builtin-sched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf.h"
   4 #include "perf-sys.h"
   5
   6 #include "util/cpumap.h"
   7 #include "util/evlist.h"
   8 #include "util/evsel.h"
   9 #include "util/evsel_fprintf.h"
  10 #include "util/symbol.h"
  11 #include "util/thread.h"
  12 #include "util/header.h"
  13 #include "util/session.h"
  14 #include "util/tool.h"
  15 #include "util/cloexec.h"
  16 #include "util/thread_map.h"
  17 #include "util/color.h"
  18 #include "util/stat.h"
  19 #include "util/string2.h"
  20 #include "util/callchain.h"
  21 #include "util/time-utils.h"
  22
  23 #include <subcmd/pager.h>
  24 #include <subcmd/parse-options.h>
  25 #include "util/trace-event.h"
  26
  27 #include "util/debug.h"
  28 #include "util/event.h"
  29
  30 #include <linux/kernel.h>
  31 #include <linux/log2.h>
  32 #include <linux/zalloc.h>
  33 #include <sys/prctl.h>
  34 #include <sys/resource.h>
  35 #include <inttypes.h>
  36
  37 #include <errno.h>
  38 #include <semaphore.h>
  39 #include <pthread.h>
  40 #include <math.h>
  41 #include <api/fs/fs.h>
  42 #include <perf/cpumap.h>
  43 #include <linux/time64.h>
  44 #include <linux/err.h>
  45
  46 #include <linux/ctype.h>
  47
  48 #define PR_SET_NAME             15               /* Set process name */
  49 #define MAX_CPUS                4096
  50 #define COMM_LEN                20
  51 #define SYM_LEN                 129
  52 #define MAX_PID                 1024000
  53
  54 static const char *cpu_list;
  55 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  56
  57 struct sched_atom;
  58
  59 struct task_desc {
  60         unsigned long           nr;
  61         unsigned long           pid;
  62         char                    comm[COMM_LEN];
  63
  64         unsigned long           nr_events;
  65         unsigned long           curr_event;
  66         struct sched_atom       **atoms;
  67
  68         pthread_t               thread;
  69         sem_t                   sleep_sem;
  70
  71         sem_t                   ready_for_work;
  72         sem_t                   work_done_sem;
  73
  74         u64                     cpu_usage;
  75 };
  76
  77 enum sched_event_type {
  78         SCHED_EVENT_RUN,
  79         SCHED_EVENT_SLEEP,
  80         SCHED_EVENT_WAKEUP,
  81         SCHED_EVENT_MIGRATION,
  82 };
  83
  84 struct sched_atom {
  85         enum sched_event_type   type;
  86         int                     specific_wait;
  87         u64                     timestamp;
  88         u64                     duration;
  89         unsigned long           nr;
  90         sem_t                   *wait_sem;
  91         struct task_desc        *wakee;
  92 };
  93
  94 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  95
  96 /* task state bitmask, copied from include/linux/sched.h */
  97 #define TASK_RUNNING            0
  98 #define TASK_INTERRUPTIBLE      1
  99 #define TASK_UNINTERRUPTIBLE    2
 100 #define __TASK_STOPPED          4
 101 #define __TASK_TRACED           8
 102 /* in tsk->exit_state */
 103 #define EXIT_DEAD               16
 104 #define EXIT_ZOMBIE             32
 105 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
 106 /* in tsk->state again */
 107 #define TASK_DEAD               64
 108 #define TASK_WAKEKILL           128
 109 #define TASK_WAKING             256
 110 #define TASK_PARKED             512
 111
 112 enum thread_state {
 113         THREAD_SLEEPING = 0,
 114         THREAD_WAIT_CPU,
 115         THREAD_SCHED_IN,
 116         THREAD_IGNORE
 117 };
 118
 119 struct work_atom {
 120         struct list_head        list;
 121         enum thread_state       state;
 122         u64                     sched_out_time;
 123         u64                     wake_up_time;
 124         u64                     sched_in_time;
 125         u64                     runtime;
 126 };
 127
 128 struct work_atoms {
 129         struct list_head        work_list;
 130         struct thread           *thread;
 131         struct rb_node          node;
 132         u64                     max_lat;
 133         u64                     max_lat_start;
 134         u64                     max_lat_end;
 135         u64                     total_lat;
 136         u64                     nb_atoms;
 137         u64                     total_runtime;
 138         int                     num_merged;
 139 };
 140
 141 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 142
 143 struct perf_sched;
 144
 145 struct trace_sched_handler {
 146         int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 147                             struct perf_sample *sample, struct machine *machine);
 148
 149         int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 150                              struct perf_sample *sample, struct machine *machine);
 151
 152         int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 153                             struct perf_sample *sample, struct machine *machine);
 154
 155         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 156         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 157                           struct machine *machine);
 158
 159         int (*migrate_task_event)(struct perf_sched *sched,
 160                                   struct evsel *evsel,
 161                                   struct perf_sample *sample,
 162                                   struct machine *machine);
 163 };
 164
 165 #define COLOR_PIDS PERF_COLOR_BLUE
 166 #define COLOR_CPUS PERF_COLOR_BG_RED
 167
 168 struct perf_sched_map {
 169         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 170         int                     *comp_cpus;
 171         bool                     comp;
 172         struct perf_thread_map *color_pids;
 173         const char              *color_pids_str;
 174         struct perf_cpu_map     *color_cpus;
 175         const char              *color_cpus_str;
 176         struct perf_cpu_map     *cpus;
 177         const char              *cpus_str;
 178 };
 179
 180 struct perf_sched {
 181         struct perf_tool tool;
 182         const char       *sort_order;
 183         unsigned long    nr_tasks;
 184         struct task_desc **pid_to_task;
 185         struct task_desc **tasks;
 186         const struct trace_sched_handler *tp_handler;
 187         pthread_mutex_t  start_work_mutex;
 188         pthread_mutex_t  work_done_wait_mutex;
 189         int              profile_cpu;
 190 /*
 191  * Track the current task - that way we can know whether there's any
 192  * weird events, such as a task being switched away that is not current.
 193  */
 194         int              max_cpu;
 195         u32              curr_pid[MAX_CPUS];
 196         struct thread    *curr_thread[MAX_CPUS];
 197         char             next_shortname1;
 198         char             next_shortname2;
 199         unsigned int     replay_repeat;
 200         unsigned long    nr_run_events;
 201         unsigned long    nr_sleep_events;
 202         unsigned long    nr_wakeup_events;
 203         unsigned long    nr_sleep_corrections;
 204         unsigned long    nr_run_events_optimized;
 205         unsigned long    targetless_wakeups;
 206         unsigned long    multitarget_wakeups;
 207         unsigned long    nr_runs;
 208         unsigned long    nr_timestamps;
 209         unsigned long    nr_unordered_timestamps;
 210         unsigned long    nr_context_switch_bugs;
 211         unsigned long    nr_events;
 212         unsigned long    nr_lost_chunks;
 213         unsigned long    nr_lost_events;
 214         u64              run_measurement_overhead;
 215         u64              sleep_measurement_overhead;
 216         u64              start_time;
 217         u64              cpu_usage;
 218         u64              runavg_cpu_usage;
 219         u64              parent_cpu_usage;
 220         u64              runavg_parent_cpu_usage;
 221         u64              sum_runtime;
 222         u64              sum_fluct;
 223         u64              run_avg;
 224         u64              all_runtime;
 225         u64              all_count;
 226         u64              cpu_last_switched[MAX_CPUS];
 227         struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 228         struct list_head sort_list, cmp_pid;
 229         bool force;
 230         bool skip_merge;
 231         struct perf_sched_map map;
 232
 233         /* options for timehist command */
 234         bool            summary;
 235         bool            summary_only;
 236         bool            idle_hist;
 237         bool            show_callchain;
 238         unsigned int    max_stack;
 239         bool            show_cpu_visual;
 240         bool            show_wakeups;
 241         bool            show_next;
 242         bool            show_migrations;
 243         bool            show_state;
 244         u64             skipped_samples;
 245         const char      *time_str;
 246         struct perf_time_interval ptime;
 247         struct perf_time_interval hist_time;
 248 };
 249
 250 /* per thread run time data */
 251 struct thread_runtime {
 252         u64 last_time;      /* time of previous sched in/out event */
 253         u64 dt_run;         /* run time */
 254         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 255         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 256         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 257         u64 dt_delay;       /* time between wakeup and sched-in */
 258         u64 ready_to_run;   /* time of wakeup */
 259
 260         struct stats run_stats;
 261         u64 total_run_time;
 262         u64 total_sleep_time;
 263         u64 total_iowait_time;
 264         u64 total_preempt_time;
 265         u64 total_delay_time;
 266
 267         int last_state;
 268
 269         char shortname[3];
 270         bool comm_changed;
 271
 272         u64 migrations;
 273 };
 274
 275 /* per event run time data */
 276 struct evsel_runtime {
 277         u64 *last_time; /* time this event was last seen per cpu */
 278         u32 ncpu;       /* highest cpu slot allocated */
 279 };
 280
 281 /* per cpu idle time data */
 282 struct idle_thread_runtime {
 283         struct thread_runtime   tr;
 284         struct thread           *last_thread;
 285         struct rb_root_cached   sorted_root;
 286         struct callchain_root   callchain;
 287         struct callchain_cursor cursor;
 288 };
 289
 290 /* track idle times per cpu */
 291 static struct thread **idle_threads;
 292 static int idle_max_cpu;
 293 static char idle_comm[] = "<idle>";
 294
 295 static u64 get_nsecs(void)
 296 {
 297         struct timespec ts;
 298
 299         clock_gettime(CLOCK_MONOTONIC, &ts);
 300
 301         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 302 }
 303
 304 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 305 {
 306         u64 T0 = get_nsecs(), T1;
 307
 308         do {
 309                 T1 = get_nsecs();
 310         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 311 }
 312
 313 static void sleep_nsecs(u64 nsecs)
 314 {
 315         struct timespec ts;
 316
 317         ts.tv_nsec = nsecs % 999999999;
 318         ts.tv_sec = nsecs / 999999999;
 319
 320         nanosleep(&ts, NULL);
 321 }
 322
 323 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 324 {
 325         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 326         int i;
 327
 328         for (i = 0; i < 10; i++) {
 329                 T0 = get_nsecs();
 330                 burn_nsecs(sched, 0);
 331                 T1 = get_nsecs();
 332                 delta = T1-T0;
 333                 min_delta = min(min_delta, delta);
 334         }
 335         sched->run_measurement_overhead = min_delta;
 336
 337         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 338 }
 339
 340 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 341 {
 342         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 343         int i;
 344
 345         for (i = 0; i < 10; i++) {
 346                 T0 = get_nsecs();
 347                 sleep_nsecs(10000);
 348                 T1 = get_nsecs();
 349                 delta = T1-T0;
 350                 min_delta = min(min_delta, delta);
 351         }
 352         min_delta -= 10000;
 353         sched->sleep_measurement_overhead = min_delta;
 354
 355         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 356 }
 357
 358 static struct sched_atom *
 359 get_new_event(struct task_desc *task, u64 timestamp)
 360 {
 361         struct sched_atom *event = zalloc(sizeof(*event));
 362         unsigned long idx = task->nr_events;
 363         size_t size;
 364
 365         event->timestamp = timestamp;
 366         event->nr = idx;
 367
 368         task->nr_events++;
 369         size = sizeof(struct sched_atom *) * task->nr_events;
 370         task->atoms = realloc(task->atoms, size);
 371         BUG_ON(!task->atoms);
 372
 373         task->atoms[idx] = event;
 374
 375         return event;
 376 }
 377
 378 static struct sched_atom *last_event(struct task_desc *task)
 379 {
 380         if (!task->nr_events)
 381                 return NULL;
 382
 383         return task->atoms[task->nr_events - 1];
 384 }
 385
 386 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 387                                 u64 timestamp, u64 duration)
 388 {
 389         struct sched_atom *event, *curr_event = last_event(task);
 390
 391         /*
 392          * optimize an existing RUN event by merging this one
 393          * to it:
 394          */
 395         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 396                 sched->nr_run_events_optimized++;
 397                 curr_event->duration += duration;
 398                 return;
 399         }
 400
 401         event = get_new_event(task, timestamp);
 402
 403         event->type = SCHED_EVENT_RUN;
 404         event->duration = duration;
 405
 406         sched->nr_run_events++;
 407 }
 408
 409 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 410                                    u64 timestamp, struct task_desc *wakee)
 411 {
 412         struct sched_atom *event, *wakee_event;
 413
 414         event = get_new_event(task, timestamp);
 415         event->type = SCHED_EVENT_WAKEUP;
 416         event->wakee = wakee;
 417
 418         wakee_event = last_event(wakee);
 419         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 420                 sched->targetless_wakeups++;
 421                 return;
 422         }
 423         if (wakee_event->wait_sem) {
 424                 sched->multitarget_wakeups++;
 425                 return;
 426         }
 427
 428         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 429         sem_init(wakee_event->wait_sem, 0, 0);
 430         wakee_event->specific_wait = 1;
 431         event->wait_sem = wakee_event->wait_sem;
 432
 433         sched->nr_wakeup_events++;
 434 }
 435
 436 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 437                                   u64 timestamp, u64 task_state __maybe_unused)
 438 {
 439         struct sched_atom *event = get_new_event(task, timestamp);
 440
 441         event->type = SCHED_EVENT_SLEEP;
 442
 443         sched->nr_sleep_events++;
