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_atom;
  54
  55 struct task_desc {
  56         unsigned long           nr;
  57         unsigned long           pid;
  58         char                    comm[COMM_LEN];
  59
  60         unsigned long           nr_events;
  61         unsigned long           curr_event;
  62         struct sched_atom       **atoms;
  63
  64         pthread_t               thread;
  65         sem_t                   sleep_sem;
  66
  67         sem_t                   ready_for_work;
  68         sem_t                   work_done_sem;
  69
  70         u64                     cpu_usage;
  71 };
  72
  73 enum sched_event_type {
  74         SCHED_EVENT_RUN,
  75         SCHED_EVENT_SLEEP,
  76         SCHED_EVENT_WAKEUP,
  77 };
  78
  79 struct sched_atom {
  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 static unsigned long            nr_timestamps;
 120 static unsigned long            unordered_timestamps;
 121
 122 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 123
 124 enum thread_state {
 125         THREAD_SLEEPING = 0,
 126         THREAD_WAIT_CPU,
 127         THREAD_SCHED_IN,
 128         THREAD_IGNORE
 129 };
 130
 131 struct work_atom {
 132         struct list_head        list;
 133         enum thread_state       state;
 134         u64                     sched_out_time;
 135         u64                     wake_up_time;
 136         u64                     sched_in_time;
 137         u64                     runtime;
 138 };
 139
 140 struct work_atoms {
 141         struct list_head        work_list;
 142         struct thread           *thread;
 143         struct rb_node          node;
 144         u64                     max_lat;
 145         u64                     total_lat;
 146         u64                     nb_atoms;
 147         u64                     total_runtime;
 148 };
 149
 150 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 151
 152 static struct rb_root           atom_root, sorted_atom_root;
 153
 154 static u64                      all_runtime;
 155 static u64                      all_count;
 156
 157 static int read_events(void);
 158
 159
 160 static u64 get_nsecs(void)
 161 {
 162         struct timespec ts;
 163
 164         clock_gettime(CLOCK_MONOTONIC, &ts);
 165
 166         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 167 }
 168
 169 static void burn_nsecs(u64 nsecs)
 170 {
 171         u64 T0 = get_nsecs(), T1;
 172
 173         do {
 174                 T1 = get_nsecs();
 175         } while (T1 + run_measurement_overhead < T0 + nsecs);
 176 }
 177
 178 static void sleep_nsecs(u64 nsecs)
 179 {
 180         struct timespec ts;
 181
 182         ts.tv_nsec = nsecs % 999999999;
 183         ts.tv_sec = nsecs / 999999999;
 184
 185         nanosleep(&ts, NULL);
 186 }
 187
 188 static void calibrate_run_measurement_overhead(void)
 189 {
 190         u64 T0, T1, delta, min_delta = 1000000000ULL;
 191         int i;
 192
 193         for (i = 0; i < 10; i++) {
 194                 T0 = get_nsecs();
 195                 burn_nsecs(0);
 196                 T1 = get_nsecs();
 197                 delta = T1-T0;
 198                 min_delta = min(min_delta, delta);
 199         }
 200         run_measurement_overhead = min_delta;
 201
 202         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 203 }
 204
 205 static void calibrate_sleep_measurement_overhead(void)
 206 {
 207         u64 T0, T1, delta, min_delta = 1000000000ULL;
 208         int i;
 209
 210         for (i = 0; i < 10; i++) {
 211                 T0 = get_nsecs();
 212                 sleep_nsecs(10000);
 213                 T1 = get_nsecs();
 214                 delta = T1-T0;
 215                 min_delta = min(min_delta, delta);
 216         }
 217         min_delta -= 10000;
 218         sleep_measurement_overhead = min_delta;
 219
 220         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 221 }
 222
 223 static struct sched_atom *
 224 get_new_event(struct task_desc *task, u64 timestamp)
 225 {
 226         struct sched_atom *event = calloc(1, sizeof(*event));
 227         unsigned long idx = task->nr_events;
 228         size_t size;
 229
 230         event->timestamp = timestamp;
 231         event->nr = idx;
 232
 233         task->nr_events++;
 234         size = sizeof(struct sched_atom *) * task->nr_events;
 235         task->atoms = realloc(task->atoms, size);
 236         BUG_ON(!task->atoms);
 237
 238         task->atoms[idx] = event;
 239
 240         return event;
 241 }
 242
 243 static struct sched_atom *last_event(struct task_desc *task)
 244 {
 245         if (!task->nr_events)
 246                 return NULL;
 247
 248         return task->atoms[task->nr_events - 1];
 249 }
 250
 251 static void
 252 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 253 {
 254         struct sched_atom *event, *curr_event = last_event(task);
 255
 256         /*
 257          * optimize an existing RUN event by merging this one
 258          * to it:
 259          */
 260         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 261                 nr_run_events_optimized++;
 262                 curr_event->duration += duration;
 263                 return;
 264         }
 265
 266         event = get_new_event(task, timestamp);
 267
 268         event->type = SCHED_EVENT_RUN;
 269         event->duration = duration;
 270
 271         nr_run_events++;
 272 }
 273
 274 static void
 275 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 276                        struct task_desc *wakee)
 277 {
 278         struct sched_atom *event, *wakee_event;
 279
 280         event = get_new_event(task, timestamp);
 281         event->type = SCHED_EVENT_WAKEUP;
 282         event->wakee = wakee;
 283
 284         wakee_event = last_event(wakee);
 285         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 286                 targetless_wakeups++;
 287                 return;
 288         }
 289         if (wakee_event->wait_sem) {
 290                 multitarget_wakeups++;
 291                 return;
 292         }
 293
 294         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
 295         sem_init(wakee_event->wait_sem, 0, 0);
 296         wakee_event->specific_wait = 1;
 297         event->wait_sem = wakee_event->wait_sem;
 298
 299         nr_wakeup_events++;
 300 }
 301
 302 static void
 303 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 304                       u64 task_state __used)
 305 {
 306         struct sched_atom *event = get_new_event(task, timestamp);
 307
 308         event->type = SCHED_EVENT_SLEEP;
 309
 310         nr_sleep_events++;
 311 }
 312
 313 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 314 {
 315         struct task_desc *task;
 316
 317         BUG_ON(pid >= MAX_PID);
 318
 319         task = pid_to_task[pid];
 320
 321         if (task)
 322                 return task;
 323
 324         task = calloc(1, sizeof(*task));
 325         task->pid = pid;
 326         task->nr = nr_tasks;
 327         strcpy(task->comm, comm);
 328         /*
 329          * every task starts in sleeping state - this gets ignored
 330          * if there's no wakeup pointing to this sleep state:
 331          */
 332         add_sched_event_sleep(task, 0, 0);
 333
 334         pid_to_task[pid] = task;
 335         nr_tasks++;
 336         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 337         BUG_ON(!tasks);
 338         tasks[task->nr] = task;
