SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
  10 #include "util/parse-options.h"
  11 #include "util/trace-event.h"
  12
  13 #include "util/debug.h"
  14 #include "util/data_map.h"
  15
  16 #include <sys/types.h>
  17 #include <sys/prctl.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 unsigned long            total_comm = 0;
  26
  27 static struct perf_header       *header;
  28 static u64                      sample_type;
  29
  30 static char                     default_sort_order[] = "avg, max, switch, runtime";
  31 static char                     *sort_order = default_sort_order;
  32
  33 static int                      profile_cpu = -1;
  34
  35 static char                     *cwd;
  36 static int                      cwdlen;
  37
  38 #define PR_SET_NAME             15               /* Set process name */
  39 #define MAX_CPUS                4096
  40
  41 #define BUG_ON(x)               assert(!(x))
  42
  43 static u64                      run_measurement_overhead;
  44 static u64                      sleep_measurement_overhead;
  45
  46 #define COMM_LEN                20
  47 #define SYM_LEN                 129
  48
  49 #define MAX_PID                 65536
  50
  51 static unsigned long            nr_tasks;
  52
  53 struct sched_atom;
  54
  55 struct task_desc {
  56         unsigned long           nr;
  57         unsigned long           pid;
  58         char                    comm[COMM_LEN];
  59
  60         unsigned long           nr_events;
  61         unsigned long           curr_event;
  62         struct sched_atom       **atoms;
  63
  64         pthread_t               thread;
  65         sem_t                   sleep_sem;
  66
  67         sem_t                   ready_for_work;
  68         sem_t                   work_done_sem;
  69
  70         u64                     cpu_usage;
  71 };
  72
  73 enum sched_event_type {
  74         SCHED_EVENT_RUN,
  75         SCHED_EVENT_SLEEP,
  76         SCHED_EVENT_WAKEUP,
  77         SCHED_EVENT_MIGRATION,
  78 };
  79
  80 struct sched_atom {
  81         enum sched_event_type   type;
  82         u64                     timestamp;
  83         u64                     duration;
  84         unsigned long           nr;
  85         int                     specific_wait;
  86         sem_t                   *wait_sem;
  87         struct task_desc        *wakee;
  88 };
  89
  90 static struct task_desc         *pid_to_task[MAX_PID];
  91
  92 static struct task_desc         **tasks;
  93
  94 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  95 static u64                      start_time;
  96
  97 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  98
  99 static unsigned long            nr_run_events;
 100 static unsigned long            nr_sleep_events;
 101 static unsigned long            nr_wakeup_events;
 102
 103 static unsigned long            nr_sleep_corrections;
 104 static unsigned long            nr_run_events_optimized;
 105
 106 static unsigned long            targetless_wakeups;
 107 static unsigned long            multitarget_wakeups;
 108
 109 static u64                      cpu_usage;
 110 static u64                      runavg_cpu_usage;
 111 static u64                      parent_cpu_usage;
 112 static u64                      runavg_parent_cpu_usage;
 113
 114 static unsigned long            nr_runs;
 115 static u64                      sum_runtime;
 116 static u64                      sum_fluct;
 117 static u64                      run_avg;
 118
 119 static unsigned long            replay_repeat = 10;
 120 static unsigned long            nr_timestamps;
 121 static unsigned long            nr_unordered_timestamps;
 122 static unsigned long            nr_state_machine_bugs;
 123 static unsigned long            nr_context_switch_bugs;
 124 static unsigned long            nr_events;
 125 static unsigned long            nr_lost_chunks;
 126 static unsigned long            nr_lost_events;
 127
 128 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 129
 130 enum thread_state {
 131         THREAD_SLEEPING = 0,
 132         THREAD_WAIT_CPU,
 133         THREAD_SCHED_IN,
 134         THREAD_IGNORE
 135 };
 136
 137 struct work_atom {
 138         struct list_head        list;
 139         enum thread_state       state;
 140         u64                     sched_out_time;
 141         u64                     wake_up_time;
 142         u64                     sched_in_time;
 143         u64                     runtime;
 144 };
 145
 146 struct work_atoms {
 147         struct list_head        work_list;
 148         struct thread           *thread;
 149         struct rb_node          node;
 150         u64                     max_lat;
 151         u64                     total_lat;
 152         u64                     nb_atoms;
 153         u64                     total_runtime;
 154 };
 155
 156 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 157
 158 static struct rb_root           atom_root, sorted_atom_root;
 159
 160 static u64                      all_runtime;
 161 static u64                      all_count;
 162
 163
 164 static u64 get_nsecs(void)
 165 {
 166         struct timespec ts;
 167
 168         clock_gettime(CLOCK_MONOTONIC, &ts);
 169
 170         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 171 }
 172
 173 static void burn_nsecs(u64 nsecs)
 174 {
 175         u64 T0 = get_nsecs(), T1;
 176
 177         do {
 178                 T1 = get_nsecs();
 179         } while (T1 + run_measurement_overhead < T0 + nsecs);
 180 }
 181
 182 static void sleep_nsecs(u64 nsecs)
 183 {
 184         struct timespec ts;
 185
 186         ts.tv_nsec = nsecs % 999999999;
 187         ts.tv_sec = nsecs / 999999999;
 188
 189         nanosleep(&ts, NULL);
 190 }
 191
 192 static void calibrate_run_measurement_overhead(void)
 193 {
 194         u64 T0, T1, delta, min_delta = 1000000000ULL;
 195         int i;
 196
 197         for (i = 0; i < 10; i++) {
 198                 T0 = get_nsecs();
 199                 burn_nsecs(0);
 200                 T1 = get_nsecs();
 201                 delta = T1-T0;
 202                 min_delta = min(min_delta, delta);
 203         }
 204         run_measurement_overhead = min_delta;
 205
 206         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 207 }
 208
 209 static void calibrate_sleep_measurement_overhead(void)
 210 {
 211         u64 T0, T1, delta, min_delta = 1000000000ULL;
 212         int i;
 213
 214         for (i = 0; i < 10; i++) {
 215                 T0 = get_nsecs();
 216                 sleep_nsecs(10000);
 217                 T1 = get_nsecs();
 218                 delta = T1-T0;
 219                 min_delta = min(min_delta, delta);
 220         }
 221         min_delta -= 10000;
 222         sleep_measurement_overhead = min_delta;
 223
 224         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 225 }
 226
 227 static struct sched_atom *
 228 get_new_event(struct task_desc *task, u64 timestamp)
 229 {
 230         struct sched_atom *event = calloc(1, sizeof(*event));
 231         unsigned long idx = task->nr_events;
 232         size_t size;
 233
 234         event->timestamp = timestamp;
 235         event->nr = idx;
 236
 237         task->nr_events++;
 238         size = sizeof(struct sched_atom *) * task->nr_events;
 239         task->atoms = realloc(task->atoms, size);
 240         BUG_ON(!task->atoms);
 241
 242         task->atoms[idx] = event;
 243
 244         return event;
 245 }
 246
 247 static struct sched_atom *last_event(struct task_desc *task)
 248 {
 249         if (!task->nr_events)
 250                 return NULL;
 251
 252         return task->atoms[task->nr_events - 1];
