tools/perf/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/evlist.h"
   7 #include "util/evsel.h"
   8 #include "util/symbol.h"
   9 #include "util/thread.h"
  10 #include "util/header.h"
  11 #include "util/session.h"
  12 #include "util/tool.h"
  13 #include "util/cloexec.h"
  14 #include "util/thread_map.h"
  15 #include "util/color.h"
  16 #include "util/stat.h"
  17 #include "util/string2.h"
  18 #include "util/callchain.h"
  19 #include "util/time-utils.h"
  20
  21 #include <subcmd/pager.h>
  22 #include <subcmd/parse-options.h>
  23 #include "util/trace-event.h"
  24
  25 #include "util/debug.h"
  26
  27 #include <linux/kernel.h>
  28 #include <linux/log2.h>
  29 #include <linux/zalloc.h>
  30 #include <sys/prctl.h>
  31 #include <sys/resource.h>
  32 #include <inttypes.h>
  33
  34 #include <errno.h>
  35 #include <semaphore.h>
  36 #include <pthread.h>
  37 #include <math.h>
  38 #include <api/fs/fs.h>
  39 #include <linux/time64.h>
  40
  41 #include <linux/ctype.h>
  42
  43 #define PR_SET_NAME             15               /* Set process name */
  44 #define MAX_CPUS                4096
  45 #define COMM_LEN                20
  46 #define SYM_LEN                 129
  47 #define MAX_PID                 1024000
  48
  49 struct sched_atom;
  50
  51 struct task_desc {
  52         unsigned long           nr;
  53         unsigned long           pid;
  54         char                    comm[COMM_LEN];
  55
  56         unsigned long           nr_events;
  57         unsigned long           curr_event;
  58         struct sched_atom       **atoms;
  59
  60         pthread_t               thread;
  61         sem_t                   sleep_sem;
  62
  63         sem_t                   ready_for_work;
  64         sem_t                   work_done_sem;
  65
  66         u64                     cpu_usage;
  67 };
  68
  69 enum sched_event_type {
  70         SCHED_EVENT_RUN,
  71         SCHED_EVENT_SLEEP,
  72         SCHED_EVENT_WAKEUP,
  73         SCHED_EVENT_MIGRATION,
  74 };
  75
  76 struct sched_atom {
  77         enum sched_event_type   type;
  78         int                     specific_wait;
  79         u64                     timestamp;
  80         u64                     duration;
  81         unsigned long           nr;
  82         sem_t                   *wait_sem;
  83         struct task_desc        *wakee;
  84 };
  85
  86 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  87
  88 /* task state bitmask, copied from include/linux/sched.h */
  89 #define TASK_RUNNING            0
  90 #define TASK_INTERRUPTIBLE      1
  91 #define TASK_UNINTERRUPTIBLE    2
  92 #define __TASK_STOPPED          4
  93 #define __TASK_TRACED           8
  94 /* in tsk->exit_state */
  95 #define EXIT_DEAD               16
  96 #define EXIT_ZOMBIE             32
  97 #define EXIT_TRACE              (EXIT_ZOMBIE | EXIT_DEAD)
  98 /* in tsk->state again */
  99 #define TASK_DEAD               64
 100 #define TASK_WAKEKILL           128
 101 #define TASK_WAKING             256
 102 #define TASK_PARKED             512
 103
 104 enum thread_state {
 105         THREAD_SLEEPING = 0,
 106         THREAD_WAIT_CPU,
 107         THREAD_SCHED_IN,
 108         THREAD_IGNORE
 109 };
 110
 111 struct work_atom {
 112         struct list_head        list;
 113         enum thread_state       state;
 114         u64                     sched_out_time;
 115         u64                     wake_up_time;
 116         u64                     sched_in_time;
 117         u64                     runtime;
 118 };
 119
 120 struct work_atoms {
 121         struct list_head        work_list;
 122         struct thread           *thread;
 123         struct rb_node          node;
 124         u64                     max_lat;
 125         u64                     max_lat_at;
 126         u64                     total_lat;
 127         u64                     nb_atoms;
 128         u64                     total_runtime;
 129         int                     num_merged;
 130 };
 131
 132 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 133
 134 struct perf_sched;
 135
 136 struct trace_sched_handler {
 137         int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 138                             struct perf_sample *sample, struct machine *machine);
 139
 140         int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 141                              struct perf_sample *sample, struct machine *machine);
 142
 143         int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 144                             struct perf_sample *sample, struct machine *machine);
 145
 146         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 147         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 148                           struct machine *machine);
 149
 150         int (*migrate_task_event)(struct perf_sched *sched,
 151                                   struct evsel *evsel,
 152                                   struct perf_sample *sample,
 153                                   struct machine *machine);
 154 };
 155
 156 #define COLOR_PIDS PERF_COLOR_BLUE
 157 #define COLOR_CPUS PERF_COLOR_BG_RED
 158
 159 struct perf_sched_map {
 160         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 161         int                     *comp_cpus;
 162         bool                     comp;
 163         struct perf_thread_map *color_pids;
 164         const char              *color_pids_str;
 165         struct perf_cpu_map     *color_cpus;
 166         const char              *color_cpus_str;
 167         struct perf_cpu_map     *cpus;
 168         const char              *cpus_str;
 169 };
 170
 171 struct perf_sched {
 172         struct perf_tool tool;
 173         const char       *sort_order;
 174         unsigned long    nr_tasks;
 175         struct task_desc **pid_to_task;
 176         struct task_desc **tasks;
 177         const struct trace_sched_handler *tp_handler;
 178         pthread_mutex_t  start_work_mutex;
 179         pthread_mutex_t  work_done_wait_mutex;
 180         int              profile_cpu;
 181 /*
 182  * Track the current task - that way we can know whether there's any
 183  * weird events, such as a task being switched away that is not current.
 184  */
 185         int              max_cpu;
 186         u32              curr_pid[MAX_CPUS];
 187         struct thread    *curr_thread[MAX_CPUS];
 188         char             next_shortname1;
 189         char             next_shortname2;
 190         unsigned int     replay_repeat;
 191         unsigned long    nr_run_events;
 192         unsigned long    nr_sleep_events;
 193         unsigned long    nr_wakeup_events;
 194         unsigned long    nr_sleep_corrections;
 195         unsigned long    nr_run_events_optimized;
 196         unsigned long    targetless_wakeups;
 197         unsigned long    multitarget_wakeups;
 198         unsigned long    nr_runs;
 199         unsigned long    nr_timestamps;
 200         unsigned long    nr_unordered_timestamps;
 201         unsigned long    nr_context_switch_bugs;
 202         unsigned long    nr_events;
 203         unsigned long    nr_lost_chunks;
 204         unsigned long    nr_lost_events;
 205         u64              run_measurement_overhead;
 206         u64              sleep_measurement_overhead;
 207         u64              start_time;
 208         u64              cpu_usage;
 209         u64              runavg_cpu_usage;
 210         u64              parent_cpu_usage;
 211         u64              runavg_parent_cpu_usage;
 212         u64              sum_runtime;
 213         u64              sum_fluct;
 214         u64              run_avg;
 215         u64              all_runtime;
 216         u64              all_count;
 217         u64              cpu_last_switched[MAX_CPUS];
 218         struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 219         struct list_head sort_list, cmp_pid;
 220         bool force;
 221         bool skip_merge;
 222         struct perf_sched_map map;
 223
 224         /* options for timehist command */
 225         bool            summary;
 226         bool            summary_only;
 227         bool            idle_hist;
 228         bool            show_callchain;
 229         unsigned int    max_stack;
 230         bool            show_cpu_visual;
 231         bool            show_wakeups;
 232         bool            show_next;
 233         bool            show_migrations;
 234         bool            show_state;
 235         u64             skipped_samples;
 236         const char      *time_str;
 237         struct perf_time_interval ptime;
 238         struct perf_time_interval hist_time;
 239 };
 240
 241 /* per thread run time data */
 242 struct thread_runtime {
 243         u64 last_time;      /* time of previous sched in/out event */
 244         u64 dt_run;         /* run time */
 245         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 246         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 247         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 248         u64 dt_delay;       /* time between wakeup and sched-in */
 249         u64 ready_to_run;   /* time of wakeup */
 250
 251         struct stats run_stats;
 252         u64 total_run_time;
 253         u64 total_sleep_time;
 254         u64 total_iowait_time;
 255         u64 total_preempt_time;
 256         u64 total_delay_time;
 257
 258         int last_state;
 259
 260         char shortname[3];
 261         bool comm_changed;
 262
 263         u64 migrations;
 264 };
 265
 266 /* per event run time data */
 267 struct evsel_runtime {
 268         u64 *last_time; /* time this event was last seen per cpu */
 269         u32 ncpu;       /* highest cpu slot allocated */
 270 };
 271
 272 /* per cpu idle time data */
 273 struct idle_thread_runtime {
 274         struct thread_runtime   tr;
 275         struct thread           *last_thread;
 276         struct rb_root_cached   sorted_root;
 277         struct callchain_root   callchain;
 278         struct callchain_cursor cursor;
 279 };
 280
 281 /* track idle times per cpu */
 282 static struct thread **idle_threads;
 283 static int idle_max_cpu;
 284 static char idle_comm[] = "<idle>";
 285
 286 static u64 get_nsecs(void)
 287 {
 288         struct timespec ts;
 289
 290         clock_gettime(CLOCK_MONOTONIC, &ts);
 291
 292         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 293 }
 294
 295 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 296 {
 297         u64 T0 = get_nsecs(), T1;
 298
 299         do {
 300                 T1 = get_nsecs();
 301         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 302 }
 303
 304 static void sleep_nsecs(u64 nsecs)
 305 {
 306         struct timespec ts;
 307
 308         ts.tv_nsec = nsecs % 999999999;
 309         ts.tv_sec = nsecs / 999999999;
 310
 311         nanosleep(&ts, NULL);
 312 }
 313
 314 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 315 {
 316         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 317         int i;
 318
 319         for (i = 0; i < 10; i++) {
 320                 T0 = get_nsecs();
 321                 burn_nsecs(sched, 0);
 322                 T1 = get_nsecs();
 323                 delta = T1-T0;
 324                 min_delta = min(min_delta, delta);
 325         }
 326         sched->run_measurement_overhead = min_delta;
 327
 328         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 329 }
 330
 331 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 332 {
 333         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 334         int i;
 335
 336         for (i = 0; i < 10; i++) {
 337                 T0 = get_nsecs();
 338                 sleep_nsecs(10000);
 339                 T1 = get_nsecs();
 340                 delta = T1-T0;
 341                 min_delta = min(min_delta, delta);
 342         }
 343         min_delta -= 10000;
 344         sched->sleep_measurement_overhead = min_delta;
 345
 346         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 347 }
 348
 349 static struct sched_atom *
 350 get_new_event(struct task_desc *task, u64 timestamp)
 351 {
 352         struct sched_atom *event = zalloc(sizeof(*event));
 353         unsigned long idx = task->nr_events;
 354         size_t size;
 355
 356         event->timestamp = timestamp;
 357         event->nr = idx;
 358
 359         task->nr_events++;
 360         size = sizeof(struct sched_atom *) * task->nr_events;
 361         task->atoms = realloc(task->atoms, size);
 362         BUG_ON(!task->atoms);
 363
 364         task->atoms[idx] = event;
 365
 366         return event;
 367 }
 368
 369 static struct sched_atom *last_event(struct task_desc *task)
 370 {
 371         if (!task->nr_events)
 372                 return NULL;
 373
 374         return task->atoms[task->nr_events - 1];
 375 }
 376
 377 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 378                                 u64 timestamp, u64 duration)
 379 {
 380         struct sched_atom *event, *curr_event = last_event(task);
 381
 382         /*
 383          * optimize an existing RUN event by merging this one
 384          * to it:
 385          */
 386         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 387                 sched->nr_run_events_optimized++;
 388                 curr_event->duration += duration;
 389                 return;
 390         }
 391
 392         event = get_new_event(task, timestamp);
 393
 394         event->type = SCHED_EVENT_RUN;
 395         event->duration = duration;
 396
 397         sched->nr_run_events++;
 398 }
 399
 400 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 401                                    u64 timestamp, struct task_desc *wakee)
 402 {
 403         struct sched_atom *event, *wakee_event;
 404
 405         event = get_new_event(task, timestamp);
 406         event->type = SCHED_EVENT_WAKEUP;
 407         event->wakee = wakee;
 408
 409         wakee_event = last_event(wakee);
 410         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 411                 sched->targetless_wakeups++;
 412                 return;
 413         }
 414         if (wakee_event->wait_sem) {
 415                 sched->multitarget_wakeups++;
 416                 return;
 417         }
 418
 419         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 420         sem_init(wakee_event->wait_sem, 0, 0);
 421         wakee_event->specific_wait = 1;
 422         event->wait_sem = wakee_event->wait_sem;
 423
 424         sched->nr_wakeup_events++;
 425 }
 426
 427 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 428                                   u64 timestamp, u64 task_state __maybe_unused)
 429 {
 430         struct sched_atom *event = get_new_event(task, timestamp);
 431
 432         event->type = SCHED_EVENT_SLEEP;
 433
 434         sched->nr_sleep_events++;
 435 }
 436
 437 static struct task_desc *register_pid(struct perf_sched *sched,
 438                                       unsigned long pid, const char *comm)
 439 {
 440         struct task_desc *task;
