SAFE public projects git trees. - safe/jmp/linux-2.6/blob - tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/cache.h"
   6 #include "util/symbol.h"
   7 #include "util/thread.h"
   8 #include "util/header.h"
   9
  10 #include "util/parse-options.h"
  11 #include "util/trace-event.h"
  12
  13 #include "util/debug.h"
  14 #include "util/data_map.h"
  15
  16 #include <sys/types.h>
  17 #include <sys/prctl.h>
  18
  19 #include <semaphore.h>
  20 #include <pthread.h>
  21 #include <math.h>
  22
  23 static char                     const *input_name = "perf.data";
  24
  25 static struct perf_header       *header;
  26 static u64                      sample_type;
  27
  28 static char                     default_sort_order[] = "avg, max, switch, runtime";
  29 static char                     *sort_order = default_sort_order;
  30
  31 static int                      profile_cpu = -1;
  32
  33 #define PR_SET_NAME             15               /* Set process name */
  34 #define MAX_CPUS                4096
  35
  36 static u64                      run_measurement_overhead;
  37 static u64                      sleep_measurement_overhead;
  38
  39 #define COMM_LEN                20
  40 #define SYM_LEN                 129
  41
  42 #define MAX_PID                 65536
  43
  44 static unsigned long            nr_tasks;
  45
  46 struct sched_atom;
  47
  48 struct task_desc {
  49         unsigned long           nr;
  50         unsigned long           pid;
  51         char                    comm[COMM_LEN];
  52
  53         unsigned long           nr_events;
  54         unsigned long           curr_event;
  55         struct sched_atom       **atoms;
  56
  57         pthread_t               thread;
  58         sem_t                   sleep_sem;
  59
  60         sem_t                   ready_for_work;
  61         sem_t                   work_done_sem;
  62
  63         u64                     cpu_usage;
  64 };
  65
  66 enum sched_event_type {
  67         SCHED_EVENT_RUN,
  68         SCHED_EVENT_SLEEP,
  69         SCHED_EVENT_WAKEUP,
  70         SCHED_EVENT_MIGRATION,
  71 };
  72
  73 struct sched_atom {
  74         enum sched_event_type   type;
  75         u64                     timestamp;
  76         u64                     duration;
  77         unsigned long           nr;
  78         int                     specific_wait;
  79         sem_t                   *wait_sem;
  80         struct task_desc        *wakee;
  81 };
  82
  83 static struct task_desc         *pid_to_task[MAX_PID];
  84
  85 static struct task_desc         **tasks;
  86
  87 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  88 static u64                      start_time;
  89
  90 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  91
  92 static unsigned long            nr_run_events;
  93 static unsigned long            nr_sleep_events;
  94 static unsigned long            nr_wakeup_events;
  95
  96 static unsigned long            nr_sleep_corrections;
  97 static unsigned long            nr_run_events_optimized;
  98
  99 static unsigned long            targetless_wakeups;
 100 static unsigned long            multitarget_wakeups;
 101
 102 static u64                      cpu_usage;
 103 static u64                      runavg_cpu_usage;
 104 static u64                      parent_cpu_usage;
 105 static u64                      runavg_parent_cpu_usage;
 106
 107 static unsigned long            nr_runs;
 108 static u64                      sum_runtime;
 109 static u64                      sum_fluct;
 110 static u64                      run_avg;
 111
 112 static unsigned long            replay_repeat = 10;
 113 static unsigned long            nr_timestamps;
 114 static unsigned long            nr_unordered_timestamps;
 115 static unsigned long            nr_state_machine_bugs;
 116 static unsigned long            nr_context_switch_bugs;
 117 static unsigned long            nr_events;
 118 static unsigned long            nr_lost_chunks;
 119 static unsigned long            nr_lost_events;
 120
 121 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 122
 123 enum thread_state {
 124         THREAD_SLEEPING = 0,
 125         THREAD_WAIT_CPU,
 126         THREAD_SCHED_IN,
 127         THREAD_IGNORE
 128 };
 129
 130 struct work_atom {
 131         struct list_head        list;
 132         enum thread_state       state;
 133         u64                     sched_out_time;
 134         u64                     wake_up_time;
 135         u64                     sched_in_time;
 136         u64                     runtime;
 137 };
 138
 139 struct work_atoms {
 140         struct list_head        work_list;
 141         struct thread           *thread;
 142         struct rb_node          node;
 143         u64                     max_lat;
 144         u64                     total_lat;
 145         u64                     nb_atoms;
 146         u64                     total_runtime;
 147 };
 148
 149 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 150
 151 static struct rb_root           atom_root, sorted_atom_root;
 152
 153 static u64                      all_runtime;
 154 static u64                      all_count;
 155
 156
 157 static u64 get_nsecs(void)
 158 {
 159         struct timespec ts;
 160
 161         clock_gettime(CLOCK_MONOTONIC, &ts);
 162
 163         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 164 }
 165
 166 static void burn_nsecs(u64 nsecs)
 167 {
 168         u64 T0 = get_nsecs(), T1;
 169
 170         do {
 171                 T1 = get_nsecs();
 172         } while (T1 + run_measurement_overhead < T0 + nsecs);
 173 }
 174
 175 static void sleep_nsecs(u64 nsecs)
 176 {
 177         struct timespec ts;
 178
 179         ts.tv_nsec = nsecs % 999999999;
 180         ts.tv_sec = nsecs / 999999999;
 181
 182         nanosleep(&ts, NULL);
 183 }
 184
 185 static void calibrate_run_measurement_overhead(void)
 186 {
 187         u64 T0, T1, delta, min_delta = 1000000000ULL;
 188         int i;
 189
 190         for (i = 0; i < 10; i++) {
 191                 T0 = get_nsecs();
 192                 burn_nsecs(0);
 193                 T1 = get_nsecs();
 194                 delta = T1-T0;
 195                 min_delta = min(min_delta, delta);
 196         }
 197         run_measurement_overhead = min_delta;
 198
 199         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 200 }
 201
 202 static void calibrate_sleep_measurement_overhead(void)
 203 {
 204         u64 T0, T1, delta, min_delta = 1000000000ULL;
 205         int i;
 206
 207         for (i = 0; i < 10; i++) {
 208                 T0 = get_nsecs();
 209                 sleep_nsecs(10000);
 210                 T1 = get_nsecs();
 211                 delta = T1-T0;
 212                 min_delta = min(min_delta, delta);
 213         }
 214         min_delta -= 10000;
 215         sleep_measurement_overhead = min_delta;
 216
 217         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 218 }
 219
 220 static struct sched_atom *
 221 get_new_event(struct task_desc *task, u64 timestamp)
 222 {
 223         struct sched_atom *event = zalloc(sizeof(*event));
 224         unsigned long idx = task->nr_events;
 225         size_t size;
 226
 227         event->timestamp = timestamp;
 228         event->nr = idx;
 229
 230         task->nr_events++;
 231         size = sizeof(struct sched_atom *) * task->nr_events;
 232         task->atoms = realloc(task->atoms, size);
 233         BUG_ON(!task->atoms);
 234
 235         task->atoms[idx] = event;
 236
 237         return event;
 238 }
 239
 240 static struct sched_atom *last_event(struct task_desc *task)
 241 {
 242         if (!task->nr_events)
 243                 return NULL;
 244
 245         return task->atoms[task->nr_events - 1];
 246 }
 247
 248 static void
 249 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 250 {
 251         struct sched_atom *event, *curr_event = last_event(task);
 252
 253         /*
 254          * optimize an existing RUN event by merging this one
