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 #include "util/session.h"
  10
  11 #include "util/parse-options.h"
  12 #include "util/trace-event.h"
  13
  14 #include "util/debug.h"
  15
  16 #include <sys/prctl.h>
  17
  18 #include <semaphore.h>
  19 #include <pthread.h>
  20 #include <math.h>
  21
  22 static char                     const *input_name = "perf.data";
  23
  24 static char                     default_sort_order[] = "avg, max, switch, runtime";
  25 static char                     *sort_order = default_sort_order;
  26
  27 static int                      profile_cpu = -1;
  28
  29 #define PR_SET_NAME             15               /* Set process name */
  30 #define MAX_CPUS                4096
  31
  32 static u64                      run_measurement_overhead;
  33 static u64                      sleep_measurement_overhead;
  34
  35 #define COMM_LEN                20
  36 #define SYM_LEN                 129
  37
  38 #define MAX_PID                 65536
  39
  40 static unsigned long            nr_tasks;
  41
  42 struct sched_atom;
  43
  44 struct task_desc {
  45         unsigned long           nr;
  46         unsigned long           pid;
  47         char                    comm[COMM_LEN];
  48
  49         unsigned long           nr_events;
  50         unsigned long           curr_event;
  51         struct sched_atom       **atoms;
  52
  53         pthread_t               thread;
  54         sem_t                   sleep_sem;
  55
  56         sem_t                   ready_for_work;
  57         sem_t                   work_done_sem;
  58
  59         u64                     cpu_usage;
  60 };
  61
  62 enum sched_event_type {
  63         SCHED_EVENT_RUN,
  64         SCHED_EVENT_SLEEP,
  65         SCHED_EVENT_WAKEUP,
  66         SCHED_EVENT_MIGRATION,
  67 };
  68
  69 struct sched_atom {
  70         enum sched_event_type   type;
  71         u64                     timestamp;
  72         u64                     duration;
  73         unsigned long           nr;
  74         int                     specific_wait;
  75         sem_t                   *wait_sem;
  76         struct task_desc        *wakee;
  77 };
  78
  79 static struct task_desc         *pid_to_task[MAX_PID];
  80
  81 static struct task_desc         **tasks;
  82
  83 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  84 static u64                      start_time;
  85
  86 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  87
  88 static unsigned long            nr_run_events;
  89 static unsigned long            nr_sleep_events;
  90 static unsigned long            nr_wakeup_events;
  91
  92 static unsigned long            nr_sleep_corrections;
  93 static unsigned long            nr_run_events_optimized;
  94
  95 static unsigned long            targetless_wakeups;
  96 static unsigned long            multitarget_wakeups;
  97
  98 static u64                      cpu_usage;
  99 static u64                      runavg_cpu_usage;
 100 static u64                      parent_cpu_usage;
 101 static u64                      runavg_parent_cpu_usage;
 102
 103 static unsigned long            nr_runs;
 104 static u64                      sum_runtime;
 105 static u64                      sum_fluct;
 106 static u64                      run_avg;
 107
 108 static unsigned long            replay_repeat = 10;
 109 static unsigned long            nr_timestamps;
 110 static unsigned long            nr_unordered_timestamps;
 111 static unsigned long            nr_state_machine_bugs;
 112 static unsigned long            nr_context_switch_bugs;
 113 static unsigned long            nr_events;
 114 static unsigned long            nr_lost_chunks;
 115 static unsigned long            nr_lost_events;
 116
 117 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 118
 119 enum thread_state {
 120         THREAD_SLEEPING = 0,
 121         THREAD_WAIT_CPU,
 122         THREAD_SCHED_IN,
 123         THREAD_IGNORE
 124 };
 125
 126 struct work_atom {
 127         struct list_head        list;
 128         enum thread_state       state;
 129         u64                     sched_out_time;
 130         u64                     wake_up_time;
 131         u64                     sched_in_time;
 132         u64                     runtime;
 133 };
 134
 135 struct work_atoms {
 136         struct list_head        work_list;
 137         struct thread           *thread;
 138         struct rb_node          node;
 139         u64                     max_lat;
 140         u64                     max_lat_at;
 141         u64                     total_lat;
 142         u64                     nb_atoms;
 143         u64                     total_runtime;
 144 };
 145
 146 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 147
 148 static struct rb_root           atom_root, sorted_atom_root;
 149
 150 static u64                      all_runtime;
 151 static u64                      all_count;
 152
 153
 154 static u64 get_nsecs(void)
 155 {
 156         struct timespec ts;
 157
 158         clock_gettime(CLOCK_MONOTONIC, &ts);
 159
 160         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 161 }
 162
 163 static void burn_nsecs(u64 nsecs)
 164 {
 165         u64 T0 = get_nsecs(), T1;
 166
 167         do {
 168                 T1 = get_nsecs();
 169         } while (T1 + run_measurement_overhead < T0 + nsecs);
 170 }
 171
 172 static void sleep_nsecs(u64 nsecs)
 173 {
 174         struct timespec ts;
 175
 176         ts.tv_nsec = nsecs % 999999999;
 177         ts.tv_sec = nsecs / 999999999;
 178
 179         nanosleep(&ts, NULL);
 180 }
 181
 182 static void calibrate_run_measurement_overhead(void)
 183 {
 184         u64 T0, T1, delta, min_delta = 1000000000ULL;
 185         int i;
 186
 187         for (i = 0; i < 10; i++) {
 188                 T0 = get_nsecs();
 189                 burn_nsecs(0);
 190                 T1 = get_nsecs();
 191                 delta = T1-T0;
 192                 min_delta = min(min_delta, delta);
 193         }
 194         run_measurement_overhead = min_delta;
 195
 196         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 197 }
 198
 199 static void calibrate_sleep_measurement_overhead(void)
 200 {
 201         u64 T0, T1, delta, min_delta = 1000000000ULL;
 202         int i;
 203
 204         for (i = 0; i < 10; i++) {
 205                 T0 = get_nsecs();
 206                 sleep_nsecs(10000);
 207                 T1 = get_nsecs();
 208                 delta = T1-T0;
 209                 min_delta = min(min_delta, delta);
 210         }
 211         min_delta -= 10000;
 212         sleep_measurement_overhead = min_delta;
 213
 214         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 215 }
 216
 217 static struct sched_atom *
 218 get_new_event(struct task_desc *task, u64 timestamp)
 219 {
 220         struct sched_atom *event = zalloc(sizeof(*event));
 221         unsigned long idx = task->nr_events;
 222         size_t size;
 223
 224         event->timestamp = timestamp;
 225         event->nr = idx;
 226
 227         task->nr_events++;
 228         size = sizeof(struct sched_atom *) * task->nr_events;
 229         task->atoms = realloc(task->atoms, size);
 230         BUG_ON(!task->atoms);
 231
 232         task->atoms[idx] = event;
 233
 234         return event;
 235 }
 236
 237 static struct sched_atom *last_event(struct task_desc *task)
 238 {
 239         if (!task->nr_events)
 240                 return NULL;
 241
 242         return task->atoms[task->nr_events - 1];
 243 }
 244
 245 static void
 246 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 247 {
 248         struct sched_atom *event, *curr_event = last_event(task);
 249
 250         /*
 251          * optimize an existing RUN event by merging this one
 252          * to it:
 253          */
 254         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 255                 nr_run_events_optimized++;
 256                 curr_event->duration += duration;
 257                 return;
 258         }
 259
 260         event = get_new_event(task, timestamp);
 261
 262         event->type = SCHED_EVENT_RUN;