 444 }
 445
 446 static struct task_desc *register_pid(struct perf_sched *sched,
 447                                       unsigned long pid, const char *comm)
 448 {
 449         struct task_desc *task;
 450         static int pid_max;
 451
 452         if (sched->pid_to_task == NULL) {
 453                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 454                         pid_max = MAX_PID;
 455                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 456         }
 457         if (pid >= (unsigned long)pid_max) {
 458                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 459                         sizeof(struct task_desc *))) == NULL);
 460                 while (pid >= (unsigned long)pid_max)
 461                         sched->pid_to_task[pid_max++] = NULL;
 462         }
 463
 464         task = sched->pid_to_task[pid];
 465
 466         if (task)
 467                 return task;
 468
 469         task = zalloc(sizeof(*task));
 470         task->pid = pid;
 471         task->nr = sched->nr_tasks;
 472         strcpy(task->comm, comm);
 473         /*
 474          * every task starts in sleeping state - this gets ignored
 475          * if there's no wakeup pointing to this sleep state:
 476          */
 477         add_sched_event_sleep(sched, task, 0, 0);
 478
 479         sched->pid_to_task[pid] = task;
 480         sched->nr_tasks++;
 481         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 482         BUG_ON(!sched->tasks);
 483         sched->tasks[task->nr] = task;
 484
 485         if (verbose > 0)
 486                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 487
 488         return task;
 489 }
 490
 491
 492 static void print_task_traces(struct perf_sched *sched)
 493 {
 494         struct task_desc *task;
 495         unsigned long i;
 496
 497         for (i = 0; i < sched->nr_tasks; i++) {
 498                 task = sched->tasks[i];
 499                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 500                         task->nr, task->comm, task->pid, task->nr_events);
 501         }
 502 }
 503
 504 static void add_cross_task_wakeups(struct perf_sched *sched)
 505 {
 506         struct task_desc *task1, *task2;
 507         unsigned long i, j;
 508
 509         for (i = 0; i < sched->nr_tasks; i++) {
 510                 task1 = sched->tasks[i];
 511                 j = i + 1;
 512                 if (j == sched->nr_tasks)
 513                         j = 0;
 514                 task2 = sched->tasks[j];
 515                 add_sched_event_wakeup(sched, task1, 0, task2);
 516         }
 517 }
 518
 519 static void perf_sched__process_event(struct perf_sched *sched,
 520                                       struct sched_atom *atom)
 521 {
 522         int ret = 0;
 523
 524         switch (atom->type) {
 525                 case SCHED_EVENT_RUN:
 526                         burn_nsecs(sched, atom->duration);
 527                         break;
 528                 case SCHED_EVENT_SLEEP:
 529                         if (atom->wait_sem)
 530                                 ret = sem_wait(atom->wait_sem);
 531                         BUG_ON(ret);
 532                         break;
 533                 case SCHED_EVENT_WAKEUP:
 534                         if (atom->wait_sem)
 535                                 ret = sem_post(atom->wait_sem);
 536                         BUG_ON(ret);
 537                         break;
 538                 case SCHED_EVENT_MIGRATION:
 539                         break;
 540                 default:
 541                         BUG_ON(1);
 542         }
 543 }
 544
 545 static u64 get_cpu_usage_nsec_parent(void)
 546 {
 547         struct rusage ru;
 548         u64 sum;
 549         int err;
 550
 551         err = getrusage(RUSAGE_SELF, &ru);
 552         BUG_ON(err);
 553
 554         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 555         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 556
 557         return sum;
 558 }
 559
 560 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 561 {
 562         struct perf_event_attr attr;
 563         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 564         int fd;
 565         struct rlimit limit;
 566         bool need_privilege = false;
 567
 568         memset(&attr, 0, sizeof(attr));
 569
 570         attr.type = PERF_TYPE_SOFTWARE;
 571         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 572
 573 force_again:
 574         fd = sys_perf_event_open(&attr, 0, -1, -1,
 575                                  perf_event_open_cloexec_flag());
 576
 577         if (fd < 0) {
 578                 if (errno == EMFILE) {
 579                         if (sched->force) {
 580                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 581                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 582                                 if (limit.rlim_cur > limit.rlim_max) {
 583                                         limit.rlim_max = limit.rlim_cur;
 584                                         need_privilege = true;
 585                                 }
 586                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 587                                         if (need_privilege && errno == EPERM)
 588                                                 strcpy(info, "Need privilege\n");
 589                                 } else
 590                                         goto force_again;
 591                         } else
 592                                 strcpy(info, "Have a try with -f option\n");
 593                 }
 594                 pr_err("Error: sys_perf_event_open() syscall returned "
 595                        "with %d (%s)\n%s", fd,
 596                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 597                 exit(EXIT_FAILURE);
 598         }
 599         return fd;
 600 }
 601
 602 static u64 get_cpu_usage_nsec_self(int fd)
 603 {
 604         u64 runtime;
 605         int ret;
 606
 607         ret = read(fd, &runtime, sizeof(runtime));
 608         BUG_ON(ret != sizeof(runtime));
 609
 610         return runtime;
 611 }
 612
 613 struct sched_thread_parms {
 614         struct task_desc  *task;
 615         struct perf_sched *sched;
 616         int fd;
 617 };
 618
 619 static void *thread_func(void *ctx)
 620 {
 621         struct sched_thread_parms *parms = ctx;
 622         struct task_desc *this_task = parms->task;
 623         struct perf_sched *sched = parms->sched;
 624         u64 cpu_usage_0, cpu_usage_1;
 625         unsigned long i, ret;
 626         char comm2[22];
 627         int fd = parms->fd;
 628
 629         zfree(&parms);
 630
 631         sprintf(comm2, ":%s", this_task->comm);
 632         prctl(PR_SET_NAME, comm2);
 633         if (fd < 0)
 634                 return NULL;
 635 again:
 636         ret = sem_post(&this_task->ready_for_work);
 637         BUG_ON(ret);
 638         ret = pthread_mutex_lock(&sched->start_work_mutex);
 639         BUG_ON(ret);
 640         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 641         BUG_ON(ret);
 642
 643         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 644
 645         for (i = 0; i < this_task->nr_events; i++) {
 646                 this_task->curr_event = i;
 647                 perf_sched__process_event(sched, this_task->atoms[i]);
 648         }
 649
 650         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 651         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 652         ret = sem_post(&this_task->work_done_sem);
 653         BUG_ON(ret);
 654
 655         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 656         BUG_ON(ret);
 657         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 658         BUG_ON(ret);
 659
 660         goto again;
 661 }
 662
 663 static void create_tasks(struct perf_sched *sched)
 664 {
 665         struct task_desc *task;
 666         pthread_attr_t attr;
 667         unsigned long i;
 668         int err;
 669
 670         err = pthread_attr_init(&attr);
 671         BUG_ON(err);
 672         err = pthread_attr_setstacksize(&attr,
 673                         (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
 674         BUG_ON(err);
 675         err = pthread_mutex_lock(&sched->start_work_mutex);
 676         BUG_ON(err);
 677         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 678         BUG_ON(err);
 679         for (i = 0; i < sched->nr_tasks; i++) {
 680                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 681                 BUG_ON(parms == NULL);
 682                 parms->task = task = sched->tasks[i];
 683                 parms->sched = sched;
 684                 parms->fd = self_open_counters(sched, i);
 685                 sem_init(&task->sleep_sem, 0, 0);
 686                 sem_init(&task->ready_for_work, 0, 0);
 687                 sem_init(&task->work_done_sem, 0, 0);
 688                 task->curr_event = 0;
 689                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 690                 BUG_ON(err);
 691         }
 692 }
 693
 694 static void wait_for_tasks(struct perf_sched *sched)
 695 {
 696         u64 cpu_usage_0, cpu_usage_1;
 697         struct task_desc *task;
 698         unsigned long i, ret;
 699
 700         sched->start_time = get_nsecs();
 701         sched->cpu_usage = 0;
 702         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 703
 704         for (i = 0; i < sched->nr_tasks; i++) {
 705                 task = sched->tasks[i];
 706                 ret = sem_wait(&task->ready_for_work);
 707                 BUG_ON(ret);
 708                 sem_init(&task->ready_for_work, 0, 0);
 709         }
 710         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 711         BUG_ON(ret);
 712
 713         cpu_usage_0 = get_cpu_usage_nsec_parent();
 714
 715         pthread_mutex_unlock(&sched->start_work_mutex);
 716
 717         for (i = 0; i < sched->nr_tasks; i++) {
 718                 task = sched->tasks[i];
 719                 ret = sem_wait(&task->work_done_sem);
 720                 BUG_ON(ret);
 721                 sem_init(&task->work_done_sem, 0, 0);
 722                 sched->cpu_usage += task->cpu_usage;
 723                 task->cpu_usage = 0;
 724         }
 725
 726         cpu_usage_1 = get_cpu_usage_nsec_parent();
 727         if (!sched->runavg_cpu_usage)
 728                 sched->runavg_cpu_usage = sched->cpu_usage;
 729         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 730
 731         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 732         if (!sched->runavg_parent_cpu_usage)
 733                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 734         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 735                                          sched->parent_cpu_usage)/sched->replay_repeat;
 736
 737         ret = pthread_mutex_lock(&sched->start_work_mutex);
 738         BUG_ON(ret);
 739
 740         for (i = 0; i < sched->nr_tasks; i++) {
 741                 task = sched->tasks[i];
 742                 sem_init(&task->sleep_sem, 0, 0);
 743                 task->curr_event = 0;
 744         }
 745 }
 746
 747 static void run_one_test(struct perf_sched *sched)
 748 {
 749         u64 T0, T1, delta, avg_delta, fluct;
 750
 751         T0 = get_nsecs();
 752         wait_for_tasks(sched);
 753         T1 = get_nsecs();
 754
 755         delta = T1 - T0;
 756         sched->sum_runtime += delta;
 757         sched->nr_runs++;
 758
 759         avg_delta = sched->sum_runtime / sched->nr_runs;
 760         if (delta < avg_delta)
 761                 fluct = avg_delta - delta;
 762         else
 763                 fluct = delta - avg_delta;
 764         sched->sum_fluct += fluct;
 765         if (!sched->run_avg)
 766                 sched->run_avg = delta;
 767         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 768
 769         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 770
 771         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 772
 773         printf("cpu: %0.2f / %0.2f",
 774                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 775
 776 #if 0
 777         /*
 778          * rusage statistics done by the parent, these are less
 779          * accurate than the sched->sum_exec_runtime based statistics:
 780          */
 781         printf(" [%0.2f / %0.2f]",
 782                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 783                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 784 #endif
 785
 786         printf("\n");
 787
 788         if (sched->nr_sleep_corrections)
 789                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 790         sched->nr_sleep_corrections = 0;
 791 }
 792
 793 static void test_calibrations(struct perf_sched *sched)
 794 {
 795         u64 T0, T1;
 796
 797         T0 = get_nsecs();
 798         burn_nsecs(sched, NSEC_PER_MSEC);
 799         T1 = get_nsecs();
 800
 801         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 802
 803         T0 = get_nsecs();
 804         sleep_nsecs(NSEC_PER_MSEC);
 805         T1 = get_nsecs();
 806
 807         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 808 }
 809
 810 static int
 811 replay_wakeup_event(struct perf_sched *sched,
 812                     struct evsel *evsel, struct perf_sample *sample,
 813                     struct machine *machine __maybe_unused)
 814 {
 815         const char *comm = evsel__strval(evsel, sample, "comm");
 816         const u32 pid    = evsel__intval(evsel, sample, "pid");
 817         struct task_desc *waker, *wakee;
 818
 819         if (verbose > 0) {
 820                 printf("sched_wakeup event %p\n", evsel);
 821
 822                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 823         }
 824
 825         waker = register_pid(sched, sample->tid, "<unknown>");
 826         wakee = register_pid(sched, pid, comm);
 827
 828         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 829         return 0;
 830 }
 831
 832 static int replay_switch_event(struct perf_sched *sched,
 833                                struct evsel *evsel,
 834                                struct perf_sample *sample,
 835                                struct machine *machine __maybe_unused)
 836 {
 837         const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 838                    *next_comm  = evsel__strval(evsel, sample, "next_comm");
 839         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 840                   next_pid = evsel__intval(evsel, sample, "next_pid");
 841         const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
 842         struct task_desc *prev, __maybe_unused *next;
 843         u64 timestamp0, timestamp = sample->time;
 844         int cpu = sample->cpu;
 845         s64 delta;
 846
 847         if (verbose > 0)
 848                 printf("sched_switch event %p\n", evsel);
 849
 850         if (cpu >= MAX_CPUS || cpu < 0)
 851                 return 0;
 852
 853         timestamp0 = sched->cpu_last_switched[cpu];
 854         if (timestamp0)
 855                 delta = timestamp - timestamp0;
 856         else
 857                 delta = 0;
