tools/perf/builtin-sched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf-sys.h"
   4
   5 #include "util/cpumap.h"
   6 #include "util/evlist.h"
   7 #include "util/evsel.h"
   8 #include "util/evsel_fprintf.h"
   9 #include "util/mutex.h"
  10 #include "util/symbol.h"
  11 #include "util/thread.h"
  12 #include "util/header.h"
  13 #include "util/session.h"
  14 #include "util/tool.h"
  15 #include "util/cloexec.h"
  16 #include "util/thread_map.h"
  17 #include "util/color.h"
  18 #include "util/stat.h"
  19 #include "util/string2.h"
  20 #include "util/callchain.h"
  21 #include "util/time-utils.h"
  22
  23 #include <subcmd/pager.h>
  24 #include <subcmd/parse-options.h>
  25 #include "util/trace-event.h"
  26
  27 #include "util/debug.h"
  28 #include "util/event.h"
  29 #include "util/util.h"
  30
  31 #include <linux/kernel.h>
  32 #include <linux/log2.h>
  33 #include <linux/zalloc.h>
  34 #include <sys/prctl.h>
  35 #include <sys/resource.h>
  36 #include <inttypes.h>
  37
  38 #include <errno.h>
  39 #include <semaphore.h>
  40 #include <pthread.h>
  41 #include <math.h>
  42 #include <api/fs/fs.h>
  43 #include <perf/cpumap.h>
  44 #include <linux/time64.h>
  45 #include <linux/err.h>
  46
  47 #include <linux/ctype.h>
  48
  49 #define PR_SET_NAME             15               /* Set process name */
  50 #define MAX_CPUS                4096
  51 #define COMM_LEN                20
  52 #define SYM_LEN                 129
  53 #define MAX_PID                 1024000
  54 #define MAX_PRIO                140
  55
  56 static const char *cpu_list;
  57 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  58
  59 struct sched_atom;
  60
  61 struct task_desc {
  62         unsigned long           nr;
  63         unsigned long           pid;
  64         char                    comm[COMM_LEN];
  65
  66         unsigned long           nr_events;
  67         unsigned long           curr_event;
  68         struct sched_atom       **atoms;
  69
  70         pthread_t               thread;
  71
  72         sem_t                   ready_for_work;
  73         sem_t                   work_done_sem;
  74
  75         u64                     cpu_usage;
  76 };
  77
  78 enum sched_event_type {
  79         SCHED_EVENT_RUN,
  80         SCHED_EVENT_SLEEP,
  81         SCHED_EVENT_WAKEUP,
  82 };
  83
  84 struct sched_atom {
  85         enum sched_event_type   type;
  86         u64                     timestamp;
  87         u64                     duration;
  88         unsigned long           nr;
  89         sem_t                   *wait_sem;
  90         struct task_desc        *wakee;
  91 };
  92
  93 enum thread_state {
  94         THREAD_SLEEPING = 0,
  95         THREAD_WAIT_CPU,
  96         THREAD_SCHED_IN,
  97         THREAD_IGNORE
  98 };
  99
 100 struct work_atom {
 101         struct list_head        list;
 102         enum thread_state       state;
 103         u64                     sched_out_time;
 104         u64                     wake_up_time;
 105         u64                     sched_in_time;
 106         u64                     runtime;
 107 };
 108
 109 struct work_atoms {
 110         struct list_head        work_list;
 111         struct thread           *thread;
 112         struct rb_node          node;
 113         u64                     max_lat;
 114         u64                     max_lat_start;
 115         u64                     max_lat_end;
 116         u64                     total_lat;
 117         u64                     nb_atoms;
 118         u64                     total_runtime;
 119         int                     num_merged;
 120 };
 121
 122 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 123
 124 struct perf_sched;
 125
 126 struct trace_sched_handler {
 127         int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 128                             struct perf_sample *sample, struct machine *machine);
 129
 130         int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 131                              struct perf_sample *sample, struct machine *machine);
 132
 133         int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 134                             struct perf_sample *sample, struct machine *machine);
 135
 136         /* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 137         int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 138                           struct machine *machine);
 139
 140         int (*migrate_task_event)(struct perf_sched *sched,
 141                                   struct evsel *evsel,
 142                                   struct perf_sample *sample,
 143                                   struct machine *machine);
 144 };
 145
 146 #define COLOR_PIDS PERF_COLOR_BLUE
 147 #define COLOR_CPUS PERF_COLOR_BG_RED
 148
 149 struct perf_sched_map {
 150         DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 151         struct perf_cpu         *comp_cpus;
 152         bool                     comp;
 153         struct perf_thread_map *color_pids;
 154         const char              *color_pids_str;
 155         struct perf_cpu_map     *color_cpus;
 156         const char              *color_cpus_str;
 157         const char              *task_name;
 158         struct strlist          *task_names;
 159         bool                    fuzzy;
 160         struct perf_cpu_map     *cpus;
 161         const char              *cpus_str;
 162 };
 163
 164 struct perf_sched {
 165         struct perf_tool tool;
 166         const char       *sort_order;
 167         unsigned long    nr_tasks;
 168         struct task_desc **pid_to_task;
 169         struct task_desc **tasks;
 170         const struct trace_sched_handler *tp_handler;
 171         struct mutex     start_work_mutex;
 172         struct mutex     work_done_wait_mutex;
 173         int              profile_cpu;
 174 /*
 175  * Track the current task - that way we can know whether there's any
 176  * weird events, such as a task being switched away that is not current.
 177  */
 178         struct perf_cpu  max_cpu;
 179         u32              *curr_pid;
 180         struct thread    **curr_thread;
 181         struct thread    **curr_out_thread;
 182         char             next_shortname1;
 183         char             next_shortname2;
 184         unsigned int     replay_repeat;
 185         unsigned long    nr_run_events;
 186         unsigned long    nr_sleep_events;
 187         unsigned long    nr_wakeup_events;
 188         unsigned long    nr_sleep_corrections;
 189         unsigned long    nr_run_events_optimized;
 190         unsigned long    targetless_wakeups;
 191         unsigned long    multitarget_wakeups;
 192         unsigned long    nr_runs;
 193         unsigned long    nr_timestamps;
 194         unsigned long    nr_unordered_timestamps;
 195         unsigned long    nr_context_switch_bugs;
 196         unsigned long    nr_events;
 197         unsigned long    nr_lost_chunks;
 198         unsigned long    nr_lost_events;
 199         u64              run_measurement_overhead;
 200         u64              sleep_measurement_overhead;
 201         u64              start_time;
 202         u64              cpu_usage;
 203         u64              runavg_cpu_usage;
 204         u64              parent_cpu_usage;
 205         u64              runavg_parent_cpu_usage;
 206         u64              sum_runtime;
 207         u64              sum_fluct;
 208         u64              run_avg;
 209         u64              all_runtime;
 210         u64              all_count;
 211         u64              *cpu_last_switched;
 212         struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 213         struct list_head sort_list, cmp_pid;
 214         bool force;
 215         bool skip_merge;
 216         struct perf_sched_map map;
 217
 218         /* options for timehist command */
 219         bool            summary;
 220         bool            summary_only;
 221         bool            idle_hist;
 222         bool            show_callchain;
 223         unsigned int    max_stack;
 224         bool            show_cpu_visual;
 225         bool            show_wakeups;
 226         bool            show_next;
 227         bool            show_migrations;
 228         bool            pre_migrations;
 229         bool            show_state;
 230         bool            show_prio;
 231         u64             skipped_samples;
 232         const char      *time_str;
 233         struct perf_time_interval ptime;
 234         struct perf_time_interval hist_time;
 235         volatile bool   thread_funcs_exit;
 236         const char      *prio_str;
 237         DECLARE_BITMAP(prio_bitmap, MAX_PRIO);
 238 };
 239
 240 /* per thread run time data */
 241 struct thread_runtime {
 242         u64 last_time;      /* time of previous sched in/out event */
 243         u64 dt_run;         /* run time */
 244         u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 245         u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 246         u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 247         u64 dt_delay;       /* time between wakeup and sched-in */
