kernel/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 #include <api/fs/fs.h>
  27
  28 #define PR_SET_NAME             15               /* Set process name */
  29 #define MAX_CPUS                4096
  30 #define COMM_LEN                20
  31 #define SYM_LEN                 129
  32 #define MAX_PID                 1024000
  33
  34 struct sched_atom;
  35
  36 struct task_desc {
  37         unsigned long           nr;
  38         unsigned long           pid;
  39         char                    comm[COMM_LEN];
  40
  41         unsigned long           nr_events;
  42         unsigned long           curr_event;
  43         struct sched_atom       **atoms;
  44
  45         pthread_t               thread;
  46         sem_t                   sleep_sem;
  47
  48         sem_t                   ready_for_work;
  49         sem_t                   work_done_sem;
  50
  51         u64                     cpu_usage;
  52 };
  53
  54 enum sched_event_type {
  55         SCHED_EVENT_RUN,
  56         SCHED_EVENT_SLEEP,
  57         SCHED_EVENT_WAKEUP,
  58         SCHED_EVENT_MIGRATION,
  59 };
  60
  61 struct sched_atom {
  62         enum sched_event_type   type;
  63         int                     specific_wait;
  64         u64                     timestamp;
  65         u64                     duration;
  66         unsigned long           nr;
  67         sem_t                   *wait_sem;
  68         struct task_desc        *wakee;
  69 };
  70
  71 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  72
  73 enum thread_state {
  74         THREAD_SLEEPING = 0,
  75         THREAD_WAIT_CPU,
  76         THREAD_SCHED_IN,
  77         THREAD_IGNORE
  78 };
  79
  80 struct work_atom {
  81         struct list_head        list;
  82         enum thread_state       state;
  83         u64                     sched_out_time;
  84         u64                     wake_up_time;
  85         u64                     sched_in_time;
  86         u64                     runtime;
  87 };
  88
  89 struct work_atoms {
  90         struct list_head        work_list;
  91         struct thread           *thread;
  92         struct rb_node          node;
  93         u64                     max_lat;
  94         u64                     max_lat_at;
  95         u64                     total_lat;
  96         u64                     nb_atoms;
  97         u64                     total_runtime;
  98 };
  99
 100 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 101
 102 struct perf_sched;
 103
 104 struct trace_sched_handler {
 105         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 106                             struct perf_sample *sample, struct machine *machine);
 107
 108         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 109                              struct perf_sample *sample, struct machine *machine);
 110
 111         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 112                             struct perf_sample *sample, struct machine *machine);
 113
 114         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 115         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 116                           struct machine *machine);
 117
 118         int (*migrate_task_event)(struct perf_sched *sched,
 119                                   struct perf_evsel *evsel,
 120                                   struct perf_sample *sample,
 121                                   struct machine *machine);
 122 };
 123
 124 struct perf_sched {
 125         struct perf_tool tool;
 126         const char       *sort_order;
 127         unsigned long    nr_tasks;
 128         struct task_desc **pid_to_task;
 129         struct task_desc **tasks;
 130         const struct trace_sched_handler *tp_handler;
 131         pthread_mutex_t  start_work_mutex;
 132         pthread_mutex_t  work_done_wait_mutex;
 133         int              profile_cpu;
 134 /*
 135  * Track the current task - that way we can know whether there's any
 136  * weird events, such as a task being switched away that is not current.
 137  */
 138         int              max_cpu;
 139         u32              curr_pid[MAX_CPUS];
 140         struct thread    *curr_thread[MAX_CPUS];
 141         char             next_shortname1;
 142         char             next_shortname2;
 143         unsigned int     replay_repeat;
 144         unsigned long    nr_run_events;
 145         unsigned long    nr_sleep_events;
 146         unsigned long    nr_wakeup_events;
 147         unsigned long    nr_sleep_corrections;
 148         unsigned long    nr_run_events_optimized;
 149         unsigned long    targetless_wakeups;
 150         unsigned long    multitarget_wakeups;
 151         unsigned long    nr_runs;
 152         unsigned long    nr_timestamps;
 153         unsigned long    nr_unordered_timestamps;
 154         unsigned long    nr_context_switch_bugs;
 155         unsigned long    nr_events;
 156         unsigned long    nr_lost_chunks;
 157         unsigned long    nr_lost_events;
 158         u64              run_measurement_overhead;
 159         u64              sleep_measurement_overhead;
 160         u64              start_time;
 161         u64              cpu_usage;
 162         u64              runavg_cpu_usage;
 163         u64              parent_cpu_usage;
 164         u64              runavg_parent_cpu_usage;
 165         u64              sum_runtime;
 166         u64              sum_fluct;
 167         u64              run_avg;
 168         u64              all_runtime;
 169         u64              all_count;
 170         u64              cpu_last_switched[MAX_CPUS];
 171         struct rb_root   atom_root, sorted_atom_root;
 172         struct list_head sort_list, cmp_pid;
 173         bool force;
 174 };
 175
 176 static u64 get_nsecs(void)
 177 {
 178         struct timespec ts;
 179
 180         clock_gettime(CLOCK_MONOTONIC, &ts);
 181
 182         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 183 }
 184
 185 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 186 {
 187         u64 T0 = get_nsecs(), T1;
 188
 189         do {
 190                 T1 = get_nsecs();
 191         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 192 }
 193
 194 static void sleep_nsecs(u64 nsecs)
 195 {
 196         struct timespec ts;
 197
 198         ts.tv_nsec = nsecs % 999999999;
 199         ts.tv_sec = nsecs / 999999999;
 200
 201         nanosleep(&ts, NULL);
 202 }
 203
 204 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 205 {
 206         u64 T0, T1, delta, min_delta = 1000000000ULL;
 207         int i;
 208
 209         for (i = 0; i < 10; i++) {
 210                 T0 = get_nsecs();
 211                 burn_nsecs(sched, 0);
 212                 T1 = get_nsecs();
 213                 delta = T1-T0;
 214                 min_delta = min(min_delta, delta);
 215         }
 216         sched->run_measurement_overhead = min_delta;
 217
 218         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 219 }
 220
 221 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 222 {
 223         u64 T0, T1, delta, min_delta = 1000000000ULL;
 224         int i;
 225
 226         for (i = 0; i < 10; i++) {
 227                 T0 = get_nsecs();
 228                 sleep_nsecs(10000);
 229                 T1 = get_nsecs();
 230                 delta = T1-T0;
 231                 min_delta = min(min_delta, delta);
 232         }
 233         min_delta -= 10000;
 234         sched->sleep_measurement_overhead = min_delta;
 235
 236         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 237 }
 238
 239 static struct sched_atom *
 240 get_new_event(struct task_desc *task, u64 timestamp)
 241 {
 242         struct sched_atom *event = zalloc(sizeof(*event));
 243         unsigned long idx = task->nr_events;