 339
 340         if (verbose)
 341                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 342
 343         return task;
 344 }
 345
 346
 347 static void print_task_traces(void)
 348 {
 349         struct task_desc *task;
 350         unsigned long i;
 351
 352         for (i = 0; i < nr_tasks; i++) {
 353                 task = tasks[i];
 354                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 355                         task->nr, task->comm, task->pid, task->nr_events);
 356         }
 357 }
 358
 359 static void add_cross_task_wakeups(void)
 360 {
 361         struct task_desc *task1, *task2;
 362         unsigned long i, j;
 363
 364         for (i = 0; i < nr_tasks; i++) {
 365                 task1 = tasks[i];
 366                 j = i + 1;
 367                 if (j == nr_tasks)
 368                         j = 0;
 369                 task2 = tasks[j];
 370                 add_sched_event_wakeup(task1, 0, task2);
 371         }
 372 }
 373
 374 static void
 375 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 376 {
 377         int ret = 0;
 378         u64 now;
 379         long long delta;
 380
 381         now = get_nsecs();
 382         delta = start_time + atom->timestamp - now;
 383
 384         switch (atom->type) {
 385                 case SCHED_EVENT_RUN:
 386                         burn_nsecs(atom->duration);
 387                         break;
 388                 case SCHED_EVENT_SLEEP:
 389                         if (atom->wait_sem)
 390                                 ret = sem_wait(atom->wait_sem);
 391                         BUG_ON(ret);
 392                         break;
 393                 case SCHED_EVENT_WAKEUP:
 394                         if (atom->wait_sem)
 395                                 ret = sem_post(atom->wait_sem);
 396                         BUG_ON(ret);
 397                         break;
 398                 default:
 399                         BUG_ON(1);
 400         }
 401 }
 402
 403 static u64 get_cpu_usage_nsec_parent(void)
 404 {
 405         struct rusage ru;
 406         u64 sum;
 407         int err;
 408
 409         err = getrusage(RUSAGE_SELF, &ru);
 410         BUG_ON(err);
 411
 412         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 413         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 414
 415         return sum;
 416 }
 417
 418 static u64 get_cpu_usage_nsec_self(void)
 419 {
 420         char filename [] = "/proc/1234567890/sched";
 421         unsigned long msecs, nsecs;
 422         char *line = NULL;
 423         u64 total = 0;
 424         size_t len = 0;
 425         ssize_t chars;
 426         FILE *file;
 427         int ret;
 428
 429         sprintf(filename, "/proc/%d/sched", getpid());
 430         file = fopen(filename, "r");
 431         BUG_ON(!file);
 432
 433         while ((chars = getline(&line, &len, file)) != -1) {
 434                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 435                         &msecs, &nsecs);
 436                 if (ret == 2) {
 437                         total = msecs*1e6 + nsecs;
 438                         break;
 439                 }
 440         }
 441         if (line)
 442                 free(line);
 443         fclose(file);
 444
 445         return total;
 446 }
 447
 448 static void *thread_func(void *ctx)
 449 {
 450         struct task_desc *this_task = ctx;
 451         u64 cpu_usage_0, cpu_usage_1;
 452         unsigned long i, ret;
 453         char comm2[22];
 454
 455         sprintf(comm2, ":%s", this_task->comm);
 456         prctl(PR_SET_NAME, comm2);
 457
 458 again:
 459         ret = sem_post(&this_task->ready_for_work);
 460         BUG_ON(ret);
 461         ret = pthread_mutex_lock(&start_work_mutex);
 462         BUG_ON(ret);
 463         ret = pthread_mutex_unlock(&start_work_mutex);
 464         BUG_ON(ret);
 465
 466         cpu_usage_0 = get_cpu_usage_nsec_self();
 467
 468         for (i = 0; i < this_task->nr_events; i++) {
 469                 this_task->curr_event = i;
 470                 process_sched_event(this_task, this_task->atoms[i]);
 471         }
 472
 473         cpu_usage_1 = get_cpu_usage_nsec_self();
 474         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 475
 476         ret = sem_post(&this_task->work_done_sem);
 477         BUG_ON(ret);
 478
 479         ret = pthread_mutex_lock(&work_done_wait_mutex);
 480         BUG_ON(ret);
 481         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 482         BUG_ON(ret);
 483
 484         goto again;
 485 }
 486
 487 static void create_tasks(void)
 488 {
 489         struct task_desc *task;
 490         pthread_attr_t attr;
 491         unsigned long i;
 492         int err;
 493
 494         err = pthread_attr_init(&attr);
 495         BUG_ON(err);
 496         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 497         BUG_ON(err);
 498         err = pthread_mutex_lock(&start_work_mutex);
 499         BUG_ON(err);
 500         err = pthread_mutex_lock(&work_done_wait_mutex);
 501         BUG_ON(err);
 502         for (i = 0; i < nr_tasks; i++) {
 503                 task = tasks[i];
 504                 sem_init(&task->sleep_sem, 0, 0);
 505                 sem_init(&task->ready_for_work, 0, 0);
 506                 sem_init(&task->work_done_sem, 0, 0);
 507                 task->curr_event = 0;
 508                 err = pthread_create(&task->thread, &attr, thread_func, task);
 509                 BUG_ON(err);
 510         }
 511 }
 512
 513 static void wait_for_tasks(void)
 514 {
 515         u64 cpu_usage_0, cpu_usage_1;
 516         struct task_desc *task;
 517         unsigned long i, ret;
 518
 519         start_time = get_nsecs();
 520         cpu_usage = 0;
 521         pthread_mutex_unlock(&work_done_wait_mutex);
 522
 523         for (i = 0; i < nr_tasks; i++) {
 524                 task = tasks[i];
 525                 ret = sem_wait(&task->ready_for_work);
 526                 BUG_ON(ret);
 527                 sem_init(&task->ready_for_work, 0, 0);
 528         }
 529         ret = pthread_mutex_lock(&work_done_wait_mutex);
 530         BUG_ON(ret);
 531
 532         cpu_usage_0 = get_cpu_usage_nsec_parent();
 533
 534         pthread_mutex_unlock(&start_work_mutex);
 535
 536         for (i = 0; i < nr_tasks; i++) {
 537                 task = tasks[i];
 538                 ret = sem_wait(&task->work_done_sem);
 539                 BUG_ON(ret);
 540                 sem_init(&task->work_done_sem, 0, 0);
 541                 cpu_usage += task->cpu_usage;
 542                 task->cpu_usage = 0;
 543         }
 544
 545         cpu_usage_1 = get_cpu_usage_nsec_parent();
 546         if (!runavg_cpu_usage)
 547                 runavg_cpu_usage = cpu_usage;
 548         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 549
 550         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 551         if (!runavg_parent_cpu_usage)
 552                 runavg_parent_cpu_usage = parent_cpu_usage;
 553         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 554                                    parent_cpu_usage)/10;
 555
 556         ret = pthread_mutex_lock(&start_work_mutex);
 557         BUG_ON(ret);
 558
 559         for (i = 0; i < nr_tasks; i++) {
 560                 task = tasks[i];
 561                 sem_init(&task->sleep_sem, 0, 0);
 562                 task->curr_event = 0;
 563         }
 564 }
 565
 566 static void run_one_test(void)