 255          * to it:
 256          */
 257         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 258                 nr_run_events_optimized++;
 259                 curr_event->duration += duration;
 260                 return;
 261         }
 262
 263         event = get_new_event(task, timestamp);
 264
 265         event->type = SCHED_EVENT_RUN;
 266         event->duration = duration;
 267
 268         nr_run_events++;
 269 }
 270
 271 static void
 272 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 273                        struct task_desc *wakee)
 274 {
 275         struct sched_atom *event, *wakee_event;
 276
 277         event = get_new_event(task, timestamp);
 278         event->type = SCHED_EVENT_WAKEUP;
 279         event->wakee = wakee;
 280
 281         wakee_event = last_event(wakee);
 282         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 283                 targetless_wakeups++;
 284                 return;
 285         }
 286         if (wakee_event->wait_sem) {
 287                 multitarget_wakeups++;
 288                 return;
 289         }
 290
 291         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 292         sem_init(wakee_event->wait_sem, 0, 0);
 293         wakee_event->specific_wait = 1;
 294         event->wait_sem = wakee_event->wait_sem;
 295
 296         nr_wakeup_events++;
 297 }
 298
 299 static void
 300 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 301                       u64 task_state __used)
 302 {
 303         struct sched_atom *event = get_new_event(task, timestamp);
 304
 305         event->type = SCHED_EVENT_SLEEP;
 306
 307         nr_sleep_events++;
 308 }
 309
 310 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 311 {
 312         struct task_desc *task;
 313
 314         BUG_ON(pid >= MAX_PID);
 315
 316         task = pid_to_task[pid];
 317
 318         if (task)
 319                 return task;
 320
 321         task = zalloc(sizeof(*task));
 322         task->pid = pid;
 323         task->nr = nr_tasks;
 324         strcpy(task->comm, comm);
 325         /*
 326          * every task starts in sleeping state - this gets ignored
 327          * if there's no wakeup pointing to this sleep state:
 328          */
 329         add_sched_event_sleep(task, 0, 0);
 330
 331         pid_to_task[pid] = task;
 332         nr_tasks++;
 333         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 334         BUG_ON(!tasks);
 335         tasks[task->nr] = task;
 336
 337         if (verbose)
 338                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 339
 340         return task;
 341 }
 342
 343
 344 static void print_task_traces(void)
 345 {
 346         struct task_desc *task;
 347         unsigned long i;
 348
 349         for (i = 0; i < nr_tasks; i++) {
 350                 task = tasks[i];
 351                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 352                         task->nr, task->comm, task->pid, task->nr_events);
 353         }
 354 }
 355
 356 static void add_cross_task_wakeups(void)
 357 {
 358         struct task_desc *task1, *task2;
 359         unsigned long i, j;
 360
 361         for (i = 0; i < nr_tasks; i++) {
 362                 task1 = tasks[i];
 363                 j = i + 1;
 364                 if (j == nr_tasks)
 365                         j = 0;
 366                 task2 = tasks[j];
 367                 add_sched_event_wakeup(task1, 0, task2);
 368         }
 369 }
 370
 371 static void
 372 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 373 {
 374         int ret = 0;
 375         u64 now;
 376         long long delta;
 377
 378         now = get_nsecs();
 379         delta = start_time + atom->timestamp - now;
 380
 381         switch (atom->type) {
 382                 case SCHED_EVENT_RUN:
 383                         burn_nsecs(atom->duration);
 384                         break;
 385                 case SCHED_EVENT_SLEEP:
 386                         if (atom->wait_sem)
 387                                 ret = sem_wait(atom->wait_sem);
 388                         BUG_ON(ret);
 389                         break;
 390                 case SCHED_EVENT_WAKEUP:
 391                         if (atom->wait_sem)
 392                                 ret = sem_post(atom->wait_sem);
 393                         BUG_ON(ret);
 394                         break;
 395                 case SCHED_EVENT_MIGRATION:
 396                         break;
 397                 default:
 398                         BUG_ON(1);
 399         }
 400 }
 401
 402 static u64 get_cpu_usage_nsec_parent(void)
 403 {
 404         struct rusage ru;
 405         u64 sum;
 406         int err;
 407
 408         err = getrusage(RUSAGE_SELF, &ru);
 409         BUG_ON(err);
 410
 411         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 412         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 413
 414         return sum;
 415 }
 416
 417 static u64 get_cpu_usage_nsec_self(void)
 418 {
 419         char filename [] = "/proc/1234567890/sched";
 420         unsigned long msecs, nsecs;
 421         char *line = NULL;
 422         u64 total = 0;
 423         size_t len = 0;
 424         ssize_t chars;
 425         FILE *file;
 426         int ret;
 427
 428         sprintf(filename, "/proc/%d/sched", getpid());
 429         file = fopen(filename, "r");
 430         BUG_ON(!file);
 431
 432         while ((chars = getline(&line, &len, file)) != -1) {
 433                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 434                         &msecs, &nsecs);
 435                 if (ret == 2) {
 436                         total = msecs*1e6 + nsecs;
 437                         break;
 438                 }
 439         }
 440         if (line)
 441                 free(line);
 442         fclose(file);
 443
 444         return total;
 445 }
 446
 447 static void *thread_func(void *ctx)
 448 {
 449         struct task_desc *this_task = ctx;
 450         u64 cpu_usage_0, cpu_usage_1;
 451         unsigned long i, ret;
 452         char comm2[22];
 453
 454         sprintf(comm2, ":%s", this_task->comm);
 455         prctl(PR_SET_NAME, comm2);
 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();
 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();
 473         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 474
 475         ret = sem_post(&this_task->work_done_sem);
 476         BUG_ON(ret);
 477
 478         ret = pthread_mutex_lock(&work_done_wait_mutex);
 479         BUG_ON(ret);
 480         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 481         BUG_ON(ret);
 482
 483         goto again;
 484 }
 485
 486 static void create_tasks(void)
 487 {
 488         struct task_desc *task;
 489         pthread_attr_t attr;
 490         unsigned long i;
 491         int err;
 492
 493         err = pthread_attr_init(&attr);
 494         BUG_ON(err);
 495         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 496         BUG_ON(err);
 497         err = pthread_mutex_lock(&start_work_mutex);
 498         BUG_ON(err);
 499         err = pthread_mutex_lock(&work_done_wait_mutex);
 500         BUG_ON(err);
 501         for (i = 0; i < nr_tasks; i++) {
 502                 task = tasks[i];
 503                 sem_init(&task->sleep_sem, 0, 0);
 504                 sem_init(&task->ready_for_work, 0, 0);
 505                 sem_init(&task->work_done_sem, 0, 0);
 506                 task->curr_event = 0;
 507                 err = pthread_create(&task->thread, &attr, thread_func, task);
 508                 BUG_ON(err);
 509         }
 510 }
 511
 512 static void wait_for_tasks(void)
 513 {
 514         u64 cpu_usage_0, cpu_usage_1;
 515         struct task_desc *task;
 516         unsigned long i, ret;
 517
 518         start_time = get_nsecs();
 519         cpu_usage = 0;
 520         pthread_mutex_unlock(&work_done_wait_mutex);
 521
 522         for (i = 0; i < nr_tasks; i++) {
 523                 task = tasks[i];
 524                 ret = sem_wait(&task->ready_for_work);