 248         u64 dt_pre_mig;     /* time between migration and wakeup */
 249         u64 ready_to_run;   /* time of wakeup */
 250         u64 migrated;       /* time when a thread is migrated */
 251
 252         struct stats run_stats;
 253         u64 total_run_time;
 254         u64 total_sleep_time;
 255         u64 total_iowait_time;
 256         u64 total_preempt_time;
 257         u64 total_delay_time;
 258         u64 total_pre_mig_time;
 259
 260         char last_state;
 261
 262         char shortname[3];
 263         bool comm_changed;
 264
 265         u64 migrations;
 266
 267         int prio;
 268 };
 269
 270 /* per event run time data */
 271 struct evsel_runtime {
 272         u64 *last_time; /* time this event was last seen per cpu */
 273         u32 ncpu;       /* highest cpu slot allocated */
 274 };
 275
 276 /* per cpu idle time data */
 277 struct idle_thread_runtime {
 278         struct thread_runtime   tr;
 279         struct thread           *last_thread;
 280         struct rb_root_cached   sorted_root;
 281         struct callchain_root   callchain;
 282         struct callchain_cursor cursor;
 283 };
 284
 285 /* track idle times per cpu */
 286 static struct thread **idle_threads;
 287 static int idle_max_cpu;
 288 static char idle_comm[] = "<idle>";
 289
 290 static u64 get_nsecs(void)
 291 {
 292         struct timespec ts;
 293
 294         clock_gettime(CLOCK_MONOTONIC, &ts);
 295
 296         return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 297 }
 298
 299 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 300 {
 301         u64 T0 = get_nsecs(), T1;
 302
 303         do {
 304                 T1 = get_nsecs();
 305         } while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 306 }
 307
 308 static void sleep_nsecs(u64 nsecs)
 309 {
 310         struct timespec ts;
 311
 312         ts.tv_nsec = nsecs % 999999999;
 313         ts.tv_sec = nsecs / 999999999;
 314
 315         nanosleep(&ts, NULL);
 316 }
 317
 318 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 319 {
 320         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 321         int i;
 322
 323         for (i = 0; i < 10; i++) {
 324                 T0 = get_nsecs();
 325                 burn_nsecs(sched, 0);
 326                 T1 = get_nsecs();
 327                 delta = T1-T0;
 328                 min_delta = min(min_delta, delta);
 329         }
 330         sched->run_measurement_overhead = min_delta;
 331
 332         printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 333 }
 334
 335 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 336 {
 337         u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 338         int i;
 339
 340         for (i = 0; i < 10; i++) {
 341                 T0 = get_nsecs();
 342                 sleep_nsecs(10000);
 343                 T1 = get_nsecs();
 344                 delta = T1-T0;
 345                 min_delta = min(min_delta, delta);
 346         }
 347         min_delta -= 10000;
 348         sched->sleep_measurement_overhead = min_delta;
 349
 350         printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 351 }
 352
 353 static struct sched_atom *
 354 get_new_event(struct task_desc *task, u64 timestamp)
 355 {
 356         struct sched_atom *event = zalloc(sizeof(*event));
 357         unsigned long idx = task->nr_events;
 358         size_t size;
 359
 360         event->timestamp = timestamp;
 361         event->nr = idx;
 362
 363         task->nr_events++;
 364         size = sizeof(struct sched_atom *) * task->nr_events;
 365         task->atoms = realloc(task->atoms, size);
 366         BUG_ON(!task->atoms);
 367
 368         task->atoms[idx] = event;
 369
 370         return event;
 371 }
 372
 373 static struct sched_atom *last_event(struct task_desc *task)
 374 {
 375         if (!task->nr_events)
 376                 return NULL;
 377
 378         return task->atoms[task->nr_events - 1];
 379 }
 380
 381 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 382                                 u64 timestamp, u64 duration)
 383 {
 384         struct sched_atom *event, *curr_event = last_event(task);
 385
 386         /*
 387          * optimize an existing RUN event by merging this one
 388          * to it:
 389          */
 390         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 391                 sched->nr_run_events_optimized++;
 392                 curr_event->duration += duration;
 393                 return;
 394         }
 395
 396         event = get_new_event(task, timestamp);
 397
 398         event->type = SCHED_EVENT_RUN;
 399         event->duration = duration;
 400
 401         sched->nr_run_events++;
 402 }
 403
 404 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 405                                    u64 timestamp, struct task_desc *wakee)
 406 {
 407         struct sched_atom *event, *wakee_event;
 408
 409         event = get_new_event(task, timestamp);
 410         event->type = SCHED_EVENT_WAKEUP;
 411         event->wakee = wakee;
 412
 413         wakee_event = last_event(wakee);
 414         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 415                 sched->targetless_wakeups++;
 416                 return;
 417         }
 418         if (wakee_event->wait_sem) {
 419                 sched->multitarget_wakeups++;
 420                 return;
 421         }
 422
 423         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 424         sem_init(wakee_event->wait_sem, 0, 0);
 425         event->wait_sem = wakee_event->wait_sem;
 426
 427         sched->nr_wakeup_events++;
 428 }
 429
 430 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 431                                   u64 timestamp)
 432 {
 433         struct sched_atom *event = get_new_event(task, timestamp);
 434
 435         event->type = SCHED_EVENT_SLEEP;
 436
 437         sched->nr_sleep_events++;
 438 }
 439
 440 static struct task_desc *register_pid(struct perf_sched *sched,
 441                                       unsigned long pid, const char *comm)
 442 {
 443         struct task_desc *task;
 444         static int pid_max;
 445
 446         if (sched->pid_to_task == NULL) {
 447                 if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 448                         pid_max = MAX_PID;
 449                 BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 450         }
 451         if (pid >= (unsigned long)pid_max) {
 452                 BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 453                         sizeof(struct task_desc *))) == NULL);
 454                 while (pid >= (unsigned long)pid_max)
 455                         sched->pid_to_task[pid_max++] = NULL;
 456         }
 457
 458         task = sched->pid_to_task[pid];
 459
 460         if (task)
 461                 return task;
 462
 463         task = zalloc(sizeof(*task));
 464         task->pid = pid;
 465         task->nr = sched->nr_tasks;
 466         strcpy(task->comm, comm);
 467         /*
 468          * every task starts in sleeping state - this gets ignored
 469          * if there's no wakeup pointing to this sleep state:
 470          */
 471         add_sched_event_sleep(sched, task, 0);
 472
 473         sched->pid_to_task[pid] = task;
 474         sched->nr_tasks++;
 475         sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 476         BUG_ON(!sched->tasks);
 477         sched->tasks[task->nr] = task;
 478
 479         if (verbose > 0)
 480                 printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 481
 482         return task;
 483 }
 484
 485
 486 static void print_task_traces(struct perf_sched *sched)
 487 {
 488         struct task_desc *task;
 489         unsigned long i;
 490
 491         for (i = 0; i < sched->nr_tasks; i++) {
 492                 task = sched->tasks[i];
 493                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 494                         task->nr, task->comm, task->pid, task->nr_events);
 495         }
 496 }
 497
 498 static void add_cross_task_wakeups(struct perf_sched *sched)
 499 {
 500         struct task_desc *task1, *task2;
 501         unsigned long i, j;
 502
 503         for (i = 0; i < sched->nr_tasks; i++) {
 504                 task1 = sched->tasks[i];
 505                 j = i + 1;
 506                 if (j == sched->nr_tasks)
 507                         j = 0;
 508                 task2 = sched->tasks[j];
 509                 add_sched_event_wakeup(sched, task1, 0, task2);
 510         }
 511 }
 512
 513 static void perf_sched__process_event(struct perf_sched *sched,
 514                                       struct sched_atom *atom)
 515 {
 516         int ret = 0;
 517
 518         switch (atom->type) {
 519                 case SCHED_EVENT_RUN:
 520                         burn_nsecs(sched, atom->duration);
 521                         break;
 522                 case SCHED_EVENT_SLEEP:
 523                         if (atom->wait_sem)