 567 {
 568         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 569
 570         T0 = get_nsecs();
 571         wait_for_tasks();
 572         T1 = get_nsecs();
 573
 574         delta = T1 - T0;
 575         sum_runtime += delta;
 576         nr_runs++;
 577
 578         avg_delta = sum_runtime / nr_runs;
 579         if (delta < avg_delta)
 580                 fluct = avg_delta - delta;
 581         else
 582                 fluct = delta - avg_delta;
 583         sum_fluct += fluct;
 584         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 585         if (!run_avg)
 586                 run_avg = delta;
 587         run_avg = (run_avg*9 + delta)/10;
 588
 589         printf("#%-3ld: %0.3f, ",
 590                 nr_runs, (double)delta/1000000.0);
 591
 592         printf("ravg: %0.2f, ",
 593                 (double)run_avg/1e6);
 594
 595         printf("cpu: %0.2f / %0.2f",
 596                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 597
 598 #if 0
 599         /*
 600          * rusage statistics done by the parent, these are less
 601          * accurate than the sum_exec_runtime based statistics:
 602          */
 603         printf(" [%0.2f / %0.2f]",
 604                 (double)parent_cpu_usage/1e6,
 605                 (double)runavg_parent_cpu_usage/1e6);
 606 #endif
 607
 608         printf("\n");
 609
 610         if (nr_sleep_corrections)
 611                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 612         nr_sleep_corrections = 0;
 613 }
 614
 615 static void test_calibrations(void)
 616 {
 617         u64 T0, T1;
 618
 619         T0 = get_nsecs();
 620         burn_nsecs(1e6);
 621         T1 = get_nsecs();
 622
 623         printf("the run test took %Ld nsecs\n", T1-T0);
 624
 625         T0 = get_nsecs();
 626         sleep_nsecs(1e6);
 627         T1 = get_nsecs();
 628
 629         printf("the sleep test took %Ld nsecs\n", T1-T0);
 630 }
 631
 632 static void __cmd_replay(void)
 633 {
 634         unsigned long i;
 635
 636         calibrate_run_measurement_overhead();
 637         calibrate_sleep_measurement_overhead();
 638
 639         test_calibrations();
 640
 641         read_events();
 642
 643         printf("nr_run_events:        %ld\n", nr_run_events);
 644         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
 645         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
 646
 647         if (targetless_wakeups)
 648                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
 649         if (multitarget_wakeups)
 650                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
 651         if (nr_run_events_optimized)
 652                 printf("run atoms optimized: %ld\n",
 653                         nr_run_events_optimized);
 654
 655         print_task_traces();
 656         add_cross_task_wakeups();
 657
 658         create_tasks();
 659         printf("------------------------------------------------------------\n");
 660         for (i = 0; i < replay_repeat; i++)
 661                 run_one_test();
 662 }
 663
 664 static int
 665 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 666 {
 667         struct thread *thread;
 668
 669         thread = threads__findnew(event->comm.pid, &threads, &last_match);
 670
 671         dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
 672                 (void *)(offset + head),
 673                 (void *)(long)(event->header.size),
 674                 event->comm.comm, event->comm.pid);
 675
 676         if (thread == NULL ||
 677             thread__set_comm(thread, event->comm.comm)) {
 678                 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
 679                 return -1;
 680         }
 681         total_comm++;
 682
 683         return 0;
 684 }
 685
 686
 687 struct raw_event_sample {
 688         u32 size;
 689         char data[0];
 690 };
 691
 692 #define FILL_FIELD(ptr, field, event, data)     \
 693         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 694
 695 #define FILL_ARRAY(ptr, array, event, data)                     \
 696 do {                                                            \
 697         void *__array = raw_field_ptr(event, #array, data);     \
 698         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 699 } while(0)
 700
 701 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 702 do {                                                            \
 703         FILL_FIELD(ptr, common_type, event, data);              \
 704         FILL_FIELD(ptr, common_flags, event, data);             \
 705         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 706         FILL_FIELD(ptr, common_pid, event, data);               \
 707         FILL_FIELD(ptr, common_tgid, event, data);              \
 708 } while (0)
 709
 710
 711
 712 struct trace_switch_event {
 713         u32 size;
 714
 715         u16 common_type;
 716         u8 common_flags;
 717         u8 common_preempt_count;
 718         u32 common_pid;
 719         u32 common_tgid;
 720
 721         char prev_comm[16];
 722         u32 prev_pid;
 723         u32 prev_prio;
 724         u64 prev_state;
 725         char next_comm[16];
 726         u32 next_pid;
 727         u32 next_prio;
 728 };
 729
 730 struct trace_runtime_event {
 731         u32 size;
 732
 733         u16 common_type;
 734         u8 common_flags;
 735         u8 common_preempt_count;
 736         u32 common_pid;
 737         u32 common_tgid;
 738
 739         char comm[16];
 740         u32 pid;
 741         u64 runtime;
 742         u64 vruntime;
 743 };
 744
 745 struct trace_wakeup_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 comm[16];
 755         u32 pid;
 756
 757         u32 prio;
 758         u32 success;
 759         u32 cpu;
 760 };
 761
 762 struct trace_fork_event {
 763         u32 size;
 764
 765         u16 common_type;
 766         u8 common_flags;
 767         u8 common_preempt_count;
 768         u32 common_pid;
 769         u32 common_tgid;
 770
 771         char parent_comm[16];
 772         u32 parent_pid;
 773         char child_comm[16];
 774         u32 child_pid;
 775 };
 776
 777 struct trace_sched_handler {
 778         void (*switch_event)(struct trace_switch_event *,
 779                              struct event *,
 780                              int cpu,
 781                              u64 timestamp,
 782                              struct thread *thread);
 783
 784         void (*runtime_event)(struct trace_runtime_event *,
 785                               struct event *,
 786                               int cpu,
 787                               u64 timestamp,
 788                               struct thread *thread);
 789
 790         void (*wakeup_event)(struct trace_wakeup_event *,
 791                              struct event *,
 792                              int cpu,
 793                              u64 timestamp,
 794                              struct thread *thread);
 795
 796         void (*fork_event)(struct trace_fork_event *,
 797                            struct event *,
 798                            int cpu,
 799                            u64 timestamp,
 800                            struct thread *thread);
 801 };
 802
 803
 804 static void
 805 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 806                     struct event *event,
 807                     int cpu __used,
 808                     u64 timestamp __used,