 525                 BUG_ON(ret);
 526                 sem_init(&task->ready_for_work, 0, 0);
 527         }
 528         ret = pthread_mutex_lock(&work_done_wait_mutex);
 529         BUG_ON(ret);
 530
 531         cpu_usage_0 = get_cpu_usage_nsec_parent();
 532
 533         pthread_mutex_unlock(&start_work_mutex);
 534
 535         for (i = 0; i < nr_tasks; i++) {
 536                 task = tasks[i];
 537                 ret = sem_wait(&task->work_done_sem);
 538                 BUG_ON(ret);
 539                 sem_init(&task->work_done_sem, 0, 0);
 540                 cpu_usage += task->cpu_usage;
 541                 task->cpu_usage = 0;
 542         }
 543
 544         cpu_usage_1 = get_cpu_usage_nsec_parent();
 545         if (!runavg_cpu_usage)
 546                 runavg_cpu_usage = cpu_usage;
 547         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 548
 549         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 550         if (!runavg_parent_cpu_usage)
 551                 runavg_parent_cpu_usage = parent_cpu_usage;
 552         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 553                                    parent_cpu_usage)/10;
 554
 555         ret = pthread_mutex_lock(&start_work_mutex);
 556         BUG_ON(ret);
 557
 558         for (i = 0; i < nr_tasks; i++) {
 559                 task = tasks[i];
 560                 sem_init(&task->sleep_sem, 0, 0);
 561                 task->curr_event = 0;
 562         }
 563 }
 564
 565 static void run_one_test(void)
 566 {
 567         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 568
 569         T0 = get_nsecs();
 570         wait_for_tasks();
 571         T1 = get_nsecs();
 572
 573         delta = T1 - T0;
 574         sum_runtime += delta;
 575         nr_runs++;
 576
 577         avg_delta = sum_runtime / nr_runs;
 578         if (delta < avg_delta)
 579                 fluct = avg_delta - delta;
 580         else
 581                 fluct = delta - avg_delta;
 582         sum_fluct += fluct;
 583         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 584         if (!run_avg)
 585                 run_avg = delta;
 586         run_avg = (run_avg*9 + delta)/10;
 587
 588         printf("#%-3ld: %0.3f, ",
 589                 nr_runs, (double)delta/1000000.0);
 590
 591         printf("ravg: %0.2f, ",
 592                 (double)run_avg/1e6);
 593
 594         printf("cpu: %0.2f / %0.2f",
 595                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 596
 597 #if 0
 598         /*
 599          * rusage statistics done by the parent, these are less
 600          * accurate than the sum_exec_runtime based statistics:
 601          */
 602         printf(" [%0.2f / %0.2f]",
 603                 (double)parent_cpu_usage/1e6,
 604                 (double)runavg_parent_cpu_usage/1e6);
 605 #endif
 606
 607         printf("\n");
 608
 609         if (nr_sleep_corrections)
 610                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 611         nr_sleep_corrections = 0;
 612 }
 613
 614 static void test_calibrations(void)
 615 {
 616         u64 T0, T1;
 617
 618         T0 = get_nsecs();
 619         burn_nsecs(1e6);
 620         T1 = get_nsecs();
 621
 622         printf("the run test took %Ld nsecs\n", T1-T0);
 623
 624         T0 = get_nsecs();
 625         sleep_nsecs(1e6);
 626         T1 = get_nsecs();
 627
 628         printf("the sleep test took %Ld nsecs\n", T1-T0);
 629 }
 630
 631 #define FILL_FIELD(ptr, field, event, data)     \
 632         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 633
 634 #define FILL_ARRAY(ptr, array, event, data)                     \
 635 do {                                                            \
 636         void *__array = raw_field_ptr(event, #array, data);     \
 637         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 638 } while(0)
 639
 640 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 641 do {                                                            \
 642         FILL_FIELD(ptr, common_type, event, data);              \
 643         FILL_FIELD(ptr, common_flags, event, data);             \
 644         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 645         FILL_FIELD(ptr, common_pid, event, data);               \
 646         FILL_FIELD(ptr, common_tgid, event, data);              \
 647 } while (0)
 648
 649
 650
 651 struct trace_switch_event {
 652         u32 size;
 653
 654         u16 common_type;
 655         u8 common_flags;
 656         u8 common_preempt_count;
 657         u32 common_pid;
 658         u32 common_tgid;
 659
 660         char prev_comm[16];
 661         u32 prev_pid;
 662         u32 prev_prio;
 663         u64 prev_state;
 664         char next_comm[16];
 665         u32 next_pid;
 666         u32 next_prio;
 667 };
 668
 669 struct trace_runtime_event {
 670         u32 size;
 671
 672         u16 common_type;
 673         u8 common_flags;
 674         u8 common_preempt_count;
 675         u32 common_pid;
 676         u32 common_tgid;
 677
 678         char comm[16];
 679         u32 pid;
 680         u64 runtime;
 681         u64 vruntime;
 682 };
 683
 684 struct trace_wakeup_event {
 685         u32 size;
 686
 687         u16 common_type;
 688         u8 common_flags;
 689         u8 common_preempt_count;
 690         u32 common_pid;
 691         u32 common_tgid;
 692
 693         char comm[16];
 694         u32 pid;
 695
 696         u32 prio;
 697         u32 success;
 698         u32 cpu;
 699 };
 700
 701 struct trace_fork_event {
 702         u32 size;
 703
 704         u16 common_type;
 705         u8 common_flags;
 706         u8 common_preempt_count;
 707         u32 common_pid;
 708         u32 common_tgid;
 709
 710         char parent_comm[16];
 711         u32 parent_pid;
 712         char child_comm[16];
 713         u32 child_pid;
 714 };
 715
 716 struct trace_migrate_task_event {
 717         u32 size;
 718
 719         u16 common_type;
 720         u8 common_flags;
 721         u8 common_preempt_count;
 722         u32 common_pid;
 723         u32 common_tgid;
 724
 725         char comm[16];
 726         u32 pid;
 727
 728         u32 prio;
 729         u32 cpu;
 730 };
 731
 732 struct trace_sched_handler {
 733         void (*switch_event)(struct trace_switch_event *,
 734                              struct event *,
 735                              int cpu,
 736                              u64 timestamp,
 737                              struct thread *thread);
 738
 739         void (*runtime_event)(struct trace_runtime_event *,
 740                               struct event *,
 741                               int cpu,
 742                               u64 timestamp,
 743                               struct thread *thread);
 744
 745         void (*wakeup_event)(struct trace_wakeup_event *,
 746                              struct event *,
 747                              int cpu,
 748                              u64 timestamp,
 749                              struct thread *thread);
 750
 751         void (*fork_event)(struct trace_fork_event *,
 752                            struct event *,
 753                            int cpu,
 754                            u64 timestamp,
 755                            struct thread *thread);
 756
 757         void (*migrate_task_event)(struct trace_migrate_task_event *,
 758                            struct event *,
 759                            int cpu,
 760                            u64 timestamp,
 761                            struct thread *thread);
 762 };
 763
 764
 765 static void
 766 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 767                     struct event *event,
 768                     int cpu __used,
 769                     u64 timestamp __used,
 770                     struct thread *thread __used)
 771 {
 772         struct task_desc *waker, *wakee;
 773
 774         if (verbose) {
 775                 printf("sched_wakeup event %p\n", event);
 776
