comm);
 324         /*
 325          * every task starts in sleeping state - this gets ignored
 326          * if there's no wakeup pointing to this sleep state:
 327          */
 328         add_sched_event_sleep(task, 0, 0);
 329
 330         pid_to_task[pid] = task;
 331         nr_tasks++;
 332         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 333         BUG_ON(!tasks);
 334         tasks[task->nr] = task;
 335
 336         if (verbose)
 337                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 338
 339         return task;
 340 }
 341
 342
 343 static void print_task_traces(void)
 344 {
 345         struct task_desc *task;
 346         unsigned long i;
 347
 348         for (i = 0; i < nr_tasks; i++) {
 349                 task = tasks[i];
 350                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 351                         task->nr, task->comm, task->pid, task->nr_events);
 352         }
 353 }
 354
 355 static void add_cross_task_wakeups(void)
 356 {
 357         struct task_desc *task1, *task2;
 358         unsigned long i, j;
 359
 360         for (i = 0; i < nr_tasks; i++) {
 361                 task1 = tasks[i];
 362                 j = i + 1;
 363                 if (j == nr_tasks)
 364                         j = 0;
 365                 task2 = tasks[j];
 366                 add_sched_event_wakeup(task1, 0, task2);
 367         }
 368 }
 369
 370 static void
 371 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 372 {
 373         int ret = 0;
 374         u64 now;
 375         long long delta;
 376
 377         now = get_nsecs();
 378         delta = start_time + atom->timestamp - now;
 379
 380         switch (atom->type) {
 381                 case SCHED_EVENT_RUN:
 382                         burn_nsecs(atom->duration);
 383                         break;
 384                 case SCHED_EVENT_SLEEP:
 385                         if (atom->wait_sem)
 386                                 ret = sem_wait(atom->wait_sem);
 387                         BUG_ON(ret);
 388                         break;
 389                 case SCHED_EVENT_WAKEUP:
 390                         if (atom->wait_sem)
 391                                 ret = sem_post(atom->wait_sem);
 392                         BUG_ON(ret);
 393                         break;
 394                 case SCHED_EVENT_MIGRATION:
 395                         break;
 396                 default:
 397                         BUG_ON(1);
 398         }
 399 }
 400
 401 static u64 get_cpu_usage_nsec_parent(void)
 402 {
 403         struct rusage ru;
 404         u64 sum;
 405         int err;
 406
 407         err = getrusage(RUSAGE_SELF, &ru);
 408         BUG_ON(err);
 409
 410         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 411         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 412
 413         return sum;
 414 }
 415
 416 static int self_open_counters(void)
 417 {
 418         struct perf_event_attr attr;
 419         int fd;
 420
 421         memset(&attr, 0, sizeof(attr));
 422
 423         attr.type = PERF_TYPE_SOFTWARE;
 424         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 425
 426         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 427
 428         if (fd < 0)
 429                 die("Error: sys_perf_event_open() syscall returned"
 430                     "with %d (%s)\n", fd, strerror(errno));
 431         return fd;
 432 }
 433
 434 static u64 get_cpu_usage_nsec_self(int fd)
 435 {
 436         u64 runtime;
 437         int ret;
 438
 439         ret = read(fd, &runtime, sizeof(runtime));
 440         BUG_ON(ret != sizeof(runtime));
 441
 442         return runtime;
 443 }
 444
 445 static void *thread_func(void *ctx)
 446 {
 447         struct task_desc *this_task = ctx;
 448         u64 cpu_usage_0, cpu_usage_1;
 449         unsigned long i, ret;
 450         char comm2[22];
 451         int fd;
 452
 453         sprintf(comm2, ":%s", this_task->comm);
 454         prctl(PR_SET_NAME, comm2);
 455         fd = self_open_counters();
 456
 457 again:
 458         ret = sem_post(&this_task->ready_for_work);
 459         BUG_ON(ret);
 460         ret = pthread_mutex_lock(&start_work_mutex);
 461         BUG_ON(ret);
 462         ret = pthread_mutex_unlock(&start_work_mutex);
 463         BUG_ON(ret);
 464
 465         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 466
 467         for (i = 0; i < this_task->nr_events; i++) {
 468                 this_task->curr_event = i;
 469                 process_sched_event(this_task, this_task->atoms[i]);
 470         }
 471
 472         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 473         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 474         ret = sem_post(&this_task->work_done_sem);
 475         BUG_ON(ret);
 476
 477         ret = pthread_mutex_lock(&work_done_wait_mutex);
 478         BUG_ON(ret);
 479         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 480         BUG_ON(ret);
 481
 482         goto again;
 483 }
 484
 485 static void create_tasks(void)
 486 {
 487         struct task_desc *task;
 488         pthread_attr_t attr;
 489         unsigned long i;
 490         int err;
 491
 492         err = pthread_attr_init(&attr);
 493         BUG_ON(err);
 494         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 495         BUG_ON(err);
 496         err = pthread_mutex_lock(&start_work_mutex);
 497         BUG_ON(err);
 498         err = pthread_mutex_lock(&work_done_wait_mutex);
 499         BUG_ON(err);
 500         for (i = 0; i < nr_tasks; i++) {
 501                 task = tasks[i];
 502                 sem_init(&task->sleep_sem, 0, 0);
 503                 sem_init(&task->ready_for_work, 0, 0);
 504                 sem_init(&task->work_done_sem, 0, 0);
 505                 task->curr_event = 0;
 506                 err = pthread_create(&task->thread, &attr, thread_func, task);
 507                 BUG_ON(err);
 508         }
 509 }
 510
 511 static void wait_for_tasks(void)
