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 #include "util/cloexec.h"
  14
  15 #include "util/parse-options.h"
  16 #include "util/trace-event.h"
  17
  18 #include "util/debug.h"
  19
  20 #include <sys/prctl.h>
  21 #include <sys/resource.h>
  22
  23 #include <semaphore.h>
  24 #include <pthread.h>
  25 #include <math.h>
  26
  27 #define PR_SET_NAME             15               /* Set process name */
  28 #define MAX_CPUS                4096
  29 #define COMM_LEN                20
  30 #define SYM_LEN                 129
  31 #define MAX_PID                 65536
  32
  33 struct sched_atom;
  34
  35 struct task_desc {
  36         unsigned long           nr;
  37         unsigned long           pid;
  38         char                    comm[COMM_LEN];
  39
  40         unsigned long           nr_events;
  41         unsigned long           curr_event;
  42         struct sched_atom       **atoms;
  43
  44         pthread_t               thread;
  45         sem_t                   sleep_sem;
  46
  47         sem_t                   ready_for_work;
  48         sem_t                   work_done_sem;
  49
  50         u64                     cpu_usage;
  51 };
  52
  53 enum sched_event_type {
  54         SCHED_EVENT_RUN,
  55         SCHED_EVENT_SLEEP,
  56         SCHED_EVENT_WAKEUP,
  57         SCHED_EVENT_MIGRATION,
  58 };
  59
  60 struct sched_atom {
  61         enum sched_event_type   type;
  62         int                     specific_wait;
  63         u64                     timestamp;
  64         u64                     duration;
  65         unsigned long           nr;
  66         sem_t                   *wait_sem;
  67         struct task_desc        *wakee;
  68 };
  69
  70 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  71
  72 enum thread_state {
  73         THREAD_SLEEPING = 0,
  74         THREAD_WAIT_CPU,
  75         THREAD_SCHED_IN,
  76         THREAD_IGNORE
  77 };
  78
  79 struct work_atom {
  80         struct list_head        list;
  81         enum thread_state       state;
  82         u64                     sched_out_time;
  83         u64                     wake_up_time;
  84         u64                     sched_in_time;
  85         u64                     runtime;
  86 };
  87
  88 struct work_atoms {
  89         struct list_head        work_list;
  90         struct thread           *thread;
  91         struct rb_node          node;
  92         u64                     max_lat;
  93         u64                     max_lat_at;
  94         u64                     total_lat;
  95         u64                     nb_atoms;
  96         u64                     total_runtime;
  97 };
  98
  99 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 100
 101 struct perf_sched;
 102
 103 struct trace_sched_handler {
 104         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 105                             struct perf_sample *sample, struct machine *machine);
 106
 107         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 108                              struct perf_sample *sample, struct machine *machine);
 109
 110         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 111                             struct perf_sample *sample, struct machine *machine);
 112
 113         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 114         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 115                           struct machine *machine);
 116
 117         int (*migrate_task_event)(struct perf_sched *sched,
 118                                   struct perf_evsel *evsel,
 119                                   struct perf_sample *sample,
 120                                   struct machine *machine);
 121 };
 122
 123 struct perf_sched {
 124         struct perf_tool tool;
 125         const char       *sort_order;
 126         unsigned long    nr_tasks;
 127         struct task_desc *pid_to_task[MAX_PID];
 128         struct task_desc **tasks;
 129         const struct trace_sched_handler *tp_handler;
 130         pthread_mutex_t  start_work_mutex;
 131         pthread_mutex_t  work_done_wait_mutex;
 132         int              profile_cpu;
 133 /*
 134  * Track the current task - that way we can know whether there's any
 135  * weird events, such as a task being switched away that is not current.
 136  */
 137         int              max_cpu;
 138         u32              curr_pid[MAX_CPUS];
 139         struct thread    *curr_thread[MAX_CPUS];
 140         char             next_shortname1;
 141         char             next_shortname2;
 142         unsigned int     replay_repeat;
 143         unsigned long    nr_run_events;
 144         unsigned long    nr_sleep_events;
 145         unsigned long    nr_wakeup_events;
 146         unsigned long    nr_sleep_corrections;
 147         unsigned long    nr_run_events_optimized;
 148         unsigned long    targetless_wakeups;
 149         unsigned long    multitarget_wakeups;
 150         unsigned long    nr_runs;
 151         unsigned long    nr_timestamps;
 152         unsigned long    nr_unordered_timestamps;
 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         char sbuf[STRERR_BUFSIZE];
 432         int fd;
 433
 434         memset(&attr, 0, sizeof(attr));
 435
 436         attr.type = PERF_TYPE_SOFTWARE;
 437         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 438
 439         fd = sys_perf_event_open(&attr, 0, -1, -1,
 440                                  perf_event_open_cloexec_flag());
 441
 442         if (fd < 0)
 443                 pr_err("Error: sys_perf_event_open() syscall returned "
 444                        "with %d (%s)\n", fd,
 445                        strerror_r(errno, sbuf, sizeof(sbuf)));
 446         return fd;
 447 }
 448
 449 static u64 get_cpu_usage_nsec_self(int fd)
 450 {
 451         u64 runtime;
 452         int ret;
 453
 454         ret = read(fd, &runtime, sizeof(runtime));
 455         BUG_ON(ret != sizeof(runtime));
 456
 457         return runtime;
 458 }
 459
 460 struct sched_thread_parms {
 461         struct task_desc  *task;
 462         struct perf_sched *sched;
 463 };
 464
 465 static void *thread_func(void *ctx)
 466 {
 467         struct sched_thread_parms *parms = ctx;
 468         struct task_desc *this_task = parms->task;
 469         struct perf_sched *sched = parms->sched;
 470         u64 cpu_usage_0, cpu_usage_1;
 471         unsigned long i, ret;
 472         char comm2[22];
 473         int fd;
 474
 475         zfree(&parms);
 476
 477         sprintf(comm2, ":%s", this_task->comm);
 478         prctl(PR_SET_NAME, comm2);
 479         fd = self_open_counters();
