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 #include "util/thread_map.h"
  15 #include "util/color.h"
  16 #include "util/stat.h"
  17 #include "util/callchain.h"
  18 #include "util/time-utils.h"
  19
  20 #include <subcmd/parse-options.h>
  21 #include "util/trace-event.h"
  22
  23 #include "util/debug.h"
  24
  25 #include <linux/log2.h>
  26 #include <sys/prctl.h>
  27 #include <sys/resource.h>
  28
  29 #include <semaphore.h>
  30 #include <pthread.h>
  31 #include <math.h>
  32 #include <api/fs/fs.h>
  33 #include <linux/time64.h>
  34
  35 #define PR_SET_NAME             15               /* Set process name */
  36 #define MAX_CPUS                4096
  37 #define COMM_LEN                20
  38 #define SYM_LEN                 129
  39 #define MAX_PID                 1024000
  40
  41 struct sched_atom;
  42
  43 struct task_desc {
  44         unsigned long           nr;
  45         unsigned long           pid;
  46         char                    comm[COMM_LEN];
  47
  48         unsigned long           nr_events;
  49         unsigned long           curr_event;
  50         struct sched_atom       **atoms;
  51
  52         pthread_t               thread;
  53         sem_t                   sleep_sem;
  54
  55         sem_t                   ready_for_work;
  56         sem_t                   work_done_sem;
  57
  58         u64                     cpu_usage;
  59 };
  60
  61 enum sched_event_type {
  62         SCHED_EVENT_RUN,
  63         SCHED_EVENT_SLEEP,
  64         SCHED_EVENT_WAKEUP,
  65         SCHED_EVENT_MIGRATION,
  66 };
  67
  68 struct sched_atom {
  69         enum sched_event_type   type;
  70         int                     specific_wait;
  71         u64                     timestamp;
  72         u64                     duration;
  73         unsigned long           nr;
  74         sem_t                   *wait_sem;
  75         struct task_desc        *wakee;
  76 };
  77
  78 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  79
  80 enum thread_state {
  81         THREAD_SLEEPING = 0,
  82         THREAD_WAIT_CPU,
  83         THREAD_SCHED_IN,
  84         THREAD_IGNORE
  85 };
  86
  87 struct work_atom {
  88         struct list_head        list;
  89         enum thread_state       state;
  90         u64                     sched_out_time;
  91         u64                     wake_up_time;
  92         u64                     sched_in_time;
  93         u64                     runtime;
  94 };
  95
  96 struct work_atoms {
  97         struct list_head        work_list;
  98         struct thread           *thread;
  99         struct rb_node          node;
 100         u64                     max_lat;
 101         u64                     max_lat_at;
 102         u64                     total_lat;
 103         u64                     nb_atoms;
 104         u64                     total_runtime;
 105         int                     num_merged;
 106 };
 107
 108 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 109
 110 struct perf_sched;
 111
 112 struct trace_sched_handler {
 113         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 114                             struct perf_sample *sample, struct machine *machine);
 115
 116         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 117                              struct perf_sample *sample, struct machine *machine);
 118
 119         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 120                             struct perf_sample *sample, struct machine *machine);
 121
 122         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 123         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 124                           struct machine *machine);
 125
 126         int (*migrate_task_event)(struct perf_sched *sched,
 127                                   struct perf_evsel *evsel,
 128                                   struct perf_sample *sample,
 129                                   struct machine *machine);
 130 };
 131
 132 #define COLOR_PIDS PERF_COLOR_BLUE
 133 #define COLOR_CPUS PERF_COLOR_BG_RED
 134
 135 struct perf_sched_map {
 136         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 137         int                     *comp_cpus;
 138         bool                     comp;
 139         struct thread_map       *color_pids;
 140         const char              *color_pids_str;
 141         struct cpu_map          *color_cpus;
 142         const char              *color_cpus_str;
 143         struct cpu_map          *cpus;
 144         const char              *cpus_str;
 145 };
 146
 147 struct perf_sched {
 148         struct perf_tool tool;
 149         const char       *sort_order;
 150         unsigned long    nr_tasks;
 151         struct task_desc **pid_to_task;
 152         struct task_desc **tasks;
 153         const struct trace_sched_handler *tp_handler;
 154         pthread_mutex_t  start_work_mutex;
 155         pthread_mutex_t  work_done_wait_mutex;
 156         int              profile_cpu;
 157 /*
 158  * Track the current task - that way we can know whether there's any
 159  * weird events, such as a task being switched away that is not current.
 160  */
 161         int              max_cpu;
 162         u32              curr_pid[MAX_CPUS];
 163         struct thread    *curr_thread[MAX_CPUS];
 164         char             next_shortname1;
 165         char             next_shortname2;
 166         unsigned int     replay_repeat;
 167         unsigned long    nr_run_events;
 168         unsigned long    nr_sleep_events;
 169         unsigned long    nr_wakeup_events;
 170         unsigned long    nr_sleep_corrections;
 171         unsigned long    nr_run_events_optimized;
 172         unsigned long    targetless_wakeups;
 173         unsigned long    multitarget_wakeups;
 174         unsigned long    nr_runs;
 175         unsigned long    nr_timestamps;
 176         unsigned long    nr_unordered_timestamps;
 177         unsigned long    nr_context_switch_bugs;
 178         unsigned long    nr_events;
 179         unsigned long    nr_lost_chunks;
 180         unsigned long    nr_lost_events;
 181         u64              run_measurement_overhead;
 182         u64              sleep_measurement_overhead;
 183         u64              start_time;
 184         u64              cpu_usage;
 185         u64              runavg_cpu_usage;
 186         u64              parent_cpu_usage;
 187         u64              runavg_parent_cpu_usage;
 188         u64              sum_runtime;
 189         u64              sum_fluct;
 190         u64              run_avg;
 191         u64              all_runtime;
 192         u64              all_count;
 193         u64              cpu_last_switched[MAX_CPUS];
 194         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 195         struct list_head sort_list, cmp_pid;
 196         bool force;
 197         bool skip_merge;
 198         struct perf_sched_map map;
 199
 200         /* options for timehist command */
 201         bool            summary;
 202         bool            summary_only;
 203         bool            idle_hist;
 204         bool            show_callchain;
 205         unsigned int    max_stack;
 206         bool            show_cpu_visual;
 207         bool            show_wakeups;
 208         bool            show_migrations;
 209         u64             skipped_samples;
 210         const char      *time_str;
 211         struct perf_time_interval ptime;
 212         struct perf_time_interval hist_time;
 213 };
 214
 215 /* per thread run time data */
 216 struct thread_runtime {
 217         u64 last_time;      /* time of previous sched in/out event */
 218         u64 dt_run;         /* run time */
 219         u64 dt_wait;        /* time between CPU access (off cpu) */
 220         u64 dt_delay;       /* time between wakeup and sched-in */
 221         u64 ready_to_run;   /* time of wakeup */
 222
 223         struct stats run_stats;
 224         u64 total_run_time;
 225
 226         u64 migrations;
 227 };
 228
 229 /* per event run time data */
 230 struct evsel_runtime {
 231         u64 *last_time; /* time this event was last seen per cpu */
 232         u32 ncpu;       /* highest cpu slot allocated */
 233 };
 234
 235 /* per cpu idle time data */
 236 struct idle_thread_runtime {
 237         struct thread_runtime   tr;
 238         struct thread           *last_thread;
 239         struct rb_root          sorted_root;
 240         struct callchain_root   callchain;
 241         struct callchain_cursor cursor;
 242 };
 243
 244 /* track idle times per cpu */
 245 static struct thread **idle_threads;
 246 static int idle_max_cpu;
 247 static char idle_comm[] = "<idle>";
 248
 249 static u64 get_nsecs(void)
 250 {
 251         struct timespec ts;
 252
 253         clock_gettime(CLOCK_MONOTONIC, &ts);
 254
 255         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 256 }
 257
 258 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 259 {
 260         u64 T0 = get_nsecs(), T1;
 261
 262         do {
 263                 T1 = get_nsecs();
 264         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 265 }
 266
 267 static void sleep_nsecs(u64 nsecs)
 268 {
 269         struct timespec ts;
 270
 271         ts.tv_nsec = nsecs % 999999999;
 272         ts.tv_sec = nsecs / 999999999;
 273
 274         nanosleep(&ts, NULL);
 275 }
 276
 277 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 278 {
 279         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 280         int i;
 281
 282         for (i = 0; i < 10; i++) {
 283                 T0 = get_nsecs();
 284                 burn_nsecs(sched, 0);
 285                 T1 = get_nsecs();
 286                 delta = T1-T0;
 287                 min_delta = min(min_delta, delta);
 288         }
 289         sched->run_measurement_overhead = min_delta;
 290
 291         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 292 }
 293
 294 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 295 {
 296         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 297         int i;
 298
 299         for (i = 0; i < 10; i++) {
 300                 T0 = get_nsecs();
 301                 sleep_nsecs(10000);
 302                 T1 = get_nsecs();
 303                 delta = T1-T0;
 304                 min_delta = min(min_delta, delta);
 305         }
 306         min_delta -= 10000;
 307         sched->sleep_measurement_overhead = min_delta;
 308
 309         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 310 }
 311
 312 static struct sched_atom *
 313 get_new_event(struct task_desc *task, u64 timestamp)
 314 {
 315         struct sched_atom *event = zalloc(sizeof(*event));
 316         unsigned long idx = task->nr_events;
 317         size_t size;
 318
 319         event->timestamp = timestamp;
 320         event->nr = idx;
 321
 322         task->nr_events++;
 323         size = sizeof(struct sched_atom *) * task->nr_events;
 324         task->atoms = realloc(task->atoms, size);
 325         BUG_ON(!task->atoms);
 326
 327         task->atoms[idx] = event;
 328
 329         return event;
 330 }
 331
 332 static struct sched_atom *last_event(struct task_desc *task)
 333 {
 334         if (!task->nr_events)
 335                 return NULL;
 336
 337         return task->atoms[task->nr_events - 1];
 338 }
 339
 340 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 341                                 u64 timestamp, u64 duration)
 342 {
 343         struct sched_atom *event, *curr_event = last_event(task);
 344
 345         /*
 346          * optimize an existing RUN event by merging this one
 347          * to it:
 348          */
 349         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 350                 sched->nr_run_events_optimized++;
 351                 curr_event->duration += duration;
 352                 return;
 353         }
 354
 355         event = get_new_event(task, timestamp);
 356
 357         event->type = SCHED_EVENT_RUN;
 358         event->duration = duration;
 359
 360         sched->nr_run_events++;
 361 }
 362
 363 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 364                                    u64 timestamp, struct task_desc *wakee)
 365 {
 366         struct sched_atom *event, *wakee_event;
 367
 368         event = get_new_event(task, timestamp);
 369         event->type = SCHED_EVENT_WAKEUP;
 370         event->wakee = wakee;
 371
 372         wakee_event = last_event(wakee);
 373         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 374                 sched->targetless_wakeups++;
 375                 return;
 376         }
 377         if (wakee_event->wait_sem) {
 378                 sched->multitarget_wakeups++;
 379                 return;
 380         }
 381
 382         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 383         sem_init(wakee_event->wait_sem, 0, 0);
 384         wakee_event->specific_wait = 1;
 385         event->wait_sem = wakee_event->wait_sem;
 386
 387         sched->nr_wakeup_events++;
 388 }
 389
 390 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 391                                   u64 timestamp, u64 task_state __maybe_unused)
 392 {
 393         struct sched_atom *event = get_new_event(task, timestamp);
 394
 395         event->type = SCHED_EVENT_SLEEP;
 396
 397         sched->nr_sleep_events++;
 398 }
 399
 400 static struct task_desc *register_pid(struct perf_sched *sched,
 401                                       unsigned long pid, const char *comm)
 402 {
 403         struct task_desc *task;
 404         static int pid_max;
 405
 406         if (sched->pid_to_task == NULL) {
 407                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 408                         pid_max = MAX_PID;
 409                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 410         }
 411         if (pid >= (unsigned long)pid_max) {
 412                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 413                         sizeof(struct task_desc *))) == NULL);
 414                 while (pid >= (unsigned long)pid_max)
 415                         sched->pid_to_task[pid_max++] = NULL;
 416         }
 417
 418         task = sched->pid_to_task[pid];
 419