 244         size_t size;
 245
 246         event->timestamp = timestamp;
 247         event->nr = idx;
 248
 249         task->nr_events++;
 250         size = sizeof(struct sched_atom *) * task->nr_events;
 251         task->atoms = realloc(task->atoms, size);
 252         BUG_ON(!task->atoms);
 253
 254         task->atoms[idx] = event;
 255
 256         return event;
 257 }
 258
 259 static struct sched_atom *last_event(struct task_desc *task)
 260 {
 261         if (!task->nr_events)
 262                 return NULL;
 263
 264         return task->atoms[task->nr_events - 1];
 265 }
 266
 267 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 268                                 u64 timestamp, u64 duration)
 269 {
 270         struct sched_atom *event, *curr_event = last_event(task);
 271
 272         /*
 273          * optimize an existing RUN event by merging this one
 274          * to it:
 275          */
 276         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 277                 sched->nr_run_events_optimized++;
 278                 curr_event->duration += duration;
 279                 return;
 280         }
 281
 282         event = get_new_event(task, timestamp);
 283
 284         event->type = SCHED_EVENT_RUN;
 285         event->duration = duration;
 286
 287         sched->nr_run_events++;
 288 }
 289
 290 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 291                                    u64 timestamp, struct task_desc *wakee)
 292 {
 293         struct sched_atom *event, *wakee_event;
 294
 295         event = get_new_event(task, timestamp);
 296         event->type = SCHED_EVENT_WAKEUP;
 297         event->wakee = wakee;
 298
 299         wakee_event = last_event(wakee);
 300         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 301                 sched->targetless_wakeups++;
 302                 return;
 303         }
 304         if (wakee_event->wait_sem) {
 305                 sched->multitarget_wakeups++;
 306                 return;
 307         }
 308
 309         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 310         sem_init(wakee_event->wait_sem, 0, 0);
 311         wakee_event->specific_wait = 1;
 312         event->wait_sem = wakee_event->wait_sem;
 313
 314         sched->nr_wakeup_events++;
 315 }
 316
 317 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 318                                   u64 timestamp, u64 task_state __maybe_unused)
 319 {
 320         struct sched_atom *event = get_new_event(task, timestamp);
 321
 322         event->type = SCHED_EVENT_SLEEP;
 323
 324         sched->nr_sleep_events++;
 325 }
 326
 327 static struct task_desc *register_pid(struct perf_sched *sched,
 328                                       unsigned long pid, const char *comm)
 329 {
 330         struct task_desc *task;
 331         static int pid_max;
 332
 333         if (sched->pid_to_task == NULL) {
 334                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 335                         pid_max = MAX_PID;
 336                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 337         }
 338         if (pid >= (unsigned long)pid_max) {
 339                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 340                         sizeof(struct task_desc *))) == NULL);
 341                 while (pid >= (unsigned long)pid_max)
 342                         sched->pid_to_task[pid_max++] = NULL;
 343         }
 344
 345         task = sched->pid_to_task[pid];
 346
 347         if (task)
 348                 return task;
 349
 350         task = zalloc(sizeof(*task));
 351         task->pid = pid;
 352         task->nr = sched->nr_tasks;
 353         strcpy(task->comm, comm);
 354         /*
 355          * every task starts in sleeping state - this gets ignored
 356          * if there's no wakeup pointing to this sleep state:
 357          */
 358         add_sched_event_sleep(sched, task, 0, 0);
 359
 360         sched->pid_to_task[pid] = task;
 361         sched->nr_tasks++;
 362         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 363         BUG_ON(!sched->tasks);
 364         sched->tasks[task->nr] = task;
 365
 366         if (verbose)
 367                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 368
 369         return task;
 370 }
 371
 372
 373 static void print_task_traces(struct perf_sched *sched)
 374 {
 375         struct task_desc *task;
 376         unsigned long i;
 377
 378         for (i = 0; i < sched->nr_tasks; i++) {
 379                 task = sched->tasks[i];
 380                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 381                         task->nr, task->comm, task->pid, task->nr_events);
 382         }
 383 }
 384
 385 static void add_cross_task_wakeups(struct perf_sched *sched)
 386 {
 387         struct task_desc *task1, *task2;
 388         unsigned long i, j;
 389
 390         for (i = 0; i < sched->nr_tasks; i++) {
 391                 task1 = sched->tasks[i];
 392                 j = i + 1;
 393                 if (j == sched->nr_tasks)
 394                         j = 0;
 395                 task2 = sched->tasks[j];
 396                 add_sched_event_wakeup(sched, task1, 0, task2);
 397         }
 398 }
 399
 400 static void perf_sched__process_event(struct perf_sched *sched,
 401                                       struct sched_atom *atom)
 402 {
 403         int ret = 0;
 404
 405         switch (atom->type) {
 406                 case SCHED_EVENT_RUN:
 407                         burn_nsecs(sched, atom->duration);
 408                         break;
 409                 case SCHED_EVENT_SLEEP:
 410                         if (atom->wait_sem)
 411                                 ret = sem_wait(atom->wait_sem);
 412                         BUG_ON(ret);
 413                         break;
 414                 case SCHED_EVENT_WAKEUP:
 415                         if (atom->wait_sem)
 416                                 ret = sem_post(atom->wait_sem);
 417                         BUG_ON(ret);
 418                         break;
 419                 case SCHED_EVENT_MIGRATION:
 420                         break;
 421                 default:
 422                         BUG_ON(1);
 423         }
 424 }
 425
 426 static u64 get_cpu_usage_nsec_parent(void)
 427 {
 428         struct rusage ru;
 429         u64 sum;
 430         int err;
 431
 432         err = getrusage(RUSAGE_SELF, &ru);
 433         BUG_ON(err);
 434
 435         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 436         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 437
 438         return sum;
 439 }
 440
 441 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 442 {
 443         struct perf_event_attr attr;
 444         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 445         int fd;
 446         struct rlimit limit;
 447         bool need_privilege = false;
 448
 449         memset(&attr, 0, sizeof(attr));
 450
 451         attr.type = PERF_TYPE_SOFTWARE;
 452         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 453
 454 force_again:
 455         fd = sys_perf_event_open(&attr, 0, -1, -1,
 456                                  perf_event_open_cloexec_flag());
 457
 458         if (fd < 0) {
 459                 if (errno == EMFILE) {
 460                         if (sched->force) {
 461                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 462                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 463                                 if (limit.rlim_cur > limit.rlim_max) {
 464                                         limit.rlim_max = limit.rlim_cur;
 465                                         need_privilege = true;
 466                                 }
 467                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 468                                         if (need_privilege && errno == EPERM)