 858
 859         if (delta < 0) {
 860                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 861                 return -1;
 862         }
 863
 864         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 865                  prev_comm, prev_pid, next_comm, next_pid, delta);
 866
 867         prev = register_pid(sched, prev_pid, prev_comm);
 868         next = register_pid(sched, next_pid, next_comm);
 869
 870         sched->cpu_last_switched[cpu] = timestamp;
 871
 872         add_sched_event_run(sched, prev, timestamp, delta);
 873         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 874
 875         return 0;
 876 }
 877
 878 static int replay_fork_event(struct perf_sched *sched,
 879                              union perf_event *event,
 880                              struct machine *machine)
 881 {
 882         struct thread *child, *parent;
 883
 884         child = machine__findnew_thread(machine, event->fork.pid,
 885                                         event->fork.tid);
 886         parent = machine__findnew_thread(machine, event->fork.ppid,
 887                                          event->fork.ptid);
 888
 889         if (child == NULL || parent == NULL) {
 890                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 891                                  child, parent);
 892                 goto out_put;
 893         }
 894
 895         if (verbose > 0) {
 896                 printf("fork event\n");
 897                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 898                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 899         }
 900
 901         register_pid(sched, parent->tid, thread__comm_str(parent));
 902         register_pid(sched, child->tid, thread__comm_str(child));
 903 out_put:
 904         thread__put(child);
 905         thread__put(parent);
 906         return 0;
 907 }
 908
 909 struct sort_dimension {
 910         const char              *name;
 911         sort_fn_t               cmp;
 912         struct list_head        list;
 913 };
 914
 915 /*
 916  * handle runtime stats saved per thread
 917  */
 918 static struct thread_runtime *thread__init_runtime(struct thread *thread)
 919 {
 920         struct thread_runtime *r;
 921
 922         r = zalloc(sizeof(struct thread_runtime));
 923         if (!r)
 924                 return NULL;
 925
 926         init_stats(&r->run_stats);
 927         thread__set_priv(thread, r);
 928
 929         return r;
 930 }
 931
 932 static struct thread_runtime *thread__get_runtime(struct thread *thread)
 933 {
 934         struct thread_runtime *tr;
 935
 936         tr = thread__priv(thread);
 937         if (tr == NULL) {
 938                 tr = thread__init_runtime(thread);
 939                 if (tr == NULL)
 940                         pr_debug("Failed to malloc memory for runtime data.\n");
 941         }
 942
 943         return tr;
 944 }
 945
 946 static int
 947 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 948 {
 949         struct sort_dimension *sort;
 950         int ret = 0;
 951
 952         BUG_ON(list_empty(list));
 953
 954         list_for_each_entry(sort, list, list) {
 955                 ret = sort->cmp(l, r);
 956                 if (ret)
 957                         return ret;
 958         }
 959
 960         return ret;
 961 }
 962
 963 static struct work_atoms *
 964 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 965                          struct list_head *sort_list)
 966 {
 967         struct rb_node *node = root->rb_root.rb_node;
 968         struct work_atoms key = { .thread = thread };
 969
 970         while (node) {
 971                 struct work_atoms *atoms;
 972                 int cmp;
 973
 974                 atoms = container_of(node, struct work_atoms, node);
 975
 976                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 977                 if (cmp > 0)
 978                         node = node->rb_left;
 979                 else if (cmp < 0)
 980                         node = node->rb_right;
 981                 else {
 982                         BUG_ON(thread != atoms->thread);
 983                         return atoms;
 984                 }
 985         }
 986         return NULL;
 987 }
 988
 989 static void
 990 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 991                          struct list_head *sort_list)
 992 {
 993         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 994         bool leftmost = true;
 995
 996         while (*new) {
 997                 struct work_atoms *this;
 998                 int cmp;
 999
1000                 this = container_of(*new, struct work_atoms, node);
1001                 parent = *new;
1002
1003                 cmp = thread_lat_cmp(sort_list, data, this);
1004
1005                 if (cmp > 0)
1006                         new = &((*new)->rb_left);
1007                 else {
1008                         new = &((*new)->rb_right);
1009                         leftmost = false;
1010                 }
1011         }
1012
1013         rb_link_node(&data->node, parent, new);
1014         rb_insert_color_cached(&data->node, root, leftmost);
1015 }
1016
1017 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1018 {
1019         struct work_atoms *atoms = zalloc(sizeof(*atoms));
1020         if (!atoms) {
1021                 pr_err("No memory at %s\n", __func__);
1022                 return -1;
1023         }
1024
1025         atoms->thread = thread__get(thread);
1026         INIT_LIST_HEAD(&atoms->work_list);
1027         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1028         return 0;
1029 }
1030
1031 static char sched_out_state(u64 prev_state)
1032 {
1033         const char *str = TASK_STATE_TO_CHAR_STR;
1034
1035         return str[prev_state];
1036 }
1037
1038 static int
1039 add_sched_out_event(struct work_atoms *atoms,
1040                     char run_state,
1041                     u64 timestamp)
1042 {
1043         struct work_atom *atom = zalloc(sizeof(*atom));
1044         if (!atom) {
1045                 pr_err("Non memory at %s", __func__);
1046                 return -1;
1047         }
1048
1049         atom->sched_out_time = timestamp;
1050
1051         if (run_state == 'R') {
1052                 atom->state = THREAD_WAIT_CPU;
1053                 atom->wake_up_time = atom->sched_out_time;
1054         }
1055
1056         list_add_tail(&atom->list, &atoms->work_list);
1057         return 0;
1058 }
1059
1060 static void
1061 add_runtime_event(struct work_atoms *atoms, u64 delta,
1062                   u64 timestamp __maybe_unused)
1063 {
1064         struct work_atom *atom;
1065
1066         BUG_ON(list_empty(&atoms->work_list));
1067
1068         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1069
1070         atom->runtime += delta;
1071         atoms->total_runtime += delta;
1072 }
1073
1074 static void
1075 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1076 {
1077         struct work_atom *atom;
1078         u64 delta;
1079
1080         if (list_empty(&atoms->work_list))
1081                 return;
1082
1083         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1084
1085         if (atom->state != THREAD_WAIT_CPU)
1086                 return;
1087
1088         if (timestamp < atom->wake_up_time) {
1089                 atom->state = THREAD_IGNORE;
1090                 return;
1091         }
1092
1093         atom->state = THREAD_SCHED_IN;
1094         atom->sched_in_time = timestamp;
1095
1096         delta = atom->sched_in_time - atom->wake_up_time;
1097         atoms->total_lat += delta;
1098         if (delta > atoms->max_lat) {
1099                 atoms->max_lat = delta;
1100                 atoms->max_lat_start = atom->wake_up_time;
1101                 atoms->max_lat_end = timestamp;
1102         }
1103         atoms->nb_atoms++;
1104 }
1105
1106 static int latency_switch_event(struct perf_sched *sched,
1107                                 struct evsel *evsel,
1108                                 struct perf_sample *sample,
1109                                 struct machine *machine)
1110 {
1111         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1112                   next_pid = evsel__intval(evsel, sample, "next_pid");
1113         const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1114         struct work_atoms *out_events, *in_events;
1115         struct thread *sched_out, *sched_in;
1116         u64 timestamp0, timestamp = sample->time;
1117         int cpu = sample->cpu, err = -1;
1118         s64 delta;
1119
1120         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1121
1122         timestamp0 = sched->cpu_last_switched[cpu];
1123         sched->cpu_last_switched[cpu] = timestamp;
1124         if (timestamp0)
1125                 delta = timestamp - timestamp0;
1126         else
1127                 delta = 0;
1128
1129         if (delta < 0) {
1130                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1131                 return -1;
1132         }
1133
1134         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1135         sched_in = machine__findnew_thread(machine, -1, next_pid);
1136         if (sched_out == NULL || sched_in == NULL)
1137                 goto out_put;
1138
1139         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1140         if (!out_events) {
1141                 if (thread_atoms_insert(sched, sched_out))
1142                         goto out_put;
1143                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1144                 if (!out_events) {
1145                         pr_err("out-event: Internal tree error");
1146                         goto out_put;
1147                 }
1148         }
1149         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1150                 return -1;
1151
1152         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1153         if (!in_events) {
1154                 if (thread_atoms_insert(sched, sched_in))
1155                         goto out_put;
1156                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1157                 if (!in_events) {
1158                         pr_err("in-event: Internal tree error");
1159                         goto out_put;
1160                 }
1161                 /*
1162                  * Take came in we have not heard about yet,
1163                  * add in an initial atom in runnable state:
1164                  */
1165                 if (add_sched_out_event(in_events, 'R', timestamp))
1166                         goto out_put;
1167         }
1168         add_sched_in_event(in_events, timestamp);
1169         err = 0;
1170 out_put:
1171         thread__put(sched_out);
1172         thread__put(sched_in);
1173         return err;
1174 }
1175
1176 static int latency_runtime_event(struct perf_sched *sched,
1177                                  struct evsel *evsel,
1178                                  struct perf_sample *sample,
1179                                  struct machine *machine)
1180 {
1181         const u32 pid      = evsel__intval(evsel, sample, "pid");
1182         const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1183         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1184         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1185         u64 timestamp = sample->time;
1186         int cpu = sample->cpu, err = -1;
1187
1188         if (thread == NULL)
1189                 return -1;
1190
1191         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1192         if (!atoms) {
1193                 if (thread_atoms_insert(sched, thread))
1194                         goto out_put;
1195                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1196                 if (!atoms) {
1197                         pr_err("in-event: Internal tree error");
1198                         goto out_put;
1199                 }
1200                 if (add_sched_out_event(atoms, 'R', timestamp))
1201                         goto out_put;
1202         }
1203
1204         add_runtime_event(atoms, runtime, timestamp);
1205         err = 0;
1206 out_put:
1207         thread__put(thread);
1208         return err;
1209 }
1210
1211 static int latency_wakeup_event(struct perf_sched *sched,
1212                                 struct evsel *evsel,
1213                                 struct perf_sample *sample,
1214                                 struct machine *machine)
1215 {
1216         const u32 pid     = evsel__intval(evsel, sample, "pid");
1217         struct work_atoms *atoms;
1218         struct work_atom *atom;
1219         struct thread *wakee;
1220         u64 timestamp = sample->time;
1221         int err = -1;
1222
1223         wakee = machine__findnew_thread(machine, -1, pid);
1224         if (wakee == NULL)
1225                 return -1;
1226         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1227         if (!atoms) {
1228                 if (thread_atoms_insert(sched, wakee))
1229                         goto out_put;
1230                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1231                 if (!atoms) {
1232                         pr_err("wakeup-event: Internal tree error");
1233                         goto out_put;
1234                 }
1235                 if (add_sched_out_event(atoms, 'S', timestamp))
1236                         goto out_put;
1237         }
1238
1239         BUG_ON(list_empty(&atoms->work_list));
1240
1241         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1242
1243         /*
1244          * As we do not guarantee the wakeup event happens when
1245          * task is out of run queue, also may happen when task is
1246          * on run queue and wakeup only change ->state to TASK_RUNNING,
1247          * then we should not set the ->wake_up_time when wake up a
1248          * task which is on run queue.
1249          *
1250          * You WILL be missing events if you've recorded only
1251          * one CPU, or are only looking at only one, so don't
1252          * skip in this case.
1253          */
1254         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1255                 goto out_ok;
1256
1257         sched->nr_timestamps++;
1258         if (atom->sched_out_time > timestamp) {
1259                 sched->nr_unordered_timestamps++;
1260                 goto out_ok;
1261         }
1262
1263         atom->state = THREAD_WAIT_CPU;
1264         atom->wake_up_time = timestamp;
1265 out_ok:
1266         err = 0;
1267 out_put:
1268         thread__put(wakee);
1269         return err;
1270 }
1271
1272 static int latency_migrate_task_event(struct perf_sched *sched,
1273                                       struct evsel *evsel,
1274                                       struct perf_sample *sample,
1275                                       struct machine *machine)
1276 {
1277         const u32 pid = evsel__intval(evsel, sample, "pid");
1278         u64 timestamp = sample->time;
1279         struct work_atoms *atoms;
1280         struct work_atom *atom;
1281         struct thread *migrant;
1282         int err = -1;
1283
1284         /*
1285          * Only need to worry about migration when profiling one CPU.