 524                                 ret = sem_wait(atom->wait_sem);
 525                         BUG_ON(ret);
 526                         break;
 527                 case SCHED_EVENT_WAKEUP:
 528                         if (atom->wait_sem)
 529                                 ret = sem_post(atom->wait_sem);
 530                         BUG_ON(ret);
 531                         break;
 532                 default:
 533                         BUG_ON(1);
 534         }
 535 }
 536
 537 static u64 get_cpu_usage_nsec_parent(void)
 538 {
 539         struct rusage ru;
 540         u64 sum;
 541         int err;
 542
 543         err = getrusage(RUSAGE_SELF, &ru);
 544         BUG_ON(err);
 545
 546         sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 547         sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 548
 549         return sum;
 550 }
 551
 552 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 553 {
 554         struct perf_event_attr attr;
 555         char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 556         int fd;
 557         struct rlimit limit;
 558         bool need_privilege = false;
 559
 560         memset(&attr, 0, sizeof(attr));
 561
 562         attr.type = PERF_TYPE_SOFTWARE;
 563         attr.config = PERF_COUNT_SW_TASK_CLOCK;
 564
 565 force_again:
 566         fd = sys_perf_event_open(&attr, 0, -1, -1,
 567                                  perf_event_open_cloexec_flag());
 568
 569         if (fd < 0) {
 570                 if (errno == EMFILE) {
 571                         if (sched->force) {
 572                                 BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 573                                 limit.rlim_cur += sched->nr_tasks - cur_task;
 574                                 if (limit.rlim_cur > limit.rlim_max) {
 575                                         limit.rlim_max = limit.rlim_cur;
 576                                         need_privilege = true;
 577                                 }
 578                                 if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 579                                         if (need_privilege && errno == EPERM)
 580                                                 strcpy(info, "Need privilege\n");
 581                                 } else
 582                                         goto force_again;
 583                         } else
 584                                 strcpy(info, "Have a try with -f option\n");
 585                 }
 586                 pr_err("Error: sys_perf_event_open() syscall returned "
 587                        "with %d (%s)\n%s", fd,
 588                        str_error_r(errno, sbuf, sizeof(sbuf)), info);
 589                 exit(EXIT_FAILURE);
 590         }
 591         return fd;
 592 }
 593
 594 static u64 get_cpu_usage_nsec_self(int fd)
 595 {
 596         u64 runtime;
 597         int ret;
 598
 599         ret = read(fd, &runtime, sizeof(runtime));
 600         BUG_ON(ret != sizeof(runtime));
 601
 602         return runtime;
 603 }
 604
 605 struct sched_thread_parms {
 606         struct task_desc  *task;
 607         struct perf_sched *sched;
 608         int fd;
 609 };
 610
 611 static void *thread_func(void *ctx)
 612 {
 613         struct sched_thread_parms *parms = ctx;
 614         struct task_desc *this_task = parms->task;
 615         struct perf_sched *sched = parms->sched;
 616         u64 cpu_usage_0, cpu_usage_1;
 617         unsigned long i, ret;
 618         char comm2[22];
 619         int fd = parms->fd;
 620
 621         zfree(&parms);
 622
 623         sprintf(comm2, ":%s", this_task->comm);
 624         prctl(PR_SET_NAME, comm2);
 625         if (fd < 0)
 626                 return NULL;
 627
 628         while (!sched->thread_funcs_exit) {
 629                 ret = sem_post(&this_task->ready_for_work);
 630                 BUG_ON(ret);
 631                 mutex_lock(&sched->start_work_mutex);
 632                 mutex_unlock(&sched->start_work_mutex);
 633
 634                 cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 635
 636                 for (i = 0; i < this_task->nr_events; i++) {
 637                         this_task->curr_event = i;
 638                         perf_sched__process_event(sched, this_task->atoms[i]);
 639                 }
 640
 641                 cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 642                 this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 643                 ret = sem_post(&this_task->work_done_sem);
 644                 BUG_ON(ret);
 645
 646                 mutex_lock(&sched->work_done_wait_mutex);
 647                 mutex_unlock(&sched->work_done_wait_mutex);
 648         }
 649         return NULL;
 650 }
 651
 652 static void create_tasks(struct perf_sched *sched)
 653         EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
 654         EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
 655 {
 656         struct task_desc *task;
 657         pthread_attr_t attr;
 658         unsigned long i;
 659         int err;
 660
 661         err = pthread_attr_init(&attr);
 662         BUG_ON(err);
 663         err = pthread_attr_setstacksize(&attr,
 664                         (size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
 665         BUG_ON(err);
 666         mutex_lock(&sched->start_work_mutex);
 667         mutex_lock(&sched->work_done_wait_mutex);
 668         for (i = 0; i < sched->nr_tasks; i++) {
 669                 struct sched_thread_parms *parms = malloc(sizeof(*parms));
 670                 BUG_ON(parms == NULL);
 671                 parms->task = task = sched->tasks[i];
 672                 parms->sched = sched;
 673                 parms->fd = self_open_counters(sched, i);
 674                 sem_init(&task->ready_for_work, 0, 0);
 675                 sem_init(&task->work_done_sem, 0, 0);
 676                 task->curr_event = 0;
 677                 err = pthread_create(&task->thread, &attr, thread_func, parms);
 678                 BUG_ON(err);
 679         }
 680 }
 681
 682 static void destroy_tasks(struct perf_sched *sched)
 683         UNLOCK_FUNCTION(sched->start_work_mutex)
 684         UNLOCK_FUNCTION(sched->work_done_wait_mutex)
 685 {
 686         struct task_desc *task;
 687         unsigned long i;
 688         int err;
 689
 690         mutex_unlock(&sched->start_work_mutex);
 691         mutex_unlock(&sched->work_done_wait_mutex);
 692         /* Get rid of threads so they won't be upset by mutex destrunction */
 693         for (i = 0; i < sched->nr_tasks; i++) {
 694                 task = sched->tasks[i];
 695                 err = pthread_join(task->thread, NULL);
 696                 BUG_ON(err);
 697                 sem_destroy(&task->ready_for_work);
 698                 sem_destroy(&task->work_done_sem);
 699         }
 700 }
 701
 702 static void wait_for_tasks(struct perf_sched *sched)
 703         EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 704         EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 705 {
 706         u64 cpu_usage_0, cpu_usage_1;
 707         struct task_desc *task;
 708         unsigned long i, ret;
 709
 710         sched->start_time = get_nsecs();
 711         sched->cpu_usage = 0;
 712         mutex_unlock(&sched->work_done_wait_mutex);
 713
 714         for (i = 0; i < sched->nr_tasks; i++) {
 715                 task = sched->tasks[i];
 716                 ret = sem_wait(&task->ready_for_work);
 717                 BUG_ON(ret);
 718                 sem_init(&task->ready_for_work, 0, 0);
 719         }
 720         mutex_lock(&sched->work_done_wait_mutex);
 721
 722         cpu_usage_0 = get_cpu_usage_nsec_parent();
 723
 724         mutex_unlock(&sched->start_work_mutex);
 725
 726         for (i = 0; i < sched->nr_tasks; i++) {
 727                 task = sched->tasks[i];
 728                 ret = sem_wait(&task->work_done_sem);
 729                 BUG_ON(ret);
 730                 sem_init(&task->work_done_sem, 0, 0);
 731                 sched->cpu_usage += task->cpu_usage;
 732                 task->cpu_usage = 0;
 733         }
 734
 735         cpu_usage_1 = get_cpu_usage_nsec_parent();
 736         if (!sched->runavg_cpu_usage)
 737                 sched->runavg_cpu_usage = sched->cpu_usage;
 738         sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 739
 740         sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 741         if (!sched->runavg_parent_cpu_usage)
 742                 sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 743         sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 744                                          sched->parent_cpu_usage)/sched->replay_repeat;
 745
 746         mutex_lock(&sched->start_work_mutex);
 747
 748         for (i = 0; i < sched->nr_tasks; i++) {
 749                 task = sched->tasks[i];
 750                 task->curr_event = 0;
 751         }
 752 }
 753
 754 static void run_one_test(struct perf_sched *sched)
 755         EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 756         EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 757 {
