tools/perf/builtin-sched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf-sys.h"
   4
   5 #include "util/cpumap.h"
   6 #include "util/evlist.h"
   7 #include "util/evsel.h"
   8 #include "util/evsel_fprintf.h"
   9 #include "util/mutex.h"
  10 #include "util/symbol.h"
  11 #include "util/thread.h"
  12 #include "util/header.h"
  13 #include "util/session.h"
  14 #include "util/tool.h"
  15 #include "util/cloexec.h"
  16 #include "util/thread_map.h"
  17 #include "util/color.h"
  18 #include "util/stat.h"
  19 #include "util/string2.h"
  20 #include "util/callchain.h"
  21 #include "util/time-utils.h"
  22
  23 #include <subcmd/pager.h>
  24 #include <subcmd/parse-options.h>
  25 #include "util/trace-event.h"
  26
  27 #include "util/debug.h"
  28 #include "util/event.h"
  29 #include "util/util.h"
  30
  31 #include <linux/kernel.h>
  32 #include <linux/log2.h>
  33 #include <linux/zalloc.h>
  34 #include <sys/prctl.h>
  35 #include <sys/resource.h>
  36 #include <inttypes.h>
  37
  38 #include <errno.h>
  39 #include <semaphore.h>
  40 #include <pthread.h>
  41 #include <math.h>
  42 #include <api/fs/fs.h>
  43 #include <perf/cpumap.h>
  44 #include <linux/time64.h>
  45 #include <linux/err.h>
  46
  47 #include <linux/ctype.h>
  48
  49 #define PR_SET_NAME             15               /* Set process name */
  50 #define MAX_CPUS                4096
  51 #define COMM_LEN                20
  52 #define SYM_LEN                 129
  53 #define MAX_PID                 1024000
  54
  55 static const char *cpu_list;
  56 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  57
  58 struct sched_atom;
  59
  60 struct task_desc {
  61         unsigned long           nr;
  62         unsigned long           pid;
  63         char                    comm[COMM_LEN];
  64
  65         unsigned long           nr_events;
  66         unsigned long           curr_event;
  67         struct sched_atom       **atoms;
  68
  69         pthread_t               thread;
  70         sem_t                   sleep_sem;
  71
  72         sem_t                   ready_for_work;
  73         sem_t                   work_done_sem;
  74
  75         u64                     cpu_usage;
  76 };
  77
  78 enum sched_event_type {
  79         SCHED_EVENT_RUN,
  80         SCHED_EVENT_SLEEP,
  81         SCHED_EVENT_WAKEUP,
  82         SCHED_EVENT_MIGRATION,
  83 };
  84
  85 struct sched_atom {
  86         enum sched_event_type   type;
  87         int                     specific_wait;
  88         u64                     timestamp;
  89         u64                     duration;
  90         unsigned long           nr;
  91         sem_t                   *wait_sem;
  92         struct task_desc        *wakee;
  93 };
  94
  95 enum thread_state {
  96         THREAD_SLEEPING = 0,
  97         THREAD_WAIT_CPU,
  98         THREAD_SCHED_IN,
  99         THREAD_IGNORE
 100 };
 101
 102 struct work_atom {
 103         struct list_head        list;
 104         enum thread_state       state;
 105         u64                     sched_out_time;
 106         u64                     wake_up_time;
 107         u64                     sched_in_time;
 108         u64                     runtime;
 109 };
 110
 111 struct work_atoms {
 112         struct list_head        work_list;
 113         struct thread           *thread;
 114         struct rb_node          node;
 115         u64                     max_lat;
 116         u64                     max_lat_start;
 117         u64                     max_lat_end;
 118         u64                     total_lat;
 119         u64                     nb_atoms;
 120         u64                     total_runtime;
 121         int                     num_merged;
 122 };
 123
 124 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 125
 126 struct perf_sched;
 127
 128 struct trace_sched_handler {
 129         int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 130                             struct perf_sample *sample, struct machine *machine);
 131
 132         int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 133                              struct perf_sample *sample, struct machine *machine);
 134
 135         int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 136                             struct perf_sample *sample, struct machine *machine);
 137
 138         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 139         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 140                           struct machine *machine);
 141
 142         int (*migrate_task_event)(struct perf_sched *sched,
 143                                   struct evsel *evsel,
 144                                   struct perf_sample *sample,
 145                                   struct machine *machine);
 146 };
 147
 148 #define COLOR_PIDS PERF_COLOR_BLUE
 149 #define COLOR_CPUS PERF_COLOR_BG_RED
 150
 151 struct perf_sched_map {
 152         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 153         struct perf_cpu         *comp_cpus;
 154         bool                     comp;
 155         struct perf_thread_map *color_pids;
 156         const char              *color_pids_str;
 157         struct perf_cpu_map     *color_cpus;
 158         const char              *color_cpus_str;
 159         struct perf_cpu_map     *cpus;
 160         const char              *cpus_str;
 161 };
 162
 163 struct perf_sched {
 164         struct perf_tool tool;
 165         const char       *sort_order;
 166         unsigned long    nr_tasks;
 167         struct task_desc **pid_to_task;
 168         struct task_desc **tasks;
 169         const struct trace_sched_handler *tp_handler;
 170         struct mutex     start_work_mutex;
 171         struct mutex     work_done_wait_mutex;
 172         int              profile_cpu;
 173 /*
 174  * Track the current task - that way we can know whether there's any
 175  * weird events, such as a task being switched away that is not current.
 176  */
 177         struct perf_cpu  max_cpu;
 178         u32              *curr_pid;
 179         struct thread    **curr_thread;
 180         char             next_shortname1;
 181         char             next_shortname2;
 182         unsigned int     replay_repeat;
 183         unsigned long    nr_run_events;
 184         unsigned long    nr_sleep_events;
 185         unsigned long    nr_wakeup_events;
 186         unsigned long    nr_sleep_corrections;
 187         unsigned long    nr_run_events_optimized;
 188         unsigned long    targetless_wakeups;
 189         unsigned long    multitarget_wakeups;
 190         unsigned long    nr_runs;
 191         unsigned long    nr_timestamps;
 192         unsigned long    nr_unordered_timestamps;
 193         unsigned long    nr_context_switch_bugs;
 194         unsigned long    nr_events;
 195         unsigned long    nr_lost_chunks;
 196         unsigned long    nr_lost_events;
 197         u64              run_measurement_overhead;
 198         u64              sleep_measurement_overhead;
 199         u64              start_time;
 200         u64              cpu_usage;
 201         u64              runavg_cpu_usage;
 202         u64              parent_cpu_usage;
 203         u64              runavg_parent_cpu_usage;
 204         u64              sum_runtime;
 205         u64              sum_fluct;
 206         u64              run_avg;
 207         u64              all_runtime;
 208         u64              all_count;
 209         u64              *cpu_last_switched;
 210         struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 211         struct list_head sort_list, cmp_pid;
 212         bool force;
 213         bool skip_merge;
 214         struct perf_sched_map map;
 215
 216         /* options for timehist command */
 217         bool            summary;
 218         bool            summary_only;
 219         bool            idle_hist;
 220         bool            show_callchain;
 221         unsigned int    max_stack;
 222         bool            show_cpu_visual;
 223         bool            show_wakeups;
 224         bool            show_next;
 225         bool            show_migrations;
 226         bool            show_state;
 227         u64             skipped_samples;
 228         const char      *time_str;
 229         struct perf_time_interval ptime;
 230         struct perf_time_interval hist_time;
 231         volatile bool   thread_funcs_exit;
 232 };
 233
 234 /* per thread run time data */
 235 struct thread_runtime {
 236         u64 last_time;      /* time of previous sched in/out event */
 237         u64 dt_run;         /* run time */
 238         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 239         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 240         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 241         u64 dt_delay;       /* time between wakeup and sched-in */
 242         u64 ready_to_run;   /* time of wakeup */
 243
 244         struct stats run_stats;
 245         u64 total_run_time;
 246         u64 total_sleep_time;
 247         u64 total_iowait_time;
 248         u64 total_preempt_time;
 249         u64 total_delay_time;
 250
 251         char last_state;
 252
 253         char shortname[3];
 254         bool comm_changed;
 255
 256         u64 migrations;
 257 };
 258
 259 /* per event run time data */
 260 struct evsel_runtime {
 261         u64 *last_time; /* time this event was last seen per cpu */
 262         u32 ncpu;       /* highest cpu slot allocated */
 263 };
 264
 265 /* per cpu idle time data */
 266 struct idle_thread_runtime {
 267         struct thread_runtime   tr;
 268         struct thread           *last_thread;
 269         struct rb_root_cached   sorted_root;
 270         struct callchain_root   callchain;
 271         struct callchain_cursor cursor;
 272 };
 273
 274 /* track idle times per cpu */
 275 static struct thread **idle_threads;
 276 static int idle_max_cpu;
 277 static char idle_comm[] = "<idle>";
 278
 279 static u64 get_nsecs(void)
 280 {
 281         struct timespec ts;
 282
 283         clock_gettime(CLOCK_MONOTONIC, &ts);
 284
 285         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 286 }
 287
 288 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 289 {
 290         u64 T0 = get_nsecs(), T1;
 291
 292         do {
 293                 T1 = get_nsecs();
 294         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 295 }
 296
 297 static void sleep_nsecs(u64 nsecs)
 298 {
 299         struct timespec ts;
 300
 301         ts.tv_nsec = nsecs % 999999999;
 302         ts.tv_sec = nsecs / 999999999;
 303
 304         nanosleep(&ts, NULL);
 305 }
 306
 307 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 308 {
 309         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 310         int i;
 311
 312         for (i = 0; i < 10; i++) {
 313                 T0 = get_nsecs();
 314                 burn_nsecs(sched, 0);
 315                 T1 = get_nsecs();
 316                 delta = T1-T0;
 317                 min_delta = min(min_delta, delta);
 318         }
 319         sched->run_measurement_overhead = min_delta;
 320
 321         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 322 }
 323
 324 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 325 {
 326         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 327         int i;
 328
 329         for (i = 0; i < 10; i++) {
 330                 T0 = get_nsecs();
 331                 sleep_nsecs(10000);
 332                 T1 = get_nsecs();
 333                 delta = T1-T0;
 334                 min_delta = min(min_delta, delta);
 335         }
 336         min_delta -= 10000;
 337         sched->sleep_measurement_overhead = min_delta;
 338
 339         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 340 }
 341
 342 static struct sched_atom *
 343 get_new_event(struct task_desc *task, u64 timestamp)
 344 {
 345         struct sched_atom *event = zalloc(sizeof(*event));
 346         unsigned long idx = task->nr_events;
 347         size_t size;
 348
 349         event->timestamp = timestamp;
 350         event->nr = idx;
 351
 352         task->nr_events++;
 353         size = sizeof(struct sched_atom *) * task->nr_events;
 354         task->atoms = realloc(task->atoms, size);
 355         BUG_ON(!task->atoms);
 356
 357         task->atoms[idx] = event;
 358
 359         return event;
 360 }
 361
 362 static struct sched_atom *last_event(struct task_desc *task)
 363 {
 364         if (!task->nr_events)
 365                 return NULL;
 366
 367         return task->atoms[task->nr_events - 1];
 368 }
 369
 370 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 371                                 u64 timestamp, u64 duration)
 372 {
 373         struct sched_atom *event, *curr_event = last_event(task);
 374
 375         /*
 376          * optimize an existing RUN event by merging this one
 377          * to it:
 378          */
 379         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 380                 sched->nr_run_events_optimized++;
 381                 curr_event->duration += duration;
 382                 return;
 383         }
 384
 385         event = get_new_event(task, timestamp);
 386
 387         event->type = SCHED_EVENT_RUN;
 388         event->duration = duration;
 389
 390         sched->nr_run_events++;
 391 }
 392
 393 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 394                                    u64 timestamp, struct task_desc *wakee)
 395 {
 396         struct sched_atom *event, *wakee_event;
 397
 398         event = get_new_event(task, timestamp);
 399         event->type = SCHED_EVENT_WAKEUP;
 400         event->wakee = wakee;
 401
 402         wakee_event = last_event(wakee);
 403         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 404                 sched->targetless_wakeups++;
 405                 return;
 406         }
 407         if (wakee_event->wait_sem) {
 408                 sched->multitarget_wakeups++;
 409                 return;
 410         }
 411
 412         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 413         sem_init(wakee_event->wait_sem, 0, 0);
 414         wakee_event->specific_wait = 1;
 415         event->wait_sem = wakee_event->wait_sem;
 416
 417         sched->nr_wakeup_events++;
 418 }
 419
 420 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 421                                   u64 timestamp, const char task_state __maybe_unused)
 422 {
 423         struct sched_atom *event = get_new_event(task, timestamp);
 424
 425         event->type = SCHED_EVENT_SLEEP;
 426
 427         sched->nr_sleep_events++;
 428 }
 429
 430 static struct task_desc *register_pid(struct perf_sched *sched,
 431                                       unsigned long pid, const char *comm)
 432 {
 433         struct task_desc *task;
 434         static int pid_max;
 435
 436         if (sched->pid_to_task == NULL) {
 437                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 438                         pid_max = MAX_PID;
 439                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 440         }
 441         if (pid >= (unsigned long)pid_max) {
 442                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 443                         sizeof(struct task_desc *))) == NULL);
 444                 while (pid >= (unsigned long)pid_max)