 263         event->duration = duration;
 264
 265         nr_run_events++;
 266 }
 267
 268 static void
 269 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 270                        struct task_desc *wakee)
 271 {
 272         struct sched_atom *event, *wakee_event;
 273
 274         event = get_new_event(task, timestamp);
 275         event->type = SCHED_EVENT_WAKEUP;
 276         event->wakee = wakee;
 277
 278         wakee_event = last_event(wakee);
 279         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 280                 targetless_wakeups++;
 281                 return;
 282         }
 283         if (wakee_event->wait_sem) {
 284                 multitarget_wakeups++;
 285                 return;
 286         }
 287
 288         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 289         sem_init(wakee_event->wait_sem, 0, 0);
 290         wakee_event->specific_wait = 1;
 291         event->wait_sem = wakee_event->wait_sem;
 292
 293         nr_wakeup_events++;
 294 }
 295
 296 static void
 297 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 298                       u64 task_state __used)
 299 {
 300         struct sched_atom *event = get_new_event(task, timestamp);
 301
 302         event->type = SCHED_EVENT_SLEEP;
 303
 304         nr_sleep_events++;
 305 }
 306
 307 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 308 {
 309         struct task_desc *task;
 310
 311         BUG_ON(pid >= MAX_PID);
 312
 313         task = pid_to_task[pid];
 314
 315         if (task)
 316                 return task;
 317
 318         task = zalloc(sizeof(*task));
 319         task->pid = pid;
 320         task->nr = nr_tasks;
 321         strcpy(task->comm, comm);
 322         /*
 323          * every task starts in sleeping state - this gets ignored
 324          * if there's no wakeup pointing to this sleep state:
 325          */
 326         add_sched_event_sleep(task, 0, 0);
 327
 328         pid_to_task[pid] = task;
 329         nr_tasks++;
 330         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 331         BUG_ON(!tasks);
 332         tasks[task->nr] = task;
 333
 334         if (verbose)
 335                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 336
 337         return task;
 338 }
 339
 340
 341 static void print_task_traces(void)
 342 {
 343         struct task_desc *task;
 344         unsigned long i;
 345
 346         for (i = 0; i < nr_tasks; i++) {
 347                 task = tasks[i];
 348                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 349                         task->nr, task->comm, task->pid, task->nr_events);
 350         }
 351 }
 352
 353 static void add_cross_task_wakeups(void)
 354 {
 355         struct task_desc *task1, *task2;
 356         unsigned long i, j;
 357
 358         for (i = 0; i < nr_tasks; i++) {
 359                 task1 = tasks[i];
 360                 j = i + 1;
 361                 if (j == nr_tasks)
 362                         j = 0;
 363                 task2 = tasks[j];
 364                 add_sched_event_wakeup(task1, 0, task2);
 365         }
 366 }
 367
 368 static void
 369 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 370 {
 371         int ret = 0;
 372         u64 now;
 373         long long delta;
 374
 375         now = get_nsecs();
 376         delta = start_time + atom->timestamp - now;
 377
 378         switch (atom->type) {
 379                 case SCHED_EVENT_RUN:
 380                         burn_nsecs(atom->duration);
 381                         break;
 382                 case SCHED_EVENT_SLEEP:
 383                         if (atom->wait_sem)
 384                                 ret = sem_wait(atom->wait_sem);
 385                         BUG_ON(ret);
 386                         break;
 387                 case SCHED_EVENT_WAKEUP:
 388                         if (atom->wait_sem)
 389                                 ret = sem_post(atom->wait_sem);
 390                         BUG_ON(ret);
 391                         break;
 392                 case SCHED_EVENT_MIGRATION:
 393                         break;
 394                 default:
 395                         BUG_ON(1);
 396         }
 397 }
 398
 399 static u64 get_cpu_usage_nsec_parent(void)
 400 {
 401         struct rusage ru;
 402         u64 sum;
 403         int err;
 404
 405         err = getrusage(RUSAGE_SELF, &ru);
 406         BUG_ON(err);
 407
 408         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 409         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 410
 411         return sum;
 412 }
 413
 414 static int self_open_counters(void)
 415 {
 416         struct perf_event_attr attr;
 417         int fd;
 418
 419         memset(&attr, 0, sizeof(attr));
 420
 421         attr.type = PERF_TYPE_SOFTWARE;
 422         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 423
 424         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 425
 426         if (fd < 0)
 427                 die("Error: sys_perf_event_open() syscall returned"
 428                     "with %d (%s)\n", fd, strerror(errno));
 429         return fd;
 430 }
 431
 432 static u64 get_cpu_usage_nsec_self(int fd)
 433 {
 434         u64 runtime;
 435         int ret;
 436
 437         ret = read(fd, &runtime, sizeof(runtime));
 438         BUG_ON(ret != sizeof(runtime));
 439
 440         return runtime;
 441 }
 442
 443 static void *thread_func(void *ctx)
 444 {
 445         struct task_desc *this_task = ctx;
 446         u64 cpu_usage_0, cpu_usage_1;
 447         unsigned long i, ret;
 448         char comm2[22];
 449         int fd;
 450
 451         sprintf(comm2, ":%s", this_task->comm);
 452         prctl(PR_SET_NAME, comm2);
 453         fd = self_open_counters();
 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(fd);
 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->atoms[i]);
 468         }
 469
 470         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 471         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 472         ret = sem_post(&this_task->work_done_sem);
 473         BUG_ON(ret);
 474
 475         ret = pthread_mutex_lock(&work_done_wait_mutex);
 476         BUG_ON(ret);
 477         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 478         BUG_ON(ret);
 479
 480         goto again;
 481 }
 482
 483 static void create_tasks(void)
 484 {
 485         struct task_desc *task;
 486         pthread_attr_t attr;
 487         unsigned long i;
 488         int err;
 489
 490         err = pthread_attr_init(&attr);
 491         BUG_ON(err);
 492         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 493         BUG_ON(err);
 494         err = pthread_mutex_lock(&start_work_mutex);
 495         BUG_ON(err);
 496         err = pthread_mutex_lock(&work_done_wait_mutex);
 497         BUG_ON(err);
 498         for (i = 0; i < nr_tasks; i++) {
 499                 task = tasks[i];
 500                 sem_init(&task->sleep_sem, 0, 0);
 501                 sem_init(&task->ready_for_work, 0, 0);
 502                 sem_init(&task->work_done_sem, 0, 0);
 503                 task->curr_event = 0;
 504                 err = pthread_create(&task->thread, &attr, thread_func, task);
 505                 BUG_ON(err);
 506         }
 507 }
 508
 509 static void wait_for_tasks(void)
 510 {
 511         u64 cpu_usage_0, cpu_usage_1;
 512         struct task_desc *task;
 513         unsigned long i, ret;
 514
 515         start_time = get_nsecs();
 516         cpu_usage = 0;
 517         pthread_mutex_unlock(&work_done_wait_mutex);
 518
 519         for (i = 0; i < nr_tasks; i++) {
 520                 task = tasks[i];
 521                 ret = sem_wait(&task->ready_for_work);
 522                 BUG_ON(ret);
 523                 sem_init(&task->ready_for_work, 0, 0);
 524         }
 525         ret = pthread_mutex_lock(&work_done_wait_mutex);
 526         BUG_ON(ret);
 527
 528         cpu_usage_0 = get_cpu_usage_nsec_parent();
 529
 530         pthread_mutex_unlock(&start_work_mutex);
 531
 532         for (i = 0; i < nr_tasks; i++) {
 533                 task = tasks[i];
 534                 ret = sem_wait(&task->work_done_sem);