 469                                                 strcpy(info, "Need privilege\n");
 470                                 } else
 471                                         goto force_again;
 472                         } else
 473                                 strcpy(info, "Have a try with -f option\n");
 474                 }
 475                 pr_err("Error: sys_perf_event_open() syscall returned "
 476                        "with %d (%s)\n%s", fd,
 477                        strerror_r(errno, sbuf, sizeof(sbuf)), info);
 478                 exit(EXIT_FAILURE);
 479         }
 480         return fd;
 481 }
 482
 483 static u64 get_cpu_usage_nsec_self(int fd)
 484 {
 485         u64 runtime;
 486         int ret;
 487
 488         ret = read(fd, &runtime, sizeof(runtime));
 489         BUG_ON(ret != sizeof(runtime));
 490
 491         return runtime;
 492 }
 493
 494 struct sched_thread_parms {
 495         struct task_desc  *task;
 496         struct perf_sched *sched;
 497         int fd;
 498 };
 499
 500 static void *thread_func(void *ctx)
 501 {
 502         struct sched_thread_parms *parms = ctx;
 503         struct task_desc *this_task = parms->task;
 504         struct perf_sched *sched = parms->sched;
 505         u64 cpu_usage_0, cpu_usage_1;
 506         unsigned long i, ret;
 507         char comm2[22];
 508         int fd = parms->fd;
 509
 510         zfree(&parms);
 511
 512         sprintf(comm2, ":%s", this_task->comm);
 513         prctl(PR_SET_NAME, comm2);
 514         if (fd < 0)
 515                 return NULL;
 516 again:
 517         ret = sem_post(&this_task->ready_for_work);
 518         BUG_ON(ret);
 519         ret = pthread_mutex_lock(&sched->start_work_mutex);
 520         BUG_ON(ret);
 521         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 522         BUG_ON(ret);
 523
 524         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 525
 526         for (i = 0; i < this_task->nr_events; i++) {
 527                 this_task->curr_event = i;
 528                 perf_sched__process_event(sched, this_task->atoms[i]);
 529         }
 530
 531         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 532         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 533         ret = sem_post(&this_task->work_done_sem);
 534         BUG_ON(ret);
 535
 536         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 537         BUG_ON(ret);
 538         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 539         BUG_ON(ret);
 540
 541         goto again;
 542 }
 543
 544 static void create_tasks(struct perf_sched *sched)
 545 {
 546         struct task_desc *task;
 547         pthread_attr_t attr;
 548         unsigned long i;
 549         int err;
 550
 551         err = pthread_attr_init(&attr);
 552         BUG_ON(err);
 553         err = pthread_attr_setstacksize(&attr,
 554                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 555         BUG_ON(err);
 556         err = pthread_mutex_lock(&sched->start_work_mutex);
 557         BUG_ON(err);
 558         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 559         BUG_ON(err);
 560         for (i = 0; i < sched->nr_tasks; i++) {
 561                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 562                 BUG_ON(parms == NULL);
 563                 parms->task = task = sched->tasks[i];
 564                 parms->sched = sched;
 565                 parms->fd = self_open_counters(sched, i);
 566                 sem_init(&task->sleep_sem, 0, 0);
 567                 sem_init(&task->ready_for_work, 0, 0);
 568                 sem_init(&task->work_done_sem, 0, 0);
 569                 task->curr_event = 0;
 570                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 571                 BUG_ON(err);
 572         }
 573 }
 574
 575 static void wait_for_tasks(struct perf_sched *sched)
 576 {
 577         u64 cpu_usage_0, cpu_usage_1;
 578         struct task_desc *task;
 579         unsigned long i, ret;
 580
 581         sched->start_time = get_nsecs();
 582         sched->cpu_usage = 0;
 583         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 584
 585         for (i = 0; i < sched->nr_tasks; i++) {
 586                 task = sched->tasks[i];
 587                 ret = sem_wait(&task->ready_for_work);
 588                 BUG_ON(ret);
 589                 sem_init(&task->ready_for_work, 0, 0);
 590         }
 591         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 592         BUG_ON(ret);
 593
 594         cpu_usage_0 = get_cpu_usage_nsec_parent();
 595
 596         pthread_mutex_unlock(&sched->start_work_mutex);
 597
 598         for (i = 0; i < sched->nr_tasks; i++) {
 599                 task = sched->tasks[i];
 600                 ret = sem_wait(&task->work_done_sem);
 601                 BUG_ON(ret);
 602                 sem_init(&task->work_done_sem, 0, 0);
 603                 sched->cpu_usage += task->cpu_usage;
 604                 task->cpu_usage = 0;
 605         }
 606
 607         cpu_usage_1 = get_cpu_usage_nsec_parent();
 608         if (!sched->runavg_cpu_usage)
 609                 sched->runavg_cpu_usage = sched->cpu_usage;
 610         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 611
 612         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 613         if (!sched->runavg_parent_cpu_usage)
 614                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 615         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 616                                          sched->parent_cpu_usage)/sched->replay_repeat;
 617
 618         ret = pthread_mutex_lock(&sched->start_work_mutex);
 619         BUG_ON(ret);
 620
 621         for (i = 0; i < sched->nr_tasks; i++) {
 622                 task = sched->tasks[i];
 623                 sem_init(&task->sleep_sem, 0, 0);
 624                 task->curr_event = 0;
 625         }
 626 }
 627
 628 static void run_one_test(struct perf_sched *sched)
 629 {
 630         u64 T0, T1, delta, avg_delta, fluct;
 631
 632         T0 = get_nsecs();
 633         wait_for_tasks(sched);
 634         T1 = get_nsecs();
 635
 636         delta = T1 - T0;
 637         sched->sum_runtime += delta;
 638         sched->nr_runs++;
 639
 640         avg_delta = sched->sum_runtime / sched->nr_runs;
 641         if (delta < avg_delta)
 642                 fluct = avg_delta - delta;
 643         else
 644                 fluct = delta - avg_delta;
 645         sched->sum_fluct += fluct;
 646         if (!sched->run_avg)
 647                 sched->run_avg = delta;
 648         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 649
 650         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 651
 652         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 653
 654         printf("cpu: %0.2f / %0.2f",
 655                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 656
 657 #if 0
 658         /*
 659          * rusage statistics done by the parent, these are less
 660          * accurate than the sched->sum_exec_runtime based statistics:
 661          */
 662         printf(" [%0.2f / %0.2f]",
 663                 (double)sched->parent_cpu_usage/1e6,
 664                 (double)sched->runavg_parent_cpu_usage/1e6);
 665 #endif
 666
 667         printf("\n");
 668
 669         if (sched->nr_sleep_corrections)
 670                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 671         sched->nr_sleep_corrections = 0;
 672 }
 673
 674 static void test_calibrations(struct perf_sched *sched)
 675 {
 676         u64 T0, T1;
 677
 678         T0 = get_nsecs();
 679         burn_nsecs(sched, 1e6);
 680         T1 = get_nsecs();
 681
 682         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 683