 445                         sched->pid_to_task[pid_max++] = NULL;
 446         }
 447
 448         task = sched->pid_to_task[pid];
 449
 450         if (task)
 451                 return task;
 452
 453         task = zalloc(sizeof(*task));
 454         task->pid = pid;
 455         task->nr = sched->nr_tasks;
 456         strcpy(task->comm, comm);
 457         /*
 458          * every task starts in sleeping state - this gets ignored
 459          * if there's no wakeup pointing to this sleep state:
 460          */
 461         add_sched_event_sleep(sched, task, 0, 0);
 462
 463         sched->pid_to_task[pid] = task;
 464         sched->nr_tasks++;
 465         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 466         BUG_ON(!sched->tasks);
 467         sched->tasks[task->nr] = task;
 468
 469         if (verbose > 0)
 470                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 471
 472         return task;
 473 }
 474
 475
 476 static void print_task_traces(struct perf_sched *sched)
 477 {
 478         struct task_desc *task;
 479         unsigned long i;
 480
 481         for (i = 0; i < sched->nr_tasks; i++) {
 482                 task = sched->tasks[i];
 483                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 484                         task->nr, task->comm, task->pid, task->nr_events);
 485         }
 486 }
 487
 488 static void add_cross_task_wakeups(struct perf_sched *sched)
 489 {
 490         struct task_desc *task1, *task2;
 491         unsigned long i, j;
 492
 493         for (i = 0; i < sched->nr_tasks; i++) {
 494                 task1 = sched->tasks[i];
 495                 j = i + 1;
 496                 if (j == sched->nr_tasks)
 497                         j = 0;
 498                 task2 = sched->tasks[j];
 499                 add_sched_event_wakeup(sched, task1, 0, task2);
 500         }
 501 }
 502
 503 static void perf_sched__process_event(struct perf_sched *sched,
 504                                       struct sched_atom *atom)
 505 {
 506         int ret = 0;
 507
 508         switch (atom->type) {
 509                 case SCHED_EVENT_RUN:
 510                         burn_nsecs(sched, atom->duration);
 511                         break;
 512                 case SCHED_EVENT_SLEEP:
 513                         if (atom->wait_sem)
 514                                 ret = sem_wait(atom->wait_sem);
 515                         BUG_ON(ret);
 516                         break;
 517                 case SCHED_EVENT_WAKEUP:
 518                         if (atom->wait_sem)
 519                                 ret = sem_post(atom->wait_sem);
 520                         BUG_ON(ret);
 521                         break;
 522                 case SCHED_EVENT_MIGRATION:
 523                         break;
 524                 default:
 525                         BUG_ON(1);
 526         }
 527 }
 528
 529 static u64 get_cpu_usage_nsec_parent(void)
 530 {
 531         struct rusage ru;
 532         u64 sum;
 533         int err;
 534
 535         err = getrusage(RUSAGE_SELF, &ru);
 536         BUG_ON(err);
 537
 538         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 539         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 540
 541         return sum;
 542 }
 543
 544 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 545 {
 546         struct perf_event_attr attr;
 547         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 548         int fd;
 549         struct rlimit limit;
 550         bool need_privilege = false;
 551
 552         memset(&attr, 0, sizeof(attr));
 553
 554         attr.type = PERF_TYPE_SOFTWARE;
 555         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 556
 557 force_again:
 558         fd = sys_perf_event_open(&attr, 0, -1, -1,
 559                                  perf_event_open_cloexec_flag());
 560
 561         if (fd < 0) {
 562                 if (errno == EMFILE) {
 563                         if (sched->force) {
 564                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 565                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 566                                 if (limit.rlim_cur > limit.rlim_max) {
 567                                         limit.rlim_max = limit.rlim_cur;
 568                                         need_privilege = true;
 569                                 }
 570                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 571                                         if (need_privilege && errno == EPERM)
 572                                                 strcpy(info, "Need privilege\n");
 573                                 } else
 574                                         goto force_again;
 575                         } else
 576                                 strcpy(info, "Have a try with -f option\n");
 577                 }
 578                 pr_err("Error: sys_perf_event_open() syscall returned "
 579                        "with %d (%s)\n%s", fd,
 580                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 581                 exit(EXIT_FAILURE);
 582         }
 583         return fd;
 584 }
 585
 586 static u64 get_cpu_usage_nsec_self(int fd)
 587 {
 588         u64 runtime;
 589         int ret;
 590
 591         ret = read(fd, &runtime, sizeof(runtime));
 592         BUG_ON(ret != sizeof(runtime));
 593
 594         return runtime;
 595 }
 596
 597 struct sched_thread_parms {
 598         struct task_desc  *task;
 599         struct perf_sched *sched;
 600         int fd;
 601 };
 602
 603 static void *thread_func(void *ctx)
 604 {
 605         struct sched_thread_parms *parms = ctx;
 606         struct task_desc *this_task = parms->task;
 607         struct perf_sched *sched = parms->sched;
 608         u64 cpu_usage_0, cpu_usage_1;
 609         unsigned long i, ret;
 610         char comm2[22];
 611         int fd = parms->fd;
 612
 613         zfree(&parms);
 614
 615         sprintf(comm2, ":%s", this_task->comm);
 616         prctl(PR_SET_NAME, comm2);
 617         if (fd < 0)
 618                 return NULL;
 619
 620         while (!sched->thread_funcs_exit) {
 621                 ret = sem_post(&this_task->ready_for_work);
 622                 BUG_ON(ret);
 623                 mutex_lock(&sched->start_work_mutex);
 624                 mutex_unlock(&sched->start_work_mutex);
 625
 626                 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 627
 628                 for (i = 0; i < this_task->nr_events; i++) {
 629                         this_task->curr_event = i;
 630                         perf_sched__process_event(sched, this_task->atoms[i]);
 631                 }
 632
 633                 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 634                 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 635                 ret = sem_post(&this_task->work_done_sem);
 636                 BUG_ON(ret);
 637
 638                 mutex_lock(&sched->work_done_wait_mutex);
 639                 mutex_unlock(&sched->work_done_wait_mutex);
 640         }
 641         return NULL;
 642 }
 643
 644 static void create_tasks(struct perf_sched *sched)
 645         EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
 646         EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
 647 {
 648         struct task_desc *task;
 649         pthread_attr_t attr;
 650         unsigned long i;
 651         int err;
 652
 653         err = pthread_attr_init(&attr);
 654         BUG_ON(err);
 655         err = pthread_attr_setstacksize(&attr,
 656                         (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
 657         BUG_ON(err);
 658         mutex_lock(&sched->start_work_mutex);
 659         mutex_lock(&sched->work_done_wait_mutex);
 660         for (i = 0; i < sched->nr_tasks; i++) {
 661                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 662                 BUG_ON(parms == NULL);
 663                 parms->task = task = sched->tasks[i];
 664                 parms->sched = sched;
 665                 parms->fd = self_open_counters(sched, i);
 666                 sem_init(&task->sleep_sem, 0, 0);
 667                 sem_init(&task->ready_for_work, 0, 0);
 668                 sem_init(&task->work_done_sem, 0, 0);
 669                 task->curr_event = 0;
 670                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 671                 BUG_ON(err);
 672         }
 673 }
 674
 675 static void destroy_tasks(struct perf_sched *sched)
 676         UNLOCK_FUNCTION(sched->start_work_mutex)
 677         UNLOCK_FUNCTION(sched->work_done_wait_mutex)
 678 {
 679         struct task_desc *task;
 680         unsigned long i;
 681         int err;
 682
 683         mutex_unlock(&sched->start_work_mutex);
 684         mutex_unlock(&sched->work_done_wait_mutex);
 685         /* Get rid of threads so they won't be upset by mutex destrunction */
 686         for (i = 0; i < sched->nr_tasks; i++) {
 687                 task = sched->tasks[i];
 688                 err = pthread_join(task->thread, NULL);
 689                 BUG_ON(err);
 690                 sem_destroy(&task->sleep_sem);
 691                 sem_destroy(&task->ready_for_work);
 692                 sem_destroy(&task->work_done_sem);
 693         }
 694 }
 695
 696 static void wait_for_tasks(struct perf_sched *sched)
 697         EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 698         EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 699 {
 700         u64 cpu_usage_0, cpu_usage_1;
 701         struct task_desc *task;
 702         unsigned long i, ret;
 703
 704         sched->start_time = get_nsecs();
 705         sched->cpu_usage = 0;
 706         mutex_unlock(&sched->work_done_wait_mutex);
 707
 708         for (i = 0; i < sched->nr_tasks; i++) {
 709                 task = sched->tasks[i];
 710                 ret = sem_wait(&task->ready_for_work);
 711                 BUG_ON(ret);
 712                 sem_init(&task->ready_for_work, 0, 0);
 713         }
 714         mutex_lock(&sched->work_done_wait_mutex);
 715
 716         cpu_usage_0 = get_cpu_usage_nsec_parent();
 717
 718         mutex_unlock(&sched->start_work_mutex);
 719
 720         for (i = 0; i < sched->nr_tasks; i++) {
 721                 task = sched->tasks[i];
 722                 ret = sem_wait(&task->work_done_sem);
 723                 BUG_ON(ret);
 724                 sem_init(&task->work_done_sem, 0, 0);
 725                 sched->cpu_usage += task->cpu_usage;
 726                 task->cpu_usage = 0;
 727         }
 728
 729         cpu_usage_1 = get_cpu_usage_nsec_parent();
 730         if (!sched->runavg_cpu_usage)
 731                 sched->runavg_cpu_usage = sched->cpu_usage;
 732         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 733
 734         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 735         if (!sched->runavg_parent_cpu_usage)
 736                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 737         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 738                                          sched->parent_cpu_usage)/sched->replay_repeat;
 739
 740         mutex_lock(&sched->start_work_mutex);
 741
 742         for (i = 0; i < sched->nr_tasks; i++) {
 743                 task = sched->tasks[i];
 744                 sem_init(&task->sleep_sem, 0, 0);
 745                 task->curr_event = 0;
 746         }
 747 }
 748
 749 static void run_one_test(struct perf_sched *sched)
 750         EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 751         EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 752 {
 753         u64 T0, T1, delta, avg_delta, fluct;
 754
 755         T0 = get_nsecs();
 756         wait_for_tasks(sched);
 757         T1 = get_nsecs();
 758
 759         delta = T1 - T0;
 760         sched->sum_runtime += delta;
 761         sched->nr_runs++;
 762
 763         avg_delta = sched->sum_runtime / sched->nr_runs;
 764         if (delta < avg_delta)
 765                 fluct = avg_delta - delta;
 766         else
 767                 fluct = delta - avg_delta;
 768         sched->sum_fluct += fluct;
 769         if (!sched->run_avg)
 770                 sched->run_avg = delta;
 771         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 772
 773         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 774
 775         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 776
 777         printf("cpu: %0.2f / %0.2f",
 778                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 779
 780 #if 0
 781         /*
 782          * rusage statistics done by the parent, these are less
 783          * accurate than the sched->sum_exec_runtime based statistics:
 784          */
 785         printf(" [%0.2f / %0.2f]",
 786                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 787                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 788 #endif
 789
 790         printf("\n");
 791
 792         if (sched->nr_sleep_corrections)
 793                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 794         sched->nr_sleep_corrections = 0;
 795 }
 796
 797 static void test_calibrations(struct perf_sched *sched)
 798 {
 799         u64 T0, T1;
 800
 801         T0 = get_nsecs();
 802         burn_nsecs(sched, NSEC_PER_MSEC);
 803         T1 = get_nsecs();
 804
 805         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 806
 807         T0 = get_nsecs();
 808         sleep_nsecs(NSEC_PER_MSEC);
 809         T1 = get_nsecs();
 810
 811         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 812 }
 813
 814 static int
 815 replay_wakeup_event(struct perf_sched *sched,
 816                     struct evsel *evsel, struct perf_sample *sample,
 817                     struct machine *machine __maybe_unused)
 818 {
 819         const char *comm = evsel__strval(evsel, sample, "comm");
 820         const u32 pid    = evsel__intval(evsel, sample, "pid");
 821         struct task_desc *waker, *wakee;
 822
 823         if (verbose > 0) {
 824                 printf("sched_wakeup event %p\n", evsel);
 825
 826                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 827         }
 828
 829         waker = register_pid(sched, sample->tid, "<unknown>");
 830         wakee = register_pid(sched, pid, comm);
 831
 832         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 833         return 0;
 834 }
 835
 836 static int replay_switch_event(struct perf_sched *sched,
 837                                struct evsel *evsel,
 838                                struct perf_sample *sample,
 839                                struct machine *machine __maybe_unused)
 840 {
 841         const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 842                    *next_comm  = evsel__strval(evsel, sample, "next_comm");
 843         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 844                   next_pid = evsel__intval(evsel, sample, "next_pid");
 845         const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
 846         struct task_desc *prev, __maybe_unused *next;
 847         u64 timestamp0, timestamp = sample->time;
 848         int cpu = sample->cpu;
 849         s64 delta;
 850
 851         if (verbose > 0)
 852                 printf("sched_switch event %p\n", evsel);
 853
