tools/perf/builtin-sched.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/evlist.h"
   6 #include "util/cache.h"
   7 #include "util/evsel.h"
   8 #include "util/symbol.h"
   9 #include "util/thread.h"
  10 #include "util/header.h"
  11 #include "util/session.h"
  12 #include "util/tool.h"
  13
  14 #include "util/parse-options.h"
  15 #include "util/trace-event.h"
  16
  17 #include "util/debug.h"
  18
  19 #include <sys/prctl.h>
  20 #include <sys/resource.h>
  21
  22 #include <semaphore.h>
  23 #include <pthread.h>
  24 #include <math.h>
  25
  26 #define PR_SET_NAME             15               /* Set process name */
  27 #define MAX_CPUS                4096
  28 #define COMM_LEN                20
  29 #define SYM_LEN                 129
  30 #define MAX_PID                 65536
  31
  32 struct sched_atom;
  33
  34 struct task_desc {
  35         unsigned long           nr;
  36         unsigned long           pid;
  37         char                    comm[COMM_LEN];
  38
  39         unsigned long           nr_events;
  40         unsigned long           curr_event;
  41         struct sched_atom       **atoms;
  42
  43         pthread_t               thread;
  44         sem_t                   sleep_sem;
  45
  46         sem_t                   ready_for_work;
  47         sem_t                   work_done_sem;
  48
  49         u64                     cpu_usage;
  50 };
  51
  52 enum sched_event_type {
  53         SCHED_EVENT_RUN,
  54         SCHED_EVENT_SLEEP,
  55         SCHED_EVENT_WAKEUP,
  56         SCHED_EVENT_MIGRATION,
  57 };
  58
  59 struct sched_atom {
  60         enum sched_event_type   type;
  61         int                     specific_wait;
  62         u64                     timestamp;
  63         u64                     duration;
  64         unsigned long           nr;
  65         sem_t                   *wait_sem;
  66         struct task_desc        *wakee;
  67 };
  68
  69 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
  70
  71 enum thread_state {
  72         THREAD_SLEEPING = 0,
  73         THREAD_WAIT_CPU,
  74         THREAD_SCHED_IN,
  75         THREAD_IGNORE
  76 };
  77
  78 struct work_atom {
  79         struct list_head        list;
  80         enum thread_state       state;
  81         u64                     sched_out_time;
  82         u64                     wake_up_time;
  83         u64                     sched_in_time;
  84         u64                     runtime;
  85 };
  86
  87 struct work_atoms {
  88         struct list_head        work_list;
  89         struct thread           *thread;
  90         struct rb_node          node;
  91         u64                     max_lat;
  92         u64                     max_lat_at;
  93         u64                     total_lat;
  94         u64                     nb_atoms;
  95         u64                     total_runtime;
  96 };
  97
  98 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
  99
 100 struct perf_sched;
 101
 102 struct trace_sched_handler {
 103         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 104                             struct perf_sample *sample, struct machine *machine);
 105
 106         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 107                              struct perf_sample *sample, struct machine *machine);
 108
 109         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 110                             struct perf_sample *sample, struct machine *machine);
 111
 112         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 113         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 114                           struct machine *machine);
 115
 116         int (*migrate_task_event)(struct perf_sched *sched,
 117                                   struct perf_evsel *evsel,
 118                                   struct perf_sample *sample,
 119                                   struct machine *machine);
 120 };
 121
 122 struct perf_sched {
 123         struct perf_tool tool;
 124         const char       *sort_order;
 125         unsigned long    nr_tasks;
 126         struct task_desc *pid_to_task[MAX_PID];
 127         struct task_desc **tasks;
 128         const struct trace_sched_handler *tp_handler;
 129         pthread_mutex_t  start_work_mutex;
 130         pthread_mutex_t  work_done_wait_mutex;
 131         int              profile_cpu;
 132 /*
 133  * Track the current task - that way we can know whether there's any
 134  * weird events, such as a task being switched away that is not current.
 135  */
 136         int              max_cpu;
 137         u32              curr_pid[MAX_CPUS];
 138         struct thread    *curr_thread[MAX_CPUS];
 139         char             next_shortname1;
 140         char             next_shortname2;
 141         unsigned int     replay_repeat;
 142         unsigned long    nr_run_events;
 143         unsigned long    nr_sleep_events;
 144         unsigned long    nr_wakeup_events;
 145         unsigned long    nr_sleep_corrections;
 146         unsigned long    nr_run_events_optimized;
 147         unsigned long    targetless_wakeups;
 148         unsigned long    multitarget_wakeups;
 149         unsigned long    nr_runs;
 150         unsigned long    nr_timestamps;
 151         unsigned long    nr_unordered_timestamps;
 152         unsigned long    nr_state_machine_bugs;
 153         unsigned long    nr_context_switch_bugs;
 154         unsigned long    nr_events;
 155         unsigned long    nr_lost_chunks;
 156         unsigned long    nr_lost_events;
 157         u64              run_measurement_overhead;
 158         u64              sleep_measurement_overhead;
 159         u64              start_time;
 160         u64              cpu_usage;
 161         u64              runavg_cpu_usage;
 162         u64              parent_cpu_usage;
 163         u64              runavg_parent_cpu_usage;
 164         u64              sum_runtime;
 165         u64              sum_fluct;
 166         u64              run_avg;
 167         u64              all_runtime;
 168         u64              all_count;
 169         u64              cpu_last_switched[MAX_CPUS];
 170         struct rb_root   atom_root, sorted_atom_root;
 171         struct list_head sort_list, cmp_pid;
 172 };
 173
 174 static u64 get_nsecs(void)
 175 {
 176         struct timespec ts;
 177
 178         clock_gettime(CLOCK_MONOTONIC, &ts);
 179
 180         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 181 }
 182
 183 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 184 {
 185         u64 T0 = get_nsecs(), T1;
 186
 187         do {
 188                 T1 = get_nsecs();
 189         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 190 }
 191
 192 static void sleep_nsecs(u64 nsecs)
 193 {
 194         struct timespec ts;
 195
 196         ts.tv_nsec = nsecs % 999999999;
 197         ts.tv_sec = nsecs / 999999999;
 198
 199         nanosleep(&ts, NULL);
 200 }
 201
 202 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 203 {
 204         u64 T0, T1, delta, min_delta = 1000000000ULL;
 205         int i;
 206
 207         for (i = 0; i < 10; i++) {
 208                 T0 = get_nsecs();
 209                 burn_nsecs(sched, 0);
 210                 T1 = get_nsecs();
 211                 delta = T1-T0;
 212                 min_delta = min(min_delta, delta);
 213         }
 214         sched->run_measurement_overhead = min_delta;
 215