 480         if (fd < 0)
 481                 return NULL;
 482 again:
 483         ret = sem_post(&this_task->ready_for_work);
 484         BUG_ON(ret);
 485         ret = pthread_mutex_lock(&sched->start_work_mutex);
 486         BUG_ON(ret);
 487         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 488         BUG_ON(ret);
 489
 490         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 491
 492         for (i = 0; i < this_task->nr_events; i++) {
 493                 this_task->curr_event = i;
 494                 perf_sched__process_event(sched, this_task->atoms[i]);
 495         }
 496
 497         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 498         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 499         ret = sem_post(&this_task->work_done_sem);
 500         BUG_ON(ret);
 501
 502         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 503         BUG_ON(ret);
 504         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 505         BUG_ON(ret);
 506
 507         goto again;
 508 }
 509
 510 static void create_tasks(struct perf_sched *sched)
 511 {
 512         struct task_desc *task;
 513         pthread_attr_t attr;
 514         unsigned long i;
 515         int err;
 516
 517         err = pthread_attr_init(&attr);
 518         BUG_ON(err);
 519         err = pthread_attr_setstacksize(&attr,
 520                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 521         BUG_ON(err);
 522         err = pthread_mutex_lock(&sched->start_work_mutex);
 523         BUG_ON(err);
 524         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 525         BUG_ON(err);
 526         for (i = 0; i < sched->nr_tasks; i++) {
 527                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 528                 BUG_ON(parms == NULL);
 529                 parms->task = task = sched->tasks[i];
 530                 parms->sched = sched;
 531                 sem_init(&task->sleep_sem, 0, 0);
 532                 sem_init(&task->ready_for_work, 0, 0);
 533                 sem_init(&task->work_done_sem, 0, 0);
 534                 task->curr_event = 0;
 535                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 536                 BUG_ON(err);
 537         }
 538 }
 539
 540 static void wait_for_tasks(struct perf_sched *sched)
 541 {
 542         u64 cpu_usage_0, cpu_usage_1;
 543         struct task_desc *task;
 544         unsigned long i, ret;
 545
 546         sched->start_time = get_nsecs();
 547         sched->cpu_usage = 0;
 548         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 549
 550         for (i = 0; i < sched->nr_tasks; i++) {
 551                 task = sched->tasks[i];
 552                 ret = sem_wait(&task->ready_for_work);
 553                 BUG_ON(ret);
 554                 sem_init(&task->ready_for_work, 0, 0);
 555         }
 556         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 557         BUG_ON(ret);
 558
 559         cpu_usage_0 = get_cpu_usage_nsec_parent();
 560
 561         pthread_mutex_unlock(&sched->start_work_mutex);
 562
 563         for (i = 0; i < sched->nr_tasks; i++) {
 564                 task = sched->tasks[i];
 565                 ret = sem_wait(&task->work_done_sem);
 566                 BUG_ON(ret);
 567                 sem_init(&task->work_done_sem, 0, 0);
 568                 sched->cpu_usage += task->cpu_usage;
 569                 task->cpu_usage = 0;
 570         }
 571
 572         cpu_usage_1 = get_cpu_usage_nsec_parent();
 573         if (!sched->runavg_cpu_usage)
 574                 sched->runavg_cpu_usage = sched->cpu_usage;
 575         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
 576
 577         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 578         if (!sched->runavg_parent_cpu_usage)
 579                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 580         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
 581                                          sched->parent_cpu_usage)/10;
 582
 583         ret = pthread_mutex_lock(&sched->start_work_mutex);
 584         BUG_ON(ret);
 585
 586         for (i = 0; i < sched->nr_tasks; i++) {
 587                 task = sched->tasks[i];
 588                 sem_init(&task->sleep_sem, 0, 0);
 589                 task->curr_event = 0;
 590         }
 591 }
 592
 593 static void run_one_test(struct perf_sched *sched)
 594 {
 595         u64 T0, T1, delta, avg_delta, fluct;
 596
 597         T0 = get_nsecs();
 598         wait_for_tasks(sched);
 599         T1 = get_nsecs();
 600
 601         delta = T1 - T0;
 602         sched->sum_runtime += delta;
 603         sched->nr_runs++;
 604
 605         avg_delta = sched->sum_runtime / sched->nr_runs;
 606         if (delta < avg_delta)
 607                 fluct = avg_delta - delta;
 608         else
 609                 fluct = delta - avg_delta;
 610         sched->sum_fluct += fluct;
 611         if (!sched->run_avg)
 612                 sched->run_avg = delta;
 613         sched->run_avg = (sched->run_avg * 9 + delta) / 10;
 614
 615         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 616
 617         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 618
 619         printf("cpu: %0.2f / %0.2f",
 620                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 621
 622 #if 0
 623         /*
 624          * rusage statistics done by the parent, these are less
 625          * accurate than the sched->sum_exec_runtime based statistics:
 626          */
 627         printf(" [%0.2f / %0.2f]",
 628                 (double)sched->parent_cpu_usage/1e6,
 629                 (double)sched->runavg_parent_cpu_usage/1e6);
 630 #endif
 631
 632         printf("\n");
 633
 634         if (sched->nr_sleep_corrections)
 635                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 636         sched->nr_sleep_corrections = 0;
 637 }
 638
 639 static void test_calibrations(struct perf_sched *sched)
 640 {
 641         u64 T0, T1;
 642
 643         T0 = get_nsecs();
 644         burn_nsecs(sched, 1e6);
 645         T1 = get_nsecs();
 646
 647         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 648
 649         T0 = get_nsecs();
 650         sleep_nsecs(1e6);
 651         T1 = get_nsecs();
 652
 653         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 654 }
 655
 656 static int
 657 replay_wakeup_event(struct perf_sched *sched,
 658                     struct perf_evsel *evsel, struct perf_sample *sample,
 659                     struct machine *machine __maybe_unused)
 660 {
 661         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 662         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 663         struct task_desc *waker, *wakee;
