Pileus Git - ~andy/linux/blob - tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/cache.h"
   6 #include "util/symbol.h"
   7 #include "util/thread.h"
   8 #include "util/header.h"
   9
  10 #include "util/parse-options.h"
  11 #include "util/trace-event.h"
  12
  13 #include "util/debug.h"
  14 #include "util/data_map.h"
  15
  16 #include <sys/types.h>
  17 #include <sys/prctl.h>
  18
  19 #include <semaphore.h>
  20 #include <pthread.h>
  21 #include <math.h>
  22
  23 static char                     const *input_name = "perf.data";
  24
  25 static unsigned long            total_comm = 0;
  26
  27 static struct rb_root           threads;
  28 static struct thread            *last_match;
  29
  30 static struct perf_header       *header;
  31 static u64                      sample_type;
  32
  33 static char                     default_sort_order[] = "avg, max, switch, runtime";
  34 static char                     *sort_order = default_sort_order;
  35
  36 static int                      profile_cpu = -1;
  37
  38 static char                     *cwd;
  39 static int                      cwdlen;
  40
  41 #define PR_SET_NAME             15               /* Set process name */
  42 #define MAX_CPUS                4096
  43
  44 #define BUG_ON(x)               assert(!(x))
  45
  46 static u64                      run_measurement_overhead;
  47 static u64                      sleep_measurement_overhead;
  48
  49 #define COMM_LEN                20
  50 #define SYM_LEN                 129
  51
  52 #define MAX_PID                 65536
  53
  54 static unsigned long            nr_tasks;
  55
  56 struct sched_atom;
  57
  58 struct task_desc {
  59         unsigned long           nr;
  60         unsigned long           pid;
  61         char                    comm[COMM_LEN];
  62
  63         unsigned long           nr_events;
  64         unsigned long           curr_event;
  65         struct sched_atom       **atoms;
  66
  67         pthread_t               thread;
  68         sem_t                   sleep_sem;
  69
  70         sem_t                   ready_for_work;
  71         sem_t                   work_done_sem;
  72
  73         u64                     cpu_usage;
  74 };
  75
  76 enum sched_event_type {
  77         SCHED_EVENT_RUN,
  78         SCHED_EVENT_SLEEP,
  79         SCHED_EVENT_WAKEUP,
  80         SCHED_EVENT_MIGRATION,
  81 };
  82
  83 struct sched_atom {
  84         enum sched_event_type   type;
  85         u64                     timestamp;
  86         u64                     duration;
  87         unsigned long           nr;
  88         int                     specific_wait;
  89         sem_t                   *wait_sem;
  90         struct task_desc        *wakee;
  91 };
  92
  93 static struct task_desc         *pid_to_task[MAX_PID];
  94
  95 static struct task_desc         **tasks;
  96
  97 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
  98 static u64                      start_time;
  99
 100 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
 101
 102 static unsigned long            nr_run_events;
 103 static unsigned long            nr_sleep_events;
 104 static unsigned long            nr_wakeup_events;
 105
 106 static unsigned long            nr_sleep_corrections;
 107 static unsigned long            nr_run_events_optimized;
 108
 109 static unsigned long            targetless_wakeups;
 110 static unsigned long            multitarget_wakeups;
 111
 112 static u64                      cpu_usage;
 113 static u64                      runavg_cpu_usage;
 114 static u64                      parent_cpu_usage;
 115 static u64                      runavg_parent_cpu_usage;
 116
 117 static unsigned long            nr_runs;
 118 static u64                      sum_runtime;
 119 static u64                      sum_fluct;
 120 static u64                      run_avg;
 121
 122 static unsigned long            replay_repeat = 10;
 123 static unsigned long            nr_timestamps;
 124 static unsigned long            nr_unordered_timestamps;
 125 static unsigned long            nr_state_machine_bugs;
 126 static unsigned long            nr_context_switch_bugs;
 127 static unsigned long            nr_events;
 128 static unsigned long            nr_lost_chunks;
 129 static unsigned long            nr_lost_events;
 130
 131 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
 132
 133 enum thread_state {
 134         THREAD_SLEEPING = 0,
 135         THREAD_WAIT_CPU,
 136         THREAD_SCHED_IN,
 137         THREAD_IGNORE
 138 };
 139
 140 struct work_atom {
 141         struct list_head        list;
 142         enum thread_state       state;
 143         u64                     sched_out_time;
 144         u64                     wake_up_time;
 145         u64                     sched_in_time;
 146         u64                     runtime;
 147 };
 148
 149 struct work_atoms {
 150         struct list_head        work_list;
 151         struct thread           *thread;
 152         struct rb_node          node;
 153         u64                     max_lat;
 154         u64                     total_lat;
 155         u64                     nb_atoms;
 156         u64                     total_runtime;
 157 };
 158
 159 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 160
 161 static struct rb_root           atom_root, sorted_atom_root;
 162
 163 static u64                      all_runtime;
 164 static u64                      all_count;
 165
 166
 167 static u64 get_nsecs(void)
 168 {
 169         struct timespec ts;
 170
 171         clock_gettime(CLOCK_MONOTONIC, &ts);
 172
 173         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 174 }
 175
 176 static void burn_nsecs(u64 nsecs)
 177 {
 178         u64 T0 = get_nsecs(), T1;
 179
 180         do {
 181                 T1 = get_nsecs();
 182         } while (T1 + run_measurement_overhead < T0 + nsecs);
 183 }
 184
 185 static void sleep_nsecs(u64 nsecs)
 186 {
 187         struct timespec ts;
 188
 189         ts.tv_nsec = nsecs % 999999999;
 190         ts.tv_sec = nsecs / 999999999;
 191
 192         nanosleep(&ts, NULL);
 193 }
 194
 195 static void calibrate_run_measurement_overhead(void)
 196 {
 197         u64 T0, T1, delta, min_delta = 1000000000ULL;
 198         int i;
 199
 200         for (i = 0; i < 10; i++) {
 201                 T0 = get_nsecs();
 202                 burn_nsecs(0);
 203                 T1 = get_nsecs();
 204                 delta = T1-T0;
 205                 min_delta = min(min_delta, delta);
 206         }
 207         run_measurement_overhead = min_delta;
 208
 209         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 210 }
 211
 212 static void calibrate_sleep_measurement_overhead(void)
 213 {
 214         u64 T0, T1, delta, min_delta = 1000000000ULL;
 215         int i;
 216
 217         for (i = 0; i < 10; i++) {
 218                 T0 = get_nsecs();
 219                 sleep_nsecs(10000);
 220                 T1 = get_nsecs();
 221                 delta = T1-T0;
 222                 min_delta = min(min_delta, delta);
 223         }
 224         min_delta -= 10000;
 225         sleep_measurement_overhead = min_delta;
 226
 227         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 228 }
 229
 230 static struct sched_atom *
 231 get_new_event(struct task_desc *task, u64 timestamp)
 232 {
 233         struct sched_atom *event = calloc(1, sizeof(*event));
 234         unsigned long idx = task->nr_events;
 235         size_t size;
 236
 237         event->timestamp = timestamp;
 238         event->nr = idx;
 239
 240         task->nr_events++;
 241         size = sizeof(struct sched_atom *) * task->nr_events;
 242         task->atoms = realloc(task->atoms, size);
 243         BUG_ON(!task->atoms);
 244