 512 {
 513         u64 cpu_usage_0, cpu_usage_1;
 514         struct task_desc *task;
 515         unsigned long i, ret;
 516
 517         start_time = get_nsecs();
 518         cpu_usage = 0;
 519         pthread_mutex_unlock(&work_done_wait_mutex);
 520
 521         for (i = 0; i < nr_tasks; i++) {
 522                 task = tasks[i];
 523                 ret = sem_wait(&task->ready_for_work);
 524                 BUG_ON(ret);
 525                 sem_init(&task->ready_for_work, 0, 0);
 526         }
 527         ret = pthread_mutex_lock(&work_done_wait_mutex);
 528         BUG_ON(ret);
 529
 530         cpu_usage_0 = get_cpu_usage_nsec_parent();
 531
 532         pthread_mutex_unlock(&start_work_mutex);
 533
 534         for (i = 0; i < nr_tasks; i++) {
 535                 task = tasks[i];
 536                 ret = sem_wait(&task->work_done_sem);
 537                 BUG_ON(ret);
 538                 sem_init(&task->work_done_sem, 0, 0);
 539                 cpu_usage += task->cpu_usage;
 540                 task->cpu_usage = 0;
 541         }
 542
 543         cpu_usage_1 = get_cpu_usage_nsec_parent();
 544         if (!runavg_cpu_usage)
 545                 runavg_cpu_usage = cpu_usage;
 546         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 547
 548         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 549         if (!runavg_parent_cpu_usage)
 550                 runavg_parent_cpu_usage = parent_cpu_usage;
 551         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 552                                    parent_cpu_usage)/10;
 553
 554         ret = pthread_mutex_lock(&start_work_mutex);
 555         BUG_ON(ret);
 556
 557         for (i = 0; i < nr_tasks; i++) {
 558                 task = tasks[i];
 559                 sem_init(&task->sleep_sem, 0, 0);
 560                 task->curr_event = 0;
 561         }
 562 }
 563
 564 static void run_one_test(void)
 565 {
 566         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 567
 568         T0 = get_nsecs();
 569         wait_for_tasks();
 570         T1 = get_nsecs();
 571
 572         delta = T1 - T0;
 573         sum_runtime += delta;
 574         nr_runs++;
 575
 576         avg_delta = sum_runtime / nr_runs;
 577         if (delta < avg_delta)
 578                 fluct = avg_delta - delta;
 579         else
 580                 fluct = delta - avg_delta;
 581         sum_fluct += fluct;
 582         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 583         if (!run_avg)
 584                 run_avg = delta;
 585         run_avg = (run_avg*9 + delta)/10;
 586
 587         printf("#%-3ld: %0.3f, ",
 588                 nr_runs, (double)delta/1000000.0);
 589
 590         printf("ravg: %0.2f, ",
 591                 (double)run_avg/1e6);
 592
 593         printf("cpu: %0.2f / %0.2f",
 594                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 595
 596 #if 0
 597         /*
 598          * rusage statistics done by the parent, these are less
 599          * accurate than the sum_exec_runtime based statistics:
 600          */
 601         printf(" [%0.2f / %0.2f]",
 602                 (double)parent_cpu_usage/1e6,
 603                 (double)runavg_parent_cpu_usage/1e6);
 604 #endif
 605
 606         printf("\n");
 607
 608         if (nr_sleep_corrections)
 609                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 610         nr_sleep_corrections = 0;
 611 }
 612
 613 static void test_calibrations(void)
 614 {
 615         u64 T0, T1;
 616
 617         T0 = get_nsecs();
 618         burn_nsecs(1e6);
 619         T1 = get_nsecs();
 620
 621         printf("the run test took %Ld nsecs\n", T1-T0);
 622
 623         T0 = get_nsecs();
 624         sleep_nsecs(1e6);
 625         T1 = get_nsecs();
 626
 627         printf("the sleep test took %Ld nsecs\n", T1-T0);
 628 }
 629
 630 #define FILL_FIELD(ptr, field, event, data)     \
 631         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 632
 633 #define FILL_ARRAY(ptr, array, event, data)                     \
 634 do {                                                            \
 635         void *__array = raw_field_ptr(event, #array, data);     \
 636         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 637 } while(0)
 638
 639 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 640 do {                                                            \
 641         FILL_FIELD(ptr, common_type, event, data);              \
 642         FILL_FIELD(ptr, common_flags, event, data);             \
 643         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 644         FILL_FIELD(ptr, common_pid, event, data);               \
 645         FILL_FIELD(ptr, common_tgid, event, data);              \
 646 } while (0)
 647
 648
 649
 650 struct trace_switch_event {
 651         u32 size;
 652
 653         u16 common_type;
 654         u8 common_flags;
 655         u8 common_preempt_count;
 656         u32 common_pid;
 657         u32 common_tgid;
 658
 659         char prev_comm[16];
 660         u32 prev_pid;
 661         u32 prev_prio;
 662         u64 prev_state;
 663         char next_comm[16];
 664         u32 next_pid;
 665         u32 next_prio;
 666 };
 667
 668 struct trace_runtime_event {
 669         u32 size;
 670
 671         u16 common_type;
 672         u8 common_flags;
 673         u8 common_preempt_count;
 674         u32 common_pid;
 675         u32 common_tgid;
 676
 677         char comm[16];
 678         u32 pid;
 679         u64 runtime;
 680         u64 vruntime;
 681 };
 682
 683 struct trace_wakeup_event {
 684         u32 size;
 685
 686         u16 common_type;
 687         u8 common_flags;
 688         u8 common_preempt_count;
 689         u32 common_pid;
 690         u32 common_tgid;
 691
 692         char comm[16];
 693         u32 pid;
 694
 695         u32 prio;
 696         u32 success;
 697         u32 cpu;
 698 };
 699
 700 struct trace_fork_event {
 701         u32 size;
 702
 703         u16 common_type;