 854         if (cpu >= MAX_CPUS || cpu < 0)
 855                 return 0;
 856
 857         timestamp0 = sched->cpu_last_switched[cpu];
 858         if (timestamp0)
 859                 delta = timestamp - timestamp0;
 860         else
 861                 delta = 0;
 862
 863         if (delta < 0) {
 864                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 865                 return -1;
 866         }
 867
 868         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 869                  prev_comm, prev_pid, next_comm, next_pid, delta);
 870
 871         prev = register_pid(sched, prev_pid, prev_comm);
 872         next = register_pid(sched, next_pid, next_comm);
 873
 874         sched->cpu_last_switched[cpu] = timestamp;
 875
 876         add_sched_event_run(sched, prev, timestamp, delta);
 877         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 878
 879         return 0;
 880 }
 881
 882 static int replay_fork_event(struct perf_sched *sched,
 883                              union perf_event *event,
 884                              struct machine *machine)
 885 {
 886         struct thread *child, *parent;
 887
 888         child = machine__findnew_thread(machine, event->fork.pid,
 889                                         event->fork.tid);
 890         parent = machine__findnew_thread(machine, event->fork.ppid,
 891                                          event->fork.ptid);
 892
 893         if (child == NULL || parent == NULL) {
 894                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 895                                  child, parent);
 896                 goto out_put;
 897         }
 898
 899         if (verbose > 0) {
 900                 printf("fork event\n");
 901                 printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
 902                 printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
 903         }
 904
 905         register_pid(sched, thread__tid(parent), thread__comm_str(parent));
 906         register_pid(sched, thread__tid(child), thread__comm_str(child));
 907 out_put:
 908         thread__put(child);
 909         thread__put(parent);
 910         return 0;
 911 }
 912
 913 struct sort_dimension {
 914         const char              *name;
 915         sort_fn_t               cmp;
 916         struct list_head        list;
 917 };
 918
 919 /*
 920  * handle runtime stats saved per thread
 921  */
 922 static struct thread_runtime *thread__init_runtime(struct thread *thread)
 923 {
 924         struct thread_runtime *r;
 925
 926         r = zalloc(sizeof(struct thread_runtime));
 927         if (!r)
 928                 return NULL;
 929
 930         init_stats(&r->run_stats);
 931         thread__set_priv(thread, r);
 932
 933         return r;
 934 }
 935
 936 static struct thread_runtime *thread__get_runtime(struct thread *thread)
 937 {
 938         struct thread_runtime *tr;
 939
 940         tr = thread__priv(thread);
 941         if (tr == NULL) {
 942                 tr = thread__init_runtime(thread);
 943                 if (tr == NULL)
 944                         pr_debug("Failed to malloc memory for runtime data.\n");
 945         }
 946
 947         return tr;
 948 }
 949
 950 static int
 951 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 952 {
 953         struct sort_dimension *sort;
 954         int ret = 0;
 955
 956         BUG_ON(list_empty(list));
 957
 958         list_for_each_entry(sort, list, list) {
 959                 ret = sort->cmp(l, r);
 960                 if (ret)
 961                         return ret;
 962         }
 963
 964         return ret;
 965 }
 966
 967 static struct work_atoms *
 968 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 969                          struct list_head *sort_list)
 970 {
 971         struct rb_node *node = root->rb_root.rb_node;
 972         struct work_atoms key = { .thread = thread };
 973
 974         while (node) {
 975                 struct work_atoms *atoms;
 976                 int cmp;
 977
 978                 atoms = container_of(node, struct work_atoms, node);
 979
 980                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 981                 if (cmp > 0)
 982                         node = node->rb_left;
 983                 else if (cmp < 0)
 984                         node = node->rb_right;
 985                 else {
 986                         BUG_ON(thread != atoms->thread);
 987                         return atoms;
 988                 }
 989         }
 990         return NULL;
 991 }
 992
 993 static void
 994 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 995                          struct list_head *sort_list)
 996 {
 997         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 998         bool leftmost = true;
 999
1000         while (*new) {
1001                 struct work_atoms *this;
1002                 int cmp;
1003
1004                 this = container_of(*new, struct work_atoms, node);
1005                 parent = *new;
1006
1007                 cmp = thread_lat_cmp(sort_list, data, this);
1008
1009                 if (cmp > 0)
1010                         new = &((*new)->rb_left);
1011                 else {
1012                         new = &((*new)->rb_right);
1013                         leftmost = false;
1014                 }
1015         }
1016
1017         rb_link_node(&data->node, parent, new);
1018         rb_insert_color_cached(&data->node, root, leftmost);
1019 }
1020
1021 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1022 {
1023         struct work_atoms *atoms = zalloc(sizeof(*atoms));
1024         if (!atoms) {
1025                 pr_err("No memory at %s\n", __func__);
1026                 return -1;
1027         }
1028
1029         atoms->thread = thread__get(thread);
1030         INIT_LIST_HEAD(&atoms->work_list);
1031         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1032         return 0;
1033 }
1034
1035 static int
1036 add_sched_out_event(struct work_atoms *atoms,
1037                     char run_state,
1038                     u64 timestamp)
1039 {
1040         struct work_atom *atom = zalloc(sizeof(*atom));
1041         if (!atom) {
1042                 pr_err("Non memory at %s", __func__);
1043                 return -1;
1044         }
1045
1046         atom->sched_out_time = timestamp;
1047
1048         if (run_state == 'R') {
1049                 atom->state = THREAD_WAIT_CPU;
1050                 atom->wake_up_time = atom->sched_out_time;
1051         }
1052
1053         list_add_tail(&atom->list, &atoms->work_list);
1054         return 0;
1055 }
1056
1057 static void
1058 add_runtime_event(struct work_atoms *atoms, u64 delta,
1059                   u64 timestamp __maybe_unused)
1060 {
1061         struct work_atom *atom;
1062
1063         BUG_ON(list_empty(&atoms->work_list));
1064
1065         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1066
1067         atom->runtime += delta;
1068         atoms->total_runtime += delta;
1069 }
1070
1071 static void
1072 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1073 {
1074         struct work_atom *atom;
1075         u64 delta;
1076
1077         if (list_empty(&atoms->work_list))
1078                 return;
1079
1080         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1081
1082         if (atom->state != THREAD_WAIT_CPU)
1083                 return;
1084
1085         if (timestamp < atom->wake_up_time) {
1086                 atom->state = THREAD_IGNORE;
1087                 return;
1088         }
1089
1090         atom->state = THREAD_SCHED_IN;
1091         atom->sched_in_time = timestamp;
1092
1093         delta = atom->sched_in_time - atom->wake_up_time;
1094         atoms->total_lat += delta;
1095         if (delta > atoms->max_lat) {
1096                 atoms->max_lat = delta;
1097                 atoms->max_lat_start = atom->wake_up_time;
1098                 atoms->max_lat_end = timestamp;
1099         }
1100         atoms->nb_atoms++;
1101 }
1102
1103 static int latency_switch_event(struct perf_sched *sched,
1104                                 struct evsel *evsel,
1105                                 struct perf_sample *sample,
1106                                 struct machine *machine)
1107 {
1108         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1109                   next_pid = evsel__intval(evsel, sample, "next_pid");
1110         const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1111         struct work_atoms *out_events, *in_events;
1112         struct thread *sched_out, *sched_in;
1113         u64 timestamp0, timestamp = sample->time;
1114         int cpu = sample->cpu, err = -1;
1115         s64 delta;
1116
1117         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1118
1119         timestamp0 = sched->cpu_last_switched[cpu];
1120         sched->cpu_last_switched[cpu] = timestamp;
1121         if (timestamp0)
1122                 delta = timestamp - timestamp0;
1123         else
1124                 delta = 0;
1125
1126         if (delta < 0) {
1127                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1128                 return -1;
1129         }
1130
1131         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1132         sched_in = machine__findnew_thread(machine, -1, next_pid);
1133         if (sched_out == NULL || sched_in == NULL)
1134                 goto out_put;
1135
1136         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1137         if (!out_events) {
1138                 if (thread_atoms_insert(sched, sched_out))
1139                         goto out_put;
1140                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1141                 if (!out_events) {
1142                         pr_err("out-event: Internal tree error");
1143                         goto out_put;
1144                 }
1145         }
1146         if (add_sched_out_event(out_events, prev_state, timestamp))
1147                 return -1;
1148
1149         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1150         if (!in_events) {
1151                 if (thread_atoms_insert(sched, sched_in))
1152                         goto out_put;
1153                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1154                 if (!in_events) {
1155                         pr_err("in-event: Internal tree error");
1156                         goto out_put;
1157                 }
1158                 /*
1159                  * Take came in we have not heard about yet,
1160                  * add in an initial atom in runnable state:
1161                  */
1162                 if (add_sched_out_event(in_events, 'R', timestamp))
1163                         goto out_put;
1164         }
1165         add_sched_in_event(in_events, timestamp);
1166         err = 0;
1167 out_put:
1168         thread__put(sched_out);
1169         thread__put(sched_in);
1170         return err;
1171 }
1172
1173 static int latency_runtime_event(struct perf_sched *sched,
1174                                  struct evsel *evsel,
1175                                  struct perf_sample *sample,
1176                                  struct machine *machine)
1177 {
1178         const u32 pid      = evsel__intval(evsel, sample, "pid");
1179         const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1180         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1181         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1182         u64 timestamp = sample->time;
1183         int cpu = sample->cpu, err = -1;
1184
1185         if (thread == NULL)
1186                 return -1;
1187
1188         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1189         if (!atoms) {
1190                 if (thread_atoms_insert(sched, thread))
1191                         goto out_put;
1192                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1193                 if (!atoms) {
1194                         pr_err("in-event: Internal tree error");
1195                         goto out_put;
1196                 }
1197                 if (add_sched_out_event(atoms, 'R', timestamp))
1198                         goto out_put;
1199         }
1200
1201         add_runtime_event(atoms, runtime, timestamp);
1202         err = 0;
1203 out_put:
1204         thread__put(thread);
1205         return err;
1206 }
1207
1208 static int latency_wakeup_event(struct perf_sched *sched,
1209                                 struct evsel *evsel,
1210                                 struct perf_sample *sample,
1211                                 struct machine *machine)
1212 {
1213         const u32 pid     = evsel__intval(evsel, sample, "pid");
1214         struct work_atoms *atoms;
1215         struct work_atom *atom;
1216         struct thread *wakee;
1217         u64 timestamp = sample->time;
1218         int err = -1;
1219
1220         wakee = machine__findnew_thread(machine, -1, pid);
1221         if (wakee == NULL)
1222                 return -1;
1223         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1224         if (!atoms) {
1225                 if (thread_atoms_insert(sched, wakee))
1226                         goto out_put;
1227                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1228                 if (!atoms) {
1229                         pr_err("wakeup-event: Internal tree error");
1230                         goto out_put;
1231                 }
1232                 if (add_sched_out_event(atoms, 'S', timestamp))
1233                         goto out_put;
1234         }
1235
1236         BUG_ON(list_empty(&atoms->work_list));
1237
1238         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1239
1240         /*
1241          * As we do not guarantee the wakeup event happens when
1242          * task is out of run queue, also may happen when task is
1243          * on run queue and wakeup only change ->state to TASK_RUNNING,
1244          * then we should not set the ->wake_up_time when wake up a
1245          * task which is on run queue.
1246          *
1247          * You WILL be missing events if you've recorded only
1248          * one CPU, or are only looking at only one, so don't
1249          * skip in this case.
1250          */
1251         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1252                 goto out_ok;
1253
1254         sched->nr_timestamps++;
1255         if (atom->sched_out_time > timestamp) {
1256                 sched->nr_unordered_timestamps++;
1257                 goto out_ok;
1258         }
1259
1260         atom->state = THREAD_WAIT_CPU;
1261         atom->wake_up_time = timestamp;
1262 out_ok:
1263         err = 0;
1264 out_put:
1265         thread__put(wakee);
1266         return err;
1267 }
1268
1269 static int latency_migrate_task_event(struct perf_sched *sched,
1270                                       struct evsel *evsel,
1271                                       struct perf_sample *sample,
1272                                       struct machine *machine)
1273 {
1274         const u32 pid = evsel__intval(evsel, sample, "pid");
1275         u64 timestamp = sample->time;
1276         struct work_atoms *atoms;
1277         struct work_atom *atom;
1278         struct thread *migrant;
1279         int err = -1;
1280
1281         /*
1282          * Only need to worry about migration when profiling one CPU.