 777                 printf(" ... pid %d woke up %s/%d\n",
 778                         wakeup_event->common_pid,
 779                         wakeup_event->comm,
 780                         wakeup_event->pid);
 781         }
 782
 783         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 784         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 785
 786         add_sched_event_wakeup(waker, timestamp, wakee);
 787 }
 788
 789 static u64 cpu_last_switched[MAX_CPUS];
 790
 791 static void
 792 replay_switch_event(struct trace_switch_event *switch_event,
 793                     struct event *event,
 794                     int cpu,
 795                     u64 timestamp,
 796                     struct thread *thread __used)
 797 {
 798         struct task_desc *prev, *next;
 799         u64 timestamp0;
 800         s64 delta;
 801
 802         if (verbose)
 803                 printf("sched_switch event %p\n", event);
 804
 805         if (cpu >= MAX_CPUS || cpu < 0)
 806                 return;
 807
 808         timestamp0 = cpu_last_switched[cpu];
 809         if (timestamp0)
 810                 delta = timestamp - timestamp0;
 811         else
 812                 delta = 0;
 813
 814         if (delta < 0)
 815                 die("hm, delta: %Ld < 0 ?\n", delta);
 816
 817         if (verbose) {
 818                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 819                         switch_event->prev_comm, switch_event->prev_pid,
 820                         switch_event->next_comm, switch_event->next_pid,
 821                         delta);
 822         }
 823
 824         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 825         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 826
 827         cpu_last_switched[cpu] = timestamp;
 828
 829         add_sched_event_run(prev, timestamp, delta);
 830         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 831 }
 832
 833
 834 static void
 835 replay_fork_event(struct trace_fork_event *fork_event,
 836                   struct event *event,
 837                   int cpu __used,
 838                   u64 timestamp __used,
 839                   struct thread *thread __used)
 840 {
 841         if (verbose) {
 842                 printf("sched_fork event %p\n", event);
 843                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 844                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 845         }
 846         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 847         register_pid(fork_event->child_pid, fork_event->child_comm);
 848 }
 849
 850 static struct trace_sched_handler replay_ops  = {
 851         .wakeup_event           = replay_wakeup_event,
 852         .switch_event           = replay_switch_event,
 853         .fork_event             = replay_fork_event,
 854 };
 855
 856 struct sort_dimension {
 857         const char              *name;
 858         sort_fn_t               cmp;
 859         struct list_head        list;
 860 };
 861
 862 static LIST_HEAD(cmp_pid);
 863
 864 static int
 865 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 866 {
 867         struct sort_dimension *sort;
 868         int ret = 0;
 869
 870         BUG_ON(list_empty(list));
 871
 872         list_for_each_entry(sort, list, list) {
 873                 ret = sort->cmp(l, r);
 874                 if (ret)
 875                         return ret;
 876         }
 877
 878         return ret;
 879 }
 880
 881 static struct work_atoms *
 882 thread_atoms_search(struct rb_root *root, struct thread *thread,
 883                          struct list_head *sort_list)
 884 {
 885         struct rb_node *node = root->rb_node;
 886         struct work_atoms key = { .thread = thread };
 887
 888         while (node) {
 889                 struct work_atoms *atoms;
 890                 int cmp;
 891
 892                 atoms = container_of(node, struct work_atoms, node);
 893
 894                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 895                 if (cmp > 0)
 896                         node = node->rb_left;
 897                 else if (cmp < 0)
 898                         node = node->rb_right;
 899                 else {
 900                         BUG_ON(thread != atoms->thread);
 901                         return atoms;
 902                 }
 903         }
 904         return NULL;
 905 }
 906
 907 static void
 908 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 909                          struct list_head *sort_list)
 910 {
 911         struct rb_node **new = &(root->rb_node), *parent = NULL;
 912
 913         while (*new) {
 914                 struct work_atoms *this;
 915                 int cmp;
 916
 917                 this = container_of(*new, struct work_atoms, node);
 918                 parent = *new;
 919
 920                 cmp = thread_lat_cmp(sort_list, data, this);
 921
 922                 if (cmp > 0)
 923                         new = &((*new)->rb_left);
 924                 else
 925                         new = &((*new)->rb_right);
 926         }
 927
 928         rb_link_node(&data->node, parent, new);
 929         rb_insert_color(&data->node, root);
 930 }
 931
 932 static void thread_atoms_insert(struct thread *thread)
 933 {
 934         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 935         if (!atoms)
 936                 die("No memory");
 937
 938         atoms->thread = thread;
 939         INIT_LIST_HEAD(&atoms->work_list);
 940         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 941 }
 942
 943 static void
 944 latency_fork_event(struct trace_fork_event *fork_event __used,
 945                    struct event *event __used,
 946                    int cpu __used,
 947                    u64 timestamp __used,
 948                    struct thread *thread __used)
 949 {
 950         /* should insert the newcomer */
 951 }
 952
 953 __used
 954 static char sched_out_state(struct trace_switch_event *switch_event)
 955 {
 956         const char *str = TASK_STATE_TO_CHAR_STR;
 957
 958         return str[switch_event->prev_state];
 959 }
 960
 961 static void
 962 add_sched_out_event(struct work_atoms *atoms,
 963                     char run_state,
 964                     u64 timestamp)
 965 {
 966         struct work_atom *atom = zalloc(sizeof(*atom));
 967         if (!atom)
 968                 die("Non memory");
 969
 970         atom->sched_out_time = timestamp;
 971
 972         if (run_state == 'R') {
 973                 atom->state = THREAD_WAIT_CPU;
 974                 atom->wake_up_time = atom->sched_out_time;
 975         }
 976
 977         list_add_tail(&atom->list, &atoms->work_list);
 978 }
 979
 980 static void
 981 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
 982 {
 983         struct work_atom *atom;
 984
 985         BUG_ON(list_empty(&atoms->work_list));
 986
 987         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 988
 989         atom->runtime += delta;
 990         atoms->total_runtime += delta;
 991 }
 992
 993 static void
 994 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 995 {
 996         struct work_atom *atom;
 997         u64 delta;
 998
 999         if (list_empty(&atoms->work_list))
1000                 return;
1001
1002         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1003
1004         if (atom->state != THREAD_WAIT_CPU)
1005                 return;
1006
1007         if (timestamp < atom->wake_up_time) {
1008                 atom->state = THREAD_IGNORE;
1009                 return;
1010         }
1011
1012         atom->state = THREAD_SCHED_IN;
1013         atom->sched_in_time = timestamp;
1014
1015         delta = atom->sched_in_time - atom->wake_up_time;
1016         atoms->total_lat += delta;
1017         if (delta > atoms->max_lat)
1018                 atoms->max_lat = delta;
1019         atoms->nb_atoms++;
1020 }
1021
1022 static void
1023 latency_switch_event(struct trace_switch_event *switch_event,
1024                      struct event *event __used,
