tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/evlist.h"
   6 #include "util/cache.h"
   7 #include "util/evsel.h"
   8 #include "util/symbol.h"
   9 #include "util/thread.h"
  10 #include "util/header.h"
  11 #include "util/session.h"
  12 #include "util/tool.h"
  13
  14 #include "util/parse-options.h"
  15 #include "util/trace-event.h"
  16
  17 #include "util/debug.h"
  18
  19 #include <sys/prctl.h>
  20 #include <sys/resource.h>
  21
  22 #include <semaphore.h>
  23 #include <pthread.h>
  24 #include <math.h>
  25
  26 static const char               *input_name;
  27
  28 static char                     default_sort_order[] = "avg, max, switch, runtime";
  29 static const char               *sort_order = default_sort_order;
  30
  31 static int                      profile_cpu = -1;
  32
  33 #define PR_SET_NAME             15               /* Set process name */
  34 #define MAX_CPUS                4096
  35
  36 static u64                      run_measurement_overhead;
  37 static u64                      sleep_measurement_overhead;
  38
  39 #define COMM_LEN                20
  40 #define SYM_LEN                 129
  41
  42 #define MAX_PID                 65536
  43
  44 static unsigned long            nr_tasks;
  45
  46 struct perf_sched {
  47         struct perf_tool    tool;
  48         struct perf_session *session;
  49 };
  50
  51 struct sched_atom;
  52
  53 struct task_desc {
  54         unsigned long           nr;
  55         unsigned long           pid;
  56         char                    comm[COMM_LEN];
  57
  58         unsigned long           nr_events;
  59         unsigned long           curr_event;
  60         struct sched_atom       **atoms;
  61
  62         pthread_t               thread;
  63         sem_t                   sleep_sem;
  64
  65         sem_t                   ready_for_work;
  66         sem_t                   work_done_sem;
  67
  68         u64                     cpu_usage;
  69 };
  70
  71 enum sched_event_type {
  72         SCHED_EVENT_RUN,
  73         SCHED_EVENT_SLEEP,
  74         SCHED_EVENT_WAKEUP,
  75         SCHED_EVENT_MIGRATION,
  76 };
  77
  78 struct sched_atom {
  79         enum sched_event_type   type;
  80         int                     specific_wait;
  81         u64                     timestamp;
  82         u64                     duration;
  83         unsigned long           nr;
  84         sem_t                   *wait_sem;
  85         struct task_desc        *wakee;
  86 };
  87
  88 static struct task_desc         *pid_to_task[MAX_PID];
  89
  90 static struct task_desc         **tasks;
  91
  92 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  93 static u64                      start_time;
  94
  95 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
  96
  97 static unsigned long            nr_run_events;
  98 static unsigned long            nr_sleep_events;
  99 static unsigned long            nr_wakeup_events;
 100
 101 static unsigned long            nr_sleep_corrections;
 102 static unsigned long            nr_run_events_optimized;
 103
 104 static unsigned long            targetless_wakeups;
 105 static unsigned long            multitarget_wakeups;
 106
 107 static u64                      cpu_usage;
 108 static u64                      runavg_cpu_usage;
 109 static u64                      parent_cpu_usage;
 110 static u64                      runavg_parent_cpu_usage;
 111
 112 static unsigned long            nr_runs;
 113 static u64                      sum_runtime;
 114 static u64                      sum_fluct;
 115 static u64                      run_avg;
 116
 117 static unsigned int             replay_repeat = 10;
 118 static unsigned long            nr_timestamps;
 119 static unsigned long            nr_unordered_timestamps;
 120 static unsigned long            nr_state_machine_bugs;
 121 static unsigned long            nr_context_switch_bugs;
 122 static unsigned long            nr_events;
 123 static unsigned long            nr_lost_chunks;
 124 static unsigned long            nr_lost_events;
 125
 126 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 127
 128 enum thread_state {
 129         THREAD_SLEEPING = 0,
 130         THREAD_WAIT_CPU,
 131         THREAD_SCHED_IN,
 132         THREAD_IGNORE
 133 };
 134
 135 struct work_atom {
 136         struct list_head        list;
 137         enum thread_state       state;
 138         u64                     sched_out_time;
 139         u64                     wake_up_time;
 140         u64                     sched_in_time;
 141         u64                     runtime;
 142 };
 143
 144 struct work_atoms {
 145         struct list_head        work_list;
 146         struct thread           *thread;
 147         struct rb_node          node;
 148         u64                     max_lat;
 149         u64                     max_lat_at;
 150         u64                     total_lat;
 151         u64                     nb_atoms;
 152         u64                     total_runtime;
 153 };
 154
 155 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 156
 157 static struct rb_root           atom_root, sorted_atom_root;
 158
 159 static u64                      all_runtime;
 160 static u64                      all_count;
 161
 162
 163 static u64 get_nsecs(void)
 164 {
 165         struct timespec ts;
 166
 167         clock_gettime(CLOCK_MONOTONIC, &ts);
 168
 169         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 170 }
 171
 172 static void burn_nsecs(u64 nsecs)
 173 {
 174         u64 T0 = get_nsecs(), T1;
 175
 176         do {
 177                 T1 = get_nsecs();
 178         } while (T1 + run_measurement_overhead < T0 + nsecs);
 179 }
 180
 181 static void sleep_nsecs(u64 nsecs)
 182 {
 183         struct timespec ts;
 184
 185         ts.tv_nsec = nsecs % 999999999;
 186         ts.tv_sec = nsecs / 999999999;
 187
 188         nanosleep(&ts, NULL);
 189 }
 190
 191 static void calibrate_run_measurement_overhead(void)
 192 {
 193         u64 T0, T1, delta, min_delta = 1000000000ULL;
 194         int i;
 195
 196         for (i = 0; i < 10; i++) {
 197                 T0 = get_nsecs();
 198                 burn_nsecs(0);
 199                 T1 = get_nsecs();
 200                 delta = T1-T0;
 201                 min_delta = min(min_delta, delta);
 202         }
 203         run_measurement_overhead = min_delta;
 204
 205         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 206 }
 207
 208 static void calibrate_sleep_measurement_overhead(void)
 209 {
 210         u64 T0, T1, delta, min_delta = 1000000000ULL;
 211         int i;
 212
 213         for (i = 0; i < 10; i++) {
 214                 T0 = get_nsecs();
 215                 sleep_nsecs(10000);
 216                 T1 = get_nsecs();
 217                 delta = T1-T0;
 218                 min_delta = min(min_delta, delta);
 219         }
 220         min_delta -= 10000;
 221         sleep_measurement_overhead = min_delta;
 222
 223         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 224 }
 225
 226 static struct sched_atom *
 227 get_new_event(struct task_desc *task, u64 timestamp)
 228 {
 229         struct sched_atom *event = zalloc(sizeof(*event));
 230         unsigned long idx = task->nr_events;
 231         size_t size;
 232
 233         event->timestamp = timestamp;
 234         event->nr = idx;
 235
 236         task->nr_events++;
 237         size = sizeof(struct sched_atom *) * task->nr_events;
 238         task->atoms = realloc(task->atoms, size);
 239         BUG_ON(!task->atoms);
 240
 241         task->atoms[idx] = event;
 242
 243         return event;
 244 }
 245
 246 static struct sched_atom *last_event(struct task_desc *task)
 247 {
 248         if (!task->nr_events)
 249                 return NULL;
 250
 251         return task->atoms[task->nr_events - 1];
 252 }
 253
 254 static void
 255 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 256 {
