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
  15 #include <sys/types.h>
  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 static int                      input;
  24 static unsigned long            page_size;
  25 static unsigned long            mmap_window = 32;
  26
  27 static unsigned long            total_comm = 0;
  28
  29 static struct rb_root           threads;
  30 static struct thread            *last_match;
  31
  32 static struct perf_header       *header;
  33 static u64                      sample_type;
  34
  35 static char                     default_sort_order[] = "avg, max, switch, runtime";
  36 static char                     *sort_order = default_sort_order;
  37
  38 #define PR_SET_NAME             15               /* Set process name */
  39 #define MAX_CPUS                4096
  40
  41 #define BUG_ON(x)               assert(!(x))
  42
  43 static u64                      run_measurement_overhead;
  44 static u64                      sleep_measurement_overhead;
  45
  46 #define COMM_LEN                20
  47 #define SYM_LEN                 129
  48
  49 #define MAX_PID                 65536
  50
  51 static unsigned long            nr_tasks;
  52
  53 struct sched_event;
  54
  55 struct task_desc {
  56         unsigned long           nr;
  57         unsigned long           pid;
  58         char                    comm[COMM_LEN];
  59
  60         unsigned long           nr_events;
  61         unsigned long           curr_event;
  62         struct sched_event      **events;
  63
  64         pthread_t               thread;
  65         sem_t                   sleep_sem;
  66
  67         sem_t                   ready_for_work;
  68         sem_t                   work_done_sem;
  69
  70         u64                     cpu_usage;
  71 };
  72
  73 enum sched_event_type {
  74         SCHED_EVENT_RUN,
  75         SCHED_EVENT_SLEEP,
  76         SCHED_EVENT_WAKEUP,
  77 };
  78
  79 struct sched_event {
  80         enum sched_event_type   type;
  81         u64                     timestamp;
  82         u64                     duration;
  83         unsigned long           nr;
  84         int                     specific_wait;
  85         sem_t                   *wait_sem;
  86         struct task_desc        *wakee;
  87 };
  88
  89 static struct task_desc         *pid_to_task[MAX_PID];
  90
  91 static struct task_desc         **tasks;
  92
  93 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  94 static u64                      start_time;
  95
  96 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  97
  98 static unsigned long            nr_run_events;
  99 static unsigned long            nr_sleep_events;
 100 static unsigned long            nr_wakeup_events;
 101
 102 static unsigned long            nr_sleep_corrections;
 103 static unsigned long            nr_run_events_optimized;
 104
 105 static unsigned long            targetless_wakeups;
 106 static unsigned long            multitarget_wakeups;
 107
 108 static u64                      cpu_usage;
 109 static u64                      runavg_cpu_usage;
 110 static u64                      parent_cpu_usage;
 111 static u64                      runavg_parent_cpu_usage;
 112
 113 static unsigned long            nr_runs;
 114 static u64                      sum_runtime;
 115 static u64                      sum_fluct;
 116 static u64                      run_avg;
 117
 118 static unsigned long            replay_repeat = 10;
 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                     wake_up_time;
 133         u64                     sched_in_time;
 134         u64                     runtime;
 135 };
 136
 137 struct task_atoms {
 138         struct list_head        atom_list;
 139         struct thread           *thread;
 140         struct rb_node          node;
 141         u64                     max_lat;
 142         u64                     total_lat;
 143         u64                     nb_atoms;
 144         u64                     total_runtime;
 145 };
 146
 147 typedef int (*sort_thread_lat)(struct task_atoms *, struct task_atoms *);
 148
 149 static struct rb_root           atom_root, sorted_atom_root;
 150
 151 static u64                      all_runtime;
 152 static u64                      all_count;
 153
 154 static int read_events(void);
 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_event *
 221 get_new_event(struct task_desc *task, u64 timestamp)
 222 {
 223         struct sched_event *event = calloc(1, 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_event *) * task->nr_events;
 232         task->events = realloc(task->events, size);
 233         BUG_ON(!task->events);
 234
 235         task->events[idx] = event;
 236
 237         return event;
 238 }
 239
 240 static struct sched_event *last_event(struct task_desc *task)
 241 {
 242         if (!task->nr_events)
 243                 return NULL;
 244
 245         return task->events[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_event *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_event *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 = calloc(1, 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_event *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 = calloc(1, 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_event *event)
 373 {
 374         int ret = 0;
 375         u64 now;
 376         long long delta;
 377
 378         now = get_nsecs();
 379         delta = start_time + event->timestamp - now;
 380
 381         switch (event->type) {
 382                 case SCHED_EVENT_RUN:
 383                         burn_nsecs(event->duration);
 384                         break;
 385                 case SCHED_EVENT_SLEEP:
 386                         if (event->wait_sem)
 387                                 ret = sem_wait(event->wait_sem);
 388                         BUG_ON(ret);
 389                         break;
 390                 case SCHED_EVENT_WAKEUP:
 391                         if (event->wait_sem)