 664
 665         if (verbose) {
 666                 printf("sched_wakeup event %p\n", evsel);
 667
 668                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 669         }
 670
 671         waker = register_pid(sched, sample->tid, "<unknown>");
 672         wakee = register_pid(sched, pid, comm);
 673
 674         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 675         return 0;
 676 }
 677
 678 static int replay_switch_event(struct perf_sched *sched,
 679                                struct perf_evsel *evsel,
 680                                struct perf_sample *sample,
 681                                struct machine *machine __maybe_unused)
 682 {
 683         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 684                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 685         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 686                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 687         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 688         struct task_desc *prev, __maybe_unused *next;
 689         u64 timestamp0, timestamp = sample->time;
 690         int cpu = sample->cpu;
 691         s64 delta;
 692
 693         if (verbose)
 694                 printf("sched_switch event %p\n", evsel);
 695
 696         if (cpu >= MAX_CPUS || cpu < 0)
 697                 return 0;
 698
 699         timestamp0 = sched->cpu_last_switched[cpu];
 700         if (timestamp0)
 701                 delta = timestamp - timestamp0;
 702         else
 703                 delta = 0;
 704
 705         if (delta < 0) {
 706                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 707                 return -1;
 708         }
 709
 710         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 711                  prev_comm, prev_pid, next_comm, next_pid, delta);
 712
 713         prev = register_pid(sched, prev_pid, prev_comm);
 714         next = register_pid(sched, next_pid, next_comm);
 715
 716         sched->cpu_last_switched[cpu] = timestamp;
 717
 718         add_sched_event_run(sched, prev, timestamp, delta);
 719         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 720
 721         return 0;
 722 }
 723
 724 static int replay_fork_event(struct perf_sched *sched,
 725                              union perf_event *event,
 726                              struct machine *machine)
 727 {
 728         struct thread *child, *parent;
 729
 730         child = machine__findnew_thread(machine, event->fork.pid,
 731                                         event->fork.tid);
 732         parent = machine__findnew_thread(machine, event->fork.ppid,
 733                                          event->fork.ptid);
 734
 735         if (child == NULL || parent == NULL) {
 736                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 737                                  child, parent);
 738                 return 0;
 739         }
 740
 741         if (verbose) {
 742                 printf("fork event\n");
 743                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 744                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 745         }
 746
 747         register_pid(sched, parent->tid, thread__comm_str(parent));
 748         register_pid(sched, child->tid, thread__comm_str(child));
 749         return 0;
 750 }
 751
 752 struct sort_dimension {
 753         const char              *name;
 754         sort_fn_t               cmp;
 755         struct list_head        list;
 756 };
 757
 758 static int
 759 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 760 {
 761         struct sort_dimension *sort;
 762         int ret = 0;
 763
 764         BUG_ON(list_empty(list));
 765
 766         list_for_each_entry(sort, list, list) {
 767                 ret = sort->cmp(l, r);
 768                 if (ret)
 769                         return ret;
 770         }
 771
 772         return ret;
 773 }
 774
 775 static struct work_atoms *
 776 thread_atoms_search(struct rb_root *root, struct thread *thread,
 777                          struct list_head *sort_list)
 778 {
 779         struct rb_node *node = root->rb_node;
 780         struct work_atoms key = { .thread = thread };
 781
 782         while (node) {
 783                 struct work_atoms *atoms;
 784                 int cmp;
 785
 786                 atoms = container_of(node, struct work_atoms, node);
 787
 788                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 789                 if (cmp > 0)
 790                         node = node->rb_left;
 791                 else if (cmp < 0)
 792                         node = node->rb_right;
 793                 else {
 794                         BUG_ON(thread != atoms->thread);
 795                         return atoms;
 796                 }
 797         }
 798         return NULL;
 799 }
 800
 801 static void
 802 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 803                          struct list_head *sort_list)
 804 {
 805         struct rb_node **new = &(root->rb_node), *parent = NULL;
 806
 807         while (*new) {
 808                 struct work_atoms *this;
 809                 int cmp;
 810
 811                 this = container_of(*new, struct work_atoms, node);
 812                 parent = *new;
 813
 814                 cmp = thread_lat_cmp(sort_list, data, this);
 815
 816                 if (cmp > 0)
 817                         new = &((*new)->rb_left);
 818                 else
 819                         new = &((*new)->rb_right);
 820         }
 821
 822         rb_link_node(&data->node, parent, new);
 823         rb_insert_color(&data->node, root);
 824 }
 825
 826 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 827 {
 828         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 829         if (!atoms) {
 830                 pr_err("No memory at %s\n", __func__);
 831                 return -1;
 832         }
 833
 834         atoms->thread = thread;
 835         INIT_LIST_HEAD(&atoms->work_list);
 836         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 837         return 0;
 838 }
 839
 840 static char sched_out_state(u64 prev_state)
 841 {
 842         const char *str = TASK_STATE_TO_CHAR_STR;
 843
 844         return str[prev_state];
 845 }
 846
 847 static int
 848 add_sched_out_event(struct work_atoms *atoms,
 849                     char run_state,
 850                     u64 timestamp)
 851 {
 852         struct work_atom *atom = zalloc(sizeof(*atom));
 853         if (!atom) {