 257         struct sched_atom *event, *curr_event = last_event(task);
 258
 259         /*
 260          * optimize an existing RUN event by merging this one
 261          * to it:
 262          */
 263         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 264                 nr_run_events_optimized++;
 265                 curr_event->duration += duration;
 266                 return;
 267         }
 268
 269         event = get_new_event(task, timestamp);
 270
 271         event->type = SCHED_EVENT_RUN;
 272         event->duration = duration;
 273
 274         nr_run_events++;
 275 }
 276
 277 static void
 278 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 279                        struct task_desc *wakee)
 280 {
 281         struct sched_atom *event, *wakee_event;
 282
 283         event = get_new_event(task, timestamp);
 284         event->type = SCHED_EVENT_WAKEUP;
 285         event->wakee = wakee;
 286
 287         wakee_event = last_event(wakee);
 288         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 289                 targetless_wakeups++;
 290                 return;
 291         }
 292         if (wakee_event->wait_sem) {
 293                 multitarget_wakeups++;
 294                 return;
 295         }
 296
 297         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 298         sem_init(wakee_event->wait_sem, 0, 0);
 299         wakee_event->specific_wait = 1;
 300         event->wait_sem = wakee_event->wait_sem;
 301
 302         nr_wakeup_events++;
 303 }
 304
 305 static void
 306 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 307                       u64 task_state __used)
 308 {
 309         struct sched_atom *event = get_new_event(task, timestamp);
 310
 311         event->type = SCHED_EVENT_SLEEP;
 312
 313         nr_sleep_events++;
 314 }
 315
 316 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 317 {
 318         struct task_desc *task;
 319
 320         BUG_ON(pid >= MAX_PID);
 321
 322         task = pid_to_task[pid];
 323
 324         if (task)
 325                 return task;
 326
 327         task = zalloc(sizeof(*task));
 328         task->pid = pid;
 329         task->nr = nr_tasks;
 330         strcpy(task->comm, comm);
 331         /*
 332          * every task starts in sleeping state - this gets ignored
 333          * if there's no wakeup pointing to this sleep state:
 334          */
 335         add_sched_event_sleep(task, 0, 0);
 336
 337         pid_to_task[pid] = task;
 338         nr_tasks++;
 339         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 340         BUG_ON(!tasks);
 341         tasks[task->nr] = task;
 342
 343         if (verbose)
 344                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 345
 346         return task;
 347 }
 348
 349
 350 static void print_task_traces(void)
 351 {
 352         struct task_desc *task;
 353         unsigned long i;
 354
 355         for (i = 0; i < nr_tasks; i++) {
 356                 task = tasks[i];
 357                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 358                         task->nr, task->comm, task->pid, task->nr_events);
 359         }
 360 }
 361
 362 static void add_cross_task_wakeups(void)
 363 {
 364         struct task_desc *task1, *task2;
 365         unsigned long i, j;
 366
 367         for (i = 0; i < nr_tasks; i++) {
 368                 task1 = tasks[i];
 369                 j = i + 1;
 370                 if (j == nr_tasks)
 371                         j = 0;
 372                 task2 = tasks[j];
 373                 add_sched_event_wakeup(task1, 0, task2);
 374         }
 375 }
 376
 377 static void
 378 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 379 {
 380         int ret = 0;
 381
 382         switch (atom->type) {
 383                 case SCHED_EVENT_RUN:
 384                         burn_nsecs(atom->duration);
 385                         break;
 386                 case SCHED_EVENT_SLEEP:
 387                         if (atom->wait_sem)
 388                                 ret = sem_wait(atom->wait_sem);
 389                         BUG_ON(ret);
 390                         break;
 391                 case SCHED_EVENT_WAKEUP:
 392                         if (atom->wait_sem)
 393                                 ret = sem_post(atom->wait_sem);
 394                         BUG_ON(ret);
 395                         break;
 396                 case SCHED_EVENT_MIGRATION:
 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 int self_open_counters(void)
 419 {
 420         struct perf_event_attr attr;
 421         int fd;
 422
 423         memset(&attr, 0, sizeof(attr));
 424
 425         attr.type = PERF_TYPE_SOFTWARE;
 426         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 427
 428         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 429
 430         if (fd < 0)
 431                 die("Error: sys_perf_event_open() syscall returned"
 432                     "with %d (%s)\n", fd, strerror(errno));
 433         return fd;
 434 }
 435
 436 static u64 get_cpu_usage_nsec_self(int fd)
 437 {
 438         u64 runtime;
 439         int ret;
 440
 441         ret = read(fd, &runtime, sizeof(runtime));
 442         BUG_ON(ret != sizeof(runtime));
 443
 444         return runtime;
 445 }
 446
 447 static void *thread_func(void *ctx)
 448 {
 449         struct task_desc *this_task = ctx;
 450         u64 cpu_usage_0, cpu_usage_1;
 451         unsigned long i, ret;
 452         char comm2[22];
 453         int fd;
 454
 455         sprintf(comm2, ":%s", this_task->comm);
 456         prctl(PR_SET_NAME, comm2);
 457         fd = self_open_counters();
 458
 459 again:
 460         ret = sem_post(&this_task->ready_for_work);
 461         BUG_ON(ret);
 462         ret = pthread_mutex_lock(&start_work_mutex);
 463         BUG_ON(ret);
 464         ret = pthread_mutex_unlock(&start_work_mutex);
 465         BUG_ON(ret);
 466
 467         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 468
 469         for (i = 0; i < this_task->nr_events; i++) {
 470                 this_task->curr_event = i;
 471                 process_sched_event(this_task, this_task->atoms[i]);
 472         }
 473
 474         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 475         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 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,
 497                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 498         BUG_ON(err);
 499         err = pthread_mutex_lock(&start_work_mutex);
 500         BUG_ON(err);
 501         err = pthread_mutex_lock(&work_done_wait_mutex);
 502         BUG_ON(err);
 503         for (i = 0; i < nr_tasks; i++) {
 504                 task = tasks[i];
 505                 sem_init(&task->sleep_sem, 0, 0);
 506                 sem_init(&task->ready_for_work, 0, 0);
 507                 sem_init(&task->work_done_sem, 0, 0);
 508                 task->curr_event = 0;
 509                 err = pthread_create(&task->thread, &attr, thread_func, task);
 510                 BUG_ON(err);
 511         }
 512 }
 513
 514 static void wait_for_tasks(void)
 515 {
 516         u64 cpu_usage_0, cpu_usage_1;
 517         struct task_desc *task;
 518         unsigned long i, ret;
 519
 520         start_time = get_nsecs();
 521         cpu_usage = 0;
 522         pthread_mutex_unlock(&work_done_wait_mutex);
 523
 524         for (i = 0; i < nr_tasks; i++) {
 525                 task = tasks[i];
 526                 ret = sem_wait(&task->ready_for_work);
 527                 BUG_ON(ret);
 528                 sem_init(&task->ready_for_work, 0, 0);
 529         }
 530         ret = pthread_mutex_lock(&work_done_wait_mutex);
 531         BUG_ON(ret);
 532
 533         cpu_usage_0 = get_cpu_usage_nsec_parent();
 534
 535         pthread_mutex_unlock(&start_work_mutex);
 536
 537         for (i = 0; i < nr_tasks; i++) {
 538                 task = tasks[i];
