SAFE public projects git trees. - safe/jmp/linux-2.6/blob - tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2
   3 #include "util/util.h"
   4 #include "util/cache.h"
   5 #include "util/symbol.h"
   6 #include "util/thread.h"
   7 #include "util/header.h"
   8
   9 #include "util/parse-options.h"
  10
  11 #include "perf.h"
  12 #include "util/debug.h"
  13
  14 #include "util/trace-event.h"
  15 #include <sys/types.h>
  16
  17 static char                     const *input_name = "perf.data";
  18 static int                      input;
  19 static unsigned long            page_size;
  20 static unsigned long            mmap_window = 32;
  21
  22 static unsigned long            total_comm = 0;
  23
  24 static struct rb_root           threads;
  25 static struct thread            *last_match;
  26
  27 static struct perf_header       *header;
  28 static u64                      sample_type;
  29
  30
  31 /*
  32  * Scheduler benchmarks
  33  */
  34 #include <sys/resource.h>
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/time.h>
  38 #include <sys/prctl.h>
  39
  40 #include <linux/unistd.h>
  41
  42 #include <semaphore.h>
  43 #include <pthread.h>
  44 #include <signal.h>
  45 #include <values.h>
  46 #include <string.h>
  47 #include <unistd.h>
  48 #include <stdlib.h>
  49 #include <assert.h>
  50 #include <fcntl.h>
  51 #include <time.h>
  52 #include <math.h>
  53
  54 #include <stdio.h>
  55
  56 #define PR_SET_NAME     15               /* Set process name */
  57
  58 #define BUG_ON(x)       assert(!(x))
  59
  60 #define DEBUG           0
  61
  62 typedef unsigned long long nsec_t;
  63
  64 static nsec_t run_measurement_overhead;
  65 static nsec_t sleep_measurement_overhead;
  66
  67 static nsec_t get_nsecs(void)
  68 {
  69         struct timespec ts;
  70
  71         clock_gettime(CLOCK_MONOTONIC, &ts);
  72
  73         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
  74 }
  75
  76 static void burn_nsecs(nsec_t nsecs)
  77 {
  78         nsec_t T0 = get_nsecs(), T1;
  79
  80         do {
  81                 T1 = get_nsecs();
  82         } while (T1 + run_measurement_overhead < T0 + nsecs);
  83 }
  84
  85 static void sleep_nsecs(nsec_t nsecs)
  86 {
  87         struct timespec ts;
  88
  89         ts.tv_nsec = nsecs % 999999999;
  90         ts.tv_sec = nsecs / 999999999;
  91
  92         nanosleep(&ts, NULL);
  93 }
  94
  95 static void calibrate_run_measurement_overhead(void)
  96 {
  97         nsec_t T0, T1, delta, min_delta = 1000000000ULL;
  98         int i;
  99
 100         for (i = 0; i < 10; i++) {
 101                 T0 = get_nsecs();
 102                 burn_nsecs(0);
 103                 T1 = get_nsecs();
 104                 delta = T1-T0;
 105                 min_delta = min(min_delta, delta);
 106         }
 107         run_measurement_overhead = min_delta;
 108
 109         printf("run measurement overhead: %Ld nsecs\n", min_delta);
 110 }
 111
 112 static void calibrate_sleep_measurement_overhead(void)
 113 {
 114         nsec_t T0, T1, delta, min_delta = 1000000000ULL;
 115         int i;
 116
 117         for (i = 0; i < 10; i++) {
 118                 T0 = get_nsecs();
 119                 sleep_nsecs(10000);
 120                 T1 = get_nsecs();
 121                 delta = T1-T0;
 122                 min_delta = min(min_delta, delta);
 123         }
 124         min_delta -= 10000;
 125         sleep_measurement_overhead = min_delta;
 126
 127         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 128 }
 129
 130 #define COMM_LEN        20
 131 #define SYM_LEN         129
 132
 133 #define MAX_PID         65536
 134
 135 static unsigned long nr_tasks;
 136
 137 struct sched_event;
 138
 139 struct task_desc {
 140         unsigned long           nr;
 141         unsigned long           pid;
 142         char                    comm[COMM_LEN];
 143
 144         unsigned long           nr_events;
 145         unsigned long           curr_event;
 146         struct sched_event      **events;
 147
 148         pthread_t               thread;
 149         sem_t                   sleep_sem;
 150
 151         sem_t                   ready_for_work;
 152         sem_t                   work_done_sem;
 153
 154         nsec_t                  cpu_usage;
 155 };
 156
 157 enum sched_event_type {
 158         SCHED_EVENT_RUN,
 159         SCHED_EVENT_SLEEP,
 160         SCHED_EVENT_WAKEUP,
 161 };
 162
 163 struct sched_event {
 164         enum sched_event_type   type;