 758         u64 T0, T1, delta, avg_delta, fluct;
 759
 760         T0 = get_nsecs();
 761         wait_for_tasks(sched);
 762         T1 = get_nsecs();
 763
 764         delta = T1 - T0;
 765         sched->sum_runtime += delta;
 766         sched->nr_runs++;
 767
 768         avg_delta = sched->sum_runtime / sched->nr_runs;
 769         if (delta < avg_delta)
 770                 fluct = avg_delta - delta;
 771         else
 772                 fluct = delta - avg_delta;
 773         sched->sum_fluct += fluct;
 774         if (!sched->run_avg)
 775                 sched->run_avg = delta;
 776         sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 777
 778         printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 779
 780         printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 781
 782         printf("cpu: %0.2f / %0.2f",
 783                 (double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 784
 785 #if 0
 786         /*
 787          * rusage statistics done by the parent, these are less
 788          * accurate than the sched->sum_exec_runtime based statistics:
 789          */
 790         printf(" [%0.2f / %0.2f]",
 791                 (double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 792                 (double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 793 #endif
 794
 795         printf("\n");
 796
 797         if (sched->nr_sleep_corrections)
 798                 printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 799         sched->nr_sleep_corrections = 0;
 800 }
 801
 802 static void test_calibrations(struct perf_sched *sched)
 803 {
 804         u64 T0, T1;
 805
 806         T0 = get_nsecs();
 807         burn_nsecs(sched, NSEC_PER_MSEC);
 808         T1 = get_nsecs();
 809
 810         printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 811
 812         T0 = get_nsecs();
 813         sleep_nsecs(NSEC_PER_MSEC);
 814         T1 = get_nsecs();
 815
 816         printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 817 }
 818
 819 static int
 820 replay_wakeup_event(struct perf_sched *sched,
 821                     struct evsel *evsel, struct perf_sample *sample,
 822                     struct machine *machine __maybe_unused)
 823 {
 824         const char *comm = evsel__strval(evsel, sample, "comm");
 825         const u32 pid    = evsel__intval(evsel, sample, "pid");
 826         struct task_desc *waker, *wakee;
 827
 828         if (verbose > 0) {
 829                 printf("sched_wakeup event %p\n", evsel);
 830
 831                 printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 832         }
 833
 834         waker = register_pid(sched, sample->tid, "<unknown>");
 835         wakee = register_pid(sched, pid, comm);
 836
 837         add_sched_event_wakeup(sched, waker, sample->time, wakee);
 838         return 0;
 839 }
 840
 841 static int replay_switch_event(struct perf_sched *sched,
 842                                struct evsel *evsel,
 843                                struct perf_sample *sample,
 844                                struct machine *machine __maybe_unused)
 845 {
 846         const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 847                    *next_comm  = evsel__strval(evsel, sample, "next_comm");
 848         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 849                   next_pid = evsel__intval(evsel, sample, "next_pid");
 850         struct task_desc *prev, __maybe_unused *next;
 851         u64 timestamp0, timestamp = sample->time;
 852         int cpu = sample->cpu;
 853         s64 delta;
 854
 855         if (verbose > 0)
 856                 printf("sched_switch event %p\n", evsel);
 857
 858         if (cpu >= MAX_CPUS || cpu < 0)
 859                 return 0;
 860
 861         timestamp0 = sched->cpu_last_switched[cpu];
 862         if (timestamp0)
 863                 delta = timestamp - timestamp0;
 864         else
 865                 delta = 0;
 866
 867         if (delta < 0) {
 868                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 869                 return -1;
 870         }
 871
 872         pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 873                  prev_comm, prev_pid, next_comm, next_pid, delta);
 874
 875         prev = register_pid(sched, prev_pid, prev_comm);
 876         next = register_pid(sched, next_pid, next_comm);
 877
 878         sched->cpu_last_switched[cpu] = timestamp;
 879
 880         add_sched_event_run(sched, prev, timestamp, delta);
 881         add_sched_event_sleep(sched, prev, timestamp);
 882
 883         return 0;
 884 }
 885
 886 static int replay_fork_event(struct perf_sched *sched,
 887                              union perf_event *event,
 888                              struct machine *machine)
 889 {
 890         struct thread *child, *parent;
 891
 892         child = machine__findnew_thread(machine, event->fork.pid,
 893                                         event->fork.tid);
 894         parent = machine__findnew_thread(machine, event->fork.ppid,
 895                                          event->fork.ptid);
 896
 897         if (child == NULL || parent == NULL) {
 898                 pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 899                                  child, parent);
 900                 goto out_put;
 901         }
 902
 903         if (verbose > 0) {
 904                 printf("fork event\n");
 905                 printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
 906                 printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
 907         }
 908
 909         register_pid(sched, thread__tid(parent), thread__comm_str(parent));
 910         register_pid(sched, thread__tid(child), thread__comm_str(child));
 911 out_put:
 912         thread__put(child);
 913         thread__put(parent);
 914         return 0;
 915 }
 916
 917 struct sort_dimension {
 918         const char              *name;
 919         sort_fn_t               cmp;
 920         struct list_head        list;
 921 };
 922
 923 static inline void init_prio(struct thread_runtime *r)
 924 {
 925         r->prio = -1;
 926 }
 927
 928 /*
 929  * handle runtime stats saved per thread
 930  */
 931 static struct thread_runtime *thread__init_runtime(struct thread *thread)
 932 {
 933         struct thread_runtime *r;
 934
 935         r = zalloc(sizeof(struct thread_runtime));
 936         if (!r)
 937                 return NULL;
 938
 939         init_stats(&r->run_stats);
 940         init_prio(r);
 941         thread__set_priv(thread, r);
 942
 943         return r;
 944 }
 945
 946 static struct thread_runtime *thread__get_runtime(struct thread *thread)
 947 {
 948         struct thread_runtime *tr;
 949
 950         tr = thread__priv(thread);
 951         if (tr == NULL) {
 952                 tr = thread__init_runtime(thread);
 953                 if (tr == NULL)
 954                         pr_debug("Failed to malloc memory for runtime data.\n");
 955         }
 956
 957         return tr;
 958 }
 959
 960 static int
 961 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 962 {
 963         struct sort_dimension *sort;
 964         int ret = 0;
 965
 966         BUG_ON(list_empty(list));
 967
 968         list_for_each_entry(sort, list, list) {
 969                 ret = sort->cmp(l, r);
 970                 if (ret)
 971                         return ret;
 972         }
 973
 974         return ret;
 975 }
 976
 977 static struct work_atoms *
 978 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 979                          struct list_head *sort_list)
 980 {
 981         struct rb_node *node = root->rb_root.rb_node;
 982         struct work_atoms key = { .thread = thread };
 983
 984         while (node) {
 985                 struct work_atoms *atoms;
 986                 int cmp;
 987
 988                 atoms = container_of(node, struct work_atoms, node);
 989
 990                 cmp = thread_lat_cmp(sort_list, &key, atoms);
 991                 if (cmp > 0)
 992                         node = node->rb_left;
 993                 else if (cmp < 0)
 994                         node = node->rb_right;
 995                 else {
 996                         BUG_ON(thread != atoms->thread);
 997                         return atoms;
 998                 }
 999         }
1000         return NULL;
1001 }
1002
1003 static void
1004 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1005                          struct list_head *sort_list)
1006 {
1007         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1008         bool leftmost = true;
1009
1010         while (*new) {
1011                 struct work_atoms *this;
1012                 int cmp;
1013
1014                 this = container_of(*new, struct work_atoms, node);
1015                 parent = *new;
1016
1017                 cmp = thread_lat_cmp(sort_list, data, this);
1018
1019                 if (cmp > 0)
1020                         new = &((*new)->rb_left);
1021                 else {