1286          */
1287         if (sched->profile_cpu == -1)
1288                 return 0;
1289
1290         migrant = machine__findnew_thread(machine, -1, pid);
1291         if (migrant == NULL)
1292                 return -1;
1293         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1294         if (!atoms) {
1295                 if (thread_atoms_insert(sched, migrant))
1296                         goto out_put;
1297                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1298                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1299                 if (!atoms) {
1300                         pr_err("migration-event: Internal tree error");
1301                         goto out_put;
1302                 }
1303                 if (add_sched_out_event(atoms, 'R', timestamp))
1304                         goto out_put;
1305         }
1306
1307         BUG_ON(list_empty(&atoms->work_list));
1308
1309         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1310         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1311
1312         sched->nr_timestamps++;
1313
1314         if (atom->sched_out_time > timestamp)
1315                 sched->nr_unordered_timestamps++;
1316         err = 0;
1317 out_put:
1318         thread__put(migrant);
1319         return err;
1320 }
1321
1322 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1323 {
1324         int i;
1325         int ret;
1326         u64 avg;
1327         char max_lat_start[32], max_lat_end[32];
1328
1329         if (!work_list->nb_atoms)
1330                 return;
1331         /*
1332          * Ignore idle threads:
1333          */
1334         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1335                 return;
1336
1337         sched->all_runtime += work_list->total_runtime;
1338         sched->all_count   += work_list->nb_atoms;
1339
1340         if (work_list->num_merged > 1)
1341                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1342         else
1343                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1344
1345         for (i = 0; i < 24 - ret; i++)
1346                 printf(" ");
1347
1348         avg = work_list->total_lat / work_list->nb_atoms;
1349         timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350         timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1351
1352         printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353               (double)work_list->total_runtime / NSEC_PER_MSEC,
1354                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355                  (double)work_list->max_lat / NSEC_PER_MSEC,
1356                  max_lat_start, max_lat_end);
1357 }
1358
1359 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1360 {
1361         if (l->thread == r->thread)
1362                 return 0;
1363         if (l->thread->tid < r->thread->tid)
1364                 return -1;
1365         if (l->thread->tid > r->thread->tid)
1366                 return 1;
1367         return (int)(l->thread - r->thread);
1368 }
1369
1370 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1371 {
1372         u64 avgl, avgr;
1373
1374         if (!l->nb_atoms)
1375                 return -1;
1376
1377         if (!r->nb_atoms)
1378                 return 1;
1379
1380         avgl = l->total_lat / l->nb_atoms;
1381         avgr = r->total_lat / r->nb_atoms;
1382
1383         if (avgl < avgr)
1384                 return -1;
1385         if (avgl > avgr)
1386                 return 1;
1387
1388         return 0;
1389 }
1390
1391 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1392 {
1393         if (l->max_lat < r->max_lat)
1394                 return -1;
1395         if (l->max_lat > r->max_lat)
1396                 return 1;
1397
1398         return 0;
1399 }
1400
1401 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1402 {
1403         if (l->nb_atoms < r->nb_atoms)
1404                 return -1;
1405         if (l->nb_atoms > r->nb_atoms)
1406                 return 1;
1407
1408         return 0;
1409 }
1410
1411 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1412 {
1413         if (l->total_runtime < r->total_runtime)
1414                 return -1;
1415         if (l->total_runtime > r->total_runtime)
1416                 return 1;
1417
1418         return 0;
1419 }
1420
1421 static int sort_dimension__add(const char *tok, struct list_head *list)
1422 {
1423         size_t i;
1424         static struct sort_dimension avg_sort_dimension = {
1425                 .name = "avg",
1426                 .cmp  = avg_cmp,
1427         };
1428         static struct sort_dimension max_sort_dimension = {
1429                 .name = "max",
1430                 .cmp  = max_cmp,
1431         };
1432         static struct sort_dimension pid_sort_dimension = {
1433                 .name = "pid",
1434                 .cmp  = pid_cmp,
1435         };
1436         static struct sort_dimension runtime_sort_dimension = {
1437                 .name = "runtime",
1438                 .cmp  = runtime_cmp,
1439         };
1440         static struct sort_dimension switch_sort_dimension = {
1441                 .name = "switch",
1442                 .cmp  = switch_cmp,
1443         };
1444         struct sort_dimension *available_sorts[] = {
1445                 &pid_sort_dimension,
1446                 &avg_sort_dimension,
1447                 &max_sort_dimension,
1448                 &switch_sort_dimension,
1449                 &runtime_sort_dimension,
1450         };
1451
1452         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1453                 if (!strcmp(available_sorts[i]->name, tok)) {
1454                         list_add_tail(&available_sorts[i]->list, list);
1455
1456                         return 0;
1457                 }
1458         }
1459
1460         return -1;
1461 }
1462
1463 static void perf_sched__sort_lat(struct perf_sched *sched)
1464 {
1465         struct rb_node *node;
1466         struct rb_root_cached *root = &sched->atom_root;
1467 again:
1468         for (;;) {
1469                 struct work_atoms *data;
1470                 node = rb_first_cached(root);
1471                 if (!node)
1472                         break;
1473
1474                 rb_erase_cached(node, root);
1475                 data = rb_entry(node, struct work_atoms, node);
1476                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1477         }
1478         if (root == &sched->atom_root) {
1479                 root = &sched->merged_atom_root;
1480                 goto again;
1481         }
1482 }
1483
1484 static int process_sched_wakeup_event(struct perf_tool *tool,
1485                                       struct evsel *evsel,
1486                                       struct perf_sample *sample,
1487                                       struct machine *machine)
1488 {
1489         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1490
1491         if (sched->tp_handler->wakeup_event)
1492                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1493
1494         return 0;
1495 }
1496
1497 union map_priv {
1498         void    *ptr;
1499         bool     color;
1500 };
1501
1502 static bool thread__has_color(struct thread *thread)
1503 {
1504         union map_priv priv = {
1505                 .ptr = thread__priv(thread),
1506         };
1507
1508         return priv.color;
1509 }
1510
1511 static struct thread*
1512 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1513 {
1514         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1515         union map_priv priv = {
1516                 .color = false,
1517         };
1518
1519         if (!sched->map.color_pids || !thread || thread__priv(thread))
1520                 return thread;
1521
1522         if (thread_map__has(sched->map.color_pids, tid))
1523                 priv.color = true;
1524
1525         thread__set_priv(thread, priv.ptr);
1526         return thread;
1527 }
1528
1529 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1530                             struct perf_sample *sample, struct machine *machine)
1531 {
1532         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1533         struct thread *sched_in;
1534         struct thread_runtime *tr;
1535         int new_shortname;
1536         u64 timestamp0, timestamp = sample->time;
1537         s64 delta;
1538         int i, this_cpu = sample->cpu;
1539         int cpus_nr;
1540         bool new_cpu = false;
1541         const char *color = PERF_COLOR_NORMAL;
1542         char stimestamp[32];
1543
1544         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1545
1546         if (this_cpu > sched->max_cpu)
1547                 sched->max_cpu = this_cpu;
1548
1549         if (sched->map.comp) {
1550                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1551                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1552                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1553                         new_cpu = true;
1554                 }
1555         } else
1556                 cpus_nr = sched->max_cpu;
1557
1558         timestamp0 = sched->cpu_last_switched[this_cpu];
1559         sched->cpu_last_switched[this_cpu] = timestamp;
1560         if (timestamp0)
1561                 delta = timestamp - timestamp0;
1562         else
1563                 delta = 0;
1564
1565         if (delta < 0) {
1566                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1567                 return -1;
1568         }
1569
1570         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1571         if (sched_in == NULL)
1572                 return -1;
1573
1574         tr = thread__get_runtime(sched_in);
1575         if (tr == NULL) {
1576                 thread__put(sched_in);
1577                 return -1;
1578         }
1579
1580         sched->curr_thread[this_cpu] = thread__get(sched_in);
1581
1582         printf("  ");
1583
1584         new_shortname = 0;
1585         if (!tr->shortname[0]) {
1586                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1587                         /*
1588                          * Don't allocate a letter-number for swapper:0
1589                          * as a shortname. Instead, we use '.' for it.
1590                          */
1591                         tr->shortname[0] = '.';
1592                         tr->shortname[1] = ' ';
1593                 } else {
1594                         tr->shortname[0] = sched->next_shortname1;
1595                         tr->shortname[1] = sched->next_shortname2;
1596
1597                         if (sched->next_shortname1 < 'Z') {
1598                                 sched->next_shortname1++;
1599                         } else {
1600                                 sched->next_shortname1 = 'A';
1601                                 if (sched->next_shortname2 < '9')
1602                                         sched->next_shortname2++;
1603                                 else
1604                                         sched->next_shortname2 = '0';
1605                         }
1606                 }
1607                 new_shortname = 1;
1608         }
1609
1610         for (i = 0; i < cpus_nr; i++) {
1611                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1612                 struct thread *curr_thread = sched->curr_thread[cpu];
1613                 struct thread_runtime *curr_tr;
1614                 const char *pid_color = color;
1615                 const char *cpu_color = color;
1616
1617                 if (curr_thread && thread__has_color(curr_thread))
1618                         pid_color = COLOR_PIDS;
1619
1620                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1621                         continue;
1622
1623                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1624                         cpu_color = COLOR_CPUS;
1625
1626                 if (cpu != this_cpu)
1627                         color_fprintf(stdout, color, " ");
1628                 else
1629                         color_fprintf(stdout, cpu_color, "*");
1630
1631                 if (sched->curr_thread[cpu]) {
1632                         curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1633                         if (curr_tr == NULL) {
1634                                 thread__put(sched_in);
1635                                 return -1;
1636                         }
1637                         color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1638                 } else
1639                         color_fprintf(stdout, color, "   ");
1640         }
1641
1642         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1643                 goto out;
1644
1645         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1646         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1647         if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1648                 const char *pid_color = color;
1649
1650                 if (thread__has_color(sched_in))
1651                         pid_color = COLOR_PIDS;
1652
1653                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1654                        tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1655                 tr->comm_changed = false;
1656         }
1657
1658         if (sched->map.comp && new_cpu)
1659                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1660
1661 out:
1662         color_fprintf(stdout, color, "\n");
1663
1664         thread__put(sched_in);
1665
1666         return 0;
1667 }
1668
1669 static int process_sched_switch_event(struct perf_tool *tool,
1670                                       struct evsel *evsel,
1671                                       struct perf_sample *sample,
1672                                       struct machine *machine)
1673 {
1674         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1675         int this_cpu = sample->cpu, err = 0;
1676         u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1677             next_pid = evsel__intval(evsel, sample, "next_pid");
1678
1679         if (sched->curr_pid[this_cpu] != (u32)-1) {
1680                 /*
1681                  * Are we trying to switch away a PID that is
1682                  * not current?
1683                  */
1684                 if (sched->curr_pid[this_cpu] != prev_pid)
1685                         sched->nr_context_switch_bugs++;
1686         }
1687
1688         if (sched->tp_handler->switch_event)
1689                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1690
1691         sched->curr_pid[this_cpu] = next_pid;
1692         return err;
1693 }
1694
1695 static int process_sched_runtime_event(struct perf_tool *tool,
1696                                        struct evsel *evsel,
1697                                        struct perf_sample *sample,
1698                                        struct machine *machine)
1699 {
1700         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1701
1702         if (sched->tp_handler->runtime_event)
1703                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1704
1705         return 0;
1706 }
1707
1708 static int perf_sched__process_fork_event(struct perf_tool *tool,
1709                                           union perf_event *event,
1710                                           struct perf_sample *sample,
1711                                           struct machine *machine)
1712 {
1713         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1714
1715         /* run the fork event through the perf machineruy */
1716         perf_event__process_fork(tool, event, sample, machine);
1717
1718         /* and then run additional processing needed for this command */
1719         if (sched->tp_handler->fork_event)
1720                 return sched->tp_handler->fork_event(sched, event, machine);
1721
1722         return 0;
1723 }
1724
1725 static int process_sched_migrate_task_event(struct perf_tool *tool,
1726                                             struct evsel *evsel,
1727                                             struct perf_sample *sample,
1728                                             struct machine *machine)
1729 {
1730         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731
1732         if (sched->tp_handler->migrate_task_event)
1733                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1734
1735         return 0;
1736 }
1737
1738 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1739                                   struct evsel *evsel,
1740                                   struct perf_sample *sample,
1741                                   struct machine *machine);
1742
1743 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1744                                                  union perf_event *event __maybe_unused,
1745                                                  struct perf_sample *sample,
1746                                                  struct evsel *evsel,
1747                                                  struct machine *machine)
1748 {
1749         int err = 0;
1750
1751         if (evsel->handler != NULL) {
1752                 tracepoint_handler f = evsel->handler;
1753                 err = f(tool, evsel, sample, machine);
1754         }
1755
1756         return err;
1757 }
1758
1759 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1760                                     union perf_event *event,
1761                                     struct perf_sample *sample,
1762                                     struct machine *machine)
1763 {
1764         struct thread *thread;
1765         struct thread_runtime *tr;
1766         int err;
1767
1768         err = perf_event__process_comm(tool, event, sample, machine);
1769         if (err)
1770                 return err;
1771
1772         thread = machine__find_thread(machine, sample->pid, sample->tid);
1773         if (!thread) {
1774                 pr_err("Internal error: can't find thread\n");
1775                 return -1;
1776         }
1777
1778         tr = thread__get_runtime(thread);
1779         if (tr == NULL) {
1780                 thread__put(thread);
1781                 return -1;
1782         }
1783
1784         tr->comm_changed = true;
1785         thread__put(thread);
1786
1787         return 0;
1788 }
1789
1790 static int perf_sched__read_events(struct perf_sched *sched)
1791 {
1792         const struct evsel_str_handler handlers[] = {
1793                 { "sched:sched_switch",       process_sched_switch_event, },
1794                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1795                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1796                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1797                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1798         };
1799         struct perf_session *session;
1800         struct perf_data data = {
1801                 .path  = input_name,
1802                 .mode  = PERF_DATA_MODE_READ,
1803                 .force = sched->force,
1804         };
1805         int rc = -1;
1806
1807         session = perf_session__new(&data, false, &sched->tool);
1808         if (IS_ERR(session)) {
1809                 pr_debug("Error creating perf session");
1810                 return PTR_ERR(session);
1811         }
1812
1813         symbol__init(&session->header.env);
1814
1815         if (perf_session__set_tracepoints_handlers(session, handlers))
1816                 goto out_delete;
1817
1818         if (perf_session__has_traces(session, "record -R")) {
1819                 int err = perf_session__process_events(session);
1820                 if (err) {
1821                         pr_err("Failed to process events, error %d", err);
1822                         goto out_delete;
1823                 }
1824
1825                 sched->nr_events      = session->evlist->stats.nr_events[0];
1826                 sched->nr_lost_events = session->evlist->stats.total_lost;
1827                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1828         }
1829
1830         rc = 0;
1831 out_delete:
1832         perf_session__delete(session);
1833         return rc;
1834 }
1835
1836 /*
1837  * scheduling times are printed as msec.usec
1838  */
1839 static inline void print_sched_time(unsigned long long nsecs, int width)
1840 {
1841         unsigned long msecs;
1842         unsigned long usecs;
1843
1844         msecs  = nsecs / NSEC_PER_MSEC;
1845         nsecs -= msecs * NSEC_PER_MSEC;
1846         usecs  = nsecs / NSEC_PER_USEC;
1847         printf("%*lu.%03lu ", width, msecs, usecs);
1848 }
1849
1850 /*
1851  * returns runtime data for event, allocating memory for it the
1852  * first time it is used.