 441         static int pid_max;
 442
 443         if (sched->pid_to_task == NULL) {
 444                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 445                         pid_max = MAX_PID;
 446                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 447         }
 448         if (pid >= (unsigned long)pid_max) {
 449                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 450                         sizeof(struct task_desc *))) == NULL);
 451                 while (pid >= (unsigned long)pid_max)
 452                         sched->pid_to_task[pid_max++] = NULL;
 453         }
 454
 455         task = sched->pid_to_task[pid];
 456
 457         if (task)
 458                 return task;
 459
 460         task = zalloc(sizeof(*task));
 461         task->pid = pid;
 462         task->nr = sched->nr_tasks;
 463         strcpy(task->comm, comm);
 464         /*
 465          * every task starts in sleeping state - this gets ignored
 466          * if there's no wakeup pointing to this sleep state:
 467          */
 468         add_sched_event_sleep(sched, task, 0, 0);
 469
 470         sched->pid_to_task[pid] = task;
 471         sched->nr_tasks++;
 472         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 473         BUG_ON(!sched->tasks);
 474         sched->tasks[task->nr] = task;
 475
 476         if (verbose > 0)
 477                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 478
 479         return task;
 480 }
 481
 482
 483 static void print_task_traces(struct perf_sched *sched)
 484 {
 485         struct task_desc *task;
 486         unsigned long i;
 487
 488         for (i = 0; i < sched->nr_tasks; i++) {
 489                 task = sched->tasks[i];
 490                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 491                         task->nr, task->comm, task->pid, task->nr_events);
 492         }
 493 }
 494
 495 static void add_cross_task_wakeups(struct perf_sched *sched)
 496 {
 497         struct task_desc *task1, *task2;
 498         unsigned long i, j;
 499
 500         for (i = 0; i < sched->nr_tasks; i++) {
 501                 task1 = sched->tasks[i];
 502                 j = i + 1;
 503                 if (j == sched->nr_tasks)
 504                         j = 0;
 505                 task2 = sched->tasks[j];
 506                 add_sched_event_wakeup(sched, task1, 0, task2);
 507         }
 508 }
 509
 510 static void perf_sched__process_event(struct perf_sched *sched,
 511                                       struct sched_atom *atom)
 512 {
 513         int ret = 0;
 514
 515         switch (atom->type) {
 516                 case SCHED_EVENT_RUN:
 517                         burn_nsecs(sched, atom->duration);
 518                         break;
 519                 case SCHED_EVENT_SLEEP:
 520                         if (atom->wait_sem)
 521                                 ret = sem_wait(atom->wait_sem);
 522                         BUG_ON(ret);
 523                         break;
 524                 case SCHED_EVENT_WAKEUP:
 525                         if (atom->wait_sem)
 526                                 ret = sem_post(atom->wait_sem);
 527                         BUG_ON(ret);
 528                         break;
 529                 case SCHED_EVENT_MIGRATION:
 530                         break;
 531                 default:
 532                         BUG_ON(1);
 533         }
 534 }
 535
 536 static u64 get_cpu_usage_nsec_parent(void)
 537 {
 538         struct rusage ru;
 539         u64 sum;
 540         int err;
 541
 542         err = getrusage(RUSAGE_SELF, &ru);
 543         BUG_ON(err);
 544
 545         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 546         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 547
 548         return sum;
 549 }
 550
 551 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 552 {
 553         struct perf_event_attr attr;
 554         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 555         int fd;
 556         struct rlimit limit;
 557         bool need_privilege = false;
 558
 559         memset(&attr, 0, sizeof(attr));
 560
 561         attr.type = PERF_TYPE_SOFTWARE;
 562         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 563
 564 force_again:
 565         fd = sys_perf_event_open(&attr, 0, -1, -1,
 566                                  perf_event_open_cloexec_flag());
 567
 568         if (fd < 0) {
 569                 if (errno == EMFILE) {
 570                         if (sched->force) {
 571                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 572                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 573                                 if (limit.rlim_cur > limit.rlim_max) {
 574                                         limit.rlim_max = limit.rlim_cur;
 575                                         need_privilege = true;
 576                                 }
 577                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 578                                         if (need_privilege && errno == EPERM)
 579                                                 strcpy(info, "Need privilege\n");
 580                                 } else
 581                                         goto force_again;
 582                         } else
 583                                 strcpy(info, "Have a try with -f option\n");
 584                 }
 585                 pr_err("Error: sys_perf_event_open() syscall returned "
 586                        "with %d (%s)\n%s", fd,
 587                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 588                 exit(EXIT_FAILURE);
 589         }
 590         return fd;
 591 }
 592
 593 static u64 get_cpu_usage_nsec_self(int fd)
 594 {
 595         u64 runtime;
 596         int ret;
 597
 598         ret = read(fd, &runtime, sizeof(runtime));
 599         BUG_ON(ret != sizeof(runtime));
 600
 601         return runtime;
 602 }
 603
 604 struct sched_thread_parms {
 605         struct task_desc  *task;
 606         struct perf_sched *sched;
 607         int fd;
 608 };
 609
 610 static void *thread_func(void *ctx)
 611 {
 612         struct sched_thread_parms *parms = ctx;
 613         struct task_desc *this_task = parms->task;
 614         struct perf_sched *sched = parms->sched;
 615         u64 cpu_usage_0, cpu_usage_1;
 616         unsigned long i, ret;
 617         char comm2[22];
 618         int fd = parms->fd;
 619
 620         zfree(&parms);
 621
 622         sprintf(comm2, ":%s", this_task->comm);
 623         prctl(PR_SET_NAME, comm2);
 624         if (fd < 0)
 625                 return NULL;
 626 again:
 627         ret = sem_post(&this_task->ready_for_work);
 628         BUG_ON(ret);
 629         ret = pthread_mutex_lock(&sched->start_work_mutex);
 630         BUG_ON(ret);
 631         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 632         BUG_ON(ret);
 633
 634         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 635
 636         for (i = 0; i < this_task->nr_events; i++) {
 637                 this_task->curr_event = i;
 638                 perf_sched__process_event(sched, this_task->atoms[i]);
 639         }
 640
 641         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 642         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 643         ret = sem_post(&this_task->work_done_sem);
 644         BUG_ON(ret);
 645
 646         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 647         BUG_ON(ret);
 648         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 649         BUG_ON(ret);
 650
 651         goto again;
 652 }
 653
 654 static void create_tasks(struct perf_sched *sched)
 655 {
 656         struct task_desc *task;
 657         pthread_attr_t attr;
 658         unsigned long i;
 659         int err;
 660
 661         err = pthread_attr_init(&attr);
 662         BUG_ON(err);
 663         err = pthread_attr_setstacksize(&attr,
 664                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 665         BUG_ON(err);
 666         err = pthread_mutex_lock(&sched->start_work_mutex);
 667         BUG_ON(err);
 668         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 669         BUG_ON(err);
 670         for (i = 0; i < sched->nr_tasks; i++) {
 671                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 672                 BUG_ON(parms == NULL);
 673                 parms->task = task = sched->tasks[i];
 674                 parms->sched = sched;
 675                 parms->fd = self_open_counters(sched, i);
 676                 sem_init(&task->sleep_sem, 0, 0);
 677                 sem_init(&task->ready_for_work, 0, 0);
 678                 sem_init(&task->work_done_sem, 0, 0);
 679                 task->curr_event = 0;
 680                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 681                 BUG_ON(err);
 682         }
 683 }
 684
 685 static void wait_for_tasks(struct perf_sched *sched)
 686 {
 687         u64 cpu_usage_0, cpu_usage_1;
 688         struct task_desc *task;
 689         unsigned long i, ret;
 690
 691         sched->start_time = get_nsecs();
 692         sched->cpu_usage = 0;
 693         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 694
 695         for (i = 0; i < sched->nr_tasks; i++) {
 696                 task = sched->tasks[i];
 697                 ret = sem_wait(&task->ready_for_work);
 698                 BUG_ON(ret);
 699                 sem_init(&task->ready_for_work, 0, 0);
 700         }
 701         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 702         BUG_ON(ret);
 703
 704         cpu_usage_0 = get_cpu_usage_nsec_parent();
 705
 706         pthread_mutex_unlock(&sched->start_work_mutex);
 707
 708         for (i = 0; i < sched->nr_tasks; i++) {
 709                 task = sched->tasks[i];
 710                 ret = sem_wait(&task->work_done_sem);
 711                 BUG_ON(ret);
 712                 sem_init(&task->work_done_sem, 0, 0);
 713                 sched->cpu_usage += task->cpu_usage;
 714                 task->cpu_usage = 0;
 715         }
 716
 717         cpu_usage_1 = get_cpu_usage_nsec_parent();
 718         if (!sched->runavg_cpu_usage)
 719                 sched->runavg_cpu_usage = sched->cpu_usage;
 720         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 721
 722         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 723         if (!sched->runavg_parent_cpu_usage)
 724                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 725         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 726                                          sched->parent_cpu_usage)/sched->replay_repeat;
 727
 728         ret = pthread_mutex_lock(&sched->start_work_mutex);
 729         BUG_ON(ret);
 730
 731         for (i = 0; i < sched->nr_tasks; i++) {
 732                 task = sched->tasks[i];
 733                 sem_init(&task->sleep_sem, 0, 0);
 734                 task->curr_event = 0;
 735         }
 736 }
 737
 738 static void run_one_test(struct perf_sched *sched)
 739 {
 740         u64 T0, T1, delta, avg_delta, fluct;
 741
 742         T0 = get_nsecs();
 743         wait_for_tasks(sched);
 744         T1 = get_nsecs();
 745
 746         delta = T1 - T0;
 747         sched->sum_runtime += delta;
 748         sched->nr_runs++;
 749
 750         avg_delta = sched->sum_runtime / sched->nr_runs;
 751         if (delta < avg_delta)
 752                 fluct = avg_delta - delta;
 753         else
 754                 fluct = delta - avg_delta;
 755         sched->sum_fluct += fluct;
 756         if (!sched->run_avg)
 757                 sched->run_avg = delta;
 758         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 759
 760         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 761
 762         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 763
 764         printf("cpu: %0.2f / %0.2f",
 765                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 766
 767 #if 0
 768         /*
 769          * rusage statistics done by the parent, these are less
 770          * accurate than the sched->sum_exec_runtime based statistics:
 771          */
 772         printf(" [%0.2f / %0.2f]",
 773                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 774                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 775 #endif
 776
 777         printf("\n");
 778
 779         if (sched->nr_sleep_corrections)
 780                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 781         sched->nr_sleep_corrections = 0;
 782 }
 783
 784 static void test_calibrations(struct perf_sched *sched)
 785 {
 786         u64 T0, T1;
 787
 788         T0 = get_nsecs();
 789         burn_nsecs(sched, NSEC_PER_MSEC);
 790         T1 = get_nsecs();
 791
 792         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 793
 794         T0 = get_nsecs();
 795         sleep_nsecs(NSEC_PER_MSEC);
 796         T1 = get_nsecs();
 797
 798         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 799 }
 800
 801 static int
 802 replay_wakeup_event(struct perf_sched *sched,
 803                     struct evsel *evsel, struct perf_sample *sample,
 804                     struct machine *machine __maybe_unused)
 805 {
 806         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 807         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 808         struct task_desc *waker, *wakee;
 809
 810         if (verbose > 0) {
 811                 printf("sched_wakeup event %p\n", evsel);
 812
 813                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 814         }
 815
 816         waker = register_pid(sched, sample->tid, "<unknown>");
 817         wakee = register_pid(sched, pid, comm);
 818
 819         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 820         return 0;
 821 }
 822
 823 static int replay_switch_event(struct perf_sched *sched,
 824                                struct evsel *evsel,
 825                                struct perf_sample *sample,
 826                                struct machine *machine __maybe_unused)
 827 {
 828         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 829                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 830         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 831                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 832         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 833         struct task_desc *prev, __maybe_unused *next;
 834         u64 timestamp0, timestamp = sample->time;
 835         int cpu = sample->cpu;
 836         s64 delta;
 837
 838         if (verbose > 0)
 839                 printf("sched_switch event %p\n", evsel);
 840
 841         if (cpu >= MAX_CPUS || cpu < 0)
 842                 return 0;
 843
 844         timestamp0 = sched->cpu_last_switched[cpu];
 845         if (timestamp0)
 846                 delta = timestamp - timestamp0;
 847         else
 848                 delta = 0;
 849
 850         if (delta < 0) {