 809                     struct thread *thread __used)
 810 {
 811         struct task_desc *waker, *wakee;
 812
 813         if (verbose) {
 814                 printf("sched_wakeup event %p\n", event);
 815
 816                 printf(" ... pid %d woke up %s/%d\n",
 817                         wakeup_event->common_pid,
 818                         wakeup_event->comm,
 819                         wakeup_event->pid);
 820         }
 821
 822         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 823         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 824
 825         add_sched_event_wakeup(waker, timestamp, wakee);
 826 }
 827
 828 static u64 cpu_last_switched[MAX_CPUS];
 829
 830 static void
 831 replay_switch_event(struct trace_switch_event *switch_event,
 832                     struct event *event,
 833                     int cpu,
 834                     u64 timestamp,
 835                     struct thread *thread __used)
 836 {
 837         struct task_desc *prev, *next;
 838         u64 timestamp0;
 839         s64 delta;
 840
 841         if (verbose)
 842                 printf("sched_switch event %p\n", event);
 843
 844         if (cpu >= MAX_CPUS || cpu < 0)
 845                 return;
 846
 847         timestamp0 = cpu_last_switched[cpu];
 848         if (timestamp0)
 849                 delta = timestamp - timestamp0;
 850         else
 851                 delta = 0;
 852
 853         if (delta < 0)
 854                 die("hm, delta: %Ld < 0 ?\n", delta);
 855
 856         if (verbose) {
 857                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 858                         switch_event->prev_comm, switch_event->prev_pid,
 859                         switch_event->next_comm, switch_event->next_pid,
 860                         delta);
 861         }
 862
 863         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 864         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 865
 866         cpu_last_switched[cpu] = timestamp;
 867
 868         add_sched_event_run(prev, timestamp, delta);
 869         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 870 }
 871
 872
 873 static void
 874 replay_fork_event(struct trace_fork_event *fork_event,
 875                   struct event *event,
 876                   int cpu __used,
 877                   u64 timestamp __used,
 878                   struct thread *thread __used)
 879 {
 880         if (verbose) {
 881                 printf("sched_fork event %p\n", event);
 882                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 883                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 884         }
 885         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 886         register_pid(fork_event->child_pid, fork_event->child_comm);
 887 }
 888
 889 static struct trace_sched_handler replay_ops  = {
 890         .wakeup_event           = replay_wakeup_event,
 891         .switch_event           = replay_switch_event,
 892         .fork_event             = replay_fork_event,
 893 };
 894
 895 struct sort_dimension {
 896         const char              *name;
 897         sort_fn_t               cmp;
 898         struct list_head        list;
 899 };
 900
 901 static LIST_HEAD(cmp_pid);
 902
 903 static int
 904 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 905 {
 906         struct sort_dimension *sort;
 907         int ret = 0;
 908
 909         BUG_ON(list_empty(list));
 910
 911         list_for_each_entry(sort, list, list) {
 912                 ret = sort->cmp(l, r);
 913                 if (ret)
 914                         return ret;
 915         }
 916
 917         return ret;
 918 }
 919
 920 static struct work_atoms *
 921 thread_atoms_search(struct rb_root *root, struct thread *thread,
 922                          struct list_head *sort_list)
 923 {
 924         struct rb_node *node = root->rb_node;
 925         struct work_atoms key = { .thread = thread };
 926
 927         while (node) {
 928                 struct work_atoms *atoms;
 929                 int cmp;
 930
 931                 atoms = container_of(node, struct work_atoms, node);
 932
 933                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 934                 if (cmp > 0)
 935                         node = node->rb_left;
 936                 else if (cmp < 0)
 937                         node = node->rb_right;
 938                 else {
 939                         BUG_ON(thread != atoms->thread);
 940                         return atoms;
 941                 }
 942         }
 943         return NULL;
 944 }
 945
 946 static void
 947 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 948                          struct list_head *sort_list)
 949 {
 950         struct rb_node **new = &(root->rb_node), *parent = NULL;
 951
 952         while (*new) {
 953                 struct work_atoms *this;
 954                 int cmp;
 955
 956                 this = container_of(*new, struct work_atoms, node);
 957                 parent = *new;
 958
 959                 cmp = thread_lat_cmp(sort_list, data, this);
 960
 961                 if (cmp > 0)
 962                         new = &((*new)->rb_left);
 963                 else
 964                         new = &((*new)->rb_right);
 965         }
 966
 967         rb_link_node(&data->node, parent, new);
 968         rb_insert_color(&data->node, root);
 969 }
 970
 971 static void thread_atoms_insert(struct thread *thread)
 972 {
 973         struct work_atoms *atoms;
 974
 975         atoms = calloc(sizeof(*atoms), 1);
 976         if (!atoms)
 977                 die("No memory");
 978
 979         atoms->thread = thread;
 980         INIT_LIST_HEAD(&atoms->work_list);
 981         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 982 }
 983
 984 static void
 985 latency_fork_event(struct trace_fork_event *fork_event __used,
 986                    struct event *event __used,
 987                    int cpu __used,
 988                    u64 timestamp __used,
 989                    struct thread *thread __used)
 990 {
 991         /* should insert the newcomer */
 992 }
 993
 994 __used
 995 static char sched_out_state(struct trace_switch_event *switch_event)
 996 {
 997         const char *str = TASK_STATE_TO_CHAR_STR;
 998
 999         return str[switch_event->prev_state];
1000 }
1001
1002 static void
1003 add_sched_out_event(struct work_atoms *atoms,
1004                     char run_state,
1005                     u64 timestamp)
1006 {
1007         struct work_atom *atom;
1008
1009         atom = calloc(sizeof(*atom), 1);
1010         if (!atom)
1011                 die("Non memory");
1012
1013         atom->sched_out_time = timestamp;
1014
1015         if (run_state == 'R') {
1016                 atom->state = THREAD_WAIT_CPU;
1017                 atom->wake_up_time = atom->sched_out_time;
1018         }
1019
1020         list_add_tail(&atom->list, &atoms->work_list);
1021 }
1022
1023 static void
1024 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
1025 {
1026         struct work_atom *atom;
1027
1028         BUG_ON(list_empty(&atoms->work_list));
1029
1030         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1031
1032         atom->runtime += delta;
1033         atoms->total_runtime += delta;
1034 }
1035
1036 static void
1037 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1038 {
1039         struct work_atom *atom;
1040         u64 delta;