 245         task->atoms[idx] = event;
 246
 247         return event;
 248 }
 249
 250 static struct sched_atom *last_event(struct task_desc *task)
 251 {
 252         if (!task->nr_events)
 253                 return NULL;
 254
 255         return task->atoms[task->nr_events - 1];
 256 }
 257
 258 static void
 259 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 260 {
 261         struct sched_atom *event, *curr_event = last_event(task);
 262
 263         /*
 264          * optimize an existing RUN event by merging this one
 265          * to it:
 266          */
 267         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 268                 nr_run_events_optimized++;
 269                 curr_event->duration += duration;
 270                 return;
 271         }
 272
 273         event = get_new_event(task, timestamp);
 274
 275         event->type = SCHED_EVENT_RUN;
 276         event->duration = duration;
 277
 278         nr_run_events++;
 279 }
 280
 281 static void
 282 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
 283                        struct task_desc *wakee)
 284 {
 285         struct sched_atom *event, *wakee_event;
 286
 287         event = get_new_event(task, timestamp);
 288         event->type = SCHED_EVENT_WAKEUP;
 289         event->wakee = wakee;
 290
 291         wakee_event = last_event(wakee);
 292         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 293                 targetless_wakeups++;
 294                 return;
 295         }
 296         if (wakee_event->wait_sem) {
 297                 multitarget_wakeups++;
 298                 return;
 299         }
 300
 301         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
 302         sem_init(wakee_event->wait_sem, 0, 0);
 303         wakee_event->specific_wait = 1;
 304         event->wait_sem = wakee_event->wait_sem;
 305
 306         nr_wakeup_events++;
 307 }
 308
 309 static void
 310 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
 311                       u64 task_state __used)
 312 {
 313         struct sched_atom *event = get_new_event(task, timestamp);
 314
 315         event->type = SCHED_EVENT_SLEEP;
 316
 317         nr_sleep_events++;
 318 }
 319
 320 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 321 {
 322         struct task_desc *task;
 323
 324         BUG_ON(pid >= MAX_PID);
 325
 326         task = pid_to_task[pid];
 327
 328         if (task)
 329                 return task;
 330
 331         task = calloc(1, sizeof(*task));
 332         task->pid = pid;
 333         task->nr = nr_tasks;
 334         strcpy(task->comm, comm);
 335         /*
 336          * every task starts in sleeping state - this gets ignored
 337          * if there's no wakeup pointing to this sleep state:
 338          */
 339         add_sched_event_sleep(task, 0, 0);
 340
 341         pid_to_task[pid] = task;
 342         nr_tasks++;
 343         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 344         BUG_ON(!tasks);
 345         tasks[task->nr] = task;
 346
 347         if (verbose)
 348                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 349
 350         return task;
 351 }
 352
 353
 354 static void print_task_traces(void)
 355 {
 356         struct task_desc *task;
 357         unsigned long i;
 358
 359         for (i = 0; i < nr_tasks; i++) {
 360                 task = tasks[i];
 361                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 362                         task->nr, task->comm, task->pid, task->nr_events);
 363         }
 364 }
 365
 366 static void add_cross_task_wakeups(void)
 367 {
 368         struct task_desc *task1, *task2;
 369         unsigned long i, j;
 370
 371         for (i = 0; i < nr_tasks; i++) {
 372                 task1 = tasks[i];
 373                 j = i + 1;
 374                 if (j == nr_tasks)
 375                         j = 0;
 376                 task2 = tasks[j];
 377                 add_sched_event_wakeup(task1, 0, task2);
 378         }
 379 }
 380
 381 static void
 382 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 383 {
 384         int ret = 0;
 385         u64 now;
 386         long long delta;
 387
 388         now = get_nsecs();
 389         delta = start_time + atom->timestamp - now;
 390
 391         switch (atom->type) {
 392                 case SCHED_EVENT_RUN:
 393                         burn_nsecs(atom->duration);
 394                         break;
 395                 case SCHED_EVENT_SLEEP:
 396                         if (atom->wait_sem)
 397                                 ret = sem_wait(atom->wait_sem);
 398                         BUG_ON(ret);
 399                         break;
 400                 case SCHED_EVENT_WAKEUP:
 401                         if (atom->wait_sem)
 402                                 ret = sem_post(atom->wait_sem);
 403                         BUG_ON(ret);
 404                         break;
 405                 case SCHED_EVENT_MIGRATION:
 406                         break;
 407                 default:
 408                         BUG_ON(1);
 409         }
 410 }
 411
 412 static u64 get_cpu_usage_nsec_parent(void)
 413 {
 414         struct rusage ru;
 415         u64 sum;
 416         int err;
 417
 418         err = getrusage(RUSAGE_SELF, &ru);
 419         BUG_ON(err);
 420
 421         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 422         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 423
 424         return sum;
 425 }
 426
 427 static u64 get_cpu_usage_nsec_self(void)
 428 {
 429         char filename [] = "/proc/1234567890/sched";
 430         unsigned long msecs, nsecs;
 431         char *line = NULL;
 432         u64 total = 0;
 433         size_t len = 0;
 434         ssize_t chars;
 435         FILE *file;
 436         int ret;
 437
 438         sprintf(filename, "/proc/%d/sched", getpid());
 439         file = fopen(filename, "r");
 440         BUG_ON(!file);
 441
 442         while ((chars = getline(&line, &len, file)) != -1) {
 443                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 444                         &msecs, &nsecs);
 445                 if (ret == 2) {
 446                         total = msecs*1e6 + nsecs;
 447                         break;
 448                 }
 449         }
 450         if (line)
 451                 free(line);
 452         fclose(file);
 453
 454         return total;
 455 }
 456
 457 static void *thread_func(void *ctx)
 458 {
 459         struct task_desc *this_task = ctx;
 460         u64 cpu_usage_0, cpu_usage_1;
 461         unsigned long i, ret;
 462         char comm2[22];
 463
 464         sprintf(comm2, ":%s", this_task->comm);
 465         prctl(PR_SET_NAME, comm2);
 466
 467 again:
 468         ret = sem_post(&this_task->ready_for_work);
 469         BUG_ON(ret);
 470         ret = pthread_mutex_lock(&start_work_mutex);
 471         BUG_ON(ret);
 472         ret = pthread_mutex_unlock(&start_work_mutex);
 473         BUG_ON(ret);
 474
 475         cpu_usage_0 = get_cpu_usage_nsec_self();
 476
 477         for (i = 0; i < this_task->nr_events; i++) {
 478                 this_task->curr_event = i;
 479                 process_sched_event(this_task, this_task->atoms[i]);
 480         }
 481
 482         cpu_usage_1 = get_cpu_usage_nsec_self();
 483         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 484
 485         ret = sem_post(&this_task->work_done_sem);
 486         BUG_ON(ret);
 487
 488         ret = pthread_mutex_lock(&work_done_wait_mutex);
 489         BUG_ON(ret);
 490         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 491         BUG_ON(ret);
 492
 493         goto again;
 494 }
 495
 496 static void create_tasks(void)
 497 {
 498         struct task_desc *task;
 499         pthread_attr_t attr;