1025                      int cpu,
1026                      u64 timestamp,
1027                      struct thread *thread __used)
1028 {
1029         struct work_atoms *out_events, *in_events;
1030         struct thread *sched_out, *sched_in;
1031         u64 timestamp0;
1032         s64 delta;
1033
1034         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1035
1036         timestamp0 = cpu_last_switched[cpu];
1037         cpu_last_switched[cpu] = timestamp;
1038         if (timestamp0)
1039                 delta = timestamp - timestamp0;
1040         else
1041                 delta = 0;
1042
1043         if (delta < 0)
1044                 die("hm, delta: %Ld < 0 ?\n", delta);
1045
1046
1047         sched_out = threads__findnew(switch_event->prev_pid);
1048         sched_in = threads__findnew(switch_event->next_pid);
1049
1050         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1051         if (!out_events) {
1052                 thread_atoms_insert(sched_out);
1053                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1054                 if (!out_events)
1055                         die("out-event: Internal tree error");
1056         }
1057         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1058
1059         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1060         if (!in_events) {
1061                 thread_atoms_insert(sched_in);
1062                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1063                 if (!in_events)
1064                         die("in-event: Internal tree error");
1065                 /*
1066                  * Take came in we have not heard about yet,
1067                  * add in an initial atom in runnable state:
1068                  */
1069                 add_sched_out_event(in_events, 'R', timestamp);
1070         }
1071         add_sched_in_event(in_events, timestamp);
1072 }
1073
1074 static void
1075 latency_runtime_event(struct trace_runtime_event *runtime_event,
1076                      struct event *event __used,
1077                      int cpu,
1078                      u64 timestamp,
1079                      struct thread *this_thread __used)
1080 {
1081         struct thread *thread = threads__findnew(runtime_event->pid);
1082         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1083
1084         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1085         if (!atoms) {
1086                 thread_atoms_insert(thread);
1087                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1088                 if (!atoms)
1089                         die("in-event: Internal tree error");
1090                 add_sched_out_event(atoms, 'R', timestamp);
1091         }
1092
1093         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1094 }
1095
1096 static void
1097 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1098                      struct event *__event __used,
1099                      int cpu __used,
1100                      u64 timestamp,
1101                      struct thread *thread __used)
1102 {
1103         struct work_atoms *atoms;
1104         struct work_atom *atom;
1105         struct thread *wakee;
1106
1107         /* Note for later, it may be interesting to observe the failing cases */
1108         if (!wakeup_event->success)
1109                 return;
1110
1111         wakee = threads__findnew(wakeup_event->pid);
1112         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1113         if (!atoms) {
1114                 thread_atoms_insert(wakee);
1115                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1116                 if (!atoms)
1117                         die("wakeup-event: Internal tree error");
1118                 add_sched_out_event(atoms, 'S', timestamp);
1119         }
1120
1121         BUG_ON(list_empty(&atoms->work_list));
1122
1123         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1124
1125         /*
1126          * You WILL be missing events if you've recorded only
1127          * one CPU, or are only looking at only one, so don't
1128          * make useless noise.
1129          */
1130         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1131                 nr_state_machine_bugs++;
1132
1133         nr_timestamps++;
1134         if (atom->sched_out_time > timestamp) {
1135                 nr_unordered_timestamps++;
1136                 return;
1137         }
1138
1139         atom->state = THREAD_WAIT_CPU;
1140         atom->wake_up_time = timestamp;
1141 }
1142
1143 static void
1144 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1145                      struct event *__event __used,
1146                      int cpu __used,
1147                      u64 timestamp,
1148                      struct thread *thread __used)
1149 {
1150         struct work_atoms *atoms;
1151         struct work_atom *atom;
1152         struct thread *migrant;
1153
1154         /*
1155          * Only need to worry about migration when profiling one CPU.
1156          */
1157         if (profile_cpu == -1)
1158                 return;
1159
1160         migrant = threads__findnew(migrate_task_event->pid);
1161         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1162         if (!atoms) {
1163                 thread_atoms_insert(migrant);
1164                 register_pid(migrant->pid, migrant->comm);
1165                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1166                 if (!atoms)
1167                         die("migration-event: Internal tree error");
1168                 add_sched_out_event(atoms, 'R', timestamp);
1169         }
1170
1171         BUG_ON(list_empty(&atoms->work_list));
1172
1173         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1174         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1175
1176         nr_timestamps++;
1177
1178         if (atom->sched_out_time > timestamp)
1179                 nr_unordered_timestamps++;
1180 }
1181
1182 static struct trace_sched_handler lat_ops  = {
1183         .wakeup_event           = latency_wakeup_event,
1184         .switch_event           = latency_switch_event,
1185         .runtime_event          = latency_runtime_event,
1186         .fork_event             = latency_fork_event,
1187         .migrate_task_event     = latency_migrate_task_event,
1188 };
1189
1190 static void output_lat_thread(struct work_atoms *work_list)
1191 {
1192         int i;
1193         int ret;
1194         u64 avg;
1195
1196         if (!work_list->nb_atoms)
1197                 return;
1198         /*
1199          * Ignore idle threads:
1200          */
1201         if (!strcmp(work_list->thread->comm, "swapper"))
1202                 return;
1203
1204         all_runtime += work_list->total_runtime;
1205         all_count += work_list->nb_atoms;
1206
1207         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1208
1209         for (i = 0; i < 24 - ret; i++)
1210                 printf(" ");
1211
1212         avg = work_list->total_lat / work_list->nb_atoms;
1213
1214         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1215               (double)work_list->total_runtime / 1e6,
1216                  work_list->nb_atoms, (double)avg / 1e6,
1217                  (double)work_list->max_lat / 1e6);
1218 }
1219
1220 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1221 {
1222         if (l->thread->pid < r->thread->pid)
1223                 return -1;
1224         if (l->thread->pid > r->thread->pid)
1225                 return 1;
1226
1227         return 0;
1228 }
1229
1230 static struct sort_dimension pid_sort_dimension = {
1231         .name                   = "pid",
1232         .cmp                    = pid_cmp,
1233 };
1234
1235 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1236 {
1237         u64 avgl, avgr;
1238
1239         if (!l->nb_atoms)
1240                 return -1;
1241
1242         if (!r->nb_atoms)
1243                 return 1;
1244
1245         avgl = l->total_lat / l->nb_atoms;