 420         if (task)
 421                 return task;
 422
 423         task = zalloc(sizeof(*task));
 424         task->pid = pid;
 425         task->nr = sched->nr_tasks;
 426         strcpy(task->comm, comm);
 427         /*
 428          * every task starts in sleeping state - this gets ignored
 429          * if there's no wakeup pointing to this sleep state:
 430          */
 431         add_sched_event_sleep(sched, task, 0, 0);
 432
 433         sched->pid_to_task[pid] = task;
 434         sched->nr_tasks++;
 435         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 436         BUG_ON(!sched->tasks);
 437         sched->tasks[task->nr] = task;
 438
 439         if (verbose)
 440                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 441
 442         return task;
 443 }
 444
 445
 446 static void print_task_traces(struct perf_sched *sched)
 447 {
 448         struct task_desc *task;
 449         unsigned long i;
 450
 451         for (i = 0; i < sched->nr_tasks; i++) {
 452                 task = sched->tasks[i];
 453                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 454                         task->nr, task->comm, task->pid, task->nr_events);
 455         }
 456 }
 457
 458 static void add_cross_task_wakeups(struct perf_sched *sched)
 459 {
 460         struct task_desc *task1, *task2;
 461         unsigned long i, j;
 462
 463         for (i = 0; i < sched->nr_tasks; i++) {
 464                 task1 = sched->tasks[i];
 465                 j = i + 1;
 466                 if (j == sched->nr_tasks)
 467                         j = 0;
 468                 task2 = sched->tasks[j];
 469                 add_sched_event_wakeup(sched, task1, 0, task2);
 470         }
 471 }
 472
 473 static void perf_sched__process_event(struct perf_sched *sched,
 474                                       struct sched_atom *atom)
 475 {
 476         int ret = 0;
 477
 478         switch (atom->type) {
 479                 case SCHED_EVENT_RUN:
 480                         burn_nsecs(sched, atom->duration);
 481                         break;
 482                 case SCHED_EVENT_SLEEP:
 483                         if (atom->wait_sem)
 484                                 ret = sem_wait(atom->wait_sem);
 485                         BUG_ON(ret);
 486                         break;
 487                 case SCHED_EVENT_WAKEUP:
 488                         if (atom->wait_sem)
 489                                 ret = sem_post(atom->wait_sem);
 490                         BUG_ON(ret);
 491                         break;
 492                 case SCHED_EVENT_MIGRATION:
 493                         break;
 494                 default:
 495                         BUG_ON(1);
 496         }
 497 }
 498
 499 static u64 get_cpu_usage_nsec_parent(void)
 500 {
 501         struct rusage ru;
 502         u64 sum;
 503         int err;
 504
 505         err = getrusage(RUSAGE_SELF, &ru);
 506         BUG_ON(err);
 507
 508         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 509         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 510
 511         return sum;
 512 }
 513
 514 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 515 {
 516         struct perf_event_attr attr;
 517         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 518         int fd;
 519         struct rlimit limit;
 520         bool need_privilege = false;
 521
 522         memset(&attr, 0, sizeof(attr));
 523
 524         attr.type = PERF_TYPE_SOFTWARE;
 525         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 526
 527 force_again:
 528         fd = sys_perf_event_open(&attr, 0, -1, -1,
 529                                  perf_event_open_cloexec_flag());
 530
 531         if (fd < 0) {
 532                 if (errno == EMFILE) {
 533                         if (sched->force) {
 534                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 535                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 536                                 if (limit.rlim_cur > limit.rlim_max) {
 537                                         limit.rlim_max = limit.rlim_cur;
 538                                         need_privilege = true;
 539                                 }
 540                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 541                                         if (need_privilege && errno == EPERM)
 542                                                 strcpy(info, "Need privilege\n");
 543                                 } else
 544                                         goto force_again;
 545                         } else
 546                                 strcpy(info, "Have a try with -f option\n");
 547                 }
 548                 pr_err("Error: sys_perf_event_open() syscall returned "
 549                        "with %d (%s)\n%s", fd,
 550                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 551                 exit(EXIT_FAILURE);
 552         }
 553         return fd;
 554 }
 555
 556 static u64 get_cpu_usage_nsec_self(int fd)
 557 {
 558         u64 runtime;
 559         int ret;
 560
 561         ret = read(fd, &runtime, sizeof(runtime));
 562         BUG_ON(ret != sizeof(runtime));
 563
 564         return runtime;
 565 }
 566
 567 struct sched_thread_parms {
 568         struct task_desc  *task;
 569         struct perf_sched *sched;
 570         int fd;
 571 };
 572
 573 static void *thread_func(void *ctx)
 574 {
 575         struct sched_thread_parms *parms = ctx;
 576         struct task_desc *this_task = parms->task;
 577         struct perf_sched *sched = parms->sched;
 578         u64 cpu_usage_0, cpu_usage_1;
 579         unsigned long i, ret;
 580         char comm2[22];
 581         int fd = parms->fd;
 582
 583         zfree(&parms);
 584
 585         sprintf(comm2, ":%s", this_task->comm);
 586         prctl(PR_SET_NAME, comm2);
 587         if (fd < 0)
 588                 return NULL;
 589 again:
 590         ret = sem_post(&this_task->ready_for_work);
 591         BUG_ON(ret);
 592         ret = pthread_mutex_lock(&sched->start_work_mutex);
 593         BUG_ON(ret);
 594         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 595         BUG_ON(ret);
 596
 597         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 598
 599         for (i = 0; i < this_task->nr_events; i++) {
 600                 this_task->curr_event = i;
 601                 perf_sched__process_event(sched, this_task->atoms[i]);
 602         }
 603
 604         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 605         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 606         ret = sem_post(&this_task->work_done_sem);
 607         BUG_ON(ret);
 608
 609         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 610         BUG_ON(ret);
 611         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 612         BUG_ON(ret);
 613
 614         goto again;
 615 }
 616
 617 static void create_tasks(struct perf_sched *sched)
 618 {
 619         struct task_desc *task;
 620         pthread_attr_t attr;
 621         unsigned long i;
 622         int err;
 623
 624         err = pthread_attr_init(&attr);
 625         BUG_ON(err);
 626         err = pthread_attr_setstacksize(&attr,
 627                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 628         BUG_ON(err);
 629         err = pthread_mutex_lock(&sched->start_work_mutex);
 630         BUG_ON(err);
 631         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 632         BUG_ON(err);
 633         for (i = 0; i < sched->nr_tasks; i++) {
 634                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 635                 BUG_ON(parms == NULL);
 636                 parms->task = task = sched->tasks[i];
 637                 parms->sched = sched;
 638                 parms->fd = self_open_counters(sched, i);
 639                 sem_init(&task->sleep_sem, 0, 0);
 640                 sem_init(&task->ready_for_work, 0, 0);
 641                 sem_init(&task->work_done_sem, 0, 0);
 642                 task->curr_event = 0;
 643                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 644                 BUG_ON(err);
 645         }
 646 }
 647
 648 static void wait_for_tasks(struct perf_sched *sched)
 649 {
 650         u64 cpu_usage_0, cpu_usage_1;
 651         struct task_desc *task;
 652         unsigned long i, ret;
 653
 654         sched->start_time = get_nsecs();
 655         sched->cpu_usage = 0;
 656         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 657
 658         for (i = 0; i < sched->nr_tasks; i++) {
 659                 task = sched->tasks[i];
 660                 ret = sem_wait(&task->ready_for_work);
 661                 BUG_ON(ret);
 662                 sem_init(&task->ready_for_work, 0, 0);
 663         }
 664         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 665         BUG_ON(ret);
 666
 667         cpu_usage_0 = get_cpu_usage_nsec_parent();
 668
 669         pthread_mutex_unlock(&sched->start_work_mutex);
 670
 671         for (i = 0; i < sched->nr_tasks; i++) {
 672                 task = sched->tasks[i];
 673                 ret = sem_wait(&task->work_done_sem);
 674                 BUG_ON(ret);
 675                 sem_init(&task->work_done_sem, 0, 0);
 676                 sched->cpu_usage += task->cpu_usage;
 677                 task->cpu_usage = 0;
 678         }
 679
 680         cpu_usage_1 = get_cpu_usage_nsec_parent();
 681         if (!sched->runavg_cpu_usage)
 682                 sched->runavg_cpu_usage = sched->cpu_usage;
 683         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 684
 685         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 686         if (!sched->runavg_parent_cpu_usage)
 687                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 688         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 689                                          sched->parent_cpu_usage)/sched->replay_repeat;
 690
 691         ret = pthread_mutex_lock(&sched->start_work_mutex);
 692         BUG_ON(ret);
 693
 694         for (i = 0; i < sched->nr_tasks; i++) {
 695                 task = sched->tasks[i];
 696                 sem_init(&task->sleep_sem, 0, 0);
 697                 task->curr_event = 0;
 698         }
 699 }
 700
 701 static void run_one_test(struct perf_sched *sched)
 702 {
 703         u64 T0, T1, delta, avg_delta, fluct;
 704
 705         T0 = get_nsecs();
 706         wait_for_tasks(sched);
 707         T1 = get_nsecs();
 708
 709         delta = T1 - T0;
 710         sched->sum_runtime += delta;
 711         sched->nr_runs++;
 712
 713         avg_delta = sched->sum_runtime / sched->nr_runs;
 714         if (delta < avg_delta)
 715                 fluct = avg_delta - delta;
 716         else
 717                 fluct = delta - avg_delta;
 718         sched->sum_fluct += fluct;
 719         if (!sched->run_avg)
 720                 sched->run_avg = delta;
 721         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 722
 723         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 724
 725         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 726
 727         printf("cpu: %0.2f / %0.2f",
 728                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 729
 730 #if 0
 731         /*
 732          * rusage statistics done by the parent, these are less
 733          * accurate than the sched->sum_exec_runtime based statistics:
 734          */
 735         printf(" [%0.2f / %0.2f]",
 736                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 737                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 738 #endif
 739
 740         printf("\n");
 741
 742         if (sched->nr_sleep_corrections)
 743                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 744         sched->nr_sleep_corrections = 0;
 745 }
 746
 747 static void test_calibrations(struct perf_sched *sched)
 748 {
 749         u64 T0, T1;
 750
 751         T0 = get_nsecs();
 752         burn_nsecs(sched, NSEC_PER_MSEC);
 753         T1 = get_nsecs();
 754
 755         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 756
 757         T0 = get_nsecs();
 758         sleep_nsecs(NSEC_PER_MSEC);
 759         T1 = get_nsecs();
 760
 761         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 762 }
 763
 764 static int
 765 replay_wakeup_event(struct perf_sched *sched,
 766                     struct perf_evsel *evsel, struct perf_sample *sample,
 767                     struct machine *machine __maybe_unused)
 768 {
 769         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 770         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 771         struct task_desc *waker, *wakee;
 772
 773         if (verbose) {
 774                 printf("sched_wakeup event %p\n", evsel);
 775
 776                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 777         }
 778
 779         waker = register_pid(sched, sample->tid, "<unknown>");
 780         wakee = register_pid(sched, pid, comm);
 781
 782         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 783         return 0;
 784 }
 785
 786 static int replay_switch_event(struct perf_sched *sched,
 787                                struct perf_evsel *evsel,
 788                                struct perf_sample *sample,
 789                                struct machine *machine __maybe_unused)
 790 {
 791         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 792                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 793         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 794                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 795         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 796         struct task_desc *prev, __maybe_unused *next;
 797         u64 timestamp0, timestamp = sample->time;
 798         int cpu = sample->cpu;
 799         s64 delta;
 800
 801         if (verbose)
 802                 printf("sched_switch event %p\n", evsel);
 803
 804         if (cpu >= MAX_CPUS || cpu < 0)
 805                 return 0;
 806
 807         timestamp0 = sched->cpu_last_switched[cpu];
 808         if (timestamp0)
 809                 delta = timestamp - timestamp0;