 500         unsigned long i;
 501         int err;
 502
 503         err = pthread_attr_init(&attr);
 504         BUG_ON(err);
 505         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 506         BUG_ON(err);
 507         err = pthread_mutex_lock(&start_work_mutex);
 508         BUG_ON(err);
 509         err = pthread_mutex_lock(&work_done_wait_mutex);
 510         BUG_ON(err);
 511         for (i = 0; i < nr_tasks; i++) {
 512                 task = tasks[i];
 513                 sem_init(&task->sleep_sem, 0, 0);
 514                 sem_init(&task->ready_for_work, 0, 0);
 515                 sem_init(&task->work_done_sem, 0, 0);
 516                 task->curr_event = 0;
 517                 err = pthread_create(&task->thread, &attr, thread_func, task);
 518                 BUG_ON(err);
 519         }
 520 }
 521
 522 static void wait_for_tasks(void)
 523 {
 524         u64 cpu_usage_0, cpu_usage_1;
 525         struct task_desc *task;
 526         unsigned long i, ret;
 527
 528         start_time = get_nsecs();
 529         cpu_usage = 0;
 530         pthread_mutex_unlock(&work_done_wait_mutex);
 531
 532         for (i = 0; i < nr_tasks; i++) {
 533                 task = tasks[i];
 534                 ret = sem_wait(&task->ready_for_work);
 535                 BUG_ON(ret);
 536                 sem_init(&task->ready_for_work, 0, 0);
 537         }
 538         ret = pthread_mutex_lock(&work_done_wait_mutex);
 539         BUG_ON(ret);
 540
 541         cpu_usage_0 = get_cpu_usage_nsec_parent();
 542
 543         pthread_mutex_unlock(&start_work_mutex);
 544
 545         for (i = 0; i < nr_tasks; i++) {
 546                 task = tasks[i];
 547                 ret = sem_wait(&task->work_done_sem);
 548                 BUG_ON(ret);
 549                 sem_init(&task->work_done_sem, 0, 0);
 550                 cpu_usage += task->cpu_usage;
 551                 task->cpu_usage = 0;
 552         }
 553
 554         cpu_usage_1 = get_cpu_usage_nsec_parent();
 555         if (!runavg_cpu_usage)
 556                 runavg_cpu_usage = cpu_usage;
 557         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 558
 559         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 560         if (!runavg_parent_cpu_usage)
 561                 runavg_parent_cpu_usage = parent_cpu_usage;
 562         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 563                                    parent_cpu_usage)/10;
 564
 565         ret = pthread_mutex_lock(&start_work_mutex);
 566         BUG_ON(ret);
 567
 568         for (i = 0; i < nr_tasks; i++) {
 569                 task = tasks[i];
 570                 sem_init(&task->sleep_sem, 0, 0);
 571                 task->curr_event = 0;
 572         }
 573 }
 574
 575 static void run_one_test(void)
 576 {
 577         u64 T0, T1, delta, avg_delta, fluct, std_dev;
 578
 579         T0 = get_nsecs();
 580         wait_for_tasks();
 581         T1 = get_nsecs();
 582
 583         delta = T1 - T0;
 584         sum_runtime += delta;
 585         nr_runs++;
 586
 587         avg_delta = sum_runtime / nr_runs;
 588         if (delta < avg_delta)
 589                 fluct = avg_delta - delta;
 590         else
 591                 fluct = delta - avg_delta;
 592         sum_fluct += fluct;
 593         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 594         if (!run_avg)
 595                 run_avg = delta;
 596         run_avg = (run_avg*9 + delta)/10;
 597
 598         printf("#%-3ld: %0.3f, ",
 599                 nr_runs, (double)delta/1000000.0);
 600
 601         printf("ravg: %0.2f, ",
 602                 (double)run_avg/1e6);
 603
 604         printf("cpu: %0.2f / %0.2f",
 605                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 606
 607 #if 0
 608         /*
 609          * rusage statistics done by the parent, these are less
 610          * accurate than the sum_exec_runtime based statistics:
 611          */
 612         printf(" [%0.2f / %0.2f]",
 613                 (double)parent_cpu_usage/1e6,
 614                 (double)runavg_parent_cpu_usage/1e6);
 615 #endif
 616
 617         printf("\n");
 618
 619         if (nr_sleep_corrections)
 620                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 621         nr_sleep_corrections = 0;
 622 }
 623
 624 static void test_calibrations(void)
 625 {
 626         u64 T0, T1;
 627
 628         T0 = get_nsecs();
 629         burn_nsecs(1e6);
 630         T1 = get_nsecs();
 631
 632         printf("the run test took %Ld nsecs\n", T1-T0);
 633
 634         T0 = get_nsecs();
 635         sleep_nsecs(1e6);
 636         T1 = get_nsecs();
 637
 638         printf("the sleep test took %Ld nsecs\n", T1-T0);
 639 }
 640
 641 static int
 642 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 643 {
 644         struct thread *thread;
 645
 646         thread = threads__findnew(event->comm.tid, &threads, &last_match);
 647
 648         dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
 649                 (void *)(offset + head),
 650                 (void *)(long)(event->header.size),
 651                 event->comm.comm, event->comm.pid);
 652
 653         if (thread == NULL ||
 654             thread__set_comm(thread, event->comm.comm)) {
 655                 dump_printf("problem processing perf_event_comm, skipping event.\n");
 656                 return -1;
 657         }
 658         total_comm++;
 659
 660         return 0;
 661 }
 662
 663
 664 struct raw_event_sample {
 665         u32 size;
 666         char data[0];
 667 };
 668
 669 #define FILL_FIELD(ptr, field, event, data)     \
 670         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 671
 672 #define FILL_ARRAY(ptr, array, event, data)                     \
 673 do {                                                            \
 674         void *__array = raw_field_ptr(event, #array, data);     \
 675         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 676 } while(0)
 677
 678 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 679 do {                                                            \
 680         FILL_FIELD(ptr, common_type, event, data);              \
 681         FILL_FIELD(ptr, common_flags, event, data);             \
 682         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 683         FILL_FIELD(ptr, common_pid, event, data);               \
 684         FILL_FIELD(ptr, common_tgid, event, data);              \
 685 } while (0)
 686
 687
 688
 689 struct trace_switch_event {
 690         u32 size;
 691
 692         u16 common_type;
 693         u8 common_flags;
 694         u8 common_preempt_count;
 695         u32 common_pid;
 696         u32 common_tgid;
 697
 698         char prev_comm[16];
 699         u32 prev_pid;
 700         u32 prev_prio;
 701         u64 prev_state;
 702         char next_comm[16];
 703         u32 next_pid;
 704         u32 next_prio;
 705 };
 706
 707 struct trace_runtime_event {
 708         u32 size;
 709
 710         u16 common_type;
 711         u8 common_flags;
 712         u8 common_preempt_count;
 713         u32 common_pid;
 714         u32 common_tgid;
 715
 716         char comm[16];
 717         u32 pid;
 718         u64 runtime;
 719         u64 vruntime;
 720 };
 721
 722 struct trace_wakeup_event {
 723         u32 size;
 724
 725         u16 common_type;
 726         u8 common_flags;
 727         u8 common_preempt_count;
 728         u32 common_pid;
 729         u32 common_tgid;
 730
 731         char comm[16];
 732         u32 pid;
 733
 734         u32 prio;
 735         u32 success;
 736         u32 cpu;
 737 };
 738
 739 struct trace_fork_event {
 740         u32 size;
 741
 742         u16 common_type;
 743         u8 common_flags;
 744         u8 common_preempt_count;
 745         u32 common_pid;
