tools/perf/builtin-sched.c

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