1041
1042         if (list_empty(&atoms->work_list))
1043                 return;
1044
1045         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1046
1047         if (atom->state != THREAD_WAIT_CPU)
1048                 return;
1049
1050         if (timestamp < atom->wake_up_time) {
1051                 atom->state = THREAD_IGNORE;
1052                 return;
1053         }
1054
1055         atom->state = THREAD_SCHED_IN;
1056         atom->sched_in_time = timestamp;
1057
1058         delta = atom->sched_in_time - atom->wake_up_time;
1059         atoms->total_lat += delta;
1060         if (delta > atoms->max_lat)
1061                 atoms->max_lat = delta;
1062         atoms->nb_atoms++;
1063 }
1064
1065 static void
1066 latency_switch_event(struct trace_switch_event *switch_event,
1067                      struct event *event __used,
1068                      int cpu,
1069                      u64 timestamp,
1070                      struct thread *thread __used)
1071 {
1072         struct work_atoms *out_events, *in_events;
1073         struct thread *sched_out, *sched_in;
1074         u64 timestamp0;
1075         s64 delta;
1076
1077         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1078
1079         timestamp0 = cpu_last_switched[cpu];
1080         cpu_last_switched[cpu] = timestamp;
1081         if (timestamp0)
1082                 delta = timestamp - timestamp0;
1083         else
1084                 delta = 0;
1085
1086         if (delta < 0)
1087                 die("hm, delta: %Ld < 0 ?\n", delta);
1088
1089
1090         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1091         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1092
1093         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1094         if (!out_events) {
1095                 thread_atoms_insert(sched_out);
1096                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1097                 if (!out_events)
1098                         die("out-event: Internal tree error");
1099         }
1100         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1101
1102         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1103         if (!in_events) {
1104                 thread_atoms_insert(sched_in);
1105                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1106                 if (!in_events)
1107                         die("in-event: Internal tree error");
1108                 /*
1109                  * Take came in we have not heard about yet,
1110                  * add in an initial atom in runnable state:
1111                  */
1112                 add_sched_out_event(in_events, 'R', timestamp);
1113         }
1114         add_sched_in_event(in_events, timestamp);
1115 }
1116
1117 static void
1118 latency_runtime_event(struct trace_runtime_event *runtime_event,
1119                      struct event *event __used,
1120                      int cpu,
1121                      u64 timestamp,
1122                      struct thread *this_thread __used)
1123 {
1124         struct work_atoms *atoms;
1125         struct thread *thread;
1126
1127         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1128
1129         thread = threads__findnew(runtime_event->pid, &threads, &last_match);
1130         atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1131         if (!atoms) {
1132                 thread_atoms_insert(thread);
1133                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1134                 if (!atoms)
1135                         die("in-event: Internal tree error");
1136                 add_sched_out_event(atoms, 'R', timestamp);
1137         }
1138
1139         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1140 }
1141
1142 static void
1143 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1144                      struct event *__event __used,
1145                      int cpu __used,
1146                      u64 timestamp,
1147                      struct thread *thread __used)
1148 {
1149         struct work_atoms *atoms;
1150         struct work_atom *atom;
1151         struct thread *wakee;
1152
1153         /* Note for later, it may be interesting to observe the failing cases */
1154         if (!wakeup_event->success)
1155                 return;
1156
1157         wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1158         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1159         if (!atoms) {
1160                 thread_atoms_insert(wakee);
1161                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1162                 if (!atoms)
1163                         die("wakeup-event: Internal tree error");
1164                 add_sched_out_event(atoms, 'S', timestamp);
1165         }
1166
1167         BUG_ON(list_empty(&atoms->work_list));
1168
1169         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1170
1171         if (atom->state != THREAD_SLEEPING) {
1172                 printf("boo2\n");
1173                 return;
1174         }
1175
1176         nr_timestamps++;
1177         if (atom->sched_out_time > timestamp) {
1178                 unordered_timestamps++;
1179                 return;
1180         }
1181
1182         atom->state = THREAD_WAIT_CPU;
1183         atom->wake_up_time = timestamp;
1184 }
1185
1186 static struct trace_sched_handler lat_ops  = {
1187         .wakeup_event           = latency_wakeup_event,
1188         .switch_event           = latency_switch_event,
1189         .runtime_event          = latency_runtime_event,
1190         .fork_event             = latency_fork_event,
1191 };
1192
1193 static void output_lat_thread(struct work_atoms *work_list)
1194 {
1195         int i;
1196         int ret;
1197         u64 avg;
1198
1199         if (!work_list->nb_atoms)
1200                 return;
1201         /*
1202          * Ignore idle threads:
1203          */
1204         if (!work_list->thread->pid)
1205                 return;
1206
1207         all_runtime += work_list->total_runtime;
1208         all_count += work_list->nb_atoms;
1209
1210         ret = printf("  %s-%d ", work_list->thread->comm, work_list->thread->pid);
1211
1212         for (i = 0; i < 24 - ret; i++)
1213                 printf(" ");
1214
1215         avg = work_list->total_lat / work_list->nb_atoms;
1216
1217         printf("|%9.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1218               (double)work_list->total_runtime / 1e6,
1219                  work_list->nb_atoms, (double)avg / 1e6,
1220                  (double)work_list->max_lat / 1e6);
1221 }
1222
1223 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1224 {
1225         if (l->thread->pid < r->thread->pid)
1226                 return -1;
1227         if (l->thread->pid > r->thread->pid)
1228                 return 1;
1229
1230         return 0;
1231 }
1232
1233 static struct sort_dimension pid_sort_dimension = {
1234         .name                   = "pid",
1235         .cmp                    = pid_cmp,
1236 };
1237
1238 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1239 {
1240         u64 avgl, avgr;
1241
1242         if (!l->nb_atoms)
1243                 return -1;
1244
1245         if (!r->nb_atoms)
1246                 return 1;
1247
1248         avgl = l->total_lat / l->nb_atoms;
1249         avgr = r->total_lat / r->nb_atoms;
1250
1251         if (avgl < avgr)
1252                 return -1;
1253         if (avgl > avgr)
1254                 return 1;