 810         else
 811                 delta = 0;
 812
 813         if (delta < 0) {
 814                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 815                 return -1;
 816         }
 817
 818         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 819                  prev_comm, prev_pid, next_comm, next_pid, delta);
 820
 821         prev = register_pid(sched, prev_pid, prev_comm);
 822         next = register_pid(sched, next_pid, next_comm);
 823
 824         sched->cpu_last_switched[cpu] = timestamp;
 825
 826         add_sched_event_run(sched, prev, timestamp, delta);
 827         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 828
 829         return 0;
 830 }
 831
 832 static int replay_fork_event(struct perf_sched *sched,
 833                              union perf_event *event,
 834                              struct machine *machine)
 835 {
 836         struct thread *child, *parent;
 837
 838         child = machine__findnew_thread(machine, event->fork.pid,
 839                                         event->fork.tid);
 840         parent = machine__findnew_thread(machine, event->fork.ppid,
 841                                          event->fork.ptid);
 842
 843         if (child == NULL || parent == NULL) {
 844                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 845                                  child, parent);
 846                 goto out_put;
 847         }
 848
 849         if (verbose) {
 850                 printf("fork event\n");
 851                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 852                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 853         }
 854
 855         register_pid(sched, parent->tid, thread__comm_str(parent));
 856         register_pid(sched, child->tid, thread__comm_str(child));
 857 out_put:
 858         thread__put(child);
 859         thread__put(parent);
 860         return 0;
 861 }
 862
 863 struct sort_dimension {
 864         const char              *name;
 865         sort_fn_t               cmp;
 866         struct list_head        list;
 867 };
 868
 869 static int
 870 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 871 {
 872         struct sort_dimension *sort;
 873         int ret = 0;
 874
 875         BUG_ON(list_empty(list));
 876
 877         list_for_each_entry(sort, list, list) {
 878                 ret = sort->cmp(l, r);
 879                 if (ret)
 880                         return ret;
 881         }
 882
 883         return ret;
 884 }
 885
 886 static struct work_atoms *
 887 thread_atoms_search(struct rb_root *root, struct thread *thread,
 888                          struct list_head *sort_list)
 889 {
 890         struct rb_node *node = root->rb_node;
 891         struct work_atoms key = { .thread = thread };
 892
 893         while (node) {
 894                 struct work_atoms *atoms;
 895                 int cmp;
 896
 897                 atoms = container_of(node, struct work_atoms, node);
 898
 899                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 900                 if (cmp > 0)
 901                         node = node->rb_left;
 902                 else if (cmp < 0)
 903                         node = node->rb_right;
 904                 else {
 905                         BUG_ON(thread != atoms->thread);
 906                         return atoms;
 907                 }
 908         }
 909         return NULL;
 910 }
 911
 912 static void
 913 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 914                          struct list_head *sort_list)
 915 {
 916         struct rb_node **new = &(root->rb_node), *parent = NULL;
 917
 918         while (*new) {
 919                 struct work_atoms *this;
 920                 int cmp;
 921
 922                 this = container_of(*new, struct work_atoms, node);
 923                 parent = *new;
 924
 925                 cmp = thread_lat_cmp(sort_list, data, this);
 926
 927                 if (cmp > 0)
 928                         new = &((*new)->rb_left);
 929                 else
 930                         new = &((*new)->rb_right);
 931         }
 932
 933         rb_link_node(&data->node, parent, new);
 934         rb_insert_color(&data->node, root);
 935 }
 936
 937 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 938 {
 939         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 940         if (!atoms) {
 941                 pr_err("No memory at %s\n", __func__);
 942                 return -1;
 943         }
 944
 945         atoms->thread = thread__get(thread);
 946         INIT_LIST_HEAD(&atoms->work_list);
 947         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 948         return 0;
 949 }
 950
 951 static char sched_out_state(u64 prev_state)
 952 {
 953         const char *str = TASK_STATE_TO_CHAR_STR;
 954
 955         return str[prev_state];
 956 }
 957
 958 static int
 959 add_sched_out_event(struct work_atoms *atoms,
 960                     char run_state,
 961                     u64 timestamp)
 962 {
 963         struct work_atom *atom = zalloc(sizeof(*atom));
 964         if (!atom) {
 965                 pr_err("Non memory at %s", __func__);
 966                 return -1;
 967         }
 968
 969         atom->sched_out_time = timestamp;
 970
 971         if (run_state == 'R') {
 972                 atom->state = THREAD_WAIT_CPU;
 973                 atom->wake_up_time = atom->sched_out_time;
 974         }
 975
 976         list_add_tail(&atom->list, &atoms->work_list);
 977         return 0;
 978 }
 979
 980 static void
 981 add_runtime_event(struct work_atoms *atoms, u64 delta,
 982                   u64 timestamp __maybe_unused)
 983 {
 984         struct work_atom *atom;
 985
 986         BUG_ON(list_empty(&atoms->work_list));
 987
 988         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 989
 990         atom->runtime += delta;
 991         atoms->total_runtime += delta;
 992 }
 993
 994 static void
 995 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 996 {
 997         struct work_atom *atom;
 998         u64 delta;
 999
1000         if (list_empty(&atoms->work_list))
1001                 return;
1002
1003         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1004
1005         if (atom->state != THREAD_WAIT_CPU)
1006                 return;
1007
1008         if (timestamp < atom->wake_up_time) {
1009                 atom->state = THREAD_IGNORE;
1010                 return;
1011         }
1012
1013         atom->state = THREAD_SCHED_IN;
1014         atom->sched_in_time = timestamp;
1015
1016         delta = atom->sched_in_time - atom->wake_up_time;
1017         atoms->total_lat += delta;
1018         if (delta > atoms->max_lat) {
1019                 atoms->max_lat = delta;
1020                 atoms->max_lat_at = timestamp;
1021         }
1022         atoms->nb_atoms++;
1023 }
1024
1025 static int latency_switch_event(struct perf_sched *sched,
1026                                 struct perf_evsel *evsel,
1027                                 struct perf_sample *sample,
1028                                 struct machine *machine)
1029 {
1030         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1031                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1032         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1033         struct work_atoms *out_events, *in_events;
1034         struct thread *sched_out, *sched_in;
1035         u64 timestamp0, timestamp = sample->time;
1036         int cpu = sample->cpu, err = -1;
1037         s64 delta;
1038
1039         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040
1041         timestamp0 = sched->cpu_last_switched[cpu];
1042         sched->cpu_last_switched[cpu] = timestamp;
1043         if (timestamp0)
1044                 delta = timestamp - timestamp0;
1045         else
1046                 delta = 0;
1047
1048         if (delta < 0) {
1049                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1050                 return -1;
1051         }
1052
1053         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1054         sched_in = machine__findnew_thread(machine, -1, next_pid);
1055         if (sched_out == NULL || sched_in == NULL)
1056                 goto out_put;
1057
1058         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1059         if (!out_events) {
1060                 if (thread_atoms_insert(sched, sched_out))
1061                         goto out_put;
1062                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1063                 if (!out_events) {
1064                         pr_err("out-event: Internal tree error");
1065                         goto out_put;
1066                 }
1067         }
1068         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1069                 return -1;
1070
1071         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1072         if (!in_events) {
1073                 if (thread_atoms_insert(sched, sched_in))
1074                         goto out_put;
1075                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1076                 if (!in_events) {
1077                         pr_err("in-event: Internal tree error");
1078                         goto out_put;
1079                 }
1080                 /*
1081                  * Take came in we have not heard about yet,
1082                  * add in an initial atom in runnable state:
1083                  */
1084                 if (add_sched_out_event(in_events, 'R', timestamp))
1085                         goto out_put;
1086         }
1087         add_sched_in_event(in_events, timestamp);
1088         err = 0;
1089 out_put:
1090         thread__put(sched_out);
1091         thread__put(sched_in);
1092         return err;
1093 }
1094
1095 static int latency_runtime_event(struct perf_sched *sched,
1096                                  struct perf_evsel *evsel,
1097                                  struct perf_sample *sample,
1098                                  struct machine *machine)
1099 {
1100         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1101         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1102         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1103         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1104         u64 timestamp = sample->time;
1105         int cpu = sample->cpu, err = -1;
1106
1107         if (thread == NULL)
1108                 return -1;
1109
1110         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1111         if (!atoms) {
1112                 if (thread_atoms_insert(sched, thread))
1113                         goto out_put;
1114                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1115                 if (!atoms) {
1116                         pr_err("in-event: Internal tree error");
1117                         goto out_put;
1118                 }
1119                 if (add_sched_out_event(atoms, 'R', timestamp))
1120                         goto out_put;
1121         }
1122
1123         add_runtime_event(atoms, runtime, timestamp);
1124         err = 0;
1125 out_put:
1126         thread__put(thread);
1127         return err;
1128 }
1129
1130 static int latency_wakeup_event(struct perf_sched *sched,
1131                                 struct perf_evsel *evsel,
1132                                 struct perf_sample *sample,
1133                                 struct machine *machine)
1134 {
1135         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1136         struct work_atoms *atoms;
1137         struct work_atom *atom;
1138         struct thread *wakee;
1139         u64 timestamp = sample->time;
1140         int err = -1;
1141
1142         wakee = machine__findnew_thread(machine, -1, pid);
1143         if (wakee == NULL)
1144                 return -1;
1145         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1146         if (!atoms) {
1147                 if (thread_atoms_insert(sched, wakee))
1148                         goto out_put;
1149                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1150                 if (!atoms) {
1151                         pr_err("wakeup-event: Internal tree error");
1152                         goto out_put;
1153                 }
1154                 if (add_sched_out_event(atoms, 'S', timestamp))
1155                         goto out_put;
1156         }
1157
1158         BUG_ON(list_empty(&atoms->work_list));
1159
1160         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1161
1162         /*
1163          * As we do not guarantee the wakeup event happens when
1164          * task is out of run queue, also may happen when task is
1165          * on run queue and wakeup only change ->state to TASK_RUNNING,
1166          * then we should not set the ->wake_up_time when wake up a
1167          * task which is on run queue.
1168          *
1169          * You WILL be missing events if you've recorded only
1170          * one CPU, or are only looking at only one, so don't
1171          * skip in this case.
1172          */
1173         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1174                 goto out_ok;
1175
1176         sched->nr_timestamps++;
1177         if (atom->sched_out_time > timestamp) {
1178                 sched->nr_unordered_timestamps++;
1179                 goto out_ok;
1180         }
1181
1182         atom->state = THREAD_WAIT_CPU;
1183         atom->wake_up_time = timestamp;
1184 out_ok:
1185         err = 0;
1186 out_put:
1187         thread__put(wakee);
1188         return err;
1189 }
1190
1191 static int latency_migrate_task_event(struct perf_sched *sched,
1192                                       struct perf_evsel *evsel,
1193                                       struct perf_sample *sample,
1194                                       struct machine *machine)
1195 {
1196         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1197         u64 timestamp = sample->time;
1198         struct work_atoms *atoms;
1199         struct work_atom *atom;
1200         struct thread *migrant;
1201         int err = -1;
1202
1203         /*
1204          * Only need to worry about migration when profiling one CPU.