1246         avgr = r->total_lat / r->nb_atoms;
1247
1248         if (avgl < avgr)
1249                 return -1;
1250         if (avgl > avgr)
1251                 return 1;
1252
1253         return 0;
1254 }
1255
1256 static struct sort_dimension avg_sort_dimension = {
1257         .name                   = "avg",
1258         .cmp                    = avg_cmp,
1259 };
1260
1261 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1262 {
1263         if (l->max_lat < r->max_lat)
1264                 return -1;
1265         if (l->max_lat > r->max_lat)
1266                 return 1;
1267
1268         return 0;
1269 }
1270
1271 static struct sort_dimension max_sort_dimension = {
1272         .name                   = "max",
1273         .cmp                    = max_cmp,
1274 };
1275
1276 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1277 {
1278         if (l->nb_atoms < r->nb_atoms)
1279                 return -1;
1280         if (l->nb_atoms > r->nb_atoms)
1281                 return 1;
1282
1283         return 0;
1284 }
1285
1286 static struct sort_dimension switch_sort_dimension = {
1287         .name                   = "switch",
1288         .cmp                    = switch_cmp,
1289 };
1290
1291 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1292 {
1293         if (l->total_runtime < r->total_runtime)
1294                 return -1;
1295         if (l->total_runtime > r->total_runtime)
1296                 return 1;
1297
1298         return 0;
1299 }
1300
1301 static struct sort_dimension runtime_sort_dimension = {
1302         .name                   = "runtime",
1303         .cmp                    = runtime_cmp,
1304 };
1305
1306 static struct sort_dimension *available_sorts[] = {
1307         &pid_sort_dimension,
1308         &avg_sort_dimension,
1309         &max_sort_dimension,
1310         &switch_sort_dimension,
1311         &runtime_sort_dimension,
1312 };
1313
1314 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1315
1316 static LIST_HEAD(sort_list);
1317
1318 static int sort_dimension__add(const char *tok, struct list_head *list)
1319 {
1320         int i;
1321
1322         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1323                 if (!strcmp(available_sorts[i]->name, tok)) {
1324                         list_add_tail(&available_sorts[i]->list, list);
1325
1326                         return 0;
1327                 }
1328         }
1329
1330         return -1;
1331 }
1332
1333 static void setup_sorting(void);
1334
1335 static void sort_lat(void)
1336 {
1337         struct rb_node *node;
1338
1339         for (;;) {
1340                 struct work_atoms *data;
1341                 node = rb_first(&atom_root);
1342                 if (!node)
1343                         break;
1344
1345                 rb_erase(node, &atom_root);
1346                 data = rb_entry(node, struct work_atoms, node);
1347                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1348         }
1349 }
1350
1351 static struct trace_sched_handler *trace_handler;
1352
1353 static void
1354 process_sched_wakeup_event(void *data,
1355                            struct event *event,
1356                            int cpu __used,
1357                            u64 timestamp __used,
1358                            struct thread *thread __used)
1359 {
1360         struct trace_wakeup_event wakeup_event;
1361
1362         FILL_COMMON_FIELDS(wakeup_event, event, data);
1363
1364         FILL_ARRAY(wakeup_event, comm, event, data);
1365         FILL_FIELD(wakeup_event, pid, event, data);
1366         FILL_FIELD(wakeup_event, prio, event, data);
1367         FILL_FIELD(wakeup_event, success, event, data);
1368         FILL_FIELD(wakeup_event, cpu, event, data);
1369
1370         if (trace_handler->wakeup_event)
1371                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1372 }
1373
1374 /*
1375  * Track the current task - that way we can know whether there's any
1376  * weird events, such as a task being switched away that is not current.
1377  */
1378 static int max_cpu;
1379
1380 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1381
1382 static struct thread *curr_thread[MAX_CPUS];
1383
1384 static char next_shortname1 = 'A';
1385 static char next_shortname2 = '0';
1386
1387 static void
1388 map_switch_event(struct trace_switch_event *switch_event,
1389                  struct event *event __used,
1390                  int this_cpu,
1391                  u64 timestamp,
1392                  struct thread *thread __used)
1393 {
1394         struct thread *sched_out, *sched_in;
1395         int new_shortname;
1396         u64 timestamp0;
1397         s64 delta;
1398         int cpu;
1399
1400         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1401
1402         if (this_cpu > max_cpu)
1403                 max_cpu = this_cpu;
1404
1405         timestamp0 = cpu_last_switched[this_cpu];
1406         cpu_last_switched[this_cpu] = timestamp;
1407         if (timestamp0)
1408                 delta = timestamp - timestamp0;
1409         else
1410                 delta = 0;
1411
1412         if (delta < 0)
1413                 die("hm, delta: %Ld < 0 ?\n", delta);
1414
1415
1416         sched_out = threads__findnew(switch_event->prev_pid);
1417         sched_in = threads__findnew(switch_event->next_pid);
1418
1419         curr_thread[this_cpu] = sched_in;
1420
1421         printf("  ");
1422
1423         new_shortname = 0;
1424         if (!sched_in->shortname[0]) {
1425                 sched_in->shortname[0] = next_shortname1;
1426                 sched_in->shortname[1] = next_shortname2;
1427
1428                 if (next_shortname1 < 'Z') {
1429                         next_shortname1++;
1430                 } else {
1431                         next_shortname1='A';
1432                         if (next_shortname2 < '9') {
1433                                 next_shortname2++;
1434                         } else {
1435                                 next_shortname2='0';
1436                         }
1437                 }
1438                 new_shortname = 1;
1439         }
1440
1441         for (cpu = 0; cpu <= max_cpu; cpu++) {
1442                 if (cpu != this_cpu)
1443                         printf(" ");
1444                 else
1445                         printf("*");
1446
1447                 if (curr_thread[cpu]) {
1448                         if (curr_thread[cpu]->pid)
1449                                 printf("%2s ", curr_thread[cpu]->shortname);
1450                         else
1451                                 printf(".  ");
1452                 } else
1453                         printf("   ");
1454         }
1455
1456         printf("  %12.6f secs ", (double)timestamp/1e9);
1457         if (new_shortname) {
1458                 printf("%s => %s:%d\n",
1459                         sched_in->shortname, sched_in->comm, sched_in->pid);
1460         } else {
1461                 printf("\n");
1462         }
1463 }
1464
1465
1466 static void
1467 process_sched_switch_event(void *data,
1468                            struct event *event,
1469                            int this_cpu,
1470                            u64 timestamp __used,
1471                            struct thread *thread __used)
1472 {
1473         struct trace_switch_event switch_event;
1474
1475         FILL_COMMON_FIELDS(switch_event, event, data);
1476
1477         FILL_ARRAY(switch_event, prev_comm, event, data);
1478         FILL_FIELD(switch_event, prev_pid, event, data);
1479         FILL_FIELD(switch_event, prev_prio, event, data);
1480         FILL_FIELD(switch_event, prev_state, event, data);
1481         FILL_ARRAY(switch_event, next_comm, event, data);
1482         FILL_FIELD(switch_event, next_pid, event, data);
1483         FILL_FIELD(switch_event, next_prio, event, data);
1484
1485         if (curr_pid[this_cpu] != (u32)-1) {
1486                 /*
1487                  * Are we trying to switch away a PID that is
1488                  * not current?