 851                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 852                 return -1;
 853         }
 854
 855         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 856                  prev_comm, prev_pid, next_comm, next_pid, delta);
 857
 858         prev = register_pid(sched, prev_pid, prev_comm);
 859         next = register_pid(sched, next_pid, next_comm);
 860
 861         sched->cpu_last_switched[cpu] = timestamp;
 862
 863         add_sched_event_run(sched, prev, timestamp, delta);
 864         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 865
 866         return 0;
 867 }
 868
 869 static int replay_fork_event(struct perf_sched *sched,
 870                              union perf_event *event,
 871                              struct machine *machine)
 872 {
 873         struct thread *child, *parent;
 874
 875         child = machine__findnew_thread(machine, event->fork.pid,
 876                                         event->fork.tid);
 877         parent = machine__findnew_thread(machine, event->fork.ppid,
 878                                          event->fork.ptid);
 879
 880         if (child == NULL || parent == NULL) {
 881                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 882                                  child, parent);
 883                 goto out_put;
 884         }
 885
 886         if (verbose > 0) {
 887                 printf("fork event\n");
 888                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 889                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 890         }
 891
 892         register_pid(sched, parent->tid, thread__comm_str(parent));
 893         register_pid(sched, child->tid, thread__comm_str(child));
 894 out_put:
 895         thread__put(child);
 896         thread__put(parent);
 897         return 0;
 898 }
 899
 900 struct sort_dimension {
 901         const char              *name;
 902         sort_fn_t               cmp;
 903         struct list_head        list;
 904 };
 905
 906 /*
 907  * handle runtime stats saved per thread
 908  */
 909 static struct thread_runtime *thread__init_runtime(struct thread *thread)
 910 {
 911         struct thread_runtime *r;
 912
 913         r = zalloc(sizeof(struct thread_runtime));
 914         if (!r)
 915                 return NULL;
 916
 917         init_stats(&r->run_stats);
 918         thread__set_priv(thread, r);
 919
 920         return r;
 921 }
 922
 923 static struct thread_runtime *thread__get_runtime(struct thread *thread)
 924 {
 925         struct thread_runtime *tr;
 926
 927         tr = thread__priv(thread);
 928         if (tr == NULL) {
 929                 tr = thread__init_runtime(thread);
 930                 if (tr == NULL)
 931                         pr_debug("Failed to malloc memory for runtime data.\n");
 932         }
 933
 934         return tr;
 935 }
 936
 937 static int
 938 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 939 {
 940         struct sort_dimension *sort;
 941         int ret = 0;
 942
 943         BUG_ON(list_empty(list));
 944
 945         list_for_each_entry(sort, list, list) {
 946                 ret = sort->cmp(l, r);
 947                 if (ret)
 948                         return ret;
 949         }
 950
 951         return ret;
 952 }
 953
 954 static struct work_atoms *
 955 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 956                          struct list_head *sort_list)
 957 {
 958         struct rb_node *node = root->rb_root.rb_node;
 959         struct work_atoms key = { .thread = thread };
 960
 961         while (node) {
 962                 struct work_atoms *atoms;
 963                 int cmp;
 964
 965                 atoms = container_of(node, struct work_atoms, node);
 966
 967                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 968                 if (cmp > 0)
 969                         node = node->rb_left;
 970                 else if (cmp < 0)
 971                         node = node->rb_right;
 972                 else {
 973                         BUG_ON(thread != atoms->thread);
 974                         return atoms;
 975                 }
 976         }
 977         return NULL;
 978 }
 979
 980 static void
 981 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 982                          struct list_head *sort_list)
 983 {
 984         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 985         bool leftmost = true;
 986
 987         while (*new) {
 988                 struct work_atoms *this;
 989                 int cmp;
 990
 991                 this = container_of(*new, struct work_atoms, node);
 992                 parent = *new;
 993
 994                 cmp = thread_lat_cmp(sort_list, data, this);
 995
 996                 if (cmp > 0)
 997                         new = &((*new)->rb_left);
 998                 else {
 999                         new = &((*new)->rb_right);
1000                         leftmost = false;
1001                 }
1002         }
1003
1004         rb_link_node(&data->node, parent, new);
1005         rb_insert_color_cached(&data->node, root, leftmost);
1006 }
1007
1008 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1009 {
1010         struct work_atoms *atoms = zalloc(sizeof(*atoms));
1011         if (!atoms) {
1012                 pr_err("No memory at %s\n", __func__);
1013                 return -1;
1014         }
1015
1016         atoms->thread = thread__get(thread);
1017         INIT_LIST_HEAD(&atoms->work_list);
1018         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1019         return 0;
1020 }
1021
1022 static char sched_out_state(u64 prev_state)
1023 {
1024         const char *str = TASK_STATE_TO_CHAR_STR;
1025
1026         return str[prev_state];
1027 }
1028
1029 static int
1030 add_sched_out_event(struct work_atoms *atoms,
1031                     char run_state,
1032                     u64 timestamp)
1033 {
1034         struct work_atom *atom = zalloc(sizeof(*atom));
1035         if (!atom) {
1036                 pr_err("Non memory at %s", __func__);
1037                 return -1;
1038         }
1039
1040         atom->sched_out_time = timestamp;
1041
1042         if (run_state == 'R') {
1043                 atom->state = THREAD_WAIT_CPU;
1044                 atom->wake_up_time = atom->sched_out_time;
1045         }
1046
1047         list_add_tail(&atom->list, &atoms->work_list);
1048         return 0;
1049 }
1050
1051 static void
1052 add_runtime_event(struct work_atoms *atoms, u64 delta,
1053                   u64 timestamp __maybe_unused)
1054 {
1055         struct work_atom *atom;
1056
1057         BUG_ON(list_empty(&atoms->work_list));
1058
1059         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1060
1061         atom->runtime += delta;
1062         atoms->total_runtime += delta;
1063 }
1064
1065 static void
1066 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1067 {
1068         struct work_atom *atom;
1069         u64 delta;
1070
1071         if (list_empty(&atoms->work_list))
1072                 return;
1073
1074         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1075
1076         if (atom->state != THREAD_WAIT_CPU)
1077                 return;
1078
1079         if (timestamp < atom->wake_up_time) {
1080                 atom->state = THREAD_IGNORE;
1081                 return;
1082         }
1083
1084         atom->state = THREAD_SCHED_IN;
1085         atom->sched_in_time = timestamp;
1086
1087         delta = atom->sched_in_time - atom->wake_up_time;
1088         atoms->total_lat += delta;
1089         if (delta > atoms->max_lat) {
1090                 atoms->max_lat = delta;
1091                 atoms->max_lat_at = timestamp;
1092         }
1093         atoms->nb_atoms++;
1094 }
1095
1096 static int latency_switch_event(struct perf_sched *sched,
1097                                 struct evsel *evsel,
1098                                 struct perf_sample *sample,
1099                                 struct machine *machine)
1100 {
1101         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1102                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1103         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1104         struct work_atoms *out_events, *in_events;
1105         struct thread *sched_out, *sched_in;
1106         u64 timestamp0, timestamp = sample->time;
1107         int cpu = sample->cpu, err = -1;
1108         s64 delta;
1109
1110         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1111
1112         timestamp0 = sched->cpu_last_switched[cpu];
1113         sched->cpu_last_switched[cpu] = timestamp;
1114         if (timestamp0)
1115                 delta = timestamp - timestamp0;
1116         else
1117                 delta = 0;
1118
1119         if (delta < 0) {
1120                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1121                 return -1;
1122         }
1123
1124         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1125         sched_in = machine__findnew_thread(machine, -1, next_pid);
1126         if (sched_out == NULL || sched_in == NULL)
1127                 goto out_put;
1128
1129         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1130         if (!out_events) {
1131                 if (thread_atoms_insert(sched, sched_out))
1132                         goto out_put;
1133                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1134                 if (!out_events) {
1135                         pr_err("out-event: Internal tree error");
1136                         goto out_put;
1137                 }
1138         }
1139         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1140                 return -1;
1141
1142         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1143         if (!in_events) {
1144                 if (thread_atoms_insert(sched, sched_in))
1145                         goto out_put;
1146                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1147                 if (!in_events) {
1148                         pr_err("in-event: Internal tree error");
1149                         goto out_put;
1150                 }
1151                 /*
1152                  * Take came in we have not heard about yet,
1153                  * add in an initial atom in runnable state:
1154                  */
1155                 if (add_sched_out_event(in_events, 'R', timestamp))
1156                         goto out_put;
1157         }
1158         add_sched_in_event(in_events, timestamp);
1159         err = 0;
1160 out_put:
1161         thread__put(sched_out);
1162         thread__put(sched_in);
1163         return err;
1164 }
1165
1166 static int latency_runtime_event(struct perf_sched *sched,
1167                                  struct evsel *evsel,
1168                                  struct perf_sample *sample,
1169                                  struct machine *machine)
1170 {
1171         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1172         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1173         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1174         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1175         u64 timestamp = sample->time;
1176         int cpu = sample->cpu, err = -1;
1177
1178         if (thread == NULL)
1179                 return -1;
1180
1181         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1182         if (!atoms) {
1183                 if (thread_atoms_insert(sched, thread))
1184                         goto out_put;
1185                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1186                 if (!atoms) {
1187                         pr_err("in-event: Internal tree error");
1188                         goto out_put;
1189                 }
1190                 if (add_sched_out_event(atoms, 'R', timestamp))
1191                         goto out_put;
1192         }
1193
1194         add_runtime_event(atoms, runtime, timestamp);
1195         err = 0;
1196 out_put:
1197         thread__put(thread);
1198         return err;
1199 }
1200
1201 static int latency_wakeup_event(struct perf_sched *sched,
1202                                 struct evsel *evsel,
1203                                 struct perf_sample *sample,
1204                                 struct machine *machine)
1205 {
1206         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1207         struct work_atoms *atoms;
1208         struct work_atom *atom;
1209         struct thread *wakee;
1210         u64 timestamp = sample->time;
1211         int err = -1;
1212
1213         wakee = machine__findnew_thread(machine, -1, pid);
1214         if (wakee == NULL)
1215                 return -1;
1216         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1217         if (!atoms) {
1218                 if (thread_atoms_insert(sched, wakee))
1219                         goto out_put;
1220                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1221                 if (!atoms) {
1222                         pr_err("wakeup-event: Internal tree error");
1223                         goto out_put;
1224                 }
1225                 if (add_sched_out_event(atoms, 'S', timestamp))
1226                         goto out_put;
1227         }
1228
1229         BUG_ON(list_empty(&atoms->work_list));
1230
1231         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1232
1233         /*
1234          * As we do not guarantee the wakeup event happens when
1235          * task is out of run queue, also may happen when task is
1236          * on run queue and wakeup only change ->state to TASK_RUNNING,
1237          * then we should not set the ->wake_up_time when wake up a
1238          * task which is on run queue.
1239          *
1240          * You WILL be missing events if you've recorded only
1241          * one CPU, or are only looking at only one, so don't
1242          * skip in this case.
1243          */
1244         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1245                 goto out_ok;
1246
1247         sched->nr_timestamps++;
1248         if (atom->sched_out_time > timestamp) {
1249                 sched->nr_unordered_timestamps++;
1250                 goto out_ok;
1251         }
1252
1253         atom->state = THREAD_WAIT_CPU;
1254         atom->wake_up_time = timestamp;
1255 out_ok:
1256         err = 0;
1257 out_put:
1258         thread__put(wakee);
1259         return err;
1260 }
1261
1262 static int latency_migrate_task_event(struct perf_sched *sched,
1263                                       struct evsel *evsel,
1264                                       struct perf_sample *sample,
1265                                       struct machine *machine)
1266 {
1267         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1268         u64 timestamp = sample->time;
1269         struct work_atoms *atoms;
1270         struct work_atom *atom;
1271         struct thread *migrant;
1272         int err = -1;
1273
1274         /*
1275          * Only need to worry about migration when profiling one CPU.