 539                 ret = sem_wait(&task->work_done_sem);
 540                 BUG_ON(ret);
 541                 sem_init(&task->work_done_sem, 0, 0);
 542                 cpu_usage += task->cpu_usage;
 543                 task->cpu_usage = 0;
 544         }
 545
 546         cpu_usage_1 = get_cpu_usage_nsec_parent();
 547         if (!runavg_cpu_usage)
 548                 runavg_cpu_usage = cpu_usage;
 549         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 550
 551         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 552         if (!runavg_parent_cpu_usage)
 553                 runavg_parent_cpu_usage = parent_cpu_usage;
 554         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 555                                    parent_cpu_usage)/10;
 556
 557         ret = pthread_mutex_lock(&start_work_mutex);
 558         BUG_ON(ret);
 559
 560         for (i = 0; i < nr_tasks; i++) {
 561                 task = tasks[i];
 562                 sem_init(&task->sleep_sem, 0, 0);
 563                 task->curr_event = 0;
 564         }
 565 }
 566
 567 static void run_one_test(void)
 568 {
 569         u64 T0, T1, delta, avg_delta, fluct;
 570
 571         T0 = get_nsecs();
 572         wait_for_tasks();
 573         T1 = get_nsecs();
 574
 575         delta = T1 - T0;
 576         sum_runtime += delta;
 577         nr_runs++;
 578
 579         avg_delta = sum_runtime / nr_runs;
 580         if (delta < avg_delta)
 581                 fluct = avg_delta - delta;
 582         else
 583                 fluct = delta - avg_delta;
 584         sum_fluct += fluct;
 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 %" PRIu64 " nsecs\n", T1 - T0);
 624
 625         T0 = get_nsecs();
 626         sleep_nsecs(1e6);
 627         T1 = get_nsecs();
 628
 629         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 630 }
 631
 632 #define FILL_FIELD(ptr, field, event, data)     \
 633         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 634
 635 #define FILL_ARRAY(ptr, array, event, data)                     \
 636 do {                                                            \
 637         void *__array = raw_field_ptr(event, #array, data);     \
 638         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 639 } while(0)
 640
 641 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 642 do {                                                            \
 643         FILL_FIELD(ptr, common_type, event, data);              \
 644         FILL_FIELD(ptr, common_flags, event, data);             \
 645         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 646         FILL_FIELD(ptr, common_pid, event, data);               \
 647         FILL_FIELD(ptr, common_tgid, event, data);              \
 648 } while (0)
 649
 650
 651
 652 struct trace_switch_event {
 653         u32 size;
 654
 655         u16 common_type;
 656         u8 common_flags;
 657         u8 common_preempt_count;
 658         u32 common_pid;
 659         u32 common_tgid;
 660
 661         char prev_comm[16];
 662         u32 prev_pid;
 663         u32 prev_prio;
 664         u64 prev_state;
 665         char next_comm[16];
 666         u32 next_pid;
 667         u32 next_prio;
 668 };
 669
 670 struct trace_runtime_event {
 671         u32 size;
 672
 673         u16 common_type;
 674         u8 common_flags;
 675         u8 common_preempt_count;
 676         u32 common_pid;
 677         u32 common_tgid;
 678
 679         char comm[16];
 680         u32 pid;
 681         u64 runtime;
 682         u64 vruntime;
 683 };
 684
 685 struct trace_wakeup_event {
 686         u32 size;
 687
 688         u16 common_type;
 689         u8 common_flags;
 690         u8 common_preempt_count;
 691         u32 common_pid;
 692         u32 common_tgid;
 693
 694         char comm[16];
 695         u32 pid;
 696
 697         u32 prio;
 698         u32 success;
 699         u32 cpu;
 700 };
 701
 702 struct trace_fork_event {
 703         u32 size;
 704
 705         u16 common_type;
 706         u8 common_flags;
 707         u8 common_preempt_count;
 708         u32 common_pid;
 709         u32 common_tgid;
 710
 711         char parent_comm[16];
 712         u32 parent_pid;
 713         char child_comm[16];
 714         u32 child_pid;
 715 };
 716
 717 struct trace_migrate_task_event {
 718         u32 size;
 719
 720         u16 common_type;
 721         u8 common_flags;
 722         u8 common_preempt_count;
 723         u32 common_pid;
 724         u32 common_tgid;
 725
 726         char comm[16];
 727         u32 pid;
 728
 729         u32 prio;
 730         u32 cpu;
 731 };
 732
 733 struct trace_sched_handler {
 734         void (*switch_event)(struct trace_switch_event *,
 735                              struct machine *,
 736                              struct event_format *,
 737                              int cpu,
 738                              u64 timestamp,
 739                              struct thread *thread);
 740
 741         void (*runtime_event)(struct trace_runtime_event *,
 742                               struct machine *,
 743                               struct event_format *,
 744                               int cpu,
 745                               u64 timestamp,
 746                               struct thread *thread);
 747
 748         void (*wakeup_event)(struct trace_wakeup_event *,
 749                              struct machine *,
 750                              struct event_format *,
 751                              int cpu,
 752                              u64 timestamp,
 753                              struct thread *thread);
 754
 755         void (*fork_event)(struct trace_fork_event *,
 756                            struct event_format *,
 757                            int cpu,
 758                            u64 timestamp,
 759                            struct thread *thread);
 760
 761         void (*migrate_task_event)(struct trace_migrate_task_event *,
 762                            struct machine *machine,
 763                            struct event_format *,
 764                            int cpu,
 765                            u64 timestamp,
 766                            struct thread *thread);
 767 };
 768
 769
 770 static void
 771 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 772                     struct machine *machine __used,
 773                     struct event_format *event,
 774                     int cpu __used,
 775                     u64 timestamp __used,
 776                     struct thread *thread __used)
 777 {
 778         struct task_desc *waker, *wakee;
 779
 780         if (verbose) {
 781                 printf("sched_wakeup event %p\n", event);
 782
 783                 printf(" ... pid %d woke up %s/%d\n",
 784                         wakeup_event->common_pid,
 785                         wakeup_event->comm,
 786                         wakeup_event->pid);
 787         }
 788
 789         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 790         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 791
 792         add_sched_event_wakeup(waker, timestamp, wakee);
 793 }
 794
 795 static u64 cpu_last_switched[MAX_CPUS];
 796
 797 static void
 798 replay_switch_event(struct trace_switch_event *switch_event,
 799                     struct machine *machine __used,
 800                     struct event_format *event,
 801                     int cpu,
 802                     u64 timestamp,
 803                     struct thread *thread __used)
 804 {
 805         struct task_desc *prev, __used *next;
 806         u64 timestamp0;
 807         s64 delta;
 808
 809         if (verbose)
 810                 printf("sched_switch event %p\n", event);
 811
 812         if (cpu >= MAX_CPUS || cpu < 0)
 813                 return;
 814
 815         timestamp0 = cpu_last_switched[cpu];
 816         if (timestamp0)