 535                 BUG_ON(ret);
 536                 sem_init(&task->work_done_sem, 0, 0);
 537                 cpu_usage += task->cpu_usage;
 538                 task->cpu_usage = 0;
 539         }
 540
 541         cpu_usage_1 = get_cpu_usage_nsec_parent();
 542         if (!runavg_cpu_usage)
 543                 runavg_cpu_usage = cpu_usage;
 544         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 545
 546         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 547         if (!runavg_parent_cpu_usage)
 548                 runavg_parent_cpu_usage = parent_cpu_usage;
 549         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 550                                    parent_cpu_usage)/10;
 551
 552         ret = pthread_mutex_lock(&start_work_mutex);
 553         BUG_ON(ret);
 554
 555         for (i = 0; i < nr_tasks; i++) {
 556                 task = tasks[i];
 557                 sem_init(&task->sleep_sem, 0, 0);
 558                 task->curr_event = 0;
 559         }
 560 }
 561
 562 static void run_one_test(void)
 563 {
 564         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 565
 566         T0 = get_nsecs();
 567         wait_for_tasks();
 568         T1 = get_nsecs();
 569
 570         delta = T1 - T0;
 571         sum_runtime += delta;
 572         nr_runs++;
 573
 574         avg_delta = sum_runtime / nr_runs;
 575         if (delta < avg_delta)
 576                 fluct = avg_delta - delta;
 577         else
 578                 fluct = delta - avg_delta;
 579         sum_fluct += fluct;
 580         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 581         if (!run_avg)
 582                 run_avg = delta;
 583         run_avg = (run_avg*9 + delta)/10;
 584
 585         printf("#%-3ld: %0.3f, ",
 586                 nr_runs, (double)delta/1000000.0);
 587
 588         printf("ravg: %0.2f, ",
 589                 (double)run_avg/1e6);
 590
 591         printf("cpu: %0.2f / %0.2f",
 592                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 593
 594 #if 0
 595         /*
 596          * rusage statistics done by the parent, these are less
 597          * accurate than the sum_exec_runtime based statistics:
 598          */
 599         printf(" [%0.2f / %0.2f]",
 600                 (double)parent_cpu_usage/1e6,
 601                 (double)runavg_parent_cpu_usage/1e6);
 602 #endif
 603
 604         printf("\n");
 605
 606         if (nr_sleep_corrections)
 607                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 608         nr_sleep_corrections = 0;
 609 }
 610
 611 static void test_calibrations(void)
 612 {
 613         u64 T0, T1;
 614
 615         T0 = get_nsecs();
 616         burn_nsecs(1e6);
 617         T1 = get_nsecs();
 618
 619         printf("the run test took %Ld nsecs\n", T1-T0);
 620
 621         T0 = get_nsecs();
 622         sleep_nsecs(1e6);
 623         T1 = get_nsecs();
 624
 625         printf("the sleep test took %Ld nsecs\n", T1-T0);
 626 }
 627
 628 #define FILL_FIELD(ptr, field, event, data)     \
 629         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 630
 631 #define FILL_ARRAY(ptr, array, event, data)                     \
 632 do {                                                            \
 633         void *__array = raw_field_ptr(event, #array, data);     \
 634         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 635 } while(0)
 636
 637 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 638 do {                                                            \
 639         FILL_FIELD(ptr, common_type, event, data);              \
 640         FILL_FIELD(ptr, common_flags, event, data);             \
 641         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 642         FILL_FIELD(ptr, common_pid, event, data);               \
 643         FILL_FIELD(ptr, common_tgid, event, data);              \
 644 } while (0)
 645
 646
 647
 648 struct trace_switch_event {
 649         u32 size;
 650
 651         u16 common_type;
 652         u8 common_flags;
 653         u8 common_preempt_count;
 654         u32 common_pid;
 655         u32 common_tgid;
 656
 657         char prev_comm[16];
 658         u32 prev_pid;
 659         u32 prev_prio;
 660         u64 prev_state;
 661         char next_comm[16];
 662         u32 next_pid;
 663         u32 next_prio;
 664 };
 665
 666 struct trace_runtime_event {
 667         u32 size;
 668
 669         u16 common_type;
 670         u8 common_flags;
 671         u8 common_preempt_count;
 672         u32 common_pid;
 673         u32 common_tgid;
 674
 675         char comm[16];
 676         u32 pid;
 677         u64 runtime;
 678         u64 vruntime;
 679 };
 680
 681 struct trace_wakeup_event {
 682         u32 size;
 683
 684         u16 common_type;
 685         u8 common_flags;
 686         u8 common_preempt_count;
 687         u32 common_pid;
 688         u32 common_tgid;
 689
 690         char comm[16];
 691         u32 pid;
 692
 693         u32 prio;
 694         u32 success;
 695         u32 cpu;
 696 };
 697
 698 struct trace_fork_event {
 699         u32 size;
 700
 701         u16 common_type;
 702         u8 common_flags;
 703         u8 common_preempt_count;
 704         u32 common_pid;
 705         u32 common_tgid;
 706
 707         char parent_comm[16];
 708         u32 parent_pid;
 709         char child_comm[16];
 710         u32 child_pid;
 711 };
 712
 713 struct trace_migrate_task_event {
 714         u32 size;
 715
 716         u16 common_type;
 717         u8 common_flags;
 718         u8 common_preempt_count;
 719         u32 common_pid;
 720         u32 common_tgid;
 721
 722         char comm[16];
 723         u32 pid;
 724
 725         u32 prio;
 726         u32 cpu;
 727 };
 728
 729 struct trace_sched_handler {
 730         void (*switch_event)(struct trace_switch_event *,
 731                              struct perf_session *,
 732                              struct event *,
 733                              int cpu,
 734                              u64 timestamp,
 735                              struct thread *thread);
 736
 737         void (*runtime_event)(struct trace_runtime_event *,
 738                               struct perf_session *,
 739                               struct event *,
 740                               int cpu,
 741                               u64 timestamp,
 742                               struct thread *thread);
 743
 744         void (*wakeup_event)(struct trace_wakeup_event *,
 745                              struct perf_session *,
 746                              struct event *,
 747                              int cpu,
 748                              u64 timestamp,
 749                              struct thread *thread);
 750
 751         void (*fork_event)(struct trace_fork_event *,
 752                            struct event *,
 753                            int cpu,
 754                            u64 timestamp,
 755                            struct thread *thread);
 756
 757         void (*migrate_task_event)(struct trace_migrate_task_event *,
 758                            struct perf_session *session,
 759                            struct event *,
 760                            int cpu,
 761                            u64 timestamp,
 762                            struct thread *thread);
 763 };
 764
 765
 766 static void
 767 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 768                     struct perf_session *session __used,
 769                     struct event *event,
 770                     int cpu __used,
 771                     u64 timestamp __used,
 772                     struct thread *thread __used)
 773 {
 774         struct task_desc *waker, *wakee;
 775
 776         if (verbose) {
 777                 printf("sched_wakeup event %p\n", event);
 778
 779                 printf(" ... pid %d woke up %s/%d\n",
 780                         wakeup_event->common_pid,
 781                         wakeup_event->comm,
 782                         wakeup_event->pid);
 783         }
 784