1283          */
1284         if (sched->profile_cpu == -1)
1285                 return 0;
1286
1287         migrant = machine__findnew_thread(machine, -1, pid);
1288         if (migrant == NULL)
1289                 return -1;
1290         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1291         if (!atoms) {
1292                 if (thread_atoms_insert(sched, migrant))
1293                         goto out_put;
1294                 register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1295                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1296                 if (!atoms) {
1297                         pr_err("migration-event: Internal tree error");
1298                         goto out_put;
1299                 }
1300                 if (add_sched_out_event(atoms, 'R', timestamp))
1301                         goto out_put;
1302         }
1303
1304         BUG_ON(list_empty(&atoms->work_list));
1305
1306         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1307         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1308
1309         sched->nr_timestamps++;
1310
1311         if (atom->sched_out_time > timestamp)
1312                 sched->nr_unordered_timestamps++;
1313         err = 0;
1314 out_put:
1315         thread__put(migrant);
1316         return err;
1317 }
1318
1319 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1320 {
1321         int i;
1322         int ret;
1323         u64 avg;
1324         char max_lat_start[32], max_lat_end[32];
1325
1326         if (!work_list->nb_atoms)
1327                 return;
1328         /*
1329          * Ignore idle threads:
1330          */
1331         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1332                 return;
1333
1334         sched->all_runtime += work_list->total_runtime;
1335         sched->all_count   += work_list->nb_atoms;
1336
1337         if (work_list->num_merged > 1) {
1338                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1339                              work_list->num_merged);
1340         } else {
1341                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1342                              thread__tid(work_list->thread));
1343         }
1344
1345         for (i = 0; i < 24 - ret; i++)
1346                 printf(" ");
1347
1348         avg = work_list->total_lat / work_list->nb_atoms;
1349         timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350         timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1351
1352         printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353               (double)work_list->total_runtime / NSEC_PER_MSEC,
1354                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355                  (double)work_list->max_lat / NSEC_PER_MSEC,
1356                  max_lat_start, max_lat_end);
1357 }
1358
1359 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1360 {
1361         pid_t l_tid, r_tid;
1362
1363         if (RC_CHK_EQUAL(l->thread, r->thread))
1364                 return 0;
1365         l_tid = thread__tid(l->thread);
1366         r_tid = thread__tid(r->thread);
1367         if (l_tid < r_tid)
1368                 return -1;
1369         if (l_tid > r_tid)
1370                 return 1;
1371         return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1372 }
1373
1374 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1375 {
1376         u64 avgl, avgr;
1377
1378         if (!l->nb_atoms)
1379                 return -1;
1380
1381         if (!r->nb_atoms)
1382                 return 1;
1383
1384         avgl = l->total_lat / l->nb_atoms;
1385         avgr = r->total_lat / r->nb_atoms;
1386
1387         if (avgl < avgr)
1388                 return -1;
1389         if (avgl > avgr)
1390                 return 1;
1391
1392         return 0;
1393 }
1394
1395 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1396 {
1397         if (l->max_lat < r->max_lat)
1398                 return -1;
1399         if (l->max_lat > r->max_lat)
1400                 return 1;
1401
1402         return 0;
1403 }
1404
1405 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1406 {
1407         if (l->nb_atoms < r->nb_atoms)
1408                 return -1;
1409         if (l->nb_atoms > r->nb_atoms)
1410                 return 1;
1411
1412         return 0;
1413 }
1414
1415 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1416 {
1417         if (l->total_runtime < r->total_runtime)
1418                 return -1;
1419         if (l->total_runtime > r->total_runtime)
1420                 return 1;
1421
1422         return 0;
1423 }
1424
1425 static int sort_dimension__add(const char *tok, struct list_head *list)
1426 {
1427         size_t i;
1428         static struct sort_dimension avg_sort_dimension = {
1429                 .name = "avg",
1430                 .cmp  = avg_cmp,
1431         };
1432         static struct sort_dimension max_sort_dimension = {
1433                 .name = "max",
1434                 .cmp  = max_cmp,
1435         };
1436         static struct sort_dimension pid_sort_dimension = {
1437                 .name = "pid",
1438                 .cmp  = pid_cmp,
1439         };
1440         static struct sort_dimension runtime_sort_dimension = {
1441                 .name = "runtime",
1442                 .cmp  = runtime_cmp,
1443         };
1444         static struct sort_dimension switch_sort_dimension = {
1445                 .name = "switch",
1446                 .cmp  = switch_cmp,
1447         };
1448         struct sort_dimension *available_sorts[] = {
1449                 &pid_sort_dimension,
1450                 &avg_sort_dimension,
1451                 &max_sort_dimension,
1452                 &switch_sort_dimension,
1453                 &runtime_sort_dimension,
1454         };
1455
1456         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1457                 if (!strcmp(available_sorts[i]->name, tok)) {
1458                         list_add_tail(&available_sorts[i]->list, list);
1459
1460                         return 0;
1461                 }
1462         }
1463
1464         return -1;
1465 }
1466
1467 static void perf_sched__sort_lat(struct perf_sched *sched)
1468 {
1469         struct rb_node *node;
1470         struct rb_root_cached *root = &sched->atom_root;
1471 again:
1472         for (;;) {
1473                 struct work_atoms *data;
1474                 node = rb_first_cached(root);
1475                 if (!node)
1476                         break;
1477
1478                 rb_erase_cached(node, root);
1479                 data = rb_entry(node, struct work_atoms, node);
1480                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1481         }
1482         if (root == &sched->atom_root) {
1483                 root = &sched->merged_atom_root;
1484                 goto again;
1485         }
1486 }
1487
1488 static int process_sched_wakeup_event(struct perf_tool *tool,
1489                                       struct evsel *evsel,
1490                                       struct perf_sample *sample,
1491                                       struct machine *machine)
1492 {
1493         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1494
1495         if (sched->tp_handler->wakeup_event)
1496                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1497
1498         return 0;
1499 }
1500
1501 static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1502                                       struct evsel *evsel __maybe_unused,
1503                                       struct perf_sample *sample __maybe_unused,
1504                                       struct machine *machine __maybe_unused)
1505 {
1506         return 0;
1507 }
1508
1509 union map_priv {
1510         void    *ptr;
1511         bool     color;
1512 };
1513
1514 static bool thread__has_color(struct thread *thread)
1515 {
1516         union map_priv priv = {
1517                 .ptr = thread__priv(thread),
1518         };
1519
1520         return priv.color;
1521 }
1522
1523 static struct thread*
1524 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1525 {
1526         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1527         union map_priv priv = {
1528                 .color = false,
1529         };
1530
1531         if (!sched->map.color_pids || !thread || thread__priv(thread))
1532                 return thread;
1533
1534         if (thread_map__has(sched->map.color_pids, tid))
1535                 priv.color = true;
1536
1537         thread__set_priv(thread, priv.ptr);
1538         return thread;
1539 }
1540
1541 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1542                             struct perf_sample *sample, struct machine *machine)
1543 {
1544         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1545         struct thread *sched_in;
1546         struct thread_runtime *tr;
1547         int new_shortname;
1548         u64 timestamp0, timestamp = sample->time;
1549         s64 delta;
1550         int i;
1551         struct perf_cpu this_cpu = {
1552                 .cpu = sample->cpu,
1553         };
1554         int cpus_nr;
1555         bool new_cpu = false;
1556         const char *color = PERF_COLOR_NORMAL;
1557         char stimestamp[32];
1558
1559         BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1560
1561         if (this_cpu.cpu > sched->max_cpu.cpu)
1562                 sched->max_cpu = this_cpu;
1563
1564         if (sched->map.comp) {
1565                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1566                 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1567                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1568                         new_cpu = true;
1569                 }
1570         } else
1571                 cpus_nr = sched->max_cpu.cpu;
1572
1573         timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1574         sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1575         if (timestamp0)
1576                 delta = timestamp - timestamp0;
1577         else
1578                 delta = 0;
1579
1580         if (delta < 0) {
1581                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1582                 return -1;
1583         }
1584
1585         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1586         if (sched_in == NULL)
1587                 return -1;
1588
1589         tr = thread__get_runtime(sched_in);
1590         if (tr == NULL) {
1591                 thread__put(sched_in);
1592                 return -1;
1593         }
1594
1595         sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1596
1597         printf("  ");
1598
1599         new_shortname = 0;
1600         if (!tr->shortname[0]) {
1601                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1602                         /*
1603                          * Don't allocate a letter-number for swapper:0
1604                          * as a shortname. Instead, we use '.' for it.
1605                          */
1606                         tr->shortname[0] = '.';
1607                         tr->shortname[1] = ' ';
1608                 } else {
1609                         tr->shortname[0] = sched->next_shortname1;
1610                         tr->shortname[1] = sched->next_shortname2;
1611
1612                         if (sched->next_shortname1 < 'Z') {
1613                                 sched->next_shortname1++;
1614                         } else {
1615                                 sched->next_shortname1 = 'A';
1616                                 if (sched->next_shortname2 < '9')
1617                                         sched->next_shortname2++;
1618                                 else
1619                                         sched->next_shortname2 = '0';
1620                         }
1621                 }
1622                 new_shortname = 1;
1623         }
1624
1625         for (i = 0; i < cpus_nr; i++) {
1626                 struct perf_cpu cpu = {
1627                         .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1628                 };
1629                 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1630                 struct thread_runtime *curr_tr;
1631                 const char *pid_color = color;
1632                 const char *cpu_color = color;
1633
1634                 if (curr_thread && thread__has_color(curr_thread))
1635                         pid_color = COLOR_PIDS;
1636
1637                 if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1638                         continue;
1639
1640                 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1641                         cpu_color = COLOR_CPUS;
1642
1643                 if (cpu.cpu != this_cpu.cpu)
1644                         color_fprintf(stdout, color, " ");
1645                 else
1646                         color_fprintf(stdout, cpu_color, "*");
1647
1648                 if (sched->curr_thread[cpu.cpu]) {
1649                         curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1650                         if (curr_tr == NULL) {
1651                                 thread__put(sched_in);
1652                                 return -1;
1653                         }
1654                         color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1655                 } else
1656                         color_fprintf(stdout, color, "   ");
1657         }
1658
1659         if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1660                 goto out;
1661
1662         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1663         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1664         if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1665                 const char *pid_color = color;
1666
1667                 if (thread__has_color(sched_in))
1668                         pid_color = COLOR_PIDS;
1669
1670                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1671                         tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1672                 tr->comm_changed = false;
1673         }
1674
1675         if (sched->map.comp && new_cpu)
1676                 color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1677
1678 out:
1679         color_fprintf(stdout, color, "\n");
1680
1681         thread__put(sched_in);
1682
1683         return 0;
1684 }
1685
1686 static int process_sched_switch_event(struct perf_tool *tool,
1687                                       struct evsel *evsel,
1688                                       struct perf_sample *sample,
1689                                       struct machine *machine)
1690 {
1691         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1692         int this_cpu = sample->cpu, err = 0;
1693         u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1694             next_pid = evsel__intval(evsel, sample, "next_pid");
1695
1696         if (sched->curr_pid[this_cpu] != (u32)-1) {
1697                 /*
1698                  * Are we trying to switch away a PID that is
1699                  * not current?
1700                  */
1701                 if (sched->curr_pid[this_cpu] != prev_pid)
1702                         sched->nr_context_switch_bugs++;
1703         }
1704
1705         if (sched->tp_handler->switch_event)
1706                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1707
1708         sched->curr_pid[this_cpu] = next_pid;
1709         return err;
1710 }
1711
1712 static int process_sched_runtime_event(struct perf_tool *tool,
1713                                        struct evsel *evsel,
1714                                        struct perf_sample *sample,
1715                                        struct machine *machine)
1716 {
1717         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1718
1719         if (sched->tp_handler->runtime_event)
1720                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1721
1722         return 0;
1723 }
1724
1725 static int perf_sched__process_fork_event(struct perf_tool *tool,
1726                                           union perf_event *event,
1727                                           struct perf_sample *sample,
1728                                           struct machine *machine)
1729 {
1730         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731
1732         /* run the fork event through the perf machinery */
1733         perf_event__process_fork(tool, event, sample, machine);
1734
1735         /* and then run additional processing needed for this command */
1736         if (sched->tp_handler->fork_event)
1737                 return sched->tp_handler->fork_event(sched, event, machine);
1738
1739         return 0;
1740 }
1741
1742 static int process_sched_migrate_task_event(struct perf_tool *tool,
1743                                             struct evsel *evsel,
1744                                             struct perf_sample *sample,
1745                                             struct machine *machine)
1746 {
1747         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1748
1749         if (sched->tp_handler->migrate_task_event)
1750                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1751
1752         return 0;
1753 }
1754
1755 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1756                                   struct evsel *evsel,
1757                                   struct perf_sample *sample,
1758                                   struct machine *machine);
1759
1760 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1761                                                  union perf_event *event __maybe_unused,
1762                                                  struct perf_sample *sample,
1763                                                  struct evsel *evsel,
1764                                                  struct machine *machine)
1765 {
1766         int err = 0;
1767
1768         if (evsel->handler != NULL) {
1769                 tracepoint_handler f = evsel->handler;
1770                 err = f(tool, evsel, sample, machine);
1771         }
1772
1773         return err;
1774 }
1775
1776 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1777                                     union perf_event *event,
1778                                     struct perf_sample *sample,
1779                                     struct machine *machine)
1780 {
1781         struct thread *thread;
1782         struct thread_runtime *tr;
1783         int err;
1784
1785         err = perf_event__process_comm(tool, event, sample, machine);
1786         if (err)
1787                 return err;
1788
1789         thread = machine__find_thread(machine, sample->pid, sample->tid);
1790         if (!thread) {
1791                 pr_err("Internal error: can't find thread\n");
1792                 return -1;
1793         }
1794
1795         tr = thread__get_runtime(thread);
1796         if (tr == NULL) {
1797                 thread__put(thread);
1798                 return -1;
1799         }
1800
1801         tr->comm_changed = true;
1802         thread__put(thread);
1803
1804         return 0;
1805 }
1806
1807 static int perf_sched__read_events(struct perf_sched *sched)
1808 {
1809         struct evsel_str_handler handlers[] = {
1810                 { "sched:sched_switch",       process_sched_switch_event, },
1811                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1812                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1813                 { "sched:sched_waking",       process_sched_wakeup_event, },
1814                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1815                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1816         };
1817         struct perf_session *session;
1818         struct perf_data data = {
1819                 .path  = input_name,
1820                 .mode  = PERF_DATA_MODE_READ,
1821                 .force = sched->force,
1822         };
1823         int rc = -1;
1824
1825         session = perf_session__new(&data, &sched->tool);
1826         if (IS_ERR(session)) {
1827                 pr_debug("Error creating perf session");
1828                 return PTR_ERR(session);
1829         }
1830
1831         symbol__init(&session->header.env);
1832
1833         /* prefer sched_waking if it is captured */
1834         if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1835                 handlers[2].handler = process_sched_wakeup_ignore;
1836
1837         if (perf_session__set_tracepoints_handlers(session, handlers))
1838                 goto out_delete;
1839
1840         if (perf_session__has_traces(session, "record -R")) {
1841                 int err = perf_session__process_events(session);
1842                 if (err) {
1843                         pr_err("Failed to process events, error %d", err);
1844                         goto out_delete;
1845                 }
1846
1847                 sched->nr_events      = session->evlist->stats.nr_events[0];
1848                 sched->nr_lost_events = session->evlist->stats.total_lost;
1849                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1850         }
1851
1852         rc = 0;
1853 out_delete:
1854         perf_session__delete(session);
1855         return rc;
1856 }
1857
1858 /*
1859  * scheduling times are printed as msec.usec
1860  */
1861 static inline void print_sched_time(unsigned long long nsecs, int width)
1862 {
1863         unsigned long msecs;
1864         unsigned long usecs;
1865
1866         msecs  = nsecs / NSEC_PER_MSEC;
1867         nsecs -= msecs * NSEC_PER_MSEC;
1868         usecs  = nsecs / NSEC_PER_USEC;
1869         printf("%*lu.%03lu ", width, msecs, usecs);
1870 }
1871
1872 /*
1873  * returns runtime data for event, allocating memory for it the
1874  * first time it is used.