 817                 delta = timestamp - timestamp0;
 818         else
 819                 delta = 0;
 820
 821         if (delta < 0)
 822                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 823
 824         if (verbose) {
 825                 printf(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 826                         switch_event->prev_comm, switch_event->prev_pid,
 827                         switch_event->next_comm, switch_event->next_pid,
 828                         delta);
 829         }
 830
 831         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 832         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 833
 834         cpu_last_switched[cpu] = timestamp;
 835
 836         add_sched_event_run(prev, timestamp, delta);
 837         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 838 }
 839
 840
 841 static void
 842 replay_fork_event(struct trace_fork_event *fork_event,
 843                   struct event_format *event,
 844                   int cpu __used,
 845                   u64 timestamp __used,
 846                   struct thread *thread __used)
 847 {
 848         if (verbose) {
 849                 printf("sched_fork event %p\n", event);
 850                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 851                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 852         }
 853         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 854         register_pid(fork_event->child_pid, fork_event->child_comm);
 855 }
 856
 857 static struct trace_sched_handler replay_ops  = {
 858         .wakeup_event           = replay_wakeup_event,
 859         .switch_event           = replay_switch_event,
 860         .fork_event             = replay_fork_event,
 861 };
 862
 863 struct sort_dimension {
 864         const char              *name;
 865         sort_fn_t               cmp;
 866         struct list_head        list;
 867 };
 868
 869 static LIST_HEAD(cmp_pid);
 870
 871 static int
 872 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 873 {
 874         struct sort_dimension *sort;
 875         int ret = 0;
 876
 877         BUG_ON(list_empty(list));
 878
 879         list_for_each_entry(sort, list, list) {
 880                 ret = sort->cmp(l, r);
 881                 if (ret)
 882                         return ret;
 883         }
 884
 885         return ret;
 886 }
 887
 888 static struct work_atoms *
 889 thread_atoms_search(struct rb_root *root, struct thread *thread,
 890                          struct list_head *sort_list)
 891 {
 892         struct rb_node *node = root->rb_node;
 893         struct work_atoms key = { .thread = thread };
 894
 895         while (node) {
 896                 struct work_atoms *atoms;
 897                 int cmp;
 898
 899                 atoms = container_of(node, struct work_atoms, node);
 900
 901                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 902                 if (cmp > 0)
 903                         node = node->rb_left;
 904                 else if (cmp < 0)
 905                         node = node->rb_right;
 906                 else {
 907                         BUG_ON(thread != atoms->thread);
 908                         return atoms;
 909                 }
 910         }
 911         return NULL;
 912 }
 913
 914 static void
 915 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 916                          struct list_head *sort_list)
 917 {
 918         struct rb_node **new = &(root->rb_node), *parent = NULL;
 919
 920         while (*new) {
 921                 struct work_atoms *this;
 922                 int cmp;
 923
 924                 this = container_of(*new, struct work_atoms, node);
 925                 parent = *new;
 926
 927                 cmp = thread_lat_cmp(sort_list, data, this);
 928
 929                 if (cmp > 0)
 930                         new = &((*new)->rb_left);
 931                 else
 932                         new = &((*new)->rb_right);
 933         }
 934
 935         rb_link_node(&data->node, parent, new);
 936         rb_insert_color(&data->node, root);
 937 }
 938
 939 static void thread_atoms_insert(struct thread *thread)
 940 {
 941         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 942         if (!atoms)
 943                 die("No memory");
 944
 945         atoms->thread = thread;
 946         INIT_LIST_HEAD(&atoms->work_list);
 947         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 948 }
 949
 950 static void
 951 latency_fork_event(struct trace_fork_event *fork_event __used,
 952                    struct event_format *event __used,
 953                    int cpu __used,
 954                    u64 timestamp __used,
 955                    struct thread *thread __used)
 956 {
 957         /* should insert the newcomer */
 958 }
 959
 960 __used
 961 static char sched_out_state(struct trace_switch_event *switch_event)
 962 {
 963         const char *str = TASK_STATE_TO_CHAR_STR;
 964
 965         return str[switch_event->prev_state];
 966 }
 967
 968 static void
 969 add_sched_out_event(struct work_atoms *atoms,
 970                     char run_state,
 971                     u64 timestamp)
 972 {
 973         struct work_atom *atom = zalloc(sizeof(*atom));
 974         if (!atom)
 975                 die("Non memory");
 976
 977         atom->sched_out_time = timestamp;
 978
 979         if (run_state == 'R') {
 980                 atom->state = THREAD_WAIT_CPU;
 981                 atom->wake_up_time = atom->sched_out_time;
 982         }
 983
 984         list_add_tail(&atom->list, &atoms->work_list);
 985 }
 986
 987 static void
 988 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
 989 {
 990         struct work_atom *atom;
 991
 992         BUG_ON(list_empty(&atoms->work_list));
 993
 994         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 995
 996         atom->runtime += delta;
 997         atoms->total_runtime += delta;
 998 }
 999
1000 static void
1001 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1002 {
1003         struct work_atom *atom;
1004         u64 delta;
1005
1006         if (list_empty(&atoms->work_list))
1007                 return;
1008
1009         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1010
1011         if (atom->state != THREAD_WAIT_CPU)
1012                 return;
1013
1014         if (timestamp < atom->wake_up_time) {
1015                 atom->state = THREAD_IGNORE;
1016                 return;
1017         }
1018
1019         atom->state = THREAD_SCHED_IN;
1020         atom->sched_in_time = timestamp;
1021
1022         delta = atom->sched_in_time - atom->wake_up_time;
1023         atoms->total_lat += delta;
1024         if (delta > atoms->max_lat) {
1025                 atoms->max_lat = delta;
1026                 atoms->max_lat_at = timestamp;
1027         }
1028         atoms->nb_atoms++;
1029 }
1030
1031 static void
1032 latency_switch_event(struct trace_switch_event *switch_event,
1033                      struct machine *machine,
1034                      struct event_format *event __used,
1035                      int cpu,
1036                      u64 timestamp,
1037                      struct thread *thread __used)
1038 {
1039         struct work_atoms *out_events, *in_events;
1040         struct thread *sched_out, *sched_in;
1041         u64 timestamp0;
1042         s64 delta;
1043
1044         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1045
1046         timestamp0 = cpu_last_switched[cpu];
1047         cpu_last_switched[cpu] = timestamp;
1048         if (timestamp0)
1049                 delta = timestamp - timestamp0;
1050         else
1051                 delta = 0;
1052
1053         if (delta < 0)
1054                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1055
1056
1057         sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
1058         sched_in = machine__findnew_thread(machine, switch_event->next_pid);