 684         T0 = get_nsecs();
 685         sleep_nsecs(1e6);
 686         T1 = get_nsecs();
 687
 688         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 689 }
 690
 691 static int
 692 replay_wakeup_event(struct perf_sched *sched,
 693                     struct perf_evsel *evsel, struct perf_sample *sample,
 694                     struct machine *machine __maybe_unused)
 695 {
 696         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 697         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 698         struct task_desc *waker, *wakee;
 699
 700         if (verbose) {
 701                 printf("sched_wakeup event %p\n", evsel);
 702
 703                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 704         }
 705
 706         waker = register_pid(sched, sample->tid, "<unknown>");
 707         wakee = register_pid(sched, pid, comm);
 708
 709         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 710         return 0;
 711 }
 712
 713 static int replay_switch_event(struct perf_sched *sched,
 714                                struct perf_evsel *evsel,
 715                                struct perf_sample *sample,
 716                                struct machine *machine __maybe_unused)
 717 {
 718         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 719                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 720         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 721                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 722         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 723         struct task_desc *prev, __maybe_unused *next;
 724         u64 timestamp0, timestamp = sample->time;
 725         int cpu = sample->cpu;
 726         s64 delta;
 727
 728         if (verbose)
 729                 printf("sched_switch event %p\n", evsel);
 730
 731         if (cpu >= MAX_CPUS || cpu < 0)
 732                 return 0;
 733
 734         timestamp0 = sched->cpu_last_switched[cpu];
 735         if (timestamp0)
 736                 delta = timestamp - timestamp0;
 737         else
 738                 delta = 0;
 739
 740         if (delta < 0) {
 741                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 742                 return -1;
 743         }
 744
 745         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 746                  prev_comm, prev_pid, next_comm, next_pid, delta);
 747
 748         prev = register_pid(sched, prev_pid, prev_comm);
 749         next = register_pid(sched, next_pid, next_comm);
 750
 751         sched->cpu_last_switched[cpu] = timestamp;
 752
 753         add_sched_event_run(sched, prev, timestamp, delta);
 754         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 755
 756         return 0;
 757 }
 758
 759 static int replay_fork_event(struct perf_sched *sched,
 760                              union perf_event *event,
 761                              struct machine *machine)
 762 {
 763         struct thread *child, *parent;
 764
 765         child = machine__findnew_thread(machine, event->fork.pid,
 766                                         event->fork.tid);
 767         parent = machine__findnew_thread(machine, event->fork.ppid,
 768                                          event->fork.ptid);
 769
 770         if (child == NULL || parent == NULL) {
 771                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 772                                  child, parent);
 773                 return 0;
 774         }
 775
 776         if (verbose) {
 777                 printf("fork event\n");
 778                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 779                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 780         }
 781
 782         register_pid(sched, parent->tid, thread__comm_str(parent));
 783         register_pid(sched, child->tid, thread__comm_str(child));
 784         return 0;
 785 }
 786
 787 struct sort_dimension {
 788         const char              *name;
 789         sort_fn_t               cmp;
 790         struct list_head        list;
 791 };
 792
 793 static int
 794 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 795 {
 796         struct sort_dimension *sort;
 797         int ret = 0;
 798
 799         BUG_ON(list_empty(list));
 800
 801         list_for_each_entry(sort, list, list) {
 802                 ret = sort->cmp(l, r);
 803                 if (ret)
 804                         return ret;
 805         }
 806
 807         return ret;
 808 }
 809
 810 static struct work_atoms *
 811 thread_atoms_search(struct rb_root *root, struct thread *thread,
 812                          struct list_head *sort_list)
 813 {
 814         struct rb_node *node = root->rb_node;
 815         struct work_atoms key = { .thread = thread };
 816
 817         while (node) {
 818                 struct work_atoms *atoms;
 819                 int cmp;
 820
 821                 atoms = container_of(node, struct work_atoms, node);
 822
 823                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 824                 if (cmp > 0)
 825                         node = node->rb_left;
 826                 else if (cmp < 0)
 827                         node = node->rb_right;
 828                 else {
 829                         BUG_ON(thread != atoms->thread);
 830                         return atoms;
 831                 }
 832         }
 833         return NULL;
 834 }
 835
 836 static void
 837 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 838                          struct list_head *sort_list)
 839 {
 840         struct rb_node **new = &(root->rb_node), *parent = NULL;
 841
 842         while (*new) {
 843                 struct work_atoms *this;
 844                 int cmp;
 845
 846                 this = container_of(*new, struct work_atoms, node);
 847                 parent = *new;
 848
 849                 cmp = thread_lat_cmp(sort_list, data, this);
 850
 851                 if (cmp > 0)
 852                         new = &((*new)->rb_left);
 853                 else
 854                         new = &((*new)->rb_right);
 855         }
 856
 857         rb_link_node(&data->node, parent, new);
 858         rb_insert_color(&data->node, root);
 859 }
 860
 861 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 862 {
 863         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 864         if (!atoms) {
 865                 pr_err("No memory at %s\n", __func__);
 866                 return -1;
 867         }
 868
 869         atoms->thread = thread__get(thread);
 870         INIT_LIST_HEAD(&atoms->work_list);
 871         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 872         return 0;
 873 }
 874
 875 static char sched_out_state(u64 prev_state)
 876 {
 877         const char *str = TASK_STATE_TO_CHAR_STR;
 878
 879         return str[prev_state];
 880 }
 881
 882 static int
 883 add_sched_out_event(struct work_atoms *atoms,
 884                     char run_state,
 885                     u64 timestamp)
 886 {
 887         struct work_atom *atom = zalloc(sizeof(*atom));
 888         if (!atom) {
 889                 pr_err("Non memory at %s", __func__);
 890                 return -1;
 891         }
 892
 893         atom->sched_out_time = timestamp;
 894
 895         if (run_state == 'R') {
 896                 atom->state = THREAD_WAIT_CPU;
 897                 atom->wake_up_time = atom->sched_out_time;
 898         }
 899
 900         list_add_tail(&atom->list, &atoms->work_list);
 901         return 0;
 902 }
 903
 904 static void
 905 add_runtime_event(struct work_atoms *atoms, u64 delta,
 906                   u64 timestamp __maybe_unused)
 907 {
 908         struct work_atom *atom;
 909
 910         BUG_ON(list_empty(&atoms->work_list));
 911
 912         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 913