 785         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 786         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 787
 788         add_sched_event_wakeup(waker, timestamp, wakee);
 789 }
 790
 791 static u64 cpu_last_switched[MAX_CPUS];
 792
 793 static void
 794 replay_switch_event(struct trace_switch_event *switch_event,
 795                     struct perf_session *session __used,
 796                     struct event *event,
 797                     int cpu,
 798                     u64 timestamp,
 799                     struct thread *thread __used)
 800 {
 801         struct task_desc *prev, *next;
 802         u64 timestamp0;
 803         s64 delta;
 804
 805         if (verbose)
 806                 printf("sched_switch event %p\n", event);
 807
 808         if (cpu >= MAX_CPUS || cpu < 0)
 809                 return;
 810
 811         timestamp0 = cpu_last_switched[cpu];
 812         if (timestamp0)
 813                 delta = timestamp - timestamp0;
 814         else
 815                 delta = 0;
 816
 817         if (delta < 0)
 818                 die("hm, delta: %Ld < 0 ?\n", delta);
 819
 820         if (verbose) {
 821                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 822                         switch_event->prev_comm, switch_event->prev_pid,
 823                         switch_event->next_comm, switch_event->next_pid,
 824                         delta);
 825         }
 826
 827         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 828         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 829
 830         cpu_last_switched[cpu] = timestamp;
 831
 832         add_sched_event_run(prev, timestamp, delta);
 833         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 834 }
 835
 836
 837 static void
 838 replay_fork_event(struct trace_fork_event *fork_event,
 839                   struct event *event,
 840                   int cpu __used,
 841                   u64 timestamp __used,
 842                   struct thread *thread __used)
 843 {
 844         if (verbose) {
 845                 printf("sched_fork event %p\n", event);
 846                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 847                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 848         }
 849         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 850         register_pid(fork_event->child_pid, fork_event->child_comm);
 851 }
 852
 853 static struct trace_sched_handler replay_ops  = {
 854         .wakeup_event           = replay_wakeup_event,
 855         .switch_event           = replay_switch_event,
 856         .fork_event             = replay_fork_event,
 857 };
 858
 859 struct sort_dimension {
 860         const char              *name;
 861         sort_fn_t               cmp;
 862         struct list_head        list;
 863 };
 864
 865 static LIST_HEAD(cmp_pid);
 866
 867 static int
 868 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 869 {
 870         struct sort_dimension *sort;
 871         int ret = 0;
 872
 873         BUG_ON(list_empty(list));
 874
 875         list_for_each_entry(sort, list, list) {
 876                 ret = sort->cmp(l, r);
 877                 if (ret)
 878                         return ret;
 879         }
 880
 881         return ret;
 882 }
 883
 884 static struct work_atoms *
 885 thread_atoms_search(struct rb_root *root, struct thread *thread,
 886                          struct list_head *sort_list)
 887 {
 888         struct rb_node *node = root->rb_node;
 889         struct work_atoms key = { .thread = thread };
 890
 891         while (node) {
 892                 struct work_atoms *atoms;
 893                 int cmp;
 894
 895                 atoms = container_of(node, struct work_atoms, node);
 896
 897                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 898                 if (cmp > 0)
 899                         node = node->rb_left;
 900                 else if (cmp < 0)
 901                         node = node->rb_right;
 902                 else {
 903                         BUG_ON(thread != atoms->thread);
 904                         return atoms;
 905                 }
 906         }
 907         return NULL;
 908 }
 909
 910 static void
 911 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 912                          struct list_head *sort_list)
 913 {
 914         struct rb_node **new = &(root->rb_node), *parent = NULL;
 915
 916         while (*new) {
 917                 struct work_atoms *this;
 918                 int cmp;
 919
 920                 this = container_of(*new, struct work_atoms, node);
 921                 parent = *new;
 922
 923                 cmp = thread_lat_cmp(sort_list, data, this);
 924
 925                 if (cmp > 0)
 926                         new = &((*new)->rb_left);
 927                 else
 928                         new = &((*new)->rb_right);
 929         }
 930
 931         rb_link_node(&data->node, parent, new);
 932         rb_insert_color(&data->node, root);
 933 }
 934
 935 static void thread_atoms_insert(struct thread *thread)
 936 {
 937         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 938         if (!atoms)
 939                 die("No memory");
 940
 941         atoms->thread = thread;
 942         INIT_LIST_HEAD(&atoms->work_list);
 943         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 944 }
 945
 946 static void
 947 latency_fork_event(struct trace_fork_event *fork_event __used,
 948                    struct event *event __used,
 949                    int cpu __used,
 950                    u64 timestamp __used,
 951                    struct thread *thread __used)
 952 {
 953         /* should insert the newcomer */
 954 }
 955
 956 __used
 957 static char sched_out_state(struct trace_switch_event *switch_event)
 958 {
 959         const char *str = TASK_STATE_TO_CHAR_STR;
 960
 961         return str[switch_event->prev_state];
 962 }
 963
 964 static void
 965 add_sched_out_event(struct work_atoms *atoms,
 966                     char run_state,
 967                     u64 timestamp)
 968 {
 969         struct work_atom *atom = zalloc(sizeof(*atom));
 970         if (!atom)
 971                 die("Non memory");
 972
 973         atom->sched_out_time = timestamp;
 974
 975         if (run_state == 'R') {
 976                 atom->state = THREAD_WAIT_CPU;
 977                 atom->wake_up_time = atom->sched_out_time;
 978         }
 979
 980         list_add_tail(&atom->list, &atoms->work_list);
 981 }
 982
 983 static void
 984 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
 985 {
 986         struct work_atom *atom;
 987
 988         BUG_ON(list_empty(&atoms->work_list));
 989
 990         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 991
 992         atom->runtime += delta;
 993         atoms->total_runtime += delta;
 994 }
 995
 996 static void
 997 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 998 {
 999         struct work_atom *atom;
1000         u64 delta;
1001
1002         if (list_empty(&atoms->work_list))
1003                 return;
1004
1005         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1006
1007         if (atom->state != THREAD_WAIT_CPU)
1008                 return;
1009
1010         if (timestamp < atom->wake_up_time) {
1011                 atom->state = THREAD_IGNORE;
1012                 return;
1013         }
1014
1015         atom->state = THREAD_SCHED_IN;
1016         atom->sched_in_time = timestamp;
1017
1018         delta = atom->sched_in_time - atom->wake_up_time;
1019         atoms->total_lat += delta;
1020         if (delta > atoms->max_lat) {
1021                 atoms->max_lat = delta;
1022                 atoms->max_lat_at = timestamp;
1023         }
1024         atoms->nb_atoms++;
1025 }
1026
1027 static void
1028 latency_switch_event(struct trace_switch_event *switch_event,
1029                      struct perf_session *session,
1030                      struct event *event __used,
1031                      int cpu,
1032                      u64 timestamp,
1033                      struct thread *thread __used)
1034 {
1035         struct work_atoms *out_events, *in_events;