1875  */
1876 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1877 {
1878         struct evsel_runtime *r = evsel->priv;
1879
1880         if (r == NULL) {
1881                 r = zalloc(sizeof(struct evsel_runtime));
1882                 evsel->priv = r;
1883         }
1884
1885         return r;
1886 }
1887
1888 /*
1889  * save last time event was seen per cpu
1890  */
1891 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1892 {
1893         struct evsel_runtime *r = evsel__get_runtime(evsel);
1894
1895         if (r == NULL)
1896                 return;
1897
1898         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1899                 int i, n = __roundup_pow_of_two(cpu+1);
1900                 void *p = r->last_time;
1901
1902                 p = realloc(r->last_time, n * sizeof(u64));
1903                 if (!p)
1904                         return;
1905
1906                 r->last_time = p;
1907                 for (i = r->ncpu; i < n; ++i)
1908                         r->last_time[i] = (u64) 0;
1909
1910                 r->ncpu = n;
1911         }
1912
1913         r->last_time[cpu] = timestamp;
1914 }
1915
1916 /* returns last time this event was seen on the given cpu */
1917 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1918 {
1919         struct evsel_runtime *r = evsel__get_runtime(evsel);
1920
1921         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1922                 return 0;
1923
1924         return r->last_time[cpu];
1925 }
1926
1927 static int comm_width = 30;
1928
1929 static char *timehist_get_commstr(struct thread *thread)
1930 {
1931         static char str[32];
1932         const char *comm = thread__comm_str(thread);
1933         pid_t tid = thread__tid(thread);
1934         pid_t pid = thread__pid(thread);
1935         int n;
1936
1937         if (pid == 0)
1938                 n = scnprintf(str, sizeof(str), "%s", comm);
1939
1940         else if (tid != pid)
1941                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1942
1943         else
1944                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1945
1946         if (n > comm_width)
1947                 comm_width = n;
1948
1949         return str;
1950 }
1951
1952 static void timehist_header(struct perf_sched *sched)
1953 {
1954         u32 ncpus = sched->max_cpu.cpu + 1;
1955         u32 i, j;
1956
1957         printf("%15s %6s ", "time", "cpu");
1958
1959         if (sched->show_cpu_visual) {
1960                 printf(" ");
1961                 for (i = 0, j = 0; i < ncpus; ++i) {
1962                         printf("%x", j++);
1963                         if (j > 15)
1964                                 j = 0;
1965                 }
1966                 printf(" ");
1967         }
1968
1969         printf(" %-*s  %9s  %9s  %9s", comm_width,
1970                 "task name", "wait time", "sch delay", "run time");
1971
1972         if (sched->show_state)
1973                 printf("  %s", "state");
1974
1975         printf("\n");
1976
1977         /*
1978          * units row
1979          */
1980         printf("%15s %-6s ", "", "");
1981
1982         if (sched->show_cpu_visual)
1983                 printf(" %*s ", ncpus, "");
1984
1985         printf(" %-*s  %9s  %9s  %9s", comm_width,
1986                "[tid/pid]", "(msec)", "(msec)", "(msec)");
1987
1988         if (sched->show_state)
1989                 printf("  %5s", "");
1990
1991         printf("\n");
1992
1993         /*
1994          * separator
1995          */
1996         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1997
1998         if (sched->show_cpu_visual)
1999                 printf(" %.*s ", ncpus, graph_dotted_line);
2000
2001         printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2002                 graph_dotted_line, graph_dotted_line, graph_dotted_line,
2003                 graph_dotted_line);
2004
2005         if (sched->show_state)
2006                 printf("  %.5s", graph_dotted_line);
2007
2008         printf("\n");
2009 }
2010
2011 static void timehist_print_sample(struct perf_sched *sched,
2012                                   struct evsel *evsel,
2013                                   struct perf_sample *sample,
2014                                   struct addr_location *al,
2015                                   struct thread *thread,
2016                                   u64 t, const char state)
2017 {
2018         struct thread_runtime *tr = thread__priv(thread);
2019         const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2020         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2021         u32 max_cpus = sched->max_cpu.cpu + 1;
2022         char tstr[64];
2023         char nstr[30];
2024         u64 wait_time;
2025
2026         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2027                 return;
2028
2029         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2030         printf("%15s [%04d] ", tstr, sample->cpu);
2031
2032         if (sched->show_cpu_visual) {
2033                 u32 i;
2034                 char c;
2035
2036                 printf(" ");
2037                 for (i = 0; i < max_cpus; ++i) {
2038                         /* flag idle times with 'i'; others are sched events */
2039                         if (i == sample->cpu)
2040                                 c = (thread__tid(thread) == 0) ? 'i' : 's';
2041                         else
2042                                 c = ' ';
2043                         printf("%c", c);
2044                 }
2045                 printf(" ");
2046         }
2047
2048         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2049
2050         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2051         print_sched_time(wait_time, 6);
2052
2053         print_sched_time(tr->dt_delay, 6);
2054         print_sched_time(tr->dt_run, 6);
2055
2056         if (sched->show_state)
2057                 printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2058
2059         if (sched->show_next) {
2060                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2061                 printf(" %-*s", comm_width, nstr);
2062         }
2063
2064         if (sched->show_wakeups && !sched->show_next)
2065                 printf("  %-*s", comm_width, "");
2066
2067         if (thread__tid(thread) == 0)
2068                 goto out;
2069
2070         if (sched->show_callchain)
2071                 printf("  ");
2072
2073         sample__fprintf_sym(sample, al, 0,
2074                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2075                             EVSEL__PRINT_CALLCHAIN_ARROW |
2076                             EVSEL__PRINT_SKIP_IGNORED,
2077                             get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2078
2079 out:
2080         printf("\n");
2081 }
2082
2083 /*
2084  * Explanation of delta-time stats:
2085  *
2086  *            t = time of current schedule out event
2087  *        tprev = time of previous sched out event
2088  *                also time of schedule-in event for current task
2089  *    last_time = time of last sched change event for current task
2090  *                (i.e, time process was last scheduled out)
2091  * ready_to_run = time of wakeup for current task
2092  *
2093  * -----|------------|------------|------------|------
2094  *    last         ready        tprev          t
2095  *    time         to run
2096  *
2097  *      |-------- dt_wait --------|
2098  *                   |- dt_delay -|-- dt_run --|
2099  *
2100  *   dt_run = run time of current task
2101  *  dt_wait = time between last schedule out event for task and tprev
2102  *            represents time spent off the cpu
2103  * dt_delay = time between wakeup and schedule-in of task
2104  */
2105
2106 static void timehist_update_runtime_stats(struct thread_runtime *r,
2107                                          u64 t, u64 tprev)
2108 {
2109         r->dt_delay   = 0;
2110         r->dt_sleep   = 0;
2111         r->dt_iowait  = 0;
2112         r->dt_preempt = 0;
2113         r->dt_run     = 0;
2114
2115         if (tprev) {
2116                 r->dt_run = t - tprev;
2117                 if (r->ready_to_run) {
2118                         if (r->ready_to_run > tprev)
2119                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2120                         else
2121                                 r->dt_delay = tprev - r->ready_to_run;
2122                 }
2123
2124                 if (r->last_time > tprev)
2125                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2126                 else if (r->last_time) {
2127                         u64 dt_wait = tprev - r->last_time;
2128
2129                         if (r->last_state == 'R')
2130                                 r->dt_preempt = dt_wait;
2131                         else if (r->last_state == 'D')
2132                                 r->dt_iowait = dt_wait;
2133                         else
2134                                 r->dt_sleep = dt_wait;
2135                 }
2136         }
2137
2138         update_stats(&r->run_stats, r->dt_run);
2139
2140         r->total_run_time     += r->dt_run;
2141         r->total_delay_time   += r->dt_delay;
2142         r->total_sleep_time   += r->dt_sleep;
2143         r->total_iowait_time  += r->dt_iowait;
2144         r->total_preempt_time += r->dt_preempt;
2145 }
2146
2147 static bool is_idle_sample(struct perf_sample *sample,
2148                            struct evsel *evsel)
2149 {
2150         /* pid 0 == swapper == idle task */
2151         if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2152                 return evsel__intval(evsel, sample, "prev_pid") == 0;
2153
2154         return sample->pid == 0;
2155 }
2156
2157 static void save_task_callchain(struct perf_sched *sched,
2158                                 struct perf_sample *sample,
2159                                 struct evsel *evsel,
2160                                 struct machine *machine)
2161 {
2162         struct callchain_cursor *cursor;
2163         struct thread *thread;
2164
2165         /* want main thread for process - has maps */
2166         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2167         if (thread == NULL) {
2168                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2169                 return;
2170         }
2171
2172         if (!sched->show_callchain || sample->callchain == NULL)
2173                 return;
2174
2175         cursor = get_tls_callchain_cursor();
2176
2177         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2178                                       NULL, NULL, sched->max_stack + 2) != 0) {
2179                 if (verbose > 0)
2180                         pr_err("Failed to resolve callchain. Skipping\n");
2181
2182                 return;
2183         }
2184
2185         callchain_cursor_commit(cursor);
2186
2187         while (true) {
2188                 struct callchain_cursor_node *node;
2189                 struct symbol *sym;
2190
2191                 node = callchain_cursor_current(cursor);
2192                 if (node == NULL)
2193                         break;
2194
2195                 sym = node->ms.sym;
2196                 if (sym) {
2197                         if (!strcmp(sym->name, "schedule") ||
2198                             !strcmp(sym->name, "__schedule") ||
2199                             !strcmp(sym->name, "preempt_schedule"))
2200                                 sym->ignore = 1;
2201                 }
2202
2203                 callchain_cursor_advance(cursor);
2204         }
2205 }
2206
2207 static int init_idle_thread(struct thread *thread)
2208 {
2209         struct idle_thread_runtime *itr;
2210
2211         thread__set_comm(thread, idle_comm, 0);
2212
2213         itr = zalloc(sizeof(*itr));
2214         if (itr == NULL)
2215                 return -ENOMEM;
2216
2217         init_stats(&itr->tr.run_stats);
2218         callchain_init(&itr->callchain);
2219         callchain_cursor_reset(&itr->cursor);
2220         thread__set_priv(thread, itr);
2221
2222         return 0;
2223 }
2224
2225 /*
2226  * Track idle stats per cpu by maintaining a local thread
2227  * struct for the idle task on each cpu.
