kernel/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
  15 #include "util/parse-options.h"
  16 #include "util/trace-event.h"
  17
  18 #include "util/debug.h"
  19
  20 #include <sys/prctl.h>
  21 #include <sys/resource.h>
  22
  23 #include <semaphore.h>
  24 #include <pthread.h>
  25 #include <math.h>
  26 #include <api/fs/fs.h>
  27
  28 #define PR_SET_NAME             15               /* Set process name */
  29 #define MAX_CPUS                4096
  30 #define COMM_LEN                20
  31 #define SYM_LEN                 129
  32 #define MAX_PID                 1024000
  33
  34 struct sched_atom;
  35
  36 struct task_desc {
  37         unsigned long           nr;
  38         unsigned long           pid;
  39         char                    comm[COMM_LEN];
  40
  41         unsigned long           nr_events;
  42         unsigned long           curr_event;
  43         struct sched_atom       **atoms;
  44
  45         pthread_t               thread;
  46         sem_t                   sleep_sem;
  47
  48         sem_t                   ready_for_work;
  49         sem_t                   work_done_sem;
  50
  51         u64                     cpu_usage;
  52 };
  53
  54 enum sched_event_type {
  55         SCHED_EVENT_RUN,
  56         SCHED_EVENT_SLEEP,
  57         SCHED_EVENT_WAKEUP,
  58         SCHED_EVENT_MIGRATION,
  59 };
  60
  61 struct sched_atom {
  62         enum sched_event_type   type;
  63         int                     specific_wait;
  64         u64                     timestamp;
  65         u64                     duration;
  66         unsigned long           nr;
  67         sem_t                   *wait_sem;
  68         struct task_desc        *wakee;
  69 };
  70
  71 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  72
  73 enum thread_state {
  74         THREAD_SLEEPING = 0,
  75         THREAD_WAIT_CPU,
  76         THREAD_SCHED_IN,
  77         THREAD_IGNORE
  78 };
  79
  80 struct work_atom {
  81         struct list_head        list;
  82         enum thread_state       state;
  83         u64                     sched_out_time;
  84         u64                     wake_up_time;
  85         u64                     sched_in_time;
  86         u64                     runtime;
  87 };
  88
  89 struct work_atoms {
  90         struct list_head        work_list;
  91         struct thread           *thread;
  92         struct rb_node          node;
  93         u64                     max_lat;
  94         u64                     max_lat_at;
  95         u64                     total_lat;
  96         u64                     nb_atoms;
  97         u64                     total_runtime;
  98         int                     num_merged;
  99 };
 100
 101 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 102
 103 struct perf_sched;
 104
 105 struct trace_sched_handler {
 106         int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 107                             struct perf_sample *sample, struct machine *machine);
 108
 109         int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 110                              struct perf_sample *sample, struct machine *machine);
 111
 112         int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 113                             struct perf_sample *sample, struct machine *machine);
 114
 115         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 116         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 117                           struct machine *machine);
 118
 119         int (*migrate_task_event)(struct perf_sched *sched,
 120                                   struct perf_evsel *evsel,
 121                                   struct perf_sample *sample,
 122                                   struct machine *machine);
 123 };
 124
 125 struct perf_sched {
 126         struct perf_tool tool;
 127         const char       *sort_order;
 128         unsigned long    nr_tasks;
 129         struct task_desc **pid_to_task;
 130         struct task_desc **tasks;
 131         const struct trace_sched_handler *tp_handler;
 132         pthread_mutex_t  start_work_mutex;
 133         pthread_mutex_t  work_done_wait_mutex;
 134         int              profile_cpu;
 135 /*
 136  * Track the current task - that way we can know whether there's any
 137  * weird events, such as a task being switched away that is not current.
 138  */
 139         int              max_cpu;
 140         u32              curr_pid[MAX_CPUS];
 141         struct thread    *curr_thread[MAX_CPUS];
 142         char             next_shortname1;
 143         char             next_shortname2;
 144         unsigned int     replay_repeat;
 145         unsigned long    nr_run_events;
 146         unsigned long    nr_sleep_events;
 147         unsigned long    nr_wakeup_events;
 148         unsigned long    nr_sleep_corrections;
 149         unsigned long    nr_run_events_optimized;
 150         unsigned long    targetless_wakeups;
 151         unsigned long    multitarget_wakeups;
 152         unsigned long    nr_runs;
 153         unsigned long    nr_timestamps;
 154         unsigned long    nr_unordered_timestamps;
 155         unsigned long    nr_context_switch_bugs;
 156         unsigned long    nr_events;
 157         unsigned long    nr_lost_chunks;
 158         unsigned long    nr_lost_events;
 159         u64              run_measurement_overhead;
 160         u64              sleep_measurement_overhead;
 161         u64              start_time;
 162         u64              cpu_usage;
 163         u64              runavg_cpu_usage;
 164         u64              parent_cpu_usage;
 165         u64              runavg_parent_cpu_usage;
 166         u64              sum_runtime;
 167         u64              sum_fluct;
 168         u64              run_avg;
 169         u64              all_runtime;
 170         u64              all_count;
 171         u64              cpu_last_switched[MAX_CPUS];
 172         struct rb_root   atom_root, sorted_atom_root, merged_atom_root;
 173         struct list_head sort_list, cmp_pid;
 174         bool force;
 175         bool skip_merge;
 176 };
 177
 178 static u64 get_nsecs(void)
 179 {
 180         struct timespec ts;
 181
 182         clock_gettime(CLOCK_MONOTONIC, &ts);
 183
 184         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
 185 }
 186
 187 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 188 {
 189         u64 T0 = get_nsecs(), T1;
 190
 191         do {
 192                 T1 = get_nsecs();
 193         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 194 }
 195
 196 static void sleep_nsecs(u64 nsecs)
 197 {
 198         struct timespec ts;
 199
 200         ts.tv_nsec = nsecs % 999999999;
 201         ts.tv_sec = nsecs / 999999999;
 202
 203         nanosleep(&ts, NULL);
 204 }
 205
 206 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 207 {
 208         u64 T0, T1, delta, min_delta = 1000000000ULL;
 209         int i;
 210
 211         for (i = 0; i < 10; i++) {
 212                 T0 = get_nsecs();
 213                 burn_nsecs(sched, 0);
 214                 T1 = get_nsecs();
 215                 delta = T1-T0;
 216                 min_delta = min(min_delta, delta);
 217         }
 218         sched->run_measurement_overhead = min_delta;
 219
 220         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 221 }
 222
 223 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 224 {
 225         u64 T0, T1, delta, min_delta = 1000000000ULL;
 226         int i;
 227
 228         for (i = 0; i < 10; i++) {
 229                 T0 = get_nsecs();
 230                 sleep_nsecs(10000);
 231                 T1 = get_nsecs();
 232                 delta = T1-T0;
 233                 min_delta = min(min_delta, delta);
 234         }
 235         min_delta -= 10000;
 236         sched->sleep_measurement_overhead = min_delta;
 237
 238         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 239 }
 240
 241 static struct sched_atom *
 242 get_new_event(struct task_desc *task, u64 timestamp)