 746         u32 common_tgid;
 747
 748         char parent_comm[16];
 749         u32 parent_pid;
 750         char child_comm[16];
 751         u32 child_pid;
 752 };
 753
 754 struct trace_migrate_task_event {
 755         u32 size;
 756
 757         u16 common_type;
 758         u8 common_flags;
 759         u8 common_preempt_count;
 760         u32 common_pid;
 761         u32 common_tgid;
 762
 763         char comm[16];
 764         u32 pid;
 765
 766         u32 prio;
 767         u32 cpu;
 768 };
 769
 770 struct trace_sched_handler {
 771         void (*switch_event)(struct trace_switch_event *,
 772                              struct event *,
 773                              int cpu,
 774                              u64 timestamp,
 775                              struct thread *thread);
 776
 777         void (*runtime_event)(struct trace_runtime_event *,
 778                               struct event *,
 779                               int cpu,
 780                               u64 timestamp,
 781                               struct thread *thread);
 782
 783         void (*wakeup_event)(struct trace_wakeup_event *,
 784                              struct event *,
 785                              int cpu,
 786                              u64 timestamp,
 787                              struct thread *thread);
 788
 789         void (*fork_event)(struct trace_fork_event *,
 790                            struct event *,
 791                            int cpu,
 792                            u64 timestamp,
 793                            struct thread *thread);
 794
 795         void (*migrate_task_event)(struct trace_migrate_task_event *,
 796                            struct event *,
 797                            int cpu,
 798                            u64 timestamp,
 799                            struct thread *thread);
 800 };
 801
 802
 803 static void
 804 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
 805                     struct event *event,
 806                     int cpu __used,
 807                     u64 timestamp __used,
 808                     struct thread *thread __used)
 809 {
 810         struct task_desc *waker, *wakee;
 811
 812         if (verbose) {
 813                 printf("sched_wakeup event %p\n", event);
 814
 815                 printf(" ... pid %d woke up %s/%d\n",
 816                         wakeup_event->common_pid,
 817                         wakeup_event->comm,
 818                         wakeup_event->pid);
 819         }
 820
 821         waker = register_pid(wakeup_event->common_pid, "<unknown>");
 822         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
 823
 824         add_sched_event_wakeup(waker, timestamp, wakee);
 825 }
 826
 827 static u64 cpu_last_switched[MAX_CPUS];
 828
 829 static void
 830 replay_switch_event(struct trace_switch_event *switch_event,
 831                     struct event *event,
 832                     int cpu,
 833                     u64 timestamp,
 834                     struct thread *thread __used)
 835 {
 836         struct task_desc *prev, *next;
 837         u64 timestamp0;
 838         s64 delta;
 839
 840         if (verbose)
 841                 printf("sched_switch event %p\n", event);
 842
 843         if (cpu >= MAX_CPUS || cpu < 0)
 844                 return;
 845
 846         timestamp0 = cpu_last_switched[cpu];
 847         if (timestamp0)
 848                 delta = timestamp - timestamp0;
 849         else
 850                 delta = 0;
 851
 852         if (delta < 0)
 853                 die("hm, delta: %Ld < 0 ?\n", delta);
 854
 855         if (verbose) {
 856                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 857                         switch_event->prev_comm, switch_event->prev_pid,
 858                         switch_event->next_comm, switch_event->next_pid,
 859                         delta);
 860         }
 861
 862         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
 863         next = register_pid(switch_event->next_pid, switch_event->next_comm);
 864
 865         cpu_last_switched[cpu] = timestamp;
 866
 867         add_sched_event_run(prev, timestamp, delta);
 868         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
 869 }
 870
 871
 872 static void
 873 replay_fork_event(struct trace_fork_event *fork_event,
 874                   struct event *event,
 875                   int cpu __used,
 876                   u64 timestamp __used,
 877                   struct thread *thread __used)
 878 {
 879         if (verbose) {
 880                 printf("sched_fork event %p\n", event);
 881                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
 882                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
 883         }
 884         register_pid(fork_event->parent_pid, fork_event->parent_comm);
 885         register_pid(fork_event->child_pid, fork_event->child_comm);
 886 }
 887
 888 static struct trace_sched_handler replay_ops  = {
 889         .wakeup_event           = replay_wakeup_event,
 890         .switch_event           = replay_switch_event,
 891         .fork_event             = replay_fork_event,
 892 };
 893
 894 struct sort_dimension {
 895         const char              *name;
 896         sort_fn_t               cmp;
 897         struct list_head        list;
 898 };
 899
 900 static LIST_HEAD(cmp_pid);
 901
 902 static int
 903 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 904 {
 905         struct sort_dimension *sort;
 906         int ret = 0;
 907
 908         BUG_ON(list_empty(list));
 909
 910         list_for_each_entry(sort, list, list) {
 911                 ret = sort->cmp(l, r);
 912                 if (ret)
 913                         return ret;
 914         }
 915
 916         return ret;
 917 }
 918
 919 static struct work_atoms *
 920 thread_atoms_search(struct rb_root *root, struct thread *thread,
 921                          struct list_head *sort_list)
 922 {
 923         struct rb_node *node = root->rb_node;
 924         struct work_atoms key = { .thread = thread };
 925
 926         while (node) {
 927                 struct work_atoms *atoms;
 928                 int cmp;
 929
 930                 atoms = container_of(node, struct work_atoms, node);
 931
 932                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 933                 if (cmp > 0)
 934                         node = node->rb_left;
 935                 else if (cmp < 0)
 936                         node = node->rb_right;
 937                 else {
 938                         BUG_ON(thread != atoms->thread);
 939                         return atoms;
 940                 }
 941         }
 942         return NULL;
 943 }
 944
 945 static void
 946 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 947                          struct list_head *sort_list)
 948 {
 949         struct rb_node **new = &(root->rb_node), *parent = NULL;
 950
 951         while (*new) {
 952                 struct work_atoms *this;
 953                 int cmp;
 954
 955                 this = container_of(*new, struct work_atoms, node);
 956                 parent = *new;
 957
 958                 cmp = thread_lat_cmp(sort_list, data, this);
 959
 960                 if (cmp > 0)
 961                         new = &((*new)->rb_left);
 962                 else
 963                         new = &((*new)->rb_right);
 964         }
 965
 966         rb_link_node(&data->node, parent, new);
 967         rb_insert_color(&data->node, root);
 968 }
 969
 970 static void thread_atoms_insert(struct thread *thread)
 971 {
 972         struct work_atoms *atoms;
 973
 974         atoms = calloc(sizeof(*atoms), 1);
 975         if (!atoms)
 976                 die("No memory");
 977