 392                                 ret = sem_post(event->wait_sem);
 393                         BUG_ON(ret);
 394                         break;
 395                 default:
 396                         BUG_ON(1);
 397         }
 398 }
 399
 400 static u64 get_cpu_usage_nsec_parent(void)
 401 {
 402         struct rusage ru;
 403         u64 sum;
 404         int err;
 405
 406         err = getrusage(RUSAGE_SELF, &ru);
 407         BUG_ON(err);
 408
 409         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 410         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 411
 412         return sum;
 413 }
 414
 415 static u64 get_cpu_usage_nsec_self(void)
 416 {
 417         char filename [] = "/proc/1234567890/sched";
 418         unsigned long msecs, nsecs;
 419         char *line = NULL;
 420         u64 total = 0;
 421         size_t len = 0;
 422         ssize_t chars;
 423         FILE *file;
 424         int ret;
 425
 426         sprintf(filename, "/proc/%d/sched", getpid());
 427         file = fopen(filename, "r");
 428         BUG_ON(!file);
 429
 430         while ((chars = getline(&line, &len, file)) != -1) {
 431                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 432                         &msecs, &nsecs);
 433                 if (ret == 2) {
 434                         total = msecs*1e6 + nsecs;
 435                         break;
 436                 }
 437         }
 438         if (line)
 439                 free(line);
 440         fclose(file);
 441
 442         return total;
 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
 452         sprintf(comm2, ":%s", this_task->comm);
 453         prctl(PR_SET_NAME, comm2);
 454
 455 again:
 456         ret = sem_post(&this_task->ready_for_work);
 457         BUG_ON(ret);
 458         ret = pthread_mutex_lock(&start_work_mutex);
 459         BUG_ON(ret);
 460         ret = pthread_mutex_unlock(&start_work_mutex);
 461         BUG_ON(ret);
 462
 463         cpu_usage_0 = get_cpu_usage_nsec_self();
 464
 465         for (i = 0; i < this_task->nr_events; i++) {
 466                 this_task->curr_event = i;
 467                 process_sched_event(this_task, this_task->events[i]);
 468         }
 469
 470         cpu_usage_1 = get_cpu_usage_nsec_self();
 471         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 472
 473         ret = sem_post(&this_task->work_done_sem);
 474         BUG_ON(ret);
 475
 476         ret = pthread_mutex_lock(&work_done_wait_mutex);
 477         BUG_ON(ret);
 478         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 479         BUG_ON(ret);
 480
 481         goto again;
 482 }
 483
 484 static void create_tasks(void)
 485 {
 486         struct task_desc *task;
 487         pthread_attr_t attr;
 488         unsigned long i;
 489         int err;
 490
 491         err = pthread_attr_init(&attr);
 492         BUG_ON(err);
 493         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 494         BUG_ON(err);
 495         err = pthread_mutex_lock(&start_work_mutex);
 496         BUG_ON(err);
 497         err = pthread_mutex_lock(&work_done_wait_mutex);
 498         BUG_ON(err);
 499         for (i = 0; i < nr_tasks; i++) {
 500                 task = tasks[i];
 501                 sem_init(&task->sleep_sem, 0, 0);
 502                 sem_init(&task->ready_for_work, 0, 0);
 503                 sem_init(&task->work_done_sem, 0, 0);
 504                 task->curr_event = 0;
 505                 err = pthread_create(&task->thread, &attr, thread_func, task);
 506                 BUG_ON(err);
 507         }
 508 }
 509
 510 static void wait_for_tasks(void)
 511 {
 512         u64 cpu_usage_0, cpu_usage_1;
 513         struct task_desc *task;
 514         unsigned long i, ret;
 515
 516         start_time = get_nsecs();
 517         cpu_usage = 0;
 518         pthread_mutex_unlock(&work_done_wait_mutex);
 519
 520         for (i = 0; i < nr_tasks; i++) {
 521                 task = tasks[i];
 522                 ret = sem_wait(&task->ready_for_work);
 523                 BUG_ON(ret);
 524                 sem_init(&task->ready_for_work, 0, 0);
 525         }
 526         ret = pthread_mutex_lock(&work_done_wait_mutex);
 527         BUG_ON(ret);
 528
 529         cpu_usage_0 = get_cpu_usage_nsec_parent();
 530
 531         pthread_mutex_unlock(&start_work_mutex);
 532
 533         for (i = 0; i < nr_tasks; i++) {
 534                 task = tasks[i];
 535                 ret = sem_wait(&task->work_done_sem);
 536                 BUG_ON(ret);
 537                 sem_init(&task->work_done_sem, 0, 0);
 538                 cpu_usage += task->cpu_usage;
 539                 task->cpu_usage = 0;
 540         }
 541
 542         cpu_usage_1 = get_cpu_usage_nsec_parent();
 543         if (!runavg_cpu_usage)
 544                 runavg_cpu_usage = cpu_usage;
 545         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 546
 547         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 548         if (!runavg_parent_cpu_usage)
 549                 runavg_parent_cpu_usage = parent_cpu_usage;
 550         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 551                                    parent_cpu_usage)/10;
 552
 553         ret = pthread_mutex_lock(&start_work_mutex);
 554         BUG_ON(ret);
 555
 556         for (i = 0; i < nr_tasks; i++) {
 557                 task = tasks[i];
 558                 sem_init(&task->sleep_sem, 0, 0);
 559                 task->curr_event = 0;
 560         }
 561 }
 562
 563 static void run_one_test(void)
 564 {
 565         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 566
 567         T0 = get_nsecs();
 568         wait_for_tasks();
 569         T1 = get_nsecs();
 570
 571         delta = T1 - T0;
 572         sum_runtime += delta;
 573         nr_runs++;
 574
 575         avg_delta = sum_runtime / nr_runs;
 576         if (delta < avg_delta)
 577                 fluct = avg_delta - delta;
 578         else
 579                 fluct = delta - avg_delta;
 580         sum_fluct += fluct;