 165         nsec_t                  timestamp;
 166         nsec_t                  duration;
 167         unsigned long           nr;
 168         int                     specific_wait;
 169         sem_t                   *wait_sem;
 170         struct task_desc        *wakee;
 171 };
 172
 173 static struct task_desc         *pid_to_task[MAX_PID];
 174
 175 static struct task_desc         **tasks;
 176
 177 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
 178 static nsec_t                   start_time;
 179
 180 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
 181
 182 static unsigned long            nr_run_events;
 183 static unsigned long            nr_sleep_events;
 184 static unsigned long            nr_wakeup_events;
 185
 186 static unsigned long            nr_sleep_corrections;
 187 static unsigned long            nr_run_events_optimized;
 188
 189 static struct sched_event *
 190 get_new_event(struct task_desc *task, nsec_t timestamp)
 191 {
 192         struct sched_event *event = calloc(1, sizeof(*event));
 193         unsigned long idx = task->nr_events;
 194         size_t size;
 195
 196         event->timestamp = timestamp;
 197         event->nr = idx;
 198
 199         task->nr_events++;
 200         size = sizeof(struct sched_event *) * task->nr_events;
 201         task->events = realloc(task->events, size);
 202         BUG_ON(!task->events);
 203
 204         task->events[idx] = event;
 205
 206         return event;
 207 }
 208
 209 static struct sched_event *last_event(struct task_desc *task)
 210 {
 211         if (!task->nr_events)
 212                 return NULL;
 213
 214         return task->events[task->nr_events - 1];
 215 }
 216
 217 static void
 218 add_sched_event_run(struct task_desc *task, nsec_t timestamp, u64 duration)
 219 {
 220         struct sched_event *event, *curr_event = last_event(task);
 221
 222         /*
 223          * optimize an existing RUN event by merging this one
 224          * to it:
 225          */
 226         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 227                 nr_run_events_optimized++;
 228                 curr_event->duration += duration;
 229                 return;
 230         }
 231
 232         event = get_new_event(task, timestamp);
 233
 234         event->type = SCHED_EVENT_RUN;
 235         event->duration = duration;
 236
 237         nr_run_events++;
 238 }
 239
 240 static unsigned long targetless_wakeups;
 241 static unsigned long multitarget_wakeups;
 242
 243 static void
 244 add_sched_event_wakeup(struct task_desc *task, nsec_t timestamp,
 245                        struct task_desc *wakee)
 246 {
 247         struct sched_event *event, *wakee_event;
 248
 249         event = get_new_event(task, timestamp);
 250         event->type = SCHED_EVENT_WAKEUP;
 251         event->wakee = wakee;
 252
 253         wakee_event = last_event(wakee);
 254         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 255                 targetless_wakeups++;
 256                 return;
 257         }
 258         if (wakee_event->wait_sem) {
 259                 multitarget_wakeups++;
 260                 return;
 261         }
 262
 263         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
 264         sem_init(wakee_event->wait_sem, 0, 0);
 265         wakee_event->specific_wait = 1;
 266         event->wait_sem = wakee_event->wait_sem;
 267
 268         nr_wakeup_events++;
 269 }
 270
 271 static void
 272 add_sched_event_sleep(struct task_desc *task, nsec_t timestamp,
 273                       u64 task_state __used)
 274 {
 275         struct sched_event *event = get_new_event(task, timestamp);
 276
 277         event->type = SCHED_EVENT_SLEEP;
 278
 279         nr_sleep_events++;
 280 }
 281
 282 static struct task_desc *register_pid(unsigned long pid, const char *comm)
 283 {
 284         struct task_desc *task;
 285
 286         BUG_ON(pid >= MAX_PID);
 287
 288         task = pid_to_task[pid];
 289
 290         if (task)
 291                 return task;
 292
 293         task = calloc(1, sizeof(*task));
 294         task->pid = pid;
 295         task->nr = nr_tasks;
 296         strcpy(task->comm, comm);
 297         /*
 298          * every task starts in sleeping state - this gets ignored
 299          * if there's no wakeup pointing to this sleep state:
 300          */
 301         add_sched_event_sleep(task, 0, 0);
 302
 303         pid_to_task[pid] = task;
 304         nr_tasks++;
 305         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