 243 {
 244         struct sched_atom *event = zalloc(sizeof(*event));
 245         unsigned long idx = task->nr_events;
 246         size_t size;
 247
 248         event->timestamp = timestamp;
 249         event->nr = idx;
 250
 251         task->nr_events++;
 252         size = sizeof(struct sched_atom *) * task->nr_events;
 253         task->atoms = realloc(task->atoms, size);
 254         BUG_ON(!task->atoms);
 255
 256         task->atoms[idx] = event;
 257
 258         return event;
 259 }
 260
 261 static struct sched_atom *last_event(struct task_desc *task)
 262 {
 263         if (!task->nr_events)
 264                 return NULL;
 265
 266         return task->atoms[task->nr_events - 1];
 267 }
 268
 269 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 270                                 u64 timestamp, u64 duration)
 271 {
 272         struct sched_atom *event, *curr_event = last_event(task);
 273
 274         /*
 275          * optimize an existing RUN event by merging this one
 276          * to it:
 277          */
 278         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 279                 sched->nr_run_events_optimized++;
 280                 curr_event->duration += duration;
 281                 return;
 282         }
 283
 284         event = get_new_event(task, timestamp);
 285
 286         event->type = SCHED_EVENT_RUN;
 287         event->duration = duration;
 288
 289         sched->nr_run_events++;
 290 }
 291
 292 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 293                                    u64 timestamp, struct task_desc *wakee)
 294 {
 295         struct sched_atom *event, *wakee_event;
 296
 297         event = get_new_event(task, timestamp);
 298         event->type = SCHED_EVENT_WAKEUP;
 299         event->wakee = wakee;
 300
 301         wakee_event = last_event(wakee);
 302         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 303                 sched->targetless_wakeups++;
 304                 return;
 305         }
 306         if (wakee_event->wait_sem) {
 307                 sched->multitarget_wakeups++;
 308                 return;
 309         }
 310
 311         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 312         sem_init(wakee_event->wait_sem, 0, 0);
 313         wakee_event->specific_wait = 1;
 314         event->wait_sem = wakee_event->wait_sem;
 315
 316         sched->nr_wakeup_events++;
 317 }
 318
 319 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 320                                   u64 timestamp, u64 task_state __maybe_unused)
 321 {
 322         struct sched_atom *event = get_new_event(task, timestamp);
 323
 324         event->type = SCHED_EVENT_SLEEP;
 325
 326         sched->nr_sleep_events++;
 327 }
 328
 329 static struct task_desc *register_pid(struct perf_sched *sched,
 330                                       unsigned long pid, const char *comm)
 331 {
 332         struct task_desc *task;
 333         static int pid_max;
 334
 335         if (sched->pid_to_task == NULL) {
 336                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 337                         pid_max = MAX_PID;
 338                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 339         }
 340         if (pid >= (unsigned long)pid_max) {
 341                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 342                         sizeof(struct task_desc *))) == NULL);
 343                 while (pid >= (unsigned long)pid_max)
 344                         sched->pid_to_task[pid_max++] = NULL;
 345         }
 346
 347         task = sched->pid_to_task[pid];
 348
 349         if (task)
 350                 return task;
 351
 352         task = zalloc(sizeof(*task));
 353         task->pid = pid;
 354         task->nr = sched->nr_tasks;
 355         strcpy(task->comm, comm);
 356         /*
 357          * every task starts in sleeping state - this gets ignored
 358          * if there's no wakeup pointing to this sleep state:
 359          */
 360         add_sched_event_sleep(sched, task, 0, 0);
 361
 362         sched->pid_to_task[pid] = task;
 363         sched->nr_tasks++;
 364         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 365         BUG_ON(!sched->tasks);
 366         sched->tasks[task->nr] = task;
 367
 368         if (verbose)
 369                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 370
 371         return task;
 372 }
 373
 374
 375 static void print_task_traces(struct perf_sched *sched)
 376 {
 377         struct task_desc *task;
 378         unsigned long i;
 379
 380         for (i = 0; i < sched->nr_tasks; i++) {
 381                 task = sched->tasks[i];
 382                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 383                         task->nr, task->comm, task->pid, task->nr_events);
 384         }
 385 }
 386
 387 static void add_cross_task_wakeups(struct perf_sched *sched)
 388 {
 389         struct task_desc *task1, *task2;
 390         unsigned long i, j;
 391
 392         for (i = 0; i < sched->nr_tasks; i++) {
 393                 task1 = sched->tasks[i];
 394                 j = i + 1;
 395                 if (j == sched->nr_tasks)
 396                         j = 0;
 397                 task2 = sched->tasks[j];
 398                 add_sched_event_wakeup(sched, task1, 0, task2);
 399         }
 400 }
 401
 402 static void perf_sched__process_event(struct perf_sched *sched,
 403                                       struct sched_atom *atom)
 404 {
 405         int ret = 0;
 406
 407         switch (atom->type) {
 408                 case SCHED_EVENT_RUN:
 409                         burn_nsecs(sched, atom->duration);
 410                         break;
 411                 case SCHED_EVENT_SLEEP:
 412                         if (atom->wait_sem)
 413                                 ret = sem_wait(atom->wait_sem);
 414                         BUG_ON(ret);
 415                         break;
 416                 case SCHED_EVENT_WAKEUP:
 417                         if (atom->wait_sem)
 418                                 ret = sem_post(atom->wait_sem);
 419                         BUG_ON(ret);
 420                         break;
 421                 case SCHED_EVENT_MIGRATION:
 422                         break;
 423                 default:
 424                         BUG_ON(1);
 425         }
 426 }
 427
 428 static u64 get_cpu_usage_nsec_parent(void)
 429 {
 430         struct rusage ru;
 431         u64 sum;
 432         int err;
 433
 434         err = getrusage(RUSAGE_SELF, &ru);
 435         BUG_ON(err);
 436
 437         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
 438         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
 439
 440         return sum;
 441 }
 442
 443 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 444 {
 445         struct perf_event_attr attr;
 446         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 447         int fd;
 448         struct rlimit limit;
 449         bool need_privilege = false;
 450
 451         memset(&attr, 0, sizeof(attr));
 452
 453         attr.type = PERF_TYPE_SOFTWARE;
 454         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 455
 456 force_again:
 457         fd = sys_perf_event_open(&attr, 0, -1, -1,
 458                                  perf_event_open_cloexec_flag());
 459
 460         if (fd < 0) {
 461                 if (errno == EMFILE) {
 462                         if (sched->force) {
 463                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 464                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 465                                 if (limit.rlim_cur > limit.rlim_max) {
 466                                         limit.rlim_max = limit.rlim_cur;
 467                                         need_privilege = true;
 468                                 }
 469                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 470                                         if (need_privilege && errno == EPERM)