 581         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 582         if (!run_avg)
 583                 run_avg = delta;
 584         run_avg = (run_avg*9 + delta)/10;
 585
 586         printf("#%-3ld: %0.3f, ",
 587                 nr_runs, (double)delta/1000000.0);
 588
 589         printf("ravg: %0.2f, ",
 590                 (double)run_avg/1e6);
 591
 592         printf("cpu: %0.2f / %0.2f",
 593                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 594
 595 #if 0
 596         /*
 597          * rusage statistics done by the parent, these are less
 598          * accurate than the sum_exec_runtime based statistics:
 599          */
 600         printf(" [%0.2f / %0.2f]",
 601                 (double)parent_cpu_usage/1e6,
 602                 (double)runavg_parent_cpu_usage/1e6);
 603 #endif
 604
 605         printf("\n");
 606
 607         if (nr_sleep_corrections)
 608                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 609         nr_sleep_corrections = 0;
 610 }
 611
 612 static void test_calibrations(void)
 613 {
 614         u64 T0, T1;
 615
 616         T0 = get_nsecs();
 617         burn_nsecs(1e6);
 618         T1 = get_nsecs();
 619
 620         printf("the run test took %Ld nsecs\n", T1-T0);
 621
 622         T0 = get_nsecs();
 623         sleep_nsecs(1e6);
 624         T1 = get_nsecs();
 625
 626         printf("the sleep test took %Ld nsecs\n", T1-T0);
 627 }
 628
 629 static void __cmd_replay(void)
 630 {
 631         unsigned long i;
 632
 633         calibrate_run_measurement_overhead();
 634         calibrate_sleep_measurement_overhead();
 635
 636         test_calibrations();
 637
 638         read_events();
 639
 640         printf("nr_run_events:        %ld\n", nr_run_events);
 641         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
 642         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
 643
 644         if (targetless_wakeups)
 645                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
 646         if (multitarget_wakeups)
 647                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
 648         if (nr_run_events_optimized)
 649                 printf("run events optimized: %ld\n",
 650                         nr_run_events_optimized);
 651
 652         print_task_traces();
 653         add_cross_task_wakeups();
 654
 655         create_tasks();
 656         printf("------------------------------------------------------------\n");
 657         for (i = 0; i < replay_repeat; i++)
 658                 run_one_test();
 659 }
 660
 661 static int
 662 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 663 {
 664         struct thread *thread;
 665
 666         thread = threads__findnew(event->comm.pid, &threads, &last_match);
 667
 668         dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
 669                 (void *)(offset + head),
 670                 (void *)(long)(event->header.size),
 671                 event->comm.comm, event->comm.pid);
 672
 673         if (thread == NULL ||
 674             thread__set_comm(thread, event->comm.comm)) {
 675                 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
 676                 return -1;
 677         }
 678         total_comm++;
 679
 680         return 0;
 681 }
 682
 683
 684 struct raw_event_sample {
 685         u32 size;
 686         char data[0];
 687 };
 688
 689 #define FILL_FIELD(ptr, field, event, data)     \
 690         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 691
 692 #define FILL_ARRAY(ptr, array, event, data)                     \
 693 do {                                                            \
 694         void *__array = raw_field_ptr(event, #array, data);     \
 695         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 696 } while(0)
 697
 698 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 699 do {                                                            \
 700         FILL_FIELD(ptr, common_type, event, data);              \
 701         FILL_FIELD(ptr, common_flags, event, data);             \
 702         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 703         FILL_FIELD(ptr, common_pid, event, data);               \
 704         FILL_FIELD(ptr, common_tgid, event, data);              \
 705 } while (0)
 706
 707
 708
 709 struct trace_switch_event {
 710         u32 size;
 711
 712         u16 common_type;
 713         u8 common_flags;
 714         u8 common_preempt_count;
 715         u32 common_pid;
 716         u32 common_tgid;
 717
 718         char prev_comm[16];
 719         u32 prev_pid;
 720         u32 prev_prio;
 721         u64 prev_state;
 722         char next_comm[16];
 723         u32 next_pid;
 724         u32 next_prio;
 725 };
 726
 727
 728 struct trace_wakeup_event {
 729         u32 size;
 730
 731         u16 common_type;
 732         u8 common_flags;
 733         u8 common_preempt_count;
 734         u32 common_pid;
 735         u32 common_tgid;
 736
 737         char comm[16];
 738         u32 pid;
 739
 740         u32 prio;
 741         u32 success;
 742         u32 cpu;
 743 };
 744
 745 struct trace_fork_event {
 746         u32 size;
 747
 748         u16 common_type;
 749         u8 common_flags;
 750         u8 common_preempt_count;
 751         u32 common_pid;
 752         u32 common_tgid;
 753
 754         char parent_comm[16];
 755         u32 parent_pid;
 756         char child_comm[16];
 757         u32 child_pid;
 758 };
 759
 760 struct trace_sched_handler {
 761         void (*switch_event)(struct trace_switch_event *,
 762                              struct event *,
 763                              int cpu,
 764                              u64 timestamp,
 765                              struct thread *thread);
 766
 767         void (*wakeup_event)(struct trace_wakeup_event *,
 768                              struct event *,
 769                              int cpu,
 770                              u64 timestamp,
 771                              struct thread *thread);
 772