 704         u8 common_flags;
 705         u8 common_preempt_count;
 706         u32 common_pid;
 707         u32 common_tgid;
 708
 709         char parent_comm[16];
 710         u32 parent_pid;
 711         char child_comm[16];
 712         u32 child_pid;
 713 };
 714
 715 struct trace_migrate_task_event {
 716         u32 size;
 717
 718         u16 common_type;
 719         u8 common_flags;
 720         u8 common_preempt_count;
 721         u32 common_pid;
 722         u32 common_tgid;
 723
 724         char comm[16];
 725         u32 pid;
 726
 727         u32 prio;
 728         u32 cpu;
 729 };
 730
 731 struct trace_sched_handler {
 732         void (*switch_event)(struct trace_switch_event *,
 733                              struct event *,
 734                              int cpu,
 735                              u64 timestamp,
 736                              struct thread *thread);
 737
 738         void (*runtime_event)(struct trace_runtime_event *,
 739                               struct event *,
 740                               int cpu,
 741                               u64 timestamp,
 742                               struct thread *thread);
 743
 744         void (*wakeup_event)(struct trace_wakeup_event *,
 745                              struct event *,
 746                              int cpu,
 747                              u64 timestamp,
 748                              struct thread *thread);
 749
 750         void (*fork_event)(struct trace_fork_event *,
 751                            struct event *,
 752                            int cpu,
 753                            u64 timestamp,
 754                            struct thread *thread);
 755
 756         void (*migrate_task_event)(struct trace_migrate_task_event *,
 757                            struct event *,
 758                            int cpu,
 759                            u64 timestamp,
 760                            struct thread *thread);
 761 };
 762
 763
 764 static void
 765 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 766                     struct event *event,
 767                     int cpu __used,
 768                     u64 timestamp __used,
 769                     struct thread *thread __used)
 770 {
 771         struct task_desc *waker, *wakee;
 772
 773         if (verbose) {
 774                 printf("sched_wakeup event %p\n", event);
 775
 776                 printf(" ... pid %d woke up %s/%d\n",
 777                         wakeup_event->common_pid,
 778                         wakeup_event->comm,
 779                         wakeup_event->pid);
 780         }
 781
 782         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 783         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 784
 785         add_sched_event_wakeup(waker, timestamp, wakee);
 786 }
 787
 788 static u64 cpu_last_switched[MAX_CPUS];
 789
 790 static void
 791 replay_switch_event(struct trace_switch_event *switch_event,
 792                     struct event *event,
 793                     int cpu,
 794                     u64 timestamp,
 795                     struct thread *thread __used)
 796 {
 797         struct task_desc *prev, *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: %Ld < 0 ?\n", delta);
 815
 816         if (verbose) {
 817                 printf(" ... switch from %s/%d to %s/%d [ran %Ld 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 event *event __used,
1026                      int cpu,
1027                      u64 timestamp,
1028                      struct thread *thread __used)
1029 {
1030         struct work_atoms *out_events, *in_events;
1031         struct thread *sched_out, *sched_in;
1032         u64 timestamp0;
1033         s64 delta;
1034
1035         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1036
1037         timestamp0 = cpu_last_switched[cpu];
1038         cpu_last_switched[cpu] = timestamp;
1039         if (timestamp0)
1040                 delta = timestamp - timestamp0;
1041         else
1042                 delta = 0;
1043
1044         if (delta < 0)
1045                 die("hm, delta: %Ld < 0 ?\n", delta);
1046
1047
1048         sched_out = threads__findnew(switch_event->prev_pid);
1049         sched_in = threads__findnew(switch_event->next_pid);
1050
1051         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1052         if (!out_events) {
1053                 thread_atoms_insert(sched_out);
1054                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1055                 if (!out_events)
1056                         die("out-event: Internal tree error");
1057         }
1058         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1059
1060         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1061         if (!in_events) {
1062                 thread_atoms_insert(sched_in);
1063                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1064                 if (!in_events)
1065                         die("in-event: Internal tree error");
1066                 /*
1067                  * Take came in we have not heard about yet,
1068                  * add in an initial atom in runnable state:
1069                  */
1070                 add_sched_out_event(in_events, 'R', timestamp);
1071         }
1072         add_sched_in_event(in_events, timestamp);
1073 }
1074
1075 static void
1076 latency_runtime_event(struct trace_runtime_event *runtime_event,
1077                      struct event *event __used,
1078                      int cpu,
1079                      u64 timestamp,
1080                      struct thread *this_thread __used)
1081 {
1082         struct thread *thread = threads__findnew(runtime_event->pid);
1083         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1084
1085         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1086         if (!atoms) {
1087                 thread_atoms_insert(thread);
1088                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1089                 if (!atoms)