 978         atoms->thread = thread;
 979         INIT_LIST_HEAD(&atoms->work_list);
 980         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 981 }
 982
 983 static void
 984 latency_fork_event(struct trace_fork_event *fork_event __used,
 985                    struct event *event __used,
 986                    int cpu __used,
 987                    u64 timestamp __used,
 988                    struct thread *thread __used)
 989 {
 990         /* should insert the newcomer */
 991 }
 992
 993 __used
 994 static char sched_out_state(struct trace_switch_event *switch_event)
 995 {
 996         const char *str = TASK_STATE_TO_CHAR_STR;
 997
 998         return str[switch_event->prev_state];
 999 }
1000
1001 static void
1002 add_sched_out_event(struct work_atoms *atoms,
1003                     char run_state,
1004                     u64 timestamp)
1005 {
1006         struct work_atom *atom;
1007
1008         atom = calloc(sizeof(*atom), 1);
1009         if (!atom)
1010                 die("Non memory");
1011
1012         atom->sched_out_time = timestamp;
1013
1014         if (run_state == 'R') {
1015                 atom->state = THREAD_WAIT_CPU;
1016                 atom->wake_up_time = atom->sched_out_time;
1017         }
1018
1019         list_add_tail(&atom->list, &atoms->work_list);
1020 }
1021
1022 static void
1023 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
1024 {
1025         struct work_atom *atom;
1026
1027         BUG_ON(list_empty(&atoms->work_list));
1028
1029         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1030
1031         atom->runtime += delta;
1032         atoms->total_runtime += delta;
1033 }
1034
1035 static void
1036 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1037 {
1038         struct work_atom *atom;
1039         u64 delta;
1040
1041         if (list_empty(&atoms->work_list))
1042                 return;
1043
1044         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1045
1046         if (atom->state != THREAD_WAIT_CPU)
1047                 return;
1048
1049         if (timestamp < atom->wake_up_time) {
1050                 atom->state = THREAD_IGNORE;
1051                 return;
1052         }
1053
1054         atom->state = THREAD_SCHED_IN;
1055         atom->sched_in_time = timestamp;
1056
1057         delta = atom->sched_in_time - atom->wake_up_time;
1058         atoms->total_lat += delta;
1059         if (delta > atoms->max_lat)
1060                 atoms->max_lat = delta;
1061         atoms->nb_atoms++;
1062 }
1063
1064 static void
1065 latency_switch_event(struct trace_switch_event *switch_event,
1066                      struct event *event __used,
1067                      int cpu,
1068                      u64 timestamp,
1069                      struct thread *thread __used)
1070 {
1071         struct work_atoms *out_events, *in_events;
1072         struct thread *sched_out, *sched_in;
1073         u64 timestamp0;
1074         s64 delta;
1075
1076         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1077
1078         timestamp0 = cpu_last_switched[cpu];
1079         cpu_last_switched[cpu] = timestamp;
1080         if (timestamp0)
1081                 delta = timestamp - timestamp0;
1082         else
1083                 delta = 0;
1084
1085         if (delta < 0)
1086                 die("hm, delta: %Ld < 0 ?\n", delta);
1087
1088
1089         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1090         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1091
1092         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1093         if (!out_events) {
1094                 thread_atoms_insert(sched_out);
1095                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1096                 if (!out_events)
1097                         die("out-event: Internal tree error");
1098         }
1099         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1100
1101         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1102         if (!in_events) {
1103                 thread_atoms_insert(sched_in);
1104                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1105                 if (!in_events)
1106                         die("in-event: Internal tree error");
1107                 /*
1108                  * Take came in we have not heard about yet,
1109                  * add in an initial atom in runnable state:
1110                  */
1111                 add_sched_out_event(in_events, 'R', timestamp);
1112         }
1113         add_sched_in_event(in_events, timestamp);
1114 }
1115
1116 static void
1117 latency_runtime_event(struct trace_runtime_event *runtime_event,
1118                      struct event *event __used,
1119                      int cpu,
1120                      u64 timestamp,
1121                      struct thread *this_thread __used)
1122 {
1123         struct work_atoms *atoms;
1124         struct thread *thread;
1125
1126         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1127
1128         thread = threads__findnew(runtime_event->pid, &threads, &last_match);
1129         atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1130         if (!atoms) {
1131                 thread_atoms_insert(thread);
1132                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1133                 if (!atoms)
1134                         die("in-event: Internal tree error");
1135                 add_sched_out_event(atoms, 'R', timestamp);
1136         }
1137
1138         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1139 }
1140
1141 static void
1142 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1143                      struct event *__event __used,
1144                      int cpu __used,
1145                      u64 timestamp,
1146                      struct thread *thread __used)
1147 {
1148         struct work_atoms *atoms;
1149         struct work_atom *atom;
1150         struct thread *wakee;
1151
1152         /* Note for later, it may be interesting to observe the failing cases */
1153         if (!wakeup_event->success)
1154                 return;
1155
1156         wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1157         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1158         if (!atoms) {
1159                 thread_atoms_insert(wakee);
1160                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1161                 if (!atoms)
1162                         die("wakeup-event: Internal tree error");
1163                 add_sched_out_event(atoms, 'S', timestamp);
1164         }
1165
1166         BUG_ON(list_empty(&atoms->work_list));
1167
1168         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1169
1170         /*
1171          * You WILL be missing events if you've recorded only
1172          * one CPU, or are only looking at only one, so don't
1173          * make useless noise.
1174          */
1175         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1176                 nr_state_machine_bugs++;
1177
1178         nr_timestamps++;
1179         if (atom->sched_out_time > timestamp) {
1180                 nr_unordered_timestamps++;
1181                 return;
1182         }
1183
1184         atom->state = THREAD_WAIT_CPU;
1185         atom->wake_up_time = timestamp;
1186 }
1187
1188 static void
1189 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1190                      struct event *__event __used,
1191                      int cpu __used,
1192                      u64 timestamp,
1193                      struct thread *thread __used)
1194 {
1195         struct work_atoms *atoms;
1196         struct work_atom *atom;
1197         struct thread *migrant;
1198
1199         /*
1200          * Only need to worry about migration when profiling one CPU.