 773         void (*fork_event)(struct trace_fork_event *,
 774                            struct event *,
 775                            int cpu,
 776                            u64 timestamp,
 777                            struct thread *thread);
 778 };
 779
 780
 781 static void
 782 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 783                     struct event *event,
 784                     int cpu __used,
 785                     u64 timestamp __used,
 786                     struct thread *thread __used)
 787 {
 788         struct task_desc *waker, *wakee;
 789
 790         if (verbose) {
 791                 printf("sched_wakeup event %p\n", event);
 792
 793                 printf(" ... pid %d woke up %s/%d\n",
 794                         wakeup_event->common_pid,
 795                         wakeup_event->comm,
 796                         wakeup_event->pid);
 797         }
 798
 799         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 800         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 801
 802         add_sched_event_wakeup(waker, timestamp, wakee);
 803 }
 804
 805 static unsigned long cpu_last_switched[MAX_CPUS];
 806
 807 static void
 808 replay_switch_event(struct trace_switch_event *switch_event,
 809                     struct event *event,
 810                     int cpu,
 811                     u64 timestamp,
 812                     struct thread *thread __used)
 813 {
 814         struct task_desc *prev, *next;
 815         u64 timestamp0;
 816         s64 delta;
 817
 818         if (verbose)
 819                 printf("sched_switch event %p\n", event);
 820
 821         if (cpu >= MAX_CPUS || cpu < 0)
 822                 return;
 823
 824         timestamp0 = cpu_last_switched[cpu];
 825         if (timestamp0)
 826                 delta = timestamp - timestamp0;
 827         else
 828                 delta = 0;
 829
 830         if (delta < 0)
 831                 die("hm, delta: %Ld < 0 ?\n", delta);
 832
 833         if (verbose) {
 834                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 835                         switch_event->prev_comm, switch_event->prev_pid,
 836                         switch_event->next_comm, switch_event->next_pid,
 837                         delta);
 838         }
 839
 840         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 841         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 842
 843         cpu_last_switched[cpu] = timestamp;
 844
 845         add_sched_event_run(prev, timestamp, delta);
 846         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 847 }
 848
 849
 850 static void
 851 replay_fork_event(struct trace_fork_event *fork_event,
 852                   struct event *event,
 853                   int cpu __used,
 854                   u64 timestamp __used,
 855                   struct thread *thread __used)
 856 {
 857         if (verbose) {
 858                 printf("sched_fork event %p\n", event);
 859                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 860                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 861         }
 862         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 863         register_pid(fork_event->child_pid, fork_event->child_comm);
 864 }
 865
 866 static struct trace_sched_handler replay_ops  = {
 867         .wakeup_event           = replay_wakeup_event,
 868         .switch_event           = replay_switch_event,
 869         .fork_event             = replay_fork_event,
 870 };
 871
 872 static struct task_atoms *
 873 thread_atoms_search(struct rb_root *root, struct thread *thread)
 874 {
 875         struct rb_node *node = root->rb_node;
 876
 877         while (node) {
 878                 struct task_atoms *atoms;
 879
 880                 atoms = container_of(node, struct task_atoms, node);
 881                 if (thread->pid > atoms->thread->pid)
 882                         node = node->rb_left;
 883                 else if (thread->pid < atoms->thread->pid)
 884                         node = node->rb_right;
 885                 else {
 886                         return atoms;
 887                 }
 888         }
 889         return NULL;
 890 }
 891
 892 struct sort_dimension {
 893         const char              *name;
 894         sort_thread_lat         cmp;
 895         struct list_head        list;
 896 };
 897
 898 static LIST_HEAD(cmp_pid);
 899
 900 static int
 901 thread_lat_cmp(struct list_head *list, struct task_atoms *l,
 902                struct task_atoms *r)
 903 {
 904         struct sort_dimension *sort;
 905         int ret = 0;
 906
 907         list_for_each_entry(sort, list, list) {
 908                 ret = sort->cmp(l, r);
 909                 if (ret)
 910                         return ret;
 911         }
 912
 913         return ret;
 914 }
 915
 916 static void
 917 __thread_latency_insert(struct rb_root *root, struct task_atoms *data,
 918                          struct list_head *sort_list)
 919 {
 920         struct rb_node **new = &(root->rb_node), *parent = NULL;
 921
 922         while (*new) {
 923                 struct task_atoms *this;
 924                 int cmp;
 925
 926                 this = container_of(*new, struct task_atoms, node);
 927                 parent = *new;
 928
 929                 cmp = thread_lat_cmp(sort_list, data, this);
 930
 931                 if (cmp > 0)
 932                         new = &((*new)->rb_left);
 933                 else
 934                         new = &((*new)->rb_right);
 935         }
 936
 937         rb_link_node(&data->node, parent, new);
 938         rb_insert_color(&data->node, root);
 939 }
 940
 941 static void thread_atoms_insert(struct thread *thread)
 942 {
 943         struct task_atoms *atoms;
 944
 945         atoms = calloc(sizeof(*atoms), 1);
 946         if (!atoms)
 947                 die("No memory");
 948
 949         atoms->thread = thread;
 950         INIT_LIST_HEAD(&atoms->atom_list);