 471                                                 strcpy(info, "Need privilege\n");
 472                                 } else
 473                                         goto force_again;
 474                         } else
 475                                 strcpy(info, "Have a try with -f option\n");
 476                 }
 477                 pr_err("Error: sys_perf_event_open() syscall returned "
 478                        "with %d (%s)\n%s", fd,
 479                        strerror_r(errno, sbuf, sizeof(sbuf)), info);
 480                 exit(EXIT_FAILURE);
 481         }
 482         return fd;
 483 }
 484
 485 static u64 get_cpu_usage_nsec_self(int fd)
 486 {
 487         u64 runtime;
 488         int ret;
 489
 490         ret = read(fd, &runtime, sizeof(runtime));
 491         BUG_ON(ret != sizeof(runtime));
 492
 493         return runtime;
 494 }
 495
 496 struct sched_thread_parms {
 497         struct task_desc  *task;
 498         struct perf_sched *sched;
 499         int fd;
 500 };
 501
 502 static void *thread_func(void *ctx)
 503 {
 504         struct sched_thread_parms *parms = ctx;
 505         struct task_desc *this_task = parms->task;
 506         struct perf_sched *sched = parms->sched;
 507         u64 cpu_usage_0, cpu_usage_1;
 508         unsigned long i, ret;
 509         char comm2[22];
 510         int fd = parms->fd;
 511
 512         zfree(&parms);
 513
 514         sprintf(comm2, ":%s", this_task->comm);
 515         prctl(PR_SET_NAME, comm2);
 516         if (fd < 0)
 517                 return NULL;
 518 again:
 519         ret = sem_post(&this_task->ready_for_work);
 520         BUG_ON(ret);
 521         ret = pthread_mutex_lock(&sched->start_work_mutex);
 522         BUG_ON(ret);
 523         ret = pthread_mutex_unlock(&sched->start_work_mutex);
 524         BUG_ON(ret);
 525
 526         cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 527
 528         for (i = 0; i < this_task->nr_events; i++) {
 529                 this_task->curr_event = i;
 530                 perf_sched__process_event(sched, this_task->atoms[i]);
 531         }
 532
 533         cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 534         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 535         ret = sem_post(&this_task->work_done_sem);
 536         BUG_ON(ret);
 537
 538         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 539         BUG_ON(ret);
 540         ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 541         BUG_ON(ret);
 542
 543         goto again;
 544 }
 545
 546 static void create_tasks(struct perf_sched *sched)
 547 {
 548         struct task_desc *task;
 549         pthread_attr_t attr;
 550         unsigned long i;
 551         int err;
 552
 553         err = pthread_attr_init(&attr);
 554         BUG_ON(err);
 555         err = pthread_attr_setstacksize(&attr,
 556                         (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 557         BUG_ON(err);
 558         err = pthread_mutex_lock(&sched->start_work_mutex);
 559         BUG_ON(err);
 560         err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 561         BUG_ON(err);
 562         for (i = 0; i < sched->nr_tasks; i++) {
 563                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 564                 BUG_ON(parms == NULL);
 565                 parms->task = task = sched->tasks[i];
 566                 parms->sched = sched;
 567                 parms->fd = self_open_counters(sched, i);
 568                 sem_init(&task->sleep_sem, 0, 0);
 569                 sem_init(&task->ready_for_work, 0, 0);
 570                 sem_init(&task->work_done_sem, 0, 0);
 571                 task->curr_event = 0;
 572                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 573                 BUG_ON(err);
 574         }
 575 }
 576
 577 static void wait_for_tasks(struct perf_sched *sched)
 578 {
 579         u64 cpu_usage_0, cpu_usage_1;
 580         struct task_desc *task;
 581         unsigned long i, ret;
 582
 583         sched->start_time = get_nsecs();
 584         sched->cpu_usage = 0;
 585         pthread_mutex_unlock(&sched->work_done_wait_mutex);
 586
 587         for (i = 0; i < sched->nr_tasks; i++) {
 588                 task = sched->tasks[i];
 589                 ret = sem_wait(&task->ready_for_work);
 590                 BUG_ON(ret);
 591                 sem_init(&task->ready_for_work, 0, 0);
 592         }
 593         ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 594         BUG_ON(ret);
 595
 596         cpu_usage_0 = get_cpu_usage_nsec_parent();
 597
 598         pthread_mutex_unlock(&sched->start_work_mutex);
 599
 600         for (i = 0; i < sched->nr_tasks; i++) {
 601                 task = sched->tasks[i];
 602                 ret = sem_wait(&task->work_done_sem);
 603                 BUG_ON(ret);
 604                 sem_init(&task->work_done_sem, 0, 0);
 605                 sched->cpu_usage += task->cpu_usage;
 606                 task->cpu_usage = 0;
 607         }
 608
 609         cpu_usage_1 = get_cpu_usage_nsec_parent();
 610         if (!sched->runavg_cpu_usage)
 611                 sched->runavg_cpu_usage = sched->cpu_usage;
 612         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 613
 614         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 615         if (!sched->runavg_parent_cpu_usage)
 616                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 617         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 618                                          sched->parent_cpu_usage)/sched->replay_repeat;
 619
 620         ret = pthread_mutex_lock(&sched->start_work_mutex);
 621         BUG_ON(ret);
 622
 623         for (i = 0; i < sched->nr_tasks; i++) {
 624                 task = sched->tasks[i];
 625                 sem_init(&task->sleep_sem, 0, 0);
 626                 task->curr_event = 0;
 627         }
 628 }
 629
 630 static void run_one_test(struct perf_sched *sched)
 631 {
 632         u64 T0, T1, delta, avg_delta, fluct;
 633
 634         T0 = get_nsecs();
 635         wait_for_tasks(sched);
 636         T1 = get_nsecs();
 637
 638         delta = T1 - T0;
 639         sched->sum_runtime += delta;
 640         sched->nr_runs++;
 641
 642         avg_delta = sched->sum_runtime / sched->nr_runs;
 643         if (delta < avg_delta)
 644                 fluct = avg_delta - delta;
 645         else
 646                 fluct = delta - avg_delta;
 647         sched->sum_fluct += fluct;
 648         if (!sched->run_avg)
 649                 sched->run_avg = delta;
 650         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 651
 652         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
 653
 654         printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
 655
 656         printf("cpu: %0.2f / %0.2f",
 657                 (double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
 658
 659 #if 0
 660         /*
 661          * rusage statistics done by the parent, these are less
 662          * accurate than the sched->sum_exec_runtime based statistics:
 663          */
 664         printf(" [%0.2f / %0.2f]",
 665                 (double)sched->parent_cpu_usage/1e6,
 666                 (double)sched->runavg_parent_cpu_usage/1e6);
 667 #endif
 668
 669         printf("\n");
 670
 671         if (sched->nr_sleep_corrections)
 672                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 673         sched->nr_sleep_corrections = 0;
 674 }
 675
 676 static void test_calibrations(struct perf_sched *sched)
 677 {
 678         u64 T0, T1;
 679
 680         T0 = get_nsecs();
 681         burn_nsecs(sched, 1e6);
 682         T1 = get_nsecs();
 683
 684         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 685
 686         T0 = get_nsecs();
 687         sleep_nsecs(1e6);
 688         T1 = get_nsecs();
 689
 690         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 691 }
 692
 693 static int
 694 replay_wakeup_event(struct perf_sched *sched,
 695                     struct perf_evsel *evsel, struct perf_sample *sample,