1036         struct thread *sched_out, *sched_in;
1037         u64 timestamp0;
1038         s64 delta;
1039
1040         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1041
1042         timestamp0 = cpu_last_switched[cpu];
1043         cpu_last_switched[cpu] = timestamp;
1044         if (timestamp0)
1045                 delta = timestamp - timestamp0;
1046         else
1047                 delta = 0;
1048
1049         if (delta < 0)
1050                 die("hm, delta: %Ld < 0 ?\n", delta);
1051
1052
1053         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1054         sched_in = perf_session__findnew(session, switch_event->next_pid);
1055
1056         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1057         if (!out_events) {
1058                 thread_atoms_insert(sched_out);
1059                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1060                 if (!out_events)
1061                         die("out-event: Internal tree error");
1062         }
1063         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1064
1065         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1066         if (!in_events) {
1067                 thread_atoms_insert(sched_in);
1068                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1069                 if (!in_events)
1070                         die("in-event: Internal tree error");
1071                 /*
1072                  * Take came in we have not heard about yet,
1073                  * add in an initial atom in runnable state:
1074                  */
1075                 add_sched_out_event(in_events, 'R', timestamp);
1076         }
1077         add_sched_in_event(in_events, timestamp);
1078 }
1079
1080 static void
1081 latency_runtime_event(struct trace_runtime_event *runtime_event,
1082                      struct perf_session *session,
1083                      struct event *event __used,
1084                      int cpu,
1085                      u64 timestamp,
1086                      struct thread *this_thread __used)
1087 {
1088         struct thread *thread = perf_session__findnew(session, runtime_event->pid);
1089         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1090
1091         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1092         if (!atoms) {
1093                 thread_atoms_insert(thread);
1094                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1095                 if (!atoms)
1096                         die("in-event: Internal tree error");
1097                 add_sched_out_event(atoms, 'R', timestamp);
1098         }
1099
1100         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1101 }
1102
1103 static void
1104 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1105                      struct perf_session *session,
1106                      struct event *__event __used,
1107                      int cpu __used,
1108                      u64 timestamp,
1109                      struct thread *thread __used)
1110 {
1111         struct work_atoms *atoms;
1112         struct work_atom *atom;
1113         struct thread *wakee;
1114
1115         /* Note for later, it may be interesting to observe the failing cases */
1116         if (!wakeup_event->success)
1117                 return;
1118
1119         wakee = perf_session__findnew(session, wakeup_event->pid);
1120         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1121         if (!atoms) {
1122                 thread_atoms_insert(wakee);
1123                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1124                 if (!atoms)
1125                         die("wakeup-event: Internal tree error");
1126                 add_sched_out_event(atoms, 'S', timestamp);
1127         }
1128
1129         BUG_ON(list_empty(&atoms->work_list));
1130
1131         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1132
1133         /*
1134          * You WILL be missing events if you've recorded only
1135          * one CPU, or are only looking at only one, so don't
1136          * make useless noise.
1137          */
1138         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1139                 nr_state_machine_bugs++;
1140
1141         nr_timestamps++;
1142         if (atom->sched_out_time > timestamp) {
1143                 nr_unordered_timestamps++;
1144                 return;
1145         }
1146
1147         atom->state = THREAD_WAIT_CPU;
1148         atom->wake_up_time = timestamp;
1149 }
1150
1151 static void
1152 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1153                      struct perf_session *session,
1154                      struct event *__event __used,
1155                      int cpu __used,
1156                      u64 timestamp,
1157                      struct thread *thread __used)
1158 {
1159         struct work_atoms *atoms;
1160         struct work_atom *atom;
1161         struct thread *migrant;
1162
1163         /*
1164          * Only need to worry about migration when profiling one CPU.
1165          */
1166         if (profile_cpu == -1)
1167                 return;
1168
1169         migrant = perf_session__findnew(session, migrate_task_event->pid);
1170         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1171         if (!atoms) {
1172                 thread_atoms_insert(migrant);
1173                 register_pid(migrant->pid, migrant->comm);
1174                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1175                 if (!atoms)
1176                         die("migration-event: Internal tree error");
1177                 add_sched_out_event(atoms, 'R', timestamp);
1178         }
1179
1180         BUG_ON(list_empty(&atoms->work_list));
1181
1182         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1183         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1184
1185         nr_timestamps++;
1186
1187         if (atom->sched_out_time > timestamp)
1188                 nr_unordered_timestamps++;
1189 }
1190
1191 static struct trace_sched_handler lat_ops  = {
1192         .wakeup_event           = latency_wakeup_event,
1193         .switch_event           = latency_switch_event,
1194         .runtime_event          = latency_runtime_event,
1195         .fork_event             = latency_fork_event,
1196         .migrate_task_event     = latency_migrate_task_event,
1197 };
1198
1199 static void output_lat_thread(struct work_atoms *work_list)
1200 {
1201         int i;
1202         int ret;
1203         u64 avg;
1204
1205         if (!work_list->nb_atoms)
1206                 return;
1207         /*
1208          * Ignore idle threads:
1209          */
1210         if (!strcmp(work_list->thread->comm, "swapper"))
1211                 return;
1212
1213         all_runtime += work_list->total_runtime;
1214         all_count += work_list->nb_atoms;
1215
1216         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1217
1218         for (i = 0; i < 24 - ret; i++)
1219                 printf(" ");
1220
1221         avg = work_list->total_lat / work_list->nb_atoms;
1222
1223         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1224               (double)work_list->total_runtime / 1e6,
1225                  work_list->nb_atoms, (double)avg / 1e6,
1226                  (double)work_list->max_lat / 1e6,
1227                  (double)work_list->max_lat_at / 1e9);
1228 }
1229
1230 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1231 {
1232         if (l->thread->pid < r->thread->pid)
1233                 return -1;
1234         if (l->thread->pid > r->thread->pid)
1235                 return 1;
1236
1237         return 0;
1238 }
1239
1240 static struct sort_dimension pid_sort_dimension = {
1241         .name                   = "pid",
1242         .cmp                    = pid_cmp,
1243 };
1244
1245 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1246 {
1247         u64 avgl, avgr;
1248
1249         if (!l->nb_atoms)
1250                 return -1;
1251
1252         if (!r->nb_atoms)
1253                 return 1;
1254
1255         avgl = l->total_lat / l->nb_atoms;
1256         avgr = r->total_lat / r->nb_atoms;
1257