 951         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 952 }
 953
 954 static void
 955 latency_fork_event(struct trace_fork_event *fork_event __used,
 956                    struct event *event __used,
 957                    int cpu __used,
 958                    u64 timestamp __used,
 959                    struct thread *thread __used)
 960 {
 961         /* should insert the newcomer */
 962 }
 963
 964 __used
 965 static char sched_out_state(struct trace_switch_event *switch_event)
 966 {
 967         const char *str = TASK_STATE_TO_CHAR_STR;
 968
 969         return str[switch_event->prev_state];
 970 }
 971
 972 static void
 973 lat_sched_out(struct task_atoms *atoms,
 974               struct trace_switch_event *switch_event __used,
 975               u64 delta,
 976               u64 timestamp)
 977 {
 978         struct work_atom *atom;
 979
 980         atom = calloc(sizeof(*atom), 1);
 981         if (!atom)
 982                 die("Non memory");
 983
 984         if (sched_out_state(switch_event) == 'R') {
 985                 atom->state = THREAD_WAIT_CPU;
 986                 atom->wake_up_time = timestamp;
 987         }
 988
 989         atom->runtime = delta;
 990         list_add_tail(&atom->list, &atoms->atom_list);
 991 }
 992
 993 static void
 994 lat_sched_in(struct task_atoms *atoms, u64 timestamp)
 995 {
 996         struct work_atom *atom;
 997         u64 delta;
 998
 999         if (list_empty(&atoms->atom_list))
1000                 return;
1001
1002         atom = list_entry(atoms->atom_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         atoms->total_runtime += atom->runtime;
1021 }
1022
1023 static void
1024 latency_switch_event(struct trace_switch_event *switch_event,
1025                      struct event *event __used,
1026                      int cpu,
1027                      u64 timestamp,
1028                      struct thread *thread __used)
1029 {
1030         struct task_atoms *out_atoms, *in_atoms;
1031         struct thread *sched_out, *sched_in;
1032         u64 timestamp0;
1033         s64 delta;
1034
1035         if (cpu >= MAX_CPUS || cpu < 0)
1036                 return;
1037
1038         timestamp0 = cpu_last_switched[cpu];
1039         cpu_last_switched[cpu] = timestamp;
1040         if (timestamp0)
1041                 delta = timestamp - timestamp0;
1042         else
1043                 delta = 0;
1044
1045         if (delta < 0)
1046                 die("hm, delta: %Ld < 0 ?\n", delta);
1047
1048
1049         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1050         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1051
1052         in_atoms = thread_atoms_search(&atom_root, sched_in);
1053         if (!in_atoms) {
1054                 thread_atoms_insert(sched_in);
1055                 in_atoms = thread_atoms_search(&atom_root, sched_in);
1056                 if (!in_atoms)
1057                         die("in-atom: Internal tree error");
1058         }
1059
1060         out_atoms = thread_atoms_search(&atom_root, sched_out);
1061         if (!out_atoms) {
1062                 thread_atoms_insert(sched_out);
1063                 out_atoms = thread_atoms_search(&atom_root, sched_out);
1064                 if (!out_atoms)
1065                         die("out-atom: Internal tree error");
1066         }
1067
1068         lat_sched_in(in_atoms, timestamp);
1069         lat_sched_out(out_atoms, switch_event, delta, timestamp);
1070 }
1071
1072 static void
1073 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1074                      struct event *event __used,
1075                      int cpu __used,
1076                      u64 timestamp,
1077                      struct thread *thread __used)
1078 {
1079         struct task_atoms *atoms;
1080         struct work_atom *atom;
1081         struct thread *wakee;
1082
1083         /* Note for later, it may be interesting to observe the failing cases */
1084         if (!wakeup_event->success)
1085                 return;
1086
1087         wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1088         atoms = thread_atoms_search(&atom_root, wakee);
1089         if (!atoms) {
1090                 thread_atoms_insert(wakee);
1091                 return;
1092         }
1093
1094         if (list_empty(&atoms->atom_list))
1095                 return;
1096
1097         atom = list_entry(atoms->atom_list.prev, struct work_atom, list);
1098
1099         if (atom->state != THREAD_SLEEPING)
1100                 return;
1101
1102         atom->state = THREAD_WAIT_CPU;
1103         atom->wake_up_time = timestamp;
1104 }
1105
1106 static struct trace_sched_handler lat_ops  = {
1107         .wakeup_event           = latency_wakeup_event,
1108         .switch_event           = latency_switch_event,
1109         .fork_event             = latency_fork_event,
1110 };
1111
1112 static void output_lat_thread(struct task_atoms *atom_list)
1113 {
1114         int i;
1115         int ret;
1116         u64 avg;
1117
1118         if (!atom_list->nb_atoms)
1119                 return;
1120
1121         all_runtime += atom_list->total_runtime;
1122         all_count += atom_list->nb_atoms;
1123
1124         ret = printf(" %s ", atom_list->thread->comm);
1125
1126         for (i = 0; i < 19 - ret; i++)
1127                 printf(" ");
1128
1129         avg = atom_list->total_lat / atom_list->nb_atoms;
1130
1131         printf("|%9.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1132               (double)atom_list->total_runtime / 1e6,
1133                  atom_list->nb_atoms, (double)avg / 1e6,
1134                  (double)atom_list->max_lat / 1e6);
1135 }
1136
1137 static int pid_cmp(struct task_atoms *l, struct task_atoms *r)
1138 {
1139
1140         if (l->thread->pid < r->thread->pid)
1141                 return -1;
1142         if (l->thread->pid > r->thread->pid)
1143                 return 1;
1144
1145         return 0;
1146 }
1147
1148 static struct sort_dimension pid_sort_dimension = {
1149         .name = "pid",
1150         .cmp = pid_cmp,
1151 };
1152