1205          */
1206         if (sched->profile_cpu == -1)
1207                 return 0;
1208
1209         migrant = machine__findnew_thread(machine, -1, pid);
1210         if (migrant == NULL)
1211                 return -1;
1212         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1213         if (!atoms) {
1214                 if (thread_atoms_insert(sched, migrant))
1215                         goto out_put;
1216                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1217                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1218                 if (!atoms) {
1219                         pr_err("migration-event: Internal tree error");
1220                         goto out_put;
1221                 }
1222                 if (add_sched_out_event(atoms, 'R', timestamp))
1223                         goto out_put;
1224         }
1225
1226         BUG_ON(list_empty(&atoms->work_list));
1227
1228         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1229         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1230
1231         sched->nr_timestamps++;
1232
1233         if (atom->sched_out_time > timestamp)
1234                 sched->nr_unordered_timestamps++;
1235         err = 0;
1236 out_put:
1237         thread__put(migrant);
1238         return err;
1239 }
1240
1241 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1242 {
1243         int i;
1244         int ret;
1245         u64 avg;
1246         char max_lat_at[32];
1247
1248         if (!work_list->nb_atoms)
1249                 return;
1250         /*
1251          * Ignore idle threads:
1252          */
1253         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1254                 return;
1255
1256         sched->all_runtime += work_list->total_runtime;
1257         sched->all_count   += work_list->nb_atoms;
1258
1259         if (work_list->num_merged > 1)
1260                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1261         else
1262                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1263
1264         for (i = 0; i < 24 - ret; i++)
1265                 printf(" ");
1266
1267         avg = work_list->total_lat / work_list->nb_atoms;
1268         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1269
1270         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1271               (double)work_list->total_runtime / NSEC_PER_MSEC,
1272                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1273                  (double)work_list->max_lat / NSEC_PER_MSEC,
1274                  max_lat_at);
1275 }
1276
1277 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1278 {
1279         if (l->thread == r->thread)
1280                 return 0;
1281         if (l->thread->tid < r->thread->tid)
1282                 return -1;
1283         if (l->thread->tid > r->thread->tid)
1284                 return 1;
1285         return (int)(l->thread - r->thread);
1286 }
1287
1288 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1289 {
1290         u64 avgl, avgr;
1291
1292         if (!l->nb_atoms)
1293                 return -1;
1294
1295         if (!r->nb_atoms)
1296                 return 1;
1297
1298         avgl = l->total_lat / l->nb_atoms;
1299         avgr = r->total_lat / r->nb_atoms;
1300
1301         if (avgl < avgr)
1302                 return -1;
1303         if (avgl > avgr)
1304                 return 1;
1305
1306         return 0;
1307 }
1308
1309 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1310 {
1311         if (l->max_lat < r->max_lat)
1312                 return -1;
1313         if (l->max_lat > r->max_lat)
1314                 return 1;
1315
1316         return 0;
1317 }
1318
1319 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1320 {
1321         if (l->nb_atoms < r->nb_atoms)
1322                 return -1;
1323         if (l->nb_atoms > r->nb_atoms)
1324                 return 1;
1325
1326         return 0;
1327 }
1328
1329 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1330 {
1331         if (l->total_runtime < r->total_runtime)
1332                 return -1;
1333         if (l->total_runtime > r->total_runtime)
1334                 return 1;
1335
1336         return 0;
1337 }
1338
1339 static int sort_dimension__add(const char *tok, struct list_head *list)
1340 {
1341         size_t i;
1342         static struct sort_dimension avg_sort_dimension = {
1343                 .name = "avg",
1344                 .cmp  = avg_cmp,
1345         };
1346         static struct sort_dimension max_sort_dimension = {
1347                 .name = "max",
1348                 .cmp  = max_cmp,
1349         };
1350         static struct sort_dimension pid_sort_dimension = {
1351                 .name = "pid",
1352                 .cmp  = pid_cmp,
1353         };
1354         static struct sort_dimension runtime_sort_dimension = {
1355                 .name = "runtime",
1356                 .cmp  = runtime_cmp,
1357         };
1358         static struct sort_dimension switch_sort_dimension = {
1359                 .name = "switch",
1360                 .cmp  = switch_cmp,
1361         };
1362         struct sort_dimension *available_sorts[] = {
1363                 &pid_sort_dimension,
1364                 &avg_sort_dimension,
1365                 &max_sort_dimension,
1366                 &switch_sort_dimension,
1367                 &runtime_sort_dimension,
1368         };
1369
1370         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1371                 if (!strcmp(available_sorts[i]->name, tok)) {
1372                         list_add_tail(&available_sorts[i]->list, list);
1373
1374                         return 0;
1375                 }
1376         }
1377
1378         return -1;
1379 }
1380
1381 static void perf_sched__sort_lat(struct perf_sched *sched)
1382 {
1383         struct rb_node *node;
1384         struct rb_root *root = &sched->atom_root;
1385 again:
1386         for (;;) {
1387                 struct work_atoms *data;
1388                 node = rb_first(root);
1389                 if (!node)
1390                         break;
1391
1392                 rb_erase(node, root);
1393                 data = rb_entry(node, struct work_atoms, node);
1394                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1395         }
1396         if (root == &sched->atom_root) {
1397                 root = &sched->merged_atom_root;
1398                 goto again;
1399         }
1400 }
1401
1402 static int process_sched_wakeup_event(struct perf_tool *tool,
1403                                       struct perf_evsel *evsel,
1404                                       struct perf_sample *sample,
1405                                       struct machine *machine)
1406 {
1407         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1408
1409         if (sched->tp_handler->wakeup_event)
1410                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1411
1412         return 0;
1413 }
1414
1415 union map_priv {
1416         void    *ptr;
1417         bool     color;
1418 };
1419
1420 static bool thread__has_color(struct thread *thread)
1421 {
1422         union map_priv priv = {
1423                 .ptr = thread__priv(thread),
1424         };
1425
1426         return priv.color;
1427 }
1428
1429 static struct thread*
1430 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1431 {
1432         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1433         union map_priv priv = {
1434                 .color = false,
1435         };
1436
1437         if (!sched->map.color_pids || !thread || thread__priv(thread))
1438                 return thread;
1439
1440         if (thread_map__has(sched->map.color_pids, tid))
1441                 priv.color = true;
1442
1443         thread__set_priv(thread, priv.ptr);
1444         return thread;
1445 }
1446
1447 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1448                             struct perf_sample *sample, struct machine *machine)
1449 {
1450         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1451         struct thread *sched_in;
1452         int new_shortname;
1453         u64 timestamp0, timestamp = sample->time;
1454         s64 delta;
1455         int i, this_cpu = sample->cpu;
1456         int cpus_nr;
1457         bool new_cpu = false;
1458         const char *color = PERF_COLOR_NORMAL;
1459         char stimestamp[32];
1460
1461         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1462
1463         if (this_cpu > sched->max_cpu)
1464                 sched->max_cpu = this_cpu;
1465
1466         if (sched->map.comp) {
1467                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1468                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1469                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1470                         new_cpu = true;
1471                 }
1472         } else
1473                 cpus_nr = sched->max_cpu;
1474
1475         timestamp0 = sched->cpu_last_switched[this_cpu];
1476         sched->cpu_last_switched[this_cpu] = timestamp;
1477         if (timestamp0)
1478                 delta = timestamp - timestamp0;
1479         else
1480                 delta = 0;
1481
1482         if (delta < 0) {
1483                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1484                 return -1;
1485         }
1486
1487         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1488         if (sched_in == NULL)
1489                 return -1;
1490
1491         sched->curr_thread[this_cpu] = thread__get(sched_in);
1492
1493         printf("  ");
1494
1495         new_shortname = 0;
1496         if (!sched_in->shortname[0]) {
1497                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1498                         /*
1499                          * Don't allocate a letter-number for swapper:0
1500                          * as a shortname. Instead, we use '.' for it.
1501                          */
1502                         sched_in->shortname[0] = '.';
1503                         sched_in->shortname[1] = ' ';
1504                 } else {
1505                         sched_in->shortname[0] = sched->next_shortname1;
1506                         sched_in->shortname[1] = sched->next_shortname2;
1507
1508                         if (sched->next_shortname1 < 'Z') {
1509                                 sched->next_shortname1++;
1510                         } else {
1511                                 sched->next_shortname1 = 'A';
1512                                 if (sched->next_shortname2 < '9')
1513                                         sched->next_shortname2++;
1514                                 else
1515                                         sched->next_shortname2 = '0';
1516                         }
1517                 }
1518                 new_shortname = 1;
1519         }
1520
1521         for (i = 0; i < cpus_nr; i++) {
1522                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1523                 struct thread *curr_thread = sched->curr_thread[cpu];
1524                 const char *pid_color = color;
1525                 const char *cpu_color = color;
1526
1527                 if (curr_thread && thread__has_color(curr_thread))
1528                         pid_color = COLOR_PIDS;
1529
1530                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1531                         continue;
1532
1533                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1534                         cpu_color = COLOR_CPUS;
1535
1536                 if (cpu != this_cpu)
1537                         color_fprintf(stdout, color, " ");
1538                 else
1539                         color_fprintf(stdout, cpu_color, "*");
1540
1541                 if (sched->curr_thread[cpu])
1542                         color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1543                 else
1544                         color_fprintf(stdout, color, "   ");
1545         }
1546
1547         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1548                 goto out;
1549
1550         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1551         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1552         if (new_shortname || (verbose && sched_in->tid)) {
1553                 const char *pid_color = color;
1554
1555                 if (thread__has_color(sched_in))
1556                         pid_color = COLOR_PIDS;
1557
1558                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1559                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1560         }
1561
1562         if (sched->map.comp && new_cpu)
1563                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1564
1565 out:
1566         color_fprintf(stdout, color, "\n");
1567
1568         thread__put(sched_in);
1569
1570         return 0;
1571 }
1572
1573 static int process_sched_switch_event(struct perf_tool *tool,
1574                                       struct perf_evsel *evsel,
1575                                       struct perf_sample *sample,
1576                                       struct machine *machine)
1577 {
1578         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1579         int this_cpu = sample->cpu, err = 0;
1580         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1581             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1582
1583         if (sched->curr_pid[this_cpu] != (u32)-1) {
1584                 /*
1585                  * Are we trying to switch away a PID that is
1586                  * not current?
1587                  */
1588                 if (sched->curr_pid[this_cpu] != prev_pid)
1589                         sched->nr_context_switch_bugs++;
1590         }
1591
1592         if (sched->tp_handler->switch_event)
1593                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1594
1595         sched->curr_pid[this_cpu] = next_pid;
1596         return err;
1597 }
1598
1599 static int process_sched_runtime_event(struct perf_tool *tool,
1600                                        struct perf_evsel *evsel,
1601                                        struct perf_sample *sample,
1602                                        struct machine *machine)
1603 {
1604         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1605
1606         if (sched->tp_handler->runtime_event)
1607                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1608
1609         return 0;
1610 }
1611
1612 static int perf_sched__process_fork_event(struct perf_tool *tool,
1613                                           union perf_event *event,
1614                                           struct perf_sample *sample,
1615                                           struct machine *machine)
1616 {
1617         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1618
1619         /* run the fork event through the perf machineruy */
1620         perf_event__process_fork(tool, event, sample, machine);
1621
1622         /* and then run additional processing needed for this command */
1623         if (sched->tp_handler->fork_event)
1624                 return sched->tp_handler->fork_event(sched, event, machine);
1625
1626         return 0;
1627 }
1628
1629 static int process_sched_migrate_task_event(struct perf_tool *tool,
1630                                             struct perf_evsel *evsel,
1631                                             struct perf_sample *sample,
1632                                             struct machine *machine)
1633 {
1634         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1635
1636         if (sched->tp_handler->migrate_task_event)
1637                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1638
1639         return 0;
1640 }
1641
1642 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1643                                   struct perf_evsel *evsel,
1644                                   struct perf_sample *sample,
1645                                   struct machine *machine);
1646
1647 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1648                                                  union perf_event *event __maybe_unused,
1649                                                  struct perf_sample *sample,
1650                                                  struct perf_evsel *evsel,
1651                                                  struct machine *machine)
1652 {
1653         int err = 0;
1654
1655         if (evsel->handler != NULL) {
1656                 tracepoint_handler f = evsel->handler;
1657                 err = f(tool, evsel, sample, machine);
1658         }
1659
1660         return err;
1661 }
1662
1663 static int perf_sched__read_events(struct perf_sched *sched)
1664 {
1665         const struct perf_evsel_str_handler handlers[] = {
1666                 { "sched:sched_switch",       process_sched_switch_event, },
1667                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1668                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1669                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1670                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1671         };
1672         struct perf_session *session;
1673         struct perf_data_file file = {
1674                 .path = input_name,
1675                 .mode = PERF_DATA_MODE_READ,
1676                 .force = sched->force,
1677         };
1678         int rc = -1;
1679
1680         session = perf_session__new(&file, false, &sched->tool);
1681         if (session == NULL) {
1682                 pr_debug("No Memory for session\n");
1683                 return -1;
1684         }
1685
1686         symbol__init(&session->header.env);
1687
1688         if (perf_session__set_tracepoints_handlers(session, handlers))
1689                 goto out_delete;
1690
1691         if (perf_session__has_traces(session, "record -R")) {
1692                 int err = perf_session__process_events(session);
1693                 if (err) {
1694                         pr_err("Failed to process events, error %d", err);
1695                         goto out_delete;
1696                 }
1697
1698                 sched->nr_events      = session->evlist->stats.nr_events[0];
1699                 sched->nr_lost_events = session->evlist->stats.total_lost;
1700                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1701         }
1702
1703         rc = 0;
1704 out_delete:
1705         perf_session__delete(session);
1706         return rc;
1707 }
1708
1709 /*
1710  * scheduling times are printed as msec.usec
1711  */
1712 static inline void print_sched_time(unsigned long long nsecs, int width)
1713 {
1714         unsigned long msecs;
1715         unsigned long usecs;
1716
1717         msecs  = nsecs / NSEC_PER_MSEC;
1718         nsecs -= msecs * NSEC_PER_MSEC;
1719         usecs  = nsecs / NSEC_PER_USEC;
1720         printf("%*lu.%03lu ", width, msecs, usecs);
1721 }
1722
1723 /*
1724  * returns runtime data for event, allocating memory for it the
1725  * first time it is used.