1276          */
1277         if (sched->profile_cpu == -1)
1278                 return 0;
1279
1280         migrant = machine__findnew_thread(machine, -1, pid);
1281         if (migrant == NULL)
1282                 return -1;
1283         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1284         if (!atoms) {
1285                 if (thread_atoms_insert(sched, migrant))
1286                         goto out_put;
1287                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1288                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1289                 if (!atoms) {
1290                         pr_err("migration-event: Internal tree error");
1291                         goto out_put;
1292                 }
1293                 if (add_sched_out_event(atoms, 'R', timestamp))
1294                         goto out_put;
1295         }
1296
1297         BUG_ON(list_empty(&atoms->work_list));
1298
1299         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1300         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1301
1302         sched->nr_timestamps++;
1303
1304         if (atom->sched_out_time > timestamp)
1305                 sched->nr_unordered_timestamps++;
1306         err = 0;
1307 out_put:
1308         thread__put(migrant);
1309         return err;
1310 }
1311
1312 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1313 {
1314         int i;
1315         int ret;
1316         u64 avg;
1317         char max_lat_at[32];
1318
1319         if (!work_list->nb_atoms)
1320                 return;
1321         /*
1322          * Ignore idle threads:
1323          */
1324         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1325                 return;
1326
1327         sched->all_runtime += work_list->total_runtime;
1328         sched->all_count   += work_list->nb_atoms;
1329
1330         if (work_list->num_merged > 1)
1331                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1332         else
1333                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1334
1335         for (i = 0; i < 24 - ret; i++)
1336                 printf(" ");
1337
1338         avg = work_list->total_lat / work_list->nb_atoms;
1339         timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1340
1341         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1342               (double)work_list->total_runtime / NSEC_PER_MSEC,
1343                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1344                  (double)work_list->max_lat / NSEC_PER_MSEC,
1345                  max_lat_at);
1346 }
1347
1348 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1349 {
1350         if (l->thread == r->thread)
1351                 return 0;
1352         if (l->thread->tid < r->thread->tid)
1353                 return -1;
1354         if (l->thread->tid > r->thread->tid)
1355                 return 1;
1356         return (int)(l->thread - r->thread);
1357 }
1358
1359 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1360 {
1361         u64 avgl, avgr;
1362
1363         if (!l->nb_atoms)
1364                 return -1;
1365
1366         if (!r->nb_atoms)
1367                 return 1;
1368
1369         avgl = l->total_lat / l->nb_atoms;
1370         avgr = r->total_lat / r->nb_atoms;
1371
1372         if (avgl < avgr)
1373                 return -1;
1374         if (avgl > avgr)
1375                 return 1;
1376
1377         return 0;
1378 }
1379
1380 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1381 {
1382         if (l->max_lat < r->max_lat)
1383                 return -1;
1384         if (l->max_lat > r->max_lat)
1385                 return 1;
1386
1387         return 0;
1388 }
1389
1390 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1391 {
1392         if (l->nb_atoms < r->nb_atoms)
1393                 return -1;
1394         if (l->nb_atoms > r->nb_atoms)
1395                 return 1;
1396
1397         return 0;
1398 }
1399
1400 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1401 {
1402         if (l->total_runtime < r->total_runtime)
1403                 return -1;
1404         if (l->total_runtime > r->total_runtime)
1405                 return 1;
1406
1407         return 0;
1408 }
1409
1410 static int sort_dimension__add(const char *tok, struct list_head *list)
1411 {
1412         size_t i;
1413         static struct sort_dimension avg_sort_dimension = {
1414                 .name = "avg",
1415                 .cmp  = avg_cmp,
1416         };
1417         static struct sort_dimension max_sort_dimension = {
1418                 .name = "max",
1419                 .cmp  = max_cmp,
1420         };
1421         static struct sort_dimension pid_sort_dimension = {
1422                 .name = "pid",
1423                 .cmp  = pid_cmp,
1424         };
1425         static struct sort_dimension runtime_sort_dimension = {
1426                 .name = "runtime",
1427                 .cmp  = runtime_cmp,
1428         };
1429         static struct sort_dimension switch_sort_dimension = {
1430                 .name = "switch",
1431                 .cmp  = switch_cmp,
1432         };
1433         struct sort_dimension *available_sorts[] = {
1434                 &pid_sort_dimension,
1435                 &avg_sort_dimension,
1436                 &max_sort_dimension,
1437                 &switch_sort_dimension,
1438                 &runtime_sort_dimension,
1439         };
1440
1441         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1442                 if (!strcmp(available_sorts[i]->name, tok)) {
1443                         list_add_tail(&available_sorts[i]->list, list);
1444
1445                         return 0;
1446                 }
1447         }
1448
1449         return -1;
1450 }
1451
1452 static void perf_sched__sort_lat(struct perf_sched *sched)
1453 {
1454         struct rb_node *node;
1455         struct rb_root_cached *root = &sched->atom_root;
1456 again:
1457         for (;;) {
1458                 struct work_atoms *data;
1459                 node = rb_first_cached(root);
1460                 if (!node)
1461                         break;
1462
1463                 rb_erase_cached(node, root);
1464                 data = rb_entry(node, struct work_atoms, node);
1465                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1466         }
1467         if (root == &sched->atom_root) {
1468                 root = &sched->merged_atom_root;
1469                 goto again;
1470         }
1471 }
1472
1473 static int process_sched_wakeup_event(struct perf_tool *tool,
1474                                       struct evsel *evsel,
1475                                       struct perf_sample *sample,
1476                                       struct machine *machine)
1477 {
1478         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1479
1480         if (sched->tp_handler->wakeup_event)
1481                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1482
1483         return 0;
1484 }
1485
1486 union map_priv {
1487         void    *ptr;
1488         bool     color;
1489 };
1490
1491 static bool thread__has_color(struct thread *thread)
1492 {
1493         union map_priv priv = {
1494                 .ptr = thread__priv(thread),
1495         };
1496
1497         return priv.color;
1498 }
1499
1500 static struct thread*
1501 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1502 {
1503         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1504         union map_priv priv = {
1505                 .color = false,
1506         };
1507
1508         if (!sched->map.color_pids || !thread || thread__priv(thread))
1509                 return thread;
1510
1511         if (thread_map__has(sched->map.color_pids, tid))
1512                 priv.color = true;
1513
1514         thread__set_priv(thread, priv.ptr);
1515         return thread;
1516 }
1517
1518 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1519                             struct perf_sample *sample, struct machine *machine)
1520 {
1521         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1522         struct thread *sched_in;
1523         struct thread_runtime *tr;
1524         int new_shortname;
1525         u64 timestamp0, timestamp = sample->time;
1526         s64 delta;
1527         int i, this_cpu = sample->cpu;
1528         int cpus_nr;
1529         bool new_cpu = false;
1530         const char *color = PERF_COLOR_NORMAL;
1531         char stimestamp[32];
1532
1533         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1534
1535         if (this_cpu > sched->max_cpu)
1536                 sched->max_cpu = this_cpu;
1537
1538         if (sched->map.comp) {
1539                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1540                 if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1541                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1542                         new_cpu = true;
1543                 }
1544         } else
1545                 cpus_nr = sched->max_cpu;
1546
1547         timestamp0 = sched->cpu_last_switched[this_cpu];
1548         sched->cpu_last_switched[this_cpu] = timestamp;
1549         if (timestamp0)
1550                 delta = timestamp - timestamp0;
1551         else
1552                 delta = 0;
1553
1554         if (delta < 0) {
1555                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1556                 return -1;
1557         }
1558
1559         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1560         if (sched_in == NULL)
1561                 return -1;
1562
1563         tr = thread__get_runtime(sched_in);
1564         if (tr == NULL) {
1565                 thread__put(sched_in);
1566                 return -1;
1567         }
1568
1569         sched->curr_thread[this_cpu] = thread__get(sched_in);
1570
1571         printf("  ");
1572
1573         new_shortname = 0;
1574         if (!tr->shortname[0]) {
1575                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1576                         /*
1577                          * Don't allocate a letter-number for swapper:0
1578                          * as a shortname. Instead, we use '.' for it.
1579                          */
1580                         tr->shortname[0] = '.';
1581                         tr->shortname[1] = ' ';
1582                 } else {
1583                         tr->shortname[0] = sched->next_shortname1;
1584                         tr->shortname[1] = sched->next_shortname2;
1585
1586                         if (sched->next_shortname1 < 'Z') {
1587                                 sched->next_shortname1++;
1588                         } else {
1589                                 sched->next_shortname1 = 'A';
1590                                 if (sched->next_shortname2 < '9')
1591                                         sched->next_shortname2++;
1592                                 else
1593                                         sched->next_shortname2 = '0';
1594                         }
1595                 }
1596                 new_shortname = 1;
1597         }
1598
1599         for (i = 0; i < cpus_nr; i++) {
1600                 int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1601                 struct thread *curr_thread = sched->curr_thread[cpu];
1602                 struct thread_runtime *curr_tr;
1603                 const char *pid_color = color;
1604                 const char *cpu_color = color;
1605
1606                 if (curr_thread && thread__has_color(curr_thread))
1607                         pid_color = COLOR_PIDS;
1608
1609                 if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1610                         continue;
1611
1612                 if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1613                         cpu_color = COLOR_CPUS;
1614
1615                 if (cpu != this_cpu)
1616                         color_fprintf(stdout, color, " ");
1617                 else
1618                         color_fprintf(stdout, cpu_color, "*");
1619
1620                 if (sched->curr_thread[cpu]) {
1621                         curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1622                         if (curr_tr == NULL) {
1623                                 thread__put(sched_in);
1624                                 return -1;
1625                         }
1626                         color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1627                 } else
1628                         color_fprintf(stdout, color, "   ");
1629         }
1630
1631         if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1632                 goto out;
1633
1634         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1635         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1636         if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1637                 const char *pid_color = color;
1638
1639                 if (thread__has_color(sched_in))
1640                         pid_color = COLOR_PIDS;
1641
1642                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1643                        tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1644                 tr->comm_changed = false;
1645         }
1646
1647         if (sched->map.comp && new_cpu)
1648                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1649
1650 out:
1651         color_fprintf(stdout, color, "\n");
1652
1653         thread__put(sched_in);
1654
1655         return 0;
1656 }
1657
1658 static int process_sched_switch_event(struct perf_tool *tool,
1659                                       struct evsel *evsel,
1660                                       struct perf_sample *sample,
1661                                       struct machine *machine)
1662 {
1663         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1664         int this_cpu = sample->cpu, err = 0;
1665         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1666             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1667
1668         if (sched->curr_pid[this_cpu] != (u32)-1) {
1669                 /*
1670                  * Are we trying to switch away a PID that is
1671                  * not current?
1672                  */
1673                 if (sched->curr_pid[this_cpu] != prev_pid)
1674                         sched->nr_context_switch_bugs++;
1675         }
1676
1677         if (sched->tp_handler->switch_event)
1678                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1679
1680         sched->curr_pid[this_cpu] = next_pid;
1681         return err;
1682 }
1683
1684 static int process_sched_runtime_event(struct perf_tool *tool,
1685                                        struct evsel *evsel,
1686                                        struct perf_sample *sample,
1687                                        struct machine *machine)
1688 {
1689         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1690
1691         if (sched->tp_handler->runtime_event)
1692                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1693
1694         return 0;
1695 }
1696
1697 static int perf_sched__process_fork_event(struct perf_tool *tool,
1698                                           union perf_event *event,
1699                                           struct perf_sample *sample,
1700                                           struct machine *machine)
1701 {
1702         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1703
1704         /* run the fork event through the perf machineruy */
1705         perf_event__process_fork(tool, event, sample, machine);
1706
1707         /* and then run additional processing needed for this command */
1708         if (sched->tp_handler->fork_event)
1709                 return sched->tp_handler->fork_event(sched, event, machine);
1710
1711         return 0;
1712 }
1713
1714 static int process_sched_migrate_task_event(struct perf_tool *tool,
1715                                             struct evsel *evsel,
1716                                             struct perf_sample *sample,
1717                                             struct machine *machine)
1718 {
1719         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1720
1721         if (sched->tp_handler->migrate_task_event)
1722                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1723
1724         return 0;
1725 }
1726
1727 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1728                                   struct evsel *evsel,
1729                                   struct perf_sample *sample,
1730                                   struct machine *machine);
1731
1732 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1733                                                  union perf_event *event __maybe_unused,
1734                                                  struct perf_sample *sample,
1735                                                  struct evsel *evsel,
1736                                                  struct machine *machine)
1737 {
1738         int err = 0;
1739
1740         if (evsel->handler != NULL) {
1741                 tracepoint_handler f = evsel->handler;
1742                 err = f(tool, evsel, sample, machine);
1743         }
1744
1745         return err;
1746 }
1747
1748 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1749                                     union perf_event *event,
1750                                     struct perf_sample *sample,
1751                                     struct machine *machine)
1752 {
1753         struct thread *thread;
1754         struct thread_runtime *tr;
1755         int err;
1756
1757         err = perf_event__process_comm(tool, event, sample, machine);
1758         if (err)
1759                 return err;
1760
1761         thread = machine__find_thread(machine, sample->pid, sample->tid);
1762         if (!thread) {
1763                 pr_err("Internal error: can't find thread\n");
1764                 return -1;
1765         }
1766
1767         tr = thread__get_runtime(thread);
1768         if (tr == NULL) {
1769                 thread__put(thread);
1770                 return -1;
1771         }
1772
1773         tr->comm_changed = true;
1774         thread__put(thread);
1775
1776         return 0;
1777 }
1778
1779 static int perf_sched__read_events(struct perf_sched *sched)
1780 {
1781         const struct evsel_str_handler handlers[] = {
1782                 { "sched:sched_switch",       process_sched_switch_event, },
1783                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1784                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1785                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1786                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1787         };
1788         struct perf_session *session;
1789         struct perf_data data = {
1790                 .path  = input_name,
1791                 .mode  = PERF_DATA_MODE_READ,
1792                 .force = sched->force,
1793         };
1794         int rc = -1;
1795
1796         session = perf_session__new(&data, false, &sched->tool);
1797         if (session == NULL) {
1798                 pr_debug("No Memory for session\n");
1799                 return -1;
1800         }
1801
1802         symbol__init(&session->header.env);
1803
1804         if (perf_session__set_tracepoints_handlers(session, handlers))
1805                 goto out_delete;
1806
1807         if (perf_session__has_traces(session, "record -R")) {
1808                 int err = perf_session__process_events(session);
1809                 if (err) {
1810                         pr_err("Failed to process events, error %d", err);
1811                         goto out_delete;
1812                 }
1813
1814                 sched->nr_events      = session->evlist->stats.nr_events[0];
1815                 sched->nr_lost_events = session->evlist->stats.total_lost;
1816                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1817         }
1818
1819         rc = 0;
1820 out_delete:
1821         perf_session__delete(session);
1822         return rc;
1823 }
1824
1825 /*
1826  * scheduling times are printed as msec.usec
1827  */
1828 static inline void print_sched_time(unsigned long long nsecs, int width)
1829 {
1830         unsigned long msecs;
1831         unsigned long usecs;
1832
1833         msecs  = nsecs / NSEC_PER_MSEC;
1834         nsecs -= msecs * NSEC_PER_MSEC;
1835         usecs  = nsecs / NSEC_PER_USEC;
1836         printf("%*lu.%03lu ", width, msecs, usecs);
1837 }
1838
1839 /*
1840  * returns runtime data for event, allocating memory for it the
1841  * first time it is used.