1258         if (avgl < avgr)
1259                 return -1;
1260         if (avgl > avgr)
1261                 return 1;
1262
1263         return 0;
1264 }
1265
1266 static struct sort_dimension avg_sort_dimension = {
1267         .name                   = "avg",
1268         .cmp                    = avg_cmp,
1269 };
1270
1271 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1272 {
1273         if (l->max_lat < r->max_lat)
1274                 return -1;
1275         if (l->max_lat > r->max_lat)
1276                 return 1;
1277
1278         return 0;
1279 }
1280
1281 static struct sort_dimension max_sort_dimension = {
1282         .name                   = "max",
1283         .cmp                    = max_cmp,
1284 };
1285
1286 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1287 {
1288         if (l->nb_atoms < r->nb_atoms)
1289                 return -1;
1290         if (l->nb_atoms > r->nb_atoms)
1291                 return 1;
1292
1293         return 0;
1294 }
1295
1296 static struct sort_dimension switch_sort_dimension = {
1297         .name                   = "switch",
1298         .cmp                    = switch_cmp,
1299 };
1300
1301 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1302 {
1303         if (l->total_runtime < r->total_runtime)
1304                 return -1;
1305         if (l->total_runtime > r->total_runtime)
1306                 return 1;
1307
1308         return 0;
1309 }
1310
1311 static struct sort_dimension runtime_sort_dimension = {
1312         .name                   = "runtime",
1313         .cmp                    = runtime_cmp,
1314 };
1315
1316 static struct sort_dimension *available_sorts[] = {
1317         &pid_sort_dimension,
1318         &avg_sort_dimension,
1319         &max_sort_dimension,
1320         &switch_sort_dimension,
1321         &runtime_sort_dimension,
1322 };
1323
1324 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1325
1326 static LIST_HEAD(sort_list);
1327
1328 static int sort_dimension__add(const char *tok, struct list_head *list)
1329 {
1330         int i;
1331
1332         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1333                 if (!strcmp(available_sorts[i]->name, tok)) {
1334                         list_add_tail(&available_sorts[i]->list, list);
1335
1336                         return 0;
1337                 }
1338         }
1339
1340         return -1;
1341 }
1342
1343 static void setup_sorting(void);
1344
1345 static void sort_lat(void)
1346 {
1347         struct rb_node *node;
1348
1349         for (;;) {
1350                 struct work_atoms *data;
1351                 node = rb_first(&atom_root);
1352                 if (!node)
1353                         break;
1354
1355                 rb_erase(node, &atom_root);
1356                 data = rb_entry(node, struct work_atoms, node);
1357                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1358         }
1359 }
1360
1361 static struct trace_sched_handler *trace_handler;
1362
1363 static void
1364 process_sched_wakeup_event(void *data, struct perf_session *session,
1365                            struct event *event,
1366                            int cpu __used,
1367                            u64 timestamp __used,
1368                            struct thread *thread __used)
1369 {
1370         struct trace_wakeup_event wakeup_event;
1371
1372         FILL_COMMON_FIELDS(wakeup_event, event, data);
1373
1374         FILL_ARRAY(wakeup_event, comm, event, data);
1375         FILL_FIELD(wakeup_event, pid, event, data);
1376         FILL_FIELD(wakeup_event, prio, event, data);
1377         FILL_FIELD(wakeup_event, success, event, data);
1378         FILL_FIELD(wakeup_event, cpu, event, data);
1379
1380         if (trace_handler->wakeup_event)
1381                 trace_handler->wakeup_event(&wakeup_event, session, event,
1382                                             cpu, timestamp, thread);
1383 }
1384
1385 /*
1386  * Track the current task - that way we can know whether there's any
1387  * weird events, such as a task being switched away that is not current.
1388  */
1389 static int max_cpu;
1390
1391 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1392
1393 static struct thread *curr_thread[MAX_CPUS];
1394
1395 static char next_shortname1 = 'A';
1396 static char next_shortname2 = '0';
1397
1398 static void
1399 map_switch_event(struct trace_switch_event *switch_event,
1400                  struct perf_session *session,
1401                  struct event *event __used,
1402                  int this_cpu,
1403                  u64 timestamp,
1404                  struct thread *thread __used)
1405 {
1406         struct thread *sched_out, *sched_in;
1407         int new_shortname;
1408         u64 timestamp0;
1409         s64 delta;
1410         int cpu;
1411
1412         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1413
1414         if (this_cpu > max_cpu)
1415                 max_cpu = this_cpu;
1416
1417         timestamp0 = cpu_last_switched[this_cpu];
1418         cpu_last_switched[this_cpu] = timestamp;
1419         if (timestamp0)
1420                 delta = timestamp - timestamp0;
1421         else
1422                 delta = 0;
1423
1424         if (delta < 0)
1425                 die("hm, delta: %Ld < 0 ?\n", delta);
1426
1427
1428         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1429         sched_in = perf_session__findnew(session, switch_event->next_pid);
1430
1431         curr_thread[this_cpu] = sched_in;
1432
1433         printf("  ");
1434
1435         new_shortname = 0;
1436         if (!sched_in->shortname[0]) {
1437                 sched_in->shortname[0] = next_shortname1;
1438                 sched_in->shortname[1] = next_shortname2;
1439
1440                 if (next_shortname1 < 'Z') {
1441                         next_shortname1++;
1442                 } else {
1443                         next_shortname1='A';
1444                         if (next_shortname2 < '9') {
1445                                 next_shortname2++;
1446                         } else {
1447                                 next_shortname2='0';
1448                         }
1449                 }
1450                 new_shortname = 1;
1451         }
1452
1453         for (cpu = 0; cpu <= max_cpu; cpu++) {
1454                 if (cpu != this_cpu)
1455                         printf(" ");
1456                 else
1457                         printf("*");
1458
1459                 if (curr_thread[cpu]) {
1460                         if (curr_thread[cpu]->pid)
1461                                 printf("%2s ", curr_thread[cpu]->shortname);
1462                         else
1463                                 printf(".  ");
1464                 } else
1465                         printf("   ");
1466         }
1467
1468         printf("  %12.6f secs ", (double)timestamp/1e9);
1469         if (new_shortname) {
1470                 printf("%s => %s:%d\n",
1471                         sched_in->shortname, sched_in->comm, sched_in->pid);
1472         } else {
1473                 printf("\n");
1474         }
1475 }
1476
1477
1478 static void
1479 process_sched_switch_event(void *data, struct perf_session *session,
1480                            struct event *event,
1481                            int this_cpu,
1482                            u64 timestamp __used,
1483                            struct thread *thread __used)
1484 {
1485         struct trace_switch_event switch_event;
1486
1487         FILL_COMMON_FIELDS(switch_event, event, data);
1488
1489         FILL_ARRAY(switch_event, prev_comm, event, data);
1490         FILL_FIELD(switch_event, prev_pid, event, data);
1491         FILL_FIELD(switch_event, prev_prio, event, data);
1492         FILL_FIELD(switch_event, prev_state, event, data);
1493         FILL_ARRAY(switch_event, next_comm, event, data);
1494         FILL_FIELD(switch_event, next_pid, event, data);
1495         FILL_FIELD(switch_event, next_prio, event, data);
1496
1497         if (curr_pid[this_cpu] != (u32)-1) {
1498                 /*
1499                  * Are we trying to switch away a PID that is
1500                  * not current?