1153 static int avg_cmp(struct task_atoms *l, struct task_atoms *r)
1154 {
1155         u64 avgl, avgr;
1156
1157         if (!l->nb_atoms)
1158                 return -1;
1159
1160         if (!r->nb_atoms)
1161                 return 1;
1162
1163         avgl = l->total_lat / l->nb_atoms;
1164         avgr = r->total_lat / r->nb_atoms;
1165
1166         if (avgl < avgr)
1167                 return -1;
1168         if (avgl > avgr)
1169                 return 1;
1170
1171         return 0;
1172 }
1173
1174 static struct sort_dimension avg_sort_dimension = {
1175         .name   = "avg",
1176         .cmp    = avg_cmp,
1177 };
1178
1179 static int max_cmp(struct task_atoms *l, struct task_atoms *r)
1180 {
1181         if (l->max_lat < r->max_lat)
1182                 return -1;
1183         if (l->max_lat > r->max_lat)
1184                 return 1;
1185
1186         return 0;
1187 }
1188
1189 static struct sort_dimension max_sort_dimension = {
1190         .name   = "max",
1191         .cmp    = max_cmp,
1192 };
1193
1194 static int switch_cmp(struct task_atoms *l, struct task_atoms *r)
1195 {
1196         if (l->nb_atoms < r->nb_atoms)
1197                 return -1;
1198         if (l->nb_atoms > r->nb_atoms)
1199                 return 1;
1200
1201         return 0;
1202 }
1203
1204 static struct sort_dimension switch_sort_dimension = {
1205         .name   = "switch",
1206         .cmp    = switch_cmp,
1207 };
1208
1209 static int runtime_cmp(struct task_atoms *l, struct task_atoms *r)
1210 {
1211         if (l->total_runtime < r->total_runtime)
1212                 return -1;
1213         if (l->total_runtime > r->total_runtime)
1214                 return 1;
1215
1216         return 0;
1217 }
1218
1219 static struct sort_dimension runtime_sort_dimension = {
1220         .name   = "runtime",
1221         .cmp    = runtime_cmp,
1222 };
1223
1224 static struct sort_dimension *available_sorts[] = {
1225         &pid_sort_dimension,
1226         &avg_sort_dimension,
1227         &max_sort_dimension,
1228         &switch_sort_dimension,
1229         &runtime_sort_dimension,
1230 };
1231
1232 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1233
1234 static LIST_HEAD(sort_list);
1235
1236 static int sort_dimension__add(char *tok, struct list_head *list)
1237 {
1238         int i;
1239
1240         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1241                 if (!strcmp(available_sorts[i]->name, tok)) {
1242                         list_add_tail(&available_sorts[i]->list, list);
1243
1244                         return 0;
1245                 }
1246         }
1247
1248         return -1;
1249 }
1250
1251 static void setup_sorting(void);
1252
1253 static void sort_lat(void)
1254 {
1255         struct rb_node *node;
1256
1257         for (;;) {
1258                 struct task_atoms *data;
1259                 node = rb_first(&atom_root);
1260                 if (!node)
1261                         break;
1262
1263                 rb_erase(node, &atom_root);
1264                 data = rb_entry(node, struct task_atoms, node);
1265                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1266         }
1267 }
1268
1269 static void __cmd_lat(void)
1270 {
1271         struct rb_node *next;
1272
1273         setup_pager();
1274         read_events();
1275         sort_lat();
1276
1277         printf("-----------------------------------------------------------------------------------\n");
1278         printf(" Task              |  Runtime ms | Switches | Average delay ms | Maximum delay ms |\n");
1279         printf("-----------------------------------------------------------------------------------\n");
1280
1281         next = rb_first(&sorted_atom_root);
1282
1283         while (next) {
1284                 struct task_atoms *atom_list;
1285
1286                 atom_list = rb_entry(next, struct task_atoms, node);
1287                 output_lat_thread(atom_list);
1288                 next = rb_next(next);
1289         }
1290
1291         printf("-----------------------------------------------------------------------------------\n");
1292         printf(" TOTAL:            |%9.3f ms |%9Ld |\n",
1293                 (double)all_runtime/1e6, all_count);
1294         printf("---------------------------------------------\n");
1295 }
1296
1297 static struct trace_sched_handler *trace_handler;
1298
1299 static void
1300 process_sched_wakeup_event(struct raw_event_sample *raw,
1301                            struct event *event,
1302                            int cpu __used,
1303                            u64 timestamp __used,
1304                            struct thread *thread __used)
1305 {
1306         struct trace_wakeup_event wakeup_event;
1307
1308         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1309
1310         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1311         FILL_FIELD(wakeup_event, pid, event, raw->data);
1312         FILL_FIELD(wakeup_event, prio, event, raw->data);
1313         FILL_FIELD(wakeup_event, success, event, raw->data);
1314         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1315
1316         trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1317 }
1318
1319 static void
1320 process_sched_switch_event(struct raw_event_sample *raw,
1321                            struct event *event,
1322                            int cpu __used,
1323                            u64 timestamp __used,
1324                            struct thread *thread __used)
1325 {
1326         struct trace_switch_event switch_event;
1327
1328         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1329
1330         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1331         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1332         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1333         FILL_FIELD(switch_event, prev_state, event, raw->data);
1334         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1335         FILL_FIELD(switch_event, next_pid, event, raw->data);