1853  */
1854 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1855 {
1856         struct evsel_runtime *r = evsel->priv;
1857
1858         if (r == NULL) {
1859                 r = zalloc(sizeof(struct evsel_runtime));
1860                 evsel->priv = r;
1861         }
1862
1863         return r;
1864 }
1865
1866 /*
1867  * save last time event was seen per cpu
1868  */
1869 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1870 {
1871         struct evsel_runtime *r = evsel__get_runtime(evsel);
1872
1873         if (r == NULL)
1874                 return;
1875
1876         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1877                 int i, n = __roundup_pow_of_two(cpu+1);
1878                 void *p = r->last_time;
1879
1880                 p = realloc(r->last_time, n * sizeof(u64));
1881                 if (!p)
1882                         return;
1883
1884                 r->last_time = p;
1885                 for (i = r->ncpu; i < n; ++i)
1886                         r->last_time[i] = (u64) 0;
1887
1888                 r->ncpu = n;
1889         }
1890
1891         r->last_time[cpu] = timestamp;
1892 }
1893
1894 /* returns last time this event was seen on the given cpu */
1895 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1896 {
1897         struct evsel_runtime *r = evsel__get_runtime(evsel);
1898
1899         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1900                 return 0;
1901
1902         return r->last_time[cpu];
1903 }
1904
1905 static int comm_width = 30;
1906
1907 static char *timehist_get_commstr(struct thread *thread)
1908 {
1909         static char str[32];
1910         const char *comm = thread__comm_str(thread);
1911         pid_t tid = thread->tid;
1912         pid_t pid = thread->pid_;
1913         int n;
1914
1915         if (pid == 0)
1916                 n = scnprintf(str, sizeof(str), "%s", comm);
1917
1918         else if (tid != pid)
1919                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1920
1921         else
1922                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1923
1924         if (n > comm_width)
1925                 comm_width = n;
1926
1927         return str;
1928 }
1929
1930 static void timehist_header(struct perf_sched *sched)
1931 {
1932         u32 ncpus = sched->max_cpu + 1;
1933         u32 i, j;
1934
1935         printf("%15s %6s ", "time", "cpu");
1936
1937         if (sched->show_cpu_visual) {
1938                 printf(" ");
1939                 for (i = 0, j = 0; i < ncpus; ++i) {
1940                         printf("%x", j++);
1941                         if (j > 15)
1942                                 j = 0;
1943                 }
1944                 printf(" ");
1945         }
1946
1947         printf(" %-*s  %9s  %9s  %9s", comm_width,
1948                 "task name", "wait time", "sch delay", "run time");
1949
1950         if (sched->show_state)
1951                 printf("  %s", "state");
1952
1953         printf("\n");
1954
1955         /*
1956          * units row
1957          */
1958         printf("%15s %-6s ", "", "");
1959
1960         if (sched->show_cpu_visual)
1961                 printf(" %*s ", ncpus, "");
1962
1963         printf(" %-*s  %9s  %9s  %9s", comm_width,
1964                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1965
1966         if (sched->show_state)
1967                 printf("  %5s", "");
1968
1969         printf("\n");
1970
1971         /*
1972          * separator
1973          */
1974         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1975
1976         if (sched->show_cpu_visual)
1977                 printf(" %.*s ", ncpus, graph_dotted_line);
1978
1979         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1980                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1981                 graph_dotted_line);
1982
1983         if (sched->show_state)
1984                 printf("  %.5s", graph_dotted_line);
1985
1986         printf("\n");
1987 }
1988
1989 static char task_state_char(struct thread *thread, int state)
1990 {
1991         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1992         unsigned bit = state ? ffs(state) : 0;
1993
1994         /* 'I' for idle */
1995         if (thread->tid == 0)
1996                 return 'I';
1997
1998         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1999 }
2000
2001 static void timehist_print_sample(struct perf_sched *sched,
2002                                   struct evsel *evsel,
2003                                   struct perf_sample *sample,
2004                                   struct addr_location *al,
2005                                   struct thread *thread,
2006                                   u64 t, int state)
2007 {
2008         struct thread_runtime *tr = thread__priv(thread);
2009         const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2010         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2011         u32 max_cpus = sched->max_cpu + 1;
2012         char tstr[64];
2013         char nstr[30];
2014         u64 wait_time;
2015
2016         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2017                 return;
2018
2019         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2020         printf("%15s [%04d] ", tstr, sample->cpu);
2021
2022         if (sched->show_cpu_visual) {
2023                 u32 i;
2024                 char c;
2025
2026                 printf(" ");
2027                 for (i = 0; i < max_cpus; ++i) {
2028                         /* flag idle times with 'i'; others are sched events */
2029                         if (i == sample->cpu)
2030                                 c = (thread->tid == 0) ? 'i' : 's';
2031                         else
2032                                 c = ' ';
2033                         printf("%c", c);
2034                 }
2035                 printf(" ");
2036         }
2037
2038         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2039
2040         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2041         print_sched_time(wait_time, 6);
2042
2043         print_sched_time(tr->dt_delay, 6);
2044         print_sched_time(tr->dt_run, 6);
2045
2046         if (sched->show_state)
2047                 printf(" %5c ", task_state_char(thread, state));
2048
2049         if (sched->show_next) {
2050                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2051                 printf(" %-*s", comm_width, nstr);
2052         }
2053
2054         if (sched->show_wakeups && !sched->show_next)
2055                 printf("  %-*s", comm_width, "");
2056
2057         if (thread->tid == 0)
2058                 goto out;
2059
2060         if (sched->show_callchain)
2061                 printf("  ");
2062
2063         sample__fprintf_sym(sample, al, 0,
2064                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2065                             EVSEL__PRINT_CALLCHAIN_ARROW |
2066                             EVSEL__PRINT_SKIP_IGNORED,
2067                             &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2068
2069 out:
2070         printf("\n");
2071 }
2072
2073 /*
2074  * Explanation of delta-time stats:
2075  *
2076  *            t = time of current schedule out event
2077  *        tprev = time of previous sched out event
2078  *                also time of schedule-in event for current task
2079  *    last_time = time of last sched change event for current task
2080  *                (i.e, time process was last scheduled out)
2081  * ready_to_run = time of wakeup for current task
2082  *
2083  * -----|------------|------------|------------|------
2084  *    last         ready        tprev          t
2085  *    time         to run
2086  *
2087  *      |-------- dt_wait --------|
2088  *                   |- dt_delay -|-- dt_run --|
2089  *
2090  *   dt_run = run time of current task
2091  *  dt_wait = time between last schedule out event for task and tprev
2092  *            represents time spent off the cpu
2093  * dt_delay = time between wakeup and schedule-in of task
2094  */
2095
2096 static void timehist_update_runtime_stats(struct thread_runtime *r,
2097                                          u64 t, u64 tprev)
2098 {
2099         r->dt_delay   = 0;
2100         r->dt_sleep   = 0;
2101         r->dt_iowait  = 0;
2102         r->dt_preempt = 0;
2103         r->dt_run     = 0;
2104
2105         if (tprev) {
2106                 r->dt_run = t - tprev;
2107                 if (r->ready_to_run) {
2108                         if (r->ready_to_run > tprev)
2109                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2110                         else
2111                                 r->dt_delay = tprev - r->ready_to_run;
2112                 }
2113
2114                 if (r->last_time > tprev)
2115                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2116                 else if (r->last_time) {
2117                         u64 dt_wait = tprev - r->last_time;
2118
2119                         if (r->last_state == TASK_RUNNING)
2120                                 r->dt_preempt = dt_wait;
2121                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2122                                 r->dt_iowait = dt_wait;
2123                         else
2124                                 r->dt_sleep = dt_wait;
2125                 }
2126         }
2127
2128         update_stats(&r->run_stats, r->dt_run);
2129
2130         r->total_run_time     += r->dt_run;
2131         r->total_delay_time   += r->dt_delay;
2132         r->total_sleep_time   += r->dt_sleep;
2133         r->total_iowait_time  += r->dt_iowait;
2134         r->total_preempt_time += r->dt_preempt;
2135 }
2136
2137 static bool is_idle_sample(struct perf_sample *sample,
2138                            struct evsel *evsel)
2139 {
2140         /* pid 0 == swapper == idle task */
2141         if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2142                 return evsel__intval(evsel, sample, "prev_pid") == 0;
2143
2144         return sample->pid == 0;
2145 }
2146
2147 static void save_task_callchain(struct perf_sched *sched,
2148                                 struct perf_sample *sample,
2149                                 struct evsel *evsel,
2150                                 struct machine *machine)
2151 {
2152         struct callchain_cursor *cursor = &callchain_cursor;
2153         struct thread *thread;
2154
2155         /* want main thread for process - has maps */
2156         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2157         if (thread == NULL) {
2158                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2159                 return;
2160         }
2161
2162         if (!sched->show_callchain || sample->callchain == NULL)
2163                 return;
2164
2165         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2166                                       NULL, NULL, sched->max_stack + 2) != 0) {
2167                 if (verbose > 0)
2168                         pr_err("Failed to resolve callchain. Skipping\n");
2169
2170                 return;
2171         }
2172
2173         callchain_cursor_commit(cursor);
2174
2175         while (true) {
2176                 struct callchain_cursor_node *node;
2177                 struct symbol *sym;
2178
2179                 node = callchain_cursor_current(cursor);
2180                 if (node == NULL)
2181                         break;
2182
2183                 sym = node->ms.sym;
2184                 if (sym) {
2185                         if (!strcmp(sym->name, "schedule") ||
2186                             !strcmp(sym->name, "__schedule") ||
2187                             !strcmp(sym->name, "preempt_schedule"))
2188                                 sym->ignore = 1;
2189                 }
2190
2191                 callchain_cursor_advance(cursor);
2192         }
2193 }
2194
2195 static int init_idle_thread(struct thread *thread)
2196 {
2197         struct idle_thread_runtime *itr;
2198
2199         thread__set_comm(thread, idle_comm, 0);
2200
2201         itr = zalloc(sizeof(*itr));
2202         if (itr == NULL)
2203                 return -ENOMEM;
2204
2205         init_stats(&itr->tr.run_stats);
2206         callchain_init(&itr->callchain);
2207         callchain_cursor_reset(&itr->cursor);
2208         thread__set_priv(thread, itr);
2209
2210         return 0;
2211 }
2212
2213 /*
2214  * Track idle stats per cpu by maintaining a local thread
2215  * struct for the idle task on each cpu.