 216         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 217 }
 218
 219 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 220 {
 221         u64 T0, T1, delta, min_delta = 1000000000ULL;
 222         int i;
 223
 224         for (i = 0; i < 10; i++) {
 225                 T0 = get_nsecs();
 226                 sleep_nsecs(10000);
 227                 T1 = get_nsecs();
 228                 delta = T1-T0;
 229                 min_delta = min(min_delta, delta);
 230         }
 231         min_delta -= 10000;
 232         sched->sleep_measurement_overhead = min_delta;
 233
 234         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 235 }
 236
 237 static struct sched_atom *
 238 get_new_event(struct task_desc *task, u64 timestamp)
 239 {
 240         struct sched_atom *event = zalloc(sizeof(*event));
 241         unsigned long idx = task->nr_events;
 242         size_t size;
 243
 244         event->timestamp = timestamp;
 245         event->nr = idx;
 246
 247         task->nr_events++;
 248         size = sizeof(struct sched_atom *) * task->nr_events;
 249         task->atoms = realloc(task->atoms, size);
 250         BUG_ON(!task->atoms);
 251
 252         task->atoms[idx] = event;
 253
 254         return event;
 255 }
 256
 257 static struct sched_atom *last_event(struct task_desc *task)
 258 {
 259         if (!task->nr_events)
 260                 return NULL;
 261
 262         return task->atoms[task->nr_events - 1];
 263 }
 264
 265 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 266                                 u64 timestamp, u64 duration)
 267 {
 268         struct sched_atom *event, *curr_event = last_event(task);
 269
 270         /*
 271          * optimize an existing RUN event by merging this one
 272          * to it:
 273          */
 274         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 275                 sched->nr_run_events_optimized++;
 276                 curr_event->duration += duration;
 277                 return;
 278         }
 279
 280         event = get_new_event(task, timestamp);
 281
 282         event->type = SCHED_EVENT_RUN;
 283         event->duration = duration;
 284
 285         sched->nr_run_events++;
 286 }
 287
 288 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 289                                    u64 timestamp, struct task_desc *wakee)
 290 {
 291         struct sched_atom *event, *wakee_event;
 292
 293         event = get_new_event(task, timestamp);
 294         event->type = SCHED_EVENT_WAKEUP;
 295         event->wakee = wakee;
 296
 297         wakee_event = last_event(wakee);
 298         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 299                 sched->targetless_wakeups++;
 300                 return;
 301         }
 302         if (wakee_event->wait_sem) {
 303                 sched->multitarget_wakeups++;
 304                 return;
 305         }
 306
 307         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 308         sem_init(wakee_event->wait_sem, 0, 0);
 309         wakee_event->specific_wait = 1;
 310         event->wait_sem = wakee_event->wait_sem;
 311
 312         sched->nr_wakeup_events++;
 313 }
 314
 315 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 316                                   u64 timestamp, u64 task_state __maybe_unused)
 317 {
 318         struct sched_atom *event = get_new_event(task, timestamp);
 319
 320         event->type = SCHED_EVENT_SLEEP;
 321
 322         sched->nr_sleep_events++;
 323 }
 324
 325 static struct task_desc *register_pid(struct perf_sched *sched,
 326                                       unsigned long pid, const char *comm)
 327 {
 328         struct task_desc *task;
 329
 330         BUG_ON(pid >= MAX_PID);
 331
 332         task = sched->pid_to_task[pid];
 333
 334         if (task)
 335                 return task;
 336
 337         task = zalloc(sizeof(*task));
 338         task->pid = pid;
 339         task->nr = sched->nr_tasks;
 340         strcpy(task->comm, comm);
 341         /*
 342          * every task starts in sleeping state - this gets ignored
 343          * if there's no wakeup pointing to this sleep state:
 344          */
 345         add_sched_event_sleep(sched, task, 0, 0);
 346
 347         sched->pid_to_task[pid] = task;
 348         sched->nr_tasks++;
 349         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_task *));
 350         BUG_ON(!sched->tasks);
 351         sched->tasks[task->nr] = task;
 352
 353         if (verbose)
 354                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 355
 356         return task;
 357 }
 358
 359
 360 static void print_task_traces(struct perf_sched *sched)
 361 {
 362         struct task_desc *task;
 363         unsigned long i;
 364
 365         for (i = 0; i < sched->nr_tasks; i++) {
 366                 task = sched->tasks[i];
 367                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 368                         task->nr, task->comm, task->pid, task->nr_events);
 369         }
 370 }
 371
 372 static void add_cross_task_wakeups(struct perf_sched *sched)
 373 {
 374         struct task_desc *task1, *task2;
 375         unsigned long i, j;
 376
 377         for (i = 0; i < sched->nr_tasks; i++) {
 378                 task1 = sched->tasks[i];
 379                 j = i + 1;
 380                 if (j == sched->nr_tasks)
 381                         j = 0;
 382                 task2 = sched->tasks[j];
 383                 add_sched_event_wakeup(sched, task1, 0, task2);
 384         }
 385 }
 386
 387 static void perf_sched__process_event(struct perf_sched *sched,
 388                                       struct sched_atom *atom)
 389 {
 390         int ret = 0;
 391
 392         switch (atom->type) {
 393                 case SCHED_EVENT_RUN:
 394                         burn_nsecs(sched, atom->duration);
 395                         break;
 396                 case SCHED_EVENT_SLEEP:
 397                         if (atom->wait_sem)
 398                                 ret = sem_wait(atom->wait_sem);
 399                         BUG_ON(ret);
 400                         break;
 401                 case SCHED_EVENT_WAKEUP:
 402                         if (atom->wait_sem)
 403                                 ret = sem_post(atom->wait_sem);
 404                         BUG_ON(ret);
 405                         break;
 406                 case SCHED_EVENT_MIGRATION:
 407                         break;
 408                 default:
 409                         BUG_ON(1);
 410         }
 411 }
 412
 413 static u64 get_cpu_usage_nsec_parent(void)
 414 {
 415         struct rusage ru;
 416         u64 sum;
 417         int err;
 418
 419         err = getrusage(RUSAGE_SELF, &ru);
 420         BUG_ON(err);
 421
 422         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 423         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 424
 425         return sum;
 426 }
 427
 428 static int self_open_counters(void)
 429 {
 430         struct perf_event_attr attr;
 431         int fd;
 432
 433         memset(&attr, 0, sizeof(attr));
 434
 435         attr.type = PERF_TYPE_SOFTWARE;
 436         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 437
 438         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 439
 440         if (fd < 0)
 441                 pr_err("Error: sys_perf_event_open() syscall returned "
 442                        "with %d (%s)\n", fd, strerror(errno));
 443         return fd;
 444 }
 445
 446 static u64 get_cpu_usage_nsec_self(int fd)