 854                 pr_err("Non memory at %s", __func__);
 855                 return -1;
 856         }
 857
 858         atom->sched_out_time = timestamp;
 859
 860         if (run_state == 'R') {
 861                 atom->state = THREAD_WAIT_CPU;
 862                 atom->wake_up_time = atom->sched_out_time;
 863         }
 864
 865         list_add_tail(&atom->list, &atoms->work_list);
 866         return 0;
 867 }
 868
 869 static void
 870 add_runtime_event(struct work_atoms *atoms, u64 delta,
 871                   u64 timestamp __maybe_unused)
 872 {
 873         struct work_atom *atom;
 874
 875         BUG_ON(list_empty(&atoms->work_list));
 876
 877         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 878
 879         atom->runtime += delta;
 880         atoms->total_runtime += delta;
 881 }
 882
 883 static void
 884 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 885 {
 886         struct work_atom *atom;
 887         u64 delta;
 888
 889         if (list_empty(&atoms->work_list))
 890                 return;
 891
 892         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 893
 894         if (atom->state != THREAD_WAIT_CPU)
 895                 return;
 896
 897         if (timestamp < atom->wake_up_time) {
 898                 atom->state = THREAD_IGNORE;
 899                 return;
 900         }
 901
 902         atom->state = THREAD_SCHED_IN;
 903         atom->sched_in_time = timestamp;
 904
 905         delta = atom->sched_in_time - atom->wake_up_time;
 906         atoms->total_lat += delta;
 907         if (delta > atoms->max_lat) {
 908                 atoms->max_lat = delta;
 909                 atoms->max_lat_at = timestamp;
 910         }
 911         atoms->nb_atoms++;
 912 }
 913
 914 static int latency_switch_event(struct perf_sched *sched,
 915                                 struct perf_evsel *evsel,
 916                                 struct perf_sample *sample,
 917                                 struct machine *machine)
 918 {
 919         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 920                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 921         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 922         struct work_atoms *out_events, *in_events;
 923         struct thread *sched_out, *sched_in;
 924         u64 timestamp0, timestamp = sample->time;
 925         int cpu = sample->cpu;
 926         s64 delta;
 927
 928         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 929
 930         timestamp0 = sched->cpu_last_switched[cpu];
 931         sched->cpu_last_switched[cpu] = timestamp;
 932         if (timestamp0)
 933                 delta = timestamp - timestamp0;
 934         else
 935                 delta = 0;
 936
 937         if (delta < 0) {
 938                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 939                 return -1;
 940         }
 941
 942         sched_out = machine__findnew_thread(machine, -1, prev_pid);
 943         sched_in = machine__findnew_thread(machine, -1, next_pid);
 944
 945         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 946         if (!out_events) {
 947                 if (thread_atoms_insert(sched, sched_out))
 948                         return -1;
 949                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 950                 if (!out_events) {
 951                         pr_err("out-event: Internal tree error");
 952                         return -1;
 953                 }
 954         }
 955         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 956                 return -1;
 957
 958         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 959         if (!in_events) {
 960                 if (thread_atoms_insert(sched, sched_in))
 961                         return -1;
 962                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 963                 if (!in_events) {
 964                         pr_err("in-event: Internal tree error");
 965                         return -1;
 966                 }
 967                 /*
 968                  * Take came in we have not heard about yet,
 969                  * add in an initial atom in runnable state:
 970                  */
 971                 if (add_sched_out_event(in_events, 'R', timestamp))
 972                         return -1;
 973         }
 974         add_sched_in_event(in_events, timestamp);
 975
 976         return 0;
 977 }
 978
 979 static int latency_runtime_event(struct perf_sched *sched,
 980                                  struct perf_evsel *evsel,
 981                                  struct perf_sample *sample,
 982                                  struct machine *machine)
 983 {
 984         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
 985         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
 986         struct thread *thread = machine__findnew_thread(machine, -1, pid);
 987         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 988         u64 timestamp = sample->time;
 989         int cpu = sample->cpu;
 990
 991         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 992         if (!atoms) {
 993                 if (thread_atoms_insert(sched, thread))
 994                         return -1;
 995                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
 996                 if (!atoms) {
 997                         pr_err("in-event: Internal tree error");
 998                         return -1;
 999                 }
1000                 if (add_sched_out_event(atoms, 'R', timestamp))
1001                         return -1;
1002         }
1003
1004         add_runtime_event(atoms, runtime, timestamp);
1005         return 0;
1006 }
1007
1008 static int latency_wakeup_event(struct perf_sched *sched,
1009                                 struct perf_evsel *evsel,
1010                                 struct perf_sample *sample,
1011                                 struct machine *machine)
1012 {
1013         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1014         struct work_atoms *atoms;
1015         struct work_atom *atom;
1016         struct thread *wakee;
1017         u64 timestamp = sample->time;
1018
1019         wakee = machine__findnew_thread(machine, -1, 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          * As we do not guarantee the wakeup event happens when
1039          * task is out of run queue, also may happen when task is
1040          * on run queue and wakeup only change ->state to TASK_RUNNING,
1041          * then we should not set the ->wake_up_time when wake up a
1042          * task which is on run queue.