1842  */
1843 static struct evsel_runtime *perf_evsel__get_runtime(struct evsel *evsel)
1844 {
1845         struct evsel_runtime *r = evsel->priv;
1846
1847         if (r == NULL) {
1848                 r = zalloc(sizeof(struct evsel_runtime));
1849                 evsel->priv = r;
1850         }
1851
1852         return r;
1853 }
1854
1855 /*
1856  * save last time event was seen per cpu
1857  */
1858 static void perf_evsel__save_time(struct evsel *evsel,
1859                                   u64 timestamp, u32 cpu)
1860 {
1861         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1862
1863         if (r == NULL)
1864                 return;
1865
1866         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1867                 int i, n = __roundup_pow_of_two(cpu+1);
1868                 void *p = r->last_time;
1869
1870                 p = realloc(r->last_time, n * sizeof(u64));
1871                 if (!p)
1872                         return;
1873
1874                 r->last_time = p;
1875                 for (i = r->ncpu; i < n; ++i)
1876                         r->last_time[i] = (u64) 0;
1877
1878                 r->ncpu = n;
1879         }
1880
1881         r->last_time[cpu] = timestamp;
1882 }
1883
1884 /* returns last time this event was seen on the given cpu */
1885 static u64 perf_evsel__get_time(struct evsel *evsel, u32 cpu)
1886 {
1887         struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1888
1889         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1890                 return 0;
1891
1892         return r->last_time[cpu];
1893 }
1894
1895 static int comm_width = 30;
1896
1897 static char *timehist_get_commstr(struct thread *thread)
1898 {
1899         static char str[32];
1900         const char *comm = thread__comm_str(thread);
1901         pid_t tid = thread->tid;
1902         pid_t pid = thread->pid_;
1903         int n;
1904
1905         if (pid == 0)
1906                 n = scnprintf(str, sizeof(str), "%s", comm);
1907
1908         else if (tid != pid)
1909                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1910
1911         else
1912                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1913
1914         if (n > comm_width)
1915                 comm_width = n;
1916
1917         return str;
1918 }
1919
1920 static void timehist_header(struct perf_sched *sched)
1921 {
1922         u32 ncpus = sched->max_cpu + 1;
1923         u32 i, j;
1924
1925         printf("%15s %6s ", "time", "cpu");
1926
1927         if (sched->show_cpu_visual) {
1928                 printf(" ");
1929                 for (i = 0, j = 0; i < ncpus; ++i) {
1930                         printf("%x", j++);
1931                         if (j > 15)
1932                                 j = 0;
1933                 }
1934                 printf(" ");
1935         }
1936
1937         printf(" %-*s  %9s  %9s  %9s", comm_width,
1938                 "task name", "wait time", "sch delay", "run time");
1939
1940         if (sched->show_state)
1941                 printf("  %s", "state");
1942
1943         printf("\n");
1944
1945         /*
1946          * units row
1947          */
1948         printf("%15s %-6s ", "", "");
1949
1950         if (sched->show_cpu_visual)
1951                 printf(" %*s ", ncpus, "");
1952
1953         printf(" %-*s  %9s  %9s  %9s", comm_width,
1954                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1955
1956         if (sched->show_state)
1957                 printf("  %5s", "");
1958
1959         printf("\n");
1960
1961         /*
1962          * separator
1963          */
1964         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1965
1966         if (sched->show_cpu_visual)
1967                 printf(" %.*s ", ncpus, graph_dotted_line);
1968
1969         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1970                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1971                 graph_dotted_line);
1972
1973         if (sched->show_state)
1974                 printf("  %.5s", graph_dotted_line);
1975
1976         printf("\n");
1977 }
1978
1979 static char task_state_char(struct thread *thread, int state)
1980 {
1981         static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1982         unsigned bit = state ? ffs(state) : 0;
1983
1984         /* 'I' for idle */
1985         if (thread->tid == 0)
1986                 return 'I';
1987
1988         return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1989 }
1990
1991 static void timehist_print_sample(struct perf_sched *sched,
1992                                   struct evsel *evsel,
1993                                   struct perf_sample *sample,
1994                                   struct addr_location *al,
1995                                   struct thread *thread,
1996                                   u64 t, int state)
1997 {
1998         struct thread_runtime *tr = thread__priv(thread);
1999         const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
2000         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
2001         u32 max_cpus = sched->max_cpu + 1;
2002         char tstr[64];
2003         char nstr[30];
2004         u64 wait_time;
2005
2006         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2007         printf("%15s [%04d] ", tstr, sample->cpu);
2008
2009         if (sched->show_cpu_visual) {
2010                 u32 i;
2011                 char c;
2012
2013                 printf(" ");
2014                 for (i = 0; i < max_cpus; ++i) {
2015                         /* flag idle times with 'i'; others are sched events */
2016                         if (i == sample->cpu)
2017                                 c = (thread->tid == 0) ? 'i' : 's';
2018                         else
2019                                 c = ' ';
2020                         printf("%c", c);
2021                 }
2022                 printf(" ");
2023         }
2024
2025         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2026
2027         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2028         print_sched_time(wait_time, 6);
2029
2030         print_sched_time(tr->dt_delay, 6);
2031         print_sched_time(tr->dt_run, 6);
2032
2033         if (sched->show_state)
2034                 printf(" %5c ", task_state_char(thread, state));
2035
2036         if (sched->show_next) {
2037                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2038                 printf(" %-*s", comm_width, nstr);
2039         }
2040
2041         if (sched->show_wakeups && !sched->show_next)
2042                 printf("  %-*s", comm_width, "");
2043
2044         if (thread->tid == 0)
2045                 goto out;
2046
2047         if (sched->show_callchain)
2048                 printf("  ");
2049
2050         sample__fprintf_sym(sample, al, 0,
2051                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2052                             EVSEL__PRINT_CALLCHAIN_ARROW |
2053                             EVSEL__PRINT_SKIP_IGNORED,
2054                             &callchain_cursor, stdout);
2055
2056 out:
2057         printf("\n");
2058 }
2059
2060 /*
2061  * Explanation of delta-time stats:
2062  *
2063  *            t = time of current schedule out event
2064  *        tprev = time of previous sched out event
2065  *                also time of schedule-in event for current task
2066  *    last_time = time of last sched change event for current task
2067  *                (i.e, time process was last scheduled out)
2068  * ready_to_run = time of wakeup for current task
2069  *
2070  * -----|------------|------------|------------|------
2071  *    last         ready        tprev          t
2072  *    time         to run
2073  *
2074  *      |-------- dt_wait --------|
2075  *                   |- dt_delay -|-- dt_run --|
2076  *
2077  *   dt_run = run time of current task
2078  *  dt_wait = time between last schedule out event for task and tprev
2079  *            represents time spent off the cpu
2080  * dt_delay = time between wakeup and schedule-in of task
2081  */
2082
2083 static void timehist_update_runtime_stats(struct thread_runtime *r,
2084                                          u64 t, u64 tprev)
2085 {
2086         r->dt_delay   = 0;
2087         r->dt_sleep   = 0;
2088         r->dt_iowait  = 0;
2089         r->dt_preempt = 0;
2090         r->dt_run     = 0;
2091
2092         if (tprev) {
2093                 r->dt_run = t - tprev;
2094                 if (r->ready_to_run) {
2095                         if (r->ready_to_run > tprev)
2096                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2097                         else
2098                                 r->dt_delay = tprev - r->ready_to_run;
2099                 }
2100
2101                 if (r->last_time > tprev)
2102                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2103                 else if (r->last_time) {
2104                         u64 dt_wait = tprev - r->last_time;
2105
2106                         if (r->last_state == TASK_RUNNING)
2107                                 r->dt_preempt = dt_wait;
2108                         else if (r->last_state == TASK_UNINTERRUPTIBLE)
2109                                 r->dt_iowait = dt_wait;
2110                         else
2111                                 r->dt_sleep = dt_wait;
2112                 }
2113         }
2114
2115         update_stats(&r->run_stats, r->dt_run);
2116
2117         r->total_run_time     += r->dt_run;
2118         r->total_delay_time   += r->dt_delay;
2119         r->total_sleep_time   += r->dt_sleep;
2120         r->total_iowait_time  += r->dt_iowait;
2121         r->total_preempt_time += r->dt_preempt;
2122 }
2123
2124 static bool is_idle_sample(struct perf_sample *sample,
2125                            struct evsel *evsel)
2126 {
2127         /* pid 0 == swapper == idle task */
2128         if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2129                 return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2130
2131         return sample->pid == 0;
2132 }
2133
2134 static void save_task_callchain(struct perf_sched *sched,
2135                                 struct perf_sample *sample,
2136                                 struct evsel *evsel,
2137                                 struct machine *machine)
2138 {
2139         struct callchain_cursor *cursor = &callchain_cursor;
2140         struct thread *thread;
2141
2142         /* want main thread for process - has maps */
2143         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2144         if (thread == NULL) {
2145                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2146                 return;
2147         }
2148
2149         if (!sched->show_callchain || sample->callchain == NULL)
2150                 return;
2151
2152         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2153                                       NULL, NULL, sched->max_stack + 2) != 0) {
2154                 if (verbose > 0)
2155                         pr_err("Failed to resolve callchain. Skipping\n");
2156
2157                 return;
2158         }
2159
2160         callchain_cursor_commit(cursor);
2161
2162         while (true) {
2163                 struct callchain_cursor_node *node;
2164                 struct symbol *sym;
2165
2166                 node = callchain_cursor_current(cursor);
2167                 if (node == NULL)
2168                         break;
2169
2170                 sym = node->sym;
2171                 if (sym) {
2172                         if (!strcmp(sym->name, "schedule") ||
2173                             !strcmp(sym->name, "__schedule") ||
2174                             !strcmp(sym->name, "preempt_schedule"))
2175                                 sym->ignore = 1;
2176                 }
2177
2178                 callchain_cursor_advance(cursor);
2179         }
2180 }
2181
2182 static int init_idle_thread(struct thread *thread)
2183 {
2184         struct idle_thread_runtime *itr;
2185
2186         thread__set_comm(thread, idle_comm, 0);
2187
2188         itr = zalloc(sizeof(*itr));
2189         if (itr == NULL)
2190                 return -ENOMEM;
2191
2192         init_stats(&itr->tr.run_stats);
2193         callchain_init(&itr->callchain);
2194         callchain_cursor_reset(&itr->cursor);
2195         thread__set_priv(thread, itr);
2196
2197         return 0;
2198 }
2199
2200 /*
2201  * Track idle stats per cpu by maintaining a local thread
2202  * struct for the idle task on each cpu.