 447 {
 448         u64 runtime;
 449         int ret;
 450
 451         ret = read(fd, &runtime, sizeof(runtime));
 452         BUG_ON(ret != sizeof(runtime));
 453
 454         return runtime;
 455 }
 456
 457 struct sched_thread_parms {
 458         struct task_desc  *task;
 459         struct perf_sched *sched;
 460 };
 461
 462 static void *thread_func(void *ctx)
 463 {
 464         struct sched_thread_parms *parms = ctx;
 465         struct task_desc *this_task = parms->task;
 466         struct perf_sched *sched = parms->sched;
 467         u64 cpu_usage_0, cpu_usage_1;
 468         unsigned long i, ret;
 469         char comm2[22];
 470         int fd;
 471
 472         free(parms);
 473
 474         sprintf(comm2, ":%s", this_task->comm);
 475         prctl(PR_SET_NAME, comm2);
 476         fd = self_open_counters();
 477         if (fd < 0)
 478                 return NULL;
 479 again:
 480         ret = sem_post(&this_task->ready_for_work);
 481         BUG_ON(ret);
 482         ret = pthread_mutex_lock(&sched->start_work_mutex);
 483         BUG_ON(ret);
 484         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 485         BUG_ON(ret);
 486
 487         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 488
 489         for (i = 0; i < this_task->nr_events; i++) {
 490                 this_task->curr_event = i;
 491                 perf_sched__process_event(sched, this_task->atoms[i]);
 492         }
 493
 494         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 495         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 496         ret = sem_post(&this_task->work_done_sem);
 497         BUG_ON(ret);
 498
 499         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 500         BUG_ON(ret);
 501         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 502         BUG_ON(ret);
 503
 504         goto again;
 505 }
 506
 507 static void create_tasks(struct perf_sched *sched)
 508 {
 509         struct task_desc *task;
 510         pthread_attr_t attr;
 511         unsigned long i;
 512         int err;
 513
 514         err = pthread_attr_init(&attr);
 515         BUG_ON(err);
 516         err = pthread_attr_setstacksize(&attr,
 517                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 518         BUG_ON(err);
 519         err = pthread_mutex_lock(&sched->start_work_mutex);
 520         BUG_ON(err);
 521         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 522         BUG_ON(err);
 523         for (i = 0; i < sched->nr_tasks; i++) {
 524                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 525                 BUG_ON(parms == NULL);
 526                 parms->task = task = sched->tasks[i];
 527                 parms->sched = sched;
 528                 sem_init(&task->sleep_sem, 0, 0);
 529                 sem_init(&task->ready_for_work, 0, 0);
 530                 sem_init(&task->work_done_sem, 0, 0);
 531                 task->curr_event = 0;
 532                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 533                 BUG_ON(err);
 534         }
 535 }
 536
 537 static void wait_for_tasks(struct perf_sched *sched)
 538 {
 539         u64 cpu_usage_0, cpu_usage_1;
 540         struct task_desc *task;
 541         unsigned long i, ret;
 542
 543         sched->start_time = get_nsecs();
 544         sched->cpu_usage = 0;
 545         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 546
 547         for (i = 0; i < sched->nr_tasks; i++) {
 548                 task = sched->tasks[i];
 549                 ret = sem_wait(&task->ready_for_work);
 550                 BUG_ON(ret);
 551                 sem_init(&task->ready_for_work, 0, 0);
 552         }
 553         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 554         BUG_ON(ret);
 555
 556         cpu_usage_0 = get_cpu_usage_nsec_parent();
 557
 558         pthread_mutex_unlock(&sched->start_work_mutex);
 559
 560         for (i = 0; i < sched->nr_tasks; i++) {
 561                 task = sched->tasks[i];
 562                 ret = sem_wait(&task->work_done_sem);
 563                 BUG_ON(ret);
 564                 sem_init(&task->work_done_sem, 0, 0);
 565                 sched->cpu_usage += task->cpu_usage;
 566                 task->cpu_usage = 0;
 567         }
 568
 569         cpu_usage_1 = get_cpu_usage_nsec_parent();
 570         if (!sched->runavg_cpu_usage)
 571                 sched->runavg_cpu_usage = sched->cpu_usage;
 572         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
 573
 574         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 575         if (!sched->runavg_parent_cpu_usage)
 576                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 577         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
 578                                          sched->parent_cpu_usage)/10;
 579
 580         ret = pthread_mutex_lock(&sched->start_work_mutex);
 581         BUG_ON(ret);
 582
 583         for (i = 0; i < sched->nr_tasks; i++) {
 584                 task = sched->tasks[i];
 585                 sem_init(&task->sleep_sem, 0, 0);
 586                 task->curr_event = 0;
 587         }
 588 }
 589
 590 static void run_one_test(struct perf_sched *sched)
 591 {
 592         u64 T0, T1, delta, avg_delta, fluct;
 593
 594         T0 = get_nsecs();
 595         wait_for_tasks(sched);
 596         T1 = get_nsecs();
 597
 598         delta = T1 - T0;
 599         sched->sum_runtime += delta;
 600         sched->nr_runs++;
 601
 602         avg_delta = sched->sum_runtime / sched->nr_runs;
 603         if (delta < avg_delta)
 604                 fluct = avg_delta - delta;
 605         else
 606                 fluct = delta - avg_delta;
 607         sched->sum_fluct += fluct;
 608         if (!sched->run_avg)
 609                 sched->run_avg = delta;
 610         sched->run_avg = (sched->run_avg * 9 + delta) / 10;
 611
 612         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 613
 614         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 615
 616         printf("cpu: %0.2f / %0.2f",
 617                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 618
 619 #if 0
 620         /*
 621          * rusage statistics done by the parent, these are less
 622          * accurate than the sched->sum_exec_runtime based statistics:
 623          */
 624         printf(" [%0.2f / %0.2f]",
 625                 (double)sched->parent_cpu_usage/1e6,
 626                 (double)sched->runavg_parent_cpu_usage/1e6);
 627 #endif
 628
 629         printf("\n");
 630
 631         if (sched->nr_sleep_corrections)
 632                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 633         sched->nr_sleep_corrections = 0;
 634 }
 635
 636 static void test_calibrations(struct perf_sched *sched)
 637 {
 638         u64 T0, T1;
 639
 640         T0 = get_nsecs();
 641         burn_nsecs(sched, 1e6);
 642         T1 = get_nsecs();
 643
 644         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 645
 646         T0 = get_nsecs();
 647         sleep_nsecs(1e6);
 648         T1 = get_nsecs();
 649
 650         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 651 }
 652
 653 static int
 654 replay_wakeup_event(struct perf_sched *sched,
 655                     struct perf_evsel *evsel, struct perf_sample *sample,
 656                     struct machine *machine __maybe_unused)
 657 {