1043          *
1044          * You WILL be missing events if you've recorded only
1045          * one CPU, or are only looking at only one, so don't
1046          * skip in this case.
1047          */
1048         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1049                 return 0;
1050
1051         sched->nr_timestamps++;
1052         if (atom->sched_out_time > timestamp) {
1053                 sched->nr_unordered_timestamps++;
1054                 return 0;
1055         }
1056
1057         atom->state = THREAD_WAIT_CPU;
1058         atom->wake_up_time = timestamp;
1059         return 0;
1060 }
1061
1062 static int latency_migrate_task_event(struct perf_sched *sched,
1063                                       struct perf_evsel *evsel,
1064                                       struct perf_sample *sample,
1065                                       struct machine *machine)
1066 {
1067         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1068         u64 timestamp = sample->time;
1069         struct work_atoms *atoms;
1070         struct work_atom *atom;
1071         struct thread *migrant;
1072
1073         /*
1074          * Only need to worry about migration when profiling one CPU.
1075          */
1076         if (sched->profile_cpu == -1)
1077                 return 0;
1078
1079         migrant = machine__findnew_thread(machine, -1, pid);
1080         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1081         if (!atoms) {
1082                 if (thread_atoms_insert(sched, migrant))
1083                         return -1;
1084                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1085                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1086                 if (!atoms) {
1087                         pr_err("migration-event: Internal tree error");
1088                         return -1;
1089                 }
1090                 if (add_sched_out_event(atoms, 'R', timestamp))
1091                         return -1;
1092         }
1093
1094         BUG_ON(list_empty(&atoms->work_list));
1095
1096         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1097         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1098
1099         sched->nr_timestamps++;
1100
1101         if (atom->sched_out_time > timestamp)
1102                 sched->nr_unordered_timestamps++;
1103
1104         return 0;
1105 }
1106
1107 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1108 {
1109         int i;
1110         int ret;
1111         u64 avg;
1112
1113         if (!work_list->nb_atoms)
1114                 return;
1115         /*
1116          * Ignore idle threads:
1117          */
1118         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1119                 return;
1120
1121         sched->all_runtime += work_list->total_runtime;
1122         sched->all_count   += work_list->nb_atoms;
1123
1124         ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1125
1126         for (i = 0; i < 24 - ret; i++)
1127                 printf(" ");
1128
1129         avg = work_list->total_lat / work_list->nb_atoms;
1130
1131         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1132               (double)work_list->total_runtime / 1e6,
1133                  work_list->nb_atoms, (double)avg / 1e6,
1134                  (double)work_list->max_lat / 1e6,
1135                  (double)work_list->max_lat_at / 1e9);
1136 }
1137
1138 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1139 {
1140         if (l->thread->tid < r->thread->tid)
1141                 return -1;
1142         if (l->thread->tid > r->thread->tid)
1143                 return 1;
1144
1145         return 0;
1146 }
1147
1148 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1149 {
1150         u64 avgl, avgr;
1151
1152         if (!l->nb_atoms)
1153                 return -1;
1154
1155         if (!r->nb_atoms)
1156                 return 1;
1157
1158         avgl = l->total_lat / l->nb_atoms;
1159         avgr = r->total_lat / r->nb_atoms;
1160
1161         if (avgl < avgr)
1162                 return -1;
1163         if (avgl > avgr)
1164                 return 1;
1165
1166         return 0;
1167 }
1168
1169 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1170 {
1171         if (l->max_lat < r->max_lat)
1172                 return -1;
1173         if (l->max_lat > r->max_lat)
1174                 return 1;
1175
1176         return 0;
1177 }
1178
1179 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1180 {
1181         if (l->nb_atoms < r->nb_atoms)
1182                 return -1;
1183         if (l->nb_atoms > r->nb_atoms)
1184                 return 1;
1185
1186         return 0;
1187 }
1188
1189 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1190 {
1191         if (l->total_runtime < r->total_runtime)
1192                 return -1;
1193         if (l->total_runtime > r->total_runtime)
1194                 return 1;
1195
1196         return 0;
1197 }
1198
1199 static int sort_dimension__add(const char *tok, struct list_head *list)
1200 {
1201         size_t i;
1202         static struct sort_dimension avg_sort_dimension = {
1203                 .name = "avg",
1204                 .cmp  = avg_cmp,
1205         };
1206         static struct sort_dimension max_sort_dimension = {
1207                 .name = "max",
1208                 .cmp  = max_cmp,
1209         };
1210         static struct sort_dimension pid_sort_dimension = {
1211                 .name = "pid",
1212                 .cmp  = pid_cmp,
1213         };
1214         static struct sort_dimension runtime_sort_dimension = {
1215                 .name = "runtime",
1216                 .cmp  = runtime_cmp,
1217         };