1501                  */
1502                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1503                         nr_context_switch_bugs++;
1504         }
1505         if (trace_handler->switch_event)
1506                 trace_handler->switch_event(&switch_event, session, event,
1507                                             this_cpu, timestamp, thread);
1508
1509         curr_pid[this_cpu] = switch_event.next_pid;
1510 }
1511
1512 static void
1513 process_sched_runtime_event(void *data, struct perf_session *session,
1514                            struct event *event,
1515                            int cpu __used,
1516                            u64 timestamp __used,
1517                            struct thread *thread __used)
1518 {
1519         struct trace_runtime_event runtime_event;
1520
1521         FILL_ARRAY(runtime_event, comm, event, data);
1522         FILL_FIELD(runtime_event, pid, event, data);
1523         FILL_FIELD(runtime_event, runtime, event, data);
1524         FILL_FIELD(runtime_event, vruntime, event, data);
1525
1526         if (trace_handler->runtime_event)
1527                 trace_handler->runtime_event(&runtime_event, session, event, cpu, timestamp, thread);
1528 }
1529
1530 static void
1531 process_sched_fork_event(void *data,
1532                          struct event *event,
1533                          int cpu __used,
1534                          u64 timestamp __used,
1535                          struct thread *thread __used)
1536 {
1537         struct trace_fork_event fork_event;
1538
1539         FILL_COMMON_FIELDS(fork_event, event, data);
1540
1541         FILL_ARRAY(fork_event, parent_comm, event, data);
1542         FILL_FIELD(fork_event, parent_pid, event, data);
1543         FILL_ARRAY(fork_event, child_comm, event, data);
1544         FILL_FIELD(fork_event, child_pid, event, data);
1545
1546         if (trace_handler->fork_event)
1547                 trace_handler->fork_event(&fork_event, event,
1548                                           cpu, timestamp, thread);
1549 }
1550
1551 static void
1552 process_sched_exit_event(struct event *event,
1553                          int cpu __used,
1554                          u64 timestamp __used,
1555                          struct thread *thread __used)
1556 {
1557         if (verbose)
1558                 printf("sched_exit event %p\n", event);
1559 }
1560
1561 static void
1562 process_sched_migrate_task_event(void *data, struct perf_session *session,
1563                            struct event *event,
1564                            int cpu __used,
1565                            u64 timestamp __used,
1566                            struct thread *thread __used)
1567 {
1568         struct trace_migrate_task_event migrate_task_event;
1569
1570         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1571
1572         FILL_ARRAY(migrate_task_event, comm, event, data);
1573         FILL_FIELD(migrate_task_event, pid, event, data);
1574         FILL_FIELD(migrate_task_event, prio, event, data);
1575         FILL_FIELD(migrate_task_event, cpu, event, data);
1576
1577         if (trace_handler->migrate_task_event)
1578                 trace_handler->migrate_task_event(&migrate_task_event, session,
1579                                                  event, cpu, timestamp, thread);
1580 }
1581
1582 static void
1583 process_raw_event(event_t *raw_event __used, struct perf_session *session,
1584                   void *data, int cpu, u64 timestamp, struct thread *thread)
1585 {
1586         struct event *event;
1587         int type;
1588
1589
1590         type = trace_parse_common_type(data);
1591         event = trace_find_event(type);
1592
1593         if (!strcmp(event->name, "sched_switch"))
1594                 process_sched_switch_event(data, session, event, cpu, timestamp, thread);
1595         if (!strcmp(event->name, "sched_stat_runtime"))
1596                 process_sched_runtime_event(data, session, event, cpu, timestamp, thread);
1597         if (!strcmp(event->name, "sched_wakeup"))
1598                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1599         if (!strcmp(event->name, "sched_wakeup_new"))
1600                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1601         if (!strcmp(event->name, "sched_process_fork"))
1602                 process_sched_fork_event(data, event, cpu, timestamp, thread);
1603         if (!strcmp(event->name, "sched_process_exit"))
1604                 process_sched_exit_event(event, cpu, timestamp, thread);
1605         if (!strcmp(event->name, "sched_migrate_task"))
1606                 process_sched_migrate_task_event(data, session, event, cpu, timestamp, thread);
1607 }
1608
1609 static int process_sample_event(event_t *event, struct perf_session *session)
1610 {
1611         struct sample_data data;
1612         struct thread *thread;
1613
1614         if (!(session->sample_type & PERF_SAMPLE_RAW))
1615                 return 0;
1616
1617         memset(&data, 0, sizeof(data));
1618         data.time = -1;
1619         data.cpu = -1;
1620         data.period = -1;
1621
1622         event__parse_sample(event, session->sample_type, &data);
1623
1624         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1625                 event->header.misc,
1626                 data.pid, data.tid,
1627                 (void *)(long)data.ip,
1628                 (long long)data.period);
1629
1630         thread = perf_session__findnew(session, data.pid);
1631         if (thread == NULL) {
1632                 pr_debug("problem processing %d event, skipping it.\n",
1633                          event->header.type);
1634                 return -1;
1635         }
1636
1637         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1638
1639         if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1640                 return 0;
1641
1642         process_raw_event(event, session, data.raw_data, data.cpu, data.time, thread);
1643
1644         return 0;
1645 }
1646
1647 static int process_lost_event(event_t *event __used,
1648                               struct perf_session *session __used)
1649 {
1650         nr_lost_chunks++;
1651         nr_lost_events += event->lost.lost;
1652
1653         return 0;
1654 }
1655
1656 static int sample_type_check(struct perf_session *session __used)
1657 {
1658         if (!(session->sample_type & PERF_SAMPLE_RAW)) {
1659                 fprintf(stderr,
1660                         "No trace sample to read. Did you call perf record "
1661                         "without -R?");
1662                 return -1;
1663         }
1664
1665         return 0;
1666 }
1667
1668 static struct perf_event_ops event_ops = {
1669         .process_sample_event   = process_sample_event,
1670         .process_comm_event     = event__process_comm,
1671         .process_lost_event     = process_lost_event,
1672         .sample_type_check      = sample_type_check,
1673 };
1674
1675 static int read_events(void)
1676 {
1677         int err;
1678         struct perf_session *session = perf_session__new(input_name, O_RDONLY,
1679                                                          0, NULL);
1680         if (session == NULL)
1681                 return -ENOMEM;
1682
1683         err = perf_session__process_events(session, &event_ops);
1684         perf_session__delete(session);
1685         return err;
1686 }
1687
1688 static void print_bad_events(void)
1689 {
1690         if (nr_unordered_timestamps && nr_timestamps) {
1691                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1692                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1693                         nr_unordered_timestamps, nr_timestamps);
1694         }
1695         if (nr_lost_events && nr_events) {
1696                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1697                         (double)nr_lost_events/(double)nr_events*100.0,
1698                         nr_lost_events, nr_events, nr_lost_chunks);
1699         }
1700         if (nr_state_machine_bugs && nr_timestamps) {
1701                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1702                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1703                         nr_state_machine_bugs, nr_timestamps);
1704                 if (nr_lost_events)
1705                         printf(" (due to lost events?)");
1706                 printf("\n");
1707         }
1708         if (nr_context_switch_bugs && nr_timestamps) {