1726  */
1727 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1728 {
1729         struct evsel_runtime *r = evsel->priv;
1730
1731         if (r == NULL) {
1732                 r = zalloc(sizeof(struct evsel_runtime));
1733                 evsel->priv = r;
1734         }
1735
1736         return r;
1737 }
1738
1739 /*
1740  * save last time event was seen per cpu
1741  */
1742 static void perf_evsel__save_time(struct perf_evsel *evsel,
1743                                   u64 timestamp, u32 cpu)
1744 {
1745         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1746
1747         if (r == NULL)
1748                 return;
1749
1750         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1751                 int i, n = __roundup_pow_of_two(cpu+1);
1752                 void *p = r->last_time;
1753
1754                 p = realloc(r->last_time, n * sizeof(u64));
1755                 if (!p)
1756                         return;
1757
1758                 r->last_time = p;
1759                 for (i = r->ncpu; i < n; ++i)
1760                         r->last_time[i] = (u64) 0;
1761
1762                 r->ncpu = n;
1763         }
1764
1765         r->last_time[cpu] = timestamp;
1766 }
1767
1768 /* returns last time this event was seen on the given cpu */
1769 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1770 {
1771         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1772
1773         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1774                 return 0;
1775
1776         return r->last_time[cpu];
1777 }
1778
1779 static int comm_width = 30;
1780
1781 static char *timehist_get_commstr(struct thread *thread)
1782 {
1783         static char str[32];
1784         const char *comm = thread__comm_str(thread);
1785         pid_t tid = thread->tid;
1786         pid_t pid = thread->pid_;
1787         int n;
1788
1789         if (pid == 0)
1790                 n = scnprintf(str, sizeof(str), "%s", comm);
1791
1792         else if (tid != pid)
1793                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1794
1795         else
1796                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1797
1798         if (n > comm_width)
1799                 comm_width = n;
1800
1801         return str;
1802 }
1803
1804 static void timehist_header(struct perf_sched *sched)
1805 {
1806         u32 ncpus = sched->max_cpu + 1;
1807         u32 i, j;
1808
1809         printf("%15s %6s ", "time", "cpu");
1810
1811         if (sched->show_cpu_visual) {
1812                 printf(" ");
1813                 for (i = 0, j = 0; i < ncpus; ++i) {
1814                         printf("%x", j++);
1815                         if (j > 15)
1816                                 j = 0;
1817                 }
1818                 printf(" ");
1819         }
1820
1821         printf(" %-*s  %9s  %9s  %9s", comm_width,
1822                 "task name", "wait time", "sch delay", "run time");
1823
1824         printf("\n");
1825
1826         /*
1827          * units row
1828          */
1829         printf("%15s %-6s ", "", "");
1830
1831         if (sched->show_cpu_visual)
1832                 printf(" %*s ", ncpus, "");
1833
1834         printf(" %-*s  %9s  %9s  %9s\n", comm_width,
1835                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1836
1837         /*
1838          * separator
1839          */
1840         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1841
1842         if (sched->show_cpu_visual)
1843                 printf(" %.*s ", ncpus, graph_dotted_line);
1844
1845         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1846                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1847                 graph_dotted_line);
1848
1849         printf("\n");
1850 }
1851
1852 static void timehist_print_sample(struct perf_sched *sched,
1853                                   struct perf_sample *sample,
1854                                   struct addr_location *al,
1855                                   struct thread *thread,
1856                                   u64 t)
1857 {
1858         struct thread_runtime *tr = thread__priv(thread);
1859         u32 max_cpus = sched->max_cpu + 1;
1860         char tstr[64];
1861
1862         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1863         printf("%15s [%04d] ", tstr, sample->cpu);
1864
1865         if (sched->show_cpu_visual) {
1866                 u32 i;
1867                 char c;
1868
1869                 printf(" ");
1870                 for (i = 0; i < max_cpus; ++i) {
1871                         /* flag idle times with 'i'; others are sched events */
1872                         if (i == sample->cpu)
1873                                 c = (thread->tid == 0) ? 'i' : 's';
1874                         else
1875                                 c = ' ';
1876                         printf("%c", c);
1877                 }
1878                 printf(" ");
1879         }
1880
1881         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1882
1883         print_sched_time(tr->dt_wait, 6);
1884         print_sched_time(tr->dt_delay, 6);
1885         print_sched_time(tr->dt_run, 6);
1886
1887         if (sched->show_wakeups)
1888                 printf("  %-*s", comm_width, "");
1889
1890         if (thread->tid == 0)
1891                 goto out;
1892
1893         if (sched->show_callchain)
1894                 printf("  ");
1895
1896         sample__fprintf_sym(sample, al, 0,
1897                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1898                             EVSEL__PRINT_CALLCHAIN_ARROW |
1899                             EVSEL__PRINT_SKIP_IGNORED,
1900                             &callchain_cursor, stdout);
1901
1902 out:
1903         printf("\n");
1904 }
1905
1906 /*
1907  * Explanation of delta-time stats:
1908  *
1909  *            t = time of current schedule out event
1910  *        tprev = time of previous sched out event
1911  *                also time of schedule-in event for current task
1912  *    last_time = time of last sched change event for current task
1913  *                (i.e, time process was last scheduled out)
1914  * ready_to_run = time of wakeup for current task
1915  *
1916  * -----|------------|------------|------------|------
1917  *    last         ready        tprev          t
1918  *    time         to run
1919  *
1920  *      |-------- dt_wait --------|
1921  *                   |- dt_delay -|-- dt_run --|
1922  *
1923  *   dt_run = run time of current task
1924  *  dt_wait = time between last schedule out event for task and tprev
1925  *            represents time spent off the cpu
1926  * dt_delay = time between wakeup and schedule-in of task
1927  */
1928
1929 static void timehist_update_runtime_stats(struct thread_runtime *r,
1930                                          u64 t, u64 tprev)
1931 {
1932         r->dt_delay   = 0;
1933         r->dt_wait    = 0;
1934         r->dt_run     = 0;
1935         if (tprev) {
1936                 r->dt_run = t - tprev;
1937                 if (r->ready_to_run) {
1938                         if (r->ready_to_run > tprev)
1939                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
1940                         else
1941                                 r->dt_delay = tprev - r->ready_to_run;
1942                 }
1943
1944                 if (r->last_time > tprev)
1945                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
1946                 else if (r->last_time)
1947                         r->dt_wait = tprev - r->last_time;
1948         }
1949
1950         update_stats(&r->run_stats, r->dt_run);
1951         r->total_run_time += r->dt_run;
1952 }
1953
1954 static bool is_idle_sample(struct perf_sample *sample,
1955                            struct perf_evsel *evsel)
1956 {
1957         /* pid 0 == swapper == idle task */
1958         if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
1959                 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
1960
1961         return sample->pid == 0;
1962 }
1963
1964 static void save_task_callchain(struct perf_sched *sched,
1965                                 struct perf_sample *sample,
1966                                 struct perf_evsel *evsel,
1967                                 struct machine *machine)
1968 {
1969         struct callchain_cursor *cursor = &callchain_cursor;
1970         struct thread *thread;
1971
1972         /* want main thread for process - has maps */
1973         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
1974         if (thread == NULL) {
1975                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
1976                 return;
1977         }
1978
1979         if (!symbol_conf.use_callchain || sample->callchain == NULL)
1980                 return;
1981
1982         if (thread__resolve_callchain(thread, cursor, evsel, sample,
1983                                       NULL, NULL, sched->max_stack + 2) != 0) {
1984                 if (verbose)
1985                         error("Failed to resolve callchain. Skipping\n");
1986
1987                 return;
1988         }
1989
1990         callchain_cursor_commit(cursor);
1991
1992         while (true) {
1993                 struct callchain_cursor_node *node;
1994                 struct symbol *sym;
1995
1996                 node = callchain_cursor_current(cursor);
1997                 if (node == NULL)
1998                         break;
1999
2000                 sym = node->sym;
2001                 if (sym && sym->name) {
2002                         if (!strcmp(sym->name, "schedule") ||
2003                             !strcmp(sym->name, "__schedule") ||
2004                             !strcmp(sym->name, "preempt_schedule"))
2005                                 sym->ignore = 1;
2006                 }
2007
2008                 callchain_cursor_advance(cursor);
2009         }
2010 }
2011
2012 static int init_idle_thread(struct thread *thread)
2013 {
2014         struct idle_thread_runtime *itr;
2015
2016         thread__set_comm(thread, idle_comm, 0);
2017
2018         itr = zalloc(sizeof(*itr));
2019         if (itr == NULL)
2020                 return -ENOMEM;
2021
2022         init_stats(&itr->tr.run_stats);
2023         callchain_init(&itr->callchain);
2024         callchain_cursor_reset(&itr->cursor);
2025         thread__set_priv(thread, itr);
2026
2027         return 0;
2028 }
2029
2030 /*
2031  * Track idle stats per cpu by maintaining a local thread
2032  * struct for the idle task on each cpu.