2228  */
2229 static int init_idle_threads(int ncpu)
2230 {
2231         int i, ret;
2232
2233         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2234         if (!idle_threads)
2235                 return -ENOMEM;
2236
2237         idle_max_cpu = ncpu;
2238
2239         /* allocate the actual thread struct if needed */
2240         for (i = 0; i < ncpu; ++i) {
2241                 idle_threads[i] = thread__new(0, 0);
2242                 if (idle_threads[i] == NULL)
2243                         return -ENOMEM;
2244
2245                 ret = init_idle_thread(idle_threads[i]);
2246                 if (ret < 0)
2247                         return ret;
2248         }
2249
2250         return 0;
2251 }
2252
2253 static void free_idle_threads(void)
2254 {
2255         int i;
2256
2257         if (idle_threads == NULL)
2258                 return;
2259
2260         for (i = 0; i < idle_max_cpu; ++i) {
2261                 if ((idle_threads[i]))
2262                         thread__delete(idle_threads[i]);
2263         }
2264
2265         free(idle_threads);
2266 }
2267
2268 static struct thread *get_idle_thread(int cpu)
2269 {
2270         /*
2271          * expand/allocate array of pointers to local thread
2272          * structs if needed
2273          */
2274         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2275                 int i, j = __roundup_pow_of_two(cpu+1);
2276                 void *p;
2277
2278                 p = realloc(idle_threads, j * sizeof(struct thread *));
2279                 if (!p)
2280                         return NULL;
2281
2282                 idle_threads = (struct thread **) p;
2283                 for (i = idle_max_cpu; i < j; ++i)
2284                         idle_threads[i] = NULL;
2285
2286                 idle_max_cpu = j;
2287         }
2288
2289         /* allocate a new thread struct if needed */
2290         if (idle_threads[cpu] == NULL) {
2291                 idle_threads[cpu] = thread__new(0, 0);
2292
2293                 if (idle_threads[cpu]) {
2294                         if (init_idle_thread(idle_threads[cpu]) < 0)
2295                                 return NULL;
2296                 }
2297         }
2298
2299         return idle_threads[cpu];
2300 }
2301
2302 static void save_idle_callchain(struct perf_sched *sched,
2303                                 struct idle_thread_runtime *itr,
2304                                 struct perf_sample *sample)
2305 {
2306         struct callchain_cursor *cursor;
2307
2308         if (!sched->show_callchain || sample->callchain == NULL)
2309                 return;
2310
2311         cursor = get_tls_callchain_cursor();
2312         if (cursor == NULL)
2313                 return;
2314
2315         callchain_cursor__copy(&itr->cursor, cursor);
2316 }
2317
2318 static struct thread *timehist_get_thread(struct perf_sched *sched,
2319                                           struct perf_sample *sample,
2320                                           struct machine *machine,
2321                                           struct evsel *evsel)
2322 {
2323         struct thread *thread;
2324
2325         if (is_idle_sample(sample, evsel)) {
2326                 thread = get_idle_thread(sample->cpu);
2327                 if (thread == NULL)
2328                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329
2330         } else {
2331                 /* there were samples with tid 0 but non-zero pid */
2332                 thread = machine__findnew_thread(machine, sample->pid,
2333                                                  sample->tid ?: sample->pid);
2334                 if (thread == NULL) {
2335                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2336                                  sample->tid);
2337                 }
2338
2339                 save_task_callchain(sched, sample, evsel, machine);
2340                 if (sched->idle_hist) {
2341                         struct thread *idle;
2342                         struct idle_thread_runtime *itr;
2343
2344                         idle = get_idle_thread(sample->cpu);
2345                         if (idle == NULL) {
2346                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2347                                 return NULL;
2348                         }
2349
2350                         itr = thread__priv(idle);
2351                         if (itr == NULL)
2352                                 return NULL;
2353
2354                         itr->last_thread = thread;
2355
2356                         /* copy task callchain when entering to idle */
2357                         if (evsel__intval(evsel, sample, "next_pid") == 0)
2358                                 save_idle_callchain(sched, itr, sample);
2359                 }
2360         }
2361
2362         return thread;
2363 }
2364
2365 static bool timehist_skip_sample(struct perf_sched *sched,
2366                                  struct thread *thread,
2367                                  struct evsel *evsel,
2368                                  struct perf_sample *sample)
2369 {
2370         bool rc = false;
2371
2372         if (thread__is_filtered(thread)) {
2373                 rc = true;
2374                 sched->skipped_samples++;
2375         }
2376
2377         if (sched->idle_hist) {
2378                 if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2379                         rc = true;
2380                 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2381                          evsel__intval(evsel, sample, "next_pid") != 0)
2382                         rc = true;
2383         }
2384
2385         return rc;
2386 }
2387
2388 static void timehist_print_wakeup_event(struct perf_sched *sched,
2389                                         struct evsel *evsel,
2390                                         struct perf_sample *sample,
2391                                         struct machine *machine,
2392                                         struct thread *awakened)
2393 {
2394         struct thread *thread;
2395         char tstr[64];
2396
2397         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2398         if (thread == NULL)
2399                 return;
2400
2401         /* show wakeup unless both awakee and awaker are filtered */
2402         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2403             timehist_skip_sample(sched, awakened, evsel, sample)) {
2404                 return;
2405         }
2406
2407         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2408         printf("%15s [%04d] ", tstr, sample->cpu);
2409         if (sched->show_cpu_visual)
2410                 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2411
2412         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2413
2414         /* dt spacer */
2415         printf("  %9s  %9s  %9s ", "", "", "");
2416
2417         printf("awakened: %s", timehist_get_commstr(awakened));
2418
2419         printf("\n");
2420 }
2421
2422 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2423                                         union perf_event *event __maybe_unused,
2424                                         struct evsel *evsel __maybe_unused,
2425                                         struct perf_sample *sample __maybe_unused,
2426                                         struct machine *machine __maybe_unused)
2427 {
2428         return 0;
2429 }
2430
2431 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2432                                        union perf_event *event __maybe_unused,
2433                                        struct evsel *evsel,
2434                                        struct perf_sample *sample,
2435                                        struct machine *machine)
2436 {
2437         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2438         struct thread *thread;
2439         struct thread_runtime *tr = NULL;
2440         /* want pid of awakened task not pid in sample */
2441         const u32 pid = evsel__intval(evsel, sample, "pid");
2442
2443         thread = machine__findnew_thread(machine, 0, pid);
2444         if (thread == NULL)
2445                 return -1;
2446
2447         tr = thread__get_runtime(thread);
2448         if (tr == NULL)
2449                 return -1;
2450
2451         if (tr->ready_to_run == 0)
2452                 tr->ready_to_run = sample->time;
2453
2454         /* show wakeups if requested */
2455         if (sched->show_wakeups &&
2456             !perf_time__skip_sample(&sched->ptime, sample->time))
2457                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2458
2459         return 0;
2460 }
2461
2462 static void timehist_print_migration_event(struct perf_sched *sched,
2463                                         struct evsel *evsel,
2464                                         struct perf_sample *sample,
2465                                         struct machine *machine,
2466                                         struct thread *migrated)
2467 {
2468         struct thread *thread;
2469         char tstr[64];
2470         u32 max_cpus;
2471         u32 ocpu, dcpu;
2472
2473         if (sched->summary_only)
2474                 return;
2475
2476         max_cpus = sched->max_cpu.cpu + 1;
2477         ocpu = evsel__intval(evsel, sample, "orig_cpu");
2478         dcpu = evsel__intval(evsel, sample, "dest_cpu");
2479
2480         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2481         if (thread == NULL)
2482                 return;
2483
2484         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2485             timehist_skip_sample(sched, migrated, evsel, sample)) {
2486                 return;
2487         }
2488
2489         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2490         printf("%15s [%04d] ", tstr, sample->cpu);
2491
2492         if (sched->show_cpu_visual) {
2493                 u32 i;
2494                 char c;
2495
2496                 printf("  ");
2497                 for (i = 0; i < max_cpus; ++i) {
2498                         c = (i == sample->cpu) ? 'm' : ' ';
2499                         printf("%c", c);
2500                 }
2501                 printf("  ");
2502         }
2503
2504         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2505
2506         /* dt spacer */
2507         printf("  %9s  %9s  %9s ", "", "", "");
2508
2509         printf("migrated: %s", timehist_get_commstr(migrated));
2510         printf(" cpu %d => %d", ocpu, dcpu);
2511
2512         printf("\n");
2513 }
2514
2515 static int timehist_migrate_task_event(struct perf_tool *tool,
2516                                        union perf_event *event __maybe_unused,
2517                                        struct evsel *evsel,
2518                                        struct perf_sample *sample,
2519                                        struct machine *machine)
2520 {
2521         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2522         struct thread *thread;
2523         struct thread_runtime *tr = NULL;
2524         /* want pid of migrated task not pid in sample */
2525         const u32 pid = evsel__intval(evsel, sample, "pid");
2526
2527         thread = machine__findnew_thread(machine, 0, pid);
2528         if (thread == NULL)
2529                 return -1;
2530
2531         tr = thread__get_runtime(thread);
2532         if (tr == NULL)
2533                 return -1;
2534
2535         tr->migrations++;
2536
2537         /* show migrations if requested */
2538         timehist_print_migration_event(sched, evsel, sample, machine, thread);
2539
2540         return 0;
2541 }
2542
2543 static int timehist_sched_change_event(struct perf_tool *tool,
2544                                        union perf_event *event,
2545                                        struct evsel *evsel,
2546                                        struct perf_sample *sample,
2547                                        struct machine *machine)
2548 {
2549         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2550         struct perf_time_interval *ptime = &sched->ptime;
2551         struct addr_location al;
2552         struct thread *thread;
2553         struct thread_runtime *tr = NULL;
2554         u64 tprev, t = sample->time;
2555         int rc = 0;
2556         const char state = evsel__taskstate(evsel, sample, "prev_state");
2557
2558         addr_location__init(&al);
2559         if (machine__resolve(machine, &al, sample) < 0) {
2560                 pr_err("problem processing %d event. skipping it\n",
2561                        event->header.type);
2562                 rc = -1;
2563                 goto out;
2564         }
2565
2566         thread = timehist_get_thread(sched, sample, machine, evsel);
2567         if (thread == NULL) {
2568                 rc = -1;
2569                 goto out;
2570         }
2571
2572         if (timehist_skip_sample(sched, thread, evsel, sample))
2573                 goto out;
2574
2575         tr = thread__get_runtime(thread);
2576         if (tr == NULL) {
2577                 rc = -1;
2578                 goto out;
2579         }
2580
2581         tprev = evsel__get_time(evsel, sample->cpu);
2582
2583         /*
2584          * If start time given:
2585          * - sample time is under window user cares about - skip sample
2586          * - tprev is under window user cares about  - reset to start of window
2587          */
2588         if (ptime->start && ptime->start > t)
2589                 goto out;
2590
2591         if (tprev && ptime->start > tprev)
2592                 tprev = ptime->start;
2593
2594         /*
2595          * If end time given:
2596          * - previous sched event is out of window - we are done
2597          * - sample time is beyond window user cares about - reset it
2598          *   to close out stats for time window interest
2599          */
2600         if (ptime->end) {
2601                 if (tprev > ptime->end)
2602                         goto out;
2603
2604                 if (t > ptime->end)
2605                         t = ptime->end;
2606         }
2607
2608         if (!sched->idle_hist || thread__tid(thread) == 0) {
2609                 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2610                         timehist_update_runtime_stats(tr, t, tprev);
2611
2612                 if (sched->idle_hist) {
2613                         struct idle_thread_runtime *itr = (void *)tr;
2614                         struct thread_runtime *last_tr;
2615
2616                         BUG_ON(thread__tid(thread) != 0);
2617
2618                         if (itr->last_thread == NULL)
2619                                 goto out;
2620
2621                         /* add current idle time as last thread's runtime */
2622                         last_tr = thread__get_runtime(itr->last_thread);
2623                         if (last_tr == NULL)
2624                                 goto out;
2625
2626                         timehist_update_runtime_stats(last_tr, t, tprev);
2627                         /*
2628                          * remove delta time of last thread as it's not updated
2629                          * and otherwise it will show an invalid value next
2630                          * time.  we only care total run time and run stat.
2631                          */
2632                         last_tr->dt_run = 0;
2633                         last_tr->dt_delay = 0;
2634                         last_tr->dt_sleep = 0;
2635                         last_tr->dt_iowait = 0;
2636                         last_tr->dt_preempt = 0;
2637
2638                         if (itr->cursor.nr)
2639                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2640
2641                         itr->last_thread = NULL;
2642                 }
2643         }
2644
2645         if (!sched->summary_only)
2646                 timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2647
2648 out:
2649         if (sched->hist_time.start == 0 && t >= ptime->start)
2650                 sched->hist_time.start = t;
2651         if (ptime->end == 0 || t <= ptime->end)
2652                 sched->hist_time.end = t;
2653
2654         if (tr) {
2655                 /* time of this sched_switch event becomes last time task seen */
2656                 tr->last_time = sample->time;
2657
2658                 /* last state is used to determine where to account wait time */
2659                 tr->last_state = state;
2660
2661                 /* sched out event for task so reset ready to run time */
2662                 tr->ready_to_run = 0;
2663         }
2664
2665         evsel__save_time(evsel, sample->time, sample->cpu);
2666
2667         addr_location__exit(&al);
2668         return rc;
2669 }
2670
2671 static int timehist_sched_switch_event(struct perf_tool *tool,
2672                              union perf_event *event,
2673                              struct evsel *evsel,
2674                              struct perf_sample *sample,
2675                              struct machine *machine __maybe_unused)
2676 {
2677         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2678 }
2679
2680 static int process_lost(struct perf_tool *tool __maybe_unused,
2681                         union perf_event *event,
2682                         struct perf_sample *sample,
2683                         struct machine *machine __maybe_unused)
2684 {
2685         char tstr[64];
2686
2687         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2688         printf("%15s ", tstr);
2689         printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2690
2691         return 0;
2692 }
2693
2694
2695 static void print_thread_runtime(struct thread *t,
2696                                  struct thread_runtime *r)
2697 {
2698         double mean = avg_stats(&r->run_stats);
2699         float stddev;
2700
2701         printf("%*s   %5d  %9" PRIu64 " ",
2702                comm_width, timehist_get_commstr(t), thread__ppid(t),
2703                (u64) r->run_stats.n);
2704
2705         print_sched_time(r->total_run_time, 8);
2706         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2707         print_sched_time(r->run_stats.min, 6);
2708         printf(" ");
2709         print_sched_time((u64) mean, 6);
2710         printf(" ");
2711         print_sched_time(r->run_stats.max, 6);
2712         printf("  ");
2713         printf("%5.2f", stddev);
2714         printf("   %5" PRIu64, r->migrations);
2715         printf("\n");
2716 }
2717
2718 static void print_thread_waittime(struct thread *t,
2719                                   struct thread_runtime *r)
2720 {
2721         printf("%*s   %5d  %9" PRIu64 " ",
2722                comm_width, timehist_get_commstr(t), thread__ppid(t),
2723                (u64) r->run_stats.n);
2724
2725         print_sched_time(r->total_run_time, 8);
2726         print_sched_time(r->total_sleep_time, 6);
2727         printf(" ");
2728         print_sched_time(r->total_iowait_time, 6);
2729         printf(" ");
2730         print_sched_time(r->total_preempt_time, 6);
2731         printf(" ");
2732         print_sched_time(r->total_delay_time, 6);
2733         printf("\n");
2734 }
2735
2736 struct total_run_stats {