1022                         new = &((*new)->rb_right);
1023                         leftmost = false;
1024                 }
1025         }
1026
1027         rb_link_node(&data->node, parent, new);
1028         rb_insert_color_cached(&data->node, root, leftmost);
1029 }
1030
1031 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1032 {
1033         struct work_atoms *atoms = zalloc(sizeof(*atoms));
1034         if (!atoms) {
1035                 pr_err("No memory at %s\n", __func__);
1036                 return -1;
1037         }
1038
1039         atoms->thread = thread__get(thread);
1040         INIT_LIST_HEAD(&atoms->work_list);
1041         __thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1042         return 0;
1043 }
1044
1045 static int
1046 add_sched_out_event(struct work_atoms *atoms,
1047                     char run_state,
1048                     u64 timestamp)
1049 {
1050         struct work_atom *atom = zalloc(sizeof(*atom));
1051         if (!atom) {
1052                 pr_err("Non memory at %s", __func__);
1053                 return -1;
1054         }
1055
1056         atom->sched_out_time = timestamp;
1057
1058         if (run_state == 'R') {
1059                 atom->state = THREAD_WAIT_CPU;
1060                 atom->wake_up_time = atom->sched_out_time;
1061         }
1062
1063         list_add_tail(&atom->list, &atoms->work_list);
1064         return 0;
1065 }
1066
1067 static void
1068 add_runtime_event(struct work_atoms *atoms, u64 delta,
1069                   u64 timestamp __maybe_unused)
1070 {
1071         struct work_atom *atom;
1072
1073         BUG_ON(list_empty(&atoms->work_list));
1074
1075         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1076
1077         atom->runtime += delta;
1078         atoms->total_runtime += delta;
1079 }
1080
1081 static void
1082 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1083 {
1084         struct work_atom *atom;
1085         u64 delta;
1086
1087         if (list_empty(&atoms->work_list))
1088                 return;
1089
1090         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091
1092         if (atom->state != THREAD_WAIT_CPU)
1093                 return;
1094
1095         if (timestamp < atom->wake_up_time) {
1096                 atom->state = THREAD_IGNORE;
1097                 return;
1098         }
1099
1100         atom->state = THREAD_SCHED_IN;
1101         atom->sched_in_time = timestamp;
1102
1103         delta = atom->sched_in_time - atom->wake_up_time;
1104         atoms->total_lat += delta;
1105         if (delta > atoms->max_lat) {
1106                 atoms->max_lat = delta;
1107                 atoms->max_lat_start = atom->wake_up_time;
1108                 atoms->max_lat_end = timestamp;
1109         }
1110         atoms->nb_atoms++;
1111 }
1112
1113 static int latency_switch_event(struct perf_sched *sched,
1114                                 struct evsel *evsel,
1115                                 struct perf_sample *sample,
1116                                 struct machine *machine)
1117 {
1118         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1119                   next_pid = evsel__intval(evsel, sample, "next_pid");
1120         const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1121         struct work_atoms *out_events, *in_events;
1122         struct thread *sched_out, *sched_in;
1123         u64 timestamp0, timestamp = sample->time;
1124         int cpu = sample->cpu, err = -1;
1125         s64 delta;
1126
1127         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1128
1129         timestamp0 = sched->cpu_last_switched[cpu];
1130         sched->cpu_last_switched[cpu] = timestamp;
1131         if (timestamp0)
1132                 delta = timestamp - timestamp0;
1133         else
1134                 delta = 0;
1135
1136         if (delta < 0) {
1137                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1138                 return -1;
1139         }
1140
1141         sched_out = machine__findnew_thread(machine, -1, prev_pid);
1142         sched_in = machine__findnew_thread(machine, -1, next_pid);
1143         if (sched_out == NULL || sched_in == NULL)
1144                 goto out_put;
1145
1146         out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1147         if (!out_events) {
1148                 if (thread_atoms_insert(sched, sched_out))
1149                         goto out_put;
1150                 out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1151                 if (!out_events) {
1152                         pr_err("out-event: Internal tree error");
1153                         goto out_put;
1154                 }
1155         }
1156         if (add_sched_out_event(out_events, prev_state, timestamp))
1157                 return -1;
1158
1159         in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1160         if (!in_events) {
1161                 if (thread_atoms_insert(sched, sched_in))
1162                         goto out_put;
1163                 in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1164                 if (!in_events) {
1165                         pr_err("in-event: Internal tree error");
1166                         goto out_put;
1167                 }
1168                 /*
1169                  * Take came in we have not heard about yet,
1170                  * add in an initial atom in runnable state:
1171                  */
1172                 if (add_sched_out_event(in_events, 'R', timestamp))
1173                         goto out_put;
1174         }
1175         add_sched_in_event(in_events, timestamp);
1176         err = 0;
1177 out_put:
1178         thread__put(sched_out);
1179         thread__put(sched_in);
1180         return err;
1181 }
1182
1183 static int latency_runtime_event(struct perf_sched *sched,
1184                                  struct evsel *evsel,
1185                                  struct perf_sample *sample,
1186                                  struct machine *machine)
1187 {
1188         const u32 pid      = evsel__intval(evsel, sample, "pid");
1189         const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1190         struct thread *thread = machine__findnew_thread(machine, -1, pid);
1191         struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1192         u64 timestamp = sample->time;
1193         int cpu = sample->cpu, err = -1;
1194
1195         if (thread == NULL)
1196                 return -1;
1197
1198         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1199         if (!atoms) {
1200                 if (thread_atoms_insert(sched, thread))
1201                         goto out_put;
1202                 atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1203                 if (!atoms) {
1204                         pr_err("in-event: Internal tree error");
1205                         goto out_put;
1206                 }
1207                 if (add_sched_out_event(atoms, 'R', timestamp))
1208                         goto out_put;
1209         }
1210
1211         add_runtime_event(atoms, runtime, timestamp);
1212         err = 0;
1213 out_put:
1214         thread__put(thread);
1215         return err;
1216 }
1217
1218 static int latency_wakeup_event(struct perf_sched *sched,
1219                                 struct evsel *evsel,
1220                                 struct perf_sample *sample,
1221                                 struct machine *machine)
1222 {
1223         const u32 pid     = evsel__intval(evsel, sample, "pid");
1224         struct work_atoms *atoms;
1225         struct work_atom *atom;
1226         struct thread *wakee;
1227         u64 timestamp = sample->time;
1228         int err = -1;
1229
1230         wakee = machine__findnew_thread(machine, -1, pid);
1231         if (wakee == NULL)
1232                 return -1;
1233         atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1234         if (!atoms) {
1235                 if (thread_atoms_insert(sched, wakee))
1236                         goto out_put;
1237                 atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1238                 if (!atoms) {
1239                         pr_err("wakeup-event: Internal tree error");
1240                         goto out_put;
1241                 }
1242                 if (add_sched_out_event(atoms, 'S', timestamp))
1243                         goto out_put;
1244         }
1245
1246         BUG_ON(list_empty(&atoms->work_list));
1247
1248         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1249
1250         /*
1251          * As we do not guarantee the wakeup event happens when
1252          * task is out of run queue, also may happen when task is
1253          * on run queue and wakeup only change ->state to TASK_RUNNING,
1254          * then we should not set the ->wake_up_time when wake up a
1255          * task which is on run queue.
1256          *
1257          * You WILL be missing events if you've recorded only
1258          * one CPU, or are only looking at only one, so don't
1259          * skip in this case.