 253 }
 254
 255 static void
 256 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 257 {
 258         struct sched_atom *event, *curr_event = last_event(task);
 259
 260         /*
 261          * optimize an existing RUN event by merging this one
 262          * to it:
 263          */
 264         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 265                 nr_run_events_optimized++;
 266                 curr_event->duration += duration;
 267                 return;
 268         }
 269
 270         event = get_new_event(task, timestamp);
 271
 272         event->type = SCHED_EVENT_RUN;
 273         event->duration = duration;
 274
 275         nr_run_events++;
 276 }
 277
 278 static void
 279 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 280                        struct task_desc *wakee)
 281 {
 282         struct sched_atom *event, *wakee_event;
 283
 284         event = get_new_event(task, timestamp);
 285         event->type = SCHED_EVENT_WAKEUP;
 286         event->wakee = wakee;
 287
 288         wakee_event = last_event(wakee);
 289         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 290                 targetless_wakeups++;
 291                 return;
 292         }
 293         if (wakee_event->wait_sem) {
 294                 multitarget_wakeups++;
 295                 return;
 296         }
 297
 298         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
 299         sem_init(wakee_event->wait_sem, 0, 0);
 300         wakee_event->specific_wait = 1;
 301         event->wait_sem = wakee_event->wait_sem;
 302
 303         nr_wakeup_events++;
 304 }
 305
 306 static void
 307 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 308                       u64 task_state __used)
 309 {
 310         struct sched_atom *event = get_new_event(task, timestamp);
 311
 312         event->type = SCHED_EVENT_SLEEP;
 313
 314         nr_sleep_events++;
 315 }
 316
 317 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 318 {
 319         struct task_desc *task;
 320
 321         BUG_ON(pid >= MAX_PID);
 322
 323         task = pid_to_task[pid];
 324
 325         if (task)
 326                 return task;
 327
 328         task = calloc(1, sizeof(*task));
 329         task->pid = pid;
 330         task->nr = nr_tasks;
 331         strcpy(task->comm, comm);
 332         /*
 333          * every task starts in sleeping state - this gets ignored
 334          * if there's no wakeup pointing to this sleep state:
 335          */
 336         add_sched_event_sleep(task, 0, 0);
 337
 338         pid_to_task[pid] = task;
 339         nr_tasks++;
 340         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 341         BUG_ON(!tasks);
 342         tasks[task->nr] = task;
 343
 344         if (verbose)
 345                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 346
 347         return task;
 348 }
 349
 350
 351 static void print_task_traces(void)
 352 {
 353         struct task_desc *task;
 354         unsigned long i;
 355
 356         for (i = 0; i < nr_tasks; i++) {
 357                 task = tasks[i];
 358                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 359                         task->nr, task->comm, task->pid, task->nr_events);
 360         }
 361 }
 362
 363 static void add_cross_task_wakeups(void)
 364 {
 365         struct task_desc *task1, *task2;
 366         unsigned long i, j;
 367
 368         for (i = 0; i < nr_tasks; i++) {
 369                 task1 = tasks[i];
 370                 j = i + 1;
 371                 if (j == nr_tasks)
 372                         j = 0;
 373                 task2 = tasks[j];
 374                 add_sched_event_wakeup(task1, 0, task2);
 375         }
 376 }
 377
 378 static void
 379 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 380 {
 381         int ret = 0;
 382         u64 now;
 383         long long delta;
 384
 385         now = get_nsecs();
 386         delta = start_time + atom->timestamp - now;
 387
 388         switch (atom->type) {
 389                 case SCHED_EVENT_RUN:
 390                         burn_nsecs(atom->duration);
 391                         break;
 392                 case SCHED_EVENT_SLEEP:
 393                         if (atom->wait_sem)
 394                                 ret = sem_wait(atom->wait_sem);
 395                         BUG_ON(ret);
 396                         break;
 397                 case SCHED_EVENT_WAKEUP:
 398                         if (atom->wait_sem)
 399                                 ret = sem_post(atom->wait_sem);
 400                         BUG_ON(ret);
 401                         break;
 402                 case SCHED_EVENT_MIGRATION:
 403                         break;
 404                 default:
 405                         BUG_ON(1);
 406         }
 407 }
 408
 409 static u64 get_cpu_usage_nsec_parent(void)
 410 {
 411         struct rusage ru;
 412         u64 sum;
 413         int err;
 414
 415         err = getrusage(RUSAGE_SELF, &ru);
 416         BUG_ON(err);
 417
 418         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 419         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 420
 421         return sum;
 422 }
 423
 424 static u64 get_cpu_usage_nsec_self(void)
 425 {
 426         char filename [] = "/proc/1234567890/sched";
 427         unsigned long msecs, nsecs;
 428         char *line = NULL;
 429         u64 total = 0;
 430         size_t len = 0;
 431         ssize_t chars;
 432         FILE *file;
 433         int ret;
 434
 435         sprintf(filename, "/proc/%d/sched", getpid());
 436         file = fopen(filename, "r");
 437         BUG_ON(!file);
 438
 439         while ((chars = getline(&line, &len, file)) != -1) {
 440                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 441                         &msecs, &nsecs);
 442                 if (ret == 2) {
 443                         total = msecs*1e6 + nsecs;
 444                         break;
 445                 }
 446         }
 447         if (line)
 448                 free(line);
 449         fclose(file);
 450
 451         return total;
 452 }
 453
 454 static void *thread_func(void *ctx)
 455 {
 456         struct task_desc *this_task = ctx;
 457         u64 cpu_usage_0, cpu_usage_1;
 458         unsigned long i, ret;
 459         char comm2[22];
 460
 461         sprintf(comm2, ":%s", this_task->comm);
 462         prctl(PR_SET_NAME, comm2);
 463
 464 again:
 465         ret = sem_post(&this_task->ready_for_work);
 466         BUG_ON(ret);
 467         ret = pthread_mutex_lock(&start_work_mutex);
 468         BUG_ON(ret);
 469         ret = pthread_mutex_unlock(&start_work_mutex);
 470         BUG_ON(ret);
 471
 472         cpu_usage_0 = get_cpu_usage_nsec_self();
 473
 474         for (i = 0; i < this_task->nr_events; i++) {
 475                 this_task->curr_event = i;
 476                 process_sched_event(this_task, this_task->atoms[i]);
 477         }
 478
 479         cpu_usage_1 = get_cpu_usage_nsec_self();
 480         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 481
 482         ret = sem_post(&this_task->work_done_sem);
 483         BUG_ON(ret);
 484
 485         ret = pthread_mutex_lock(&work_done_wait_mutex);
 486         BUG_ON(ret);
 487         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 488         BUG_ON(ret);
 489
 490         goto again;
 491 }
 492
 493 static void create_tasks(void)
 494 {
 495         struct task_desc *task;
 496         pthread_attr_t attr;
 497         unsigned long i;
 498         int err;
 499
 500         err = pthread_attr_init(&attr);
 501         BUG_ON(err);