2033  */
2034 static int init_idle_threads(int ncpu)
2035 {
2036         int i, ret;
2037
2038         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2039         if (!idle_threads)
2040                 return -ENOMEM;
2041
2042         idle_max_cpu = ncpu;
2043
2044         /* allocate the actual thread struct if needed */
2045         for (i = 0; i < ncpu; ++i) {
2046                 idle_threads[i] = thread__new(0, 0);
2047                 if (idle_threads[i] == NULL)
2048                         return -ENOMEM;
2049
2050                 ret = init_idle_thread(idle_threads[i]);
2051                 if (ret < 0)
2052                         return ret;
2053         }
2054
2055         return 0;
2056 }
2057
2058 static void free_idle_threads(void)
2059 {
2060         int i;
2061
2062         if (idle_threads == NULL)
2063                 return;
2064
2065         for (i = 0; i < idle_max_cpu; ++i) {
2066                 if ((idle_threads[i]))
2067                         thread__delete(idle_threads[i]);
2068         }
2069
2070         free(idle_threads);
2071 }
2072
2073 static struct thread *get_idle_thread(int cpu)
2074 {
2075         /*
2076          * expand/allocate array of pointers to local thread
2077          * structs if needed
2078          */
2079         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2080                 int i, j = __roundup_pow_of_two(cpu+1);
2081                 void *p;
2082
2083                 p = realloc(idle_threads, j * sizeof(struct thread *));
2084                 if (!p)
2085                         return NULL;
2086
2087                 idle_threads = (struct thread **) p;
2088                 for (i = idle_max_cpu; i < j; ++i)
2089                         idle_threads[i] = NULL;
2090
2091                 idle_max_cpu = j;
2092         }
2093
2094         /* allocate a new thread struct if needed */
2095         if (idle_threads[cpu] == NULL) {
2096                 idle_threads[cpu] = thread__new(0, 0);
2097
2098                 if (idle_threads[cpu]) {
2099                         if (init_idle_thread(idle_threads[cpu]) < 0)
2100                                 return NULL;
2101                 }
2102         }
2103
2104         return idle_threads[cpu];
2105 }
2106
2107 static void save_idle_callchain(struct idle_thread_runtime *itr,
2108                                 struct perf_sample *sample)
2109 {
2110         if (!symbol_conf.use_callchain || sample->callchain == NULL)
2111                 return;
2112
2113         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2114 }
2115
2116 /*
2117  * handle runtime stats saved per thread
2118  */
2119 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2120 {
2121         struct thread_runtime *r;
2122
2123         r = zalloc(sizeof(struct thread_runtime));
2124         if (!r)
2125                 return NULL;
2126
2127         init_stats(&r->run_stats);
2128         thread__set_priv(thread, r);
2129
2130         return r;
2131 }
2132
2133 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2134 {
2135         struct thread_runtime *tr;
2136
2137         tr = thread__priv(thread);
2138         if (tr == NULL) {
2139                 tr = thread__init_runtime(thread);
2140                 if (tr == NULL)
2141                         pr_debug("Failed to malloc memory for runtime data.\n");
2142         }
2143
2144         return tr;
2145 }
2146
2147 static struct thread *timehist_get_thread(struct perf_sched *sched,
2148                                           struct perf_sample *sample,
2149                                           struct machine *machine,
2150                                           struct perf_evsel *evsel)
2151 {
2152         struct thread *thread;
2153
2154         if (is_idle_sample(sample, evsel)) {
2155                 thread = get_idle_thread(sample->cpu);
2156                 if (thread == NULL)
2157                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2158
2159         } else {
2160                 /* there were samples with tid 0 but non-zero pid */
2161                 thread = machine__findnew_thread(machine, sample->pid,
2162                                                  sample->tid ?: sample->pid);
2163                 if (thread == NULL) {
2164                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2165                                  sample->tid);
2166                 }
2167
2168                 save_task_callchain(sched, sample, evsel, machine);
2169                 if (sched->idle_hist) {
2170                         struct thread *idle;
2171                         struct idle_thread_runtime *itr;
2172
2173                         idle = get_idle_thread(sample->cpu);
2174                         if (idle == NULL) {
2175                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2176                                 return NULL;
2177                         }
2178
2179                         itr = thread__priv(idle);
2180                         if (itr == NULL)
2181                                 return NULL;
2182
2183                         itr->last_thread = thread;
2184
2185                         /* copy task callchain when entering to idle */
2186                         if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2187                                 save_idle_callchain(itr, sample);
2188                 }
2189         }
2190
2191         return thread;
2192 }
2193
2194 static bool timehist_skip_sample(struct perf_sched *sched,
2195                                  struct thread *thread,
2196                                  struct perf_evsel *evsel,
2197                                  struct perf_sample *sample)
2198 {
2199         bool rc = false;
2200
2201         if (thread__is_filtered(thread)) {
2202                 rc = true;
2203                 sched->skipped_samples++;
2204         }
2205
2206         if (sched->idle_hist) {
2207                 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2208                         rc = true;
2209                 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2210                          perf_evsel__intval(evsel, sample, "next_pid") != 0)
2211                         rc = true;
2212         }
2213
2214         return rc;
2215 }
2216
2217 static void timehist_print_wakeup_event(struct perf_sched *sched,
2218                                         struct perf_evsel *evsel,
2219                                         struct perf_sample *sample,
2220                                         struct machine *machine,
2221                                         struct thread *awakened)
2222 {
2223         struct thread *thread;
2224         char tstr[64];
2225
2226         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2227         if (thread == NULL)
2228                 return;
2229
2230         /* show wakeup unless both awakee and awaker are filtered */
2231         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2232             timehist_skip_sample(sched, awakened, evsel, sample)) {
2233                 return;
2234         }
2235
2236         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2237         printf("%15s [%04d] ", tstr, sample->cpu);
2238         if (sched->show_cpu_visual)
2239                 printf(" %*s ", sched->max_cpu + 1, "");
2240
2241         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2242
2243         /* dt spacer */
2244         printf("  %9s  %9s  %9s ", "", "", "");
2245
2246         printf("awakened: %s", timehist_get_commstr(awakened));
2247
2248         printf("\n");
2249 }
2250
2251 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2252                                        union perf_event *event __maybe_unused,
2253                                        struct perf_evsel *evsel,
2254                                        struct perf_sample *sample,
2255                                        struct machine *machine)
2256 {
2257         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2258         struct thread *thread;
2259         struct thread_runtime *tr = NULL;
2260         /* want pid of awakened task not pid in sample */
2261         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2262
2263         thread = machine__findnew_thread(machine, 0, pid);
2264         if (thread == NULL)
2265                 return -1;
2266
2267         tr = thread__get_runtime(thread);
2268         if (tr == NULL)
2269                 return -1;
2270
2271         if (tr->ready_to_run == 0)
2272                 tr->ready_to_run = sample->time;
2273
2274         /* show wakeups if requested */
2275         if (sched->show_wakeups &&
2276             !perf_time__skip_sample(&sched->ptime, sample->time))
2277                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2278
2279         return 0;
2280 }
2281
2282 static void timehist_print_migration_event(struct perf_sched *sched,
2283                                         struct perf_evsel *evsel,
2284                                         struct perf_sample *sample,
2285                                         struct machine *machine,
2286                                         struct thread *migrated)
2287 {
2288         struct thread *thread;
2289         char tstr[64];
2290         u32 max_cpus = sched->max_cpu + 1;
2291         u32 ocpu, dcpu;
2292
2293         if (sched->summary_only)
2294                 return;
2295
2296         max_cpus = sched->max_cpu + 1;
2297         ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2298         dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2299
2300         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2301         if (thread == NULL)
2302                 return;
2303
2304         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2305             timehist_skip_sample(sched, migrated, evsel, sample)) {
2306                 return;
2307         }
2308
2309         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2310         printf("%15s [%04d] ", tstr, sample->cpu);
2311
2312         if (sched->show_cpu_visual) {
2313                 u32 i;
2314                 char c;
2315
2316                 printf("  ");
2317                 for (i = 0; i < max_cpus; ++i) {
2318                         c = (i == sample->cpu) ? 'm' : ' ';
2319                         printf("%c", c);
2320                 }
2321                 printf("  ");
2322         }
2323
2324         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2325
2326         /* dt spacer */
2327         printf("  %9s  %9s  %9s ", "", "", "");
2328
2329         printf("migrated: %s", timehist_get_commstr(migrated));
2330         printf(" cpu %d => %d", ocpu, dcpu);
2331
2332         printf("\n");
2333 }
2334
2335 static int timehist_migrate_task_event(struct perf_tool *tool,
2336                                        union perf_event *event __maybe_unused,
2337                                        struct perf_evsel *evsel,
2338                                        struct perf_sample *sample,
2339                                        struct machine *machine)
2340 {
2341         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2342         struct thread *thread;
2343         struct thread_runtime *tr = NULL;
2344         /* want pid of migrated task not pid in sample */
2345         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2346
2347         thread = machine__findnew_thread(machine, 0, pid);
2348         if (thread == NULL)
2349                 return -1;
2350
2351         tr = thread__get_runtime(thread);
2352         if (tr == NULL)
2353                 return -1;
2354
2355         tr->migrations++;
2356
2357         /* show migrations if requested */
2358         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2359
2360         return 0;
2361 }
2362
2363 static int timehist_sched_change_event(struct perf_tool *tool,
2364                                        union perf_event *event,
2365                                        struct perf_evsel *evsel,
2366                                        struct perf_sample *sample,
2367                                        struct machine *machine)
2368 {
2369         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2370         struct perf_time_interval *ptime = &sched->ptime;
2371         struct addr_location al;
2372         struct thread *thread;
2373         struct thread_runtime *tr = NULL;
2374         u64 tprev, t = sample->time;
2375         int rc = 0;
2376
2377         if (machine__resolve(machine, &al, sample) < 0) {
2378                 pr_err("problem processing %d event. skipping it\n",
2379                        event->header.type);
2380                 rc = -1;
2381                 goto out;
2382         }
2383
2384         thread = timehist_get_thread(sched, sample, machine, evsel);
2385         if (thread == NULL) {
2386                 rc = -1;
2387                 goto out;
2388         }
2389
2390         if (timehist_skip_sample(sched, thread, evsel, sample))
2391                 goto out;
2392
2393         tr = thread__get_runtime(thread);
2394         if (tr == NULL) {
2395                 rc = -1;
2396                 goto out;
2397         }
2398
2399         tprev = perf_evsel__get_time(evsel, sample->cpu);
2400
2401         /*
2402          * If start time given:
2403          * - sample time is under window user cares about - skip sample
2404          * - tprev is under window user cares about  - reset to start of window
2405          */
2406         if (ptime->start && ptime->start > t)
2407                 goto out;
2408
2409         if (tprev && ptime->start > tprev)
2410                 tprev = ptime->start;
2411
2412         /*
2413          * If end time given:
2414          * - previous sched event is out of window - we are done
2415          * - sample time is beyond window user cares about - reset it
2416          *   to close out stats for time window interest
2417          */
2418         if (ptime->end) {
2419                 if (tprev > ptime->end)
2420                         goto out;
2421
2422                 if (t > ptime->end)
2423                         t = ptime->end;
2424         }
2425
2426         if (!sched->idle_hist || thread->tid == 0) {
2427                 timehist_update_runtime_stats(tr, t, tprev);
2428
2429                 if (sched->idle_hist) {
2430                         struct idle_thread_runtime *itr = (void *)tr;
2431                         struct thread_runtime *last_tr;
2432
2433                         BUG_ON(thread->tid != 0);
2434
2435                         if (itr->last_thread == NULL)
2436                                 goto out;
2437
2438                         /* add current idle time as last thread's runtime */
2439                         last_tr = thread__get_runtime(itr->last_thread);
2440                         if (last_tr == NULL)
2441                                 goto out;
2442
2443                         timehist_update_runtime_stats(last_tr, t, tprev);
2444                         /*
2445                          * remove delta time of last thread as it's not updated
2446                          * and otherwise it will show an invalid value next
2447                          * time.  we only care total run time and run stat.