 306         BUG_ON(!tasks);
 307         tasks[task->nr] = task;
 308
 309         if (verbose)
 310                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
 311
 312         return task;
 313 }
 314
 315
 316 static void print_task_traces(void)
 317 {
 318         struct task_desc *task;
 319         unsigned long i;
 320
 321         for (i = 0; i < nr_tasks; i++) {
 322                 task = tasks[i];
 323                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 324                         task->nr, task->comm, task->pid, task->nr_events);
 325         }
 326 }
 327
 328 static void add_cross_task_wakeups(void)
 329 {
 330         struct task_desc *task1, *task2;
 331         unsigned long i, j;
 332
 333         for (i = 0; i < nr_tasks; i++) {
 334                 task1 = tasks[i];
 335                 j = i + 1;
 336                 if (j == nr_tasks)
 337                         j = 0;
 338                 task2 = tasks[j];
 339                 add_sched_event_wakeup(task1, 0, task2);
 340         }
 341 }
 342
 343 static void
 344 process_sched_event(struct task_desc *this_task __used, struct sched_event *event)
 345 {
 346         int ret = 0;
 347         nsec_t now;
 348         long long delta;
 349
 350         now = get_nsecs();
 351         delta = start_time + event->timestamp - now;
 352
 353         switch (event->type) {
 354                 case SCHED_EVENT_RUN:
 355                         burn_nsecs(event->duration);
 356                         break;
 357                 case SCHED_EVENT_SLEEP:
 358                         if (event->wait_sem)
 359                                 ret = sem_wait(event->wait_sem);
 360                         BUG_ON(ret);
 361                         break;
 362                 case SCHED_EVENT_WAKEUP:
 363                         if (event->wait_sem)
 364                                 ret = sem_post(event->wait_sem);
 365                         BUG_ON(ret);
 366                         break;
 367                 default:
 368                         BUG_ON(1);
 369         }
 370 }
 371
 372 static nsec_t get_cpu_usage_nsec_parent(void)
 373 {
 374         struct rusage ru;
 375         nsec_t sum;
 376         int err;
 377
 378         err = getrusage(RUSAGE_SELF, &ru);
 379         BUG_ON(err);
 380
 381         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 382         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 383
 384         return sum;
 385 }
 386
 387 static nsec_t get_cpu_usage_nsec_self(void)
 388 {
 389         char filename [] = "/proc/1234567890/sched";
 390         unsigned long msecs, nsecs;
 391         char *line = NULL;
 392         nsec_t total = 0;
 393         size_t len = 0;
 394         ssize_t chars;
 395         FILE *file;
 396         int ret;
 397
 398         sprintf(filename, "/proc/%d/sched", getpid());
 399         file = fopen(filename, "r");
 400         BUG_ON(!file);
 401
 402         while ((chars = getline(&line, &len, file)) != -1) {
 403                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
 404                         &msecs, &nsecs);
 405                 if (ret == 2) {
 406                         total = msecs*1e6 + nsecs;
 407                         break;
 408                 }
 409         }
 410         if (line)
 411                 free(line);
 412         fclose(file);
 413
 414         return total;
 415 }
 416
 417 static void *thread_func(void *ctx)
 418 {
 419         struct task_desc *this_task = ctx;
 420         nsec_t cpu_usage_0, cpu_usage_1;
 421         unsigned long i, ret;
 422         char comm2[22];
 423
 424         sprintf(comm2, ":%s", this_task->comm);
 425         prctl(PR_SET_NAME, comm2);
 426
 427 again:
 428         ret = sem_post(&this_task->ready_for_work);
 429         BUG_ON(ret);
 430         ret = pthread_mutex_lock(&start_work_mutex);
 431         BUG_ON(ret);
 432         ret = pthread_mutex_unlock(&start_work_mutex);
 433         BUG_ON(ret);
 434
 435         cpu_usage_0 = get_cpu_usage_nsec_self();
 436
 437         for (i = 0; i < this_task->nr_events; i++) {
 438                 this_task->curr_event = i;
 439                 process_sched_event(this_task, this_task->events[i]);
 440         }
 441
 442         cpu_usage_1 = get_cpu_usage_nsec_self();
 443         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 444
 445         ret = sem_post(&this_task->work_done_sem);
 446         BUG_ON(ret);
 447