 658         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 659         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 660         struct task_desc *waker, *wakee;
 661
 662         if (verbose) {
 663                 printf("sched_wakeup event %p\n", evsel);
 664
 665                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 666         }
 667
 668         waker = register_pid(sched, sample->tid, "<unknown>");
 669         wakee = register_pid(sched, pid, comm);
 670
 671         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 672         return 0;
 673 }
 674
 675 static int replay_switch_event(struct perf_sched *sched,
 676                                struct perf_evsel *evsel,
 677                                struct perf_sample *sample,
 678                                struct machine *machine __maybe_unused)
 679 {
 680         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 681                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 682         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 683                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 684         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 685         struct task_desc *prev, __maybe_unused *next;
 686         u64 timestamp0, timestamp = sample->time;
 687         int cpu = sample->cpu;
 688         s64 delta;
 689
 690         if (verbose)
 691                 printf("sched_switch event %p\n", evsel);
 692
 693         if (cpu >= MAX_CPUS || cpu < 0)
 694                 return 0;
 695
 696         timestamp0 = sched->cpu_last_switched[cpu];
 697         if (timestamp0)
 698                 delta = timestamp - timestamp0;
 699         else
 700                 delta = 0;
 701
 702         if (delta < 0) {
 703                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 704                 return -1;
 705         }
 706
 707         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 708                  prev_comm, prev_pid, next_comm, next_pid, delta);
 709
 710         prev = register_pid(sched, prev_pid, prev_comm);
 711         next = register_pid(sched, next_pid, next_comm);
 712
 713         sched->cpu_last_switched[cpu] = timestamp;
 714
 715         add_sched_event_run(sched, prev, timestamp, delta);
 716         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 717
 718         return 0;
 719 }
 720
 721 static int replay_fork_event(struct perf_sched *sched,
 722                              union perf_event *event,
 723                              struct machine *machine)
 724 {
 725         struct thread *child, *parent;
 726
 727         child = machine__findnew_thread(machine, event->fork.tid);
 728         parent = machine__findnew_thread(machine, event->fork.ptid);
 729
 730         if (child == NULL || parent == NULL) {
 731                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 732                                  child, parent);
 733                 return 0;
 734         }
 735
 736         if (verbose) {
 737                 printf("fork event\n");
 738                 printf("... parent: %s/%d\n", parent->comm, parent->tid);
 739                 printf("...  child: %s/%d\n", child->comm, child->tid);
 740         }
 741
 742         register_pid(sched, parent->tid, parent->comm);
 743         register_pid(sched, child->tid, child->comm);
 744         return 0;
 745 }
 746
 747 struct sort_dimension {
 748         const char              *name;
 749         sort_fn_t               cmp;
 750         struct list_head        list;
 751 };
 752
 753 static int
 754 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 755 {
 756         struct sort_dimension *sort;
 757         int ret = 0;
 758
 759         BUG_ON(list_empty(list));
 760
 761         list_for_each_entry(sort, list, list) {
 762                 ret = sort->cmp(l, r);
 763                 if (ret)
 764                         return ret;
 765         }
 766
 767         return ret;
 768 }
 769
 770 static struct work_atoms *
 771 thread_atoms_search(struct rb_root *root, struct thread *thread,
 772                          struct list_head *sort_list)
 773 {
 774         struct rb_node *node = root->rb_node;
 775         struct work_atoms key = { .thread = thread };
 776
 777         while (node) {
 778                 struct work_atoms *atoms;
 779                 int cmp;
 780
 781                 atoms = container_of(node, struct work_atoms, node);
 782
 783                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 784                 if (cmp > 0)
 785                         node = node->rb_left;
 786                 else if (cmp < 0)
 787                         node = node->rb_right;
 788                 else {
 789                         BUG_ON(thread != atoms->thread);
 790                         return atoms;
 791                 }
 792         }
 793         return NULL;
 794 }
 795
 796 static void
 797 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 798                          struct list_head *sort_list)
 799 {
 800         struct rb_node **new = &(root->rb_node), *parent = NULL;
 801
 802         while (*new) {
 803                 struct work_atoms *this;
 804                 int cmp;
 805
 806                 this = container_of(*new, struct work_atoms, node);
 807                 parent = *new;
 808
 809                 cmp = thread_lat_cmp(sort_list, data, this);
 810
 811                 if (cmp > 0)
 812                         new = &((*new)->rb_left);
 813                 else
 814                         new = &((*new)->rb_right);
 815         }
 816
 817         rb_link_node(&data->node, parent, new);
 818         rb_insert_color(&data->node, root);
 819 }
 820
 821 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 822 {
 823         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 824         if (!atoms) {
 825                 pr_err("No memory at %s\n", __func__);
 826                 return -1;
 827         }
 828
 829         atoms->thread = thread;
 830         INIT_LIST_HEAD(&atoms->work_list);
 831         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 832         return 0;
 833 }
 834
 835 static char sched_out_state(u64 prev_state)
 836 {
 837         const char *str = TASK_STATE_TO_CHAR_STR;
 838
 839         return str[prev_state];
 840 }
 841
 842 static int
 843 add_sched_out_event(struct work_atoms *atoms,
 844                     char run_state,
 845                     u64 timestamp)
 846 {
 847         struct work_atom *atom = zalloc(sizeof(*atom));
 848         if (!atom) {
 849                 pr_err("Non memory at %s", __func__);
 850                 return -1;
 851         }
 852
 853         atom->sched_out_time = timestamp;
 854
 855         if (run_state == 'R') {
 856                 atom->state = THREAD_WAIT_CPU;
 857                 atom->wake_up_time = atom->sched_out_time;
 858         }
 859
 860         list_add_tail(&atom->list, &atoms->work_list);
 861         return 0;
 862 }
 863
 864 static void
 865 add_runtime_event(struct work_atoms *atoms, u64 delta,
 866                   u64 timestamp __maybe_unused)
 867 {
 868         struct work_atom *atom;
 869
 870         BUG_ON(list_empty(&atoms->work_list));
 871
 872         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 873
 874         atom->runtime += delta;
 875         atoms->total_runtime += delta;
 876 }
 877
 878 static void
 879 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 880 {
 881         struct work_atom *atom;