2737         struct perf_sched *sched;
2738         u64  sched_count;
2739         u64  task_count;
2740         u64  total_run_time;
2741 };
2742
2743 static int show_thread_runtime(struct thread *t, void *priv)
2744 {
2745         struct total_run_stats *stats = priv;
2746         struct thread_runtime *r;
2747
2748         if (thread__is_filtered(t))
2749                 return 0;
2750
2751         r = thread__priv(t);
2752         if (r && r->run_stats.n) {
2753                 stats->task_count++;
2754                 stats->sched_count += r->run_stats.n;
2755                 stats->total_run_time += r->total_run_time;
2756
2757                 if (stats->sched->show_state)
2758                         print_thread_waittime(t, r);
2759                 else
2760                         print_thread_runtime(t, r);
2761         }
2762
2763         return 0;
2764 }
2765
2766 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2767 {
2768         const char *sep = " <- ";
2769         struct callchain_list *chain;
2770         size_t ret = 0;
2771         char bf[1024];
2772         bool first;
2773
2774         if (node == NULL)
2775                 return 0;
2776
2777         ret = callchain__fprintf_folded(fp, node->parent);
2778         first = (ret == 0);
2779
2780         list_for_each_entry(chain, &node->val, list) {
2781                 if (chain->ip >= PERF_CONTEXT_MAX)
2782                         continue;
2783                 if (chain->ms.sym && chain->ms.sym->ignore)
2784                         continue;
2785                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2786                                callchain_list__sym_name(chain, bf, sizeof(bf),
2787                                                         false));
2788                 first = false;
2789         }
2790
2791         return ret;
2792 }
2793
2794 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2795 {
2796         size_t ret = 0;
2797         FILE *fp = stdout;
2798         struct callchain_node *chain;
2799         struct rb_node *rb_node = rb_first_cached(root);
2800
2801         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2802         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2803                graph_dotted_line);
2804
2805         while (rb_node) {
2806                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
2807                 rb_node = rb_next(rb_node);
2808
2809                 ret += fprintf(fp, "  ");
2810                 print_sched_time(chain->hit, 12);
2811                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2812                 ret += fprintf(fp, " %8d  ", chain->count);
2813                 ret += callchain__fprintf_folded(fp, chain);
2814                 ret += fprintf(fp, "\n");
2815         }
2816
2817         return ret;
2818 }
2819
2820 static void timehist_print_summary(struct perf_sched *sched,
2821                                    struct perf_session *session)
2822 {
2823         struct machine *m = &session->machines.host;
2824         struct total_run_stats totals;
2825         u64 task_count;
2826         struct thread *t;
2827         struct thread_runtime *r;
2828         int i;
2829         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2830
2831         memset(&totals, 0, sizeof(totals));
2832         totals.sched = sched;
2833
2834         if (sched->idle_hist) {
2835                 printf("\nIdle-time summary\n");
2836                 printf("%*s  parent  sched-out  ", comm_width, "comm");
2837                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2838         } else if (sched->show_state) {
2839                 printf("\nWait-time summary\n");
2840                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2841                 printf("   run-time      sleep      iowait     preempt       delay\n");
2842         } else {
2843                 printf("\nRuntime summary\n");
2844                 printf("%*s  parent   sched-in  ", comm_width, "comm");
2845                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2846         }
2847         printf("%*s            (count)  ", comm_width, "");
2848         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2849                sched->show_state ? "(msec)" : "%");
2850         printf("%.117s\n", graph_dotted_line);
2851
2852         machine__for_each_thread(m, show_thread_runtime, &totals);
2853         task_count = totals.task_count;
2854         if (!task_count)
2855                 printf("<no still running tasks>\n");
2856
2857         /* CPU idle stats not tracked when samples were skipped */
2858         if (sched->skipped_samples && !sched->idle_hist)
2859                 return;
2860
2861         printf("\nIdle stats:\n");
2862         for (i = 0; i < idle_max_cpu; ++i) {
2863                 if (cpu_list && !test_bit(i, cpu_bitmap))
2864                         continue;
2865
2866                 t = idle_threads[i];
2867                 if (!t)
2868                         continue;
2869
2870                 r = thread__priv(t);
2871                 if (r && r->run_stats.n) {
2872                         totals.sched_count += r->run_stats.n;
2873                         printf("    CPU %2d idle for ", i);
2874                         print_sched_time(r->total_run_time, 6);
2875                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2876                 } else
2877                         printf("    CPU %2d idle entire time window\n", i);
2878         }
2879
2880         if (sched->idle_hist && sched->show_callchain) {
2881                 callchain_param.mode  = CHAIN_FOLDED;
2882                 callchain_param.value = CCVAL_PERIOD;
2883
2884                 callchain_register_param(&callchain_param);
2885
2886                 printf("\nIdle stats by callchain:\n");
2887                 for (i = 0; i < idle_max_cpu; ++i) {
2888                         struct idle_thread_runtime *itr;
2889
2890                         t = idle_threads[i];
2891                         if (!t)
2892                                 continue;
2893
2894                         itr = thread__priv(t);
2895                         if (itr == NULL)
2896                                 continue;
2897
2898                         callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2899                                              0, &callchain_param);
2900
2901                         printf("  CPU %2d:", i);
2902                         print_sched_time(itr->tr.total_run_time, 6);
2903                         printf(" msec\n");
2904                         timehist_print_idlehist_callchain(&itr->sorted_root);
2905                         printf("\n");
2906                 }
2907         }
2908
2909         printf("\n"
2910                "    Total number of unique tasks: %" PRIu64 "\n"
2911                "Total number of context switches: %" PRIu64 "\n",
2912                totals.task_count, totals.sched_count);
2913
2914         printf("           Total run time (msec): ");
2915         print_sched_time(totals.total_run_time, 2);
2916         printf("\n");
2917
2918         printf("    Total scheduling time (msec): ");
2919         print_sched_time(hist_time, 2);
2920         printf(" (x %d)\n", sched->max_cpu.cpu);
2921 }
2922
2923 typedef int (*sched_handler)(struct perf_tool *tool,
2924                           union perf_event *event,
2925                           struct evsel *evsel,
2926                           struct perf_sample *sample,
2927                           struct machine *machine);
2928
2929 static int perf_timehist__process_sample(struct perf_tool *tool,
2930                                          union perf_event *event,
2931                                          struct perf_sample *sample,
2932                                          struct evsel *evsel,
2933                                          struct machine *machine)
2934 {
2935         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2936         int err = 0;
2937         struct perf_cpu this_cpu = {
2938                 .cpu = sample->cpu,
2939         };
2940
2941         if (this_cpu.cpu > sched->max_cpu.cpu)
2942                 sched->max_cpu = this_cpu;
2943
2944         if (evsel->handler != NULL) {
2945                 sched_handler f = evsel->handler;
2946
2947                 err = f(tool, event, evsel, sample, machine);
2948         }
2949
2950         return err;
2951 }
2952
2953 static int timehist_check_attr(struct perf_sched *sched,
2954                                struct evlist *evlist)
2955 {
2956         struct evsel *evsel;
2957         struct evsel_runtime *er;
2958
2959         list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2960                 er = evsel__get_runtime(evsel);
2961                 if (er == NULL) {
2962                         pr_err("Failed to allocate memory for evsel runtime data\n");
2963                         return -1;
2964                 }
2965
2966                 if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2967                         pr_info("Samples do not have callchains.\n");
2968                         sched->show_callchain = 0;
2969                         symbol_conf.use_callchain = 0;
2970                 }
2971         }
2972
2973         return 0;
2974 }
2975
2976 static int perf_sched__timehist(struct perf_sched *sched)
2977 {
2978         struct evsel_str_handler handlers[] = {
2979                 { "sched:sched_switch",       timehist_sched_switch_event, },
2980                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
2981                 { "sched:sched_waking",       timehist_sched_wakeup_event, },
2982                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2983         };
2984         const struct evsel_str_handler migrate_handlers[] = {
2985                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
2986         };
2987         struct perf_data data = {
2988                 .path  = input_name,
2989                 .mode  = PERF_DATA_MODE_READ,
2990                 .force = sched->force,
2991         };
2992
2993         struct perf_session *session;
2994         struct evlist *evlist;
2995         int err = -1;
2996
2997         /*
2998          * event handlers for timehist option
2999          */
3000         sched->tool.sample       = perf_timehist__process_sample;
3001         sched->tool.mmap         = perf_event__process_mmap;
3002         sched->tool.comm         = perf_event__process_comm;
3003         sched->tool.exit         = perf_event__process_exit;
3004         sched->tool.fork         = perf_event__process_fork;
3005         sched->tool.lost         = process_lost;
3006         sched->tool.attr         = perf_event__process_attr;
3007         sched->tool.tracing_data = perf_event__process_tracing_data;
3008         sched->tool.build_id     = perf_event__process_build_id;
3009
3010         sched->tool.ordered_events = true;
3011         sched->tool.ordering_requires_timestamps = true;
3012
3013         symbol_conf.use_callchain = sched->show_callchain;
3014
3015         session = perf_session__new(&data, &sched->tool);
3016         if (IS_ERR(session))
3017                 return PTR_ERR(session);
3018
3019         if (cpu_list) {
3020                 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3021                 if (err < 0)
3022                         goto out;
3023         }
3024
3025         evlist = session->evlist;
3026
3027         symbol__init(&session->header.env);
3028
3029         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3030                 pr_err("Invalid time string\n");
3031                 return -EINVAL;
3032         }
3033
3034         if (timehist_check_attr(sched, evlist) != 0)
3035                 goto out;
3036
3037         setup_pager();
3038
3039         /* prefer sched_waking if it is captured */
3040         if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3041                 handlers[1].handler = timehist_sched_wakeup_ignore;
3042
3043         /* setup per-evsel handlers */
3044         if (perf_session__set_tracepoints_handlers(session, handlers))
3045                 goto out;
3046
3047         /* sched_switch event at a minimum needs to exist */
3048         if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3049                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3050                 goto out;
3051         }
3052
3053         if (sched->show_migrations &&
3054             perf_session__set_tracepoints_handlers(session, migrate_handlers))
3055                 goto out;
3056
3057         /* pre-allocate struct for per-CPU idle stats */
3058         sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3059         if (sched->max_cpu.cpu == 0)
3060                 sched->max_cpu.cpu = 4;
3061         if (init_idle_threads(sched->max_cpu.cpu))
3062                 goto out;
3063
3064         /* summary_only implies summary option, but don't overwrite summary if set */
3065         if (sched->summary_only)
3066                 sched->summary = sched->summary_only;
3067
3068         if (!sched->summary_only)
3069                 timehist_header(sched);
3070
3071         err = perf_session__process_events(session);
3072         if (err) {
3073                 pr_err("Failed to process events, error %d", err);
3074                 goto out;
3075         }
3076
3077         sched->nr_events      = evlist->stats.nr_events[0];
3078         sched->nr_lost_events = evlist->stats.total_lost;
3079         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3080
3081         if (sched->summary)
3082                 timehist_print_summary(sched, session);
3083
3084 out:
3085         free_idle_threads();
3086         perf_session__delete(session);
3087
3088         return err;
3089 }
3090
3091
3092 static void print_bad_events(struct perf_sched *sched)
3093 {
3094         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3095                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3096                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3097                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3098         }
3099         if (sched->nr_lost_events && sched->nr_events) {
3100                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3101                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3102                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3103         }
3104         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3105                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3106                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3107                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3108                 if (sched->nr_lost_events)
3109                         printf(" (due to lost events?)");
3110                 printf("\n");
3111         }
3112 }
3113
3114 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3115 {
3116         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3117         struct work_atoms *this;
3118         const char *comm = thread__comm_str(data->thread), *this_comm;
3119         bool leftmost = true;
3120
3121         while (*new) {
3122                 int cmp;
3123
3124                 this = container_of(*new, struct work_atoms, node);
3125                 parent = *new;
3126
3127                 this_comm = thread__comm_str(this->thread);
3128                 cmp = strcmp(comm, this_comm);
3129                 if (cmp > 0) {
3130                         new = &((*new)->rb_left);
3131                 } else if (cmp < 0) {
3132                         new = &((*new)->rb_right);
3133                         leftmost = false;
3134                 } else {
3135                         this->num_merged++;
3136                         this->total_runtime += data->total_runtime;
3137                         this->nb_atoms += data->nb_atoms;
3138                         this->total_lat += data->total_lat;
3139                         list_splice(&data->work_list, &this->work_list);
3140                         if (this->max_lat < data->max_lat) {
3141                                 this->max_lat = data->max_lat;
3142                                 this->max_lat_start = data->max_lat_start;
3143                                 this->max_lat_end = data->max_lat_end;
3144                         }
3145                         zfree(&data);
3146                         return;
3147                 }
3148         }
3149
3150         data->num_merged++;
3151         rb_link_node(&data->node, parent, new);
3152         rb_insert_color_cached(&data->node, root, leftmost);
3153 }
3154
3155 static void perf_sched__merge_lat(struct perf_sched *sched)
3156 {
3157         struct work_atoms *data;
3158         struct rb_node *node;
3159
3160         if (sched->skip_merge)
3161                 return;
3162
3163         while ((node = rb_first_cached(&sched->atom_root))) {
3164                 rb_erase_cached(node, &sched->atom_root);
3165                 data = rb_entry(node, struct work_atoms, node);
3166                 __merge_work_atoms(&sched->merged_atom_root, data);
3167         }
3168 }
3169
3170 static int setup_cpus_switch_event(struct perf_sched *sched)
3171 {
3172         unsigned int i;
3173
3174         sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3175         if (!sched->cpu_last_switched)
3176                 return -1;
3177
3178         sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3179         if (!sched->curr_pid) {
3180                 zfree(&sched->cpu_last_switched);
3181                 return -1;
3182         }
3183
3184         for (i = 0; i < MAX_CPUS; i++)
3185                 sched->curr_pid[i] = -1;
3186
3187         return 0;
3188 }
3189
3190 static void free_cpus_switch_event(struct perf_sched *sched)
3191 {
3192         zfree(&sched->curr_pid);
3193         zfree(&sched->cpu_last_switched);
3194 }
3195
3196 static int perf_sched__lat(struct perf_sched *sched)
3197 {
3198         int rc = -1;
3199         struct rb_node *next;
3200
3201         setup_pager();
3202
3203         if (setup_cpus_switch_event(sched))
3204                 return rc;
3205
3206         if (perf_sched__read_events(sched))
3207                 goto out_free_cpus_switch_event;
3208
3209         perf_sched__merge_lat(sched);
3210         perf_sched__sort_lat(sched);
3211
3212         printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3213         printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3214         printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3215
3216         next = rb_first_cached(&sched->sorted_atom_root);
3217
3218         while (next) {
3219                 struct work_atoms *work_list;
3220
3221                 work_list = rb_entry(next, struct work_atoms, node);
3222                 output_lat_thread(sched, work_list);
3223                 next = rb_next(next);
3224                 thread__zput(work_list->thread);
3225         }
3226
3227         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3228         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3229                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3230