 696                     struct machine *machine __maybe_unused)
 697 {
 698         const char *comm = perf_evsel__strval(evsel, sample, "comm");
 699         const u32 pid    = perf_evsel__intval(evsel, sample, "pid");
 700         struct task_desc *waker, *wakee;
 701
 702         if (verbose) {
 703                 printf("sched_wakeup event %p\n", evsel);
 704
 705                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 706         }
 707
 708         waker = register_pid(sched, sample->tid, "<unknown>");
 709         wakee = register_pid(sched, pid, comm);
 710
 711         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 712         return 0;
 713 }
 714
 715 static int replay_switch_event(struct perf_sched *sched,
 716                                struct perf_evsel *evsel,
 717                                struct perf_sample *sample,
 718                                struct machine *machine __maybe_unused)
 719 {
 720         const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 721                    *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 722         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 723                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 724         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 725         struct task_desc *prev, __maybe_unused *next;
 726         u64 timestamp0, timestamp = sample->time;
 727         int cpu = sample->cpu;
 728         s64 delta;
 729
 730         if (verbose)
 731                 printf("sched_switch event %p\n", evsel);
 732
 733         if (cpu >= MAX_CPUS || cpu < 0)
 734                 return 0;
 735
 736         timestamp0 = sched->cpu_last_switched[cpu];
 737         if (timestamp0)
 738                 delta = timestamp - timestamp0;
 739         else
 740                 delta = 0;
 741
 742         if (delta < 0) {
 743                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 744                 return -1;
 745         }
 746
 747         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 748                  prev_comm, prev_pid, next_comm, next_pid, delta);
 749
 750         prev = register_pid(sched, prev_pid, prev_comm);
 751         next = register_pid(sched, next_pid, next_comm);
 752
 753         sched->cpu_last_switched[cpu] = timestamp;
 754
 755         add_sched_event_run(sched, prev, timestamp, delta);
 756         add_sched_event_sleep(sched, prev, timestamp, prev_state);
 757
 758         return 0;
 759 }
 760
 761 static int replay_fork_event(struct perf_sched *sched,
 762                              union perf_event *event,
 763                              struct machine *machine)
 764 {
 765         struct thread *child, *parent;
 766
 767         child = machine__findnew_thread(machine, event->fork.pid,
 768                                         event->fork.tid);
 769         parent = machine__findnew_thread(machine, event->fork.ppid,
 770                                          event->fork.ptid);
 771
 772         if (child == NULL || parent == NULL) {
 773                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 774                                  child, parent);
 775                 goto out_put;
 776         }
 777
 778         if (verbose) {
 779                 printf("fork event\n");
 780                 printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 781                 printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 782         }
 783
 784         register_pid(sched, parent->tid, thread__comm_str(parent));
 785         register_pid(sched, child->tid, thread__comm_str(child));
 786 out_put:
 787         thread__put(child);
 788         thread__put(parent);
 789         return 0;
 790 }
 791
 792 struct sort_dimension {
 793         const char              *name;
 794         sort_fn_t               cmp;
 795         struct list_head        list;
 796 };
 797
 798 static int
 799 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 800 {
 801         struct sort_dimension *sort;
 802         int ret = 0;
 803
 804         BUG_ON(list_empty(list));
 805
 806         list_for_each_entry(sort, list, list) {
 807                 ret = sort->cmp(l, r);
 808                 if (ret)
 809                         return ret;
 810         }
 811
 812         return ret;
 813 }
 814
 815 static struct work_atoms *
 816 thread_atoms_search(struct rb_root *root, struct thread *thread,
 817                          struct list_head *sort_list)
 818 {
 819         struct rb_node *node = root->rb_node;
 820         struct work_atoms key = { .thread = thread };
 821
 822         while (node) {
 823                 struct work_atoms *atoms;
 824                 int cmp;
 825
 826                 atoms = container_of(node, struct work_atoms, node);
 827
 828                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 829                 if (cmp > 0)
 830                         node = node->rb_left;
 831                 else if (cmp < 0)
 832                         node = node->rb_right;
 833                 else {
 834                         BUG_ON(thread != atoms->thread);
 835                         return atoms;
 836                 }
 837         }
 838         return NULL;
 839 }
 840
 841 static void
 842 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 843                          struct list_head *sort_list)
 844 {
 845         struct rb_node **new = &(root->rb_node), *parent = NULL;
 846
 847         while (*new) {
 848                 struct work_atoms *this;
 849                 int cmp;
 850
 851                 this = container_of(*new, struct work_atoms, node);
 852                 parent = *new;
 853
 854                 cmp = thread_lat_cmp(sort_list, data, this);
 855
 856                 if (cmp > 0)
 857                         new = &((*new)->rb_left);
 858                 else
 859                         new = &((*new)->rb_right);
 860         }
 861
 862         rb_link_node(&data->node, parent, new);
 863         rb_insert_color(&data->node, root);
 864 }
 865
 866 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
 867 {
 868         struct work_atoms *atoms = zalloc(sizeof(*atoms));
 869         if (!atoms) {
 870                 pr_err("No memory at %s\n", __func__);
 871                 return -1;
 872         }
 873
 874         atoms->thread = thread__get(thread);
 875         INIT_LIST_HEAD(&atoms->work_list);
 876         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
 877         return 0;
 878 }
 879
 880 static char sched_out_state(u64 prev_state)
 881 {
 882         const char *str = TASK_STATE_TO_CHAR_STR;
 883
 884         return str[prev_state];
 885 }
 886
 887 static int
 888 add_sched_out_event(struct work_atoms *atoms,
 889                     char run_state,
 890                     u64 timestamp)
 891 {
 892         struct work_atom *atom = zalloc(sizeof(*atom));
 893         if (!atom) {
 894                 pr_err("Non memory at %s", __func__);
 895                 return -1;
 896         }
 897
 898         atom->sched_out_time = timestamp;
 899
 900         if (run_state == 'R') {
 901                 atom->state = THREAD_WAIT_CPU;
 902                 atom->wake_up_time = atom->sched_out_time;
 903         }
 904
 905         list_add_tail(&atom->list, &atoms->work_list);
 906         return 0;
 907 }
 908
 909 static void
 910 add_runtime_event(struct work_atoms *atoms, u64 delta,
 911                   u64 timestamp __maybe_unused)
 912 {
 913         struct work_atom *atom;
 914
 915         BUG_ON(list_empty(&atoms->work_list));
 916
 917         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 918
 919         atom->runtime += delta;
 920         atoms->total_runtime += delta;
 921 }
 922
 923 static void
 924 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 925 {
 926         struct work_atom *atom;
 927         u64 delta;
 928
 929         if (list_empty(&atoms->work_list))
 930                 return;
 931
 932         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 933