1090                         die("in-event: Internal tree error");
1091                 add_sched_out_event(atoms, 'R', timestamp);
1092         }
1093
1094         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1095 }
1096
1097 static void
1098 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1099                      struct event *__event __used,
1100                      int cpu __used,
1101                      u64 timestamp,
1102                      struct thread *thread __used)
1103 {
1104         struct work_atoms *atoms;
1105         struct work_atom *atom;
1106         struct thread *wakee;
1107
1108         /* Note for later, it may be interesting to observe the failing cases */
1109         if (!wakeup_event->success)
1110                 return;
1111
1112         wakee = threads__findnew(wakeup_event->pid);
1113         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1114         if (!atoms) {
1115                 thread_atoms_insert(wakee);
1116                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1117                 if (!atoms)
1118                         die("wakeup-event: Internal tree error");
1119                 add_sched_out_event(atoms, 'S', timestamp);
1120         }
1121
1122         BUG_ON(list_empty(&atoms->work_list));
1123
1124         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1125
1126         /*
1127          * You WILL be missing events if you've recorded only
1128          * one CPU, or are only looking at only one, so don't
1129          * make useless noise.
1130          */
1131         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1132                 nr_state_machine_bugs++;
1133
1134         nr_timestamps++;
1135         if (atom->sched_out_time > timestamp) {
1136                 nr_unordered_timestamps++;
1137                 return;
1138         }
1139
1140         atom->state = THREAD_WAIT_CPU;
1141         atom->wake_up_time = timestamp;
1142 }
1143
1144 static void
1145 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1146                      struct event *__event __used,
1147                      int cpu __used,
1148                      u64 timestamp,
1149                      struct thread *thread __used)
1150 {
1151         struct work_atoms *atoms;
1152         struct work_atom *atom;
1153         struct thread *migrant;
1154
1155         /*
1156          * Only need to worry about migration when profiling one CPU.
1157          */
1158         if (profile_cpu == -1)
1159                 return;
1160
1161         migrant = threads__findnew(migrate_task_event->pid);
1162         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1163         if (!atoms) {
1164                 thread_atoms_insert(migrant);
1165                 register_pid(migrant->pid, migrant->comm);
1166                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1167                 if (!atoms)
1168                         die("migration-event: Internal tree error");
1169                 add_sched_out_event(atoms, 'R', timestamp);
1170         }
1171
1172         BUG_ON(list_empty(&atoms->work_list));
1173
1174         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1175         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1176
1177         nr_timestamps++;
1178
1179         if (atom->sched_out_time > timestamp)
1180                 nr_unordered_timestamps++;
1181 }
1182
1183 static struct trace_sched_handler lat_ops  = {
1184         .wakeup_event           = latency_wakeup_event,
1185         .switch_event           = latency_switch_event,
1186         .runtime_event          = latency_runtime_event,
1187         .fork_event             = latency_fork_event,
1188         .migrate_task_event     = latency_migrate_task_event,
1189 };
1190
1191 static void output_lat_thread(struct work_atoms *work_list)
1192 {
1193         int i;
1194         int ret;
1195         u64 avg;
1196
1197         if (!work_list->nb_atoms)
1198                 return;
1199         /*
1200          * Ignore idle threads:
1201          */
1202         if (!strcmp(work_list->thread->comm, "swapper"))
1203                 return;
1204
1205         all_runtime += work_list->total_runtime;
1206         all_count += work_list->nb_atoms;
1207
1208         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1209
1210         for (i = 0; i < 24 - ret; i++)
1211                 printf(" ");
1212
1213         avg = work_list->total_lat / work_list->nb_atoms;
1214
1215         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1216               (double)work_list->total_runtime / 1e6,
1217                  work_list->nb_atoms, (double)avg / 1e6,
1218                  (double)work_list->max_lat / 1e6,
1219                  (double)work_list->max_lat_at / 1e9);
1220 }
1221
1222 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1223 {
1224         if (l->thread->pid < r->thread->pid)
1225                 return -1;
1226         if (l->thread->pid > r->thread->pid)
1227                 return 1;
1228
1229         return 0;
1230 }
1231
1232 static struct sort_dimension pid_sort_dimension = {
1233         .name                   = "pid",
1234         .cmp                    = pid_cmp,
1235 };
1236
1237 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1238 {
1239         u64 avgl, avgr;
1240
1241         if (!l->nb_atoms)
1242                 return -1;
1243
1244         if (!r->nb_atoms)
1245                 return 1;
1246
1247         avgl = l->total_lat / l->nb_atoms;
1248         avgr = r->total_lat / r->nb_atoms;
1249
1250         if (avgl < avgr)
1251                 return -1;
1252         if (avgl > avgr)
1253                 return 1;
1254
1255         return 0;
1256 }
1257
1258 static struct sort_dimension avg_sort_dimension = {
1259         .name                   = "avg",
1260         .cmp                    = avg_cmp,
1261 };
1262
1263 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1264 {
1265         if (l->max_lat < r->max_lat)
1266                 return -1;