1218         static struct sort_dimension switch_sort_dimension = {
1219                 .name = "switch",
1220                 .cmp  = switch_cmp,
1221         };
1222         struct sort_dimension *available_sorts[] = {
1223                 &pid_sort_dimension,
1224                 &avg_sort_dimension,
1225                 &max_sort_dimension,
1226                 &switch_sort_dimension,
1227                 &runtime_sort_dimension,
1228         };
1229
1230         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1231                 if (!strcmp(available_sorts[i]->name, tok)) {
1232                         list_add_tail(&available_sorts[i]->list, list);
1233
1234                         return 0;
1235                 }
1236         }
1237
1238         return -1;
1239 }
1240
1241 static void perf_sched__sort_lat(struct perf_sched *sched)
1242 {
1243         struct rb_node *node;
1244
1245         for (;;) {
1246                 struct work_atoms *data;
1247                 node = rb_first(&sched->atom_root);
1248                 if (!node)
1249                         break;
1250
1251                 rb_erase(node, &sched->atom_root);
1252                 data = rb_entry(node, struct work_atoms, node);
1253                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1254         }
1255 }
1256
1257 static int process_sched_wakeup_event(struct perf_tool *tool,
1258                                       struct perf_evsel *evsel,
1259                                       struct perf_sample *sample,
1260                                       struct machine *machine)
1261 {
1262         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1263
1264         if (sched->tp_handler->wakeup_event)
1265                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1266
1267         return 0;
1268 }
1269
1270 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1271                             struct perf_sample *sample, struct machine *machine)
1272 {
1273         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1274         struct thread *sched_in;
1275         int new_shortname;
1276         u64 timestamp0, timestamp = sample->time;
1277         s64 delta;
1278         int cpu, this_cpu = sample->cpu;
1279
1280         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1281
1282         if (this_cpu > sched->max_cpu)
1283                 sched->max_cpu = this_cpu;
1284
1285         timestamp0 = sched->cpu_last_switched[this_cpu];
1286         sched->cpu_last_switched[this_cpu] = timestamp;
1287         if (timestamp0)
1288                 delta = timestamp - timestamp0;
1289         else
1290                 delta = 0;
1291
1292         if (delta < 0) {
1293                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1294                 return -1;
1295         }
1296
1297         sched_in = machine__findnew_thread(machine, -1, next_pid);
1298
1299         sched->curr_thread[this_cpu] = sched_in;
1300
1301         printf("  ");
1302
1303         new_shortname = 0;
1304         if (!sched_in->shortname[0]) {
1305                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1306                         /*
1307                          * Don't allocate a letter-number for swapper:0
1308                          * as a shortname. Instead, we use '.' for it.
1309                          */
1310                         sched_in->shortname[0] = '.';
1311                         sched_in->shortname[1] = ' ';
1312                 } else {
1313                         sched_in->shortname[0] = sched->next_shortname1;
1314                         sched_in->shortname[1] = sched->next_shortname2;
1315
1316                         if (sched->next_shortname1 < 'Z') {
1317                                 sched->next_shortname1++;
1318                         } else {
1319                                 sched->next_shortname1 = 'A';
1320                                 if (sched->next_shortname2 < '9')
1321                                         sched->next_shortname2++;
1322                                 else
1323                                         sched->next_shortname2 = '0';
1324                         }
1325                 }
1326                 new_shortname = 1;
1327         }
1328
1329         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1330                 if (cpu != this_cpu)
1331                         printf(" ");
1332                 else
1333                         printf("*");
1334
1335                 if (sched->curr_thread[cpu])
1336                         printf("%2s ", sched->curr_thread[cpu]->shortname);
1337                 else
1338                         printf("   ");
1339         }
1340
1341         printf("  %12.6f secs ", (double)timestamp/1e9);
1342         if (new_shortname) {
1343                 printf("%s => %s:%d\n",
1344                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1345         } else {
1346                 printf("\n");
1347         }
1348
1349         return 0;
1350 }
1351
1352 static int process_sched_switch_event(struct perf_tool *tool,
1353                                       struct perf_evsel *evsel,
1354                                       struct perf_sample *sample,
1355                                       struct machine *machine)
1356 {
1357         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1358         int this_cpu = sample->cpu, err = 0;
1359         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1360             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1361
1362         if (sched->curr_pid[this_cpu] != (u32)-1) {
1363                 /*
1364                  * Are we trying to switch away a PID that is
1365                  * not current?