 914         atom->runtime += delta;
 915         atoms->total_runtime += delta;
 916 }
 917
 918 static void
 919 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 920 {
 921         struct work_atom *atom;
 922         u64 delta;
 923
 924         if (list_empty(&atoms->work_list))
 925                 return;
 926
 927         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 928
 929         if (atom->state != THREAD_WAIT_CPU)
 930                 return;
 931
 932         if (timestamp < atom->wake_up_time) {
 933                 atom->state = THREAD_IGNORE;
 934                 return;
 935         }
 936
 937         atom->state = THREAD_SCHED_IN;
 938         atom->sched_in_time = timestamp;
 939
 940         delta = atom->sched_in_time - atom->wake_up_time;
 941         atoms->total_lat += delta;
 942         if (delta > atoms->max_lat) {
 943                 atoms->max_lat = delta;
 944                 atoms->max_lat_at = timestamp;
 945         }
 946         atoms->nb_atoms++;
 947 }
 948
 949 static int latency_switch_event(struct perf_sched *sched,
 950                                 struct perf_evsel *evsel,
 951                                 struct perf_sample *sample,
 952                                 struct machine *machine)
 953 {
 954         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 955                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 956         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 957         struct work_atoms *out_events, *in_events;
 958         struct thread *sched_out, *sched_in;
 959         u64 timestamp0, timestamp = sample->time;
 960         int cpu = sample->cpu;
 961         s64 delta;
 962
 963         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 964
 965         timestamp0 = sched->cpu_last_switched[cpu];
 966         sched->cpu_last_switched[cpu] = timestamp;
 967         if (timestamp0)
 968                 delta = timestamp - timestamp0;
 969         else
 970                 delta = 0;
 971
 972         if (delta < 0) {
 973                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 974                 return -1;
 975         }
 976
 977         sched_out = machine__findnew_thread(machine, -1, prev_pid);
 978         sched_in = machine__findnew_thread(machine, -1, next_pid);
 979
 980         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 981         if (!out_events) {
 982                 if (thread_atoms_insert(sched, sched_out))
 983                         return -1;
 984                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 985                 if (!out_events) {
 986                         pr_err("out-event: Internal tree error");
 987                         return -1;
 988                 }
 989         }
 990         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 991                 return -1;
 992
 993         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 994         if (!in_events) {
 995                 if (thread_atoms_insert(sched, sched_in))
 996                         return -1;
 997                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
 998                 if (!in_events) {
 999                         pr_err("in-event: Internal tree error");
1000                         return -1;
1001                 }
1002                 /*
1003                  * Take came in we have not heard about yet,
1004                  * add in an initial atom in runnable state:
1005                  */
1006                 if (add_sched_out_event(in_events, 'R', timestamp))
1007                         return -1;
1008         }
1009         add_sched_in_event(in_events, timestamp);
1010
1011         return 0;
1012 }
1013
1014 static int latency_runtime_event(struct perf_sched *sched,
1015                                  struct perf_evsel *evsel,
1016                                  struct perf_sample *sample,
1017                                  struct machine *machine)
1018 {
1019         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1020         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1021         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1022         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1023         u64 timestamp = sample->time;
1024         int cpu = sample->cpu;
1025
1026         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1027         if (!atoms) {
1028                 if (thread_atoms_insert(sched, thread))
1029                         return -1;
1030                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1031                 if (!atoms) {
1032                         pr_err("in-event: Internal tree error");
1033                         return -1;
1034                 }
1035                 if (add_sched_out_event(atoms, 'R', timestamp))
1036                         return -1;
1037         }
1038
1039         add_runtime_event(atoms, runtime, timestamp);
1040         return 0;
1041 }
1042
1043 static int latency_wakeup_event(struct perf_sched *sched,
1044                                 struct perf_evsel *evsel,
1045                                 struct perf_sample *sample,
1046                                 struct machine *machine)
1047 {
1048         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1049         struct work_atoms *atoms;
1050         struct work_atom *atom;
1051         struct thread *wakee;
1052         u64 timestamp = sample->time;
1053
1054         wakee = machine__findnew_thread(machine, -1, pid);
1055         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1056         if (!atoms) {
1057                 if (thread_atoms_insert(sched, wakee))
1058                         return -1;
1059                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1060                 if (!atoms) {
1061                         pr_err("wakeup-event: Internal tree error");
1062                         return -1;
1063                 }
1064                 if (add_sched_out_event(atoms, 'S', timestamp))
1065                         return -1;
1066         }
1067
1068         BUG_ON(list_empty(&atoms->work_list));
1069
1070         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1071
1072         /*
1073          * As we do not guarantee the wakeup event happens when
1074          * task is out of run queue, also may happen when task is
1075          * on run queue and wakeup only change ->state to TASK_RUNNING,
1076          * then we should not set the ->wake_up_time when wake up a
1077          * task which is on run queue.
1078          *
1079          * You WILL be missing events if you've recorded only
1080          * one CPU, or are only looking at only one, so don't
1081          * skip in this case.
1082          */
1083         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1084                 return 0;
1085
1086         sched->nr_timestamps++;
1087         if (atom->sched_out_time > timestamp) {
1088                 sched->nr_unordered_timestamps++;
1089                 return 0;
1090         }
1091
1092         atom->state = THREAD_WAIT_CPU;
1093         atom->wake_up_time = timestamp;
1094         return 0;
1095 }
1096
1097 static int latency_migrate_task_event(struct perf_sched *sched,
1098                                       struct perf_evsel *evsel,
1099                                       struct perf_sample *sample,
1100                                       struct machine *machine)
1101 {
1102         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1103         u64 timestamp = sample->time;
1104         struct work_atoms *atoms;
1105         struct work_atom *atom;
1106         struct thread *migrant;
1107
1108         /*
1109          * Only need to worry about migration when profiling one CPU.