 882         u64 delta;
 883
 884         if (list_empty(&atoms->work_list))
 885                 return;
 886
 887         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 888
 889         if (atom->state != THREAD_WAIT_CPU)
 890                 return;
 891
 892         if (timestamp < atom->wake_up_time) {
 893                 atom->state = THREAD_IGNORE;
 894                 return;
 895         }
 896
 897         atom->state = THREAD_SCHED_IN;
 898         atom->sched_in_time = timestamp;
 899
 900         delta = atom->sched_in_time - atom->wake_up_time;
 901         atoms->total_lat += delta;
 902         if (delta > atoms->max_lat) {
 903                 atoms->max_lat = delta;
 904                 atoms->max_lat_at = timestamp;
 905         }
 906         atoms->nb_atoms++;
 907 }
 908
 909 static int latency_switch_event(struct perf_sched *sched,
 910                                 struct perf_evsel *evsel,
 911                                 struct perf_sample *sample,
 912                                 struct machine *machine)
 913 {
 914         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 915                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 916         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 917         struct work_atoms *out_events, *in_events;
 918         struct thread *sched_out, *sched_in;
 919         u64 timestamp0, timestamp = sample->time;
 920         int cpu = sample->cpu;
 921         s64 delta;
 922
 923         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 924
 925         timestamp0 = sched->cpu_last_switched[cpu];
 926         sched->cpu_last_switched[cpu] = timestamp;
 927         if (timestamp0)
 928                 delta = timestamp - timestamp0;
 929         else
 930                 delta = 0;
 931
 932         if (delta < 0) {
 933                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 934                 return -1;
 935         }
 936
 937         sched_out = machine__findnew_thread(machine, prev_pid);
 938         sched_in = machine__findnew_thread(machine, next_pid);
 939
 940         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 941         if (!out_events) {
 942                 if (thread_atoms_insert(sched, sched_out))
 943                         return -1;
 944                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 945                 if (!out_events) {
 946                         pr_err("out-event: Internal tree error");
 947                         return -1;
 948                 }
 949         }
 950         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 951                 return -1;
 952
 953         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 954         if (!in_events) {
 955                 if (thread_atoms_insert(sched, sched_in))
 956                         return -1;
 957                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 958                 if (!in_events) {
 959                         pr_err("in-event: Internal tree error");
 960                         return -1;
 961                 }
 962                 /*
 963                  * Take came in we have not heard about yet,
 964                  * add in an initial atom in runnable state:
 965                  */
 966                 if (add_sched_out_event(in_events, 'R', timestamp))
 967                         return -1;
 968         }
 969         add_sched_in_event(in_events, timestamp);
 970
 971         return 0;
 972 }
 973
 974 static int latency_runtime_event(struct perf_sched *sched,
 975                                  struct perf_evsel *evsel,
 976                                  struct perf_sample *sample,
 977                                  struct machine *machine)
 978 {
 979         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
 980         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
 981         struct thread *thread = machine__findnew_thread(machine, pid);
 982         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 983         u64 timestamp = sample->time;
 984         int cpu = sample->cpu;
 985
 986         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 987         if (!atoms) {
 988                 if (thread_atoms_insert(sched, thread))
 989                         return -1;
 990                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 991                 if (!atoms) {
 992                         pr_err("in-event: Internal tree error");
 993                         return -1;
 994                 }
 995                 if (add_sched_out_event(atoms, 'R', timestamp))
 996                         return -1;
 997         }
 998
 999         add_runtime_event(atoms, runtime, timestamp);
1000         return 0;
1001 }
1002
1003 static int latency_wakeup_event(struct perf_sched *sched,
1004                                 struct perf_evsel *evsel,
1005                                 struct perf_sample *sample,
1006                                 struct machine *machine)
1007 {
1008         const u32 pid     = perf_evsel__intval(evsel, sample, "pid"),
1009                   success = perf_evsel__intval(evsel, sample, "success");
1010         struct work_atoms *atoms;
1011         struct work_atom *atom;
1012         struct thread *wakee;
1013         u64 timestamp = sample->time;
1014
1015         /* Note for later, it may be interesting to observe the failing cases */
1016         if (!success)
1017                 return 0;
1018
1019         wakee = machine__findnew_thread(machine, pid);
1020         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1021         if (!atoms) {
1022                 if (thread_atoms_insert(sched, wakee))
1023                         return -1;
1024                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1025                 if (!atoms) {
1026                         pr_err("wakeup-event: Internal tree error");
1027                         return -1;
1028                 }
1029                 if (add_sched_out_event(atoms, 'S', timestamp))
1030                         return -1;
1031         }
1032
1033         BUG_ON(list_empty(&atoms->work_list));
1034
1035         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1036
1037         /*
1038          * You WILL be missing events if you've recorded only
1039          * one CPU, or are only looking at only one, so don't
1040          * make useless noise.
1041          */
1042         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1043                 sched->nr_state_machine_bugs++;
1044
1045         sched->nr_timestamps++;
1046         if (atom->sched_out_time > timestamp) {
1047                 sched->nr_unordered_timestamps++;
1048                 return 0;
1049         }
1050
1051         atom->state = THREAD_WAIT_CPU;
1052         atom->wake_up_time = timestamp;
1053         return 0;
1054 }
1055
1056 static int latency_migrate_task_event(struct perf_sched *sched,
1057                                       struct perf_evsel *evsel,
1058                                       struct perf_sample *sample,
1059                                       struct machine *machine)
1060 {
1061         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1062         u64 timestamp = sample->time;
1063         struct work_atoms *atoms;
1064         struct work_atom *atom;
1065         struct thread *migrant;
1066
1067         /*
1068          * Only need to worry about migration when profiling one CPU.