1260          */
1261         if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1262                 goto out_ok;
1263
1264         sched->nr_timestamps++;
1265         if (atom->sched_out_time > timestamp) {
1266                 sched->nr_unordered_timestamps++;
1267                 goto out_ok;
1268         }
1269
1270         atom->state = THREAD_WAIT_CPU;
1271         atom->wake_up_time = timestamp;
1272 out_ok:
1273         err = 0;
1274 out_put:
1275         thread__put(wakee);
1276         return err;
1277 }
1278
1279 static int latency_migrate_task_event(struct perf_sched *sched,
1280                                       struct evsel *evsel,
1281                                       struct perf_sample *sample,
1282                                       struct machine *machine)
1283 {
1284         const u32 pid = evsel__intval(evsel, sample, "pid");
1285         u64 timestamp = sample->time;
1286         struct work_atoms *atoms;
1287         struct work_atom *atom;
1288         struct thread *migrant;
1289         int err = -1;
1290
1291         /*
1292          * Only need to worry about migration when profiling one CPU.
1293          */
1294         if (sched->profile_cpu == -1)
1295                 return 0;
1296
1297         migrant = machine__findnew_thread(machine, -1, pid);
1298         if (migrant == NULL)
1299                 return -1;
1300         atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1301         if (!atoms) {
1302                 if (thread_atoms_insert(sched, migrant))
1303                         goto out_put;
1304                 register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1305                 atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1306                 if (!atoms) {
1307                         pr_err("migration-event: Internal tree error");
1308                         goto out_put;
1309                 }
1310                 if (add_sched_out_event(atoms, 'R', timestamp))
1311                         goto out_put;
1312         }
1313
1314         BUG_ON(list_empty(&atoms->work_list));
1315
1316         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1317         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1318
1319         sched->nr_timestamps++;
1320
1321         if (atom->sched_out_time > timestamp)
1322                 sched->nr_unordered_timestamps++;
1323         err = 0;
1324 out_put:
1325         thread__put(migrant);
1326         return err;
1327 }
1328
1329 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1330 {
1331         int i;
1332         int ret;
1333         u64 avg;
1334         char max_lat_start[32], max_lat_end[32];
1335
1336         if (!work_list->nb_atoms)
1337                 return;
1338         /*
1339          * Ignore idle threads:
1340          */
1341         if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1342                 return;
1343
1344         sched->all_runtime += work_list->total_runtime;
1345         sched->all_count   += work_list->nb_atoms;
1346
1347         if (work_list->num_merged > 1) {
1348                 ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1349                              work_list->num_merged);
1350         } else {
1351                 ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1352                              thread__tid(work_list->thread));
1353         }
1354
1355         for (i = 0; i < 24 - ret; i++)
1356                 printf(" ");
1357
1358         avg = work_list->total_lat / work_list->nb_atoms;
1359         timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1360         timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1361
1362         printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1363               (double)work_list->total_runtime / NSEC_PER_MSEC,
1364                  work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1365                  (double)work_list->max_lat / NSEC_PER_MSEC,
1366                  max_lat_start, max_lat_end);
1367 }
1368
1369 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1370 {
1371         pid_t l_tid, r_tid;
1372
1373         if (RC_CHK_EQUAL(l->thread, r->thread))
1374                 return 0;
1375         l_tid = thread__tid(l->thread);
1376         r_tid = thread__tid(r->thread);
1377         if (l_tid < r_tid)
1378                 return -1;
1379         if (l_tid > r_tid)
1380                 return 1;
1381         return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1382 }
1383
1384 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1385 {
1386         u64 avgl, avgr;
1387
1388         if (!l->nb_atoms)
1389                 return -1;
1390
1391         if (!r->nb_atoms)
1392                 return 1;
1393
1394         avgl = l->total_lat / l->nb_atoms;
1395         avgr = r->total_lat / r->nb_atoms;
1396
1397         if (avgl < avgr)
1398                 return -1;
1399         if (avgl > avgr)
1400                 return 1;
1401
1402         return 0;
1403 }
1404
1405 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1406 {
1407         if (l->max_lat < r->max_lat)
1408                 return -1;
1409         if (l->max_lat > r->max_lat)
1410                 return 1;
1411
1412         return 0;
1413 }
1414
1415 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1416 {
1417         if (l->nb_atoms < r->nb_atoms)
1418                 return -1;
1419         if (l->nb_atoms > r->nb_atoms)
1420                 return 1;
1421
1422         return 0;
1423 }
1424
1425 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1426 {
1427         if (l->total_runtime < r->total_runtime)
1428                 return -1;
1429         if (l->total_runtime > r->total_runtime)
1430                 return 1;
1431
1432         return 0;
1433 }
1434
1435 static int sort_dimension__add(const char *tok, struct list_head *list)
1436 {
1437         size_t i;
1438         static struct sort_dimension avg_sort_dimension = {
1439                 .name = "avg",
1440                 .cmp  = avg_cmp,
1441         };
1442         static struct sort_dimension max_sort_dimension = {
1443                 .name = "max",
1444                 .cmp  = max_cmp,
1445         };
1446         static struct sort_dimension pid_sort_dimension = {
1447                 .name = "pid",
1448                 .cmp  = pid_cmp,
1449         };
1450         static struct sort_dimension runtime_sort_dimension = {
1451                 .name = "runtime",
1452                 .cmp  = runtime_cmp,
1453         };
1454         static struct sort_dimension switch_sort_dimension = {
1455                 .name = "switch",
1456                 .cmp  = switch_cmp,
1457         };
1458         struct sort_dimension *available_sorts[] = {
1459                 &pid_sort_dimension,
1460                 &avg_sort_dimension,
1461                 &max_sort_dimension,
1462                 &switch_sort_dimension,
1463                 &runtime_sort_dimension,
1464         };
1465
1466         for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1467                 if (!strcmp(available_sorts[i]->name, tok)) {
1468                         list_add_tail(&available_sorts[i]->list, list);
1469
1470                         return 0;
1471                 }
1472         }
1473
1474         return -1;
1475 }
1476
1477 static void perf_sched__sort_lat(struct perf_sched *sched)
1478 {
1479         struct rb_node *node;
1480         struct rb_root_cached *root = &sched->atom_root;
1481 again:
1482         for (;;) {
1483                 struct work_atoms *data;
1484                 node = rb_first_cached(root);
1485                 if (!node)
1486                         break;
1487
1488                 rb_erase_cached(node, root);
1489                 data = rb_entry(node, struct work_atoms, node);
1490                 __thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1491         }
1492         if (root == &sched->atom_root) {
1493                 root = &sched->merged_atom_root;
1494                 goto again;
1495         }
1496 }
1497
1498 static int process_sched_wakeup_event(const struct perf_tool *tool,
1499                                       struct evsel *evsel,
1500                                       struct perf_sample *sample,
1501                                       struct machine *machine)
1502 {
1503         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1504
1505         if (sched->tp_handler->wakeup_event)
1506                 return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1507
1508         return 0;
1509 }
1510
1511 static int process_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
1512                                       struct evsel *evsel __maybe_unused,
1513                                       struct perf_sample *sample __maybe_unused,
1514                                       struct machine *machine __maybe_unused)
1515 {
1516         return 0;
1517 }
1518
1519 union map_priv {
1520         void    *ptr;
1521         bool     color;
1522 };
1523
1524 static bool thread__has_color(struct thread *thread)
1525 {
1526         union map_priv priv = {
1527                 .ptr = thread__priv(thread),
1528         };
1529
1530         return priv.color;
1531 }
1532
1533 static struct thread*
1534 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1535 {
1536         struct thread *thread = machine__findnew_thread(machine, pid, tid);
1537         union map_priv priv = {
1538                 .color = false,
1539         };
1540
1541         if (!sched->map.color_pids || !thread || thread__priv(thread))
1542                 return thread;
1543
1544         if (thread_map__has(sched->map.color_pids, tid))
1545                 priv.color = true;
1546
1547         thread__set_priv(thread, priv.ptr);
1548         return thread;
1549 }
1550
1551 static bool sched_match_task(struct perf_sched *sched, const char *comm_str)
1552 {
1553         bool fuzzy_match = sched->map.fuzzy;
1554         struct strlist *task_names = sched->map.task_names;
1555         struct str_node *node;
1556
1557         strlist__for_each_entry(node, task_names) {