 502         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 503         BUG_ON(err);
 504         err = pthread_mutex_lock(&start_work_mutex);
 505         BUG_ON(err);
 506         err = pthread_mutex_lock(&work_done_wait_mutex);
 507         BUG_ON(err);
 508         for (i = 0; i < nr_tasks; i++) {
 509                 task = tasks[i];
 510                 sem_init(&task->sleep_sem, 0, 0);
 511                 sem_init(&task->ready_for_work, 0, 0);
 512                 sem_init(&task->work_done_sem, 0, 0);
 513                 task->curr_event = 0;
 514                 err = pthread_create(&task->thread, &attr, thread_func, task);
 515                 BUG_ON(err);
 516         }
 517 }
 518
 519 static void wait_for_tasks(void)
 520 {
 521         u64 cpu_usage_0, cpu_usage_1;
 522         struct task_desc *task;
 523         unsigned long i, ret;
 524
 525         start_time = get_nsecs();
 526         cpu_usage = 0;
 527         pthread_mutex_unlock(&work_done_wait_mutex);
 528
 529         for (i = 0; i < nr_tasks; i++) {
 530                 task = tasks[i];
 531                 ret = sem_wait(&task->ready_for_work);
 532                 BUG_ON(ret);
 533                 sem_init(&task->ready_for_work, 0, 0);
 534         }
 535         ret = pthread_mutex_lock(&work_done_wait_mutex);
 536         BUG_ON(ret);
 537
 538         cpu_usage_0 = get_cpu_usage_nsec_parent();
 539
 540         pthread_mutex_unlock(&start_work_mutex);
 541
 542         for (i = 0; i < nr_tasks; i++) {
 543                 task = tasks[i];
 544                 ret = sem_wait(&task->work_done_sem);
 545                 BUG_ON(ret);
 546                 sem_init(&task->work_done_sem, 0, 0);
 547                 cpu_usage += task->cpu_usage;
 548                 task->cpu_usage = 0;
 549         }
 550
 551         cpu_usage_1 = get_cpu_usage_nsec_parent();
 552         if (!runavg_cpu_usage)
 553                 runavg_cpu_usage = cpu_usage;
 554         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 555
 556         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 557         if (!runavg_parent_cpu_usage)
 558                 runavg_parent_cpu_usage = parent_cpu_usage;
 559         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 560                                    parent_cpu_usage)/10;
 561
 562         ret = pthread_mutex_lock(&start_work_mutex);
 563         BUG_ON(ret);
 564
 565         for (i = 0; i < nr_tasks; i++) {
 566                 task = tasks[i];
 567                 sem_init(&task->sleep_sem, 0, 0);
 568                 task->curr_event = 0;
 569         }
 570 }
 571
 572 static void run_one_test(void)
 573 {
 574         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 575
 576         T0 = get_nsecs();
 577         wait_for_tasks();
 578         T1 = get_nsecs();
 579
 580         delta = T1 - T0;
 581         sum_runtime += delta;
 582         nr_runs++;
 583
 584         avg_delta = sum_runtime / nr_runs;
 585         if (delta < avg_delta)
 586                 fluct = avg_delta - delta;
 587         else
 588                 fluct = delta - avg_delta;
 589         sum_fluct += fluct;
 590         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 591         if (!run_avg)
 592                 run_avg = delta;
 593         run_avg = (run_avg*9 + delta)/10;
 594
 595         printf("#%-3ld: %0.3f, ",
 596                 nr_runs, (double)delta/1000000.0);
 597
 598         printf("ravg: %0.2f, ",
 599                 (double)run_avg/1e6);
 600
 601         printf("cpu: %0.2f / %0.2f",
 602                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 603
 604 #if 0
 605         /*
 606          * rusage statistics done by the parent, these are less
 607          * accurate than the sum_exec_runtime based statistics:
 608          */
 609         printf(" [%0.2f / %0.2f]",
 610                 (double)parent_cpu_usage/1e6,
 611                 (double)runavg_parent_cpu_usage/1e6);
 612 #endif
 613
 614         printf("\n");
 615
 616         if (nr_sleep_corrections)
 617                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 618         nr_sleep_corrections = 0;
 619 }
 620
 621 static void test_calibrations(void)
 622 {
 623         u64 T0, T1;
 624
 625         T0 = get_nsecs();
 626         burn_nsecs(1e6);
 627         T1 = get_nsecs();
 628
 629         printf("the run test took %Ld nsecs\n", T1-T0);
 630
 631         T0 = get_nsecs();
 632         sleep_nsecs(1e6);
 633         T1 = get_nsecs();
 634
 635         printf("the sleep test took %Ld nsecs\n", T1-T0);
 636 }
 637
 638 static int
 639 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 640 {
 641         struct thread *thread = threads__findnew(event->comm.tid);
 642
 643         dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
 644                 (void *)(offset + head),
 645                 (void *)(long)(event->header.size),
 646                 event->comm.comm, event->comm.pid);
 647
 648         if (thread == NULL ||
 649             thread__set_comm(thread, event->comm.comm)) {
 650                 dump_printf("problem processing perf_event_comm, skipping event.\n");
 651                 return -1;
 652         }
 653         total_comm++;
 654
 655         return 0;
 656 }
 657
 658
 659 struct raw_event_sample {
 660         u32 size;
 661         char data[0];
 662 };
 663
 664 #define FILL_FIELD(ptr, field, event, data)     \
 665         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 666
 667 #define FILL_ARRAY(ptr, array, event, data)                     \
 668 do {                                                            \
 669         void *__array = raw_field_ptr(event, #array, data);     \
 670         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 671 } while(0)
 672
 673 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 674 do {                                                            \
 675         FILL_FIELD(ptr, common_type, event, data);              \
 676         FILL_FIELD(ptr, common_flags, event, data);             \
 677         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 678         FILL_FIELD(ptr, common_pid, event, data);               \
 679         FILL_FIELD(ptr, common_tgid, event, data);              \
 680 } while (0)
 681
 682
 683
 684 struct trace_switch_event {
 685         u32 size;
 686
 687         u16 common_type;
 688         u8 common_flags;
 689         u8 common_preempt_count;
 690         u32 common_pid;
 691         u32 common_tgid;
 692
 693         char prev_comm[16];
 694         u32 prev_pid;
 695         u32 prev_prio;
 696         u64 prev_state;
 697         char next_comm[16];
 698         u32 next_pid;
 699         u32 next_prio;
 700 };
 701
 702 struct trace_runtime_event {
 703         u32 size;
 704
 705         u16 common_type;
 706         u8 common_flags;
 707         u8 common_preempt_count;
 708         u32 common_pid;
 709         u32 common_tgid;
 710
 711         char comm[16];
 712         u32 pid;
 713         u64 runtime;
 714         u64 vruntime;
 715 };
 716
 717 struct trace_wakeup_event {
 718         u32 size;
 719
 720         u16 common_type;
 721         u8 common_flags;
 722         u8 common_preempt_count;
 723         u32 common_pid;
 724         u32 common_tgid;
 725
 726         char comm[16];
 727         u32 pid;
 728
 729         u32 prio;
 730         u32 success;
 731         u32 cpu;
 732 };
 733
 734 struct trace_fork_event {
 735         u32 size;
 736
 737         u16 common_type;
 738         u8 common_flags;
 739         u8 common_preempt_count;
 740         u32 common_pid;
 741         u32 common_tgid;
 742
 743         char parent_comm[16];