2216  */
2217 static int init_idle_threads(int ncpu)
2218 {
2219         int i, ret;
2220
2221         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2222         if (!idle_threads)
2223                 return -ENOMEM;
2224
2225         idle_max_cpu = ncpu;
2226
2227         /* allocate the actual thread struct if needed */
2228         for (i = 0; i < ncpu; ++i) {
2229                 idle_threads[i] = thread__new(0, 0);
2230                 if (idle_threads[i] == NULL)
2231                         return -ENOMEM;
2232
2233                 ret = init_idle_thread(idle_threads[i]);
2234                 if (ret < 0)
2235                         return ret;
2236         }
2237
2238         return 0;
2239 }
2240
2241 static void free_idle_threads(void)
2242 {
2243         int i;
2244
2245         if (idle_threads == NULL)
2246                 return;
2247
2248         for (i = 0; i < idle_max_cpu; ++i) {
2249                 if ((idle_threads[i]))
2250                         thread__delete(idle_threads[i]);
2251         }
2252
2253         free(idle_threads);
2254 }
2255
2256 static struct thread *get_idle_thread(int cpu)
2257 {
2258         /*
2259          * expand/allocate array of pointers to local thread
2260          * structs if needed
2261          */
2262         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2263                 int i, j = __roundup_pow_of_two(cpu+1);
2264                 void *p;
2265
2266                 p = realloc(idle_threads, j * sizeof(struct thread *));
2267                 if (!p)
2268                         return NULL;
2269
2270                 idle_threads = (struct thread **) p;
2271                 for (i = idle_max_cpu; i < j; ++i)
2272                         idle_threads[i] = NULL;
2273
2274                 idle_max_cpu = j;
2275         }
2276
2277         /* allocate a new thread struct if needed */
2278         if (idle_threads[cpu] == NULL) {
2279                 idle_threads[cpu] = thread__new(0, 0);
2280
2281                 if (idle_threads[cpu]) {
2282                         if (init_idle_thread(idle_threads[cpu]) < 0)
2283                                 return NULL;
2284                 }
2285         }
2286
2287         return idle_threads[cpu];
2288 }
2289
2290 static void save_idle_callchain(struct perf_sched *sched,
2291                                 struct idle_thread_runtime *itr,
2292                                 struct perf_sample *sample)
2293 {
2294         if (!sched->show_callchain || sample->callchain == NULL)
2295                 return;
2296
2297         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2298 }
2299
2300 static struct thread *timehist_get_thread(struct perf_sched *sched,
2301                                           struct perf_sample *sample,
2302                                           struct machine *machine,
2303                                           struct evsel *evsel)
2304 {
2305         struct thread *thread;
2306
2307         if (is_idle_sample(sample, evsel)) {
2308                 thread = get_idle_thread(sample->cpu);
2309                 if (thread == NULL)
2310                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2311
2312         } else {
2313                 /* there were samples with tid 0 but non-zero pid */
2314                 thread = machine__findnew_thread(machine, sample->pid,
2315                                                  sample->tid ?: sample->pid);
2316                 if (thread == NULL) {
2317                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2318                                  sample->tid);
2319                 }
2320
2321                 save_task_callchain(sched, sample, evsel, machine);
2322                 if (sched->idle_hist) {
2323                         struct thread *idle;
2324                         struct idle_thread_runtime *itr;
2325
2326                         idle = get_idle_thread(sample->cpu);
2327                         if (idle == NULL) {
2328                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329                                 return NULL;
2330                         }
2331
2332                         itr = thread__priv(idle);
2333                         if (itr == NULL)
2334                                 return NULL;
2335
2336                         itr->last_thread = thread;
2337
2338                         /* copy task callchain when entering to idle */
2339                         if (evsel__intval(evsel, sample, "next_pid") == 0)
2340                                 save_idle_callchain(sched, itr, sample);
2341                 }
2342         }
2343
2344         return thread;
2345 }
2346
2347 static bool timehist_skip_sample(struct perf_sched *sched,
2348                                  struct thread *thread,
2349                                  struct evsel *evsel,
2350                                  struct perf_sample *sample)
2351 {
2352         bool rc = false;
2353
2354         if (thread__is_filtered(thread)) {
2355                 rc = true;
2356                 sched->skipped_samples++;
2357         }
2358
2359         if (sched->idle_hist) {
2360                 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2361                         rc = true;
2362                 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2363                          evsel__intval(evsel, sample, "next_pid") != 0)
2364                         rc = true;
2365         }
2366
2367         return rc;
2368 }
2369
2370 static void timehist_print_wakeup_event(struct perf_sched *sched,
2371                                         struct evsel *evsel,
2372                                         struct perf_sample *sample,
2373                                         struct machine *machine,
2374                                         struct thread *awakened)
2375 {
2376         struct thread *thread;
2377         char tstr[64];
2378
2379         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2380         if (thread == NULL)
2381                 return;
2382
2383         /* show wakeup unless both awakee and awaker are filtered */
2384         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2385             timehist_skip_sample(sched, awakened, evsel, sample)) {
2386                 return;
2387         }
2388
2389         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2390         printf("%15s [%04d] ", tstr, sample->cpu);
2391         if (sched->show_cpu_visual)
2392                 printf(" %*s ", sched->max_cpu + 1, "");
2393
2394         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2395
2396         /* dt spacer */
2397         printf("  %9s  %9s  %9s ", "", "", "");
2398
2399         printf("awakened: %s", timehist_get_commstr(awakened));
2400
2401         printf("\n");
2402 }
2403
2404 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2405                                         union perf_event *event __maybe_unused,
2406                                         struct evsel *evsel __maybe_unused,
2407                                         struct perf_sample *sample __maybe_unused,
2408                                         struct machine *machine __maybe_unused)
2409 {
2410         return 0;
2411 }
2412
2413 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2414                                        union perf_event *event __maybe_unused,
2415                                        struct evsel *evsel,
2416                                        struct perf_sample *sample,
2417                                        struct machine *machine)
2418 {
2419         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2420         struct thread *thread;
2421         struct thread_runtime *tr = NULL;
2422         /* want pid of awakened task not pid in sample */
2423         const u32 pid = evsel__intval(evsel, sample, "pid");
2424
2425         thread = machine__findnew_thread(machine, 0, pid);
2426         if (thread == NULL)
2427                 return -1;
2428
2429         tr = thread__get_runtime(thread);
2430         if (tr == NULL)
2431                 return -1;
2432
2433         if (tr->ready_to_run == 0)
2434                 tr->ready_to_run = sample->time;
2435
2436         /* show wakeups if requested */
2437         if (sched->show_wakeups &&
2438             !perf_time__skip_sample(&sched->ptime, sample->time))
2439                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2440
2441         return 0;
2442 }
2443
2444 static void timehist_print_migration_event(struct perf_sched *sched,
2445                                         struct evsel *evsel,
2446                                         struct perf_sample *sample,
2447                                         struct machine *machine,
2448                                         struct thread *migrated)
2449 {
2450         struct thread *thread;
2451         char tstr[64];
2452         u32 max_cpus = sched->max_cpu + 1;
2453         u32 ocpu, dcpu;
2454
2455         if (sched->summary_only)
2456                 return;
2457
2458         max_cpus = sched->max_cpu + 1;
2459         ocpu = evsel__intval(evsel, sample, "orig_cpu");
2460         dcpu = evsel__intval(evsel, sample, "dest_cpu");
2461
2462         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2463         if (thread == NULL)
2464                 return;
2465
2466         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2467             timehist_skip_sample(sched, migrated, evsel, sample)) {
2468                 return;
2469         }
2470
2471         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2472         printf("%15s [%04d] ", tstr, sample->cpu);
2473
2474         if (sched->show_cpu_visual) {
2475                 u32 i;
2476                 char c;
2477
2478                 printf("  ");
2479                 for (i = 0; i < max_cpus; ++i) {
2480                         c = (i == sample->cpu) ? 'm' : ' ';
2481                         printf("%c", c);
2482                 }
2483                 printf("  ");
2484         }
2485
2486         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2487
2488         /* dt spacer */
2489         printf("  %9s  %9s  %9s ", "", "", "");
2490
2491         printf("migrated: %s", timehist_get_commstr(migrated));
2492         printf(" cpu %d => %d", ocpu, dcpu);
2493
2494         printf("\n");
2495 }
2496
2497 static int timehist_migrate_task_event(struct perf_tool *tool,
2498                                        union perf_event *event __maybe_unused,
2499                                        struct evsel *evsel,
2500                                        struct perf_sample *sample,
2501                                        struct machine *machine)
2502 {
2503         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2504         struct thread *thread;
2505         struct thread_runtime *tr = NULL;
2506         /* want pid of migrated task not pid in sample */
2507         const u32 pid = evsel__intval(evsel, sample, "pid");
2508
2509         thread = machine__findnew_thread(machine, 0, pid);
2510         if (thread == NULL)
2511                 return -1;
2512
2513         tr = thread__get_runtime(thread);
2514         if (tr == NULL)
2515                 return -1;
2516
2517         tr->migrations++;
2518
2519         /* show migrations if requested */
2520         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2521
2522         return 0;
2523 }
2524
2525 static int timehist_sched_change_event(struct perf_tool *tool,
2526                                        union perf_event *event,
2527                                        struct evsel *evsel,
2528                                        struct perf_sample *sample,
2529                                        struct machine *machine)
2530 {
2531         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2532         struct perf_time_interval *ptime = &sched->ptime;
2533         struct addr_location al;
2534         struct thread *thread;
2535         struct thread_runtime *tr = NULL;
2536         u64 tprev, t = sample->time;
2537         int rc = 0;
2538         int state = evsel__intval(evsel, sample, "prev_state");
2539
2540         if (machine__resolve(machine, &al, sample) < 0) {
2541                 pr_err("problem processing %d event. skipping it\n",
2542                        event->header.type);
2543                 rc = -1;
2544                 goto out;
2545         }
2546
2547         thread = timehist_get_thread(sched, sample, machine, evsel);
2548         if (thread == NULL) {
2549                 rc = -1;
2550                 goto out;
2551         }
2552
2553         if (timehist_skip_sample(sched, thread, evsel, sample))
2554                 goto out;
2555
2556         tr = thread__get_runtime(thread);
2557         if (tr == NULL) {
2558                 rc = -1;
2559                 goto out;
2560         }
2561
2562         tprev = evsel__get_time(evsel, sample->cpu);
2563
2564         /*
2565          * If start time given:
2566          * - sample time is under window user cares about - skip sample
2567          * - tprev is under window user cares about  - reset to start of window
2568          */
2569         if (ptime->start && ptime->start > t)
2570                 goto out;
2571
2572         if (tprev && ptime->start > tprev)
2573                 tprev = ptime->start;
2574
2575         /*
2576          * If end time given:
2577          * - previous sched event is out of window - we are done
2578          * - sample time is beyond window user cares about - reset it
2579          *   to close out stats for time window interest
2580          */
2581         if (ptime->end) {
2582                 if (tprev > ptime->end)
2583                         goto out;
2584
2585                 if (t > ptime->end)
2586                         t = ptime->end;
2587         }
2588
2589         if (!sched->idle_hist || thread->tid == 0) {
2590                 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2591                         timehist_update_runtime_stats(tr, t, tprev);
2592
2593                 if (sched->idle_hist) {
2594                         struct idle_thread_runtime *itr = (void *)tr;
2595                         struct thread_runtime *last_tr;
2596
2597                         BUG_ON(thread->tid != 0);
2598
2599                         if (itr->last_thread == NULL)
2600                                 goto out;
2601
2602                         /* add current idle time as last thread's runtime */
2603                         last_tr = thread__get_runtime(itr->last_thread);
2604                         if (last_tr == NULL)
2605                                 goto out;
2606
2607                         timehist_update_runtime_stats(last_tr, t, tprev);
2608                         /*
2609                          * remove delta time of last thread as it's not updated
2610                          * and otherwise it will show an invalid value next
2611                          * time.  we only care total run time and run stat.