1255
1256         return 0;
1257 }
1258
1259 static struct sort_dimension avg_sort_dimension = {
1260         .name                   = "avg",
1261         .cmp                    = avg_cmp,
1262 };
1263
1264 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1265 {
1266         if (l->max_lat < r->max_lat)
1267                 return -1;
1268         if (l->max_lat > r->max_lat)
1269                 return 1;
1270
1271         return 0;
1272 }
1273
1274 static struct sort_dimension max_sort_dimension = {
1275         .name                   = "max",
1276         .cmp                    = max_cmp,
1277 };
1278
1279 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1280 {
1281         if (l->nb_atoms < r->nb_atoms)
1282                 return -1;
1283         if (l->nb_atoms > r->nb_atoms)
1284                 return 1;
1285
1286         return 0;
1287 }
1288
1289 static struct sort_dimension switch_sort_dimension = {
1290         .name                   = "switch",
1291         .cmp                    = switch_cmp,
1292 };
1293
1294 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1295 {
1296         if (l->total_runtime < r->total_runtime)
1297                 return -1;
1298         if (l->total_runtime > r->total_runtime)
1299                 return 1;
1300
1301         return 0;
1302 }
1303
1304 static struct sort_dimension runtime_sort_dimension = {
1305         .name                   = "runtime",
1306         .cmp                    = runtime_cmp,
1307 };
1308
1309 static struct sort_dimension *available_sorts[] = {
1310         &pid_sort_dimension,
1311         &avg_sort_dimension,
1312         &max_sort_dimension,
1313         &switch_sort_dimension,
1314         &runtime_sort_dimension,
1315 };
1316
1317 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1318
1319 static LIST_HEAD(sort_list);
1320
1321 static int sort_dimension__add(char *tok, struct list_head *list)
1322 {
1323         int i;
1324
1325         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1326                 if (!strcmp(available_sorts[i]->name, tok)) {
1327                         list_add_tail(&available_sorts[i]->list, list);
1328
1329                         return 0;
1330                 }
1331         }
1332
1333         return -1;
1334 }
1335
1336 static void setup_sorting(void);
1337
1338 static void sort_lat(void)
1339 {
1340         struct rb_node *node;
1341
1342         for (;;) {
1343                 struct work_atoms *data;
1344                 node = rb_first(&atom_root);
1345                 if (!node)
1346                         break;
1347
1348                 rb_erase(node, &atom_root);
1349                 data = rb_entry(node, struct work_atoms, node);
1350                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1351         }
1352 }
1353
1354 static void __cmd_lat(void)
1355 {
1356         struct rb_node *next;
1357
1358         setup_pager();
1359         read_events();
1360         sort_lat();
1361
1362         printf("\n ---------------------------------------------------------------------------------------\n");
1363         printf("  Task                  |  Runtime ms | Switches | Average delay ms | Maximum delay ms |\n");
1364         printf(" ---------------------------------------------------------------------------------------\n");
1365
1366         next = rb_first(&sorted_atom_root);
1367
1368         while (next) {
1369                 struct work_atoms *work_list;
1370
1371                 work_list = rb_entry(next, struct work_atoms, node);
1372                 output_lat_thread(work_list);
1373                 next = rb_next(next);
1374         }
1375
1376         printf(" ---------------------------------------------------------------------------------------\n");
1377         printf("  TOTAL:                |%9.3f ms |%9Ld |",
1378                 (double)all_runtime/1e6, all_count);
1379
1380         if (unordered_timestamps && nr_timestamps) {
1381                 printf(" INFO: %.2f%% unordered events.\n",
1382                         (double)unordered_timestamps/(double)nr_timestamps*100.0);
1383         } else {
1384                 printf("\n");
1385         }
1386
1387         printf(" -------------------------------------------------\n\n");
1388 }
1389
1390 static struct trace_sched_handler *trace_handler;
1391
1392 static void
1393 process_sched_wakeup_event(struct raw_event_sample *raw,
1394                            struct event *event,
1395                            int cpu __used,
1396                            u64 timestamp __used,
1397                            struct thread *thread __used)
1398 {
1399         struct trace_wakeup_event wakeup_event;
1400
1401         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1402
1403         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1404         FILL_FIELD(wakeup_event, pid, event, raw->data);
1405         FILL_FIELD(wakeup_event, prio, event, raw->data);
1406         FILL_FIELD(wakeup_event, success, event, raw->data);
1407         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1408
1409         trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1410 }
1411
1412 static void
1413 process_sched_switch_event(struct raw_event_sample *raw,
1414                            struct event *event,
1415                            int cpu __used,
1416                            u64 timestamp __used,
1417                            struct thread *thread __used)
1418 {
1419         struct trace_switch_event switch_event;
1420
1421         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1422
1423         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1424         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1425         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1426         FILL_FIELD(switch_event, prev_state, event, raw->data);
1427         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1428         FILL_FIELD(switch_event, next_pid, event, raw->data);
1429         FILL_FIELD(switch_event, next_prio, event, raw->data);
1430
1431         trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
1432 }
1433
1434 static void
1435 process_sched_runtime_event(struct raw_event_sample *raw,
1436                            struct event *event,
1437                            int cpu __used,
1438                            u64 timestamp __used,
1439                            struct thread *thread __used)
1440 {
1441         struct trace_runtime_event runtime_event;
1442
1443         FILL_ARRAY(runtime_event, comm, event, raw->data);
1444         FILL_FIELD(runtime_event, pid, event, raw->data);
1445         FILL_FIELD(runtime_event, runtime, event, raw->data);
1446         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1447
1448         trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1449 }
1450
1451 static void
1452 process_sched_fork_event(struct raw_event_sample *raw,
1453                          struct event *event,
1454                          int cpu __used,
1455                          u64 timestamp __used,
1456                          struct thread *thread __used)
1457 {
1458         struct trace_fork_event fork_event;
1459
1460         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1461
1462         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1463         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1464         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1465         FILL_FIELD(fork_event, child_pid, event, raw->data);