 744         u32 parent_pid;
 745         char child_comm[16];
 746         u32 child_pid;
 747 };
 748
 749 struct trace_migrate_task_event {
 750         u32 size;
 751
 752         u16 common_type;
 753         u8 common_flags;
 754         u8 common_preempt_count;
 755         u32 common_pid;
 756         u32 common_tgid;
 757
 758         char comm[16];
 759         u32 pid;
 760
 761         u32 prio;
 762         u32 cpu;
 763 };
 764
 765 struct trace_sched_handler {
 766         void (*switch_event)(struct trace_switch_event *,
 767                              struct event *,
 768                              int cpu,
 769                              u64 timestamp,
 770                              struct thread *thread);
 771
 772         void (*runtime_event)(struct trace_runtime_event *,
 773                               struct event *,
 774                               int cpu,
 775                               u64 timestamp,
 776                               struct thread *thread);
 777
 778         void (*wakeup_event)(struct trace_wakeup_event *,
 779                              struct event *,
 780                              int cpu,
 781                              u64 timestamp,
 782                              struct thread *thread);
 783
 784         void (*fork_event)(struct trace_fork_event *,
 785                            struct event *,
 786                            int cpu,
 787                            u64 timestamp,
 788                            struct thread *thread);
 789
 790         void (*migrate_task_event)(struct trace_migrate_task_event *,
 791                            struct event *,
 792                            int cpu,
 793                            u64 timestamp,
 794                            struct thread *thread);
 795 };
 796
 797
 798 static void
 799 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 800                     struct event *event,
 801                     int cpu __used,
 802                     u64 timestamp __used,
 803                     struct thread *thread __used)
 804 {
 805         struct task_desc *waker, *wakee;
 806
 807         if (verbose) {
 808                 printf("sched_wakeup event %p\n", event);
 809
 810                 printf(" ... pid %d woke up %s/%d\n",
 811                         wakeup_event->common_pid,
 812                         wakeup_event->comm,
 813                         wakeup_event->pid);
 814         }
 815
 816         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 817         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 818
 819         add_sched_event_wakeup(waker, timestamp, wakee);
 820 }
 821
 822 static u64 cpu_last_switched[MAX_CPUS];
 823
 824 static void
 825 replay_switch_event(struct trace_switch_event *switch_event,
 826                     struct event *event,
 827                     int cpu,
 828                     u64 timestamp,
 829                     struct thread *thread __used)
 830 {
 831         struct task_desc *prev, *next;
 832         u64 timestamp0;
 833         s64 delta;
 834
 835         if (verbose)
 836                 printf("sched_switch event %p\n", event);
 837
 838         if (cpu >= MAX_CPUS || cpu < 0)
 839                 return;
 840
 841         timestamp0 = cpu_last_switched[cpu];
 842         if (timestamp0)
 843                 delta = timestamp - timestamp0;
 844         else
 845                 delta = 0;
 846
 847         if (delta < 0)
 848                 die("hm, delta: %Ld < 0 ?\n", delta);
 849
 850         if (verbose) {
 851                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 852                         switch_event->prev_comm, switch_event->prev_pid,
 853                         switch_event->next_comm, switch_event->next_pid,
 854                         delta);
 855         }
 856
 857         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 858         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 859
 860         cpu_last_switched[cpu] = timestamp;
 861
 862         add_sched_event_run(prev, timestamp, delta);
 863         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 864 }
 865
 866
 867 static void
 868 replay_fork_event(struct trace_fork_event *fork_event,
 869                   struct event *event,
 870                   int cpu __used,
 871                   u64 timestamp __used,
 872                   struct thread *thread __used)
 873 {
 874         if (verbose) {
 875                 printf("sched_fork event %p\n", event);
 876                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 877                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 878         }
 879         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 880         register_pid(fork_event->child_pid, fork_event->child_comm);
 881 }
 882
 883 static struct trace_sched_handler replay_ops  = {
 884         .wakeup_event           = replay_wakeup_event,
 885         .switch_event           = replay_switch_event,
 886         .fork_event             = replay_fork_event,
 887 };
 888
 889 struct sort_dimension {
 890         const char              *name;
 891         sort_fn_t               cmp;
 892         struct list_head        list;
 893 };
 894
 895 static LIST_HEAD(cmp_pid);
 896
 897 static int
 898 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 899 {
 900         struct sort_dimension *sort;
 901         int ret = 0;
 902
 903         BUG_ON(list_empty(list));
 904
 905         list_for_each_entry(sort, list, list) {
 906                 ret = sort->cmp(l, r);
 907                 if (ret)
 908                         return ret;
 909         }
 910
 911         return ret;
 912 }
 913
 914 static struct work_atoms *
 915 thread_atoms_search(struct rb_root *root, struct thread *thread,
 916                          struct list_head *sort_list)
 917 {
 918         struct rb_node *node = root->rb_node;
 919         struct work_atoms key = { .thread = thread };
 920
 921         while (node) {
 922                 struct work_atoms *atoms;
 923                 int cmp;
 924
 925                 atoms = container_of(node, struct work_atoms, node);
 926
 927                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 928                 if (cmp > 0)
 929                         node = node->rb_left;
 930                 else if (cmp < 0)
 931                         node = node->rb_right;
 932                 else {
 933                         BUG_ON(thread != atoms->thread);
 934                         return atoms;
 935                 }
 936         }
 937         return NULL;
 938 }
 939
 940 static void
 941 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 942                          struct list_head *sort_list)
 943 {
 944         struct rb_node **new = &(root->rb_node), *parent = NULL;
 945
 946         while (*new) {
 947                 struct work_atoms *this;
 948                 int cmp;
 949
 950                 this = container_of(*new, struct work_atoms, node);
 951                 parent = *new;
 952
 953                 cmp = thread_lat_cmp(sort_list, data, this);
 954
 955                 if (cmp > 0)
 956                         new = &((*new)->rb_left);
 957                 else
 958                         new = &((*new)->rb_right);
 959         }
 960
 961         rb_link_node(&data->node, parent, new);
 962         rb_insert_color(&data->node, root);
 963 }
 964
 965 static void thread_atoms_insert(struct thread *thread)
 966 {
 967         struct work_atoms *atoms;
 968
 969         atoms = calloc(sizeof(*atoms), 1);
 970         if (!atoms)
 971                 die("No memory");
 972