 448         ret = pthread_mutex_lock(&work_done_wait_mutex);
 449         BUG_ON(ret);
 450         ret = pthread_mutex_unlock(&work_done_wait_mutex);
 451         BUG_ON(ret);
 452
 453         goto again;
 454 }
 455
 456 static void create_tasks(void)
 457 {
 458         struct task_desc *task;
 459         pthread_attr_t attr;
 460         unsigned long i;
 461         int err;
 462
 463         err = pthread_attr_init(&attr);
 464         BUG_ON(err);
 465         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
 466         BUG_ON(err);
 467         err = pthread_mutex_lock(&start_work_mutex);
 468         BUG_ON(err);
 469         err = pthread_mutex_lock(&work_done_wait_mutex);
 470         BUG_ON(err);
 471         for (i = 0; i < nr_tasks; i++) {
 472                 task = tasks[i];
 473                 sem_init(&task->sleep_sem, 0, 0);
 474                 sem_init(&task->ready_for_work, 0, 0);
 475                 sem_init(&task->work_done_sem, 0, 0);
 476                 task->curr_event = 0;
 477                 err = pthread_create(&task->thread, &attr, thread_func, task);
 478                 BUG_ON(err);
 479         }
 480 }
 481
 482 static nsec_t cpu_usage;
 483 static nsec_t runavg_cpu_usage;
 484 static nsec_t parent_cpu_usage;
 485 static nsec_t runavg_parent_cpu_usage;
 486
 487 static void wait_for_tasks(void)
 488 {
 489         nsec_t cpu_usage_0, cpu_usage_1;
 490         struct task_desc *task;
 491         unsigned long i, ret;
 492
 493         start_time = get_nsecs();
 494         cpu_usage = 0;
 495         pthread_mutex_unlock(&work_done_wait_mutex);
 496
 497         for (i = 0; i < nr_tasks; i++) {
 498                 task = tasks[i];
 499                 ret = sem_wait(&task->ready_for_work);
 500                 BUG_ON(ret);
 501                 sem_init(&task->ready_for_work, 0, 0);
 502         }
 503         ret = pthread_mutex_lock(&work_done_wait_mutex);
 504         BUG_ON(ret);
 505
 506         cpu_usage_0 = get_cpu_usage_nsec_parent();
 507
 508         pthread_mutex_unlock(&start_work_mutex);
 509
 510         for (i = 0; i < nr_tasks; i++) {
 511                 task = tasks[i];
 512                 ret = sem_wait(&task->work_done_sem);
 513                 BUG_ON(ret);
 514                 sem_init(&task->work_done_sem, 0, 0);
 515                 cpu_usage += task->cpu_usage;
 516                 task->cpu_usage = 0;
 517         }
 518
 519         cpu_usage_1 = get_cpu_usage_nsec_parent();
 520         if (!runavg_cpu_usage)
 521                 runavg_cpu_usage = cpu_usage;
 522         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
 523
 524         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 525         if (!runavg_parent_cpu_usage)
 526                 runavg_parent_cpu_usage = parent_cpu_usage;
 527         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
 528                                    parent_cpu_usage)/10;
 529
 530         ret = pthread_mutex_lock(&start_work_mutex);
 531         BUG_ON(ret);
 532
 533         for (i = 0; i < nr_tasks; i++) {
 534                 task = tasks[i];
 535                 sem_init(&task->sleep_sem, 0, 0);
 536                 task->curr_event = 0;
 537         }
 538 }
 539
 540 static int __cmd_sched(void);
 541
 542 static void parse_trace(void)
 543 {
 544         __cmd_sched();
 545
 546         printf("nr_run_events:        %ld\n", nr_run_events);
 547         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
 548         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
 549
 550         if (targetless_wakeups)
 551                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
 552         if (multitarget_wakeups)
 553                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
 554         if (nr_run_events_optimized)
 555                 printf("run events optimized: %ld\n",
 556                         nr_run_events_optimized);
 557 }
 558
 559 static unsigned long nr_runs;
 560 static nsec_t sum_runtime;
 561 static nsec_t sum_fluct;
 562 static nsec_t run_avg;
 563
 564 static void run_one_test(void)
 565 {
 566         nsec_t T0, T1, delta, avg_delta, fluct, std_dev;
 567
 568         T0 = get_nsecs();
 569         wait_for_tasks();
 570         T1 = get_nsecs();
 571
 572         delta = T1 - T0;
 573         sum_runtime += delta;
 574         nr_runs++;
 575
 576         avg_delta = sum_runtime / nr_runs;
 577         if (delta < avg_delta)
 578                 fluct = avg_delta - delta;
 579         else