1489                  */
1490                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1491                         nr_context_switch_bugs++;
1492         }
1493         if (trace_handler->switch_event)
1494                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1495
1496         curr_pid[this_cpu] = switch_event.next_pid;
1497 }
1498
1499 static void
1500 process_sched_runtime_event(void *data,
1501                            struct event *event,
1502                            int cpu __used,
1503                            u64 timestamp __used,
1504                            struct thread *thread __used)
1505 {
1506         struct trace_runtime_event runtime_event;
1507
1508         FILL_ARRAY(runtime_event, comm, event, data);
1509         FILL_FIELD(runtime_event, pid, event, data);
1510         FILL_FIELD(runtime_event, runtime, event, data);
1511         FILL_FIELD(runtime_event, vruntime, event, data);
1512
1513         if (trace_handler->runtime_event)
1514                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1515 }
1516
1517 static void
1518 process_sched_fork_event(void *data,
1519                          struct event *event,
1520                          int cpu __used,
1521                          u64 timestamp __used,
1522                          struct thread *thread __used)
1523 {
1524         struct trace_fork_event fork_event;
1525
1526         FILL_COMMON_FIELDS(fork_event, event, data);
1527
1528         FILL_ARRAY(fork_event, parent_comm, event, data);
1529         FILL_FIELD(fork_event, parent_pid, event, data);
1530         FILL_ARRAY(fork_event, child_comm, event, data);
1531         FILL_FIELD(fork_event, child_pid, event, data);
1532
1533         if (trace_handler->fork_event)
1534                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1535 }
1536
1537 static void
1538 process_sched_exit_event(struct event *event,
1539                          int cpu __used,
1540                          u64 timestamp __used,
1541                          struct thread *thread __used)
1542 {
1543         if (verbose)
1544                 printf("sched_exit event %p\n", event);
1545 }
1546
1547 static void
1548 process_sched_migrate_task_event(void *data,
1549                            struct event *event,
1550                            int cpu __used,
1551                            u64 timestamp __used,
1552                            struct thread *thread __used)
1553 {
1554         struct trace_migrate_task_event migrate_task_event;
1555
1556         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1557
1558         FILL_ARRAY(migrate_task_event, comm, event, data);
1559         FILL_FIELD(migrate_task_event, pid, event, data);
1560         FILL_FIELD(migrate_task_event, prio, event, data);
1561         FILL_FIELD(migrate_task_event, cpu, event, data);
1562
1563         if (trace_handler->migrate_task_event)
1564                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1565 }
1566
1567 static void
1568 process_raw_event(event_t *raw_event __used, void *data,
1569                   int cpu, u64 timestamp, struct thread *thread)
1570 {
1571         struct event *event;
1572         int type;
1573
1574
1575         type = trace_parse_common_type(data);
1576         event = trace_find_event(type);
1577
1578         if (!strcmp(event->name, "sched_switch"))
1579                 process_sched_switch_event(data, event, cpu, timestamp, thread);
1580         if (!strcmp(event->name, "sched_stat_runtime"))
1581                 process_sched_runtime_event(data, event, cpu, timestamp, thread);
1582         if (!strcmp(event->name, "sched_wakeup"))
1583                 process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1584         if (!strcmp(event->name, "sched_wakeup_new"))
1585                 process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1586         if (!strcmp(event->name, "sched_process_fork"))
1587                 process_sched_fork_event(data, event, cpu, timestamp, thread);
1588         if (!strcmp(event->name, "sched_process_exit"))
1589                 process_sched_exit_event(event, cpu, timestamp, thread);
1590         if (!strcmp(event->name, "sched_migrate_task"))
1591                 process_sched_migrate_task_event(data, event, cpu, timestamp, thread);
1592 }
1593
1594 static int process_sample_event(event_t *event)
1595 {
1596         struct sample_data data;
1597         struct thread *thread;
1598
1599         if (!(sample_type & PERF_SAMPLE_RAW))
1600                 return 0;
1601
1602         memset(&data, 0, sizeof(data));
1603         data.time = -1;
1604         data.cpu = -1;
1605         data.period = -1;
1606
1607         event__parse_sample(event, sample_type, &data);
1608
1609         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1610                 event->header.misc,
1611                 data.pid, data.tid,
1612                 (void *)(long)data.ip,
1613                 (long long)data.period);
1614
1615         thread = threads__findnew(data.pid);
1616         if (thread == NULL) {
1617                 pr_debug("problem processing %d event, skipping it.\n",
1618                          event->header.type);
1619                 return -1;
1620         }
1621
1622         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1623
1624         if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1625                 return 0;
1626
1627         process_raw_event(event, data.raw_data, data.cpu, data.time, thread);
1628
1629         return 0;
1630 }
1631
1632 static int process_lost_event(event_t *event __used)
1633 {
1634         nr_lost_chunks++;
1635         nr_lost_events += event->lost.lost;
1636
1637         return 0;
1638 }
1639
1640 static int sample_type_check(u64 type)
1641 {
1642         sample_type = type;
1643
1644         if (!(sample_type & PERF_SAMPLE_RAW)) {
1645                 fprintf(stderr,
1646                         "No trace sample to read. Did you call perf record "
1647                         "without -R?");
1648                 return -1;
1649         }
1650
1651         return 0;
1652 }
1653
1654 static struct perf_file_handler file_handler = {
1655         .process_sample_event   = process_sample_event,
1656         .process_comm_event     = event__process_comm,
1657         .process_lost_event     = process_lost_event,
1658         .sample_type_check      = sample_type_check,
1659 };
1660
1661 static int read_events(void)
1662 {
1663         register_idle_thread();
1664         register_perf_file_handler(&file_handler);
1665
1666         return mmap_dispatch_perf_file(&header, input_name, 0, 0,
1667                                        &event__cwdlen, &event__cwd);
1668 }
1669
1670 static void print_bad_events(void)
1671 {
1672         if (nr_unordered_timestamps && nr_timestamps) {
1673                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1674                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1675                         nr_unordered_timestamps, nr_timestamps);
1676         }
1677         if (nr_lost_events && nr_events) {
1678                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1679                         (double)nr_lost_events/(double)nr_events*100.0,
1680                         nr_lost_events, nr_events, nr_lost_chunks);
1681         }
1682         if (nr_state_machine_bugs && nr_timestamps) {
1683                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1684                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1685                         nr_state_machine_bugs, nr_timestamps);
1686                 if (nr_lost_events)
1687                         printf(" (due to lost events?)");
1688                 printf("\n");
1689         }
1690         if (nr_context_switch_bugs && nr_timestamps) {
1691                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1692                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1693                         nr_context_switch_bugs, nr_timestamps);
1694                 if (nr_lost_events)
1695                         printf(" (due to lost events?)");
1696                 printf("\n");
1697         }
1698 }
1699