2612                          */
2613                         last_tr->dt_run = 0;
2614                         last_tr->dt_delay = 0;
2615                         last_tr->dt_sleep = 0;
2616                         last_tr->dt_iowait = 0;
2617                         last_tr->dt_preempt = 0;
2618
2619                         if (itr->cursor.nr)
2620                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2621
2622                         itr->last_thread = NULL;
2623                 }
2624         }
2625
2626         if (!sched->summary_only)
2627                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2628
2629 out:
2630         if (sched->hist_time.start == 0 && t >= ptime->start)
2631                 sched->hist_time.start = t;
2632         if (ptime->end == 0 || t <= ptime->end)
2633                 sched->hist_time.end = t;
2634
2635         if (tr) {
2636                 /* time of this sched_switch event becomes last time task seen */
2637                 tr->last_time = sample->time;
2638
2639                 /* last state is used to determine where to account wait time */
2640                 tr->last_state = state;
2641
2642                 /* sched out event for task so reset ready to run time */
2643                 tr->ready_to_run = 0;
2644         }
2645
2646         evsel__save_time(evsel, sample->time, sample->cpu);
2647
2648         return rc;
2649 }
2650
2651 static int timehist_sched_switch_event(struct perf_tool *tool,
2652                              union perf_event *event,
2653                              struct evsel *evsel,
2654                              struct perf_sample *sample,
2655                              struct machine *machine __maybe_unused)
2656 {
2657         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2658 }
2659
2660 static int process_lost(struct perf_tool *tool __maybe_unused,
2661                         union perf_event *event,
2662                         struct perf_sample *sample,
2663                         struct machine *machine __maybe_unused)
2664 {
2665         char tstr[64];
2666
2667         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2668         printf("%15s ", tstr);
2669         printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2670
2671         return 0;
2672 }
2673
2674
2675 static void print_thread_runtime(struct thread *t,
2676                                  struct thread_runtime *r)
2677 {
2678         double mean = avg_stats(&r->run_stats);
2679         float stddev;
2680
2681         printf("%*s   %5d  %9" PRIu64 " ",
2682                comm_width, timehist_get_commstr(t), t->ppid,
2683                (u64) r->run_stats.n);
2684
2685         print_sched_time(r->total_run_time, 8);
2686         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2687         print_sched_time(r->run_stats.min, 6);
2688         printf(" ");
2689         print_sched_time((u64) mean, 6);
2690         printf(" ");
2691         print_sched_time(r->run_stats.max, 6);
2692         printf("  ");
2693         printf("%5.2f", stddev);
2694         printf("   %5" PRIu64, r->migrations);
2695         printf("\n");
2696 }
2697
2698 static void print_thread_waittime(struct thread *t,
2699                                   struct thread_runtime *r)
2700 {
2701         printf("%*s   %5d  %9" PRIu64 " ",
2702                comm_width, timehist_get_commstr(t), t->ppid,
2703                (u64) r->run_stats.n);
2704
2705         print_sched_time(r->total_run_time, 8);
2706         print_sched_time(r->total_sleep_time, 6);
2707         printf(" ");
2708         print_sched_time(r->total_iowait_time, 6);
2709         printf(" ");
2710         print_sched_time(r->total_preempt_time, 6);
2711         printf(" ");
2712         print_sched_time(r->total_delay_time, 6);
2713         printf("\n");
2714 }
2715
2716 struct total_run_stats {
2717         struct perf_sched *sched;
2718         u64  sched_count;
2719         u64  task_count;
2720         u64  total_run_time;
2721 };
2722
2723 static int __show_thread_runtime(struct thread *t, void *priv)
2724 {
2725         struct total_run_stats *stats = priv;
2726         struct thread_runtime *r;
2727
2728         if (thread__is_filtered(t))
2729                 return 0;
2730
2731         r = thread__priv(t);
2732         if (r && r->run_stats.n) {
2733                 stats->task_count++;
2734                 stats->sched_count += r->run_stats.n;
2735                 stats->total_run_time += r->total_run_time;
2736
2737                 if (stats->sched->show_state)
2738                         print_thread_waittime(t, r);
2739                 else
2740                         print_thread_runtime(t, r);
2741         }
2742
2743         return 0;
2744 }
2745
2746 static int show_thread_runtime(struct thread *t, void *priv)
2747 {
2748         if (t->dead)
2749                 return 0;
2750
2751         return __show_thread_runtime(t, priv);
2752 }
2753
2754 static int show_deadthread_runtime(struct thread *t, void *priv)
2755 {
2756         if (!t->dead)
2757                 return 0;
2758
2759         return __show_thread_runtime(t, priv);
2760 }
2761
2762 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2763 {
2764         const char *sep = " <- ";
2765         struct callchain_list *chain;
2766         size_t ret = 0;
2767         char bf[1024];
2768         bool first;
2769
2770         if (node == NULL)
2771                 return 0;
2772
2773         ret = callchain__fprintf_folded(fp, node->parent);
2774         first = (ret == 0);
2775
2776         list_for_each_entry(chain, &node->val, list) {
2777                 if (chain->ip >= PERF_CONTEXT_MAX)
2778                         continue;
2779                 if (chain->ms.sym && chain->ms.sym->ignore)
2780                         continue;
2781                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2782                                callchain_list__sym_name(chain, bf, sizeof(bf),
2783                                                         false));
2784                 first = false;
2785         }
2786
2787         return ret;
2788 }
2789
2790 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2791 {
2792         size_t ret = 0;
2793         FILE *fp = stdout;
2794         struct callchain_node *chain;
2795         struct rb_node *rb_node = rb_first_cached(root);
2796
2797         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2798         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2799                graph_dotted_line);
2800
2801         while (rb_node) {
2802                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2803                 rb_node = rb_next(rb_node);
2804
2805                 ret += fprintf(fp, "  ");
2806                 print_sched_time(chain->hit, 12);
2807                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2808                 ret += fprintf(fp, " %8d  ", chain->count);
2809                 ret += callchain__fprintf_folded(fp, chain);
2810                 ret += fprintf(fp, "\n");
2811         }
2812
2813         return ret;
2814 }
2815
2816 static void timehist_print_summary(struct perf_sched *sched,
2817                                    struct perf_session *session)
2818 {
2819         struct machine *m = &session->machines.host;
2820         struct total_run_stats totals;
2821         u64 task_count;
2822         struct thread *t;
2823         struct thread_runtime *r;
2824         int i;
2825         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2826
2827         memset(&totals, 0, sizeof(totals));
2828         totals.sched = sched;
2829
2830         if (sched->idle_hist) {
2831                 printf("\nIdle-time summary\n");
2832                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2833                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2834         } else if (sched->show_state) {
2835                 printf("\nWait-time summary\n");
2836                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2837                 printf("   run-time      sleep      iowait     preempt       delay\n");
2838         } else {
2839                 printf("\nRuntime summary\n");
2840                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2841                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2842         }
2843         printf("%*s            (count)  ", comm_width, "");
2844         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2845                sched->show_state ? "(msec)" : "%");
2846         printf("%.117s\n", graph_dotted_line);
2847
2848         machine__for_each_thread(m, show_thread_runtime, &totals);
2849         task_count = totals.task_count;
2850         if (!task_count)
2851                 printf("<no still running tasks>\n");
2852
2853         printf("\nTerminated tasks:\n");
2854         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2855         if (task_count == totals.task_count)
2856                 printf("<no terminated tasks>\n");
2857
2858         /* CPU idle stats not tracked when samples were skipped */
2859         if (sched->skipped_samples && !sched->idle_hist)
2860                 return;
2861
2862         printf("\nIdle stats:\n");
2863         for (i = 0; i < idle_max_cpu; ++i) {
2864                 if (cpu_list && !test_bit(i, cpu_bitmap))
2865                         continue;
2866
2867                 t = idle_threads[i];
2868                 if (!t)
2869                         continue;
2870
2871                 r = thread__priv(t);
2872                 if (r && r->run_stats.n) {
2873                         totals.sched_count += r->run_stats.n;
2874                         printf("    CPU %2d idle for ", i);
2875                         print_sched_time(r->total_run_time, 6);
2876                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2877                 } else
2878                         printf("    CPU %2d idle entire time window\n", i);
2879         }
2880
2881         if (sched->idle_hist && sched->show_callchain) {
2882                 callchain_param.mode  = CHAIN_FOLDED;
2883                 callchain_param.value = CCVAL_PERIOD;
2884
2885                 callchain_register_param(&callchain_param);
2886
2887                 printf("\nIdle stats by callchain:\n");
2888                 for (i = 0; i < idle_max_cpu; ++i) {
2889                         struct idle_thread_runtime *itr;
2890
2891                         t = idle_threads[i];
2892                         if (!t)
2893                                 continue;
2894
2895                         itr = thread__priv(t);
2896                         if (itr == NULL)
2897                                 continue;
2898
2899                         callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2900                                              0, &callchain_param);
2901
2902                         printf("  CPU %2d:", i);
2903                         print_sched_time(itr->tr.total_run_time, 6);
2904                         printf(" msec\n");
2905                         timehist_print_idlehist_callchain(&itr->sorted_root);
2906                         printf("\n");
2907                 }
2908         }
2909
2910         printf("\n"
2911                "    Total number of unique tasks: %" PRIu64 "\n"
2912                "Total number of context switches: %" PRIu64 "\n",
2913                totals.task_count, totals.sched_count);
2914
2915         printf("           Total run time (msec): ");
2916         print_sched_time(totals.total_run_time, 2);
2917         printf("\n");
2918
2919         printf("    Total scheduling time (msec): ");
2920         print_sched_time(hist_time, 2);
2921         printf(" (x %d)\n", sched->max_cpu);
2922 }
2923
2924 typedef int (*sched_handler)(struct perf_tool *tool,
2925                           union perf_event *event,
2926                           struct evsel *evsel,
2927                           struct perf_sample *sample,
2928                           struct machine *machine);
2929
2930 static int perf_timehist__process_sample(struct perf_tool *tool,
2931                                          union perf_event *event,
2932                                          struct perf_sample *sample,
2933                                          struct evsel *evsel,
2934                                          struct machine *machine)
2935 {
2936         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2937         int err = 0;
2938         int this_cpu = sample->cpu;
2939
2940         if (this_cpu > sched->max_cpu)
2941                 sched->max_cpu = this_cpu;
2942
2943         if (evsel->handler != NULL) {
2944                 sched_handler f = evsel->handler;
2945
2946                 err = f(tool, event, evsel, sample, machine);
2947         }
2948
2949         return err;
2950 }
2951
2952 static int timehist_check_attr(struct perf_sched *sched,
2953                                struct evlist *evlist)
2954 {
2955         struct evsel *evsel;
2956         struct evsel_runtime *er;
2957
2958         list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2959                 er = evsel__get_runtime(evsel);
2960                 if (er == NULL) {
2961                         pr_err("Failed to allocate memory for evsel runtime data\n");
2962                         return -1;
2963                 }
2964
2965                 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2966                         pr_info("Samples do not have callchains.\n");
2967                         sched->show_callchain = 0;
2968                         symbol_conf.use_callchain = 0;
2969                 }
2970         }
2971
2972         return 0;
2973 }
2974
2975 static int perf_sched__timehist(struct perf_sched *sched)
2976 {
2977         struct evsel_str_handler handlers[] = {
2978                 { "sched:sched_switch",       timehist_sched_switch_event, },
2979                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2980                 { "sched:sched_waking",       timehist_sched_wakeup_event, },
2981                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2982         };
2983         const struct evsel_str_handler migrate_handlers[] = {
2984                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2985         };
2986         struct perf_data data = {
2987                 .path  = input_name,
2988                 .mode  = PERF_DATA_MODE_READ,
2989                 .force = sched->force,
2990         };
2991
2992         struct perf_session *session;
2993         struct evlist *evlist;
2994         int err = -1;
2995
2996         /*
2997          * event handlers for timehist option
2998          */
2999         sched->tool.sample       = perf_timehist__process_sample;
3000         sched->tool.mmap         = perf_event__process_mmap;
3001         sched->tool.comm         = perf_event__process_comm;
3002         sched->tool.exit         = perf_event__process_exit;
3003         sched->tool.fork         = perf_event__process_fork;
3004         sched->tool.lost         = process_lost;
3005         sched->tool.attr         = perf_event__process_attr;
3006         sched->tool.tracing_data = perf_event__process_tracing_data;
3007         sched->tool.build_id     = perf_event__process_build_id;
3008
3009         sched->tool.ordered_events = true;
3010         sched->tool.ordering_requires_timestamps = true;
3011
3012         symbol_conf.use_callchain = sched->show_callchain;
3013
3014         session = perf_session__new(&data, false, &sched->tool);
3015         if (IS_ERR(session))
3016                 return PTR_ERR(session);
3017
3018         if (cpu_list) {
3019                 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3020                 if (err < 0)
3021                         goto out;
3022         }
3023
3024         evlist = session->evlist;
3025
3026         symbol__init(&session->header.env);
3027
3028         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3029                 pr_err("Invalid time string\n");
3030                 return -EINVAL;
3031         }
3032
3033         if (timehist_check_attr(sched, evlist) != 0)
3034                 goto out;
3035
3036         setup_pager();
3037
3038         /* prefer sched_waking if it is captured */
3039         if (perf_evlist__find_tracepoint_by_name(session->evlist,
3040                                                   "sched:sched_waking"))
3041                 handlers[1].handler = timehist_sched_wakeup_ignore;
3042
3043         /* setup per-evsel handlers */
3044         if (perf_session__set_tracepoints_handlers(session, handlers))
3045                 goto out;
3046
3047         /* sched_switch event at a minimum needs to exist */
3048         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3049                                                   "sched:sched_switch")) {
3050                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3051                 goto out;
3052         }
3053
3054         if (sched->show_migrations &&
3055             perf_session__set_tracepoints_handlers(session, migrate_handlers))
3056                 goto out;
3057
3058         /* pre-allocate struct for per-CPU idle stats */
3059         sched->max_cpu = session->header.env.nr_cpus_online;
3060         if (sched->max_cpu == 0)
3061                 sched->max_cpu = 4;
3062         if (init_idle_threads(sched->max_cpu))
3063                 goto out;
3064
3065         /* summary_only implies summary option, but don't overwrite summary if set */
3066         if (sched->summary_only)
3067                 sched->summary = sched->summary_only;
3068
3069         if (!sched->summary_only)
3070                 timehist_header(sched);
3071
3072         err = perf_session__process_events(session);
3073         if (err) {
3074                 pr_err("Failed to process events, error %d", err);
3075                 goto out;
3076         }
3077
3078         sched->nr_events      = evlist->stats.nr_events[0];
3079         sched->nr_lost_events = evlist->stats.total_lost;
3080         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3081
3082         if (sched->summary)
3083                 timehist_print_summary(sched, session);
3084
3085 out:
3086         free_idle_threads();
3087         perf_session__delete(session);
3088
3089         return err;
3090 }
3091
3092
3093 static void print_bad_events(struct perf_sched *sched)
3094 {
3095         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3096                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3097                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3098                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3099         }
3100         if (sched->nr_lost_events && sched->nr_events) {
3101                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3102                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3103                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3104         }
3105         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3106                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3107                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3108                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3109                 if (sched->nr_lost_events)
3110                         printf(" (due to lost events?)");
3111                 printf("\n");
3112         }
3113 }
3114
3115 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3116 {
3117         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3118         struct work_atoms *this;
3119         const char *comm = thread__comm_str(data->thread), *this_comm;
3120         bool leftmost = true;
3121
3122         while (*new) {
3123                 int cmp;
3124
3125                 this = container_of(*new, struct work_atoms, node);
3126                 parent = *new;
3127
3128                 this_comm = thread__comm_str(this->thread);
3129                 cmp = strcmp(comm, this_comm);