1110          */
1111         if (sched->profile_cpu == -1)
1112                 return 0;
1113
1114         migrant = machine__findnew_thread(machine, -1, pid);
1115         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1116         if (!atoms) {
1117                 if (thread_atoms_insert(sched, migrant))
1118                         return -1;
1119                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1120                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1121                 if (!atoms) {
1122                         pr_err("migration-event: Internal tree error");
1123                         return -1;
1124                 }
1125                 if (add_sched_out_event(atoms, 'R', timestamp))
1126                         return -1;
1127         }
1128
1129         BUG_ON(list_empty(&atoms->work_list));
1130
1131         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1132         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1133
1134         sched->nr_timestamps++;
1135
1136         if (atom->sched_out_time > timestamp)
1137                 sched->nr_unordered_timestamps++;
1138
1139         return 0;
1140 }
1141
1142 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1143 {
1144         int i;
1145         int ret;
1146         u64 avg;
1147
1148         if (!work_list->nb_atoms)
1149                 return;
1150         /*
1151          * Ignore idle threads:
1152          */
1153         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1154                 return;
1155
1156         sched->all_runtime += work_list->total_runtime;
1157         sched->all_count   += work_list->nb_atoms;
1158
1159         ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1160
1161         for (i = 0; i < 24 - ret; i++)
1162                 printf(" ");
1163
1164         avg = work_list->total_lat / work_list->nb_atoms;
1165
1166         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1167               (double)work_list->total_runtime / 1e6,
1168                  work_list->nb_atoms, (double)avg / 1e6,
1169                  (double)work_list->max_lat / 1e6,
1170                  (double)work_list->max_lat_at / 1e9);
1171 }
1172
1173 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1174 {
1175         if (l->thread->tid < r->thread->tid)
1176                 return -1;
1177         if (l->thread->tid > r->thread->tid)
1178                 return 1;
1179
1180         return 0;
1181 }
1182
1183 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1184 {
1185         u64 avgl, avgr;
1186
1187         if (!l->nb_atoms)
1188                 return -1;
1189
1190         if (!r->nb_atoms)
1191                 return 1;
1192
1193         avgl = l->total_lat / l->nb_atoms;
1194         avgr = r->total_lat / r->nb_atoms;
1195
1196         if (avgl < avgr)
1197                 return -1;
1198         if (avgl > avgr)
1199                 return 1;
1200
1201         return 0;
1202 }
1203
1204 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1205 {
1206         if (l->max_lat < r->max_lat)
1207                 return -1;
1208         if (l->max_lat > r->max_lat)
1209                 return 1;
1210
1211         return 0;
1212 }
1213
1214 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1215 {
1216         if (l->nb_atoms < r->nb_atoms)
1217                 return -1;
1218         if (l->nb_atoms > r->nb_atoms)
1219                 return 1;
1220
1221         return 0;
1222 }
1223
1224 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1225 {
1226         if (l->total_runtime < r->total_runtime)
1227                 return -1;
1228         if (l->total_runtime > r->total_runtime)
1229                 return 1;
1230
1231         return 0;
1232 }
1233
1234 static int sort_dimension__add(const char *tok, struct list_head *list)
1235 {
1236         size_t i;
1237         static struct sort_dimension avg_sort_dimension = {
1238                 .name = "avg",
1239                 .cmp  = avg_cmp,
1240         };
1241         static struct sort_dimension max_sort_dimension = {
1242                 .name = "max",
1243                 .cmp  = max_cmp,
1244         };
1245         static struct sort_dimension pid_sort_dimension = {
1246                 .name = "pid",
1247                 .cmp  = pid_cmp,
1248         };
1249         static struct sort_dimension runtime_sort_dimension = {
1250                 .name = "runtime",
1251                 .cmp  = runtime_cmp,
1252         };
1253         static struct sort_dimension switch_sort_dimension = {
1254                 .name = "switch",
1255                 .cmp  = switch_cmp,
1256         };
1257         struct sort_dimension *available_sorts[] = {
1258                 &pid_sort_dimension,
1259                 &avg_sort_dimension,
1260                 &max_sort_dimension,
1261                 &switch_sort_dimension,
1262                 &runtime_sort_dimension,
1263         };
1264
1265         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1266                 if (!strcmp(available_sorts[i]->name, tok)) {
1267                         list_add_tail(&available_sorts[i]->list, list);
1268
1269                         return 0;
1270                 }
1271         }
1272
1273         return -1;
1274 }
1275
1276 static void perf_sched__sort_lat(struct perf_sched *sched)
1277 {
1278         struct rb_node *node;
1279
1280         for (;;) {
1281                 struct work_atoms *data;
1282                 node = rb_first(&sched->atom_root);
1283                 if (!node)
1284                         break;
1285
1286                 rb_erase(node, &sched->atom_root);
1287                 data = rb_entry(node, struct work_atoms, node);
1288                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1289         }
1290 }
1291
1292 static int process_sched_wakeup_event(struct perf_tool *tool,
1293                                       struct perf_evsel *evsel,
1294                                       struct perf_sample *sample,
1295                                       struct machine *machine)
1296 {
1297         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1298
1299         if (sched->tp_handler->wakeup_event)
1300                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1301
1302         return 0;
1303 }
1304
1305 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1306                             struct perf_sample *sample, struct machine *machine)
1307 {
1308         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1309         struct thread *sched_in;
1310         int new_shortname;
1311         u64 timestamp0, timestamp = sample->time;
1312         s64 delta;
1313         int cpu, this_cpu = sample->cpu;
1314
1315         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1316
1317         if (this_cpu > sched->max_cpu)
1318                 sched->max_cpu = this_cpu;
1319
1320         timestamp0 = sched->cpu_last_switched[this_cpu];
1321         sched->cpu_last_switched[this_cpu] = timestamp;
1322         if (timestamp0)
1323                 delta = timestamp - timestamp0;
1324         else
1325                 delta = 0;
1326
1327         if (delta < 0) {
1328                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1329                 return -1;
1330         }
1331
1332         sched_in = machine__findnew_thread(machine, -1, next_pid);
1333
1334         sched->curr_thread[this_cpu] = sched_in;
1335
1336         printf("  ");
1337
1338         new_shortname = 0;
1339         if (!sched_in->shortname[0]) {
1340                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1341                         /*
1342                          * Don't allocate a letter-number for swapper:0
1343                          * as a shortname. Instead, we use '.' for it.