 934         if (atom->state != THREAD_WAIT_CPU)
 935                 return;
 936
 937         if (timestamp < atom->wake_up_time) {
 938                 atom->state = THREAD_IGNORE;
 939                 return;
 940         }
 941
 942         atom->state = THREAD_SCHED_IN;
 943         atom->sched_in_time = timestamp;
 944
 945         delta = atom->sched_in_time - atom->wake_up_time;
 946         atoms->total_lat += delta;
 947         if (delta > atoms->max_lat) {
 948                 atoms->max_lat = delta;
 949                 atoms->max_lat_at = timestamp;
 950         }
 951         atoms->nb_atoms++;
 952 }
 953
 954 static int latency_switch_event(struct perf_sched *sched,
 955                                 struct perf_evsel *evsel,
 956                                 struct perf_sample *sample,
 957                                 struct machine *machine)
 958 {
 959         const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 960                   next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 961         const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 962         struct work_atoms *out_events, *in_events;
 963         struct thread *sched_out, *sched_in;
 964         u64 timestamp0, timestamp = sample->time;
 965         int cpu = sample->cpu, err = -1;
 966         s64 delta;
 967
 968         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 969
 970         timestamp0 = sched->cpu_last_switched[cpu];
 971         sched->cpu_last_switched[cpu] = timestamp;
 972         if (timestamp0)
 973                 delta = timestamp - timestamp0;
 974         else
 975                 delta = 0;
 976
 977         if (delta < 0) {
 978                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 979                 return -1;
 980         }
 981
 982         sched_out = machine__findnew_thread(machine, -1, prev_pid);
 983         sched_in = machine__findnew_thread(machine, -1, next_pid);
 984         if (sched_out == NULL || sched_in == NULL)
 985                 goto out_put;
 986
 987         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 988         if (!out_events) {
 989                 if (thread_atoms_insert(sched, sched_out))
 990                         goto out_put;
 991                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
 992                 if (!out_events) {
 993                         pr_err("out-event: Internal tree error");
 994                         goto out_put;
 995                 }
 996         }
 997         if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
 998                 return -1;
 999
1000         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1001         if (!in_events) {
1002                 if (thread_atoms_insert(sched, sched_in))
1003                         goto out_put;
1004                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1005                 if (!in_events) {
1006                         pr_err("in-event: Internal tree error");
1007                         goto out_put;
1008                 }
1009                 /*
1010                  * Take came in we have not heard about yet,
1011                  * add in an initial atom in runnable state:
1012                  */
1013                 if (add_sched_out_event(in_events, 'R', timestamp))
1014                         goto out_put;
1015         }
1016         add_sched_in_event(in_events, timestamp);
1017         err = 0;
1018 out_put:
1019         thread__put(sched_out);
1020         thread__put(sched_in);
1021         return err;
1022 }
1023
1024 static int latency_runtime_event(struct perf_sched *sched,
1025                                  struct perf_evsel *evsel,
1026                                  struct perf_sample *sample,
1027                                  struct machine *machine)
1028 {
1029         const u32 pid      = perf_evsel__intval(evsel, sample, "pid");
1030         const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1031         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1032         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1033         u64 timestamp = sample->time;
1034         int cpu = sample->cpu, err = -1;
1035
1036         if (thread == NULL)
1037                 return -1;
1038
1039         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040         if (!atoms) {
1041                 if (thread_atoms_insert(sched, thread))
1042                         goto out_put;
1043                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1044                 if (!atoms) {
1045                         pr_err("in-event: Internal tree error");
1046                         goto out_put;
1047                 }
1048                 if (add_sched_out_event(atoms, 'R', timestamp))
1049                         goto out_put;
1050         }
1051
1052         add_runtime_event(atoms, runtime, timestamp);
1053         err = 0;
1054 out_put:
1055         thread__put(thread);
1056         return err;
1057 }
1058
1059 static int latency_wakeup_event(struct perf_sched *sched,
1060                                 struct perf_evsel *evsel,
1061                                 struct perf_sample *sample,
1062                                 struct machine *machine)
1063 {
1064         const u32 pid     = perf_evsel__intval(evsel, sample, "pid");
1065         struct work_atoms *atoms;
1066         struct work_atom *atom;
1067         struct thread *wakee;
1068         u64 timestamp = sample->time;
1069         int err = -1;
1070
1071         wakee = machine__findnew_thread(machine, -1, pid);
1072         if (wakee == NULL)
1073                 return -1;
1074         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1075         if (!atoms) {
1076                 if (thread_atoms_insert(sched, wakee))
1077                         goto out_put;
1078                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1079                 if (!atoms) {
1080                         pr_err("wakeup-event: Internal tree error");
1081                         goto out_put;
1082                 }
1083                 if (add_sched_out_event(atoms, 'S', timestamp))
1084                         goto out_put;
1085         }
1086
1087         BUG_ON(list_empty(&atoms->work_list));
1088
1089         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1090
1091         /*
1092          * As we do not guarantee the wakeup event happens when
1093          * task is out of run queue, also may happen when task is
1094          * on run queue and wakeup only change ->state to TASK_RUNNING,
1095          * then we should not set the ->wake_up_time when wake up a
1096          * task which is on run queue.
1097          *
1098          * You WILL be missing events if you've recorded only
1099          * one CPU, or are only looking at only one, so don't
1100          * skip in this case.
1101          */
1102         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1103                 goto out_ok;
1104
1105         sched->nr_timestamps++;
1106         if (atom->sched_out_time > timestamp) {
1107                 sched->nr_unordered_timestamps++;
1108                 goto out_ok;
1109         }
1110
1111         atom->state = THREAD_WAIT_CPU;
1112         atom->wake_up_time = timestamp;
1113 out_ok:
1114         err = 0;
1115 out_put:
1116         thread__put(wakee);
1117         return err;
1118 }
1119
1120 static int latency_migrate_task_event(struct perf_sched *sched,
1121                                       struct perf_evsel *evsel,
1122                                       struct perf_sample *sample,
1123                                       struct machine *machine)
1124 {
1125         const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1126         u64 timestamp = sample->time;
1127         struct work_atoms *atoms;
1128         struct work_atom *atom;
1129         struct thread *migrant;
1130         int err = -1;
1131
1132         /*
1133          * Only need to worry about migration when profiling one CPU.