1069          */
1070         if (sched->profile_cpu == -1)
1071                 return 0;
1072
1073         migrant = machine__findnew_thread(machine, pid);
1074         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1075         if (!atoms) {
1076                 if (thread_atoms_insert(sched, migrant))
1077                         return -1;
1078                 register_pid(sched, migrant->tid, migrant->comm);
1079                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1080                 if (!atoms) {
1081                         pr_err("migration-event: Internal tree error");
1082                         return -1;
1083                 }
1084                 if (add_sched_out_event(atoms, 'R', timestamp))
1085                         return -1;
1086         }
1087
1088         BUG_ON(list_empty(&atoms->work_list));
1089
1090         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1092
1093         sched->nr_timestamps++;
1094
1095         if (atom->sched_out_time > timestamp)
1096                 sched->nr_unordered_timestamps++;
1097
1098         return 0;
1099 }
1100
1101 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1102 {
1103         int i;
1104         int ret;
1105         u64 avg;
1106
1107         if (!work_list->nb_atoms)
1108                 return;
1109         /*
1110          * Ignore idle threads:
1111          */
1112         if (!strcmp(work_list->thread->comm, "swapper"))
1113                 return;
1114
1115         sched->all_runtime += work_list->total_runtime;
1116         sched->all_count   += work_list->nb_atoms;
1117
1118         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->tid);
1119
1120         for (i = 0; i < 24 - ret; i++)
1121                 printf(" ");
1122
1123         avg = work_list->total_lat / work_list->nb_atoms;
1124
1125         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1126               (double)work_list->total_runtime / 1e6,
1127                  work_list->nb_atoms, (double)avg / 1e6,
1128                  (double)work_list->max_lat / 1e6,
1129                  (double)work_list->max_lat_at / 1e9);
1130 }
1131
1132 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1133 {
1134         if (l->thread->tid < r->thread->tid)
1135                 return -1;
1136         if (l->thread->tid > r->thread->tid)
1137                 return 1;
1138
1139         return 0;
1140 }
1141
1142 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1143 {
1144         u64 avgl, avgr;
1145
1146         if (!l->nb_atoms)
1147                 return -1;
1148
1149         if (!r->nb_atoms)
1150                 return 1;
1151
1152         avgl = l->total_lat / l->nb_atoms;
1153         avgr = r->total_lat / r->nb_atoms;
1154
1155         if (avgl < avgr)
1156                 return -1;
1157         if (avgl > avgr)
1158                 return 1;
1159
1160         return 0;
1161 }
1162
1163 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1164 {
1165         if (l->max_lat < r->max_lat)
1166                 return -1;
1167         if (l->max_lat > r->max_lat)
1168                 return 1;
1169
1170         return 0;
1171 }
1172
1173 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1174 {
1175         if (l->nb_atoms < r->nb_atoms)
1176                 return -1;
1177         if (l->nb_atoms > r->nb_atoms)
1178                 return 1;
1179
1180         return 0;
1181 }
1182
1183 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1184 {
1185         if (l->total_runtime < r->total_runtime)
1186                 return -1;
1187         if (l->total_runtime > r->total_runtime)
1188                 return 1;
1189
1190         return 0;
1191 }
1192
1193 static int sort_dimension__add(const char *tok, struct list_head *list)
1194 {
1195         size_t i;
1196         static struct sort_dimension avg_sort_dimension = {
1197                 .name = "avg",
1198                 .cmp  = avg_cmp,
1199         };
1200         static struct sort_dimension max_sort_dimension = {
1201                 .name = "max",
1202                 .cmp  = max_cmp,
1203         };
1204         static struct sort_dimension pid_sort_dimension = {
1205                 .name = "pid",
1206                 .cmp  = pid_cmp,
1207         };
1208         static struct sort_dimension runtime_sort_dimension = {
1209                 .name = "runtime",
1210                 .cmp  = runtime_cmp,
1211         };
1212         static struct sort_dimension switch_sort_dimension = {
1213                 .name = "switch",
1214                 .cmp  = switch_cmp,
1215         };
1216         struct sort_dimension *available_sorts[] = {
1217                 &pid_sort_dimension,
1218                 &avg_sort_dimension,
1219                 &max_sort_dimension,
1220                 &switch_sort_dimension,
1221                 &runtime_sort_dimension,
1222         };
1223
1224         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1225                 if (!strcmp(available_sorts[i]->name, tok)) {
1226                         list_add_tail(&available_sorts[i]->list, list);
1227
1228                         return 0;
1229                 }
1230         }
1231
1232         return -1;
1233 }
1234
1235 static void perf_sched__sort_lat(struct perf_sched *sched)
1236 {
1237         struct rb_node *node;
1238
1239         for (;;) {
1240                 struct work_atoms *data;
1241                 node = rb_first(&sched->atom_root);
1242                 if (!node)
1243                         break;
1244
1245                 rb_erase(node, &sched->atom_root);
1246                 data = rb_entry(node, struct work_atoms, node);
1247                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1248         }
1249 }
1250
1251 static int process_sched_wakeup_event(struct perf_tool *tool,
1252                                       struct perf_evsel *evsel,
1253                                       struct perf_sample *sample,
1254                                       struct machine *machine)
1255 {
1256         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1257
1258         if (sched->tp_handler->wakeup_event)
1259                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1260
1261         return 0;
1262 }
1263
1264 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1265                             struct perf_sample *sample, struct machine *machine)
1266 {
1267         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1268                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1269         struct thread *sched_out __maybe_unused, *sched_in;
1270         int new_shortname;
1271         u64 timestamp0, timestamp = sample->time;
1272         s64 delta;
1273         int cpu, this_cpu = sample->cpu;
1274
1275         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1276
1277         if (this_cpu > sched->max_cpu)
1278                 sched->max_cpu = this_cpu;
1279
1280         timestamp0 = sched->cpu_last_switched[this_cpu];
1281         sched->cpu_last_switched[this_cpu] = timestamp;
1282         if (timestamp0)
1283                 delta = timestamp - timestamp0;
1284         else
1285                 delta = 0;
1286
1287         if (delta < 0) {
1288                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1289                 return -1;
1290         }
1291
1292         sched_out = machine__findnew_thread(machine, prev_pid);
1293         sched_in = machine__findnew_thread(machine, next_pid);
1294
1295         sched->curr_thread[this_cpu] = sched_in;
1296
1297         printf("  ");
1298
1299         new_shortname = 0;
1300         if (!sched_in->shortname[0]) {
1301                 sched_in->shortname[0] = sched->next_shortname1;
1302                 sched_in->shortname[1] = sched->next_shortname2;
1303
1304                 if (sched->next_shortname1 < 'Z') {
1305                         sched->next_shortname1++;
1306                 } else {
1307                         sched->next_shortname1='A';
1308                         if (sched->next_shortname2 < '9') {
1309                                 sched->next_shortname2++;
1310                         } else {
1311                                 sched->next_shortname2='0';
1312                         }
1313                 }
1314                 new_shortname = 1;
1315         }
1316
1317         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1318                 if (cpu != this_cpu)
1319                         printf(" ");
1320                 else
1321                         printf("*");
1322
1323                 if (sched->curr_thread[cpu]) {
1324                         if (sched->curr_thread[cpu]->tid)
1325                                 printf("%2s ", sched->curr_thread[cpu]->shortname);
1326                         else
1327                                 printf(".  ");
1328                 } else
1329                         printf("   ");
1330         }
1331
1332         printf("  %12.6f secs ", (double)timestamp/1e9);
1333         if (new_shortname) {
1334                 printf("%s => %s:%d\n",
1335                         sched_in->shortname, sched_in->comm, sched_in->tid);
1336         } else {
1337                 printf("\n");
1338         }
1339
1340         return 0;
1341 }
1342
1343 static int process_sched_switch_event(struct perf_tool *tool,
1344                                       struct perf_evsel *evsel,
1345                                       struct perf_sample *sample,
1346                                       struct machine *machine)
1347 {
1348         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1349         int this_cpu = sample->cpu, err = 0;
1350         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1351             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1352
1353         if (sched->curr_pid[this_cpu] != (u32)-1) {
1354                 /*
1355                  * Are we trying to switch away a PID that is
1356                  * not current?