 973         atoms->thread = thread;
 974         INIT_LIST_HEAD(&atoms->work_list);
 975         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 976 }
 977
 978 static void
 979 latency_fork_event(struct trace_fork_event *fork_event __used,
 980                    struct event *event __used,
 981                    int cpu __used,
 982                    u64 timestamp __used,
 983                    struct thread *thread __used)
 984 {
 985         /* should insert the newcomer */
 986 }
 987
 988 __used
 989 static char sched_out_state(struct trace_switch_event *switch_event)
 990 {
 991         const char *str = TASK_STATE_TO_CHAR_STR;
 992
 993         return str[switch_event->prev_state];
 994 }
 995
 996 static void
 997 add_sched_out_event(struct work_atoms *atoms,
 998                     char run_state,
 999                     u64 timestamp)
1000 {
1001         struct work_atom *atom;
1002
1003         atom = calloc(sizeof(*atom), 1);
1004         if (!atom)
1005                 die("Non memory");
1006
1007         atom->sched_out_time = timestamp;
1008
1009         if (run_state == 'R') {
1010                 atom->state = THREAD_WAIT_CPU;
1011                 atom->wake_up_time = atom->sched_out_time;
1012         }
1013
1014         list_add_tail(&atom->list, &atoms->work_list);
1015 }
1016
1017 static void
1018 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
1019 {
1020         struct work_atom *atom;
1021
1022         BUG_ON(list_empty(&atoms->work_list));
1023
1024         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1025
1026         atom->runtime += delta;
1027         atoms->total_runtime += delta;
1028 }
1029
1030 static void
1031 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1032 {
1033         struct work_atom *atom;
1034         u64 delta;
1035
1036         if (list_empty(&atoms->work_list))
1037                 return;
1038
1039         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1040
1041         if (atom->state != THREAD_WAIT_CPU)
1042                 return;
1043
1044         if (timestamp < atom->wake_up_time) {
1045                 atom->state = THREAD_IGNORE;
1046                 return;
1047         }
1048
1049         atom->state = THREAD_SCHED_IN;
1050         atom->sched_in_time = timestamp;
1051
1052         delta = atom->sched_in_time - atom->wake_up_time;
1053         atoms->total_lat += delta;
1054         if (delta > atoms->max_lat)
1055                 atoms->max_lat = delta;
1056         atoms->nb_atoms++;
1057 }
1058
1059 static void
1060 latency_switch_event(struct trace_switch_event *switch_event,
1061                      struct event *event __used,
1062                      int cpu,
1063                      u64 timestamp,
1064                      struct thread *thread __used)
1065 {
1066         struct work_atoms *out_events, *in_events;
1067         struct thread *sched_out, *sched_in;
1068         u64 timestamp0;
1069         s64 delta;
1070
1071         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1072
1073         timestamp0 = cpu_last_switched[cpu];
1074         cpu_last_switched[cpu] = timestamp;
1075         if (timestamp0)
1076                 delta = timestamp - timestamp0;
1077         else
1078                 delta = 0;
1079
1080         if (delta < 0)
1081                 die("hm, delta: %Ld < 0 ?\n", delta);
1082
1083
1084         sched_out = threads__findnew(switch_event->prev_pid);
1085         sched_in = threads__findnew(switch_event->next_pid);
1086
1087         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1088         if (!out_events) {
1089                 thread_atoms_insert(sched_out);
1090                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1091                 if (!out_events)
1092                         die("out-event: Internal tree error");
1093         }
1094         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1095
1096         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1097         if (!in_events) {
1098                 thread_atoms_insert(sched_in);
1099                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1100                 if (!in_events)
1101                         die("in-event: Internal tree error");
1102                 /*
1103                  * Take came in we have not heard about yet,
1104                  * add in an initial atom in runnable state:
1105                  */
1106                 add_sched_out_event(in_events, 'R', timestamp);
1107         }
1108         add_sched_in_event(in_events, timestamp);
1109 }
1110
1111 static void
1112 latency_runtime_event(struct trace_runtime_event *runtime_event,
1113                      struct event *event __used,
1114                      int cpu,
1115                      u64 timestamp,
1116                      struct thread *this_thread __used)
1117 {
1118         struct thread *thread = threads__findnew(runtime_event->pid);
1119         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1120
1121         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1122         if (!atoms) {
1123                 thread_atoms_insert(thread);
1124                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1125                 if (!atoms)
1126                         die("in-event: Internal tree error");
1127                 add_sched_out_event(atoms, 'R', timestamp);
1128         }
1129
1130         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1131 }
1132
1133 static void
1134 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1135                      struct event *__event __used,
1136                      int cpu __used,
1137                      u64 timestamp,
1138                      struct thread *thread __used)
1139 {
1140         struct work_atoms *atoms;
1141         struct work_atom *atom;
1142         struct thread *wakee;
1143
1144         /* Note for later, it may be interesting to observe the failing cases */
1145         if (!wakeup_event->success)
1146                 return;
1147
1148         wakee = threads__findnew(wakeup_event->pid);
1149         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1150         if (!atoms) {
1151                 thread_atoms_insert(wakee);
1152                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1153                 if (!atoms)
1154                         die("wakeup-event: Internal tree error");
1155                 add_sched_out_event(atoms, 'S', timestamp);
1156         }
1157
1158         BUG_ON(list_empty(&atoms->work_list));
1159
1160         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1161
1162         /*
1163          * You WILL be missing events if you've recorded only
1164          * one CPU, or are only looking at only one, so don't
1165          * make useless noise.
1166          */
1167         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1168                 nr_state_machine_bugs++;
1169
1170         nr_timestamps++;
1171         if (atom->sched_out_time > timestamp) {
1172                 nr_unordered_timestamps++;
1173                 return;
1174         }
1175
1176         atom->state = THREAD_WAIT_CPU;
1177         atom->wake_up_time = timestamp;
1178 }
1179
1180 static void
1181 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1182                      struct event *__event __used,
1183                      int cpu __used,
1184                      u64 timestamp,
1185                      struct thread *thread __used)
1186 {
1187         struct work_atoms *atoms;
1188         struct work_atom *atom;
1189         struct thread *migrant;
1190
1191         /*
1192          * Only need to worry about migration when profiling one CPU.