 580                 fluct = delta - avg_delta;
 581         sum_fluct += fluct;
 582         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
 583         if (!run_avg)
 584                 run_avg = delta;
 585         run_avg = (run_avg*9 + delta)/10;
 586
 587         printf("#%-3ld: %0.3f, ",
 588                 nr_runs, (double)delta/1000000.0);
 589
 590 #if 0
 591         printf("%0.2f +- %0.2f, ",
 592                 (double)avg_delta/1e6, (double)std_dev/1e6);
 593 #endif
 594         printf("ravg: %0.2f, ",
 595                 (double)run_avg/1e6);
 596
 597         printf("cpu: %0.2f / %0.2f",
 598                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
 599
 600 #if 0
 601         /*
 602          * rusage statistics done by the parent, these are less
 603          * accurate than the sum_exec_runtime based statistics:
 604          */
 605         printf(" [%0.2f / %0.2f]",
 606                 (double)parent_cpu_usage/1e6,
 607                 (double)runavg_parent_cpu_usage/1e6);
 608 #endif
 609
 610         printf("\n");
 611
 612         if (nr_sleep_corrections)
 613                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
 614         nr_sleep_corrections = 0;
 615 }
 616
 617 static void test_calibrations(void)
 618 {
 619         nsec_t T0, T1;
 620
 621         T0 = get_nsecs();
 622         burn_nsecs(1e6);
 623         T1 = get_nsecs();
 624
 625         printf("the run test took %Ld nsecs\n", T1-T0);
 626
 627         T0 = get_nsecs();
 628         sleep_nsecs(1e6);
 629         T1 = get_nsecs();
 630
 631         printf("the sleep test took %Ld nsecs\n", T1-T0);
 632 }
 633
 634 static int
 635 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 636 {
 637         struct thread *thread;
 638
 639         thread = threads__findnew(event->comm.pid, &threads, &last_match);
 640
 641         dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
 642                 (void *)(offset + head),
 643                 (void *)(long)(event->header.size),
 644                 event->comm.comm, event->comm.pid);
 645
 646         if (thread == NULL ||
 647             thread__set_comm(thread, event->comm.comm)) {
 648                 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
 649                 return -1;
 650         }
 651         total_comm++;
 652
 653         return 0;
 654 }
 655
 656
 657 struct raw_event_sample {
 658         u32 size;
 659         char data[0];
 660 };
 661
 662 #define FILL_FIELD(ptr, field, event, data)     \
 663         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
 664
 665 #define FILL_ARRAY(ptr, array, event, data)                     \
 666 do {                                                            \
 667         void *__array = raw_field_ptr(event, #array, data);     \
 668         memcpy(ptr.array, __array, sizeof(ptr.array));  \
 669 } while(0)
 670
 671 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
 672 do {                                                            \
 673         FILL_FIELD(ptr, common_type, event, data);              \
 674         FILL_FIELD(ptr, common_flags, event, data);             \
 675         FILL_FIELD(ptr, common_preempt_count, event, data);     \
 676         FILL_FIELD(ptr, common_pid, event, data);               \
 677         FILL_FIELD(ptr, common_tgid, event, data);              \
 678 } while (0)
 679
 680 struct trace_wakeup_event {
 681         u32 size;
 682
 683         u16 common_type;
 684         u8 common_flags;
 685         u8 common_preempt_count;
 686         u32 common_pid;
 687         u32 common_tgid;
 688
 689         char comm[16];
 690         u32 pid;
 691
 692         u32 prio;
 693         u32 success;
 694         u32 cpu;
 695 };
 696
 697 static void
 698 process_sched_wakeup_event(struct raw_event_sample *raw, struct event *event,
 699                   int cpu __used, u64 timestamp __used, struct thread *thread __used)
 700 {
 701         struct task_desc *waker, *wakee;
 702         struct trace_wakeup_event wakeup_event;
 703
 704         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
 705
 706         FILL_ARRAY(wakeup_event, comm, event, raw->data);
 707         FILL_FIELD(wakeup_event, pid, event, raw->data);
 708         FILL_FIELD(wakeup_event, prio, event, raw->data);
 709         FILL_FIELD(wakeup_event, success, event, raw->data);
 710         FILL_FIELD(wakeup_event, cpu, event, raw->data);
 711
 712
 713         if (verbose) {
 714                 printf("sched_wakeup event %p\n", event);