1201          */
1202         if (profile_cpu == -1)
1203                 return;
1204
1205         migrant = threads__findnew(migrate_task_event->pid, &threads, &last_match);
1206         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1207         if (!atoms) {
1208                 thread_atoms_insert(migrant);
1209                 register_pid(migrant->pid, migrant->comm);
1210                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1211                 if (!atoms)
1212                         die("migration-event: Internal tree error");
1213                 add_sched_out_event(atoms, 'R', timestamp);
1214         }
1215
1216         BUG_ON(list_empty(&atoms->work_list));
1217
1218         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1219         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1220
1221         nr_timestamps++;
1222
1223         if (atom->sched_out_time > timestamp)
1224                 nr_unordered_timestamps++;
1225 }
1226
1227 static struct trace_sched_handler lat_ops  = {
1228         .wakeup_event           = latency_wakeup_event,
1229         .switch_event           = latency_switch_event,
1230         .runtime_event          = latency_runtime_event,
1231         .fork_event             = latency_fork_event,
1232         .migrate_task_event     = latency_migrate_task_event,
1233 };
1234
1235 static void output_lat_thread(struct work_atoms *work_list)
1236 {
1237         int i;
1238         int ret;
1239         u64 avg;
1240
1241         if (!work_list->nb_atoms)
1242                 return;
1243         /*
1244          * Ignore idle threads:
1245          */
1246         if (!strcmp(work_list->thread->comm, "swapper"))
1247                 return;
1248
1249         all_runtime += work_list->total_runtime;
1250         all_count += work_list->nb_atoms;
1251
1252         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1253
1254         for (i = 0; i < 24 - ret; i++)
1255                 printf(" ");
1256
1257         avg = work_list->total_lat / work_list->nb_atoms;
1258
1259         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1260               (double)work_list->total_runtime / 1e6,
1261                  work_list->nb_atoms, (double)avg / 1e6,
1262                  (double)work_list->max_lat / 1e6);
1263 }
1264
1265 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1266 {
1267         if (l->thread->pid < r->thread->pid)
1268                 return -1;
1269         if (l->thread->pid > r->thread->pid)
1270                 return 1;
1271
1272         return 0;
1273 }
1274
1275 static struct sort_dimension pid_sort_dimension = {
1276         .name                   = "pid",
1277         .cmp                    = pid_cmp,
1278 };
1279
1280 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1281 {
1282         u64 avgl, avgr;
1283
1284         if (!l->nb_atoms)
1285                 return -1;
1286
1287         if (!r->nb_atoms)
1288                 return 1;
1289
1290         avgl = l->total_lat / l->nb_atoms;
1291         avgr = r->total_lat / r->nb_atoms;
1292
1293         if (avgl < avgr)
1294                 return -1;
1295         if (avgl > avgr)
1296                 return 1;
1297
1298         return 0;
1299 }
1300
1301 static struct sort_dimension avg_sort_dimension = {
1302         .name                   = "avg",
1303         .cmp                    = avg_cmp,
1304 };
1305
1306 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1307 {
1308         if (l->max_lat < r->max_lat)
1309                 return -1;
1310         if (l->max_lat > r->max_lat)
1311                 return 1;
1312
1313         return 0;
1314 }
1315
1316 static struct sort_dimension max_sort_dimension = {
1317         .name                   = "max",
1318         .cmp                    = max_cmp,
1319 };
1320
1321 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1322 {
1323         if (l->nb_atoms < r->nb_atoms)
1324                 return -1;
1325         if (l->nb_atoms > r->nb_atoms)
1326                 return 1;
1327
1328         return 0;
1329 }
1330
1331 static struct sort_dimension switch_sort_dimension = {
1332         .name                   = "switch",
1333         .cmp                    = switch_cmp,
1334 };
1335
1336 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1337 {
1338         if (l->total_runtime < r->total_runtime)
1339                 return -1;
1340         if (l->total_runtime > r->total_runtime)
1341                 return 1;
1342
1343         return 0;
1344 }
1345
1346 static struct sort_dimension runtime_sort_dimension = {
1347         .name                   = "runtime",
1348         .cmp                    = runtime_cmp,
1349 };
1350
1351 static struct sort_dimension *available_sorts[] = {
1352         &pid_sort_dimension,
1353         &avg_sort_dimension,
1354         &max_sort_dimension,
1355         &switch_sort_dimension,
1356         &runtime_sort_dimension,
1357 };
1358
1359 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1360
1361 static LIST_HEAD(sort_list);
1362
1363 static int sort_dimension__add(const char *tok, struct list_head *list)
1364 {
1365         int i;
1366
1367         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1368                 if (!strcmp(available_sorts[i]->name, tok)) {
1369                         list_add_tail(&available_sorts[i]->list, list);
1370
1371                         return 0;
1372                 }
1373         }
1374
1375         return -1;
1376 }
1377
1378 static void setup_sorting(void);
1379
1380 static void sort_lat(void)
1381 {
1382         struct rb_node *node;
1383
1384         for (;;) {
1385                 struct work_atoms *data;
1386                 node = rb_first(&atom_root);
1387                 if (!node)
1388                         break;
1389
1390                 rb_erase(node, &atom_root);
1391                 data = rb_entry(node, struct work_atoms, node);
1392                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1393         }
1394 }
1395
1396 static struct trace_sched_handler *trace_handler;
1397
1398 static void
1399 process_sched_wakeup_event(struct raw_event_sample *raw,
1400                            struct event *event,
1401                            int cpu __used,
1402                            u64 timestamp __used,
1403                            struct thread *thread __used)
1404 {
1405         struct trace_wakeup_event wakeup_event;
1406
1407         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1408
1409         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1410         FILL_FIELD(wakeup_event, pid, event, raw->data);
1411         FILL_FIELD(wakeup_event, prio, event, raw->data);
1412         FILL_FIELD(wakeup_event, success, event, raw->data);
1413         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1414
1415         if (trace_handler->wakeup_event)
1416                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1417 }
1418
1419 /*
1420  * Track the current task - that way we can know whether there's any
1421  * weird events, such as a task being switched away that is not current.
1422  */
1423 static int max_cpu;
1424
1425 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1426
1427 static struct thread *curr_thread[MAX_CPUS];
1428
1429 static char next_shortname1 = 'A';
1430 static char next_shortname2 = '0';
1431
1432 static void
1433 map_switch_event(struct trace_switch_event *switch_event,
1434                  struct event *event __used,
1435                  int this_cpu,
1436                  u64 timestamp,
1437                  struct thread *thread __used)
1438 {
1439         struct thread *sched_out, *sched_in;
1440         int new_shortname;
1441         u64 timestamp0;
1442         s64 delta;
1443         int cpu;
1444
1445         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1446
1447         if (this_cpu > max_cpu)
1448                 max_cpu = this_cpu;
1449
1450         timestamp0 = cpu_last_switched[this_cpu];
1451         cpu_last_switched[this_cpu] = timestamp;
1452         if (timestamp0)
1453                 delta = timestamp - timestamp0;
1454         else
1455                 delta = 0;
1456
1457         if (delta < 0)
1458                 die("hm, delta: %Ld < 0 ?\n", delta);
1459
1460
1461         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1462         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1463
1464         curr_thread[this_cpu] = sched_in;
1465
1466         printf("  ");
1467
1468         new_shortname = 0;
1469         if (!sched_in->shortname[0]) {
1470                 sched_in->shortname[0] = next_shortname1;
1471                 sched_in->shortname[1] = next_shortname2;
1472
1473                 if (next_shortname1 < 'Z') {
1474                         next_shortname1++;
1475                 } else {
1476                         next_shortname1='A';
1477                         if (next_shortname2 < '9') {
1478                                 next_shortname2++;
1479                         } else {
1480                                 next_shortname2='0';
1481                         }
1482                 }
1483                 new_shortname = 1;
1484         }
1485
1486         for (cpu = 0; cpu <= max_cpu; cpu++) {
1487                 if (cpu != this_cpu)
1488                         printf(" ");
1489                 else
1490                         printf("*");
1491
1492                 if (curr_thread[cpu]) {
1493                         if (curr_thread[cpu]->pid)
1494                                 printf("%2s ", curr_thread[cpu]->shortname);
1495                         else
1496                                 printf(".  ");
1497                 } else
1498                         printf("   ");
1499         }
1500
1501         printf("  %12.6f secs ", (double)timestamp/1e9);
1502         if (new_shortname) {
1503                 printf("%s => %s:%d\n",
1504                         sched_in->shortname, sched_in->comm, sched_in->pid);
1505         } else {
1506                 printf("\n");
1507         }
1508 }
1509
1510
1511 static void
1512 process_sched_switch_event(struct raw_event_sample *raw,
1513                            struct event *event,
1514                            int this_cpu,
1515                            u64 timestamp __used,
1516                            struct thread *thread __used)
1517 {
1518         struct trace_switch_event switch_event;
1519
1520         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1521
1522         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1523         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1524         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1525         FILL_FIELD(switch_event, prev_state, event, raw->data);
1526         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1527         FILL_FIELD(switch_event, next_pid, event, raw->data);
1528         FILL_FIELD(switch_event, next_prio, event, raw->data);
1529
1530         if (curr_pid[this_cpu] != (u32)-1) {
1531                 /*
1532                  * Are we trying to switch away a PID that is
1533                  * not current?