1193          */
1194         if (profile_cpu == -1)
1195                 return;
1196
1197         migrant = threads__findnew(migrate_task_event->pid);
1198         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1199         if (!atoms) {
1200                 thread_atoms_insert(migrant);
1201                 register_pid(migrant->pid, migrant->comm);
1202                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1203                 if (!atoms)
1204                         die("migration-event: Internal tree error");
1205                 add_sched_out_event(atoms, 'R', timestamp);
1206         }
1207
1208         BUG_ON(list_empty(&atoms->work_list));
1209
1210         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1211         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1212
1213         nr_timestamps++;
1214
1215         if (atom->sched_out_time > timestamp)
1216                 nr_unordered_timestamps++;
1217 }
1218
1219 static struct trace_sched_handler lat_ops  = {
1220         .wakeup_event           = latency_wakeup_event,
1221         .switch_event           = latency_switch_event,
1222         .runtime_event          = latency_runtime_event,
1223         .fork_event             = latency_fork_event,
1224         .migrate_task_event     = latency_migrate_task_event,
1225 };
1226
1227 static void output_lat_thread(struct work_atoms *work_list)
1228 {
1229         int i;
1230         int ret;
1231         u64 avg;
1232
1233         if (!work_list->nb_atoms)
1234                 return;
1235         /*
1236          * Ignore idle threads:
1237          */
1238         if (!strcmp(work_list->thread->comm, "swapper"))
1239                 return;
1240
1241         all_runtime += work_list->total_runtime;
1242         all_count += work_list->nb_atoms;
1243
1244         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1245
1246         for (i = 0; i < 24 - ret; i++)
1247                 printf(" ");
1248
1249         avg = work_list->total_lat / work_list->nb_atoms;
1250
1251         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1252               (double)work_list->total_runtime / 1e6,
1253                  work_list->nb_atoms, (double)avg / 1e6,
1254                  (double)work_list->max_lat / 1e6);
1255 }
1256
1257 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1258 {
1259         if (l->thread->pid < r->thread->pid)
1260                 return -1;
1261         if (l->thread->pid > r->thread->pid)
1262                 return 1;
1263
1264         return 0;
1265 }
1266
1267 static struct sort_dimension pid_sort_dimension = {
1268         .name                   = "pid",
1269         .cmp                    = pid_cmp,
1270 };
1271
1272 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1273 {
1274         u64 avgl, avgr;
1275
1276         if (!l->nb_atoms)
1277                 return -1;
1278
1279         if (!r->nb_atoms)
1280                 return 1;
1281
1282         avgl = l->total_lat / l->nb_atoms;
1283         avgr = r->total_lat / r->nb_atoms;
1284
1285         if (avgl < avgr)
1286                 return -1;
1287         if (avgl > avgr)
1288                 return 1;
1289
1290         return 0;
1291 }
1292
1293 static struct sort_dimension avg_sort_dimension = {
1294         .name                   = "avg",
1295         .cmp                    = avg_cmp,
1296 };
1297
1298 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1299 {
1300         if (l->max_lat < r->max_lat)
1301                 return -1;
1302         if (l->max_lat > r->max_lat)
1303                 return 1;
1304
1305         return 0;
1306 }
1307
1308 static struct sort_dimension max_sort_dimension = {
1309         .name                   = "max",
1310         .cmp                    = max_cmp,
1311 };
1312
1313 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1314 {
1315         if (l->nb_atoms < r->nb_atoms)
1316                 return -1;
1317         if (l->nb_atoms > r->nb_atoms)
1318                 return 1;
1319
1320         return 0;
1321 }
1322
1323 static struct sort_dimension switch_sort_dimension = {
1324         .name                   = "switch",
1325         .cmp                    = switch_cmp,
1326 };
1327
1328 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1329 {
1330         if (l->total_runtime < r->total_runtime)
1331                 return -1;
1332         if (l->total_runtime > r->total_runtime)
1333                 return 1;
1334
1335         return 0;
1336 }
1337
1338 static struct sort_dimension runtime_sort_dimension = {
1339         .name                   = "runtime",
1340         .cmp                    = runtime_cmp,
1341 };
1342
1343 static struct sort_dimension *available_sorts[] = {
1344         &pid_sort_dimension,
1345         &avg_sort_dimension,
1346         &max_sort_dimension,
1347         &switch_sort_dimension,
1348         &runtime_sort_dimension,
1349 };
1350
1351 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1352
1353 static LIST_HEAD(sort_list);
1354
1355 static int sort_dimension__add(const char *tok, struct list_head *list)
1356 {
1357         int i;
1358
1359         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1360                 if (!strcmp(available_sorts[i]->name, tok)) {
1361                         list_add_tail(&available_sorts[i]->list, list);
1362
1363                         return 0;
1364                 }
1365         }
1366
1367         return -1;
1368 }
1369
1370 static void setup_sorting(void);
1371
1372 static void sort_lat(void)
1373 {
1374         struct rb_node *node;
1375
1376         for (;;) {
1377                 struct work_atoms *data;
1378                 node = rb_first(&atom_root);
1379                 if (!node)
1380                         break;
1381
1382                 rb_erase(node, &atom_root);
1383                 data = rb_entry(node, struct work_atoms, node);
1384                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1385         }
1386 }
1387
1388 static struct trace_sched_handler *trace_handler;
1389
1390 static void
1391 process_sched_wakeup_event(struct raw_event_sample *raw,
1392                            struct event *event,
1393                            int cpu __used,
1394                            u64 timestamp __used,
1395                            struct thread *thread __used)
1396 {
1397         struct trace_wakeup_event wakeup_event;
1398
1399         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1400
1401         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1402         FILL_FIELD(wakeup_event, pid, event, raw->data);
1403         FILL_FIELD(wakeup_event, prio, event, raw->data);
1404         FILL_FIELD(wakeup_event, success, event, raw->data);
1405         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1406
1407         if (trace_handler->wakeup_event)
1408                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1409 }
1410
1411 /*
1412  * Track the current task - that way we can know whether there's any
1413  * weird events, such as a task being switched away that is not current.
1414  */
1415 static int max_cpu;
1416
1417 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1418
1419 static struct thread *curr_thread[MAX_CPUS];
1420
1421 static char next_shortname1 = 'A';
1422 static char next_shortname2 = '0';
1423
1424 static void
1425 map_switch_event(struct trace_switch_event *switch_event,
1426                  struct event *event __used,
1427                  int this_cpu,
1428                  u64 timestamp,
1429                  struct thread *thread __used)
1430 {
1431         struct thread *sched_out, *sched_in;
1432         int new_shortname;
1433         u64 timestamp0;
1434         s64 delta;
1435         int cpu;
1436
1437         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1438
1439         if (this_cpu > max_cpu)
1440                 max_cpu = this_cpu;
1441
1442         timestamp0 = cpu_last_switched[this_cpu];
1443         cpu_last_switched[this_cpu] = timestamp;
1444         if (timestamp0)
1445                 delta = timestamp - timestamp0;
1446         else
1447                 delta = 0;
1448
1449         if (delta < 0)
1450                 die("hm, delta: %Ld < 0 ?\n", delta);
1451
1452
1453         sched_out = threads__findnew(switch_event->prev_pid);
1454         sched_in = threads__findnew(switch_event->next_pid);
1455
1456         curr_thread[this_cpu] = sched_in;
1457
1458         printf("  ");
1459
1460         new_shortname = 0;
1461         if (!sched_in->shortname[0]) {
1462                 sched_in->shortname[0] = next_shortname1;
1463                 sched_in->shortname[1] = next_shortname2;
1464
1465                 if (next_shortname1 < 'Z') {
1466                         next_shortname1++;
1467                 } else {
1468                         next_shortname1='A';
1469                         if (next_shortname2 < '9') {
1470                                 next_shortname2++;
1471                         } else {
1472                                 next_shortname2='0';
1473                         }
1474                 }
1475                 new_shortname = 1;
1476         }
1477
1478         for (cpu = 0; cpu <= max_cpu; cpu++) {
1479                 if (cpu != this_cpu)
1480                         printf(" ");
1481                 else
1482                         printf("*");
1483
1484                 if (curr_thread[cpu]) {
1485                         if (curr_thread[cpu]->pid)
1486                                 printf("%2s ", curr_thread[cpu]->shortname);
1487                         else
1488                                 printf(".  ");
1489                 } else
1490                         printf("   ");
1491         }
1492
1493         printf("  %12.6f secs ", (double)timestamp/1e9);
1494         if (new_shortname) {
1495                 printf("%s => %s:%d\n",
1496                         sched_in->shortname, sched_in->comm, sched_in->pid);
1497         } else {
1498                 printf("\n");
1499         }
1500 }
1501
1502
1503 static void
1504 process_sched_switch_event(struct raw_event_sample *raw,
1505                            struct event *event,
1506                            int this_cpu,
1507                            u64 timestamp __used,
1508                            struct thread *thread __used)
1509 {
1510         struct trace_switch_event switch_event;
1511
1512         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1513
1514         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1515         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1516         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1517         FILL_FIELD(switch_event, prev_state, event, raw->data);
1518         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1519         FILL_FIELD(switch_event, next_pid, event, raw->data);
1520         FILL_FIELD(switch_event, next_prio, event, raw->data);
1521
1522         if (curr_pid[this_cpu] != (u32)-1) {
1523                 /*
1524                  * Are we trying to switch away a PID that is
1525                  * not current?