1059
1060         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1061         if (!out_events) {
1062                 thread_atoms_insert(sched_out);
1063                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1064                 if (!out_events)
1065                         die("out-event: Internal tree error");
1066         }
1067         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1068
1069         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1070         if (!in_events) {
1071                 thread_atoms_insert(sched_in);
1072                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1073                 if (!in_events)
1074                         die("in-event: Internal tree error");
1075                 /*
1076                  * Take came in we have not heard about yet,
1077                  * add in an initial atom in runnable state:
1078                  */
1079                 add_sched_out_event(in_events, 'R', timestamp);
1080         }
1081         add_sched_in_event(in_events, timestamp);
1082 }
1083
1084 static void
1085 latency_runtime_event(struct trace_runtime_event *runtime_event,
1086                      struct machine *machine,
1087                      struct event_format *event __used,
1088                      int cpu,
1089                      u64 timestamp,
1090                      struct thread *this_thread __used)
1091 {
1092         struct thread *thread = machine__findnew_thread(machine, runtime_event->pid);
1093         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1094
1095         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1096         if (!atoms) {
1097                 thread_atoms_insert(thread);
1098                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1099                 if (!atoms)
1100                         die("in-event: Internal tree error");
1101                 add_sched_out_event(atoms, 'R', timestamp);
1102         }
1103
1104         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1105 }
1106
1107 static void
1108 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1109                      struct machine *machine,
1110                      struct event_format *__event __used,
1111                      int cpu __used,
1112                      u64 timestamp,
1113                      struct thread *thread __used)
1114 {
1115         struct work_atoms *atoms;
1116         struct work_atom *atom;
1117         struct thread *wakee;
1118
1119         /* Note for later, it may be interesting to observe the failing cases */
1120         if (!wakeup_event->success)
1121                 return;
1122
1123         wakee = machine__findnew_thread(machine, wakeup_event->pid);
1124         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1125         if (!atoms) {
1126                 thread_atoms_insert(wakee);
1127                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1128                 if (!atoms)
1129                         die("wakeup-event: Internal tree error");
1130                 add_sched_out_event(atoms, 'S', timestamp);
1131         }
1132
1133         BUG_ON(list_empty(&atoms->work_list));
1134
1135         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1136
1137         /*
1138          * You WILL be missing events if you've recorded only
1139          * one CPU, or are only looking at only one, so don't
1140          * make useless noise.
1141          */
1142         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1143                 nr_state_machine_bugs++;
1144
1145         nr_timestamps++;
1146         if (atom->sched_out_time > timestamp) {
1147                 nr_unordered_timestamps++;
1148                 return;
1149         }
1150
1151         atom->state = THREAD_WAIT_CPU;
1152         atom->wake_up_time = timestamp;
1153 }
1154
1155 static void
1156 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1157                      struct machine *machine,
1158                      struct event_format *__event __used,
1159                      int cpu __used,
1160                      u64 timestamp,
1161                      struct thread *thread __used)
1162 {
1163         struct work_atoms *atoms;
1164         struct work_atom *atom;
1165         struct thread *migrant;
1166
1167         /*
1168          * Only need to worry about migration when profiling one CPU.
1169          */
1170         if (profile_cpu == -1)
1171                 return;
1172
1173         migrant = machine__findnew_thread(machine, migrate_task_event->pid);
1174         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1175         if (!atoms) {
1176                 thread_atoms_insert(migrant);
1177                 register_pid(migrant->pid, migrant->comm);
1178                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1179                 if (!atoms)
1180                         die("migration-event: Internal tree error");
1181                 add_sched_out_event(atoms, 'R', timestamp);
1182         }
1183
1184         BUG_ON(list_empty(&atoms->work_list));
1185
1186         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1187         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1188
1189         nr_timestamps++;
1190
1191         if (atom->sched_out_time > timestamp)
1192                 nr_unordered_timestamps++;
1193 }
1194
1195 static struct trace_sched_handler lat_ops  = {
1196         .wakeup_event           = latency_wakeup_event,
1197         .switch_event           = latency_switch_event,
1198         .runtime_event          = latency_runtime_event,
1199         .fork_event             = latency_fork_event,
1200         .migrate_task_event     = latency_migrate_task_event,
1201 };
1202
1203 static void output_lat_thread(struct work_atoms *work_list)
1204 {
1205         int i;
1206         int ret;
1207         u64 avg;
1208
1209         if (!work_list->nb_atoms)
1210                 return;
1211         /*
1212          * Ignore idle threads:
1213          */
1214         if (!strcmp(work_list->thread->comm, "swapper"))
1215                 return;
1216
1217         all_runtime += work_list->total_runtime;
1218         all_count += work_list->nb_atoms;
1219
1220         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1221
1222         for (i = 0; i < 24 - ret; i++)
1223                 printf(" ");
1224
1225         avg = work_list->total_lat / work_list->nb_atoms;
1226
1227         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1228               (double)work_list->total_runtime / 1e6,
1229                  work_list->nb_atoms, (double)avg / 1e6,
1230                  (double)work_list->max_lat / 1e6,
1231                  (double)work_list->max_lat_at / 1e9);
1232 }
1233
1234 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1235 {
1236         if (l->thread->pid < r->thread->pid)
1237                 return -1;
1238         if (l->thread->pid > r->thread->pid)
1239                 return 1;
1240
1241         return 0;
1242 }
1243
1244 static struct sort_dimension pid_sort_dimension = {
1245         .name                   = "pid",
1246         .cmp                    = pid_cmp,
1247 };
1248
1249 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1250 {
1251         u64 avgl, avgr;
1252
1253         if (!l->nb_atoms)
1254                 return -1;
1255
1256         if (!r->nb_atoms)
1257                 return 1;
1258
1259         avgl = l->total_lat / l->nb_atoms;
1260         avgr = r->total_lat / r->nb_atoms;
1261
1262         if (avgl < avgr)
1263                 return -1;
1264         if (avgl > avgr)
1265                 return 1;
1266
1267         return 0;
1268 }
1269
1270 static struct sort_dimension avg_sort_dimension = {
1271         .name                   = "avg",
1272         .cmp                    = avg_cmp,
1273 };
1274
1275 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1276 {