 715
 716                 printf(" ... pid %d woke up %s/%d\n",
 717                         wakeup_event.common_pid,
 718                         wakeup_event.comm,
 719                         wakeup_event.pid);
 720         }
 721
 722         waker = register_pid(wakeup_event.common_pid, "<unknown>");
 723         wakee = register_pid(wakeup_event.pid, wakeup_event.comm);
 724
 725         add_sched_event_wakeup(waker, timestamp, wakee);
 726 }
 727
 728 struct trace_switch_event {
 729         u32 size;
 730
 731         u16 common_type;
 732         u8 common_flags;
 733         u8 common_preempt_count;
 734         u32 common_pid;
 735         u32 common_tgid;
 736
 737         char prev_comm[16];
 738         u32 prev_pid;
 739         u32 prev_prio;
 740         u64 prev_state;
 741         char next_comm[16];
 742         u32 next_pid;
 743         u32 next_prio;
 744 };
 745
 746 #define MAX_CPUS 4096
 747
 748 unsigned long cpu_last_switched[MAX_CPUS];
 749
 750 static void
 751 process_sched_switch_event(struct raw_event_sample *raw, struct event *event,
 752                   int cpu __used, u64 timestamp __used, struct thread *thread __used)
 753 {
 754         struct trace_switch_event switch_event;
 755         struct task_desc *prev, *next;
 756         u64 timestamp0;
 757         s64 delta;
 758
 759         FILL_COMMON_FIELDS(switch_event, event, raw->data);
 760
 761         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
 762         FILL_FIELD(switch_event, prev_pid, event, raw->data);
 763         FILL_FIELD(switch_event, prev_prio, event, raw->data);
 764         FILL_FIELD(switch_event, prev_state, event, raw->data);
 765         FILL_ARRAY(switch_event, next_comm, event, raw->data);
 766         FILL_FIELD(switch_event, next_pid, event, raw->data);
 767         FILL_FIELD(switch_event, next_prio, event, raw->data);
 768
 769         if (verbose)
 770                 printf("sched_switch event %p\n", event);
 771
 772         if (cpu >= MAX_CPUS || cpu < 0)
 773                 return;
 774
 775         timestamp0 = cpu_last_switched[cpu];
 776         if (timestamp0)
 777                 delta = timestamp - timestamp0;
 778         else
 779                 delta = 0;
 780
 781         if (delta < 0)
 782                 die("hm, delta: %Ld < 0 ?\n", delta);
 783
 784         if (verbose) {
 785                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
 786                         switch_event.prev_comm, switch_event.prev_pid,
 787                         switch_event.next_comm, switch_event.next_pid,
 788                         delta);
 789         }
 790
 791         prev = register_pid(switch_event.prev_pid, switch_event.prev_comm);
 792         next = register_pid(switch_event.next_pid, switch_event.next_comm);
 793
 794         cpu_last_switched[cpu] = timestamp;
 795
 796         add_sched_event_run(prev, timestamp, delta);
 797         add_sched_event_sleep(prev, timestamp, switch_event.prev_state);
 798 }
 799
 800 struct trace_fork_event {
 801         u32 size;
 802
 803         u16 common_type;
 804         u8 common_flags;
 805         u8 common_preempt_count;
 806         u32 common_pid;
 807         u32 common_tgid;
 808
 809         char parent_comm[16];
 810         u32 parent_pid;
 811         char child_comm[16];
 812         u32 child_pid;
 813 };
 814
 815 static void
 816 process_sched_fork_event(struct raw_event_sample *raw, struct event *event,
 817                   int cpu __used, u64 timestamp __used, struct thread *thread __used)
 818 {
 819         struct trace_fork_event fork_event;
 820
 821         FILL_COMMON_FIELDS(fork_event, event, raw->data);
 822
 823         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
 824         FILL_FIELD(fork_event, parent_pid, event, raw->data);
 825         FILL_ARRAY(fork_event, child_comm, event, raw->data);
 826         FILL_FIELD(fork_event, child_pid, event, raw->data);
 827
 828         if (verbose) {
 829                 printf("sched_fork event %p\n", event);
 830                 printf("... parent: %s/%d\n", fork_event.parent_comm, fork_event.parent_pid);
 831                 printf("...  child: %s/%d\n", fork_event.child_comm, fork_event.child_pid);
 832         }
 833         register_pid(fork_event.parent_pid, fork_event.parent_comm);
 834         register_pid(fork_event.child_pid, fork_event.child_comm);
 835 }
 836
 837 static void process_sched_exit_event(struct event *event,