1526                  */
1527                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1528                         nr_context_switch_bugs++;
1529         }
1530         if (trace_handler->switch_event)
1531                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1532
1533         curr_pid[this_cpu] = switch_event.next_pid;
1534 }
1535
1536 static void
1537 process_sched_runtime_event(struct raw_event_sample *raw,
1538                            struct event *event,
1539                            int cpu __used,
1540                            u64 timestamp __used,
1541                            struct thread *thread __used)
1542 {
1543         struct trace_runtime_event runtime_event;
1544
1545         FILL_ARRAY(runtime_event, comm, event, raw->data);
1546         FILL_FIELD(runtime_event, pid, event, raw->data);
1547         FILL_FIELD(runtime_event, runtime, event, raw->data);
1548         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1549
1550         if (trace_handler->runtime_event)
1551                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1552 }
1553
1554 static void
1555 process_sched_fork_event(struct raw_event_sample *raw,
1556                          struct event *event,
1557                          int cpu __used,
1558                          u64 timestamp __used,
1559                          struct thread *thread __used)
1560 {
1561         struct trace_fork_event fork_event;
1562
1563         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1564
1565         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1566         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1567         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1568         FILL_FIELD(fork_event, child_pid, event, raw->data);
1569
1570         if (trace_handler->fork_event)
1571                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1572 }
1573
1574 static void
1575 process_sched_exit_event(struct event *event,
1576                          int cpu __used,
1577                          u64 timestamp __used,
1578                          struct thread *thread __used)
1579 {
1580         if (verbose)
1581                 printf("sched_exit event %p\n", event);
1582 }
1583
1584 static void
1585 process_sched_migrate_task_event(struct raw_event_sample *raw,
1586                            struct event *event,
1587                            int cpu __used,
1588                            u64 timestamp __used,
1589                            struct thread *thread __used)
1590 {
1591         struct trace_migrate_task_event migrate_task_event;
1592
1593         FILL_COMMON_FIELDS(migrate_task_event, event, raw->data);
1594
1595         FILL_ARRAY(migrate_task_event, comm, event, raw->data);
1596         FILL_FIELD(migrate_task_event, pid, event, raw->data);
1597         FILL_FIELD(migrate_task_event, prio, event, raw->data);
1598         FILL_FIELD(migrate_task_event, cpu, event, raw->data);
1599
1600         if (trace_handler->migrate_task_event)
1601                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1602 }
1603
1604 static void
1605 process_raw_event(event_t *raw_event __used, void *more_data,
1606                   int cpu, u64 timestamp, struct thread *thread)
1607 {
1608         struct raw_event_sample *raw = more_data;
1609         struct event *event;
1610         int type;
1611
1612         type = trace_parse_common_type(raw->data);
1613         event = trace_find_event(type);
1614
1615         if (!strcmp(event->name, "sched_switch"))
1616                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1617         if (!strcmp(event->name, "sched_stat_runtime"))
1618                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1619         if (!strcmp(event->name, "sched_wakeup"))
1620                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1621         if (!strcmp(event->name, "sched_wakeup_new"))
1622                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1623         if (!strcmp(event->name, "sched_process_fork"))
1624                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1625         if (!strcmp(event->name, "sched_process_exit"))
1626                 process_sched_exit_event(event, cpu, timestamp, thread);
1627         if (!strcmp(event->name, "sched_migrate_task"))
1628                 process_sched_migrate_task_event(raw, event, cpu, timestamp, thread);
1629 }
1630
1631 static int
1632 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1633 {
1634         struct thread *thread;
1635         u64 ip = event->ip.ip;
1636         u64 timestamp = -1;
1637         u32 cpu = -1;
1638         u64 period = 1;
1639         void *more_data = event->ip.__more_data;
1640
1641         if (!(sample_type & PERF_SAMPLE_RAW))
1642                 return 0;
1643
1644         thread = threads__findnew(event->ip.pid);
1645
1646         if (sample_type & PERF_SAMPLE_TIME) {
1647                 timestamp = *(u64 *)more_data;
1648                 more_data += sizeof(u64);
1649         }
1650
1651         if (sample_type & PERF_SAMPLE_CPU) {
1652                 cpu = *(u32 *)more_data;
1653                 more_data += sizeof(u32);
1654                 more_data += sizeof(u32); /* reserved */
1655         }
1656
1657         if (sample_type & PERF_SAMPLE_PERIOD) {
1658                 period = *(u64 *)more_data;
1659                 more_data += sizeof(u64);
1660         }
1661
1662         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1663                 (void *)(offset + head),
1664                 (void *)(long)(event->header.size),
1665                 event->header.misc,
1666                 event->ip.pid, event->ip.tid,
1667                 (void *)(long)ip,
1668                 (long long)period);
1669
1670         if (thread == NULL) {
1671                 eprintf("problem processing %d event, skipping it.\n",
1672                         event->header.type);
1673                 return -1;
1674         }
1675
1676         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1677
1678         if (profile_cpu != -1 && profile_cpu != (int) cpu)
1679                 return 0;
1680
1681         process_raw_event(event, more_data, cpu, timestamp, thread);
1682
1683         return 0;
1684 }
1685
1686 static int
1687 process_lost_event(event_t *event __used,
1688                    unsigned long offset __used,
1689                    unsigned long head __used)
1690 {
1691         nr_lost_chunks++;
1692         nr_lost_events += event->lost.lost;
1693
1694         return 0;
1695 }
1696
1697 static int sample_type_check(u64 type)
1698 {
1699         sample_type = type;
1700
1701         if (!(sample_type & PERF_SAMPLE_RAW)) {
1702                 fprintf(stderr,
1703                         "No trace sample to read. Did you call perf record "
1704                         "without -R?");
1705                 return -1;
1706         }
1707
1708         return 0;
1709 }
1710
1711 static struct perf_file_handler file_handler = {
1712         .process_sample_event   = process_sample_event,
1713         .process_comm_event     = process_comm_event,
1714         .process_lost_event     = process_lost_event,
1715         .sample_type_check      = sample_type_check,
1716 };
1717
1718 static int read_events(void)
1719 {
1720         register_idle_thread();
1721         register_perf_file_handler(&file_handler);
1722
1723         return mmap_dispatch_perf_file(&header, input_name, 0, 0, &cwdlen, &cwd);
1724 }
1725
1726 static void print_bad_events(void)
1727 {
1728         if (nr_unordered_timestamps && nr_timestamps) {
1729                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1730                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1731                         nr_unordered_timestamps, nr_timestamps);
1732         }
1733         if (nr_lost_events && nr_events) {
1734                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1735                         (double)nr_lost_events/(double)nr_events*100.0,
1736                         nr_lost_events, nr_events, nr_lost_chunks);
1737         }
1738         if (nr_state_machine_bugs && nr_timestamps) {
1739                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1740                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1741                         nr_state_machine_bugs, nr_timestamps);
1742                 if (nr_lost_events)
1743                         printf(" (due to lost events?)");
1744                 printf("\n");
1745         }
1746         if (nr_context_switch_bugs && nr_timestamps) {