1534                  */
1535                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1536                         nr_context_switch_bugs++;
1537         }
1538         if (trace_handler->switch_event)
1539                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1540
1541         curr_pid[this_cpu] = switch_event.next_pid;
1542 }
1543
1544 static void
1545 process_sched_runtime_event(struct raw_event_sample *raw,
1546                            struct event *event,
1547                            int cpu __used,
1548                            u64 timestamp __used,
1549                            struct thread *thread __used)
1550 {
1551         struct trace_runtime_event runtime_event;
1552
1553         FILL_ARRAY(runtime_event, comm, event, raw->data);
1554         FILL_FIELD(runtime_event, pid, event, raw->data);
1555         FILL_FIELD(runtime_event, runtime, event, raw->data);
1556         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1557
1558         if (trace_handler->runtime_event)
1559                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1560 }
1561
1562 static void
1563 process_sched_fork_event(struct raw_event_sample *raw,
1564                          struct event *event,
1565                          int cpu __used,
1566                          u64 timestamp __used,
1567                          struct thread *thread __used)
1568 {
1569         struct trace_fork_event fork_event;
1570
1571         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1572
1573         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1574         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1575         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1576         FILL_FIELD(fork_event, child_pid, event, raw->data);
1577
1578         if (trace_handler->fork_event)
1579                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1580 }
1581
1582 static void
1583 process_sched_exit_event(struct event *event,
1584                          int cpu __used,
1585                          u64 timestamp __used,
1586                          struct thread *thread __used)
1587 {
1588         if (verbose)
1589                 printf("sched_exit event %p\n", event);
1590 }
1591
1592 static void
1593 process_sched_migrate_task_event(struct raw_event_sample *raw,
1594                            struct event *event,
1595                            int cpu __used,
1596                            u64 timestamp __used,
1597                            struct thread *thread __used)
1598 {
1599         struct trace_migrate_task_event migrate_task_event;
1600
1601         FILL_COMMON_FIELDS(migrate_task_event, event, raw->data);
1602
1603         FILL_ARRAY(migrate_task_event, comm, event, raw->data);
1604         FILL_FIELD(migrate_task_event, pid, event, raw->data);
1605         FILL_FIELD(migrate_task_event, prio, event, raw->data);
1606         FILL_FIELD(migrate_task_event, cpu, event, raw->data);
1607
1608         if (trace_handler->migrate_task_event)
1609                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1610 }
1611
1612 static void
1613 process_raw_event(event_t *raw_event __used, void *more_data,
1614                   int cpu, u64 timestamp, struct thread *thread)
1615 {
1616         struct raw_event_sample *raw = more_data;
1617         struct event *event;
1618         int type;
1619
1620         type = trace_parse_common_type(raw->data);
1621         event = trace_find_event(type);
1622
1623         if (!strcmp(event->name, "sched_switch"))
1624                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1625         if (!strcmp(event->name, "sched_stat_runtime"))
1626                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1627         if (!strcmp(event->name, "sched_wakeup"))
1628                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1629         if (!strcmp(event->name, "sched_wakeup_new"))
1630                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1631         if (!strcmp(event->name, "sched_process_fork"))
1632                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1633         if (!strcmp(event->name, "sched_process_exit"))
1634                 process_sched_exit_event(event, cpu, timestamp, thread);
1635         if (!strcmp(event->name, "sched_migrate_task"))
1636                 process_sched_migrate_task_event(raw, event, cpu, timestamp, thread);
1637 }
1638
1639 static int
1640 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1641 {
1642         struct thread *thread;
1643         u64 ip = event->ip.ip;
1644         u64 timestamp = -1;
1645         u32 cpu = -1;
1646         u64 period = 1;
1647         void *more_data = event->ip.__more_data;
1648
1649         if (!(sample_type & PERF_SAMPLE_RAW))
1650                 return 0;
1651
1652         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1653
1654         if (sample_type & PERF_SAMPLE_TIME) {
1655                 timestamp = *(u64 *)more_data;
1656                 more_data += sizeof(u64);
1657         }
1658
1659         if (sample_type & PERF_SAMPLE_CPU) {
1660                 cpu = *(u32 *)more_data;
1661                 more_data += sizeof(u32);
1662                 more_data += sizeof(u32); /* reserved */
1663         }
1664
1665         if (sample_type & PERF_SAMPLE_PERIOD) {
1666                 period = *(u64 *)more_data;
1667                 more_data += sizeof(u64);
1668         }
1669
1670         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1671                 (void *)(offset + head),
1672                 (void *)(long)(event->header.size),
1673                 event->header.misc,
1674                 event->ip.pid, event->ip.tid,
1675                 (void *)(long)ip,
1676                 (long long)period);
1677
1678         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1679
1680         if (thread == NULL) {
1681                 eprintf("problem processing %d event, skipping it.\n",
1682                         event->header.type);
1683                 return -1;
1684         }
1685
1686         if (profile_cpu != -1 && profile_cpu != (int) cpu)
1687                 return 0;
1688
1689         process_raw_event(event, more_data, cpu, timestamp, thread);
1690
1691         return 0;
1692 }
1693
1694 static int
1695 process_lost_event(event_t *event __used,
1696                    unsigned long offset __used,
1697                    unsigned long head __used)
1698 {
1699         nr_lost_chunks++;
1700         nr_lost_events += event->lost.lost;
1701
1702         return 0;
1703 }
1704
1705 static int sample_type_check(u64 type)
1706 {
1707         sample_type = type;
1708
1709         if (!(sample_type & PERF_SAMPLE_RAW)) {
1710                 fprintf(stderr,
1711                         "No trace sample to read. Did you call perf record "
1712                         "without -R?");
1713                 return -1;
1714         }
1715
1716         return 0;
1717 }
1718
1719 static struct perf_file_handler file_handler = {
1720         .process_sample_event   = process_sample_event,
1721         .process_comm_event     = process_comm_event,
1722         .process_lost_event     = process_lost_event,
1723         .sample_type_check      = sample_type_check,
1724 };
1725
1726 static int read_events(void)
1727 {
1728         register_idle_thread(&threads, &last_match);
1729         register_perf_file_handler(&file_handler);
1730
1731         return mmap_dispatch_perf_file(&header, input_name, 0, 0, &cwdlen, &cwd);
1732 }
1733
1734 static void print_bad_events(void)
1735 {
1736         if (nr_unordered_timestamps && nr_timestamps) {
1737                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1738                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1739                         nr_unordered_timestamps, nr_timestamps);
1740         }
1741         if (nr_lost_events && nr_events) {
1742                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1743                         (double)nr_lost_events/(double)nr_events*100.0,
1744                         nr_lost_events, nr_events, nr_lost_chunks);
1745         }
1746         if (nr_state_machine_bugs && nr_timestamps) {
1747                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1748                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1749                         nr_state_machine_bugs, nr_timestamps);
1750                 if (nr_lost_events)
1751                         printf(" (due to lost events?)");
1752                 printf("\n");
1753         }
1754         if (nr_context_switch_bugs && nr_timestamps) {