1267         if (l->max_lat > r->max_lat)
1268                 return 1;
1269
1270         return 0;
1271 }
1272
1273 static struct sort_dimension max_sort_dimension = {
1274         .name                   = "max",
1275         .cmp                    = max_cmp,
1276 };
1277
1278 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1279 {
1280         if (l->nb_atoms < r->nb_atoms)
1281                 return -1;
1282         if (l->nb_atoms > r->nb_atoms)
1283                 return 1;
1284
1285         return 0;
1286 }
1287
1288 static struct sort_dimension switch_sort_dimension = {
1289         .name                   = "switch",
1290         .cmp                    = switch_cmp,
1291 };
1292
1293 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1294 {
1295         if (l->total_runtime < r->total_runtime)
1296                 return -1;
1297         if (l->total_runtime > r->total_runtime)
1298                 return 1;
1299
1300         return 0;
1301 }
1302
1303 static struct sort_dimension runtime_sort_dimension = {
1304         .name                   = "runtime",
1305         .cmp                    = runtime_cmp,
1306 };
1307
1308 static struct sort_dimension *available_sorts[] = {
1309         &pid_sort_dimension,
1310         &avg_sort_dimension,
1311         &max_sort_dimension,
1312         &switch_sort_dimension,
1313         &runtime_sort_dimension,
1314 };
1315
1316 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1317
1318 static LIST_HEAD(sort_list);
1319
1320 static int sort_dimension__add(const char *tok, struct list_head *list)
1321 {
1322         int i;
1323
1324         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1325                 if (!strcmp(available_sorts[i]->name, tok)) {
1326                         list_add_tail(&available_sorts[i]->list, list);
1327
1328                         return 0;
1329                 }
1330         }
1331
1332         return -1;
1333 }
1334
1335 static void setup_sorting(void);
1336
1337 static void sort_lat(void)
1338 {
1339         struct rb_node *node;
1340
1341         for (;;) {
1342                 struct work_atoms *data;
1343                 node = rb_first(&atom_root);
1344                 if (!node)
1345                         break;
1346
1347                 rb_erase(node, &atom_root);
1348                 data = rb_entry(node, struct work_atoms, node);
1349                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1350         }
1351 }
1352
1353 static struct trace_sched_handler *trace_handler;
1354
1355 static void
1356 process_sched_wakeup_event(void *data,
1357                            struct event *event,
1358                            int cpu __used,
1359                            u64 timestamp __used,
1360                            struct thread *thread __used)
1361 {
1362         struct trace_wakeup_event wakeup_event;
1363
1364         FILL_COMMON_FIELDS(wakeup_event, event, data);
1365
1366         FILL_ARRAY(wakeup_event, comm, event, data);
1367         FILL_FIELD(wakeup_event, pid, event, data);
1368         FILL_FIELD(wakeup_event, prio, event, data);
1369         FILL_FIELD(wakeup_event, success, event, data);
1370         FILL_FIELD(wakeup_event, cpu, event, data);
1371
1372         if (trace_handler->wakeup_event)
1373                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1374 }
1375
1376 /*
1377  * Track the current task - that way we can know whether there's any
1378  * weird events, such as a task being switched away that is not current.
1379  */
1380 static int max_cpu;
1381
1382 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1383
1384 static struct thread *curr_thread[MAX_CPUS];
1385
1386 static char next_shortname1 = 'A';
1387 static char next_shortname2 = '0';
1388
1389 static void
1390 map_switch_event(struct trace_switch_event *switch_event,
1391                  struct event *event __used,
1392                  int this_cpu,
1393                  u64 timestamp,
1394                  struct thread *thread __used)
1395 {
1396         struct thread *sched_out, *sched_in;
1397         int new_shortname;
1398         u64 timestamp0;
1399         s64 delta;
1400         int cpu;
1401
1402         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1403
1404         if (this_cpu > max_cpu)
1405                 max_cpu = this_cpu;
1406
1407         timestamp0 = cpu_last_switched[this_cpu];
1408         cpu_last_switched[this_cpu] = timestamp;
1409         if (timestamp0)
1410                 delta = timestamp - timestamp0;
1411         else
1412                 delta = 0;
1413
1414         if (delta < 0)
1415                 die("hm, delta: %Ld < 0 ?\n", delta);
1416
1417
1418         sched_out = threads__findnew(switch_event->prev_pid);
1419         sched_in = threads__findnew(switch_event->next_pid);
1420
1421         curr_thread[this_cpu] = sched_in;
1422
1423         printf("  ");
1424
1425         new_shortname = 0;
1426         if (!sched_in->shortname[0]) {
1427                 sched_in->shortname[0] = next_shortname1;
1428                 sched_in->shortname[1] = next_shortname2;
1429
1430                 if (next_shortname1 < 'Z') {
1431                         next_shortname1++;
1432                 } else {
1433                         next_shortname1='A';
1434                         if (next_shortname2 < '9') {
1435                                 next_shortname2++;
1436                         } else {
1437                                 next_shortname2='0';
1438                         }
1439                 }
1440                 new_shortname = 1;
1441         }
1442
1443         for (cpu = 0; cpu <= max_cpu; cpu++) {
1444                 if (cpu != this_cpu)
1445                         printf(" ");
1446                 else
1447                         printf("*");
1448
1449                 if (curr_thread[cpu]) {
1450                         if (curr_thread[cpu]->pid)
1451                                 printf("%2s ", curr_thread[cpu]->shortname);
1452                         else