3231         printf(" ---------------------------------------------------\n");
3232
3233         print_bad_events(sched);
3234         printf("\n");
3235
3236         rc = 0;
3237
3238 out_free_cpus_switch_event:
3239         free_cpus_switch_event(sched);
3240         return rc;
3241 }
3242
3243 static int setup_map_cpus(struct perf_sched *sched)
3244 {
3245         sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3246
3247         if (sched->map.comp) {
3248                 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3249                 if (!sched->map.comp_cpus)
3250                         return -1;
3251         }
3252
3253         if (sched->map.cpus_str) {
3254                 sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3255                 if (!sched->map.cpus) {
3256                         pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3257                         zfree(&sched->map.comp_cpus);
3258                         return -1;
3259                 }
3260         }
3261
3262         return 0;
3263 }
3264
3265 static int setup_color_pids(struct perf_sched *sched)
3266 {
3267         struct perf_thread_map *map;
3268
3269         if (!sched->map.color_pids_str)
3270                 return 0;
3271
3272         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3273         if (!map) {
3274                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3275                 return -1;
3276         }
3277
3278         sched->map.color_pids = map;
3279         return 0;
3280 }
3281
3282 static int setup_color_cpus(struct perf_sched *sched)
3283 {
3284         struct perf_cpu_map *map;
3285
3286         if (!sched->map.color_cpus_str)
3287                 return 0;
3288
3289         map = perf_cpu_map__new(sched->map.color_cpus_str);
3290         if (!map) {
3291                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3292                 return -1;
3293         }
3294
3295         sched->map.color_cpus = map;
3296         return 0;
3297 }
3298
3299 static int perf_sched__map(struct perf_sched *sched)
3300 {
3301         int rc = -1;
3302
3303         sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3304         if (!sched->curr_thread)
3305                 return rc;
3306
3307         if (setup_cpus_switch_event(sched))
3308                 goto out_free_curr_thread;
3309
3310         if (setup_map_cpus(sched))
3311                 goto out_free_cpus_switch_event;
3312
3313         if (setup_color_pids(sched))
3314                 goto out_put_map_cpus;
3315
3316         if (setup_color_cpus(sched))
3317                 goto out_put_color_pids;
3318
3319         setup_pager();
3320         if (perf_sched__read_events(sched))
3321                 goto out_put_color_cpus;
3322
3323         rc = 0;
3324         print_bad_events(sched);
3325
3326 out_put_color_cpus:
3327         perf_cpu_map__put(sched->map.color_cpus);
3328
3329 out_put_color_pids:
3330         perf_thread_map__put(sched->map.color_pids);
3331
3332 out_put_map_cpus:
3333         zfree(&sched->map.comp_cpus);
3334         perf_cpu_map__put(sched->map.cpus);
3335
3336 out_free_cpus_switch_event:
3337         free_cpus_switch_event(sched);
3338
3339 out_free_curr_thread:
3340         zfree(&sched->curr_thread);
3341         return rc;
3342 }
3343
3344 static int perf_sched__replay(struct perf_sched *sched)
3345 {
3346         int ret;
3347         unsigned long i;
3348
3349         mutex_init(&sched->start_work_mutex);
3350         mutex_init(&sched->work_done_wait_mutex);
3351
3352         ret = setup_cpus_switch_event(sched);
3353         if (ret)
3354                 goto out_mutex_destroy;
3355
3356         calibrate_run_measurement_overhead(sched);
3357         calibrate_sleep_measurement_overhead(sched);
3358
3359         test_calibrations(sched);
3360
3361         ret = perf_sched__read_events(sched);
3362         if (ret)
3363                 goto out_free_cpus_switch_event;
3364
3365         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3366         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3367         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3368
3369         if (sched->targetless_wakeups)
3370                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3371         if (sched->multitarget_wakeups)
3372                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3373         if (sched->nr_run_events_optimized)
3374                 printf("run atoms optimized: %ld\n",
3375                         sched->nr_run_events_optimized);
3376
3377         print_task_traces(sched);
3378         add_cross_task_wakeups(sched);
3379
3380         sched->thread_funcs_exit = false;
3381         create_tasks(sched);
3382         printf("------------------------------------------------------------\n");
3383         for (i = 0; i < sched->replay_repeat; i++)
3384                 run_one_test(sched);
3385
3386         sched->thread_funcs_exit = true;
3387         destroy_tasks(sched);
3388
3389 out_free_cpus_switch_event:
3390         free_cpus_switch_event(sched);
3391
3392 out_mutex_destroy:
3393         mutex_destroy(&sched->start_work_mutex);
3394         mutex_destroy(&sched->work_done_wait_mutex);
3395         return ret;
3396 }
3397
3398 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3399                           const char * const usage_msg[])
3400 {
3401         char *tmp, *tok, *str = strdup(sched->sort_order);
3402
3403         for (tok = strtok_r(str, ", ", &tmp);
3404                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3405                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3406                         usage_with_options_msg(usage_msg, options,
3407                                         "Unknown --sort key: `%s'", tok);
3408                 }
3409         }
3410
3411         free(str);
3412
3413         sort_dimension__add("pid", &sched->cmp_pid);
3414 }
3415
3416 static bool schedstat_events_exposed(void)
3417 {
3418         /*
3419          * Select "sched:sched_stat_wait" event to check
3420          * whether schedstat tracepoints are exposed.
3421          */
3422         return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3423                 false : true;
3424 }
3425
3426 static int __cmd_record(int argc, const char **argv)
3427 {
3428         unsigned int rec_argc, i, j;
3429         char **rec_argv;
3430         const char **rec_argv_copy;
3431         const char * const record_args[] = {
3432                 "record",
3433                 "-a",
3434                 "-R",
3435                 "-m", "1024",
3436                 "-c", "1",
3437                 "-e", "sched:sched_switch",
3438                 "-e", "sched:sched_stat_runtime",
3439                 "-e", "sched:sched_process_fork",
3440                 "-e", "sched:sched_wakeup_new",
3441                 "-e", "sched:sched_migrate_task",
3442         };
3443
3444         /*
3445          * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3446          * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3447          * to prevent "perf sched record" execution failure, determine
3448          * whether to record schedstat events according to actual situation.
3449          */
3450         const char * const schedstat_args[] = {
3451                 "-e", "sched:sched_stat_wait",
3452                 "-e", "sched:sched_stat_sleep",
3453                 "-e", "sched:sched_stat_iowait",
3454         };
3455         unsigned int schedstat_argc = schedstat_events_exposed() ?
3456                 ARRAY_SIZE(schedstat_args) : 0;
3457
3458         struct tep_event *waking_event;
3459         int ret;
3460
3461         /*
3462          * +2 for either "-e", "sched:sched_wakeup" or
3463          * "-e", "sched:sched_waking"
3464          */
3465         rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3466         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3467         if (rec_argv == NULL)
3468                 return -ENOMEM;
3469         rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3470         if (rec_argv_copy == NULL) {
3471                 free(rec_argv);
3472                 return -ENOMEM;
3473         }
3474
3475         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3476                 rec_argv[i] = strdup(record_args[i]);
3477
3478         rec_argv[i++] = strdup("-e");
3479         waking_event = trace_event__tp_format("sched", "sched_waking");
3480         if (!IS_ERR(waking_event))
3481                 rec_argv[i++] = strdup("sched:sched_waking");
3482         else
3483                 rec_argv[i++] = strdup("sched:sched_wakeup");
3484
3485         for (j = 0; j < schedstat_argc; j++)
3486                 rec_argv[i++] = strdup(schedstat_args[j]);
3487
3488         for (j = 1; j < (unsigned int)argc; j++, i++)
3489                 rec_argv[i] = strdup(argv[j]);
3490
3491         BUG_ON(i != rec_argc);
3492
3493         memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3494         ret = cmd_record(rec_argc, rec_argv_copy);
3495
3496         for (i = 0; i < rec_argc; i++)
3497                 free(rec_argv[i]);
3498         free(rec_argv);
3499         free(rec_argv_copy);
3500
3501         return ret;
3502 }
3503
3504 int cmd_sched(int argc, const char **argv)
3505 {
3506         static const char default_sort_order[] = "avg, max, switch, runtime";
3507         struct perf_sched sched = {
3508                 .tool = {
3509                         .sample          = perf_sched__process_tracepoint_sample,
3510                         .comm            = perf_sched__process_comm,
3511                         .namespaces      = perf_event__process_namespaces,
3512                         .lost            = perf_event__process_lost,
3513                         .fork            = perf_sched__process_fork_event,
3514                         .ordered_events = true,
3515                 },
3516                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3517                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3518                 .sort_order           = default_sort_order,
3519                 .replay_repeat        = 10,
3520                 .profile_cpu          = -1,
3521                 .next_shortname1      = 'A',
3522                 .next_shortname2      = '0',
3523                 .skip_merge           = 0,
3524                 .show_callchain       = 1,
3525                 .max_stack            = 5,
3526         };
3527         const struct option sched_options[] = {
3528         OPT_STRING('i', "input", &input_name, "file",
3529                     "input file name"),
3530         OPT_INCR('v', "verbose", &verbose,
3531                     "be more verbose (show symbol address, etc)"),
3532         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3533                     "dump raw trace in ASCII"),
3534         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3535         OPT_END()
3536         };
3537         const struct option latency_options[] = {
3538         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3539                    "sort by key(s): runtime, switch, avg, max"),
3540         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3541                     "CPU to profile on"),
3542         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3543                     "latency stats per pid instead of per comm"),
3544         OPT_PARENT(sched_options)
3545         };
3546         const struct option replay_options[] = {
3547         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3548                      "repeat the workload replay N times (-1: infinite)"),
3549         OPT_PARENT(sched_options)
3550         };
3551         const struct option map_options[] = {
3552         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3553                     "map output in compact mode"),
3554         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3555                    "highlight given pids in map"),
3556         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3557                     "highlight given CPUs in map"),
3558         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3559                     "display given CPUs in map"),
3560         OPT_PARENT(sched_options)
3561         };
3562         const struct option timehist_options[] = {
3563         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3564                    "file", "vmlinux pathname"),
3565         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3566                    "file", "kallsyms pathname"),
3567         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3568                     "Display call chains if present (default on)"),
3569         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3570                    "Maximum number of functions to display backtrace."),
3571         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3572                     "Look for files with symbols relative to this directory"),
3573         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3574                     "Show only syscall summary with statistics"),
3575         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3576                     "Show all syscalls and summary with statistics"),
3577         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3578         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3579         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3580         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3581         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3582         OPT_STRING(0, "time", &sched.time_str, "str",
3583                    "Time span for analysis (start,stop)"),
3584         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3585         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3586                    "analyze events only for given process id(s)"),
3587         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3588                    "analyze events only for given thread id(s)"),
3589         OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3590         OPT_PARENT(sched_options)
3591         };
3592
3593         const char * const latency_usage[] = {
3594                 "perf sched latency [<options>]",
3595                 NULL
3596         };
3597         const char * const replay_usage[] = {
3598                 "perf sched replay [<options>]",
3599                 NULL
3600         };
3601         const char * const map_usage[] = {
3602                 "perf sched map [<options>]",
3603                 NULL
3604         };
3605         const char * const timehist_usage[] = {
3606                 "perf sched timehist [<options>]",
3607                 NULL
3608         };
3609         const char *const sched_subcommands[] = { "record", "latency", "map",
3610                                                   "replay", "script",
3611                                                   "timehist", NULL };
3612         const char *sched_usage[] = {
3613                 NULL,
3614                 NULL
3615         };
3616         struct trace_sched_handler lat_ops  = {
3617                 .wakeup_event       = latency_wakeup_event,
3618                 .switch_event       = latency_switch_event,
3619                 .runtime_event      = latency_runtime_event,
3620                 .migrate_task_event = latency_migrate_task_event,
3621         };
3622         struct trace_sched_handler map_ops  = {
3623                 .switch_event       = map_switch_event,
3624         };
3625         struct trace_sched_handler replay_ops  = {
3626                 .wakeup_event       = replay_wakeup_event,
3627                 .switch_event       = replay_switch_event,
3628                 .fork_event         = replay_fork_event,
3629         };
3630         int ret;
3631
3632         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3633                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3634         if (!argc)
3635                 usage_with_options(sched_usage, sched_options);
3636
3637         /*
3638          * Aliased to 'perf script' for now:
3639          */
3640         if (!strcmp(argv[0], "script")) {
3641                 return cmd_script(argc, argv);
3642         } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3643                 return __cmd_record(argc, argv);
3644         } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3645                 sched.tp_handler = &lat_ops;
3646                 if (argc > 1) {
3647                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3648                         if (argc)
3649                                 usage_with_options(latency_usage, latency_options);
3650                 }
3651                 setup_sorting(&sched, latency_options, latency_usage);
3652                 return perf_sched__lat(&sched);
3653         } else if (!strcmp(argv[0], "map")) {
3654                 if (argc) {
3655                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3656                         if (argc)
3657                                 usage_with_options(map_usage, map_options);
3658                 }
3659                 sched.tp_handler = &map_ops;
3660                 setup_sorting(&sched, latency_options, latency_usage);
3661                 return perf_sched__map(&sched);
3662         } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3663                 sched.tp_handler = &replay_ops;
3664                 if (argc) {
3665                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3666                         if (argc)
3667                                 usage_with_options(replay_usage, replay_options);
3668                 }
3669                 return perf_sched__replay(&sched);
3670         } else if (!strcmp(argv[0], "timehist")) {
3671                 if (argc) {
3672                         argc = parse_options(argc, argv, timehist_options,
3673                                              timehist_usage, 0);
3674                         if (argc)
3675                                 usage_with_options(timehist_usage, timehist_options);
3676                 }
3677                 if ((sched.show_wakeups || sched.show_next) &&
3678                     sched.summary_only) {
3679                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3680                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3681                         if (sched.show_wakeups)
3682                                 parse_options_usage(NULL, timehist_options, "w", true);
3683                         if (sched.show_next)
3684                                 parse_options_usage(NULL, timehist_options, "n", true);
3685                         return -EINVAL;
3686                 }
3687                 ret = symbol__validate_sym_arguments();
3688                 if (ret)
3689                         return ret;
3690
3691                 return perf_sched__timehist(&sched);
3692         } else {
3693                 usage_with_options(sched_usage, sched_options);
3694         }
3695
3696         return 0;
3697 }