1357                  */
1358                 if (sched->curr_pid[this_cpu] != prev_pid)
1359                         sched->nr_context_switch_bugs++;
1360         }
1361
1362         if (sched->tp_handler->switch_event)
1363                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1364
1365         sched->curr_pid[this_cpu] = next_pid;
1366         return err;
1367 }
1368
1369 static int process_sched_runtime_event(struct perf_tool *tool,
1370                                        struct perf_evsel *evsel,
1371                                        struct perf_sample *sample,
1372                                        struct machine *machine)
1373 {
1374         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1375
1376         if (sched->tp_handler->runtime_event)
1377                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1378
1379         return 0;
1380 }
1381
1382 static int perf_sched__process_fork_event(struct perf_tool *tool,
1383                                           union perf_event *event,
1384                                           struct perf_sample *sample,
1385                                           struct machine *machine)
1386 {
1387         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1388
1389         /* run the fork event through the perf machineruy */
1390         perf_event__process_fork(tool, event, sample, machine);
1391
1392         /* and then run additional processing needed for this command */
1393         if (sched->tp_handler->fork_event)
1394                 return sched->tp_handler->fork_event(sched, event, machine);
1395
1396         return 0;
1397 }
1398
1399 static int process_sched_migrate_task_event(struct perf_tool *tool,
1400                                             struct perf_evsel *evsel,
1401                                             struct perf_sample *sample,
1402                                             struct machine *machine)
1403 {
1404         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1405
1406         if (sched->tp_handler->migrate_task_event)
1407                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1408
1409         return 0;
1410 }
1411
1412 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1413                                   struct perf_evsel *evsel,
1414                                   struct perf_sample *sample,
1415                                   struct machine *machine);
1416
1417 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1418                                                  union perf_event *event __maybe_unused,
1419                                                  struct perf_sample *sample,
1420                                                  struct perf_evsel *evsel,
1421                                                  struct machine *machine)
1422 {
1423         int err = 0;
1424
1425         evsel->hists.stats.total_period += sample->period;
1426         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
1427
1428         if (evsel->handler.func != NULL) {
1429                 tracepoint_handler f = evsel->handler.func;
1430                 err = f(tool, evsel, sample, machine);
1431         }
1432
1433         return err;
1434 }
1435
1436 static int perf_sched__read_events(struct perf_sched *sched,
1437                                    struct perf_session **psession)
1438 {
1439         const struct perf_evsel_str_handler handlers[] = {
1440                 { "sched:sched_switch",       process_sched_switch_event, },
1441                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1442                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1443                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1444                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1445         };
1446         struct perf_session *session;
1447
1448         session = perf_session__new(input_name, O_RDONLY, 0, false, &sched->tool);
1449         if (session == NULL) {
1450                 pr_debug("No Memory for session\n");
1451                 return -1;
1452         }
1453
1454         if (perf_session__set_tracepoints_handlers(session, handlers))
1455                 goto out_delete;
1456
1457         if (perf_session__has_traces(session, "record -R")) {
1458                 int err = perf_session__process_events(session, &sched->tool);
1459                 if (err) {
1460                         pr_err("Failed to process events, error %d", err);
1461                         goto out_delete;
1462                 }
1463
1464                 sched->nr_events      = session->stats.nr_events[0];
1465                 sched->nr_lost_events = session->stats.total_lost;
1466                 sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST];
1467         }
1468
1469         if (psession)
1470                 *psession = session;
1471         else
1472                 perf_session__delete(session);
1473
1474         return 0;
1475
1476 out_delete:
1477         perf_session__delete(session);
1478         return -1;
1479 }
1480
1481 static void print_bad_events(struct perf_sched *sched)
1482 {
1483         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1484                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1485                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1486                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1487         }
1488         if (sched->nr_lost_events && sched->nr_events) {
1489                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1490                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1491                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1492         }
1493         if (sched->nr_state_machine_bugs && sched->nr_timestamps) {
1494                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1495                         (double)sched->nr_state_machine_bugs/(double)sched->nr_timestamps*100.0,
1496                         sched->nr_state_machine_bugs, sched->nr_timestamps);
1497                 if (sched->nr_lost_events)
1498                         printf(" (due to lost events?)");
1499                 printf("\n");
1500         }
1501         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1502                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1503                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1504                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1505                 if (sched->nr_lost_events)
1506                         printf(" (due to lost events?)");
1507                 printf("\n");
1508         }
1509 }
1510
1511 static int perf_sched__lat(struct perf_sched *sched)
1512 {
1513         struct rb_node *next;
1514         struct perf_session *session;
1515
1516         setup_pager();
1517
1518         /* save session -- references to threads are held in work_list */
1519         if (perf_sched__read_events(sched, &session))
1520                 return -1;
1521
1522         perf_sched__sort_lat(sched);
1523
1524         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1525         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1526         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1527
1528         next = rb_first(&sched->sorted_atom_root);
1529
1530         while (next) {
1531                 struct work_atoms *work_list;
1532
1533                 work_list = rb_entry(next, struct work_atoms, node);
1534                 output_lat_thread(sched, work_list);
1535                 next = rb_next(next);
1536         }
1537
1538         printf(" -----------------------------------------------------------------------------------------\n");
1539         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1540                 (double)sched->all_runtime / 1e6, sched->all_count);
1541
1542         printf(" ---------------------------------------------------\n");
1543
1544         print_bad_events(sched);
1545         printf("\n");
1546
1547         perf_session__delete(session);
1548         return 0;
1549 }
1550
1551 static int perf_sched__map(struct perf_sched *sched)
1552 {
1553         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1554
1555         setup_pager();
1556         if (perf_sched__read_events(sched, NULL))
1557                 return -1;
1558         print_bad_events(sched);
1559         return 0;
1560 }
1561
1562 static int perf_sched__replay(struct perf_sched *sched)
1563 {
1564         unsigned long i;
1565
1566         calibrate_run_measurement_overhead(sched);
1567         calibrate_sleep_measurement_overhead(sched);
1568
1569         test_calibrations(sched);