1453                                 printf(".  ");
1454                 } else
1455                         printf("   ");
1456         }
1457
1458         printf("  %12.6f secs ", (double)timestamp/1e9);
1459         if (new_shortname) {
1460                 printf("%s => %s:%d\n",
1461                         sched_in->shortname, sched_in->comm, sched_in->pid);
1462         } else {
1463                 printf("\n");
1464         }
1465 }
1466
1467
1468 static void
1469 process_sched_switch_event(void *data,
1470                            struct event *event,
1471                            int this_cpu,
1472                            u64 timestamp __used,
1473                            struct thread *thread __used)
1474 {
1475         struct trace_switch_event switch_event;
1476
1477         FILL_COMMON_FIELDS(switch_event, event, data);
1478
1479         FILL_ARRAY(switch_event, prev_comm, event, data);
1480         FILL_FIELD(switch_event, prev_pid, event, data);
1481         FILL_FIELD(switch_event, prev_prio, event, data);
1482         FILL_FIELD(switch_event, prev_state, event, data);
1483         FILL_ARRAY(switch_event, next_comm, event, data);
1484         FILL_FIELD(switch_event, next_pid, event, data);
1485         FILL_FIELD(switch_event, next_prio, event, data);
1486
1487         if (curr_pid[this_cpu] != (u32)-1) {
1488                 /*
1489                  * Are we trying to switch away a PID that is
1490                  * not current?
1491                  */
1492                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1493                         nr_context_switch_bugs++;
1494         }
1495         if (trace_handler->switch_event)
1496                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1497
1498         curr_pid[this_cpu] = switch_event.next_pid;
1499 }
1500
1501 static void
1502 process_sched_runtime_event(void *data,
1503                            struct event *event,
1504                            int cpu __used,
1505                            u64 timestamp __used,
1506                            struct thread *thread __used)
1507 {
1508         struct trace_runtime_event runtime_event;
1509
1510         FILL_ARRAY(runtime_event, comm, event, data);
1511         FILL_FIELD(runtime_event, pid, event, data);
1512         FILL_FIELD(runtime_event, runtime, event, data);
1513         FILL_FIELD(runtime_event, vruntime, event, data);
1514
1515         if (trace_handler->runtime_event)
1516                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1517 }
1518
1519 static void
1520 process_sched_fork_event(void *data,
1521                          struct event *event,
1522                          int cpu __used,
1523                          u64 timestamp __used,
1524                          struct thread *thread __used)
1525 {
1526         struct trace_fork_event fork_event;
1527
1528         FILL_COMMON_FIELDS(fork_event, event, data);
1529
1530         FILL_ARRAY(fork_event, parent_comm, event, data);
1531         FILL_FIELD(fork_event, parent_pid, event, data);
1532         FILL_ARRAY(fork_event, child_comm, event, data);
1533         FILL_FIELD(fork_event, child_pid, event, data);
1534
1535         if (trace_handler->fork_event)
1536                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1537 }
1538
1539 static void
1540 process_sched_exit_event(struct event *event,
1541                          int cpu __used,
1542                          u64 timestamp __used,
1543                          struct thread *thread __used)
1544 {
1545         if (verbose)
1546                 printf("sched_exit event %p\n", event);
1547 }
1548
1549 static void
1550 process_sched_migrate_task_event(void *data,
1551                            struct event *event,
1552                            int cpu __used,
1553                            u64 timestamp __used,
1554                            struct thread *thread __used)
1555 {
1556         struct trace_migrate_task_event migrate_task_event;
1557
1558         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1559
1560         FILL_ARRAY(migrate_task_event, comm, event, data);
1561         FILL_FIELD(migrate_task_event, pid, event, data);
1562         FILL_FIELD(migrate_task_event, prio, event, data);
1563         FILL_FIELD(migrate_task_event, cpu, event, data);
1564
1565         if (trace_handler->migrate_task_event)
1566                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1567 }
1568
1569 static void
1570 process_raw_event(event_t *raw_event __used, void *data,
1571                   int cpu, u64 timestamp, struct thread *thread)
1572 {
1573         struct event *event;
1574         int type;
1575
1576
1577         type = trace_parse_common_type(data);
1578         event = trace_find_event(type);
1579
1580         if (!strcmp(event->name, "sched_switch"))
1581                 process_sched_switch_event(data, event, cpu, timestamp, thread);
1582         if (!strcmp(event->name, "sched_stat_runtime"))
1583                 process_sched_runtime_event(data, event, cpu, timestamp, thread);
1584         if (!strcmp(event->name, "sched_wakeup"))
1585                 process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1586         if (!strcmp(event->name, "sched_wakeup_new"))
1587                 process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1588         if (!strcmp(event->name, "sched_process_fork"))
1589                 process_sched_fork_event(data, event, cpu, timestamp, thread);
1590         if (!strcmp(event->name, "sched_process_exit"))
1591                 process_sched_exit_event(event, cpu, timestamp, thread);
1592         if (!strcmp(event->name, "sched_migrate_task"))
1593                 process_sched_migrate_task_event(data, event, cpu, timestamp, thread);
1594 }
1595
1596 static int process_sample_event(event_t *event,
1597                                 struct perf_session *session __used)
1598 {
1599         struct sample_data data;
1600         struct thread *thread;