1336         FILL_FIELD(switch_event, next_prio, event, raw->data);
1337
1338         trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
1339 }
1340
1341 static void
1342 process_sched_fork_event(struct raw_event_sample *raw,
1343                          struct event *event,
1344                          int cpu __used,
1345                          u64 timestamp __used,
1346                          struct thread *thread __used)
1347 {
1348         struct trace_fork_event fork_event;
1349
1350         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1351
1352         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1353         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1354         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1355         FILL_FIELD(fork_event, child_pid, event, raw->data);
1356
1357         trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1358 }
1359
1360 static void
1361 process_sched_exit_event(struct event *event,
1362                          int cpu __used,
1363                          u64 timestamp __used,
1364                          struct thread *thread __used)
1365 {
1366         if (verbose)
1367                 printf("sched_exit event %p\n", event);
1368 }
1369
1370 static void
1371 process_raw_event(event_t *raw_event __used, void *more_data,
1372                   int cpu, u64 timestamp, struct thread *thread)
1373 {
1374         struct raw_event_sample *raw = more_data;
1375         struct event *event;
1376         int type;
1377
1378         type = trace_parse_common_type(raw->data);
1379         event = trace_find_event(type);
1380
1381         if (!strcmp(event->name, "sched_switch"))
1382                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1383         if (!strcmp(event->name, "sched_wakeup"))
1384                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1385         if (!strcmp(event->name, "sched_wakeup_new"))
1386                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1387         if (!strcmp(event->name, "sched_process_fork"))
1388                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1389         if (!strcmp(event->name, "sched_process_exit"))
1390                 process_sched_exit_event(event, cpu, timestamp, thread);
1391 }
1392
1393 static int
1394 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1395 {
1396         char level;
1397         int show = 0;
1398         struct dso *dso = NULL;
1399         struct thread *thread;
1400         u64 ip = event->ip.ip;
1401         u64 timestamp = -1;
1402         u32 cpu = -1;
1403         u64 period = 1;
1404         void *more_data = event->ip.__more_data;
1405         int cpumode;
1406
1407         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1408
1409         if (sample_type & PERF_SAMPLE_TIME) {
1410                 timestamp = *(u64 *)more_data;
1411                 more_data += sizeof(u64);
1412         }
1413
1414         if (sample_type & PERF_SAMPLE_CPU) {
1415                 cpu = *(u32 *)more_data;
1416                 more_data += sizeof(u32);
1417                 more_data += sizeof(u32); /* reserved */
1418         }
1419
1420         if (sample_type & PERF_SAMPLE_PERIOD) {
1421                 period = *(u64 *)more_data;
1422                 more_data += sizeof(u64);
1423         }
1424
1425         dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1426                 (void *)(offset + head),
1427                 (void *)(long)(event->header.size),
1428                 event->header.misc,
1429                 event->ip.pid, event->ip.tid,
1430                 (void *)(long)ip,
1431                 (long long)period);
1432
1433         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1434
1435         if (thread == NULL) {
1436                 eprintf("problem processing %d event, skipping it.\n",
1437                         event->header.type);
1438                 return -1;
1439         }
1440
1441         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1442
1443         if (cpumode == PERF_EVENT_MISC_KERNEL) {
1444                 show = SHOW_KERNEL;
1445                 level = 'k';
1446
1447                 dso = kernel_dso;
1448
1449                 dump_printf(" ...... dso: %s\n", dso->name);
1450
1451         } else if (cpumode == PERF_EVENT_MISC_USER) {
1452
1453                 show = SHOW_USER;
1454                 level = '.';
1455
1456         } else {
1457                 show = SHOW_HV;
1458                 level = 'H';
1459
1460                 dso = hypervisor_dso;
1461
1462                 dump_printf(" ...... dso: [hypervisor]\n");
1463         }
1464
1465         if (sample_type & PERF_SAMPLE_RAW)
1466                 process_raw_event(event, more_data, cpu, timestamp, thread);
1467
1468         return 0;
1469 }
1470
1471 static int
1472 process_event(event_t *event, unsigned long offset, unsigned long head)
1473 {
1474         trace_event(event);
1475
1476         switch (event->header.type) {
1477         case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
1478                 return 0;
1479
1480         case PERF_EVENT_COMM:
1481                 return process_comm_event(event, offset, head);
1482
1483         case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1484                 return 0;
1485
1486         case PERF_EVENT_SAMPLE:
1487                 return process_sample_event(event, offset, head);
1488
1489         case PERF_EVENT_MAX:
1490         default:
1491                 return -1;
1492         }
1493
1494         return 0;
1495 }
1496
1497 static int read_events(void)
1498 {
1499         int ret, rc = EXIT_FAILURE;
1500         unsigned long offset = 0;
1501         unsigned long head = 0;
1502         struct stat perf_stat;
1503         event_t *event;
1504         uint32_t size;
1505         char *buf;
1506
1507         trace_report();
1508         register_idle_thread(&threads, &last_match);
1509
1510         input = open(input_name, O_RDONLY);
1511         if (input < 0) {
1512                 perror("failed to open file");