1466
1467         trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1468 }
1469
1470 static void
1471 process_sched_exit_event(struct event *event,
1472                          int cpu __used,
1473                          u64 timestamp __used,
1474                          struct thread *thread __used)
1475 {
1476         if (verbose)
1477                 printf("sched_exit event %p\n", event);
1478 }
1479
1480 static void
1481 process_raw_event(event_t *raw_event __used, void *more_data,
1482                   int cpu, u64 timestamp, struct thread *thread)
1483 {
1484         struct raw_event_sample *raw = more_data;
1485         struct event *event;
1486         int type;
1487
1488         type = trace_parse_common_type(raw->data);
1489         event = trace_find_event(type);
1490
1491         if (!strcmp(event->name, "sched_switch"))
1492                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1493         if (!strcmp(event->name, "sched_stat_runtime"))
1494                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1495         if (!strcmp(event->name, "sched_wakeup"))
1496                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1497         if (!strcmp(event->name, "sched_wakeup_new"))
1498                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1499         if (!strcmp(event->name, "sched_process_fork"))
1500                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1501         if (!strcmp(event->name, "sched_process_exit"))
1502                 process_sched_exit_event(event, cpu, timestamp, thread);
1503 }
1504
1505 static int
1506 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1507 {
1508         char level;
1509         int show = 0;
1510         struct dso *dso = NULL;
1511         struct thread *thread;
1512         u64 ip = event->ip.ip;
1513         u64 timestamp = -1;
1514         u32 cpu = -1;
1515         u64 period = 1;
1516         void *more_data = event->ip.__more_data;
1517         int cpumode;
1518
1519         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1520
1521         if (sample_type & PERF_SAMPLE_TIME) {
1522                 timestamp = *(u64 *)more_data;
1523                 more_data += sizeof(u64);
1524         }
1525
1526         if (sample_type & PERF_SAMPLE_CPU) {
1527                 cpu = *(u32 *)more_data;
1528                 more_data += sizeof(u32);
1529                 more_data += sizeof(u32); /* reserved */
1530         }
1531
1532         if (sample_type & PERF_SAMPLE_PERIOD) {
1533                 period = *(u64 *)more_data;
1534                 more_data += sizeof(u64);
1535         }
1536
1537         dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1538                 (void *)(offset + head),
1539                 (void *)(long)(event->header.size),
1540                 event->header.misc,
1541                 event->ip.pid, event->ip.tid,
1542                 (void *)(long)ip,
1543                 (long long)period);
1544
1545         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1546
1547         if (thread == NULL) {
1548                 eprintf("problem processing %d event, skipping it.\n",
1549                         event->header.type);
1550                 return -1;
1551         }
1552
1553         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1554
1555         if (cpumode == PERF_EVENT_MISC_KERNEL) {
1556                 show = SHOW_KERNEL;
1557                 level = 'k';
1558
1559                 dso = kernel_dso;
1560
1561                 dump_printf(" ...... dso: %s\n", dso->name);
1562
1563         } else if (cpumode == PERF_EVENT_MISC_USER) {
1564
1565                 show = SHOW_USER;
1566                 level = '.';
1567
1568         } else {
1569                 show = SHOW_HV;
1570                 level = 'H';
1571
1572                 dso = hypervisor_dso;
1573
1574                 dump_printf(" ...... dso: [hypervisor]\n");
1575         }
1576
1577         if (sample_type & PERF_SAMPLE_RAW)
1578                 process_raw_event(event, more_data, cpu, timestamp, thread);
1579
1580         return 0;
1581 }
1582
1583 static int
1584 process_event(event_t *event, unsigned long offset, unsigned long head)
1585 {
1586         trace_event(event);
1587
1588         switch (event->header.type) {
1589         case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
1590                 return 0;
1591
1592         case PERF_EVENT_COMM:
1593                 return process_comm_event(event, offset, head);
1594
1595         case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1596                 return 0;
1597
1598         case PERF_EVENT_SAMPLE:
1599                 return process_sample_event(event, offset, head);
1600
1601         case PERF_EVENT_MAX:
1602         default:
1603                 return -1;
1604         }
1605
1606         return 0;
1607 }
1608
1609 static int read_events(void)
1610 {
1611         int ret, rc = EXIT_FAILURE;
1612         unsigned long offset = 0;
1613         unsigned long head = 0;
1614         struct stat perf_stat;
1615         event_t *event;
1616         uint32_t size;
1617         char *buf;
1618
1619         trace_report();
1620         register_idle_thread(&threads, &last_match);
1621
1622         input = open(input_name, O_RDONLY);
1623         if (input < 0) {
1624                 perror("failed to open file");
1625                 exit(-1);
1626         }
1627
1628         ret = fstat(input, &perf_stat);
1629         if (ret < 0) {
1630                 perror("failed to stat file");
1631                 exit(-1);
1632         }
1633
1634         if (!perf_stat.st_size) {
1635                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1636                 exit(0);
1637         }
1638         header = perf_header__read(input);
1639         head = header->data_offset;
1640         sample_type = perf_header__sample_type(header);
1641
1642         if (!(sample_type & PERF_SAMPLE_RAW))
1643                 die("No trace sample to read. Did you call perf record "
1644                     "without -R?");
1645
1646         if (load_kernel() < 0) {
1647                 perror("failed to load kernel symbols");
1648                 return EXIT_FAILURE;
1649         }
1650
1651 remap:
1652         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1653                            MAP_SHARED, input, offset);
1654         if (buf == MAP_FAILED) {
1655                 perror("failed to mmap file");
1656                 exit(-1);
1657         }
1658
1659 more:
1660         event = (event_t *)(buf + head);
1661
1662         size = event->header.size;
1663         if (!size)
1664                 size = 8;
1665
1666         if (head + event->header.size >= page_size * mmap_window) {
1667                 unsigned long shift = page_size * (head / page_size);
1668                 int res;
1669
1670                 res = munmap(buf, page_size * mmap_window);
1671                 assert(res == 0);
1672
1673                 offset += shift;
1674                 head -= shift;
1675                 goto remap;
1676         }
1677
1678         size = event->header.size;
1679
1680
1681         if (!size || process_event(event, offset, head) < 0) {
1682
1683                 /*
1684                  * assume we lost track of the stream, check alignment, and
1685                  * increment a single u64 in the hope to catch on again 'soon'.