1601
1602         if (!(sample_type & PERF_SAMPLE_RAW))
1603                 return 0;
1604
1605         memset(&data, 0, sizeof(data));
1606         data.time = -1;
1607         data.cpu = -1;
1608         data.period = -1;
1609
1610         event__parse_sample(event, sample_type, &data);
1611
1612         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1613                 event->header.misc,
1614                 data.pid, data.tid,
1615                 (void *)(long)data.ip,
1616                 (long long)data.period);
1617
1618         thread = threads__findnew(data.pid);
1619         if (thread == NULL) {
1620                 pr_debug("problem processing %d event, skipping it.\n",
1621                          event->header.type);
1622                 return -1;
1623         }
1624
1625         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1626
1627         if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1628                 return 0;
1629
1630         process_raw_event(event, data.raw_data, data.cpu, data.time, thread);
1631
1632         return 0;
1633 }
1634
1635 static int process_lost_event(event_t *event __used,
1636                               struct perf_session *session __used)
1637 {
1638         nr_lost_chunks++;
1639         nr_lost_events += event->lost.lost;
1640
1641         return 0;
1642 }
1643
1644 static int sample_type_check(u64 type)
1645 {
1646         sample_type = type;
1647
1648         if (!(sample_type & PERF_SAMPLE_RAW)) {
1649                 fprintf(stderr,
1650                         "No trace sample to read. Did you call perf record "
1651                         "without -R?");
1652                 return -1;
1653         }
1654
1655         return 0;
1656 }
1657
1658 static struct perf_event_ops event_ops = {
1659         .process_sample_event   = process_sample_event,
1660         .process_comm_event     = event__process_comm,
1661         .process_lost_event     = process_lost_event,
1662         .sample_type_check      = sample_type_check,
1663 };
1664
1665 static int read_events(void)
1666 {
1667         int err;
1668         struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
1669
1670         if (session == NULL)
1671                 return -ENOMEM;
1672
1673         err = perf_session__process_events(session, &event_ops);
1674         perf_session__delete(session);
1675         return err;
1676 }
1677
1678 static void print_bad_events(void)
1679 {
1680         if (nr_unordered_timestamps && nr_timestamps) {
1681                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1682                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1683                         nr_unordered_timestamps, nr_timestamps);
1684         }
1685         if (nr_lost_events && nr_events) {
1686                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1687                         (double)nr_lost_events/(double)nr_events*100.0,
1688                         nr_lost_events, nr_events, nr_lost_chunks);
1689         }
1690         if (nr_state_machine_bugs && nr_timestamps) {
1691                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1692                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1693                         nr_state_machine_bugs, nr_timestamps);
1694                 if (nr_lost_events)
1695                         printf(" (due to lost events?)");
1696                 printf("\n");
1697         }
1698         if (nr_context_switch_bugs && nr_timestamps) {
1699                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1700                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1701                         nr_context_switch_bugs, nr_timestamps);
1702                 if (nr_lost_events)
1703                         printf(" (due to lost events?)");
1704                 printf("\n");
1705         }
1706 }
1707
1708 static void __cmd_lat(void)
1709 {
1710         struct rb_node *next;
1711
1712         setup_pager();
1713         read_events();
1714         sort_lat();
1715
1716         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1717         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1718         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1719
1720         next = rb_first(&sorted_atom_root);
1721
1722         while (next) {
1723                 struct work_atoms *work_list;
1724
1725                 work_list = rb_entry(next, struct work_atoms, node);
1726                 output_lat_thread(work_list);
1727                 next = rb_next(next);
1728         }
1729
1730         printf(" -----------------------------------------------------------------------------------------\n");
1731         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1732                 (double)all_runtime/1e6, all_count);
1733
1734         printf(" ---------------------------------------------------\n");
1735
1736         print_bad_events();
1737         printf("\n");
1738
1739 }
1740
1741 static struct trace_sched_handler map_ops  = {
1742         .wakeup_event           = NULL,
1743         .switch_event           = map_switch_event,
1744         .runtime_event          = NULL,
1745         .fork_event             = NULL,
1746 };
1747
1748 static void __cmd_map(void)
1749 {
1750         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1751
1752         setup_pager();
1753         read_events();
1754         print_bad_events();
1755 }
1756
1757 static void __cmd_replay(void)
1758 {
1759         unsigned long i;
1760
1761         calibrate_run_measurement_overhead();
1762         calibrate_sleep_measurement_overhead();
1763
1764         test_calibrations();
1765
1766         read_events();
1767
1768         printf("nr_run_events:        %ld\n", nr_run_events);
1769         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1770         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1771
1772         if (targetless_wakeups)
1773                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1774         if (multitarget_wakeups)