1366                  */
1367                 if (sched->curr_pid[this_cpu] != prev_pid)
1368                         sched->nr_context_switch_bugs++;
1369         }
1370
1371         if (sched->tp_handler->switch_event)
1372                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1373
1374         sched->curr_pid[this_cpu] = next_pid;
1375         return err;
1376 }
1377
1378 static int process_sched_runtime_event(struct perf_tool *tool,
1379                                        struct perf_evsel *evsel,
1380                                        struct perf_sample *sample,
1381                                        struct machine *machine)
1382 {
1383         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1384
1385         if (sched->tp_handler->runtime_event)
1386                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1387
1388         return 0;
1389 }
1390
1391 static int perf_sched__process_fork_event(struct perf_tool *tool,
1392                                           union perf_event *event,
1393                                           struct perf_sample *sample,
1394                                           struct machine *machine)
1395 {
1396         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1397
1398         /* run the fork event through the perf machineruy */
1399         perf_event__process_fork(tool, event, sample, machine);
1400
1401         /* and then run additional processing needed for this command */
1402         if (sched->tp_handler->fork_event)
1403                 return sched->tp_handler->fork_event(sched, event, machine);
1404
1405         return 0;
1406 }
1407
1408 static int process_sched_migrate_task_event(struct perf_tool *tool,
1409                                             struct perf_evsel *evsel,
1410                                             struct perf_sample *sample,
1411                                             struct machine *machine)
1412 {
1413         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1414
1415         if (sched->tp_handler->migrate_task_event)
1416                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1417
1418         return 0;
1419 }
1420
1421 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1422                                   struct perf_evsel *evsel,
1423                                   struct perf_sample *sample,
1424                                   struct machine *machine);
1425
1426 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1427                                                  union perf_event *event __maybe_unused,
1428                                                  struct perf_sample *sample,
1429                                                  struct perf_evsel *evsel,
1430                                                  struct machine *machine)
1431 {
1432         int err = 0;
1433
1434         if (evsel->handler != NULL) {
1435                 tracepoint_handler f = evsel->handler;
1436                 err = f(tool, evsel, sample, machine);
1437         }
1438
1439         return err;
1440 }
1441
1442 static int perf_sched__read_events(struct perf_sched *sched,
1443                                    struct perf_session **psession)
1444 {
1445         const struct perf_evsel_str_handler handlers[] = {
1446                 { "sched:sched_switch",       process_sched_switch_event, },
1447                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1448                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1449                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1450                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1451         };
1452         struct perf_session *session;
1453         struct perf_data_file file = {
1454                 .path = input_name,
1455                 .mode = PERF_DATA_MODE_READ,
1456         };
1457
1458         session = perf_session__new(&file, false, &sched->tool);
1459         if (session == NULL) {
1460                 pr_debug("No Memory for session\n");
1461                 return -1;
1462         }
1463
1464         symbol__init(&session->header.env);
1465
1466         if (perf_session__set_tracepoints_handlers(session, handlers))
1467                 goto out_delete;
1468
1469         if (perf_session__has_traces(session, "record -R")) {
1470                 int err = perf_session__process_events(session, &sched->tool);
1471                 if (err) {
1472                         pr_err("Failed to process events, error %d", err);
1473                         goto out_delete;
1474                 }
1475
1476                 sched->nr_events      = session->stats.nr_events[0];
1477                 sched->nr_lost_events = session->stats.total_lost;
1478                 sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST];
1479         }
1480
1481         if (psession)
1482                 *psession = session;
1483         else
1484                 perf_session__delete(session);
1485
1486         return 0;
1487
1488 out_delete:
1489         perf_session__delete(session);
1490         return -1;
1491 }
1492
1493 static void print_bad_events(struct perf_sched *sched)
1494 {
1495         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1496                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1497                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1498                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1499         }
1500         if (sched->nr_lost_events && sched->nr_events) {
1501                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1502                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1503                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1504         }
1505         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1506                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1507                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1508                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1509                 if (sched->nr_lost_events)
1510                         printf(" (due to lost events?)");
1511                 printf("\n");
1512         }
1513 }
1514
1515 static int perf_sched__lat(struct perf_sched *sched)
1516 {
1517         struct rb_node *next;
1518         struct perf_session *session;
1519
1520         setup_pager();
1521
1522         /* save session -- references to threads are held in work_list */
1523         if (perf_sched__read_events(sched, &session))
1524                 return -1;
1525
1526         perf_sched__sort_lat(sched);
1527
1528         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1529         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1530         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1531
1532         next = rb_first(&sched->sorted_atom_root);
1533
1534         while (next) {
1535                 struct work_atoms *work_list;
1536
1537                 work_list = rb_entry(next, struct work_atoms, node);
1538                 output_lat_thread(sched, work_list);
1539                 next = rb_next(next);
1540         }
1541
1542         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1543         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1544                 (double)sched->all_runtime / 1e6, sched->all_count);
1545
1546         printf(" ---------------------------------------------------\n");
1547
1548         print_bad_events(sched);
1549         printf("\n");
1550
1551         perf_session__delete(session);
1552         return 0;
1553 }
1554
1555 static int perf_sched__map(struct perf_sched *sched)
1556 {
1557         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1558
1559         setup_pager();
1560         if (perf_sched__read_events(sched, NULL))
1561                 return -1;
1562         print_bad_events(sched);
1563         return 0;
1564 }
1565
1566 static int perf_sched__replay(struct perf_sched *sched)
1567 {
1568         unsigned long i;
1569
1570         calibrate_run_measurement_overhead(sched);
1571         calibrate_sleep_measurement_overhead(sched);
1572
1573         test_calibrations(sched);
1574
1575         if (perf_sched__read_events(sched, NULL))
1576                 return -1;
1577
1578         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1579         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1580         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1581
1582         if (sched->targetless_wakeups)
1583                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1584         if (sched->multitarget_wakeups)