1277         if (l->max_lat < r->max_lat)
1278                 return -1;
1279         if (l->max_lat > r->max_lat)
1280                 return 1;
1281
1282         return 0;
1283 }
1284
1285 static struct sort_dimension max_sort_dimension = {
1286         .name                   = "max",
1287         .cmp                    = max_cmp,
1288 };
1289
1290 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1291 {
1292         if (l->nb_atoms < r->nb_atoms)
1293                 return -1;
1294         if (l->nb_atoms > r->nb_atoms)
1295                 return 1;
1296
1297         return 0;
1298 }
1299
1300 static struct sort_dimension switch_sort_dimension = {
1301         .name                   = "switch",
1302         .cmp                    = switch_cmp,
1303 };
1304
1305 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1306 {
1307         if (l->total_runtime < r->total_runtime)
1308                 return -1;
1309         if (l->total_runtime > r->total_runtime)
1310                 return 1;
1311
1312         return 0;
1313 }
1314
1315 static struct sort_dimension runtime_sort_dimension = {
1316         .name                   = "runtime",
1317         .cmp                    = runtime_cmp,
1318 };
1319
1320 static struct sort_dimension *available_sorts[] = {
1321         &pid_sort_dimension,
1322         &avg_sort_dimension,
1323         &max_sort_dimension,
1324         &switch_sort_dimension,
1325         &runtime_sort_dimension,
1326 };
1327
1328 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1329
1330 static LIST_HEAD(sort_list);
1331
1332 static int sort_dimension__add(const char *tok, struct list_head *list)
1333 {
1334         int i;
1335
1336         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1337                 if (!strcmp(available_sorts[i]->name, tok)) {
1338                         list_add_tail(&available_sorts[i]->list, list);
1339
1340                         return 0;
1341                 }
1342         }
1343
1344         return -1;
1345 }
1346
1347 static void setup_sorting(void);
1348
1349 static void sort_lat(void)
1350 {
1351         struct rb_node *node;
1352
1353         for (;;) {
1354                 struct work_atoms *data;
1355                 node = rb_first(&atom_root);
1356                 if (!node)
1357                         break;
1358
1359                 rb_erase(node, &atom_root);
1360                 data = rb_entry(node, struct work_atoms, node);
1361                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1362         }
1363 }
1364
1365 static struct trace_sched_handler *trace_handler;
1366
1367 static void
1368 process_sched_wakeup_event(struct perf_tool *tool __used,
1369                            struct event_format *event,
1370                            struct perf_sample *sample,
1371                            struct machine *machine,
1372                            struct thread *thread)
1373 {
1374         void *data = sample->raw_data;
1375         struct trace_wakeup_event wakeup_event;
1376
1377         FILL_COMMON_FIELDS(wakeup_event, event, data);
1378
1379         FILL_ARRAY(wakeup_event, comm, event, data);
1380         FILL_FIELD(wakeup_event, pid, event, data);
1381         FILL_FIELD(wakeup_event, prio, event, data);
1382         FILL_FIELD(wakeup_event, success, event, data);
1383         FILL_FIELD(wakeup_event, cpu, event, data);
1384
1385         if (trace_handler->wakeup_event)
1386                 trace_handler->wakeup_event(&wakeup_event, machine, event,
1387                                             sample->cpu, sample->time, thread);
1388 }
1389
1390 /*
1391  * Track the current task - that way we can know whether there's any
1392  * weird events, such as a task being switched away that is not current.
1393  */
1394 static int max_cpu;
1395
1396 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1397
1398 static struct thread *curr_thread[MAX_CPUS];
1399
1400 static char next_shortname1 = 'A';
1401 static char next_shortname2 = '0';
1402
1403 static void
1404 map_switch_event(struct trace_switch_event *switch_event,
1405                  struct machine *machine,
1406                  struct event_format *event __used,
1407                  int this_cpu,
1408                  u64 timestamp,
1409                  struct thread *thread __used)
1410 {
1411         struct thread *sched_out __used, *sched_in;
1412         int new_shortname;
1413         u64 timestamp0;
1414         s64 delta;
1415         int cpu;
1416
1417         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1418
1419         if (this_cpu > max_cpu)
1420                 max_cpu = this_cpu;
1421
1422         timestamp0 = cpu_last_switched[this_cpu];
1423         cpu_last_switched[this_cpu] = timestamp;
1424         if (timestamp0)
1425                 delta = timestamp - timestamp0;
1426         else
1427                 delta = 0;
1428
1429         if (delta < 0)
1430                 die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1431
1432
1433         sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
1434         sched_in = machine__findnew_thread(machine, switch_event->next_pid);
1435
1436         curr_thread[this_cpu] = sched_in;
1437
1438         printf("  ");
1439
1440         new_shortname = 0;
1441         if (!sched_in->shortname[0]) {
1442                 sched_in->shortname[0] = next_shortname1;
1443                 sched_in->shortname[1] = next_shortname2;
1444
1445                 if (next_shortname1 < 'Z') {
1446                         next_shortname1++;
1447                 } else {
1448                         next_shortname1='A';
1449                         if (next_shortname2 < '9') {
1450                                 next_shortname2++;
1451                         } else {
1452                                 next_shortname2='0';
1453                         }
1454                 }
1455                 new_shortname = 1;
1456         }
1457
1458         for (cpu = 0; cpu <= max_cpu; cpu++) {
1459                 if (cpu != this_cpu)
1460                         printf(" ");
1461                 else
1462                         printf("*");
1463
1464                 if (curr_thread[cpu]) {
1465                         if (curr_thread[cpu]->pid)
1466                                 printf("%2s ", curr_thread[cpu]->shortname);
1467                         else
1468                                 printf(".  ");
1469                 } else
1470                         printf("   ");
1471         }
1472
1473         printf("  %12.6f secs ", (double)timestamp/1e9);
1474         if (new_shortname) {
1475                 printf("%s => %s:%d\n",
1476                         sched_in->shortname, sched_in->comm, sched_in->pid);
1477         } else {
1478                 printf("\n");
1479         }
1480 }
1481
1482 static void
1483 process_sched_switch_event(struct perf_tool *tool __used,
1484                            struct event_format *event,
1485                            struct perf_sample *sample,
1486                            struct machine *machine,
1487                            struct thread *thread)
1488 {
1489         int this_cpu = sample->cpu;
1490         void *data = sample->raw_data;
1491         struct trace_switch_event switch_event;
1492
1493         FILL_COMMON_FIELDS(switch_event, event, data);
1494
1495         FILL_ARRAY(switch_event, prev_comm, event, data);
1496         FILL_FIELD(switch_event, prev_pid, event, data);
1497         FILL_FIELD(switch_event, prev_prio, event, data);
1498         FILL_FIELD(switch_event, prev_state, event, data);
1499         FILL_ARRAY(switch_event, next_comm, event, data);
1500         FILL_FIELD(switch_event, next_pid, event, data);
1501         FILL_FIELD(switch_event, next_prio, event, data);
1502
1503         if (curr_pid[this_cpu] != (u32)-1) {
1504                 /*
1505                  * Are we trying to switch away a PID that is
1506                  * not current?
