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