2448                          */
2449                         last_tr->dt_run = 0;
2450                         last_tr->dt_wait = 0;
2451                         last_tr->dt_delay = 0;
2452
2453                         if (itr->cursor.nr)
2454                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2455
2456                         itr->last_thread = NULL;
2457                 }
2458         }
2459
2460         if (!sched->summary_only)
2461                 timehist_print_sample(sched, sample, &al, thread, t);
2462
2463 out:
2464         if (sched->hist_time.start == 0 && t >= ptime->start)
2465                 sched->hist_time.start = t;
2466         if (ptime->end == 0 || t <= ptime->end)
2467                 sched->hist_time.end = t;
2468
2469         if (tr) {
2470                 /* time of this sched_switch event becomes last time task seen */
2471                 tr->last_time = sample->time;
2472
2473                 /* sched out event for task so reset ready to run time */
2474                 tr->ready_to_run = 0;
2475         }
2476
2477         perf_evsel__save_time(evsel, sample->time, sample->cpu);
2478
2479         return rc;
2480 }
2481
2482 static int timehist_sched_switch_event(struct perf_tool *tool,
2483                              union perf_event *event,
2484                              struct perf_evsel *evsel,
2485                              struct perf_sample *sample,
2486                              struct machine *machine __maybe_unused)
2487 {
2488         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2489 }
2490
2491 static int process_lost(struct perf_tool *tool __maybe_unused,
2492                         union perf_event *event,
2493                         struct perf_sample *sample,
2494                         struct machine *machine __maybe_unused)
2495 {
2496         char tstr[64];
2497
2498         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2499         printf("%15s ", tstr);
2500         printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2501
2502         return 0;
2503 }
2504
2505
2506 static void print_thread_runtime(struct thread *t,
2507                                  struct thread_runtime *r)
2508 {
2509         double mean = avg_stats(&r->run_stats);
2510         float stddev;
2511
2512         printf("%*s   %5d  %9" PRIu64 " ",
2513                comm_width, timehist_get_commstr(t), t->ppid,
2514                (u64) r->run_stats.n);
2515
2516         print_sched_time(r->total_run_time, 8);
2517         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2518         print_sched_time(r->run_stats.min, 6);
2519         printf(" ");
2520         print_sched_time((u64) mean, 6);
2521         printf(" ");
2522         print_sched_time(r->run_stats.max, 6);
2523         printf("  ");
2524         printf("%5.2f", stddev);
2525         printf("   %5" PRIu64, r->migrations);
2526         printf("\n");
2527 }
2528
2529 struct total_run_stats {
2530         u64  sched_count;
2531         u64  task_count;
2532         u64  total_run_time;
2533 };
2534
2535 static int __show_thread_runtime(struct thread *t, void *priv)
2536 {
2537         struct total_run_stats *stats = priv;
2538         struct thread_runtime *r;
2539
2540         if (thread__is_filtered(t))
2541                 return 0;
2542
2543         r = thread__priv(t);
2544         if (r && r->run_stats.n) {
2545                 stats->task_count++;
2546                 stats->sched_count += r->run_stats.n;
2547                 stats->total_run_time += r->total_run_time;
2548                 print_thread_runtime(t, r);
2549         }
2550
2551         return 0;
2552 }
2553
2554 static int show_thread_runtime(struct thread *t, void *priv)
2555 {
2556         if (t->dead)
2557                 return 0;
2558
2559         return __show_thread_runtime(t, priv);
2560 }
2561
2562 static int show_deadthread_runtime(struct thread *t, void *priv)
2563 {
2564         if (!t->dead)
2565                 return 0;
2566
2567         return __show_thread_runtime(t, priv);
2568 }
2569
2570 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2571 {
2572         const char *sep = " <- ";
2573         struct callchain_list *chain;
2574         size_t ret = 0;
2575         char bf[1024];
2576         bool first;
2577
2578         if (node == NULL)
2579                 return 0;
2580
2581         ret = callchain__fprintf_folded(fp, node->parent);
2582         first = (ret == 0);
2583
2584         list_for_each_entry(chain, &node->val, list) {
2585                 if (chain->ip >= PERF_CONTEXT_MAX)
2586                         continue;
2587                 if (chain->ms.sym && chain->ms.sym->ignore)
2588                         continue;
2589                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2590                                callchain_list__sym_name(chain, bf, sizeof(bf),
2591                                                         false));
2592                 first = false;
2593         }
2594
2595         return ret;
2596 }
2597
2598 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2599 {
2600         size_t ret = 0;
2601         FILE *fp = stdout;
2602         struct callchain_node *chain;
2603         struct rb_node *rb_node = rb_first(root);
2604
2605         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2606         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2607                graph_dotted_line);
2608
2609         while (rb_node) {
2610                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2611                 rb_node = rb_next(rb_node);
2612
2613                 ret += fprintf(fp, "  ");
2614                 print_sched_time(chain->hit, 12);
2615                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2616                 ret += fprintf(fp, " %8d  ", chain->count);
2617                 ret += callchain__fprintf_folded(fp, chain);
2618                 ret += fprintf(fp, "\n");
2619         }
2620
2621         return ret;
2622 }
2623
2624 static void timehist_print_summary(struct perf_sched *sched,
2625                                    struct perf_session *session)
2626 {
2627         struct machine *m = &session->machines.host;
2628         struct total_run_stats totals;
2629         u64 task_count;
2630         struct thread *t;
2631         struct thread_runtime *r;
2632         int i;
2633         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2634
2635         memset(&totals, 0, sizeof(totals));
2636
2637         if (sched->idle_hist) {
2638                 printf("\nIdle-time summary\n");
2639                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2640                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2641         } else {
2642                 printf("\nRuntime summary\n");
2643                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2644                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2645         }
2646         printf("%*s            (count)  ", comm_width, "");
2647         printf("     (msec)     (msec)      (msec)      (msec)       %%\n");
2648         printf("%.117s\n", graph_dotted_line);
2649
2650         machine__for_each_thread(m, show_thread_runtime, &totals);
2651         task_count = totals.task_count;
2652         if (!task_count)
2653                 printf("<no still running tasks>\n");
2654
2655         printf("\nTerminated tasks:\n");
2656         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2657         if (task_count == totals.task_count)
2658                 printf("<no terminated tasks>\n");
2659
2660         /* CPU idle stats not tracked when samples were skipped */
2661         if (sched->skipped_samples && !sched->idle_hist)
2662                 return;
2663
2664         printf("\nIdle stats:\n");
2665         for (i = 0; i < idle_max_cpu; ++i) {
2666                 t = idle_threads[i];
2667                 if (!t)
2668                         continue;
2669
2670                 r = thread__priv(t);
2671                 if (r && r->run_stats.n) {
2672                         totals.sched_count += r->run_stats.n;
2673                         printf("    CPU %2d idle for ", i);
2674                         print_sched_time(r->total_run_time, 6);
2675                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2676                 } else
2677                         printf("    CPU %2d idle entire time window\n", i);
2678         }
2679
2680         if (sched->idle_hist && symbol_conf.use_callchain) {
2681                 callchain_param.mode  = CHAIN_FOLDED;
2682                 callchain_param.value = CCVAL_PERIOD;
2683
2684                 callchain_register_param(&callchain_param);
2685
2686                 printf("\nIdle stats by callchain:\n");
2687                 for (i = 0; i < idle_max_cpu; ++i) {
2688                         struct idle_thread_runtime *itr;
2689
2690                         t = idle_threads[i];
2691                         if (!t)
2692                                 continue;
2693
2694                         itr = thread__priv(t);
2695                         if (itr == NULL)
2696                                 continue;
2697
2698                         callchain_param.sort(&itr->sorted_root, &itr->callchain,
2699                                              0, &callchain_param);
2700
2701                         printf("  CPU %2d:", i);
2702                         print_sched_time(itr->tr.total_run_time, 6);
2703                         printf(" msec\n");
2704                         timehist_print_idlehist_callchain(&itr->sorted_root);
2705                         printf("\n");
2706                 }
2707         }
2708
2709         printf("\n"
2710                "    Total number of unique tasks: %" PRIu64 "\n"
2711                "Total number of context switches: %" PRIu64 "\n",
2712                totals.task_count, totals.sched_count);
2713
2714         printf("           Total run time (msec): ");
2715         print_sched_time(totals.total_run_time, 2);
2716         printf("\n");
2717
2718         printf("    Total scheduling time (msec): ");
2719         print_sched_time(hist_time, 2);
2720         printf(" (x %d)\n", sched->max_cpu);
2721 }
2722
2723 typedef int (*sched_handler)(struct perf_tool *tool,
2724                           union perf_event *event,
2725                           struct perf_evsel *evsel,
2726                           struct perf_sample *sample,
2727                           struct machine *machine);
2728
2729 static int perf_timehist__process_sample(struct perf_tool *tool,
2730                                          union perf_event *event,
2731                                          struct perf_sample *sample,
2732                                          struct perf_evsel *evsel,
2733                                          struct machine *machine)
2734 {
2735         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2736         int err = 0;
2737         int this_cpu = sample->cpu;
2738
2739         if (this_cpu > sched->max_cpu)
2740                 sched->max_cpu = this_cpu;
2741
2742         if (evsel->handler != NULL) {
2743                 sched_handler f = evsel->handler;
2744
2745                 err = f(tool, event, evsel, sample, machine);
2746         }
2747
2748         return err;
2749 }
2750
2751 static int timehist_check_attr(struct perf_sched *sched,
2752                                struct perf_evlist *evlist)
2753 {
2754         struct perf_evsel *evsel;
2755         struct evsel_runtime *er;
2756
2757         list_for_each_entry(evsel, &evlist->entries, node) {
2758                 er = perf_evsel__get_runtime(evsel);
2759                 if (er == NULL) {
2760                         pr_err("Failed to allocate memory for evsel runtime data\n");
2761                         return -1;
2762                 }
2763
2764                 if (sched->show_callchain &&
2765                     !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2766                         pr_info("Samples do not have callchains.\n");
2767                         sched->show_callchain = 0;
2768                         symbol_conf.use_callchain = 0;
2769                 }
2770         }
2771
2772         return 0;
2773 }
2774
2775 static int perf_sched__timehist(struct perf_sched *sched)
2776 {
2777         const struct perf_evsel_str_handler handlers[] = {
2778                 { "sched:sched_switch",       timehist_sched_switch_event, },
2779                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2780                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2781         };
2782         const struct perf_evsel_str_handler migrate_handlers[] = {
2783                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2784         };
2785         struct perf_data_file file = {
2786                 .path = input_name,
2787                 .mode = PERF_DATA_MODE_READ,
2788                 .force = sched->force,
2789         };
2790
2791         struct perf_session *session;
2792         struct perf_evlist *evlist;
2793         int err = -1;
2794
2795         /*
2796          * event handlers for timehist option
2797          */
2798         sched->tool.sample       = perf_timehist__process_sample;
2799         sched->tool.mmap         = perf_event__process_mmap;
2800         sched->tool.comm         = perf_event__process_comm;
2801         sched->tool.exit         = perf_event__process_exit;
2802         sched->tool.fork         = perf_event__process_fork;
2803         sched->tool.lost         = process_lost;
2804         sched->tool.attr         = perf_event__process_attr;
2805         sched->tool.tracing_data = perf_event__process_tracing_data;
2806         sched->tool.build_id     = perf_event__process_build_id;
2807
2808         sched->tool.ordered_events = true;
2809         sched->tool.ordering_requires_timestamps = true;
2810
2811         symbol_conf.use_callchain = sched->show_callchain;
2812
2813         session = perf_session__new(&file, false, &sched->tool);
2814         if (session == NULL)
2815                 return -ENOMEM;
2816
2817         evlist = session->evlist;
2818
2819         symbol__init(&session->header.env);
2820
2821         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2822                 pr_err("Invalid time string\n");
2823                 return -EINVAL;
2824         }
2825
2826         if (timehist_check_attr(sched, evlist) != 0)
2827                 goto out;
2828
2829         setup_pager();
2830
2831         /* setup per-evsel handlers */
2832         if (perf_session__set_tracepoints_handlers(session, handlers))
2833                 goto out;
2834
2835         /* sched_switch event at a minimum needs to exist */
2836         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2837                                                   "sched:sched_switch")) {
2838                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2839                 goto out;
2840         }
2841
2842         if (sched->show_migrations &&
2843             perf_session__set_tracepoints_handlers(session, migrate_handlers))
2844                 goto out;
2845
2846         /* pre-allocate struct for per-CPU idle stats */
2847         sched->max_cpu = session->header.env.nr_cpus_online;
2848         if (sched->max_cpu == 0)
2849                 sched->max_cpu = 4;
2850         if (init_idle_threads(sched->max_cpu))
2851                 goto out;
2852
2853         /* summary_only implies summary option, but don't overwrite summary if set */
2854         if (sched->summary_only)
2855                 sched->summary = sched->summary_only;
2856
2857         if (!sched->summary_only)
2858                 timehist_header(sched);
2859
2860         err = perf_session__process_events(session);
2861         if (err) {
2862                 pr_err("Failed to process events, error %d", err);
2863                 goto out;
2864         }
2865
2866         sched->nr_events      = evlist->stats.nr_events[0];
2867         sched->nr_lost_events = evlist->stats.total_lost;
2868         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2869
2870         if (sched->summary)
2871                 timehist_print_summary(sched, session);
2872
2873 out:
2874         free_idle_threads();
2875         perf_session__delete(session);
2876
2877         return err;
2878 }
2879
2880
2881 static void print_bad_events(struct perf_sched *sched)
2882 {
2883         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
2884                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
2885                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
2886                         sched->nr_unordered_timestamps, sched->nr_timestamps);
2887         }
2888         if (sched->nr_lost_events && sched->nr_events) {
2889                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
2890                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
2891                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
2892         }
2893         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
2894                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
2895                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
2896                         sched->nr_context_switch_bugs, sched->nr_timestamps);
2897                 if (sched->nr_lost_events)
2898                         printf(" (due to lost events?)");
2899                 printf("\n");
2900         }
2901 }
2902
2903 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
2904 {
2905         struct rb_node **new = &(root->rb_node), *parent = NULL;
2906         struct work_atoms *this;
2907         const char *comm = thread__comm_str(data->thread), *this_comm;
2908
2909         while (*new) {
2910                 int cmp;
2911
2912                 this = container_of(*new, struct work_atoms, node);
2913                 parent = *new;