1558                 bool match_found = fuzzy_match ? !!strstr(comm_str, node->s) :
1559                                                         !strcmp(comm_str, node->s);
1560                 if (match_found)
1561                         return true;
1562         }
1563
1564         return false;
1565 }
1566
1567 static void print_sched_map(struct perf_sched *sched, struct perf_cpu this_cpu, int cpus_nr,
1568                                                                 const char *color, bool sched_out)
1569 {
1570         for (int i = 0; i < cpus_nr; i++) {
1571                 struct perf_cpu cpu = {
1572                         .cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1573                 };
1574                 struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1575                 struct thread *curr_out_thread = sched->curr_out_thread[cpu.cpu];
1576                 struct thread_runtime *curr_tr;
1577                 const char *pid_color = color;
1578                 const char *cpu_color = color;
1579                 char symbol = ' ';
1580                 struct thread *thread_to_check = sched_out ? curr_out_thread : curr_thread;
1581
1582                 if (thread_to_check && thread__has_color(thread_to_check))
1583                         pid_color = COLOR_PIDS;
1584
1585                 if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1586                         cpu_color = COLOR_CPUS;
1587
1588                 if (cpu.cpu == this_cpu.cpu)
1589                         symbol = '*';
1590
1591                 color_fprintf(stdout, cpu.cpu != this_cpu.cpu ? color : cpu_color, "%c", symbol);
1592
1593                 thread_to_check = sched_out ? sched->curr_out_thread[cpu.cpu] :
1594                                                                 sched->curr_thread[cpu.cpu];
1595
1596                 if (thread_to_check) {
1597                         curr_tr = thread__get_runtime(thread_to_check);
1598                         if (curr_tr == NULL)
1599                                 return;
1600
1601                         if (sched_out) {
1602                                 if (cpu.cpu == this_cpu.cpu)
1603                                         color_fprintf(stdout, color, "-  ");
1604                                 else {
1605                                         curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1606                                         if (curr_tr != NULL)
1607                                                 color_fprintf(stdout, pid_color, "%2s ",
1608                                                                                 curr_tr->shortname);
1609                                 }
1610                         } else
1611                                 color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1612                 } else
1613                         color_fprintf(stdout, color, "   ");
1614         }
1615 }
1616
1617 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1618                             struct perf_sample *sample, struct machine *machine)
1619 {
1620         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1621         const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid");
1622         struct thread *sched_in, *sched_out;
1623         struct thread_runtime *tr;
1624         int new_shortname;
1625         u64 timestamp0, timestamp = sample->time;
1626         s64 delta;
1627         struct perf_cpu this_cpu = {
1628                 .cpu = sample->cpu,
1629         };
1630         int cpus_nr;
1631         int proceed;
1632         bool new_cpu = false;
1633         const char *color = PERF_COLOR_NORMAL;
1634         char stimestamp[32];
1635         const char *str;
1636
1637         BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1638
1639         if (this_cpu.cpu > sched->max_cpu.cpu)
1640                 sched->max_cpu = this_cpu;
1641
1642         if (sched->map.comp) {
1643                 cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1644                 if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1645                         sched->map.comp_cpus[cpus_nr++] = this_cpu;
1646                         new_cpu = true;
1647                 }
1648         } else
1649                 cpus_nr = sched->max_cpu.cpu;
1650
1651         timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1652         sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1653         if (timestamp0)
1654                 delta = timestamp - timestamp0;
1655         else
1656                 delta = 0;
1657
1658         if (delta < 0) {
1659                 pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1660                 return -1;
1661         }
1662
1663         sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1664         sched_out = map__findnew_thread(sched, machine, -1, prev_pid);
1665         if (sched_in == NULL || sched_out == NULL)
1666                 return -1;
1667
1668         tr = thread__get_runtime(sched_in);
1669         if (tr == NULL) {
1670                 thread__put(sched_in);
1671                 return -1;
1672         }
1673
1674         sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1675         sched->curr_out_thread[this_cpu.cpu] = thread__get(sched_out);
1676
1677         str = thread__comm_str(sched_in);
1678         new_shortname = 0;
1679         if (!tr->shortname[0]) {
1680                 if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1681                         /*
1682                          * Don't allocate a letter-number for swapper:0
1683                          * as a shortname. Instead, we use '.' for it.
1684                          */
1685                         tr->shortname[0] = '.';
1686                         tr->shortname[1] = ' ';
1687                 } else if (!sched->map.task_name || sched_match_task(sched, str)) {
1688                         tr->shortname[0] = sched->next_shortname1;
1689                         tr->shortname[1] = sched->next_shortname2;
1690
1691                         if (sched->next_shortname1 < 'Z') {
1692                                 sched->next_shortname1++;
1693                         } else {
1694                                 sched->next_shortname1 = 'A';
1695                                 if (sched->next_shortname2 < '9')
1696                                         sched->next_shortname2++;
1697                                 else
1698                                         sched->next_shortname2 = '0';
1699                         }
1700                 } else {
1701                         tr->shortname[0] = '-';
1702                         tr->shortname[1] = ' ';
1703                 }
1704                 new_shortname = 1;
1705         }
1706
1707         if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1708                 goto out;
1709
1710         proceed = 0;
1711         str = thread__comm_str(sched_in);
1712         /*
1713          * Check which of sched_in and sched_out matches the passed --task-name
1714          * arguments and call the corresponding print_sched_map.
1715          */
1716         if (sched->map.task_name && !sched_match_task(sched, str)) {
1717                 if (!sched_match_task(sched, thread__comm_str(sched_out)))
1718                         goto out;
1719                 else
1720                         goto sched_out;
1721
1722         } else {
1723                 str = thread__comm_str(sched_out);
1724                 if (!(sched->map.task_name && !sched_match_task(sched, str)))
1725                         proceed = 1;
1726         }
1727
1728         printf("  ");
1729
1730         print_sched_map(sched, this_cpu, cpus_nr, color, false);
1731
1732         timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1733         color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1734         if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1735                 const char *pid_color = color;
1736
1737                 if (thread__has_color(sched_in))
1738                         pid_color = COLOR_PIDS;
1739
1740                 color_fprintf(stdout, pid_color, "%s => %s:%d",
1741                         tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1742                 tr->comm_changed = false;
1743         }
1744
1745         if (sched->map.comp && new_cpu)
1746                 color_fprintf(stdout, color, " (CPU %d)", this_cpu.cpu);
1747
1748         if (proceed != 1) {
1749                 color_fprintf(stdout, color, "\n");
1750                 goto out;
1751         }
1752
1753 sched_out:
1754         if (sched->map.task_name) {
1755                 tr = thread__get_runtime(sched->curr_out_thread[this_cpu.cpu]);
1756                 if (strcmp(tr->shortname, "") == 0)
1757                         goto out;
1758
1759                 if (proceed == 1)
1760                         color_fprintf(stdout, color, "\n");
1761
1762                 printf("  ");
1763                 print_sched_map(sched, this_cpu, cpus_nr, color, true);
1764                 timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1765                 color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1766         }
1767
1768         color_fprintf(stdout, color, "\n");
1769
1770 out:
1771         if (sched->map.task_name)
1772                 thread__put(sched_out);
1773
1774         thread__put(sched_in);
1775
1776         return 0;
1777 }
1778
1779 static int process_sched_switch_event(const struct perf_tool *tool,
1780                                       struct evsel *evsel,
1781                                       struct perf_sample *sample,
1782                                       struct machine *machine)
1783 {
1784         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1785         int this_cpu = sample->cpu, err = 0;
1786         u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1787             next_pid = evsel__intval(evsel, sample, "next_pid");
1788
1789         if (sched->curr_pid[this_cpu] != (u32)-1) {
1790                 /*
1791                  * Are we trying to switch away a PID that is
1792                  * not current?