1344                          */
1345                         sched_in->shortname[0] = '.';
1346                         sched_in->shortname[1] = ' ';
1347                 } else {
1348                         sched_in->shortname[0] = sched->next_shortname1;
1349                         sched_in->shortname[1] = sched->next_shortname2;
1350
1351                         if (sched->next_shortname1 < 'Z') {
1352                                 sched->next_shortname1++;
1353                         } else {
1354                                 sched->next_shortname1 = 'A';
1355                                 if (sched->next_shortname2 < '9')
1356                                         sched->next_shortname2++;
1357                                 else
1358                                         sched->next_shortname2 = '0';
1359                         }
1360                 }
1361                 new_shortname = 1;
1362         }
1363
1364         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1365                 if (cpu != this_cpu)
1366                         printf(" ");
1367                 else
1368                         printf("*");
1369
1370                 if (sched->curr_thread[cpu])
1371                         printf("%2s ", sched->curr_thread[cpu]->shortname);
1372                 else
1373                         printf("   ");
1374         }
1375
1376         printf("  %12.6f secs ", (double)timestamp/1e9);
1377         if (new_shortname) {
1378                 printf("%s => %s:%d\n",
1379                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1380         } else {
1381                 printf("\n");
1382         }
1383
1384         return 0;
1385 }
1386
1387 static int process_sched_switch_event(struct perf_tool *tool,
1388                                       struct perf_evsel *evsel,
1389                                       struct perf_sample *sample,
1390                                       struct machine *machine)
1391 {
1392         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1393         int this_cpu = sample->cpu, err = 0;
1394         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1395             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1396
1397         if (sched->curr_pid[this_cpu] != (u32)-1) {
1398                 /*
1399                  * Are we trying to switch away a PID that is
1400                  * not current?
1401                  */
1402                 if (sched->curr_pid[this_cpu] != prev_pid)
1403                         sched->nr_context_switch_bugs++;
1404         }
1405
1406         if (sched->tp_handler->switch_event)
1407                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1408
1409         sched->curr_pid[this_cpu] = next_pid;
1410         return err;
1411 }
1412
1413 static int process_sched_runtime_event(struct perf_tool *tool,
1414                                        struct perf_evsel *evsel,
1415                                        struct perf_sample *sample,
1416                                        struct machine *machine)
1417 {
1418         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1419
1420         if (sched->tp_handler->runtime_event)
1421                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1422
1423         return 0;
1424 }
1425
1426 static int perf_sched__process_fork_event(struct perf_tool *tool,
1427                                           union perf_event *event,
1428                                           struct perf_sample *sample,
1429                                           struct machine *machine)
1430 {
1431         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1432
1433         /* run the fork event through the perf machineruy */
1434         perf_event__process_fork(tool, event, sample, machine);
1435
1436         /* and then run additional processing needed for this command */
1437         if (sched->tp_handler->fork_event)
1438                 return sched->tp_handler->fork_event(sched, event, machine);
1439
1440         return 0;
1441 }
1442
1443 static int process_sched_migrate_task_event(struct perf_tool *tool,
1444                                             struct perf_evsel *evsel,
1445                                             struct perf_sample *sample,
1446                                             struct machine *machine)
1447 {
1448         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1449
1450         if (sched->tp_handler->migrate_task_event)
1451                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1452
1453         return 0;
1454 }
1455
1456 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1457                                   struct perf_evsel *evsel,
1458                                   struct perf_sample *sample,
1459                                   struct machine *machine);
1460
1461 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1462                                                  union perf_event *event __maybe_unused,
1463                                                  struct perf_sample *sample,
1464                                                  struct perf_evsel *evsel,
1465                                                  struct machine *machine)
1466 {
1467         int err = 0;
1468
1469         if (evsel->handler != NULL) {
1470                 tracepoint_handler f = evsel->handler;
1471                 err = f(tool, evsel, sample, machine);
1472         }
1473
1474         return err;
1475 }
1476
1477 static int perf_sched__read_events(struct perf_sched *sched)
1478 {
1479         const struct perf_evsel_str_handler handlers[] = {
1480                 { "sched:sched_switch",       process_sched_switch_event, },
1481                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1482                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1483                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1484                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1485         };
1486         struct perf_session *session;
1487         struct perf_data_file file = {
1488                 .path = input_name,
1489                 .mode = PERF_DATA_MODE_READ,
1490                 .force = sched->force,
1491         };
1492         int rc = -1;
1493
1494         session = perf_session__new(&file, false, &sched->tool);
1495         if (session == NULL) {
1496                 pr_debug("No Memory for session\n");
1497                 return -1;
1498         }
1499
1500         symbol__init(&session->header.env);
1501
1502         if (perf_session__set_tracepoints_handlers(session, handlers))
1503                 goto out_delete;
1504
1505         if (perf_session__has_traces(session, "record -R")) {
1506                 int err = perf_session__process_events(session);
1507                 if (err) {
1508                         pr_err("Failed to process events, error %d", err);
1509                         goto out_delete;
1510                 }
1511
1512                 sched->nr_events      = session->evlist->stats.nr_events[0];
1513                 sched->nr_lost_events = session->evlist->stats.total_lost;
1514                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1515         }
1516
1517         rc = 0;
1518 out_delete:
1519         perf_session__delete(session);
1520         return rc;
1521 }
1522
1523 static void print_bad_events(struct perf_sched *sched)
1524 {
1525         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1526                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1527                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1528                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1529         }
1530         if (sched->nr_lost_events && sched->nr_events) {
1531                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1532                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1533                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1534         }
1535         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1536                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1537                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1538                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1539                 if (sched->nr_lost_events)
1540                         printf(" (due to lost events?)");
1541                 printf("\n");
1542         }
1543 }
1544
1545 static int perf_sched__lat(struct perf_sched *sched)
1546 {
1547         struct rb_node *next;
1548
1549         setup_pager();
1550
1551         if (perf_sched__read_events(sched))
1552                 return -1;
1553
1554         perf_sched__sort_lat(sched);
1555
1556         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1557         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1558         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1559
1560         next = rb_first(&sched->sorted_atom_root);
1561
1562         while (next) {
1563                 struct work_atoms *work_list;
1564
1565                 work_list = rb_entry(next, struct work_atoms, node);
1566                 output_lat_thread(sched, work_list);
1567                 next = rb_next(next);
1568                 thread__zput(work_list->thread);
1569         }
1570
1571         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1572         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1573                 (double)sched->all_runtime / 1e6, sched->all_count);
1574
1575         printf(" ---------------------------------------------------\n");
1576
1577         print_bad_events(sched);