1686                  */
1687
1688                 if (unlikely(head & 7))
1689                         head &= ~7ULL;
1690
1691                 size = 8;
1692         }
1693
1694         head += size;
1695
1696         if (offset + head < (unsigned long)perf_stat.st_size)
1697                 goto more;
1698
1699         rc = EXIT_SUCCESS;
1700         close(input);
1701
1702         return rc;
1703 }
1704
1705 static const char * const sched_usage[] = {
1706         "perf sched [<options>] {record|latency|replay|trace}",
1707         NULL
1708 };
1709
1710 static const struct option sched_options[] = {
1711         OPT_BOOLEAN('v', "verbose", &verbose,
1712                     "be more verbose (show symbol address, etc)"),
1713         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1714                     "dump raw trace in ASCII"),
1715         OPT_END()
1716 };
1717
1718 static const char * const latency_usage[] = {
1719         "perf sched latency [<options>]",
1720         NULL
1721 };
1722
1723 static const struct option latency_options[] = {
1724         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1725                    "sort by key(s): runtime, switch, avg, max"),
1726         OPT_BOOLEAN('v', "verbose", &verbose,
1727                     "be more verbose (show symbol address, etc)"),
1728         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1729                     "dump raw trace in ASCII"),
1730         OPT_END()
1731 };
1732
1733 static const char * const replay_usage[] = {
1734         "perf sched replay [<options>]",
1735         NULL
1736 };
1737
1738 static const struct option replay_options[] = {
1739         OPT_INTEGER('r', "repeat", &replay_repeat,
1740                     "repeat the workload replay N times (-1: infinite)"),
1741         OPT_BOOLEAN('v', "verbose", &verbose,
1742                     "be more verbose (show symbol address, etc)"),
1743         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1744                     "dump raw trace in ASCII"),
1745         OPT_END()
1746 };
1747
1748 static void setup_sorting(void)
1749 {
1750         char *tmp, *tok, *str = strdup(sort_order);
1751
1752         for (tok = strtok_r(str, ", ", &tmp);
1753                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1754                 if (sort_dimension__add(tok, &sort_list) < 0) {
1755                         error("Unknown --sort key: `%s'", tok);
1756                         usage_with_options(latency_usage, latency_options);
1757                 }
1758         }
1759
1760         free(str);
1761
1762         sort_dimension__add((char *)"pid", &cmp_pid);
1763 }
1764
1765 static const char *record_args[] = {
1766         "record",
1767         "-a",
1768         "-R",
1769         "-M",
1770         "-f",
1771         "-c", "1",
1772         "-e", "sched:sched_switch:r",
1773         "-e", "sched:sched_stat_wait:r",
1774         "-e", "sched:sched_stat_sleep:r",
1775         "-e", "sched:sched_stat_iowait:r",
1776         "-e", "sched:sched_stat_runtime:r",
1777         "-e", "sched:sched_process_exit:r",
1778         "-e", "sched:sched_process_fork:r",
1779         "-e", "sched:sched_wakeup:r",
1780         "-e", "sched:sched_migrate_task:r",
1781 };
1782
1783 static int __cmd_record(int argc, const char **argv)
1784 {
1785         unsigned int rec_argc, i, j;
1786         const char **rec_argv;
1787
1788         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1789         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1790
1791         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1792                 rec_argv[i] = strdup(record_args[i]);
1793
1794         for (j = 1; j < (unsigned int)argc; j++, i++)
1795                 rec_argv[i] = argv[j];
1796
1797         BUG_ON(i != rec_argc);
1798
1799         return cmd_record(i, rec_argv, NULL);
1800 }
1801
1802 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1803 {
1804         symbol__init();
1805         page_size = getpagesize();
1806
1807         argc = parse_options(argc, argv, sched_options, sched_usage,
1808                              PARSE_OPT_STOP_AT_NON_OPTION);
1809         if (!argc)
1810                 usage_with_options(sched_usage, sched_options);
1811
1812         if (!strncmp(argv[0], "rec", 3)) {
1813                 return __cmd_record(argc, argv);
1814         } else if (!strncmp(argv[0], "lat", 3)) {
1815                 trace_handler = &lat_ops;
1816                 if (argc > 1) {
1817                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1818                         if (argc)
1819                                 usage_with_options(latency_usage, latency_options);
1820                 }
1821                 setup_sorting();
1822                 __cmd_lat();
1823         } else if (!strncmp(argv[0], "rep", 3)) {
1824                 trace_handler = &replay_ops;
1825                 if (argc) {
1826                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1827                         if (argc)
1828                                 usage_with_options(replay_usage, replay_options);
1829                 }
1830                 __cmd_replay();
1831         } else if (!strcmp(argv[0], "trace")) {
1832                 /*
1833                  * Aliased to 'perf trace' for now:
1834                  */
1835                 return cmd_trace(argc, argv, prefix);
1836         } else {
1837                 usage_with_options(sched_usage, sched_options);
1838         }
1839
1840         return 0;
1841 }