1134          */
1135         if (sched->profile_cpu == -1)
1136                 return 0;
1137
1138         migrant = machine__findnew_thread(machine, -1, pid);
1139         if (migrant == NULL)
1140                 return -1;
1141         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1142         if (!atoms) {
1143                 if (thread_atoms_insert(sched, migrant))
1144                         goto out_put;
1145                 register_pid(sched, migrant->tid, thread__comm_str(migrant));
1146                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1147                 if (!atoms) {
1148                         pr_err("migration-event: Internal tree error");
1149                         goto out_put;
1150                 }
1151                 if (add_sched_out_event(atoms, 'R', timestamp))
1152                         goto out_put;
1153         }
1154
1155         BUG_ON(list_empty(&atoms->work_list));
1156
1157         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1158         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1159
1160         sched->nr_timestamps++;
1161
1162         if (atom->sched_out_time > timestamp)
1163                 sched->nr_unordered_timestamps++;
1164         err = 0;
1165 out_put:
1166         thread__put(migrant);
1167         return err;
1168 }
1169
1170 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1171 {
1172         int i;
1173         int ret;
1174         u64 avg;
1175
1176         if (!work_list->nb_atoms)
1177                 return;
1178         /*
1179          * Ignore idle threads:
1180          */
1181         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1182                 return;
1183
1184         sched->all_runtime += work_list->total_runtime;
1185         sched->all_count   += work_list->nb_atoms;
1186
1187         if (work_list->num_merged > 1)
1188                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1189         else
1190                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1191
1192         for (i = 0; i < 24 - ret; i++)
1193                 printf(" ");
1194
1195         avg = work_list->total_lat / work_list->nb_atoms;
1196
1197         printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1198               (double)work_list->total_runtime / 1e6,
1199                  work_list->nb_atoms, (double)avg / 1e6,
1200                  (double)work_list->max_lat / 1e6,
1201                  (double)work_list->max_lat_at / 1e9);
1202 }
1203
1204 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1205 {
1206         if (l->thread == r->thread)
1207                 return 0;
1208         if (l->thread->tid < r->thread->tid)
1209                 return -1;
1210         if (l->thread->tid > r->thread->tid)
1211                 return 1;
1212         return (int)(l->thread - r->thread);
1213 }
1214
1215 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1216 {
1217         u64 avgl, avgr;
1218
1219         if (!l->nb_atoms)
1220                 return -1;
1221
1222         if (!r->nb_atoms)
1223                 return 1;
1224
1225         avgl = l->total_lat / l->nb_atoms;
1226         avgr = r->total_lat / r->nb_atoms;
1227
1228         if (avgl < avgr)
1229                 return -1;
1230         if (avgl > avgr)
1231                 return 1;
1232
1233         return 0;
1234 }
1235
1236 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1237 {
1238         if (l->max_lat < r->max_lat)
1239                 return -1;
1240         if (l->max_lat > r->max_lat)
1241                 return 1;
1242
1243         return 0;
1244 }
1245
1246 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1247 {
1248         if (l->nb_atoms < r->nb_atoms)
1249                 return -1;
1250         if (l->nb_atoms > r->nb_atoms)
1251                 return 1;
1252
1253         return 0;
1254 }
1255
1256 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1257 {
1258         if (l->total_runtime < r->total_runtime)
1259                 return -1;
1260         if (l->total_runtime > r->total_runtime)
1261                 return 1;
1262
1263         return 0;
1264 }
1265
1266 static int sort_dimension__add(const char *tok, struct list_head *list)
1267 {
1268         size_t i;
1269         static struct sort_dimension avg_sort_dimension = {
1270                 .name = "avg",
1271                 .cmp  = avg_cmp,
1272         };
1273         static struct sort_dimension max_sort_dimension = {
1274                 .name = "max",
1275                 .cmp  = max_cmp,
1276         };
1277         static struct sort_dimension pid_sort_dimension = {
1278                 .name = "pid",
1279                 .cmp  = pid_cmp,
1280         };
1281         static struct sort_dimension runtime_sort_dimension = {
1282                 .name = "runtime",
1283                 .cmp  = runtime_cmp,
1284         };
1285         static struct sort_dimension switch_sort_dimension = {
1286                 .name = "switch",
1287                 .cmp  = switch_cmp,
1288         };
1289         struct sort_dimension *available_sorts[] = {
1290                 &pid_sort_dimension,
1291                 &avg_sort_dimension,
1292                 &max_sort_dimension,
1293                 &switch_sort_dimension,
1294                 &runtime_sort_dimension,
1295         };
1296
1297         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1298                 if (!strcmp(available_sorts[i]->name, tok)) {
1299                         list_add_tail(&available_sorts[i]->list, list);
1300
1301                         return 0;
1302                 }
1303         }
1304
1305         return -1;
1306 }
1307
1308 static void perf_sched__sort_lat(struct perf_sched *sched)
1309 {
1310         struct rb_node *node;
1311         struct rb_root *root = &sched->atom_root;
1312 again:
1313         for (;;) {
1314                 struct work_atoms *data;
1315                 node = rb_first(root);
1316                 if (!node)
1317                         break;
1318
1319                 rb_erase(node, root);
1320                 data = rb_entry(node, struct work_atoms, node);
1321                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1322         }
1323         if (root == &sched->atom_root) {
1324                 root = &sched->merged_atom_root;
1325                 goto again;
1326         }
1327 }
1328
1329 static int process_sched_wakeup_event(struct perf_tool *tool,
1330                                       struct perf_evsel *evsel,
1331                                       struct perf_sample *sample,
1332                                       struct machine *machine)
1333 {
1334         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1335
1336         if (sched->tp_handler->wakeup_event)
1337                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1338
1339         return 0;
1340 }
1341
1342 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1343                             struct perf_sample *sample, struct machine *machine)
1344 {
1345         const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1346         struct thread *sched_in;
1347         int new_shortname;
1348         u64 timestamp0, timestamp = sample->time;
1349         s64 delta;
1350         int cpu, this_cpu = sample->cpu;
1351
1352         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1353
1354         if (this_cpu > sched->max_cpu)
1355                 sched->max_cpu = this_cpu;
1356
1357         timestamp0 = sched->cpu_last_switched[this_cpu];
1358         sched->cpu_last_switched[this_cpu] = timestamp;
1359         if (timestamp0)
1360                 delta = timestamp - timestamp0;
1361         else
1362                 delta = 0;
1363
1364         if (delta < 0) {
1365                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1366                 return -1;
1367         }
1368
1369         sched_in = machine__findnew_thread(machine, -1, next_pid);
1370         if (sched_in == NULL)
1371                 return -1;
1372
1373         sched->curr_thread[this_cpu] = thread__get(sched_in);
1374
1375         printf("  ");
1376
1377         new_shortname = 0;
1378         if (!sched_in->shortname[0]) {
1379                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1380                         /*
1381                          * Don't allocate a letter-number for swapper:0
1382                          * as a shortname. Instead, we use '.' for it.
1383                          */
1384                         sched_in->shortname[0] = '.';
1385                         sched_in->shortname[1] = ' ';
1386                 } else {
1387                         sched_in->shortname[0] = sched->next_shortname1;
1388                         sched_in->shortname[1] = sched->next_shortname2;
1389
1390                         if (sched->next_shortname1 < 'Z') {
1391                                 sched->next_shortname1++;
1392                         } else {
1393                                 sched->next_shortname1 = 'A';
1394                                 if (sched->next_shortname2 < '9')
1395                                         sched->next_shortname2++;
1396                                 else
1397                                         sched->next_shortname2 = '0';
1398                         }
1399                 }
1400                 new_shortname = 1;
1401         }
1402
1403         for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1404                 if (cpu != this_cpu)
1405                         printf(" ");
1406                 else
1407                         printf("*");
1408
1409                 if (sched->curr_thread[cpu])
1410                         printf("%2s ", sched->curr_thread[cpu]->shortname);
1411                 else
1412                         printf("   ");
1413         }
1414
1415         printf("  %12.6f secs ", (double)timestamp/1e9);
1416         if (new_shortname) {
1417                 printf("%s => %s:%d\n",
1418                        sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1419         } else {
1420                 printf("\n");
1421         }
1422
1423         thread__put(sched_in);
1424
1425         return 0;
1426 }
1427
1428 static int process_sched_switch_event(struct perf_tool *tool,
1429                                       struct perf_evsel *evsel,
1430                                       struct perf_sample *sample,
1431                                       struct machine *machine)
1432 {
1433         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1434         int this_cpu = sample->cpu, err = 0;
1435         u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1436             next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1437
1438         if (sched->curr_pid[this_cpu] != (u32)-1) {
1439                 /*
1440                  * Are we trying to switch away a PID that is
1441                  * not current?