2914
2915                 this_comm = thread__comm_str(this->thread);
2916                 cmp = strcmp(comm, this_comm);
2917                 if (cmp > 0) {
2918                         new = &((*new)->rb_left);
2919                 } else if (cmp < 0) {
2920                         new = &((*new)->rb_right);
2921                 } else {
2922                         this->num_merged++;
2923                         this->total_runtime += data->total_runtime;
2924                         this->nb_atoms += data->nb_atoms;
2925                         this->total_lat += data->total_lat;
2926                         list_splice(&data->work_list, &this->work_list);
2927                         if (this->max_lat < data->max_lat) {
2928                                 this->max_lat = data->max_lat;
2929                                 this->max_lat_at = data->max_lat_at;
2930                         }
2931                         zfree(&data);
2932                         return;
2933                 }
2934         }
2935
2936         data->num_merged++;
2937         rb_link_node(&data->node, parent, new);
2938         rb_insert_color(&data->node, root);
2939 }
2940
2941 static void perf_sched__merge_lat(struct perf_sched *sched)
2942 {
2943         struct work_atoms *data;
2944         struct rb_node *node;
2945
2946         if (sched->skip_merge)
2947                 return;
2948
2949         while ((node = rb_first(&sched->atom_root))) {
2950                 rb_erase(node, &sched->atom_root);
2951                 data = rb_entry(node, struct work_atoms, node);
2952                 __merge_work_atoms(&sched->merged_atom_root, data);
2953         }
2954 }
2955
2956 static int perf_sched__lat(struct perf_sched *sched)
2957 {
2958         struct rb_node *next;
2959
2960         setup_pager();
2961
2962         if (perf_sched__read_events(sched))
2963                 return -1;
2964
2965         perf_sched__merge_lat(sched);
2966         perf_sched__sort_lat(sched);
2967
2968         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
2969         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
2970         printf(" -----------------------------------------------------------------------------------------------------------------\n");
2971
2972         next = rb_first(&sched->sorted_atom_root);
2973
2974         while (next) {
2975                 struct work_atoms *work_list;
2976
2977                 work_list = rb_entry(next, struct work_atoms, node);
2978                 output_lat_thread(sched, work_list);
2979                 next = rb_next(next);
2980                 thread__zput(work_list->thread);
2981         }
2982
2983         printf(" -----------------------------------------------------------------------------------------------------------------\n");
2984         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
2985                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
2986
2987         printf(" ---------------------------------------------------\n");
2988
2989         print_bad_events(sched);
2990         printf("\n");
2991
2992         return 0;
2993 }
2994
2995 static int setup_map_cpus(struct perf_sched *sched)
2996 {
2997         struct cpu_map *map;
2998
2999         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3000
3001         if (sched->map.comp) {
3002                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3003                 if (!sched->map.comp_cpus)
3004                         return -1;
3005         }
3006
3007         if (!sched->map.cpus_str)
3008                 return 0;
3009
3010         map = cpu_map__new(sched->map.cpus_str);
3011         if (!map) {
3012                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3013                 return -1;
3014         }
3015
3016         sched->map.cpus = map;
3017         return 0;
3018 }
3019
3020 static int setup_color_pids(struct perf_sched *sched)
3021 {
3022         struct thread_map *map;
3023
3024         if (!sched->map.color_pids_str)
3025                 return 0;
3026
3027         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3028         if (!map) {
3029                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3030                 return -1;
3031         }
3032
3033         sched->map.color_pids = map;
3034         return 0;
3035 }
3036
3037 static int setup_color_cpus(struct perf_sched *sched)
3038 {
3039         struct cpu_map *map;
3040
3041         if (!sched->map.color_cpus_str)
3042                 return 0;
3043
3044         map = cpu_map__new(sched->map.color_cpus_str);
3045         if (!map) {
3046                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3047                 return -1;
3048         }
3049
3050         sched->map.color_cpus = map;
3051         return 0;
3052 }
3053
3054 static int perf_sched__map(struct perf_sched *sched)
3055 {
3056         if (setup_map_cpus(sched))
3057                 return -1;
3058
3059         if (setup_color_pids(sched))
3060                 return -1;
3061
3062         if (setup_color_cpus(sched))
3063                 return -1;
3064
3065         setup_pager();
3066         if (perf_sched__read_events(sched))
3067                 return -1;
3068         print_bad_events(sched);
3069         return 0;
3070 }
3071
3072 static int perf_sched__replay(struct perf_sched *sched)
3073 {
3074         unsigned long i;
3075
3076         calibrate_run_measurement_overhead(sched);
3077         calibrate_sleep_measurement_overhead(sched);
3078
3079         test_calibrations(sched);
3080
3081         if (perf_sched__read_events(sched))
3082                 return -1;
3083
3084         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3085         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3086         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3087
3088         if (sched->targetless_wakeups)
3089                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3090         if (sched->multitarget_wakeups)
3091                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3092         if (sched->nr_run_events_optimized)
3093                 printf("run atoms optimized: %ld\n",
3094                         sched->nr_run_events_optimized);
3095
3096         print_task_traces(sched);
3097         add_cross_task_wakeups(sched);
3098
3099         create_tasks(sched);
3100         printf("------------------------------------------------------------\n");
3101         for (i = 0; i < sched->replay_repeat; i++)
3102                 run_one_test(sched);
3103
3104         return 0;
3105 }
3106
3107 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3108                           const char * const usage_msg[])
3109 {
3110         char *tmp, *tok, *str = strdup(sched->sort_order);
3111
3112         for (tok = strtok_r(str, ", ", &tmp);
3113                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3114                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3115                         usage_with_options_msg(usage_msg, options,
3116                                         "Unknown --sort key: `%s'", tok);
3117                 }
3118         }
3119
3120         free(str);
3121
3122         sort_dimension__add("pid", &sched->cmp_pid);
3123 }
3124
3125 static int __cmd_record(int argc, const char **argv)
3126 {
3127         unsigned int rec_argc, i, j;
3128         const char **rec_argv;
3129         const char * const record_args[] = {
3130                 "record",
3131                 "-a",
3132                 "-R",
3133                 "-m", "1024",
3134                 "-c", "1",
3135                 "-e", "sched:sched_switch",
3136                 "-e", "sched:sched_stat_wait",
3137                 "-e", "sched:sched_stat_sleep",
3138                 "-e", "sched:sched_stat_iowait",
3139                 "-e", "sched:sched_stat_runtime",
3140                 "-e", "sched:sched_process_fork",
3141                 "-e", "sched:sched_wakeup",
3142                 "-e", "sched:sched_wakeup_new",
3143                 "-e", "sched:sched_migrate_task",
3144         };
3145
3146         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3147         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3148
3149         if (rec_argv == NULL)
3150                 return -ENOMEM;
3151
3152         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3153                 rec_argv[i] = strdup(record_args[i]);
3154
3155         for (j = 1; j < (unsigned int)argc; j++, i++)
3156                 rec_argv[i] = argv[j];
3157
3158         BUG_ON(i != rec_argc);
3159
3160         return cmd_record(i, rec_argv, NULL);
3161 }
3162
3163 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
3164 {
3165         const char default_sort_order[] = "avg, max, switch, runtime";
3166         struct perf_sched sched = {
3167                 .tool = {
3168                         .sample          = perf_sched__process_tracepoint_sample,
3169                         .comm            = perf_event__process_comm,
3170                         .lost            = perf_event__process_lost,
3171                         .fork            = perf_sched__process_fork_event,
3172                         .ordered_events = true,
3173                 },
3174                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3175                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3176                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3177                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3178                 .sort_order           = default_sort_order,
3179                 .replay_repeat        = 10,
3180                 .profile_cpu          = -1,
3181                 .next_shortname1      = 'A',
3182                 .next_shortname2      = '0',
3183                 .skip_merge           = 0,
3184                 .show_callchain       = 1,
3185                 .max_stack            = 5,
3186         };
3187         const struct option sched_options[] = {
3188         OPT_STRING('i', "input", &input_name, "file",
3189                     "input file name"),
3190         OPT_INCR('v', "verbose", &verbose,
3191                     "be more verbose (show symbol address, etc)"),
3192         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3193                     "dump raw trace in ASCII"),
3194         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3195         OPT_END()
3196         };
3197         const struct option latency_options[] = {
3198         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3199                    "sort by key(s): runtime, switch, avg, max"),
3200         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3201                     "CPU to profile on"),
3202         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3203                     "latency stats per pid instead of per comm"),
3204         OPT_PARENT(sched_options)
3205         };
3206         const struct option replay_options[] = {
3207         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3208                      "repeat the workload replay N times (-1: infinite)"),
3209         OPT_PARENT(sched_options)
3210         };
3211         const struct option map_options[] = {
3212         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3213                     "map output in compact mode"),
3214         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3215                    "highlight given pids in map"),
3216         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3217                     "highlight given CPUs in map"),
3218         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3219                     "display given CPUs in map"),
3220         OPT_PARENT(sched_options)
3221         };
3222         const struct option timehist_options[] = {
3223         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3224                    "file", "vmlinux pathname"),
3225         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3226                    "file", "kallsyms pathname"),
3227         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3228                     "Display call chains if present (default on)"),
3229         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3230                    "Maximum number of functions to display backtrace."),
3231         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3232                     "Look for files with symbols relative to this directory"),
3233         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3234                     "Show only syscall summary with statistics"),
3235         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3236                     "Show all syscalls and summary with statistics"),
3237         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3238         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3239         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3240         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3241         OPT_STRING(0, "time", &sched.time_str, "str",
3242                    "Time span for analysis (start,stop)"),
3243         OPT_PARENT(sched_options)
3244         };
3245
3246         const char * const latency_usage[] = {
3247                 "perf sched latency [<options>]",
3248                 NULL
3249         };
3250         const char * const replay_usage[] = {
3251                 "perf sched replay [<options>]",
3252                 NULL
3253         };
3254         const char * const map_usage[] = {
3255                 "perf sched map [<options>]",
3256                 NULL
3257         };
3258         const char * const timehist_usage[] = {
3259                 "perf sched timehist [<options>]",
3260                 NULL
3261         };
3262         const char *const sched_subcommands[] = { "record", "latency", "map",
3263                                                   "replay", "script",
3264                                                   "timehist", NULL };
3265         const char *sched_usage[] = {
3266                 NULL,
3267                 NULL
3268         };
3269         struct trace_sched_handler lat_ops  = {
3270                 .wakeup_event       = latency_wakeup_event,
3271                 .switch_event       = latency_switch_event,
3272                 .runtime_event      = latency_runtime_event,
3273                 .migrate_task_event = latency_migrate_task_event,
3274         };
3275         struct trace_sched_handler map_ops  = {
3276                 .switch_event       = map_switch_event,
3277         };
3278         struct trace_sched_handler replay_ops  = {
3279                 .wakeup_event       = replay_wakeup_event,
3280                 .switch_event       = replay_switch_event,
3281                 .fork_event         = replay_fork_event,
3282         };
3283         unsigned int i;
3284
3285         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3286                 sched.curr_pid[i] = -1;
3287
3288         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3289                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3290         if (!argc)
3291                 usage_with_options(sched_usage, sched_options);
3292
3293         /*
3294          * Aliased to 'perf script' for now:
3295          */
3296         if (!strcmp(argv[0], "script"))
3297                 return cmd_script(argc, argv, prefix);
3298
3299         if (!strncmp(argv[0], "rec", 3)) {
3300                 return __cmd_record(argc, argv);
3301         } else if (!strncmp(argv[0], "lat", 3)) {
3302                 sched.tp_handler = &lat_ops;
3303                 if (argc > 1) {
3304                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3305                         if (argc)
3306                                 usage_with_options(latency_usage, latency_options);
3307                 }
3308                 setup_sorting(&sched, latency_options, latency_usage);
3309                 return perf_sched__lat(&sched);
3310         } else if (!strcmp(argv[0], "map")) {
3311                 if (argc) {
3312                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3313                         if (argc)
3314                                 usage_with_options(map_usage, map_options);
3315                 }
3316                 sched.tp_handler = &map_ops;
3317                 setup_sorting(&sched, latency_options, latency_usage);
3318                 return perf_sched__map(&sched);
3319         } else if (!strncmp(argv[0], "rep", 3)) {
3320                 sched.tp_handler = &replay_ops;
3321                 if (argc) {
3322                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3323                         if (argc)
3324                                 usage_with_options(replay_usage, replay_options);
3325                 }
3326                 return perf_sched__replay(&sched);
3327         } else if (!strcmp(argv[0], "timehist")) {
3328                 if (argc) {
3329                         argc = parse_options(argc, argv, timehist_options,
3330                                              timehist_usage, 0);
3331                         if (argc)
3332                                 usage_with_options(timehist_usage, timehist_options);
3333                 }
3334                 if (sched.show_wakeups && sched.summary_only) {
3335                         pr_err(" Error: -s and -w are mutually exclusive.\n");
3336                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3337                         parse_options_usage(NULL, timehist_options, "w", true);
3338                         return -EINVAL;
3339                 }
3340
3341                 return perf_sched__timehist(&sched);
3342         } else {
3343                 usage_with_options(sched_usage, sched_options);
3344         }
3345
3346         return 0;
3347 }