 838                   int cpu __used, u64 timestamp __used, struct thread *thread __used)
 839 {
 840         if (verbose)
 841                 printf("sched_exit event %p\n", event);
 842 }
 843
 844 static void
 845 process_raw_event(event_t *raw_event __used, void *more_data,
 846                   int cpu, u64 timestamp, struct thread *thread)
 847 {
 848         struct raw_event_sample *raw = more_data;
 849         struct event *event;
 850         int type;
 851
 852         type = trace_parse_common_type(raw->data);
 853         event = trace_find_event(type);
 854
 855         if (!strcmp(event->name, "sched_switch"))
 856                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
 857         if (!strcmp(event->name, "sched_wakeup"))
 858                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
 859         if (!strcmp(event->name, "sched_wakeup_new"))
 860                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
 861         if (!strcmp(event->name, "sched_process_fork"))
 862                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
 863         if (!strcmp(event->name, "sched_process_exit"))
 864                 process_sched_exit_event(event, cpu, timestamp, thread);
 865 }
 866
 867 static int
 868 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 869 {
 870         char level;
 871         int show = 0;
 872         struct dso *dso = NULL;
 873         struct thread *thread;
 874         u64 ip = event->ip.ip;
 875         u64 timestamp = -1;
 876         u32 cpu = -1;
 877         u64 period = 1;
 878         void *more_data = event->ip.__more_data;
 879         int cpumode;
 880
 881         thread = threads__findnew(event->ip.pid, &threads, &last_match);
 882
 883         if (sample_type & PERF_SAMPLE_TIME) {
 884                 timestamp = *(u64 *)more_data;
 885                 more_data += sizeof(u64);
 886         }
 887
 888         if (sample_type & PERF_SAMPLE_CPU) {
 889                 cpu = *(u32 *)more_data;
 890                 more_data += sizeof(u32);
 891                 more_data += sizeof(u32); /* reserved */
 892         }
 893
 894         if (sample_type & PERF_SAMPLE_PERIOD) {
 895                 period = *(u64 *)more_data;
 896                 more_data += sizeof(u64);
 897         }
 898
 899         dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
 900                 (void *)(offset + head),
 901                 (void *)(long)(event->header.size),
 902                 event->header.misc,
 903                 event->ip.pid, event->ip.tid,
 904                 (void *)(long)ip,
 905                 (long long)period);
 906
 907         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 908
 909         if (thread == NULL) {
 910                 eprintf("problem processing %d event, skipping it.\n",
 911                         event->header.type);
 912                 return -1;
 913         }
 914
 915         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
 916
 917         if (cpumode == PERF_EVENT_MISC_KERNEL) {
 918                 show = SHOW_KERNEL;
 919                 level = 'k';
 920
 921                 dso = kernel_dso;
 922
 923                 dump_printf(" ...... dso: %s\n", dso->name);
 924
 925         } else if (cpumode == PERF_EVENT_MISC_USER) {
 926
 927                 show = SHOW_USER;
 928                 level = '.';
 929
 930         } else {
 931                 show = SHOW_HV;
 932                 level = 'H';
 933
 934                 dso = hypervisor_dso;
 935
 936                 dump_printf(" ...... dso: [hypervisor]\n");
 937         }
 938
 939         if (sample_type & PERF_SAMPLE_RAW)
 940                 process_raw_event(event, more_data, cpu, timestamp, thread);
 941
 942         return 0;
 943 }
 944
 945 static int
 946 process_event(event_t *event, unsigned long offset, unsigned long head)
 947 {
 948         trace_event(event);
 949
 950         switch (event->header.type) {
 951         case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
 952                 return 0;
 953
 954         case PERF_EVENT_COMM:
 955                 return process_comm_event(event, offset, head);
 956
 957         case PERF_EVENT_EXIT ... PERF_EVENT_READ:
 958                 return 0;
 959
 960         case PERF_EVENT_SAMPLE:
 961                 return process_sample_event(event, offset, head);
 962
 963         case PERF_EVENT_MAX:
 964         default:
 965                 return -1;
 966         }
 967
 968         return 0;
 969 }
 970