1578         printf("\n");
1579
1580         return 0;
1581 }
1582
1583 static int perf_sched__map(struct perf_sched *sched)
1584 {
1585         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1586
1587         setup_pager();
1588         if (perf_sched__read_events(sched))
1589                 return -1;
1590         print_bad_events(sched);
1591         return 0;
1592 }
1593
1594 static int perf_sched__replay(struct perf_sched *sched)
1595 {
1596         unsigned long i;
1597
1598         calibrate_run_measurement_overhead(sched);
1599         calibrate_sleep_measurement_overhead(sched);
1600
1601         test_calibrations(sched);
1602
1603         if (perf_sched__read_events(sched))
1604                 return -1;
1605
1606         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1607         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1608         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1609
1610         if (sched->targetless_wakeups)
1611                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1612         if (sched->multitarget_wakeups)
1613                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1614         if (sched->nr_run_events_optimized)
1615                 printf("run atoms optimized: %ld\n",
1616                         sched->nr_run_events_optimized);
1617
1618         print_task_traces(sched);
1619         add_cross_task_wakeups(sched);
1620
1621         create_tasks(sched);
1622         printf("------------------------------------------------------------\n");
1623         for (i = 0; i < sched->replay_repeat; i++)
1624                 run_one_test(sched);
1625
1626         return 0;
1627 }
1628
1629 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1630                           const char * const usage_msg[])
1631 {
1632         char *tmp, *tok, *str = strdup(sched->sort_order);
1633
1634         for (tok = strtok_r(str, ", ", &tmp);
1635                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1636                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1637                         error("Unknown --sort key: `%s'", tok);
1638                         usage_with_options(usage_msg, options);
1639                 }
1640         }
1641
1642         free(str);
1643
1644         sort_dimension__add("pid", &sched->cmp_pid);
1645 }
1646
1647 static int __cmd_record(int argc, const char **argv)
1648 {
1649         unsigned int rec_argc, i, j;
1650         const char **rec_argv;
1651         const char * const record_args[] = {
1652                 "record",
1653                 "-a",
1654                 "-R",
1655                 "-m", "1024",
1656                 "-c", "1",
1657                 "-e", "sched:sched_switch",
1658                 "-e", "sched:sched_stat_wait",
1659                 "-e", "sched:sched_stat_sleep",
1660                 "-e", "sched:sched_stat_iowait",
1661                 "-e", "sched:sched_stat_runtime",
1662                 "-e", "sched:sched_process_fork",
1663                 "-e", "sched:sched_wakeup",
1664                 "-e", "sched:sched_wakeup_new",
1665                 "-e", "sched:sched_migrate_task",
1666         };
1667
1668         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1669         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1670
1671         if (rec_argv == NULL)
1672                 return -ENOMEM;
1673
1674         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1675                 rec_argv[i] = strdup(record_args[i]);
1676
1677         for (j = 1; j < (unsigned int)argc; j++, i++)
1678                 rec_argv[i] = argv[j];
1679
1680         BUG_ON(i != rec_argc);
1681
1682         return cmd_record(i, rec_argv, NULL);
1683 }
1684
1685 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1686 {
1687         const char default_sort_order[] = "avg, max, switch, runtime";
1688         struct perf_sched sched = {
1689                 .tool = {
1690                         .sample          = perf_sched__process_tracepoint_sample,
1691                         .comm            = perf_event__process_comm,
1692                         .lost            = perf_event__process_lost,
1693                         .fork            = perf_sched__process_fork_event,
1694                         .ordered_events = true,
1695                 },
1696                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1697                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1698                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1699                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1700                 .sort_order           = default_sort_order,
1701                 .replay_repeat        = 10,
1702                 .profile_cpu          = -1,
1703                 .next_shortname1      = 'A',
1704                 .next_shortname2      = '0',
1705         };
1706         const struct option latency_options[] = {
1707         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1708                    "sort by key(s): runtime, switch, avg, max"),
1709         OPT_INCR('v', "verbose", &verbose,
1710                     "be more verbose (show symbol address, etc)"),
1711         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1712                     "CPU to profile on"),
1713         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1714                     "dump raw trace in ASCII"),
1715         OPT_END()
1716         };
1717         const struct option replay_options[] = {
1718         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1719                      "repeat the workload replay N times (-1: infinite)"),
1720         OPT_INCR('v', "verbose", &verbose,
1721                     "be more verbose (show symbol address, etc)"),
1722         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1723                     "dump raw trace in ASCII"),
1724         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1725         OPT_END()
1726         };
1727         const struct option sched_options[] = {
1728         OPT_STRING('i', "input", &input_name, "file",
1729                     "input file name"),
1730         OPT_INCR('v', "verbose", &verbose,
1731                     "be more verbose (show symbol address, etc)"),
1732         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1733                     "dump raw trace in ASCII"),
1734         OPT_END()
1735         };
1736         const char * const latency_usage[] = {
1737                 "perf sched latency [<options>]",
1738                 NULL
1739         };
1740         const char * const replay_usage[] = {
1741                 "perf sched replay [<options>]",
1742                 NULL
1743         };
1744         const char *const sched_subcommands[] = { "record", "latency", "map",
1745                                                   "replay", "script", NULL };
1746         const char *sched_usage[] = {
1747                 NULL,
1748                 NULL
1749         };
1750         struct trace_sched_handler lat_ops  = {
1751                 .wakeup_event       = latency_wakeup_event,
1752                 .switch_event       = latency_switch_event,
1753                 .runtime_event      = latency_runtime_event,
1754                 .migrate_task_event = latency_migrate_task_event,
1755         };
1756         struct trace_sched_handler map_ops  = {
1757                 .switch_event       = map_switch_event,
1758         };
1759         struct trace_sched_handler replay_ops  = {
1760                 .wakeup_event       = replay_wakeup_event,
1761                 .switch_event       = replay_switch_event,
1762                 .fork_event         = replay_fork_event,
1763         };
1764         unsigned int i;
1765
1766         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1767                 sched.curr_pid[i] = -1;
1768
1769         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1770                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1771         if (!argc)
1772                 usage_with_options(sched_usage, sched_options);
1773
1774         /*
1775          * Aliased to 'perf script' for now:
1776          */
1777         if (!strcmp(argv[0], "script"))
1778                 return cmd_script(argc, argv, prefix);
1779
1780         if (!strncmp(argv[0], "rec", 3)) {
1781                 return __cmd_record(argc, argv);
1782         } else if (!strncmp(argv[0], "lat", 3)) {
1783                 sched.tp_handler = &lat_ops;
1784                 if (argc > 1) {
1785                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1786                         if (argc)
1787                                 usage_with_options(latency_usage, latency_options);
1788                 }
1789                 setup_sorting(&sched, latency_options, latency_usage);
1790                 return perf_sched__lat(&sched);
1791         } else if (!strcmp(argv[0], "map")) {
1792                 sched.tp_handler = &map_ops;
1793                 setup_sorting(&sched, latency_options, latency_usage);
1794                 return perf_sched__map(&sched);
1795         } else if (!strncmp(argv[0], "rep", 3)) {
1796                 sched.tp_handler = &replay_ops;
1797                 if (argc) {
1798                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1799                         if (argc)
1800                                 usage_with_options(replay_usage, replay_options);
1801                 }
1802                 return perf_sched__replay(&sched);
1803         } else {
1804                 usage_with_options(sched_usage, sched_options);
1805         }
1806
1807         return 0;
1808 }