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