1507                  */
1508                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1509                         nr_context_switch_bugs++;
1510         }
1511         if (trace_handler->switch_event)
1512                 trace_handler->switch_event(&switch_event, machine, event,
1513                                             this_cpu, sample->time, thread);
1514
1515         curr_pid[this_cpu] = switch_event.next_pid;
1516 }
1517
1518 static void
1519 process_sched_runtime_event(struct perf_tool *tool __used,
1520                             struct event_format *event,
1521                             struct perf_sample *sample,
1522                             struct machine *machine,
1523                             struct thread *thread)
1524 {
1525         void *data = sample->raw_data;
1526         struct trace_runtime_event runtime_event;
1527
1528         FILL_ARRAY(runtime_event, comm, event, data);
1529         FILL_FIELD(runtime_event, pid, event, data);
1530         FILL_FIELD(runtime_event, runtime, event, data);
1531         FILL_FIELD(runtime_event, vruntime, event, data);
1532
1533         if (trace_handler->runtime_event)
1534                 trace_handler->runtime_event(&runtime_event, machine, event,
1535                                              sample->cpu, sample->time, thread);
1536 }
1537
1538 static void
1539 process_sched_fork_event(struct perf_tool *tool __used,
1540                          struct event_format *event,
1541                          struct perf_sample *sample,
1542                          struct machine *machine __used,
1543                          struct thread *thread)
1544 {
1545         void *data = sample->raw_data;
1546         struct trace_fork_event fork_event;
1547
1548         FILL_COMMON_FIELDS(fork_event, event, data);
1549
1550         FILL_ARRAY(fork_event, parent_comm, event, data);
1551         FILL_FIELD(fork_event, parent_pid, event, data);
1552         FILL_ARRAY(fork_event, child_comm, event, data);
1553         FILL_FIELD(fork_event, child_pid, event, data);
1554
1555         if (trace_handler->fork_event)
1556                 trace_handler->fork_event(&fork_event, event,
1557                                           sample->cpu, sample->time, thread);
1558 }
1559
1560 static void
1561 process_sched_exit_event(struct perf_tool *tool __used,
1562                          struct event_format *event,
1563                          struct perf_sample *sample __used,
1564                          struct machine *machine __used,
1565                          struct thread *thread __used)
1566 {
1567         if (verbose)
1568                 printf("sched_exit event %p\n", event);
1569 }
1570
1571 static void
1572 process_sched_migrate_task_event(struct perf_tool *tool __used,
1573                                  struct event_format *event,
1574                                  struct perf_sample *sample,
1575                                  struct machine *machine,
1576                                  struct thread *thread)
1577 {
1578         void *data = sample->raw_data;
1579         struct trace_migrate_task_event migrate_task_event;
1580
1581         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1582
1583         FILL_ARRAY(migrate_task_event, comm, event, data);
1584         FILL_FIELD(migrate_task_event, pid, event, data);
1585         FILL_FIELD(migrate_task_event, prio, event, data);
1586         FILL_FIELD(migrate_task_event, cpu, event, data);
1587
1588         if (trace_handler->migrate_task_event)
1589                 trace_handler->migrate_task_event(&migrate_task_event, machine,
1590                                                   event, sample->cpu,
1591                                                   sample->time, thread);
1592 }
1593
1594 typedef void (*tracepoint_handler)(struct perf_tool *tool, struct event_format *event,
1595                                    struct perf_sample *sample,
1596                                    struct machine *machine,
1597                                    struct thread *thread);
1598
1599 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool,
1600                                                  union perf_event *event __used,
1601                                                  struct perf_sample *sample,
1602                                                  struct perf_evsel *evsel,
1603                                                  struct machine *machine)
1604 {
1605         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1606         struct pevent *pevent = sched->session->pevent;
1607         struct thread *thread = machine__findnew_thread(machine, sample->pid);
1608
1609         if (thread == NULL) {
1610                 pr_debug("problem processing %s event, skipping it.\n",
1611                          perf_evsel__name(evsel));
1612                 return -1;
1613         }
1614
1615         evsel->hists.stats.total_period += sample->period;
1616         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
1617
1618         if (evsel->handler.func != NULL) {
1619                 tracepoint_handler f = evsel->handler.func;
1620
1621                 if (evsel->handler.data == NULL)
1622                         evsel->handler.data = pevent_find_event(pevent,
1623                                                           evsel->attr.config);
1624
1625                 f(tool, evsel->handler.data, sample, machine, thread);
1626         }
1627
1628         return 0;
1629 }
1630
1631 static struct perf_sched perf_sched = {
1632         .tool = {
1633                 .sample          = perf_sched__process_tracepoint_sample,
1634                 .comm            = perf_event__process_comm,
1635                 .lost            = perf_event__process_lost,
1636                 .fork            = perf_event__process_task,
1637                 .ordered_samples = true,
1638         },
1639 };
1640
1641 static void read_events(bool destroy, struct perf_session **psession)
1642 {
1643         int err = -EINVAL;
1644         const struct perf_evsel_str_handler handlers[] = {
1645                 { "sched:sched_switch",       process_sched_switch_event, },
1646                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1647                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1648                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1649                 { "sched:sched_process_fork", process_sched_fork_event, },
1650                 { "sched:sched_process_exit", process_sched_exit_event, },
1651                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1652         };
1653         struct perf_session *session;
1654
1655         session = perf_session__new(input_name, O_RDONLY, 0, false,
1656                                     &perf_sched.tool);
1657         if (session == NULL)
1658                 die("No Memory");
1659
1660         perf_sched.session = session;
1661
1662         err = perf_session__set_tracepoints_handlers(session, handlers);
1663         assert(err == 0);
1664
1665         if (perf_session__has_traces(session, "record -R")) {
1666                 err = perf_session__process_events(session, &perf_sched.tool);
1667                 if (err)
1668                         die("Failed to process events, error %d", err);
1669
1670                 nr_events      = session->hists.stats.nr_events[0];
1671                 nr_lost_events = session->hists.stats.total_lost;
1672                 nr_lost_chunks = session->hists.stats.nr_events[PERF_RECORD_LOST];
1673         }
1674
1675         if (destroy)
1676                 perf_session__delete(session);
1677
1678         if (psession)
1679                 *psession = session;
1680 }
1681
1682 static void print_bad_events(void)
1683 {
1684         if (nr_unordered_timestamps && nr_timestamps) {
1685                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1686                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1687                         nr_unordered_timestamps, nr_timestamps);
1688         }
1689         if (nr_lost_events && nr_events) {
1690                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1691                         (double)nr_lost_events/(double)nr_events*100.0,
1692                         nr_lost_events, nr_events, nr_lost_chunks);
1693         }
1694         if (nr_state_machine_bugs && nr_timestamps) {
1695                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1696                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1697                         nr_state_machine_bugs, nr_timestamps);
1698                 if (nr_lost_events)
1699                         printf(" (due to lost events?)");
1700                 printf("\n");