1747                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1748                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1749                         nr_context_switch_bugs, nr_timestamps);
1750                 if (nr_lost_events)
1751                         printf(" (due to lost events?)");
1752                 printf("\n");
1753         }
1754 }
1755
1756 static void __cmd_lat(void)
1757 {
1758         struct rb_node *next;
1759
1760         setup_pager();
1761         read_events();
1762         sort_lat();
1763
1764         printf("\n -----------------------------------------------------------------------------------------\n");
1765         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1766         printf(" -----------------------------------------------------------------------------------------\n");
1767
1768         next = rb_first(&sorted_atom_root);
1769
1770         while (next) {
1771                 struct work_atoms *work_list;
1772
1773                 work_list = rb_entry(next, struct work_atoms, node);
1774                 output_lat_thread(work_list);
1775                 next = rb_next(next);
1776         }
1777
1778         printf(" -----------------------------------------------------------------------------------------\n");
1779         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1780                 (double)all_runtime/1e6, all_count);
1781
1782         printf(" ---------------------------------------------------\n");
1783
1784         print_bad_events();
1785         printf("\n");
1786
1787 }
1788
1789 static struct trace_sched_handler map_ops  = {
1790         .wakeup_event           = NULL,
1791         .switch_event           = map_switch_event,
1792         .runtime_event          = NULL,
1793         .fork_event             = NULL,
1794 };
1795
1796 static void __cmd_map(void)
1797 {
1798         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1799
1800         setup_pager();
1801         read_events();
1802         print_bad_events();
1803 }
1804
1805 static void __cmd_replay(void)
1806 {
1807         unsigned long i;
1808
1809         calibrate_run_measurement_overhead();
1810         calibrate_sleep_measurement_overhead();
1811
1812         test_calibrations();
1813
1814         read_events();
1815
1816         printf("nr_run_events:        %ld\n", nr_run_events);
1817         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1818         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1819
1820         if (targetless_wakeups)
1821                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1822         if (multitarget_wakeups)
1823                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1824         if (nr_run_events_optimized)
1825                 printf("run atoms optimized: %ld\n",
1826                         nr_run_events_optimized);
1827
1828         print_task_traces();
1829         add_cross_task_wakeups();
1830
1831         create_tasks();
1832         printf("------------------------------------------------------------\n");
1833         for (i = 0; i < replay_repeat; i++)
1834                 run_one_test();
1835 }
1836
1837
1838 static const char * const sched_usage[] = {
1839         "perf sched [<options>] {record|latency|map|replay|trace}",
1840         NULL
1841 };
1842
1843 static const struct option sched_options[] = {
1844         OPT_STRING('i', "input", &input_name, "file",
1845                     "input file name"),
1846         OPT_BOOLEAN('v', "verbose", &verbose,
1847                     "be more verbose (show symbol address, etc)"),
1848         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1849                     "dump raw trace in ASCII"),
1850         OPT_END()
1851 };
1852
1853 static const char * const latency_usage[] = {
1854         "perf sched latency [<options>]",
1855         NULL
1856 };
1857
1858 static const struct option latency_options[] = {
1859         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1860                    "sort by key(s): runtime, switch, avg, max"),
1861         OPT_BOOLEAN('v', "verbose", &verbose,
1862                     "be more verbose (show symbol address, etc)"),
1863         OPT_INTEGER('C', "CPU", &profile_cpu,
1864                     "CPU to profile on"),
1865         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1866                     "dump raw trace in ASCII"),
1867         OPT_END()
1868 };
1869
1870 static const char * const replay_usage[] = {
1871         "perf sched replay [<options>]",
1872         NULL
1873 };
1874
1875 static const struct option replay_options[] = {
1876         OPT_INTEGER('r', "repeat", &replay_repeat,
1877                     "repeat the workload replay N times (-1: infinite)"),
1878         OPT_BOOLEAN('v', "verbose", &verbose,
1879                     "be more verbose (show symbol address, etc)"),
1880         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1881                     "dump raw trace in ASCII"),
1882         OPT_END()
1883 };
1884
1885 static void setup_sorting(void)
1886 {
1887         char *tmp, *tok, *str = strdup(sort_order);
1888
1889         for (tok = strtok_r(str, ", ", &tmp);
1890                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1891                 if (sort_dimension__add(tok, &sort_list) < 0) {
1892                         error("Unknown --sort key: `%s'", tok);
1893                         usage_with_options(latency_usage, latency_options);
1894                 }
1895         }
1896
1897         free(str);
1898
1899         sort_dimension__add("pid", &cmp_pid);
1900 }
1901
1902 static const char *record_args[] = {
1903         "record",
1904         "-a",
1905         "-R",
1906         "-M",
1907         "-f",
1908         "-m", "1024",
1909         "-c", "1",
1910         "-e", "sched:sched_switch:r",
1911         "-e", "sched:sched_stat_wait:r",
1912         "-e", "sched:sched_stat_sleep:r",
1913         "-e", "sched:sched_stat_iowait:r",
1914         "-e", "sched:sched_stat_runtime:r",
1915         "-e", "sched:sched_process_exit:r",
1916         "-e", "sched:sched_process_fork:r",
1917         "-e", "sched:sched_wakeup:r",
1918         "-e", "sched:sched_migrate_task:r",
1919 };
1920
1921 static int __cmd_record(int argc, const char **argv)
1922 {
1923         unsigned int rec_argc, i, j;
1924         const char **rec_argv;
1925
1926         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1927         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1928
1929         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1930                 rec_argv[i] = strdup(record_args[i]);
1931
1932         for (j = 1; j < (unsigned int)argc; j++, i++)
1933                 rec_argv[i] = argv[j];
1934
1935         BUG_ON(i != rec_argc);
1936
1937         return cmd_record(i, rec_argv, NULL);
1938 }
1939
1940 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1941 {
1942         symbol__init();
1943
1944         argc = parse_options(argc, argv, sched_options, sched_usage,
1945                              PARSE_OPT_STOP_AT_NON_OPTION);
1946         if (!argc)
1947                 usage_with_options(sched_usage, sched_options);
1948
1949         if (!strncmp(argv[0], "rec", 3)) {
1950                 return __cmd_record(argc, argv);
1951         } else if (!strncmp(argv[0], "lat", 3)) {
1952                 trace_handler = &lat_ops;
1953                 if (argc > 1) {
1954                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1955                         if (argc)
1956                                 usage_with_options(latency_usage, latency_options);
1957                 }
1958                 setup_sorting();
1959                 __cmd_lat();
1960         } else if (!strcmp(argv[0], "map")) {
1961                 trace_handler = &map_ops;
1962                 setup_sorting();
1963                 __cmd_map();
1964         } else if (!strncmp(argv[0], "rep", 3)) {
1965                 trace_handler = &replay_ops;
1966                 if (argc) {
1967                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1968                         if (argc)
1969                                 usage_with_options(replay_usage, replay_options);
1970                 }
1971                 __cmd_replay();
1972         } else if (!strcmp(argv[0], "trace")) {
1973                 /*
1974                  * Aliased to 'perf trace' for now:
1975                  */
1976                 return cmd_trace(argc, argv, prefix);
1977         } else {
1978                 usage_with_options(sched_usage, sched_options);
1979         }
1980
1981         return 0;
1982 }