2203  */
2204 static int init_idle_threads(int ncpu)
2205 {
2206         int i, ret;
2207
2208         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2209         if (!idle_threads)
2210                 return -ENOMEM;
2211
2212         idle_max_cpu = ncpu;
2213
2214         /* allocate the actual thread struct if needed */
2215         for (i = 0; i < ncpu; ++i) {
2216                 idle_threads[i] = thread__new(0, 0);
2217                 if (idle_threads[i] == NULL)
2218                         return -ENOMEM;
2219
2220                 ret = init_idle_thread(idle_threads[i]);
2221                 if (ret < 0)
2222                         return ret;
2223         }
2224
2225         return 0;
2226 }
2227
2228 static void free_idle_threads(void)
2229 {
2230         int i;
2231
2232         if (idle_threads == NULL)
2233                 return;
2234
2235         for (i = 0; i < idle_max_cpu; ++i) {
2236                 if ((idle_threads[i]))
2237                         thread__delete(idle_threads[i]);
2238         }
2239
2240         free(idle_threads);
2241 }
2242
2243 static struct thread *get_idle_thread(int cpu)
2244 {
2245         /*
2246          * expand/allocate array of pointers to local thread
2247          * structs if needed
2248          */
2249         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2250                 int i, j = __roundup_pow_of_two(cpu+1);
2251                 void *p;
2252
2253                 p = realloc(idle_threads, j * sizeof(struct thread *));
2254                 if (!p)
2255                         return NULL;
2256
2257                 idle_threads = (struct thread **) p;
2258                 for (i = idle_max_cpu; i < j; ++i)
2259                         idle_threads[i] = NULL;
2260
2261                 idle_max_cpu = j;
2262         }
2263
2264         /* allocate a new thread struct if needed */
2265         if (idle_threads[cpu] == NULL) {
2266                 idle_threads[cpu] = thread__new(0, 0);
2267
2268                 if (idle_threads[cpu]) {
2269                         if (init_idle_thread(idle_threads[cpu]) < 0)
2270                                 return NULL;
2271                 }
2272         }
2273
2274         return idle_threads[cpu];
2275 }
2276
2277 static void save_idle_callchain(struct perf_sched *sched,
2278                                 struct idle_thread_runtime *itr,
2279                                 struct perf_sample *sample)
2280 {
2281         if (!sched->show_callchain || sample->callchain == NULL)
2282                 return;
2283
2284         callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2285 }
2286
2287 static struct thread *timehist_get_thread(struct perf_sched *sched,
2288                                           struct perf_sample *sample,
2289                                           struct machine *machine,
2290                                           struct evsel *evsel)
2291 {
2292         struct thread *thread;
2293
2294         if (is_idle_sample(sample, evsel)) {
2295                 thread = get_idle_thread(sample->cpu);
2296                 if (thread == NULL)
2297                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2298
2299         } else {
2300                 /* there were samples with tid 0 but non-zero pid */
2301                 thread = machine__findnew_thread(machine, sample->pid,
2302                                                  sample->tid ?: sample->pid);
2303                 if (thread == NULL) {
2304                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2305                                  sample->tid);
2306                 }
2307
2308                 save_task_callchain(sched, sample, evsel, machine);
2309                 if (sched->idle_hist) {
2310                         struct thread *idle;
2311                         struct idle_thread_runtime *itr;
2312
2313                         idle = get_idle_thread(sample->cpu);
2314                         if (idle == NULL) {
2315                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2316                                 return NULL;
2317                         }
2318
2319                         itr = thread__priv(idle);
2320                         if (itr == NULL)
2321                                 return NULL;
2322
2323                         itr->last_thread = thread;
2324
2325                         /* copy task callchain when entering to idle */
2326                         if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2327                                 save_idle_callchain(sched, itr, sample);
2328                 }
2329         }
2330
2331         return thread;
2332 }
2333
2334 static bool timehist_skip_sample(struct perf_sched *sched,
2335                                  struct thread *thread,
2336                                  struct evsel *evsel,
2337                                  struct perf_sample *sample)
2338 {
2339         bool rc = false;
2340
2341         if (thread__is_filtered(thread)) {
2342                 rc = true;
2343                 sched->skipped_samples++;
2344         }
2345
2346         if (sched->idle_hist) {
2347                 if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2348                         rc = true;
2349                 else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2350                          perf_evsel__intval(evsel, sample, "next_pid") != 0)
2351                         rc = true;
2352         }
2353
2354         return rc;
2355 }
2356
2357 static void timehist_print_wakeup_event(struct perf_sched *sched,
2358                                         struct evsel *evsel,
2359                                         struct perf_sample *sample,
2360                                         struct machine *machine,
2361                                         struct thread *awakened)
2362 {
2363         struct thread *thread;
2364         char tstr[64];
2365
2366         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2367         if (thread == NULL)
2368                 return;
2369
2370         /* show wakeup unless both awakee and awaker are filtered */
2371         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2372             timehist_skip_sample(sched, awakened, evsel, sample)) {
2373                 return;
2374         }
2375
2376         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2377         printf("%15s [%04d] ", tstr, sample->cpu);
2378         if (sched->show_cpu_visual)
2379                 printf(" %*s ", sched->max_cpu + 1, "");
2380
2381         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2382
2383         /* dt spacer */
2384         printf("  %9s  %9s  %9s ", "", "", "");
2385
2386         printf("awakened: %s", timehist_get_commstr(awakened));
2387
2388         printf("\n");
2389 }
2390
2391 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2392                                        union perf_event *event __maybe_unused,
2393                                        struct evsel *evsel,
2394                                        struct perf_sample *sample,
2395                                        struct machine *machine)
2396 {
2397         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2398         struct thread *thread;
2399         struct thread_runtime *tr = NULL;
2400         /* want pid of awakened task not pid in sample */
2401         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2402
2403         thread = machine__findnew_thread(machine, 0, pid);
2404         if (thread == NULL)
2405                 return -1;
2406
2407         tr = thread__get_runtime(thread);
2408         if (tr == NULL)
2409                 return -1;
2410
2411         if (tr->ready_to_run == 0)
2412                 tr->ready_to_run = sample->time;
2413
2414         /* show wakeups if requested */
2415         if (sched->show_wakeups &&
2416             !perf_time__skip_sample(&sched->ptime, sample->time))
2417                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2418
2419         return 0;
2420 }
2421
2422 static void timehist_print_migration_event(struct perf_sched *sched,
2423                                         struct evsel *evsel,
2424                                         struct perf_sample *sample,
2425                                         struct machine *machine,
2426                                         struct thread *migrated)
2427 {
2428         struct thread *thread;
2429         char tstr[64];
2430         u32 max_cpus = sched->max_cpu + 1;
2431         u32 ocpu, dcpu;
2432
2433         if (sched->summary_only)
2434                 return;
2435
2436         max_cpus = sched->max_cpu + 1;
2437         ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2438         dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2439
2440         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2441         if (thread == NULL)
2442                 return;
2443
2444         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2445             timehist_skip_sample(sched, migrated, evsel, sample)) {
2446                 return;
2447         }
2448
2449         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2450         printf("%15s [%04d] ", tstr, sample->cpu);
2451
2452         if (sched->show_cpu_visual) {
2453                 u32 i;
2454                 char c;
2455
2456                 printf("  ");
2457                 for (i = 0; i < max_cpus; ++i) {
2458                         c = (i == sample->cpu) ? 'm' : ' ';
2459                         printf("%c", c);
2460                 }
2461                 printf("  ");
2462         }
2463
2464         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2465
2466         /* dt spacer */
2467         printf("  %9s  %9s  %9s ", "", "", "");
2468
2469         printf("migrated: %s", timehist_get_commstr(migrated));
2470         printf(" cpu %d => %d", ocpu, dcpu);
2471
2472         printf("\n");
2473 }
2474
2475 static int timehist_migrate_task_event(struct perf_tool *tool,
2476                                        union perf_event *event __maybe_unused,
2477                                        struct evsel *evsel,
2478                                        struct perf_sample *sample,
2479                                        struct machine *machine)
2480 {
2481         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2482         struct thread *thread;
2483         struct thread_runtime *tr = NULL;
2484         /* want pid of migrated task not pid in sample */
2485         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2486
2487         thread = machine__findnew_thread(machine, 0, pid);
2488         if (thread == NULL)
2489                 return -1;
2490
2491         tr = thread__get_runtime(thread);
2492         if (tr == NULL)
2493                 return -1;
2494
2495         tr->migrations++;
2496
2497         /* show migrations if requested */
2498         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2499
2500         return 0;
2501 }
2502
2503 static int timehist_sched_change_event(struct perf_tool *tool,
2504                                        union perf_event *event,
2505                                        struct evsel *evsel,
2506                                        struct perf_sample *sample,
2507                                        struct machine *machine)
2508 {
2509         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2510         struct perf_time_interval *ptime = &sched->ptime;
2511         struct addr_location al;
2512         struct thread *thread;
2513         struct thread_runtime *tr = NULL;
2514         u64 tprev, t = sample->time;
2515         int rc = 0;
2516         int state = perf_evsel__intval(evsel, sample, "prev_state");
2517
2518
2519         if (machine__resolve(machine, &al, sample) < 0) {
2520                 pr_err("problem processing %d event. skipping it\n",
2521                        event->header.type);
2522                 rc = -1;
2523                 goto out;
2524         }
2525
2526         thread = timehist_get_thread(sched, sample, machine, evsel);
2527         if (thread == NULL) {
2528                 rc = -1;
2529                 goto out;
2530         }
2531
2532         if (timehist_skip_sample(sched, thread, evsel, sample))
2533                 goto out;
2534
2535         tr = thread__get_runtime(thread);
2536         if (tr == NULL) {
2537                 rc = -1;
2538                 goto out;
2539         }
2540
2541         tprev = perf_evsel__get_time(evsel, sample->cpu);
2542
2543         /*
2544          * If start time given:
2545          * - sample time is under window user cares about - skip sample
2546          * - tprev is under window user cares about  - reset to start of window
2547          */
2548         if (ptime->start && ptime->start > t)
2549                 goto out;
2550
2551         if (tprev && ptime->start > tprev)
2552                 tprev = ptime->start;
2553
2554         /*
2555          * If end time given:
2556          * - previous sched event is out of window - we are done
2557          * - sample time is beyond window user cares about - reset it
2558          *   to close out stats for time window interest
2559          */
2560         if (ptime->end) {
2561                 if (tprev > ptime->end)
2562                         goto out;
2563
2564                 if (t > ptime->end)
2565                         t = ptime->end;
2566         }
2567
2568         if (!sched->idle_hist || thread->tid == 0) {
2569                 timehist_update_runtime_stats(tr, t, tprev);
2570
2571                 if (sched->idle_hist) {
2572                         struct idle_thread_runtime *itr = (void *)tr;
2573                         struct thread_runtime *last_tr;
2574
2575                         BUG_ON(thread->tid != 0);
2576
2577                         if (itr->last_thread == NULL)
2578                                 goto out;
2579
2580                         /* add current idle time as last thread's runtime */
2581                         last_tr = thread__get_runtime(itr->last_thread);
2582                         if (last_tr == NULL)
2583                                 goto out;
2584
2585                         timehist_update_runtime_stats(last_tr, t, tprev);
2586                         /*
2587                          * remove delta time of last thread as it's not updated
2588                          * and otherwise it will show an invalid value next
2589                          * time.  we only care total run time and run stat.