1700 static void __cmd_lat(void)
1701 {
1702         struct rb_node *next;
1703
1704         setup_pager();
1705         read_events();
1706         sort_lat();
1707
1708         printf("\n -----------------------------------------------------------------------------------------\n");
1709         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1710         printf(" -----------------------------------------------------------------------------------------\n");
1711
1712         next = rb_first(&sorted_atom_root);
1713
1714         while (next) {
1715                 struct work_atoms *work_list;
1716
1717                 work_list = rb_entry(next, struct work_atoms, node);
1718                 output_lat_thread(work_list);
1719                 next = rb_next(next);
1720         }
1721
1722         printf(" -----------------------------------------------------------------------------------------\n");
1723         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1724                 (double)all_runtime/1e6, all_count);
1725
1726         printf(" ---------------------------------------------------\n");
1727
1728         print_bad_events();
1729         printf("\n");
1730
1731 }
1732
1733 static struct trace_sched_handler map_ops  = {
1734         .wakeup_event           = NULL,
1735         .switch_event           = map_switch_event,
1736         .runtime_event          = NULL,
1737         .fork_event             = NULL,
1738 };
1739
1740 static void __cmd_map(void)
1741 {
1742         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1743
1744         setup_pager();
1745         read_events();
1746         print_bad_events();
1747 }
1748
1749 static void __cmd_replay(void)
1750 {
1751         unsigned long i;
1752
1753         calibrate_run_measurement_overhead();
1754         calibrate_sleep_measurement_overhead();
1755
1756         test_calibrations();
1757
1758         read_events();
1759
1760         printf("nr_run_events:        %ld\n", nr_run_events);
1761         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1762         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1763
1764         if (targetless_wakeups)
1765                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1766         if (multitarget_wakeups)
1767                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1768         if (nr_run_events_optimized)
1769                 printf("run atoms optimized: %ld\n",
1770                         nr_run_events_optimized);
1771
1772         print_task_traces();
1773         add_cross_task_wakeups();
1774
1775         create_tasks();
1776         printf("------------------------------------------------------------\n");
1777         for (i = 0; i < replay_repeat; i++)
1778                 run_one_test();
1779 }
1780
1781
1782 static const char * const sched_usage[] = {
1783         "perf sched [<options>] {record|latency|map|replay|trace}",
1784         NULL
1785 };
1786
1787 static const struct option sched_options[] = {
1788         OPT_STRING('i', "input", &input_name, "file",
1789                     "input file name"),
1790         OPT_BOOLEAN('v', "verbose", &verbose,
1791                     "be more verbose (show symbol address, etc)"),
1792         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1793                     "dump raw trace in ASCII"),
1794         OPT_END()
1795 };
1796
1797 static const char * const latency_usage[] = {
1798         "perf sched latency [<options>]",
1799         NULL
1800 };
1801
1802 static const struct option latency_options[] = {
1803         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1804                    "sort by key(s): runtime, switch, avg, max"),
1805         OPT_BOOLEAN('v', "verbose", &verbose,
1806                     "be more verbose (show symbol address, etc)"),
1807         OPT_INTEGER('C', "CPU", &profile_cpu,
1808                     "CPU to profile on"),
1809         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1810                     "dump raw trace in ASCII"),
1811         OPT_END()
1812 };
1813
1814 static const char * const replay_usage[] = {
1815         "perf sched replay [<options>]",
1816         NULL
1817 };
1818
1819 static const struct option replay_options[] = {
1820         OPT_INTEGER('r', "repeat", &replay_repeat,
1821                     "repeat the workload replay N times (-1: infinite)"),
1822         OPT_BOOLEAN('v', "verbose", &verbose,
1823                     "be more verbose (show symbol address, etc)"),
1824         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1825                     "dump raw trace in ASCII"),
1826         OPT_END()
1827 };
1828
1829 static void setup_sorting(void)
1830 {
1831         char *tmp, *tok, *str = strdup(sort_order);
1832
1833         for (tok = strtok_r(str, ", ", &tmp);
1834                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1835                 if (sort_dimension__add(tok, &sort_list) < 0) {
1836                         error("Unknown --sort key: `%s'", tok);
1837                         usage_with_options(latency_usage, latency_options);
1838                 }
1839         }
1840
1841         free(str);
1842
1843         sort_dimension__add("pid", &cmp_pid);
1844 }
1845
1846 static const char *record_args[] = {
1847         "record",
1848         "-a",
1849         "-R",
1850         "-M",
1851         "-f",
1852         "-m", "1024",
1853         "-c", "1",
1854         "-e", "sched:sched_switch:r",
1855         "-e", "sched:sched_stat_wait:r",
1856         "-e", "sched:sched_stat_sleep:r",
1857         "-e", "sched:sched_stat_iowait:r",
1858         "-e", "sched:sched_stat_runtime:r",
1859         "-e", "sched:sched_process_exit:r",
1860         "-e", "sched:sched_process_fork:r",
1861         "-e", "sched:sched_wakeup:r",
1862         "-e", "sched:sched_migrate_task:r",
1863 };
1864
1865 static int __cmd_record(int argc, const char **argv)
1866 {
1867         unsigned int rec_argc, i, j;
1868         const char **rec_argv;
1869
1870         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1871         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1872
1873         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1874                 rec_argv[i] = strdup(record_args[i]);
1875
1876         for (j = 1; j < (unsigned int)argc; j++, i++)
1877                 rec_argv[i] = argv[j];
1878
1879         BUG_ON(i != rec_argc);
1880
1881         return cmd_record(i, rec_argv, NULL);
1882 }
1883
1884 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1885 {
1886         argc = parse_options(argc, argv, sched_options, sched_usage,
1887                              PARSE_OPT_STOP_AT_NON_OPTION);
1888         if (!argc)
1889                 usage_with_options(sched_usage, sched_options);
1890
1891         /*
1892          * Aliased to 'perf trace' for now:
1893          */
1894         if (!strcmp(argv[0], "trace"))
1895                 return cmd_trace(argc, argv, prefix);
1896
1897         symbol__init(0);
1898         if (!strncmp(argv[0], "rec", 3)) {
1899                 return __cmd_record(argc, argv);
1900         } else if (!strncmp(argv[0], "lat", 3)) {
1901                 trace_handler = &lat_ops;
1902                 if (argc > 1) {
1903                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1904                         if (argc)
1905                                 usage_with_options(latency_usage, latency_options);
1906                 }
1907                 setup_sorting();
1908                 __cmd_lat();
1909         } else if (!strcmp(argv[0], "map")) {
1910                 trace_handler = &map_ops;
1911                 setup_sorting();
1912                 __cmd_map();
1913         } else if (!strncmp(argv[0], "rep", 3)) {
1914                 trace_handler = &replay_ops;
1915                 if (argc) {
1916                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1917                         if (argc)
1918                                 usage_with_options(replay_usage, replay_options);
1919                 }
1920                 __cmd_replay();
1921         } else {
1922                 usage_with_options(sched_usage, sched_options);
1923         }
1924
1925         return 0;
1926 }