1775                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1776         if (nr_run_events_optimized)
1777                 printf("run atoms optimized: %ld\n",
1778                         nr_run_events_optimized);
1779
1780         print_task_traces();
1781         add_cross_task_wakeups();
1782
1783         create_tasks();
1784         printf("------------------------------------------------------------\n");
1785         for (i = 0; i < replay_repeat; i++)
1786                 run_one_test();
1787 }
1788
1789
1790 static const char * const sched_usage[] = {
1791         "perf sched [<options>] {record|latency|map|replay|trace}",
1792         NULL
1793 };
1794
1795 static const struct option sched_options[] = {
1796         OPT_STRING('i', "input", &input_name, "file",
1797                     "input file name"),
1798         OPT_BOOLEAN('v', "verbose", &verbose,
1799                     "be more verbose (show symbol address, etc)"),
1800         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1801                     "dump raw trace in ASCII"),
1802         OPT_END()
1803 };
1804
1805 static const char * const latency_usage[] = {
1806         "perf sched latency [<options>]",
1807         NULL
1808 };
1809
1810 static const struct option latency_options[] = {
1811         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1812                    "sort by key(s): runtime, switch, avg, max"),
1813         OPT_BOOLEAN('v', "verbose", &verbose,
1814                     "be more verbose (show symbol address, etc)"),
1815         OPT_INTEGER('C', "CPU", &profile_cpu,
1816                     "CPU to profile on"),
1817         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1818                     "dump raw trace in ASCII"),
1819         OPT_END()
1820 };
1821
1822 static const char * const replay_usage[] = {
1823         "perf sched replay [<options>]",
1824         NULL
1825 };
1826
1827 static const struct option replay_options[] = {
1828         OPT_INTEGER('r', "repeat", &replay_repeat,
1829                     "repeat the workload replay N times (-1: infinite)"),
1830         OPT_BOOLEAN('v', "verbose", &verbose,
1831                     "be more verbose (show symbol address, etc)"),
1832         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1833                     "dump raw trace in ASCII"),
1834         OPT_END()
1835 };
1836
1837 static void setup_sorting(void)
1838 {
1839         char *tmp, *tok, *str = strdup(sort_order);
1840
1841         for (tok = strtok_r(str, ", ", &tmp);
1842                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1843                 if (sort_dimension__add(tok, &sort_list) < 0) {
1844                         error("Unknown --sort key: `%s'", tok);
1845                         usage_with_options(latency_usage, latency_options);
1846                 }
1847         }
1848
1849         free(str);
1850
1851         sort_dimension__add("pid", &cmp_pid);
1852 }
1853
1854 static const char *record_args[] = {
1855         "record",
1856         "-a",
1857         "-R",
1858         "-M",
1859         "-f",
1860         "-m", "1024",
1861         "-c", "1",
1862         "-e", "sched:sched_switch:r",
1863         "-e", "sched:sched_stat_wait:r",
1864         "-e", "sched:sched_stat_sleep:r",
1865         "-e", "sched:sched_stat_iowait:r",
1866         "-e", "sched:sched_stat_runtime:r",
1867         "-e", "sched:sched_process_exit:r",
1868         "-e", "sched:sched_process_fork:r",
1869         "-e", "sched:sched_wakeup:r",
1870         "-e", "sched:sched_migrate_task:r",
1871 };
1872
1873 static int __cmd_record(int argc, const char **argv)
1874 {
1875         unsigned int rec_argc, i, j;
1876         const char **rec_argv;
1877
1878         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1879         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1880
1881         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1882                 rec_argv[i] = strdup(record_args[i]);
1883
1884         for (j = 1; j < (unsigned int)argc; j++, i++)
1885                 rec_argv[i] = argv[j];
1886
1887         BUG_ON(i != rec_argc);
1888
1889         return cmd_record(i, rec_argv, NULL);
1890 }
1891
1892 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1893 {
1894         argc = parse_options(argc, argv, sched_options, sched_usage,
1895                              PARSE_OPT_STOP_AT_NON_OPTION);
1896         if (!argc)
1897                 usage_with_options(sched_usage, sched_options);
1898
1899         /*
1900          * Aliased to 'perf trace' for now:
1901          */
1902         if (!strcmp(argv[0], "trace"))
1903                 return cmd_trace(argc, argv, prefix);
1904
1905         symbol__init(0);
1906         if (!strncmp(argv[0], "rec", 3)) {
1907                 return __cmd_record(argc, argv);
1908         } else if (!strncmp(argv[0], "lat", 3)) {
1909                 trace_handler = &lat_ops;
1910                 if (argc > 1) {
1911                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1912                         if (argc)
1913                                 usage_with_options(latency_usage, latency_options);
1914                 }
1915                 setup_sorting();
1916                 __cmd_lat();
1917         } else if (!strcmp(argv[0], "map")) {
1918                 trace_handler = &map_ops;
1919                 setup_sorting();
1920                 __cmd_map();
1921         } else if (!strncmp(argv[0], "rep", 3)) {
1922                 trace_handler = &replay_ops;
1923                 if (argc) {
1924                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1925                         if (argc)
1926                                 usage_with_options(replay_usage, replay_options);
1927                 }
1928                 __cmd_replay();
1929         } else {
1930                 usage_with_options(sched_usage, sched_options);
1931         }
1932
1933         return 0;
1934 }