2590                          */
2591                         last_tr->dt_run = 0;
2592                         last_tr->dt_delay = 0;
2593                         last_tr->dt_sleep = 0;
2594                         last_tr->dt_iowait = 0;
2595                         last_tr->dt_preempt = 0;
2596
2597                         if (itr->cursor.nr)
2598                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2599
2600                         itr->last_thread = NULL;
2601                 }
2602         }
2603
2604         if (!sched->summary_only)
2605                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2606
2607 out:
2608         if (sched->hist_time.start == 0 && t >= ptime->start)
2609                 sched->hist_time.start = t;
2610         if (ptime->end == 0 || t <= ptime->end)
2611                 sched->hist_time.end = t;
2612
2613         if (tr) {
2614                 /* time of this sched_switch event becomes last time task seen */
2615                 tr->last_time = sample->time;
2616
2617                 /* last state is used to determine where to account wait time */
2618                 tr->last_state = state;
2619
2620                 /* sched out event for task so reset ready to run time */
2621                 tr->ready_to_run = 0;
2622         }
2623
2624         perf_evsel__save_time(evsel, sample->time, sample->cpu);
2625
2626         return rc;
2627 }
2628
2629 static int timehist_sched_switch_event(struct perf_tool *tool,
2630                              union perf_event *event,
2631                              struct evsel *evsel,
2632                              struct perf_sample *sample,
2633                              struct machine *machine __maybe_unused)
2634 {
2635         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2636 }
2637
2638 static int process_lost(struct perf_tool *tool __maybe_unused,
2639                         union perf_event *event,
2640                         struct perf_sample *sample,
2641                         struct machine *machine __maybe_unused)
2642 {
2643         char tstr[64];
2644
2645         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2646         printf("%15s ", tstr);
2647         printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2648
2649         return 0;
2650 }
2651
2652
2653 static void print_thread_runtime(struct thread *t,
2654                                  struct thread_runtime *r)
2655 {
2656         double mean = avg_stats(&r->run_stats);
2657         float stddev;
2658
2659         printf("%*s   %5d  %9" PRIu64 " ",
2660                comm_width, timehist_get_commstr(t), t->ppid,
2661                (u64) r->run_stats.n);
2662
2663         print_sched_time(r->total_run_time, 8);
2664         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2665         print_sched_time(r->run_stats.min, 6);
2666         printf(" ");
2667         print_sched_time((u64) mean, 6);
2668         printf(" ");
2669         print_sched_time(r->run_stats.max, 6);
2670         printf("  ");
2671         printf("%5.2f", stddev);
2672         printf("   %5" PRIu64, r->migrations);
2673         printf("\n");
2674 }
2675
2676 static void print_thread_waittime(struct thread *t,
2677                                   struct thread_runtime *r)
2678 {
2679         printf("%*s   %5d  %9" PRIu64 " ",
2680                comm_width, timehist_get_commstr(t), t->ppid,
2681                (u64) r->run_stats.n);
2682
2683         print_sched_time(r->total_run_time, 8);
2684         print_sched_time(r->total_sleep_time, 6);
2685         printf(" ");
2686         print_sched_time(r->total_iowait_time, 6);
2687         printf(" ");
2688         print_sched_time(r->total_preempt_time, 6);
2689         printf(" ");
2690         print_sched_time(r->total_delay_time, 6);
2691         printf("\n");
2692 }
2693
2694 struct total_run_stats {
2695         struct perf_sched *sched;
2696         u64  sched_count;
2697         u64  task_count;
2698         u64  total_run_time;
2699 };
2700
2701 static int __show_thread_runtime(struct thread *t, void *priv)
2702 {
2703         struct total_run_stats *stats = priv;
2704         struct thread_runtime *r;
2705
2706         if (thread__is_filtered(t))
2707                 return 0;
2708
2709         r = thread__priv(t);
2710         if (r && r->run_stats.n) {
2711                 stats->task_count++;
2712                 stats->sched_count += r->run_stats.n;
2713                 stats->total_run_time += r->total_run_time;
2714
2715                 if (stats->sched->show_state)
2716                         print_thread_waittime(t, r);
2717                 else
2718                         print_thread_runtime(t, r);
2719         }
2720
2721         return 0;
2722 }
2723
2724 static int show_thread_runtime(struct thread *t, void *priv)
2725 {
2726         if (t->dead)
2727                 return 0;
2728
2729         return __show_thread_runtime(t, priv);
2730 }
2731
2732 static int show_deadthread_runtime(struct thread *t, void *priv)
2733 {
2734         if (!t->dead)
2735                 return 0;
2736
2737         return __show_thread_runtime(t, priv);
2738 }
2739
2740 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2741 {
2742         const char *sep = " <- ";
2743         struct callchain_list *chain;
2744         size_t ret = 0;
2745         char bf[1024];
2746         bool first;
2747
2748         if (node == NULL)
2749                 return 0;
2750
2751         ret = callchain__fprintf_folded(fp, node->parent);
2752         first = (ret == 0);
2753
2754         list_for_each_entry(chain, &node->val, list) {
2755                 if (chain->ip >= PERF_CONTEXT_MAX)
2756                         continue;
2757                 if (chain->ms.sym && chain->ms.sym->ignore)
2758                         continue;
2759                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2760                                callchain_list__sym_name(chain, bf, sizeof(bf),
2761                                                         false));
2762                 first = false;
2763         }
2764
2765         return ret;
2766 }
2767
2768 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2769 {
2770         size_t ret = 0;
2771         FILE *fp = stdout;
2772         struct callchain_node *chain;
2773         struct rb_node *rb_node = rb_first_cached(root);
2774
2775         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2776         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2777                graph_dotted_line);
2778
2779         while (rb_node) {
2780                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2781                 rb_node = rb_next(rb_node);
2782
2783                 ret += fprintf(fp, "  ");
2784                 print_sched_time(chain->hit, 12);
2785                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2786                 ret += fprintf(fp, " %8d  ", chain->count);
2787                 ret += callchain__fprintf_folded(fp, chain);
2788                 ret += fprintf(fp, "\n");
2789         }
2790
2791         return ret;
2792 }
2793
2794 static void timehist_print_summary(struct perf_sched *sched,
2795                                    struct perf_session *session)
2796 {
2797         struct machine *m = &session->machines.host;
2798         struct total_run_stats totals;
2799         u64 task_count;
2800         struct thread *t;
2801         struct thread_runtime *r;
2802         int i;
2803         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2804
2805         memset(&totals, 0, sizeof(totals));
2806         totals.sched = sched;
2807
2808         if (sched->idle_hist) {
2809                 printf("\nIdle-time summary\n");
2810                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2811                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2812         } else if (sched->show_state) {
2813                 printf("\nWait-time summary\n");
2814                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2815                 printf("   run-time      sleep      iowait     preempt       delay\n");
2816         } else {
2817                 printf("\nRuntime summary\n");
2818                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2819                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2820         }
2821         printf("%*s            (count)  ", comm_width, "");
2822         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2823                sched->show_state ? "(msec)" : "%");
2824         printf("%.117s\n", graph_dotted_line);
2825
2826         machine__for_each_thread(m, show_thread_runtime, &totals);
2827         task_count = totals.task_count;
2828         if (!task_count)
2829                 printf("<no still running tasks>\n");
2830
2831         printf("\nTerminated tasks:\n");
2832         machine__for_each_thread(m, show_deadthread_runtime, &totals);
2833         if (task_count == totals.task_count)
2834                 printf("<no terminated tasks>\n");
2835
2836         /* CPU idle stats not tracked when samples were skipped */
2837         if (sched->skipped_samples && !sched->idle_hist)
2838                 return;
2839
2840         printf("\nIdle stats:\n");
2841         for (i = 0; i < idle_max_cpu; ++i) {
2842                 t = idle_threads[i];
2843                 if (!t)
2844                         continue;
2845
2846                 r = thread__priv(t);
2847                 if (r && r->run_stats.n) {
2848                         totals.sched_count += r->run_stats.n;
2849                         printf("    CPU %2d idle for ", i);
2850                         print_sched_time(r->total_run_time, 6);
2851                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2852                 } else
2853                         printf("    CPU %2d idle entire time window\n", i);
2854         }
2855
2856         if (sched->idle_hist && sched->show_callchain) {
2857                 callchain_param.mode  = CHAIN_FOLDED;
2858                 callchain_param.value = CCVAL_PERIOD;
2859
2860                 callchain_register_param(&callchain_param);
2861
2862                 printf("\nIdle stats by callchain:\n");
2863                 for (i = 0; i < idle_max_cpu; ++i) {
2864                         struct idle_thread_runtime *itr;
2865
2866                         t = idle_threads[i];
2867                         if (!t)
2868                                 continue;
2869
2870                         itr = thread__priv(t);
2871                         if (itr == NULL)
2872                                 continue;
2873
2874                         callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2875                                              0, &callchain_param);
2876
2877                         printf("  CPU %2d:", i);
2878                         print_sched_time(itr->tr.total_run_time, 6);
2879                         printf(" msec\n");
2880                         timehist_print_idlehist_callchain(&itr->sorted_root);
2881                         printf("\n");
2882                 }
2883         }
2884
2885         printf("\n"
2886                "    Total number of unique tasks: %" PRIu64 "\n"
2887                "Total number of context switches: %" PRIu64 "\n",
2888                totals.task_count, totals.sched_count);
2889
2890         printf("           Total run time (msec): ");
2891         print_sched_time(totals.total_run_time, 2);
2892         printf("\n");
2893
2894         printf("    Total scheduling time (msec): ");
2895         print_sched_time(hist_time, 2);
2896         printf(" (x %d)\n", sched->max_cpu);
2897 }
2898
2899 typedef int (*sched_handler)(struct perf_tool *tool,
2900                           union perf_event *event,
2901                           struct evsel *evsel,
2902                           struct perf_sample *sample,
2903                           struct machine *machine);
2904
2905 static int perf_timehist__process_sample(struct perf_tool *tool,
2906                                          union perf_event *event,
2907                                          struct perf_sample *sample,
2908                                          struct evsel *evsel,
2909                                          struct machine *machine)
2910 {
2911         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2912         int err = 0;
2913         int this_cpu = sample->cpu;
2914
2915         if (this_cpu > sched->max_cpu)
2916                 sched->max_cpu = this_cpu;
2917
2918         if (evsel->handler != NULL) {
2919                 sched_handler f = evsel->handler;
2920
2921                 err = f(tool, event, evsel, sample, machine);
2922         }
2923
2924         return err;
2925 }
2926
2927 static int timehist_check_attr(struct perf_sched *sched,
2928                                struct evlist *evlist)
2929 {
2930         struct evsel *evsel;
2931         struct evsel_runtime *er;
2932
2933         list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2934                 er = perf_evsel__get_runtime(evsel);
2935                 if (er == NULL) {
2936                         pr_err("Failed to allocate memory for evsel runtime data\n");
2937                         return -1;
2938                 }
2939
2940                 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2941                         pr_info("Samples do not have callchains.\n");
2942                         sched->show_callchain = 0;
2943                         symbol_conf.use_callchain = 0;
2944                 }
2945         }
2946
2947         return 0;
2948 }
2949
2950 static int perf_sched__timehist(struct perf_sched *sched)
2951 {
2952         const struct evsel_str_handler handlers[] = {
2953                 { "sched:sched_switch",       timehist_sched_switch_event, },
2954                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2955                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2956         };
2957         const struct evsel_str_handler migrate_handlers[] = {
2958                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2959         };
2960         struct perf_data data = {
2961                 .path  = input_name,
2962                 .mode  = PERF_DATA_MODE_READ,
2963                 .force = sched->force,
2964         };
2965
2966         struct perf_session *session;
2967         struct evlist *evlist;
2968         int err = -1;
2969
2970         /*
2971          * event handlers for timehist option
2972          */
2973         sched->tool.sample       = perf_timehist__process_sample;
2974         sched->tool.mmap         = perf_event__process_mmap;
2975         sched->tool.comm         = perf_event__process_comm;
2976         sched->tool.exit         = perf_event__process_exit;
2977         sched->tool.fork         = perf_event__process_fork;
2978         sched->tool.lost         = process_lost;
2979         sched->tool.attr         = perf_event__process_attr;
2980         sched->tool.tracing_data = perf_event__process_tracing_data;
2981         sched->tool.build_id     = perf_event__process_build_id;
2982
2983         sched->tool.ordered_events = true;
2984         sched->tool.ordering_requires_timestamps = true;
2985
2986         symbol_conf.use_callchain = sched->show_callchain;
2987
2988         session = perf_session__new(&data, false, &sched->tool);
2989         if (session == NULL)
2990                 return -ENOMEM;
2991
2992         evlist = session->evlist;
2993
2994         symbol__init(&session->header.env);
2995
2996         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2997                 pr_err("Invalid time string\n");
2998                 return -EINVAL;
2999         }
3000
3001         if (timehist_check_attr(sched, evlist) != 0)
3002                 goto out;
3003
3004         setup_pager();
3005
3006         /* setup per-evsel handlers */
3007         if (perf_session__set_tracepoints_handlers(session, handlers))
3008                 goto out;
3009
3010         /* sched_switch event at a minimum needs to exist */
3011         if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3012                                                   "sched:sched_switch")) {
3013                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3014                 goto out;
3015         }
3016
3017         if (sched->show_migrations &&
3018             perf_session__set_tracepoints_handlers(session, migrate_handlers))
3019                 goto out;
3020
3021         /* pre-allocate struct for per-CPU idle stats */
3022         sched->max_cpu = session->header.env.nr_cpus_online;
3023         if (sched->max_cpu == 0)
3024                 sched->max_cpu = 4;
3025         if (init_idle_threads(sched->max_cpu))
3026                 goto out;
3027
3028         /* summary_only implies summary option, but don't overwrite summary if set */
3029         if (sched->summary_only)
3030                 sched->summary = sched->summary_only;
3031
3032         if (!sched->summary_only)
3033                 timehist_header(sched);
3034
3035         err = perf_session__process_events(session);
3036         if (err) {
3037                 pr_err("Failed to process events, error %d", err);
3038                 goto out;
3039         }
3040
3041         sched->nr_events      = evlist->stats.nr_events[0];
3042         sched->nr_lost_events = evlist->stats.total_lost;
3043         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3044
3045         if (sched->summary)
3046                 timehist_print_summary(sched, session);
3047
3048 out:
3049         free_idle_threads();
3050         perf_session__delete(session);
3051
3052         return err;
3053 }
3054
3055
3056 static void print_bad_events(struct perf_sched *sched)
3057 {
3058         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3059                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3060                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3061                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3062         }
3063         if (sched->nr_lost_events && sched->nr_events) {
3064                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3065                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3066                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3067         }
3068         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3069                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3070                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3071                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3072                 if (sched->nr_lost_events)
3073                         printf(" (due to lost events?)");
3074                 printf("\n");
3075         }
3076 }
3077
3078 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3079 {