3130                 if (cmp > 0) {
3131                         new = &((*new)->rb_left);
3132                 } else if (cmp < 0) {
3133                         new = &((*new)->rb_right);
3134                         leftmost = false;
3135                 } else {
3136                         this->num_merged++;
3137                         this->total_runtime += data->total_runtime;
3138                         this->nb_atoms += data->nb_atoms;
3139                         this->total_lat += data->total_lat;
3140                         list_splice(&data->work_list, &this->work_list);
3141                         if (this->max_lat < data->max_lat) {
3142                                 this->max_lat = data->max_lat;
3143                                 this->max_lat_start = data->max_lat_start;
3144                                 this->max_lat_end = data->max_lat_end;
3145                         }
3146                         zfree(&data);
3147                         return;
3148                 }
3149         }
3150
3151         data->num_merged++;
3152         rb_link_node(&data->node, parent, new);
3153         rb_insert_color_cached(&data->node, root, leftmost);
3154 }
3155
3156 static void perf_sched__merge_lat(struct perf_sched *sched)
3157 {
3158         struct work_atoms *data;
3159         struct rb_node *node;
3160
3161         if (sched->skip_merge)
3162                 return;
3163
3164         while ((node = rb_first_cached(&sched->atom_root))) {
3165                 rb_erase_cached(node, &sched->atom_root);
3166                 data = rb_entry(node, struct work_atoms, node);
3167                 __merge_work_atoms(&sched->merged_atom_root, data);
3168         }
3169 }
3170
3171 static int perf_sched__lat(struct perf_sched *sched)
3172 {
3173         struct rb_node *next;
3174
3175         setup_pager();
3176
3177         if (perf_sched__read_events(sched))
3178                 return -1;
3179
3180         perf_sched__merge_lat(sched);
3181         perf_sched__sort_lat(sched);
3182
3183         printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3184         printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3185         printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3186
3187         next = rb_first_cached(&sched->sorted_atom_root);
3188
3189         while (next) {
3190                 struct work_atoms *work_list;
3191
3192                 work_list = rb_entry(next, struct work_atoms, node);
3193                 output_lat_thread(sched, work_list);
3194                 next = rb_next(next);
3195                 thread__zput(work_list->thread);
3196         }
3197
3198         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3199         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3200                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3201
3202         printf(" ---------------------------------------------------\n");
3203
3204         print_bad_events(sched);
3205         printf("\n");
3206
3207         return 0;
3208 }
3209
3210 static int setup_map_cpus(struct perf_sched *sched)
3211 {
3212         struct perf_cpu_map *map;
3213
3214         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3215
3216         if (sched->map.comp) {
3217                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3218                 if (!sched->map.comp_cpus)
3219                         return -1;
3220         }
3221
3222         if (!sched->map.cpus_str)
3223                 return 0;
3224
3225         map = perf_cpu_map__new(sched->map.cpus_str);
3226         if (!map) {
3227                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3228                 return -1;
3229         }
3230
3231         sched->map.cpus = map;
3232         return 0;
3233 }
3234
3235 static int setup_color_pids(struct perf_sched *sched)
3236 {
3237         struct perf_thread_map *map;
3238
3239         if (!sched->map.color_pids_str)
3240                 return 0;
3241
3242         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3243         if (!map) {
3244                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3245                 return -1;
3246         }
3247
3248         sched->map.color_pids = map;
3249         return 0;
3250 }
3251
3252 static int setup_color_cpus(struct perf_sched *sched)
3253 {
3254         struct perf_cpu_map *map;
3255
3256         if (!sched->map.color_cpus_str)
3257                 return 0;
3258
3259         map = perf_cpu_map__new(sched->map.color_cpus_str);
3260         if (!map) {
3261                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3262                 return -1;
3263         }
3264
3265         sched->map.color_cpus = map;
3266         return 0;
3267 }
3268
3269 static int perf_sched__map(struct perf_sched *sched)
3270 {
3271         if (setup_map_cpus(sched))
3272                 return -1;
3273
3274         if (setup_color_pids(sched))
3275                 return -1;
3276
3277         if (setup_color_cpus(sched))
3278                 return -1;
3279
3280         setup_pager();
3281         if (perf_sched__read_events(sched))
3282                 return -1;
3283         print_bad_events(sched);
3284         return 0;
3285 }
3286
3287 static int perf_sched__replay(struct perf_sched *sched)
3288 {
3289         unsigned long i;
3290
3291         calibrate_run_measurement_overhead(sched);
3292         calibrate_sleep_measurement_overhead(sched);
3293
3294         test_calibrations(sched);
3295
3296         if (perf_sched__read_events(sched))
3297                 return -1;
3298
3299         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3300         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3301         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3302
3303         if (sched->targetless_wakeups)
3304                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3305         if (sched->multitarget_wakeups)
3306                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3307         if (sched->nr_run_events_optimized)
3308                 printf("run atoms optimized: %ld\n",
3309                         sched->nr_run_events_optimized);
3310
3311         print_task_traces(sched);
3312         add_cross_task_wakeups(sched);
3313
3314         create_tasks(sched);
3315         printf("------------------------------------------------------------\n");
3316         for (i = 0; i < sched->replay_repeat; i++)
3317                 run_one_test(sched);
3318
3319         return 0;
3320 }
3321
3322 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3323                           const char * const usage_msg[])
3324 {
3325         char *tmp, *tok, *str = strdup(sched->sort_order);
3326
3327         for (tok = strtok_r(str, ", ", &tmp);
3328                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3329                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3330                         usage_with_options_msg(usage_msg, options,
3331                                         "Unknown --sort key: `%s'", tok);
3332                 }
3333         }
3334
3335         free(str);
3336
3337         sort_dimension__add("pid", &sched->cmp_pid);
3338 }
3339
3340 static bool schedstat_events_exposed(void)
3341 {
3342         /*
3343          * Select "sched:sched_stat_wait" event to check
3344          * whether schedstat tracepoints are exposed.
3345          */
3346         return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3347                 false : true;
3348 }
3349
3350 static int __cmd_record(int argc, const char **argv)
3351 {
3352         unsigned int rec_argc, i, j;
3353         const char **rec_argv;
3354         const char * const record_args[] = {
3355                 "record",
3356                 "-a",
3357                 "-R",
3358                 "-m", "1024",
3359                 "-c", "1",
3360                 "-e", "sched:sched_switch",
3361                 "-e", "sched:sched_stat_runtime",
3362                 "-e", "sched:sched_process_fork",
3363                 "-e", "sched:sched_wakeup_new",
3364                 "-e", "sched:sched_migrate_task",
3365         };
3366
3367         /*
3368          * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3369          * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3370          * to prevent "perf sched record" execution failure, determine
3371          * whether to record schedstat events according to actual situation.
3372          */
3373         const char * const schedstat_args[] = {
3374                 "-e", "sched:sched_stat_wait",
3375                 "-e", "sched:sched_stat_sleep",
3376                 "-e", "sched:sched_stat_iowait",
3377         };
3378         unsigned int schedstat_argc = schedstat_events_exposed() ?
3379                 ARRAY_SIZE(schedstat_args) : 0;
3380
3381         struct tep_event *waking_event;
3382
3383         /*
3384          * +2 for either "-e", "sched:sched_wakeup" or
3385          * "-e", "sched:sched_waking"
3386          */
3387         rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3388         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3389
3390         if (rec_argv == NULL)
3391                 return -ENOMEM;
3392
3393         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3394                 rec_argv[i] = strdup(record_args[i]);
3395
3396         rec_argv[i++] = "-e";
3397         waking_event = trace_event__tp_format("sched", "sched_waking");
3398         if (!IS_ERR(waking_event))
3399                 rec_argv[i++] = strdup("sched:sched_waking");
3400         else
3401                 rec_argv[i++] = strdup("sched:sched_wakeup");
3402
3403         for (j = 0; j < schedstat_argc; j++)
3404                 rec_argv[i++] = strdup(schedstat_args[j]);
3405
3406         for (j = 1; j < (unsigned int)argc; j++, i++)
3407                 rec_argv[i] = argv[j];
3408
3409         BUG_ON(i != rec_argc);
3410
3411         return cmd_record(i, rec_argv);
3412 }
3413
3414 int cmd_sched(int argc, const char **argv)
3415 {
3416         static const char default_sort_order[] = "avg, max, switch, runtime";
3417         struct perf_sched sched = {
3418                 .tool = {
3419                         .sample          = perf_sched__process_tracepoint_sample,
3420                         .comm            = perf_sched__process_comm,
3421                         .namespaces      = perf_event__process_namespaces,
3422                         .lost            = perf_event__process_lost,
3423                         .fork            = perf_sched__process_fork_event,
3424                         .ordered_events = true,
3425                 },
3426                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3427                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3428                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3429                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3430                 .sort_order           = default_sort_order,
3431                 .replay_repeat        = 10,
3432                 .profile_cpu          = -1,
3433                 .next_shortname1      = 'A',
3434                 .next_shortname2      = '0',
3435                 .skip_merge           = 0,
3436                 .show_callchain       = 1,
3437                 .max_stack            = 5,
3438         };
3439         const struct option sched_options[] = {
3440         OPT_STRING('i', "input", &input_name, "file",
3441                     "input file name"),
3442         OPT_INCR('v', "verbose", &verbose,
3443                     "be more verbose (show symbol address, etc)"),
3444         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3445                     "dump raw trace in ASCII"),
3446         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3447         OPT_END()
3448         };
3449         const struct option latency_options[] = {
3450         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3451                    "sort by key(s): runtime, switch, avg, max"),
3452         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3453                     "CPU to profile on"),
3454         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3455                     "latency stats per pid instead of per comm"),
3456         OPT_PARENT(sched_options)
3457         };
3458         const struct option replay_options[] = {
3459         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3460                      "repeat the workload replay N times (-1: infinite)"),
3461         OPT_PARENT(sched_options)
3462         };
3463         const struct option map_options[] = {
3464         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3465                     "map output in compact mode"),
3466         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3467                    "highlight given pids in map"),
3468         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3469                     "highlight given CPUs in map"),
3470         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3471                     "display given CPUs in map"),
3472         OPT_PARENT(sched_options)
3473         };
3474         const struct option timehist_options[] = {
3475         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3476                    "file", "vmlinux pathname"),
3477         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3478                    "file", "kallsyms pathname"),
3479         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3480                     "Display call chains if present (default on)"),
3481         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3482                    "Maximum number of functions to display backtrace."),
3483         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3484                     "Look for files with symbols relative to this directory"),
3485         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3486                     "Show only syscall summary with statistics"),
3487         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3488                     "Show all syscalls and summary with statistics"),
3489         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3490         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3491         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3492         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3493         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3494         OPT_STRING(0, "time", &sched.time_str, "str",
3495                    "Time span for analysis (start,stop)"),
3496         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3497         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3498                    "analyze events only for given process id(s)"),
3499         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3500                    "analyze events only for given thread id(s)"),
3501         OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3502         OPT_PARENT(sched_options)
3503         };
3504
3505         const char * const latency_usage[] = {
3506                 "perf sched latency [<options>]",
3507                 NULL
3508         };
3509         const char * const replay_usage[] = {
3510                 "perf sched replay [<options>]",
3511                 NULL
3512         };
3513         const char * const map_usage[] = {
3514                 "perf sched map [<options>]",
3515                 NULL
3516         };
3517         const char * const timehist_usage[] = {
3518                 "perf sched timehist [<options>]",
3519                 NULL
3520         };
3521         const char *const sched_subcommands[] = { "record", "latency", "map",
3522                                                   "replay", "script",
3523                                                   "timehist", NULL };
3524         const char *sched_usage[] = {
3525                 NULL,
3526                 NULL
3527         };
3528         struct trace_sched_handler lat_ops  = {
3529                 .wakeup_event       = latency_wakeup_event,
3530                 .switch_event       = latency_switch_event,
3531                 .runtime_event      = latency_runtime_event,
3532                 .migrate_task_event = latency_migrate_task_event,
3533         };
3534         struct trace_sched_handler map_ops  = {
3535                 .switch_event       = map_switch_event,
3536         };
3537         struct trace_sched_handler replay_ops  = {
3538                 .wakeup_event       = replay_wakeup_event,
3539                 .switch_event       = replay_switch_event,
3540                 .fork_event         = replay_fork_event,
3541         };
3542         unsigned int i;
3543
3544         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3545                 sched.curr_pid[i] = -1;
3546
3547         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3548                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3549         if (!argc)
3550                 usage_with_options(sched_usage, sched_options);
3551
3552         /*
3553          * Aliased to 'perf script' for now:
3554          */
3555         if (!strcmp(argv[0], "script"))
3556                 return cmd_script(argc, argv);
3557
3558         if (!strncmp(argv[0], "rec", 3)) {
3559                 return __cmd_record(argc, argv);
3560         } else if (!strncmp(argv[0], "lat", 3)) {
3561                 sched.tp_handler = &lat_ops;
3562                 if (argc > 1) {
3563                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3564                         if (argc)
3565                                 usage_with_options(latency_usage, latency_options);
3566                 }
3567                 setup_sorting(&sched, latency_options, latency_usage);
3568                 return perf_sched__lat(&sched);
3569         } else if (!strcmp(argv[0], "map")) {
3570                 if (argc) {
3571                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3572                         if (argc)
3573                                 usage_with_options(map_usage, map_options);
3574                 }
3575                 sched.tp_handler = &map_ops;
3576                 setup_sorting(&sched, latency_options, latency_usage);
3577                 return perf_sched__map(&sched);
3578         } else if (!strncmp(argv[0], "rep", 3)) {
3579                 sched.tp_handler = &replay_ops;
3580                 if (argc) {
3581                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3582                         if (argc)
3583                                 usage_with_options(replay_usage, replay_options);
3584                 }
3585                 return perf_sched__replay(&sched);
3586         } else if (!strcmp(argv[0], "timehist")) {
3587                 if (argc) {
3588                         argc = parse_options(argc, argv, timehist_options,
3589                                              timehist_usage, 0);
3590                         if (argc)
3591                                 usage_with_options(timehist_usage, timehist_options);
3592                 }
3593                 if ((sched.show_wakeups || sched.show_next) &&
3594                     sched.summary_only) {
3595                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3596                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3597                         if (sched.show_wakeups)
3598                                 parse_options_usage(NULL, timehist_options, "w", true);
3599                         if (sched.show_next)
3600                                 parse_options_usage(NULL, timehist_options, "n", true);
3601                         return -EINVAL;
3602                 }
3603
3604                 return perf_sched__timehist(&sched);
3605         } else {
3606                 usage_with_options(sched_usage, sched_options);
3607         }
3608
3609         return 0;
3610 }