 971 static int __cmd_sched(void)
 972 {
 973         int ret, rc = EXIT_FAILURE;
 974         unsigned long offset = 0;
 975         unsigned long head = 0;
 976         struct stat perf_stat;
 977         event_t *event;
 978         uint32_t size;
 979         char *buf;
 980
 981         trace_report();
 982         register_idle_thread(&threads, &last_match);
 983
 984         input = open(input_name, O_RDONLY);
 985         if (input < 0) {
 986                 perror("failed to open file");
 987                 exit(-1);
 988         }
 989
 990         ret = fstat(input, &perf_stat);
 991         if (ret < 0) {
 992                 perror("failed to stat file");
 993                 exit(-1);
 994         }
 995
 996         if (!perf_stat.st_size) {
 997                 fprintf(stderr, "zero-sized file, nothing to do!\n");
 998                 exit(0);
 999         }
1000         header = perf_header__read(input);
1001         head = header->data_offset;
1002         sample_type = perf_header__sample_type(header);
1003
1004         if (!(sample_type & PERF_SAMPLE_RAW))
1005                 die("No trace sample to read. Did you call perf record "
1006                     "without -R?");
1007
1008         if (load_kernel() < 0) {
1009                 perror("failed to load kernel symbols");
1010                 return EXIT_FAILURE;
1011         }
1012
1013 remap:
1014         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1015                            MAP_SHARED, input, offset);
1016         if (buf == MAP_FAILED) {
1017                 perror("failed to mmap file");
1018                 exit(-1);
1019         }
1020
1021 more:
1022         event = (event_t *)(buf + head);
1023
1024         size = event->header.size;
1025         if (!size)
1026                 size = 8;
1027
1028         if (head + event->header.size >= page_size * mmap_window) {
1029                 unsigned long shift = page_size * (head / page_size);
1030                 int res;
1031
1032                 res = munmap(buf, page_size * mmap_window);
1033                 assert(res == 0);
1034
1035                 offset += shift;
1036                 head -= shift;
1037                 goto remap;
1038         }
1039
1040         size = event->header.size;
1041
1042
1043         if (!size || process_event(event, offset, head) < 0) {
1044
1045                 /*
1046                  * assume we lost track of the stream, check alignment, and
1047                  * increment a single u64 in the hope to catch on again 'soon'.
1048                  */
1049
1050                 if (unlikely(head & 7))
1051                         head &= ~7ULL;
1052
1053                 size = 8;
1054         }
1055
1056         head += size;
1057
1058         if (offset + head < (unsigned long)perf_stat.st_size)
1059                 goto more;
1060
1061         rc = EXIT_SUCCESS;
1062         close(input);
1063
1064         return rc;
1065 }
1066
1067 static const char * const annotate_usage[] = {
1068         "perf trace [<options>] <command>",
1069         NULL
1070 };
1071
1072 static const struct option options[] = {
1073         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1074                     "dump raw trace in ASCII"),
1075         OPT_BOOLEAN('v', "verbose", &verbose,
1076                     "be more verbose (show symbol address, etc)"),
1077         OPT_END()
1078 };
1079
1080 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1081 {
1082         long nr_iterations = 10, i;
1083
1084         symbol__init();
1085         page_size = getpagesize();
1086
1087         argc = parse_options(argc, argv, options, annotate_usage, 0);
1088         if (argc) {
1089                 /*
1090                  * Special case: if there's an argument left then assume tha
1091                  * it's a symbol filter:
1092                  */
1093                 if (argc > 1)
1094                         usage_with_options(annotate_usage, options);
1095         }
1096
1097 //      setup_pager();
1098
1099         calibrate_run_measurement_overhead();
1100         calibrate_sleep_measurement_overhead();
1101
1102         test_calibrations();
1103
1104         parse_trace();
1105         print_task_traces();
1106         add_cross_task_wakeups();
1107
1108         create_tasks();
1109         printf("------------------------------------------------------------\n");
1110         for (i = 0; i < nr_iterations; i++)
1111                 run_one_test();
1112
1113         return 0;
1114 }