3080         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3081         struct work_atoms *this;
3082         const char *comm = thread__comm_str(data->thread), *this_comm;
3083         bool leftmost = true;
3084
3085         while (*new) {
3086                 int cmp;
3087
3088                 this = container_of(*new, struct work_atoms, node);
3089                 parent = *new;
3090
3091                 this_comm = thread__comm_str(this->thread);
3092                 cmp = strcmp(comm, this_comm);
3093                 if (cmp > 0) {
3094                         new = &((*new)->rb_left);
3095                 } else if (cmp < 0) {
3096                         new = &((*new)->rb_right);
3097                         leftmost = false;
3098                 } else {
3099                         this->num_merged++;
3100                         this->total_runtime += data->total_runtime;
3101                         this->nb_atoms += data->nb_atoms;
3102                         this->total_lat += data->total_lat;
3103                         list_splice(&data->work_list, &this->work_list);
3104                         if (this->max_lat < data->max_lat) {
3105                                 this->max_lat = data->max_lat;
3106                                 this->max_lat_at = data->max_lat_at;
3107                         }
3108                         zfree(&data);
3109                         return;
3110                 }
3111         }
3112
3113         data->num_merged++;
3114         rb_link_node(&data->node, parent, new);
3115         rb_insert_color_cached(&data->node, root, leftmost);
3116 }
3117
3118 static void perf_sched__merge_lat(struct perf_sched *sched)
3119 {
3120         struct work_atoms *data;
3121         struct rb_node *node;
3122
3123         if (sched->skip_merge)
3124                 return;
3125
3126         while ((node = rb_first_cached(&sched->atom_root))) {
3127                 rb_erase_cached(node, &sched->atom_root);
3128                 data = rb_entry(node, struct work_atoms, node);
3129                 __merge_work_atoms(&sched->merged_atom_root, data);
3130         }
3131 }
3132
3133 static int perf_sched__lat(struct perf_sched *sched)
3134 {
3135         struct rb_node *next;
3136
3137         setup_pager();
3138
3139         if (perf_sched__read_events(sched))
3140                 return -1;
3141
3142         perf_sched__merge_lat(sched);
3143         perf_sched__sort_lat(sched);
3144
3145         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3146         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3147         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3148
3149         next = rb_first_cached(&sched->sorted_atom_root);
3150
3151         while (next) {
3152                 struct work_atoms *work_list;
3153
3154                 work_list = rb_entry(next, struct work_atoms, node);
3155                 output_lat_thread(sched, work_list);
3156                 next = rb_next(next);
3157                 thread__zput(work_list->thread);
3158         }
3159
3160         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3161         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3162                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3163
3164         printf(" ---------------------------------------------------\n");
3165
3166         print_bad_events(sched);
3167         printf("\n");
3168
3169         return 0;
3170 }
3171
3172 static int setup_map_cpus(struct perf_sched *sched)
3173 {
3174         struct perf_cpu_map *map;
3175
3176         sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3177
3178         if (sched->map.comp) {
3179                 sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3180                 if (!sched->map.comp_cpus)
3181                         return -1;
3182         }
3183
3184         if (!sched->map.cpus_str)
3185                 return 0;
3186
3187         map = perf_cpu_map__new(sched->map.cpus_str);
3188         if (!map) {
3189                 pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3190                 return -1;
3191         }
3192
3193         sched->map.cpus = map;
3194         return 0;
3195 }
3196
3197 static int setup_color_pids(struct perf_sched *sched)
3198 {
3199         struct perf_thread_map *map;
3200
3201         if (!sched->map.color_pids_str)
3202                 return 0;
3203
3204         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3205         if (!map) {
3206                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3207                 return -1;
3208         }
3209
3210         sched->map.color_pids = map;
3211         return 0;
3212 }
3213
3214 static int setup_color_cpus(struct perf_sched *sched)
3215 {
3216         struct perf_cpu_map *map;
3217
3218         if (!sched->map.color_cpus_str)
3219                 return 0;
3220
3221         map = perf_cpu_map__new(sched->map.color_cpus_str);
3222         if (!map) {
3223                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3224                 return -1;
3225         }
3226
3227         sched->map.color_cpus = map;
3228         return 0;
3229 }
3230
3231 static int perf_sched__map(struct perf_sched *sched)
3232 {
3233         if (setup_map_cpus(sched))
3234                 return -1;
3235
3236         if (setup_color_pids(sched))
3237                 return -1;
3238
3239         if (setup_color_cpus(sched))
3240                 return -1;
3241
3242         setup_pager();
3243         if (perf_sched__read_events(sched))
3244                 return -1;
3245         print_bad_events(sched);
3246         return 0;
3247 }
3248
3249 static int perf_sched__replay(struct perf_sched *sched)
3250 {
3251         unsigned long i;
3252
3253         calibrate_run_measurement_overhead(sched);
3254         calibrate_sleep_measurement_overhead(sched);
3255
3256         test_calibrations(sched);
3257
3258         if (perf_sched__read_events(sched))
3259                 return -1;
3260
3261         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3262         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3263         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3264
3265         if (sched->targetless_wakeups)
3266                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3267         if (sched->multitarget_wakeups)
3268                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3269         if (sched->nr_run_events_optimized)
3270                 printf("run atoms optimized: %ld\n",
3271                         sched->nr_run_events_optimized);
3272
3273         print_task_traces(sched);
3274         add_cross_task_wakeups(sched);
3275
3276         create_tasks(sched);
3277         printf("------------------------------------------------------------\n");
3278         for (i = 0; i < sched->replay_repeat; i++)
3279                 run_one_test(sched);
3280
3281         return 0;
3282 }
3283
3284 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3285                           const char * const usage_msg[])
3286 {
3287         char *tmp, *tok, *str = strdup(sched->sort_order);
3288
3289         for (tok = strtok_r(str, ", ", &tmp);
3290                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3291                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3292                         usage_with_options_msg(usage_msg, options,
3293                                         "Unknown --sort key: `%s'", tok);
3294                 }
3295         }
3296
3297         free(str);
3298
3299         sort_dimension__add("pid", &sched->cmp_pid);
3300 }
3301
3302 static int __cmd_record(int argc, const char **argv)
3303 {
3304         unsigned int rec_argc, i, j;
3305         const char **rec_argv;
3306         const char * const record_args[] = {
3307                 "record",
3308                 "-a",
3309                 "-R",
3310                 "-m", "1024",
3311                 "-c", "1",
3312                 "-e", "sched:sched_switch",
3313                 "-e", "sched:sched_stat_wait",
3314                 "-e", "sched:sched_stat_sleep",
3315                 "-e", "sched:sched_stat_iowait",
3316                 "-e", "sched:sched_stat_runtime",
3317                 "-e", "sched:sched_process_fork",
3318                 "-e", "sched:sched_wakeup",
3319                 "-e", "sched:sched_wakeup_new",
3320                 "-e", "sched:sched_migrate_task",
3321         };
3322
3323         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3324         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3325
3326         if (rec_argv == NULL)
3327                 return -ENOMEM;
3328
3329         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3330                 rec_argv[i] = strdup(record_args[i]);
3331
3332         for (j = 1; j < (unsigned int)argc; j++, i++)
3333                 rec_argv[i] = argv[j];
3334
3335         BUG_ON(i != rec_argc);
3336
3337         return cmd_record(i, rec_argv);
3338 }
3339
3340 int cmd_sched(int argc, const char **argv)
3341 {
3342         static const char default_sort_order[] = "avg, max, switch, runtime";
3343         struct perf_sched sched = {
3344                 .tool = {
3345                         .sample          = perf_sched__process_tracepoint_sample,
3346                         .comm            = perf_sched__process_comm,
3347                         .namespaces      = perf_event__process_namespaces,
3348                         .lost            = perf_event__process_lost,
3349                         .fork            = perf_sched__process_fork_event,
3350                         .ordered_events = true,
3351                 },
3352                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3353                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3354                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3355                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3356                 .sort_order           = default_sort_order,
3357                 .replay_repeat        = 10,
3358                 .profile_cpu          = -1,
3359                 .next_shortname1      = 'A',
3360                 .next_shortname2      = '0',
3361                 .skip_merge           = 0,
3362                 .show_callchain       = 1,
3363                 .max_stack            = 5,
3364         };
3365         const struct option sched_options[] = {
3366         OPT_STRING('i', "input", &input_name, "file",
3367                     "input file name"),
3368         OPT_INCR('v', "verbose", &verbose,
3369                     "be more verbose (show symbol address, etc)"),
3370         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3371                     "dump raw trace in ASCII"),
3372         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3373         OPT_END()
3374         };
3375         const struct option latency_options[] = {
3376         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3377                    "sort by key(s): runtime, switch, avg, max"),
3378         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3379                     "CPU to profile on"),
3380         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3381                     "latency stats per pid instead of per comm"),
3382         OPT_PARENT(sched_options)
3383         };
3384         const struct option replay_options[] = {
3385         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3386                      "repeat the workload replay N times (-1: infinite)"),
3387         OPT_PARENT(sched_options)
3388         };
3389         const struct option map_options[] = {
3390         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3391                     "map output in compact mode"),
3392         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3393                    "highlight given pids in map"),
3394         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3395                     "highlight given CPUs in map"),
3396         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3397                     "display given CPUs in map"),
3398         OPT_PARENT(sched_options)
3399         };
3400         const struct option timehist_options[] = {
3401         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3402                    "file", "vmlinux pathname"),
3403         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3404                    "file", "kallsyms pathname"),
3405         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3406                     "Display call chains if present (default on)"),
3407         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3408                    "Maximum number of functions to display backtrace."),
3409         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3410                     "Look for files with symbols relative to this directory"),
3411         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3412                     "Show only syscall summary with statistics"),
3413         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3414                     "Show all syscalls and summary with statistics"),
3415         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3416         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3417         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3418         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3419         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3420         OPT_STRING(0, "time", &sched.time_str, "str",
3421                    "Time span for analysis (start,stop)"),
3422         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3423         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3424                    "analyze events only for given process id(s)"),
3425         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3426                    "analyze events only for given thread id(s)"),
3427         OPT_PARENT(sched_options)
3428         };
3429
3430         const char * const latency_usage[] = {
3431                 "perf sched latency [<options>]",
3432                 NULL
3433         };
3434         const char * const replay_usage[] = {
3435                 "perf sched replay [<options>]",
3436                 NULL
3437         };
3438         const char * const map_usage[] = {
3439                 "perf sched map [<options>]",
3440                 NULL
3441         };
3442         const char * const timehist_usage[] = {
3443                 "perf sched timehist [<options>]",
3444                 NULL
3445         };
3446         const char *const sched_subcommands[] = { "record", "latency", "map",
3447                                                   "replay", "script",
3448                                                   "timehist", NULL };
3449         const char *sched_usage[] = {
3450                 NULL,
3451                 NULL
3452         };
3453         struct trace_sched_handler lat_ops  = {
3454                 .wakeup_event       = latency_wakeup_event,
3455                 .switch_event       = latency_switch_event,
3456                 .runtime_event      = latency_runtime_event,
3457                 .migrate_task_event = latency_migrate_task_event,
3458         };
3459         struct trace_sched_handler map_ops  = {
3460                 .switch_event       = map_switch_event,
3461         };
3462         struct trace_sched_handler replay_ops  = {
3463                 .wakeup_event       = replay_wakeup_event,
3464                 .switch_event       = replay_switch_event,
3465                 .fork_event         = replay_fork_event,
3466         };
3467         unsigned int i;
3468
3469         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3470                 sched.curr_pid[i] = -1;
3471
3472         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3473                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3474         if (!argc)
3475                 usage_with_options(sched_usage, sched_options);
3476
3477         /*
3478          * Aliased to 'perf script' for now:
3479          */
3480         if (!strcmp(argv[0], "script"))
3481                 return cmd_script(argc, argv);
3482
3483         if (!strncmp(argv[0], "rec", 3)) {
3484                 return __cmd_record(argc, argv);
3485         } else if (!strncmp(argv[0], "lat", 3)) {
3486                 sched.tp_handler = &lat_ops;
3487                 if (argc > 1) {
3488                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3489                         if (argc)
3490                                 usage_with_options(latency_usage, latency_options);
3491                 }
3492                 setup_sorting(&sched, latency_options, latency_usage);
3493                 return perf_sched__lat(&sched);
3494         } else if (!strcmp(argv[0], "map")) {
3495                 if (argc) {
3496                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3497                         if (argc)
3498                                 usage_with_options(map_usage, map_options);
3499                 }
3500                 sched.tp_handler = &map_ops;
3501                 setup_sorting(&sched, latency_options, latency_usage);
3502                 return perf_sched__map(&sched);
3503         } else if (!strncmp(argv[0], "rep", 3)) {
3504                 sched.tp_handler = &replay_ops;
3505                 if (argc) {
3506                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3507                         if (argc)
3508                                 usage_with_options(replay_usage, replay_options);
3509                 }
3510                 return perf_sched__replay(&sched);
3511         } else if (!strcmp(argv[0], "timehist")) {
3512                 if (argc) {
3513                         argc = parse_options(argc, argv, timehist_options,
3514                                              timehist_usage, 0);
3515                         if (argc)
3516                                 usage_with_options(timehist_usage, timehist_options);
3517                 }
3518                 if ((sched.show_wakeups || sched.show_next) &&
3519                     sched.summary_only) {
3520                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3521                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3522                         if (sched.show_wakeups)
3523                                 parse_options_usage(NULL, timehist_options, "w", true);
3524                         if (sched.show_next)
3525                                 parse_options_usage(NULL, timehist_options, "n", true);
3526                         return -EINVAL;
3527                 }
3528
3529                 return perf_sched__timehist(&sched);
3530         } else {
3531                 usage_with_options(sched_usage, sched_options);
3532         }
3533
3534         return 0;
3535 }