1585                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1586         if (sched->nr_run_events_optimized)
1587                 printf("run atoms optimized: %ld\n",
1588                         sched->nr_run_events_optimized);
1589
1590         print_task_traces(sched);
1591         add_cross_task_wakeups(sched);
1592
1593         create_tasks(sched);
1594         printf("------------------------------------------------------------\n");
1595         for (i = 0; i < sched->replay_repeat; i++)
1596                 run_one_test(sched);
1597
1598         return 0;
1599 }
1600
1601 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1602                           const char * const usage_msg[])
1603 {
1604         char *tmp, *tok, *str = strdup(sched->sort_order);
1605
1606         for (tok = strtok_r(str, ", ", &tmp);
1607                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1608                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1609                         error("Unknown --sort key: `%s'", tok);
1610                         usage_with_options(usage_msg, options);
1611                 }
1612         }
1613
1614         free(str);
1615
1616         sort_dimension__add("pid", &sched->cmp_pid);
1617 }
1618
1619 static int __cmd_record(int argc, const char **argv)
1620 {
1621         unsigned int rec_argc, i, j;
1622         const char **rec_argv;
1623         const char * const record_args[] = {
1624                 "record",
1625                 "-a",
1626                 "-R",
1627                 "-m", "1024",
1628                 "-c", "1",
1629                 "-e", "sched:sched_switch",
1630                 "-e", "sched:sched_stat_wait",
1631                 "-e", "sched:sched_stat_sleep",
1632                 "-e", "sched:sched_stat_iowait",
1633                 "-e", "sched:sched_stat_runtime",
1634                 "-e", "sched:sched_process_fork",
1635                 "-e", "sched:sched_wakeup",
1636                 "-e", "sched:sched_wakeup_new",
1637                 "-e", "sched:sched_migrate_task",
1638         };
1639
1640         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1641         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1642
1643         if (rec_argv == NULL)
1644                 return -ENOMEM;
1645
1646         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1647                 rec_argv[i] = strdup(record_args[i]);
1648
1649         for (j = 1; j < (unsigned int)argc; j++, i++)
1650                 rec_argv[i] = argv[j];
1651
1652         BUG_ON(i != rec_argc);
1653
1654         return cmd_record(i, rec_argv, NULL);
1655 }
1656
1657 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1658 {
1659         const char default_sort_order[] = "avg, max, switch, runtime";
1660         struct perf_sched sched = {
1661                 .tool = {
1662                         .sample          = perf_sched__process_tracepoint_sample,
1663                         .comm            = perf_event__process_comm,
1664                         .lost            = perf_event__process_lost,
1665                         .fork            = perf_sched__process_fork_event,
1666                         .ordered_events = true,
1667                 },
1668                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1669                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1670                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1671                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1672                 .sort_order           = default_sort_order,
1673                 .replay_repeat        = 10,
1674                 .profile_cpu          = -1,
1675                 .next_shortname1      = 'A',
1676                 .next_shortname2      = '0',
1677         };
1678         const struct option latency_options[] = {
1679         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1680                    "sort by key(s): runtime, switch, avg, max"),
1681         OPT_INCR('v', "verbose", &verbose,
1682                     "be more verbose (show symbol address, etc)"),
1683         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1684                     "CPU to profile on"),
1685         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1686                     "dump raw trace in ASCII"),
1687         OPT_END()
1688         };
1689         const struct option replay_options[] = {
1690         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1691                      "repeat the workload replay N times (-1: infinite)"),
1692         OPT_INCR('v', "verbose", &verbose,
1693                     "be more verbose (show symbol address, etc)"),
1694         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1695                     "dump raw trace in ASCII"),
1696         OPT_END()
1697         };
1698         const struct option sched_options[] = {
1699         OPT_STRING('i', "input", &input_name, "file",
1700                     "input file name"),
1701         OPT_INCR('v', "verbose", &verbose,
1702                     "be more verbose (show symbol address, etc)"),
1703         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1704                     "dump raw trace in ASCII"),
1705         OPT_END()
1706         };
1707         const char * const latency_usage[] = {
1708                 "perf sched latency [<options>]",
1709                 NULL
1710         };
1711         const char * const replay_usage[] = {
1712                 "perf sched replay [<options>]",
1713                 NULL
1714         };
1715         const char *const sched_subcommands[] = { "record", "latency", "map",
1716                                                   "replay", "script", NULL };
1717         const char *sched_usage[] = {
1718                 NULL,
1719                 NULL
1720         };
1721         struct trace_sched_handler lat_ops  = {
1722                 .wakeup_event       = latency_wakeup_event,
1723                 .switch_event       = latency_switch_event,
1724                 .runtime_event      = latency_runtime_event,
1725                 .migrate_task_event = latency_migrate_task_event,
1726         };
1727         struct trace_sched_handler map_ops  = {
1728                 .switch_event       = map_switch_event,
1729         };
1730         struct trace_sched_handler replay_ops  = {
1731                 .wakeup_event       = replay_wakeup_event,
1732                 .switch_event       = replay_switch_event,
1733                 .fork_event         = replay_fork_event,
1734         };
1735         unsigned int i;
1736
1737         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1738                 sched.curr_pid[i] = -1;
1739
1740         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1741                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1742         if (!argc)
1743                 usage_with_options(sched_usage, sched_options);
1744
1745         /*
1746          * Aliased to 'perf script' for now:
1747          */
1748         if (!strcmp(argv[0], "script"))
1749                 return cmd_script(argc, argv, prefix);
1750
1751         if (!strncmp(argv[0], "rec", 3)) {
1752                 return __cmd_record(argc, argv);
1753         } else if (!strncmp(argv[0], "lat", 3)) {
1754                 sched.tp_handler = &lat_ops;
1755                 if (argc > 1) {
1756                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1757                         if (argc)
1758                                 usage_with_options(latency_usage, latency_options);
1759                 }
1760                 setup_sorting(&sched, latency_options, latency_usage);
1761                 return perf_sched__lat(&sched);
1762         } else if (!strcmp(argv[0], "map")) {
1763                 sched.tp_handler = &map_ops;
1764                 setup_sorting(&sched, latency_options, latency_usage);
1765                 return perf_sched__map(&sched);
1766         } else if (!strncmp(argv[0], "rep", 3)) {
1767                 sched.tp_handler = &replay_ops;
1768                 if (argc) {
1769                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1770                         if (argc)
1771                                 usage_with_options(replay_usage, replay_options);
1772                 }
1773                 return perf_sched__replay(&sched);
1774         } else {
1775                 usage_with_options(sched_usage, sched_options);
1776         }
1777
1778         return 0;
1779 }