1755                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1756                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1757                         nr_context_switch_bugs, nr_timestamps);
1758                 if (nr_lost_events)
1759                         printf(" (due to lost events?)");
1760                 printf("\n");
1761         }
1762 }
1763
1764 static void __cmd_lat(void)
1765 {
1766         struct rb_node *next;
1767
1768         setup_pager();
1769         read_events();
1770         sort_lat();
1771
1772         printf("\n -----------------------------------------------------------------------------------------\n");
1773         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1774         printf(" -----------------------------------------------------------------------------------------\n");
1775
1776         next = rb_first(&sorted_atom_root);
1777
1778         while (next) {
1779                 struct work_atoms *work_list;
1780
1781                 work_list = rb_entry(next, struct work_atoms, node);
1782                 output_lat_thread(work_list);
1783                 next = rb_next(next);
1784         }
1785
1786         printf(" -----------------------------------------------------------------------------------------\n");
1787         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1788                 (double)all_runtime/1e6, all_count);
1789
1790         printf(" ---------------------------------------------------\n");
1791
1792         print_bad_events();
1793         printf("\n");
1794
1795 }
1796
1797 static struct trace_sched_handler map_ops  = {
1798         .wakeup_event           = NULL,
1799         .switch_event           = map_switch_event,
1800         .runtime_event          = NULL,
1801         .fork_event             = NULL,
1802 };
1803
1804 static void __cmd_map(void)
1805 {
1806         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1807
1808         setup_pager();
1809         read_events();
1810         print_bad_events();
1811 }
1812
1813 static void __cmd_replay(void)
1814 {
1815         unsigned long i;
1816
1817         calibrate_run_measurement_overhead();
1818         calibrate_sleep_measurement_overhead();
1819
1820         test_calibrations();
1821
1822         read_events();
1823
1824         printf("nr_run_events:        %ld\n", nr_run_events);
1825         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1826         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1827
1828         if (targetless_wakeups)
1829                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1830         if (multitarget_wakeups)
1831                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1832         if (nr_run_events_optimized)
1833                 printf("run atoms optimized: %ld\n",
1834                         nr_run_events_optimized);
1835
1836         print_task_traces();
1837         add_cross_task_wakeups();
1838
1839         create_tasks();
1840         printf("------------------------------------------------------------\n");
1841         for (i = 0; i < replay_repeat; i++)
1842                 run_one_test();
1843 }
1844
1845
1846 static const char * const sched_usage[] = {
1847         "perf sched [<options>] {record|latency|map|replay|trace}",
1848         NULL
1849 };
1850
1851 static const struct option sched_options[] = {
1852         OPT_STRING('i', "input", &input_name, "file",
1853                     "input file name"),
1854         OPT_BOOLEAN('v', "verbose", &verbose,
1855                     "be more verbose (show symbol address, etc)"),
1856         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1857                     "dump raw trace in ASCII"),
1858         OPT_END()
1859 };
1860
1861 static const char * const latency_usage[] = {
1862         "perf sched latency [<options>]",
1863         NULL
1864 };
1865
1866 static const struct option latency_options[] = {
1867         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1868                    "sort by key(s): runtime, switch, avg, max"),
1869         OPT_BOOLEAN('v', "verbose", &verbose,
1870                     "be more verbose (show symbol address, etc)"),
1871         OPT_INTEGER('C', "CPU", &profile_cpu,
1872                     "CPU to profile on"),
1873         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1874                     "dump raw trace in ASCII"),
1875         OPT_END()
1876 };
1877
1878 static const char * const replay_usage[] = {
1879         "perf sched replay [<options>]",
1880         NULL
1881 };
1882
1883 static const struct option replay_options[] = {
1884         OPT_INTEGER('r', "repeat", &replay_repeat,
1885                     "repeat the workload replay N times (-1: infinite)"),
1886         OPT_BOOLEAN('v', "verbose", &verbose,
1887                     "be more verbose (show symbol address, etc)"),
1888         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1889                     "dump raw trace in ASCII"),
1890         OPT_END()
1891 };
1892
1893 static void setup_sorting(void)
1894 {
1895         char *tmp, *tok, *str = strdup(sort_order);
1896
1897         for (tok = strtok_r(str, ", ", &tmp);
1898                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1899                 if (sort_dimension__add(tok, &sort_list) < 0) {
1900                         error("Unknown --sort key: `%s'", tok);
1901                         usage_with_options(latency_usage, latency_options);
1902                 }
1903         }
1904
1905         free(str);
1906
1907         sort_dimension__add("pid", &cmp_pid);
1908 }
1909
1910 static const char *record_args[] = {
1911         "record",
1912         "-a",
1913         "-R",
1914         "-M",
1915         "-f",
1916         "-m", "1024",
1917         "-c", "1",
1918         "-e", "sched:sched_switch:r",
1919         "-e", "sched:sched_stat_wait:r",
1920         "-e", "sched:sched_stat_sleep:r",
1921         "-e", "sched:sched_stat_iowait:r",
1922         "-e", "sched:sched_stat_runtime:r",
1923         "-e", "sched:sched_process_exit:r",
1924         "-e", "sched:sched_process_fork:r",
1925         "-e", "sched:sched_wakeup:r",
1926         "-e", "sched:sched_migrate_task:r",
1927 };
1928
1929 static int __cmd_record(int argc, const char **argv)
1930 {
1931         unsigned int rec_argc, i, j;
1932         const char **rec_argv;
1933
1934         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1935         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1936
1937         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1938                 rec_argv[i] = strdup(record_args[i]);
1939
1940         for (j = 1; j < (unsigned int)argc; j++, i++)
1941                 rec_argv[i] = argv[j];
1942
1943         BUG_ON(i != rec_argc);
1944
1945         return cmd_record(i, rec_argv, NULL);
1946 }
1947
1948 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1949 {
1950         symbol__init();
1951
1952         argc = parse_options(argc, argv, sched_options, sched_usage,
1953                              PARSE_OPT_STOP_AT_NON_OPTION);
1954         if (!argc)
1955                 usage_with_options(sched_usage, sched_options);
1956
1957         if (!strncmp(argv[0], "rec", 3)) {
1958                 return __cmd_record(argc, argv);
1959         } else if (!strncmp(argv[0], "lat", 3)) {
1960                 trace_handler = &lat_ops;
1961                 if (argc > 1) {
1962                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1963                         if (argc)
1964                                 usage_with_options(latency_usage, latency_options);
1965                 }
1966                 setup_sorting();
1967                 __cmd_lat();
1968         } else if (!strcmp(argv[0], "map")) {
1969                 trace_handler = &map_ops;
1970                 setup_sorting();
1971                 __cmd_map();
1972         } else if (!strncmp(argv[0], "rep", 3)) {
1973                 trace_handler = &replay_ops;
1974                 if (argc) {
1975                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1976                         if (argc)
1977                                 usage_with_options(replay_usage, replay_options);
1978                 }
1979                 __cmd_replay();
1980         } else if (!strcmp(argv[0], "trace")) {
1981                 /*
1982                  * Aliased to 'perf trace' for now:
1983                  */
1984                 return cmd_trace(argc, argv, prefix);
1985         } else {
1986                 usage_with_options(sched_usage, sched_options);
1987         }
1988
1989         return 0;
1990 }