1442                  */
1443                 if (sched->curr_pid[this_cpu] != prev_pid)
1444                         sched->nr_context_switch_bugs++;
1445         }
1446
1447         if (sched->tp_handler->switch_event)
1448                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1449
1450         sched->curr_pid[this_cpu] = next_pid;
1451         return err;
1452 }
1453
1454 static int process_sched_runtime_event(struct perf_tool *tool,
1455                                        struct perf_evsel *evsel,
1456                                        struct perf_sample *sample,
1457                                        struct machine *machine)
1458 {
1459         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1460
1461         if (sched->tp_handler->runtime_event)
1462                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1463
1464         return 0;
1465 }
1466
1467 static int perf_sched__process_fork_event(struct perf_tool *tool,
1468                                           union perf_event *event,
1469                                           struct perf_sample *sample,
1470                                           struct machine *machine)
1471 {
1472         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1473
1474         /* run the fork event through the perf machineruy */
1475         perf_event__process_fork(tool, event, sample, machine);
1476
1477         /* and then run additional processing needed for this command */
1478         if (sched->tp_handler->fork_event)
1479                 return sched->tp_handler->fork_event(sched, event, machine);
1480
1481         return 0;
1482 }
1483
1484 static int process_sched_migrate_task_event(struct perf_tool *tool,
1485                                             struct perf_evsel *evsel,
1486                                             struct perf_sample *sample,
1487                                             struct machine *machine)
1488 {
1489         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1490
1491         if (sched->tp_handler->migrate_task_event)
1492                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1493
1494         return 0;
1495 }
1496
1497 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1498                                   struct perf_evsel *evsel,
1499                                   struct perf_sample *sample,
1500                                   struct machine *machine);
1501
1502 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1503                                                  union perf_event *event __maybe_unused,
1504                                                  struct perf_sample *sample,
1505                                                  struct perf_evsel *evsel,
1506                                                  struct machine *machine)
1507 {
1508         int err = 0;
1509
1510         if (evsel->handler != NULL) {
1511                 tracepoint_handler f = evsel->handler;
1512                 err = f(tool, evsel, sample, machine);
1513         }
1514
1515         return err;
1516 }
1517
1518 static int perf_sched__read_events(struct perf_sched *sched)
1519 {
1520         const struct perf_evsel_str_handler handlers[] = {
1521                 { "sched:sched_switch",       process_sched_switch_event, },
1522                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1523                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1524                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1525                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1526         };
1527         struct perf_session *session;
1528         struct perf_data_file file = {
1529                 .path = input_name,
1530                 .mode = PERF_DATA_MODE_READ,
1531                 .force = sched->force,
1532         };
1533         int rc = -1;
1534
1535         session = perf_session__new(&file, false, &sched->tool);
1536         if (session == NULL) {
1537                 pr_debug("No Memory for session\n");
1538                 return -1;
1539         }
1540
1541         symbol__init(&session->header.env);
1542
1543         if (perf_session__set_tracepoints_handlers(session, handlers))
1544                 goto out_delete;
1545
1546         if (perf_session__has_traces(session, "record -R")) {
1547                 int err = perf_session__process_events(session);
1548                 if (err) {
1549                         pr_err("Failed to process events, error %d", err);
1550                         goto out_delete;
1551                 }
1552
1553                 sched->nr_events      = session->evlist->stats.nr_events[0];
1554                 sched->nr_lost_events = session->evlist->stats.total_lost;
1555                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1556         }
1557
1558         rc = 0;
1559 out_delete:
1560         perf_session__delete(session);
1561         return rc;
1562 }
1563
1564 static void print_bad_events(struct perf_sched *sched)
1565 {
1566         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1567                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1568                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1569                         sched->nr_unordered_timestamps, sched->nr_timestamps);
1570         }
1571         if (sched->nr_lost_events && sched->nr_events) {
1572                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1573                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1574                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1575         }
1576         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1577                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1578                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1579                         sched->nr_context_switch_bugs, sched->nr_timestamps);
1580                 if (sched->nr_lost_events)
1581                         printf(" (due to lost events?)");
1582                 printf("\n");
1583         }
1584 }
1585
1586 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
1587 {
1588         struct rb_node **new = &(root->rb_node), *parent = NULL;
1589         struct work_atoms *this;
1590         const char *comm = thread__comm_str(data->thread), *this_comm;
1591
1592         while (*new) {
1593                 int cmp;
1594
1595                 this = container_of(*new, struct work_atoms, node);
1596                 parent = *new;
1597
1598                 this_comm = thread__comm_str(this->thread);
1599                 cmp = strcmp(comm, this_comm);
1600                 if (cmp > 0) {
1601                         new = &((*new)->rb_left);
1602                 } else if (cmp < 0) {
1603                         new = &((*new)->rb_right);
1604                 } else {
1605                         this->num_merged++;
1606                         this->total_runtime += data->total_runtime;
1607                         this->nb_atoms += data->nb_atoms;
1608                         this->total_lat += data->total_lat;
1609                         list_splice(&data->work_list, &this->work_list);
1610                         if (this->max_lat < data->max_lat) {
1611                                 this->max_lat = data->max_lat;
1612                                 this->max_lat_at = data->max_lat_at;
1613                         }
1614                         zfree(&data);
1615                         return;
1616                 }
1617         }
1618
1619         data->num_merged++;
1620         rb_link_node(&data->node, parent, new);
1621         rb_insert_color(&data->node, root);
1622 }
1623
1624 static void perf_sched__merge_lat(struct perf_sched *sched)
1625 {
1626         struct work_atoms *data;
1627         struct rb_node *node;
1628
1629         if (sched->skip_merge)
1630                 return;
1631
1632         while ((node = rb_first(&sched->atom_root))) {
1633                 rb_erase(node, &sched->atom_root);
1634                 data = rb_entry(node, struct work_atoms, node);
1635                 __merge_work_atoms(&sched->merged_atom_root, data);
1636         }
1637 }
1638
1639 static int perf_sched__lat(struct perf_sched *sched)
1640 {
1641         struct rb_node *next;
1642
1643         setup_pager();
1644
1645         if (perf_sched__read_events(sched))
1646                 return -1;
1647
1648         perf_sched__merge_lat(sched);
1649         perf_sched__sort_lat(sched);
1650
1651         printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1652         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1653         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1654
1655         next = rb_first(&sched->sorted_atom_root);
1656
1657         while (next) {
1658                 struct work_atoms *work_list;
1659
1660                 work_list = rb_entry(next, struct work_atoms, node);
1661                 output_lat_thread(sched, work_list);
1662                 next = rb_next(next);
1663                 thread__zput(work_list->thread);
1664         }
1665
1666         printf(" -----------------------------------------------------------------------------------------------------------------\n");
1667         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1668                 (double)sched->all_runtime / 1e6, sched->all_count);
1669
1670         printf(" ---------------------------------------------------\n");
1671
1672         print_bad_events(sched);
1673         printf("\n");
1674
1675         return 0;
1676 }
1677
1678 static int perf_sched__map(struct perf_sched *sched)
1679 {
1680         sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1681