1513                 exit(-1);
1514         }
1515
1516         ret = fstat(input, &perf_stat);
1517         if (ret < 0) {
1518                 perror("failed to stat file");
1519                 exit(-1);
1520         }
1521
1522         if (!perf_stat.st_size) {
1523                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1524                 exit(0);
1525         }
1526         header = perf_header__read(input);
1527         head = header->data_offset;
1528         sample_type = perf_header__sample_type(header);
1529
1530         if (!(sample_type & PERF_SAMPLE_RAW))
1531                 die("No trace sample to read. Did you call perf record "
1532                     "without -R?");
1533
1534         if (load_kernel() < 0) {
1535                 perror("failed to load kernel symbols");
1536                 return EXIT_FAILURE;
1537         }
1538
1539 remap:
1540         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1541                            MAP_SHARED, input, offset);
1542         if (buf == MAP_FAILED) {
1543                 perror("failed to mmap file");
1544                 exit(-1);
1545         }
1546
1547 more:
1548         event = (event_t *)(buf + head);
1549
1550         size = event->header.size;
1551         if (!size)
1552                 size = 8;
1553
1554         if (head + event->header.size >= page_size * mmap_window) {
1555                 unsigned long shift = page_size * (head / page_size);
1556                 int res;
1557
1558                 res = munmap(buf, page_size * mmap_window);
1559                 assert(res == 0);
1560
1561                 offset += shift;
1562                 head -= shift;
1563                 goto remap;
1564         }
1565
1566         size = event->header.size;
1567
1568
1569         if (!size || process_event(event, offset, head) < 0) {
1570
1571                 /*
1572                  * assume we lost track of the stream, check alignment, and
1573                  * increment a single u64 in the hope to catch on again 'soon'.
1574                  */
1575
1576                 if (unlikely(head & 7))
1577                         head &= ~7ULL;
1578
1579                 size = 8;
1580         }
1581
1582         head += size;
1583
1584         if (offset + head < (unsigned long)perf_stat.st_size)
1585                 goto more;
1586
1587         rc = EXIT_SUCCESS;
1588         close(input);
1589
1590         return rc;
1591 }
1592
1593 static const char * const sched_usage[] = {
1594         "perf sched [<options>] {record|latency|replay}",
1595         NULL
1596 };
1597
1598 static const struct option sched_options[] = {
1599         OPT_BOOLEAN('v', "verbose", &verbose,
1600                     "be more verbose (show symbol address, etc)"),
1601         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1602                     "dump raw trace in ASCII"),
1603         OPT_END()
1604 };
1605
1606 static const char * const latency_usage[] = {
1607         "perf sched latency [<options>]",
1608         NULL
1609 };
1610
1611 static const struct option latency_options[] = {
1612         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1613                    "sort by key(s): runtime, switch, avg, max"),
1614         OPT_BOOLEAN('v', "verbose", &verbose,
1615                     "be more verbose (show symbol address, etc)"),
1616         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1617                     "dump raw trace in ASCII"),
1618         OPT_END()
1619 };
1620
1621 static const char * const replay_usage[] = {
1622         "perf sched replay [<options>]",
1623         NULL
1624 };
1625
1626 static const struct option replay_options[] = {
1627         OPT_INTEGER('r', "repeat", &replay_repeat,
1628                     "repeat the workload replay N times (-1: infinite)"),
1629         OPT_BOOLEAN('v', "verbose", &verbose,
1630                     "be more verbose (show symbol address, etc)"),
1631         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1632                     "dump raw trace in ASCII"),
1633         OPT_END()
1634 };
1635
1636 static void setup_sorting(void)
1637 {
1638         char *tmp, *tok, *str = strdup(sort_order);
1639
1640         for (tok = strtok_r(str, ", ", &tmp);
1641                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1642                 if (sort_dimension__add(tok, &sort_list) < 0) {
1643                         error("Unknown --sort key: `%s'", tok);
1644                         usage_with_options(latency_usage, latency_options);
1645                 }
1646         }
1647
1648         free(str);
1649
1650         sort_dimension__add((char *)"pid", &cmp_pid);
1651 }
1652
1653 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1654 {
1655         symbol__init();
1656         page_size = getpagesize();
1657
1658         argc = parse_options(argc, argv, sched_options, sched_usage,
1659                              PARSE_OPT_STOP_AT_NON_OPTION);
1660         if (!argc)
1661                 usage_with_options(sched_usage, sched_options);
1662
1663         if (!strncmp(argv[0], "lat", 3)) {
1664                 trace_handler = &lat_ops;
1665                 if (argc > 1) {
1666                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1667                         if (argc)
1668                                 usage_with_options(latency_usage, latency_options);
1669                         setup_sorting();
1670                 }
1671                 __cmd_lat();
1672         } else if (!strncmp(argv[0], "rep", 3)) {
1673                 trace_handler = &replay_ops;
1674                 if (argc) {
1675                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1676                         if (argc)
1677                                 usage_with_options(replay_usage, replay_options);
1678                 }
1679                 __cmd_replay();
1680         } else {
1681                 usage_with_options(sched_usage, sched_options);
1682         }
1683
1684
1685         return 0;
1686 }