1709                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1710                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1711                         nr_context_switch_bugs, nr_timestamps);
1712                 if (nr_lost_events)
1713                         printf(" (due to lost events?)");
1714                 printf("\n");
1715         }
1716 }
1717
1718 static void __cmd_lat(void)
1719 {
1720         struct rb_node *next;
1721
1722         setup_pager();
1723         read_events();
1724         sort_lat();
1725
1726         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1727         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1728         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1729
1730         next = rb_first(&sorted_atom_root);
1731
1732         while (next) {
1733                 struct work_atoms *work_list;
1734
1735                 work_list = rb_entry(next, struct work_atoms, node);
1736                 output_lat_thread(work_list);
1737                 next = rb_next(next);
1738         }
1739
1740         printf(" -----------------------------------------------------------------------------------------\n");
1741         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1742                 (double)all_runtime/1e6, all_count);
1743
1744         printf(" ---------------------------------------------------\n");
1745
1746         print_bad_events();
1747         printf("\n");
1748
1749 }
1750
1751 static struct trace_sched_handler map_ops  = {
1752         .wakeup_event           = NULL,
1753         .switch_event           = map_switch_event,
1754         .runtime_event          = NULL,
1755         .fork_event             = NULL,
1756 };
1757
1758 static void __cmd_map(void)
1759 {
1760         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1761
1762         setup_pager();
1763         read_events();
1764         print_bad_events();
1765 }
1766
1767 static void __cmd_replay(void)
1768 {
1769         unsigned long i;
1770
1771         calibrate_run_measurement_overhead();
1772         calibrate_sleep_measurement_overhead();
1773
1774         test_calibrations();
1775
1776         read_events();
1777
1778         printf("nr_run_events:        %ld\n", nr_run_events);
1779         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1780         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1781
1782         if (targetless_wakeups)
1783                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1784         if (multitarget_wakeups)
1785                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1786         if (nr_run_events_optimized)
1787                 printf("run atoms optimized: %ld\n",
1788                         nr_run_events_optimized);
1789
1790         print_task_traces();
1791         add_cross_task_wakeups();
1792
1793         create_tasks();
1794         printf("------------------------------------------------------------\n");
1795         for (i = 0; i < replay_repeat; i++)
1796                 run_one_test();
1797 }
1798
1799
1800 static const char * const sched_usage[] = {
1801         "perf sched [<options>] {record|latency|map|replay|trace}",
1802         NULL
1803 };
1804
1805 static const struct option sched_options[] = {
1806         OPT_STRING('i', "input", &input_name, "file",
1807                     "input file name"),
1808         OPT_BOOLEAN('v', "verbose", &verbose,
1809                     "be more verbose (show symbol address, etc)"),
1810         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1811                     "dump raw trace in ASCII"),
1812         OPT_END()
1813 };
1814
1815 static const char * const latency_usage[] = {
1816         "perf sched latency [<options>]",
1817         NULL
1818 };
1819
1820 static const struct option latency_options[] = {
1821         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1822                    "sort by key(s): runtime, switch, avg, max"),
1823         OPT_BOOLEAN('v', "verbose", &verbose,
1824                     "be more verbose (show symbol address, etc)"),
1825         OPT_INTEGER('C', "CPU", &profile_cpu,
1826                     "CPU to profile on"),
1827         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1828                     "dump raw trace in ASCII"),
1829         OPT_END()
1830 };
1831
1832 static const char * const replay_usage[] = {
1833         "perf sched replay [<options>]",
1834         NULL
1835 };
1836
1837 static const struct option replay_options[] = {
1838         OPT_INTEGER('r', "repeat", &replay_repeat,
1839                     "repeat the workload replay N times (-1: infinite)"),
1840         OPT_BOOLEAN('v', "verbose", &verbose,
1841                     "be more verbose (show symbol address, etc)"),
1842         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1843                     "dump raw trace in ASCII"),
1844         OPT_END()
1845 };
1846
1847 static void setup_sorting(void)
1848 {
1849         char *tmp, *tok, *str = strdup(sort_order);
1850
1851         for (tok = strtok_r(str, ", ", &tmp);
1852                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1853                 if (sort_dimension__add(tok, &sort_list) < 0) {
1854                         error("Unknown --sort key: `%s'", tok);
1855                         usage_with_options(latency_usage, latency_options);
1856                 }
1857         }
1858
1859         free(str);
1860
1861         sort_dimension__add("pid", &cmp_pid);
1862 }
1863
1864 static const char *record_args[] = {
1865         "record",
1866         "-a",
1867         "-R",
1868         "-M",
1869         "-f",
1870         "-m", "1024",
1871         "-c", "1",
1872         "-e", "sched:sched_switch:r",
1873         "-e", "sched:sched_stat_wait:r",
1874         "-e", "sched:sched_stat_sleep:r",
1875         "-e", "sched:sched_stat_iowait:r",
1876         "-e", "sched:sched_stat_runtime:r",
1877         "-e", "sched:sched_process_exit:r",
1878         "-e", "sched:sched_process_fork:r",
1879         "-e", "sched:sched_wakeup:r",
1880         "-e", "sched:sched_migrate_task:r",
1881 };
1882
1883 static int __cmd_record(int argc, const char **argv)
1884 {
1885         unsigned int rec_argc, i, j;
1886         const char **rec_argv;
1887
1888         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1889         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1890
1891         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1892                 rec_argv[i] = strdup(record_args[i]);
1893
1894         for (j = 1; j < (unsigned int)argc; j++, i++)
1895                 rec_argv[i] = argv[j];
1896
1897         BUG_ON(i != rec_argc);
1898
1899         return cmd_record(i, rec_argv, NULL);
1900 }
1901
1902 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1903 {
1904         argc = parse_options(argc, argv, sched_options, sched_usage,
1905                              PARSE_OPT_STOP_AT_NON_OPTION);
1906         if (!argc)
1907                 usage_with_options(sched_usage, sched_options);
1908
1909         /*
1910          * Aliased to 'perf trace' for now:
1911          */
1912         if (!strcmp(argv[0], "trace"))
1913                 return cmd_trace(argc, argv, prefix);
1914
1915         symbol__init(0);
1916         if (!strncmp(argv[0], "rec", 3)) {
1917                 return __cmd_record(argc, argv);
1918         } else if (!strncmp(argv[0], "lat", 3)) {
1919                 trace_handler = &lat_ops;
1920                 if (argc > 1) {
1921                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1922                         if (argc)
1923                                 usage_with_options(latency_usage, latency_options);
1924                 }
1925                 setup_sorting();
1926                 __cmd_lat();
1927         } else if (!strcmp(argv[0], "map")) {
1928                 trace_handler = &map_ops;
1929                 setup_sorting();
1930                 __cmd_map();
1931         } else if (!strncmp(argv[0], "rep", 3)) {
1932                 trace_handler = &replay_ops;
1933                 if (argc) {
1934                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1935                         if (argc)
1936                                 usage_with_options(replay_usage, replay_options);
1937                 }
1938                 __cmd_replay();
1939         } else {
1940                 usage_with_options(sched_usage, sched_options);
1941         }
1942
1943         return 0;
1944 }