1570
1571         if (perf_sched__read_events(sched, NULL))
1572                 return -1;
1573
1574         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1575         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1576         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1577
1578         if (sched->targetless_wakeups)
1579                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1580         if (sched->multitarget_wakeups)
1581                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1582         if (sched->nr_run_events_optimized)
1583                 printf("run atoms optimized: %ld\n",
1584                         sched->nr_run_events_optimized);
1585
1586         print_task_traces(sched);
1587         add_cross_task_wakeups(sched);
1588
1589         create_tasks(sched);
1590         printf("------------------------------------------------------------\n");
1591         for (i = 0; i < sched->replay_repeat; i++)
1592                 run_one_test(sched);
1593
1594         return 0;
1595 }
1596
1597 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1598                           const char * const usage_msg[])
1599 {
1600         char *tmp, *tok, *str = strdup(sched->sort_order);
1601
1602         for (tok = strtok_r(str, ", ", &tmp);
1603                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1604                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1605                         error("Unknown --sort key: `%s'", tok);
1606                         usage_with_options(usage_msg, options);
1607                 }
1608         }
1609
1610         free(str);
1611
1612         sort_dimension__add("pid", &sched->cmp_pid);
1613 }
1614
1615 static int __cmd_record(int argc, const char **argv)
1616 {
1617         unsigned int rec_argc, i, j;
1618         const char **rec_argv;
1619         const char * const record_args[] = {
1620                 "record",
1621                 "-a",
1622                 "-R",
1623                 "-m", "1024",
1624                 "-c", "1",
1625                 "-e", "sched:sched_switch",
1626                 "-e", "sched:sched_stat_wait",
1627                 "-e", "sched:sched_stat_sleep",
1628                 "-e", "sched:sched_stat_iowait",
1629                 "-e", "sched:sched_stat_runtime",
1630                 "-e", "sched:sched_process_fork",
1631                 "-e", "sched:sched_wakeup",
1632                 "-e", "sched:sched_migrate_task",
1633         };
1634
1635         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1636         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1637
1638         if (rec_argv == NULL)
1639                 return -ENOMEM;
1640
1641         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1642                 rec_argv[i] = strdup(record_args[i]);
1643
1644         for (j = 1; j < (unsigned int)argc; j++, i++)
1645                 rec_argv[i] = argv[j];
1646
1647         BUG_ON(i != rec_argc);
1648
1649         return cmd_record(i, rec_argv, NULL);
1650 }
1651
1652 static const char default_sort_order[] = "avg, max, switch, runtime";
1653 static struct perf_sched sched = {
1654         .tool = {
1655                 .sample          = perf_sched__process_tracepoint_sample,
1656                 .comm            = perf_event__process_comm,
1657                 .lost            = perf_event__process_lost,
1658                 .fork            = perf_sched__process_fork_event,
1659                 .ordered_samples = true,
1660         },
1661         .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1662         .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1663         .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1664         .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1665         .curr_pid             = { [0 ... MAX_CPUS - 1] = -1 },
1666         .sort_order           = default_sort_order,
1667         .replay_repeat        = 10,
1668         .profile_cpu          = -1,
1669         .next_shortname1      = 'A',
1670         .next_shortname2      = '0',
1671 };
1672
1673 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1674 {
1675         const struct option latency_options[] = {
1676         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1677                    "sort by key(s): runtime, switch, avg, max"),
1678         OPT_INCR('v', "verbose", &verbose,
1679                     "be more verbose (show symbol address, etc)"),
1680         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1681                     "CPU to profile on"),
1682         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1683                     "dump raw trace in ASCII"),
1684         OPT_END()
1685         };
1686         const struct option replay_options[] = {
1687         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1688                      "repeat the workload replay N times (-1: infinite)"),
1689         OPT_INCR('v', "verbose", &verbose,
1690                     "be more verbose (show symbol address, etc)"),
1691         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1692                     "dump raw trace in ASCII"),
1693         OPT_END()
1694         };
1695         const struct option sched_options[] = {
1696         OPT_STRING('i', "input", &input_name, "file",
1697                     "input file name"),
1698         OPT_INCR('v', "verbose", &verbose,
1699                     "be more verbose (show symbol address, etc)"),
1700         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1701                     "dump raw trace in ASCII"),
1702         OPT_END()
1703         };
1704         const char * const latency_usage[] = {
1705                 "perf sched latency [<options>]",
1706                 NULL
1707         };
1708         const char * const replay_usage[] = {
1709                 "perf sched replay [<options>]",
1710                 NULL
1711         };
1712         const char * const sched_usage[] = {
1713                 "perf sched [<options>] {record|latency|map|replay|script}",
1714                 NULL
1715         };
1716         struct trace_sched_handler lat_ops  = {
1717                 .wakeup_event       = latency_wakeup_event,
1718                 .switch_event       = latency_switch_event,
1719                 .runtime_event      = latency_runtime_event,
1720                 .migrate_task_event = latency_migrate_task_event,
1721         };
1722         struct trace_sched_handler map_ops  = {
1723                 .switch_event       = map_switch_event,
1724         };
1725         struct trace_sched_handler replay_ops  = {
1726                 .wakeup_event       = replay_wakeup_event,
1727                 .switch_event       = replay_switch_event,
1728                 .fork_event         = replay_fork_event,
1729         };
1730
1731         argc = parse_options(argc, argv, sched_options, sched_usage,
1732                              PARSE_OPT_STOP_AT_NON_OPTION);
1733         if (!argc)
1734                 usage_with_options(sched_usage, sched_options);
1735
1736         /*
1737          * Aliased to 'perf script' for now:
1738          */
1739         if (!strcmp(argv[0], "script"))
1740                 return cmd_script(argc, argv, prefix);
1741
1742         symbol__init();
1743         if (!strncmp(argv[0], "rec", 3)) {
1744                 return __cmd_record(argc, argv);
1745         } else if (!strncmp(argv[0], "lat", 3)) {
1746                 sched.tp_handler = &lat_ops;
1747                 if (argc > 1) {
1748                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1749                         if (argc)
1750                                 usage_with_options(latency_usage, latency_options);
1751                 }
1752                 setup_sorting(&sched, latency_options, latency_usage);
1753                 return perf_sched__lat(&sched);
1754         } else if (!strcmp(argv[0], "map")) {
1755                 sched.tp_handler = &map_ops;
1756                 setup_sorting(&sched, latency_options, latency_usage);
1757                 return perf_sched__map(&sched);
1758         } else if (!strncmp(argv[0], "rep", 3)) {
1759                 sched.tp_handler = &replay_ops;
1760                 if (argc) {
1761                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1762                         if (argc)
1763                                 usage_with_options(replay_usage, replay_options);
1764                 }
1765                 return perf_sched__replay(&sched);
1766         } else {
1767                 usage_with_options(sched_usage, sched_options);
1768         }
1769
1770         return 0;
1771 }