1682         setup_pager();
1683         if (perf_sched__read_events(sched))
1684                 return -1;
1685         print_bad_events(sched);
1686         return 0;
1687 }
1688
1689 static int perf_sched__replay(struct perf_sched *sched)
1690 {
1691         unsigned long i;
1692
1693         calibrate_run_measurement_overhead(sched);
1694         calibrate_sleep_measurement_overhead(sched);
1695
1696         test_calibrations(sched);
1697
1698         if (perf_sched__read_events(sched))
1699                 return -1;
1700
1701         printf("nr_run_events:        %ld\n", sched->nr_run_events);
1702         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1703         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1704
1705         if (sched->targetless_wakeups)
1706                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1707         if (sched->multitarget_wakeups)
1708                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1709         if (sched->nr_run_events_optimized)
1710                 printf("run atoms optimized: %ld\n",
1711                         sched->nr_run_events_optimized);
1712
1713         print_task_traces(sched);
1714         add_cross_task_wakeups(sched);
1715
1716         create_tasks(sched);
1717         printf("------------------------------------------------------------\n");
1718         for (i = 0; i < sched->replay_repeat; i++)
1719                 run_one_test(sched);
1720
1721         return 0;
1722 }
1723
1724 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1725                           const char * const usage_msg[])
1726 {
1727         char *tmp, *tok, *str = strdup(sched->sort_order);
1728
1729         for (tok = strtok_r(str, ", ", &tmp);
1730                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1731                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1732                         usage_with_options_msg(usage_msg, options,
1733                                         "Unknown --sort key: `%s'", tok);
1734                 }
1735         }
1736
1737         free(str);
1738
1739         sort_dimension__add("pid", &sched->cmp_pid);
1740 }
1741
1742 static int __cmd_record(int argc, const char **argv)
1743 {
1744         unsigned int rec_argc, i, j;
1745         const char **rec_argv;
1746         const char * const record_args[] = {
1747                 "record",
1748                 "-a",
1749                 "-R",
1750                 "-m", "1024",
1751                 "-c", "1",
1752                 "-e", "sched:sched_switch",
1753                 "-e", "sched:sched_stat_wait",
1754                 "-e", "sched:sched_stat_sleep",
1755                 "-e", "sched:sched_stat_iowait",
1756                 "-e", "sched:sched_stat_runtime",
1757                 "-e", "sched:sched_process_fork",
1758                 "-e", "sched:sched_wakeup",
1759                 "-e", "sched:sched_wakeup_new",
1760                 "-e", "sched:sched_migrate_task",
1761         };
1762
1763         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1764         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1765
1766         if (rec_argv == NULL)
1767                 return -ENOMEM;
1768
1769         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1770                 rec_argv[i] = strdup(record_args[i]);
1771
1772         for (j = 1; j < (unsigned int)argc; j++, i++)
1773                 rec_argv[i] = argv[j];
1774
1775         BUG_ON(i != rec_argc);
1776
1777         return cmd_record(i, rec_argv, NULL);
1778 }
1779
1780 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1781 {
1782         const char default_sort_order[] = "avg, max, switch, runtime";
1783         struct perf_sched sched = {
1784                 .tool = {
1785                         .sample          = perf_sched__process_tracepoint_sample,
1786                         .comm            = perf_event__process_comm,
1787                         .lost            = perf_event__process_lost,
1788                         .fork            = perf_sched__process_fork_event,
1789                         .ordered_events = true,
1790                 },
1791                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
1792                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
1793                 .start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1794                 .work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1795                 .sort_order           = default_sort_order,
1796                 .replay_repeat        = 10,
1797                 .profile_cpu          = -1,
1798                 .next_shortname1      = 'A',
1799                 .next_shortname2      = '0',
1800                 .skip_merge           = 0,
1801         };
1802         const struct option latency_options[] = {
1803         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1804                    "sort by key(s): runtime, switch, avg, max"),
1805         OPT_INCR('v', "verbose", &verbose,
1806                     "be more verbose (show symbol address, etc)"),
1807         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1808                     "CPU to profile on"),
1809         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1810                     "dump raw trace in ASCII"),
1811         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
1812                     "latency stats per pid instead of per comm"),
1813         OPT_END()
1814         };
1815         const struct option replay_options[] = {
1816         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1817                      "repeat the workload replay N times (-1: infinite)"),
1818         OPT_INCR('v', "verbose", &verbose,
1819                     "be more verbose (show symbol address, etc)"),
1820         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1821                     "dump raw trace in ASCII"),
1822         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
1823         OPT_END()
1824         };
1825         const struct option sched_options[] = {
1826         OPT_STRING('i', "input", &input_name, "file",
1827                     "input file name"),
1828         OPT_INCR('v', "verbose", &verbose,
1829                     "be more verbose (show symbol address, etc)"),
1830         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1831                     "dump raw trace in ASCII"),
1832         OPT_END()
1833         };
1834         const char * const latency_usage[] = {
1835                 "perf sched latency [<options>]",
1836                 NULL
1837         };
1838         const char * const replay_usage[] = {
1839                 "perf sched replay [<options>]",
1840                 NULL
1841         };
1842         const char *const sched_subcommands[] = { "record", "latency", "map",
1843                                                   "replay", "script", NULL };
1844         const char *sched_usage[] = {
1845                 NULL,
1846                 NULL
1847         };
1848         struct trace_sched_handler lat_ops  = {
1849                 .wakeup_event       = latency_wakeup_event,
1850                 .switch_event       = latency_switch_event,
1851                 .runtime_event      = latency_runtime_event,
1852                 .migrate_task_event = latency_migrate_task_event,
1853         };
1854         struct trace_sched_handler map_ops  = {
1855                 .switch_event       = map_switch_event,
1856         };
1857         struct trace_sched_handler replay_ops  = {
1858                 .wakeup_event       = replay_wakeup_event,
1859                 .switch_event       = replay_switch_event,
1860                 .fork_event         = replay_fork_event,
1861         };
1862         unsigned int i;
1863
1864         for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1865                 sched.curr_pid[i] = -1;
1866
1867         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1868                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1869         if (!argc)
1870                 usage_with_options(sched_usage, sched_options);
1871
1872         /*
1873          * Aliased to 'perf script' for now:
1874          */
1875         if (!strcmp(argv[0], "script"))
1876                 return cmd_script(argc, argv, prefix);
1877
1878         if (!strncmp(argv[0], "rec", 3)) {
1879                 return __cmd_record(argc, argv);
1880         } else if (!strncmp(argv[0], "lat", 3)) {
1881                 sched.tp_handler = &lat_ops;
1882                 if (argc > 1) {
1883                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1884                         if (argc)
1885                                 usage_with_options(latency_usage, latency_options);
1886                 }
1887                 setup_sorting(&sched, latency_options, latency_usage);
1888                 return perf_sched__lat(&sched);
1889         } else if (!strcmp(argv[0], "map")) {
1890                 sched.tp_handler = &map_ops;
1891                 setup_sorting(&sched, latency_options, latency_usage);
1892                 return perf_sched__map(&sched);
1893         } else if (!strncmp(argv[0], "rep", 3)) {
1894                 sched.tp_handler = &replay_ops;
1895                 if (argc) {
1896                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1897                         if (argc)
1898                                 usage_with_options(replay_usage, replay_options);
1899                 }
1900                 return perf_sched__replay(&sched);
1901         } else {
1902                 usage_with_options(sched_usage, sched_options);
1903         }
1904
1905         return 0;
1906 }