1701         }
1702         if (nr_context_switch_bugs && nr_timestamps) {
1703                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1704                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1705                         nr_context_switch_bugs, nr_timestamps);
1706                 if (nr_lost_events)
1707                         printf(" (due to lost events?)");
1708                 printf("\n");
1709         }
1710 }
1711
1712 static void __cmd_lat(void)
1713 {
1714         struct rb_node *next;
1715         struct perf_session *session;
1716
1717         setup_pager();
1718         read_events(false, &session);
1719         sort_lat();
1720
1721         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1722         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1723         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1724
1725         next = rb_first(&sorted_atom_root);
1726
1727         while (next) {
1728                 struct work_atoms *work_list;
1729
1730                 work_list = rb_entry(next, struct work_atoms, node);
1731                 output_lat_thread(work_list);
1732                 next = rb_next(next);
1733         }
1734
1735         printf(" -----------------------------------------------------------------------------------------\n");
1736         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1737                 (double)all_runtime/1e6, all_count);
1738
1739         printf(" ---------------------------------------------------\n");
1740
1741         print_bad_events();
1742         printf("\n");
1743
1744         perf_session__delete(session);
1745 }
1746
1747 static struct trace_sched_handler map_ops  = {
1748         .wakeup_event           = NULL,
1749         .switch_event           = map_switch_event,
1750         .runtime_event          = NULL,
1751         .fork_event             = NULL,
1752 };
1753
1754 static void __cmd_map(void)
1755 {
1756         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1757
1758         setup_pager();
1759         read_events(true, NULL);
1760         print_bad_events();
1761 }
1762
1763 static void __cmd_replay(void)
1764 {
1765         unsigned long i;
1766
1767         calibrate_run_measurement_overhead();
1768         calibrate_sleep_measurement_overhead();
1769
1770         test_calibrations();
1771
1772         read_events(true, NULL);
1773
1774         printf("nr_run_events:        %ld\n", nr_run_events);
1775         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1776         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1777
1778         if (targetless_wakeups)
1779                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1780         if (multitarget_wakeups)
1781                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1782         if (nr_run_events_optimized)
1783                 printf("run atoms optimized: %ld\n",
1784                         nr_run_events_optimized);
1785
1786         print_task_traces();
1787         add_cross_task_wakeups();
1788
1789         create_tasks();
1790         printf("------------------------------------------------------------\n");
1791         for (i = 0; i < replay_repeat; i++)
1792                 run_one_test();
1793 }
1794
1795
1796 static const char * const sched_usage[] = {
1797         "perf sched [<options>] {record|latency|map|replay|script}",
1798         NULL
1799 };
1800
1801 static const struct option sched_options[] = {
1802         OPT_STRING('i', "input", &input_name, "file",
1803                     "input file name"),
1804         OPT_INCR('v', "verbose", &verbose,
1805                     "be more verbose (show symbol address, etc)"),
1806         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1807                     "dump raw trace in ASCII"),
1808         OPT_END()
1809 };
1810
1811 static const char * const latency_usage[] = {
1812         "perf sched latency [<options>]",
1813         NULL
1814 };
1815
1816 static const struct option latency_options[] = {
1817         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1818                    "sort by key(s): runtime, switch, avg, max"),
1819         OPT_INCR('v', "verbose", &verbose,
1820                     "be more verbose (show symbol address, etc)"),
1821         OPT_INTEGER('C', "CPU", &profile_cpu,
1822                     "CPU to profile on"),
1823         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1824                     "dump raw trace in ASCII"),
1825         OPT_END()
1826 };
1827
1828 static const char * const replay_usage[] = {
1829         "perf sched replay [<options>]",
1830         NULL
1831 };
1832
1833 static const struct option replay_options[] = {
1834         OPT_UINTEGER('r', "repeat", &replay_repeat,
1835                      "repeat the workload replay N times (-1: infinite)"),
1836         OPT_INCR('v', "verbose", &verbose,
1837                     "be more verbose (show symbol address, etc)"),
1838         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1839                     "dump raw trace in ASCII"),
1840         OPT_END()
1841 };
1842
1843 static void setup_sorting(void)
1844 {
1845         char *tmp, *tok, *str = strdup(sort_order);
1846
1847         for (tok = strtok_r(str, ", ", &tmp);
1848                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1849                 if (sort_dimension__add(tok, &sort_list) < 0) {
1850                         error("Unknown --sort key: `%s'", tok);
1851                         usage_with_options(latency_usage, latency_options);
1852                 }
1853         }
1854
1855         free(str);
1856
1857         sort_dimension__add("pid", &cmp_pid);
1858 }
1859
1860 static const char *record_args[] = {
1861         "record",
1862         "-a",
1863         "-R",
1864         "-f",
1865         "-m", "1024",
1866         "-c", "1",
1867         "-e", "sched:sched_switch",
1868         "-e", "sched:sched_stat_wait",
1869         "-e", "sched:sched_stat_sleep",
1870         "-e", "sched:sched_stat_iowait",
1871         "-e", "sched:sched_stat_runtime",
1872         "-e", "sched:sched_process_exit",
1873         "-e", "sched:sched_process_fork",
1874         "-e", "sched:sched_wakeup",
1875         "-e", "sched:sched_migrate_task",
1876 };
1877
1878 static int __cmd_record(int argc, const char **argv)
1879 {
1880         unsigned int rec_argc, i, j;
1881         const char **rec_argv;
1882
1883         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1884         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1885
1886         if (rec_argv == NULL)
1887                 return -ENOMEM;
1888
1889         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1890                 rec_argv[i] = strdup(record_args[i]);
1891
1892         for (j = 1; j < (unsigned int)argc; j++, i++)
1893                 rec_argv[i] = argv[j];
1894
1895         BUG_ON(i != rec_argc);
1896
1897         return cmd_record(i, rec_argv, NULL);
1898 }
1899
1900 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1901 {
1902         argc = parse_options(argc, argv, sched_options, sched_usage,
1903                              PARSE_OPT_STOP_AT_NON_OPTION);
1904         if (!argc)
1905                 usage_with_options(sched_usage, sched_options);
1906
1907         /*
1908          * Aliased to 'perf script' for now:
1909          */
1910         if (!strcmp(argv[0], "script"))
1911                 return cmd_script(argc, argv, prefix);
1912
1913         symbol__init();
1914         if (!strncmp(argv[0], "rec", 3)) {
1915                 return __cmd_record(argc, argv);
1916         } else if (!strncmp(argv[0], "lat", 3)) {
1917                 trace_handler = &lat_ops;
1918                 if (argc > 1) {
1919                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1920                         if (argc)
1921                                 usage_with_options(latency_usage, latency_options);
1922                 }
1923                 setup_sorting();
1924                 __cmd_lat();
1925         } else if (!strcmp(argv[0], "map")) {
1926                 trace_handler = &map_ops;
1927                 setup_sorting();
1928                 __cmd_map();
1929         } else if (!strncmp(argv[0], "rep", 3)) {
1930                 trace_handler = &replay_ops;
1931                 if (argc) {
1932                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1933                         if (argc)
1934                                 usage_with_options(replay_usage, replay_options);
1935                 }
1936                 __cmd_replay();
1937         } else {
1938                 usage_with_options(sched_usage, sched_options);
1939         }
1940
1941         return 0;
1942 }