1793                  */
1794                 if (sched->curr_pid[this_cpu] != prev_pid)
1795                         sched->nr_context_switch_bugs++;
1796         }
1797
1798         if (sched->tp_handler->switch_event)
1799                 err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1800
1801         sched->curr_pid[this_cpu] = next_pid;
1802         return err;
1803 }
1804
1805 static int process_sched_runtime_event(const struct perf_tool *tool,
1806                                        struct evsel *evsel,
1807                                        struct perf_sample *sample,
1808                                        struct machine *machine)
1809 {
1810         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1811
1812         if (sched->tp_handler->runtime_event)
1813                 return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1814
1815         return 0;
1816 }
1817
1818 static int perf_sched__process_fork_event(const struct perf_tool *tool,
1819                                           union perf_event *event,
1820                                           struct perf_sample *sample,
1821                                           struct machine *machine)
1822 {
1823         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1824
1825         /* run the fork event through the perf machinery */
1826         perf_event__process_fork(tool, event, sample, machine);
1827
1828         /* and then run additional processing needed for this command */
1829         if (sched->tp_handler->fork_event)
1830                 return sched->tp_handler->fork_event(sched, event, machine);
1831
1832         return 0;
1833 }
1834
1835 static int process_sched_migrate_task_event(const struct perf_tool *tool,
1836                                             struct evsel *evsel,
1837                                             struct perf_sample *sample,
1838                                             struct machine *machine)
1839 {
1840         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1841
1842         if (sched->tp_handler->migrate_task_event)
1843                 return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1844
1845         return 0;
1846 }
1847
1848 typedef int (*tracepoint_handler)(const struct perf_tool *tool,
1849                                   struct evsel *evsel,
1850                                   struct perf_sample *sample,
1851                                   struct machine *machine);
1852
1853 static int perf_sched__process_tracepoint_sample(const struct perf_tool *tool __maybe_unused,
1854                                                  union perf_event *event __maybe_unused,
1855                                                  struct perf_sample *sample,
1856                                                  struct evsel *evsel,
1857                                                  struct machine *machine)
1858 {
1859         int err = 0;
1860
1861         if (evsel->handler != NULL) {
1862                 tracepoint_handler f = evsel->handler;
1863                 err = f(tool, evsel, sample, machine);
1864         }
1865
1866         return err;
1867 }
1868
1869 static int perf_sched__process_comm(const struct perf_tool *tool __maybe_unused,
1870                                     union perf_event *event,
1871                                     struct perf_sample *sample,
1872                                     struct machine *machine)
1873 {
1874         struct thread *thread;
1875         struct thread_runtime *tr;
1876         int err;
1877
1878         err = perf_event__process_comm(tool, event, sample, machine);
1879         if (err)
1880                 return err;
1881
1882         thread = machine__find_thread(machine, sample->pid, sample->tid);
1883         if (!thread) {
1884                 pr_err("Internal error: can't find thread\n");
1885                 return -1;
1886         }
1887
1888         tr = thread__get_runtime(thread);
1889         if (tr == NULL) {
1890                 thread__put(thread);
1891                 return -1;
1892         }
1893
1894         tr->comm_changed = true;
1895         thread__put(thread);
1896
1897         return 0;
1898 }
1899
1900 static int perf_sched__read_events(struct perf_sched *sched)
1901 {
1902         struct evsel_str_handler handlers[] = {
1903                 { "sched:sched_switch",       process_sched_switch_event, },
1904                 { "sched:sched_stat_runtime", process_sched_runtime_event, },
1905                 { "sched:sched_wakeup",       process_sched_wakeup_event, },
1906                 { "sched:sched_waking",       process_sched_wakeup_event, },
1907                 { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1908                 { "sched:sched_migrate_task", process_sched_migrate_task_event, },
1909         };
1910         struct perf_session *session;
1911         struct perf_data data = {
1912                 .path  = input_name,
1913                 .mode  = PERF_DATA_MODE_READ,
1914                 .force = sched->force,
1915         };
1916         int rc = -1;
1917
1918         session = perf_session__new(&data, &sched->tool);
1919         if (IS_ERR(session)) {
1920                 pr_debug("Error creating perf session");
1921                 return PTR_ERR(session);
1922         }
1923
1924         symbol__init(&session->header.env);
1925
1926         /* prefer sched_waking if it is captured */
1927         if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1928                 handlers[2].handler = process_sched_wakeup_ignore;
1929
1930         if (perf_session__set_tracepoints_handlers(session, handlers))
1931                 goto out_delete;
1932
1933         if (perf_session__has_traces(session, "record -R")) {
1934                 int err = perf_session__process_events(session);
1935                 if (err) {
1936                         pr_err("Failed to process events, error %d", err);
1937                         goto out_delete;
1938                 }
1939
1940                 sched->nr_events      = session->evlist->stats.nr_events[0];
1941                 sched->nr_lost_events = session->evlist->stats.total_lost;
1942                 sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1943         }
1944
1945         rc = 0;
1946 out_delete:
1947         perf_session__delete(session);
1948         return rc;
1949 }
1950
1951 /*
1952  * scheduling times are printed as msec.usec
1953  */
1954 static inline void print_sched_time(unsigned long long nsecs, int width)
1955 {
1956         unsigned long msecs;
1957         unsigned long usecs;
1958
1959         msecs  = nsecs / NSEC_PER_MSEC;
1960         nsecs -= msecs * NSEC_PER_MSEC;
1961         usecs  = nsecs / NSEC_PER_USEC;
1962         printf("%*lu.%03lu ", width, msecs, usecs);
1963 }
1964
1965 /*
1966  * returns runtime data for event, allocating memory for it the
1967  * first time it is used.
1968  */
1969 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1970 {
1971         struct evsel_runtime *r = evsel->priv;
1972
1973         if (r == NULL) {
1974                 r = zalloc(sizeof(struct evsel_runtime));
1975                 evsel->priv = r;
1976         }
1977
1978         return r;
1979 }
1980
1981 /*
1982  * save last time event was seen per cpu
1983  */
1984 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1985 {
1986         struct evsel_runtime *r = evsel__get_runtime(evsel);
1987
1988         if (r == NULL)
1989                 return;
1990
1991         if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1992                 int i, n = __roundup_pow_of_two(cpu+1);
1993                 void *p = r->last_time;
1994
1995                 p = realloc(r->last_time, n * sizeof(u64));
1996                 if (!p)
1997                         return;
1998
1999                 r->last_time = p;
2000                 for (i = r->ncpu; i < n; ++i)
2001                         r->last_time[i] = (u64) 0;
2002
2003                 r->ncpu = n;
2004         }
2005
2006         r->last_time[cpu] = timestamp;
2007 }
2008
2009 /* returns last time this event was seen on the given cpu */
2010 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
2011 {
2012         struct evsel_runtime *r = evsel__get_runtime(evsel);
2013
2014         if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
2015                 return 0;
2016
2017         return r->last_time[cpu];
2018 }
2019
2020 static int comm_width = 30;
2021
2022 static char *timehist_get_commstr(struct thread *thread)
2023 {
2024         static char str[32];
2025         const char *comm = thread__comm_str(thread);
2026         pid_t tid = thread__tid(thread);
2027         pid_t pid = thread__pid(thread);
2028         int n;
2029
2030         if (pid == 0)
2031                 n = scnprintf(str, sizeof(str), "%s", comm);
2032
2033         else if (tid != pid)
2034                 n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
2035
2036         else
2037                 n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
2038
2039         if (n > comm_width)
2040                 comm_width = n;
2041
2042         return str;
2043 }
2044
2045 /* prio field format: xxx or xxx->yyy */
2046 #define MAX_PRIO_STR_LEN 8
2047 static char *timehist_get_priostr(struct evsel *evsel,
2048                                   struct thread *thread,
2049                                   struct perf_sample *sample)
2050 {
2051         static char prio_str[16];
2052         int prev_prio = (int)evsel__intval(evsel, sample, "prev_prio");
2053         struct thread_runtime *tr = thread__priv(thread);
2054
2055         if (tr->prio != prev_prio && tr->prio != -1)
2056                 scnprintf(prio_str, sizeof(prio_str), "%d->%d", tr->prio, prev_prio);
2057         else
2058                 scnprintf(prio_str, sizeof(prio_str), "%d", prev_prio);
2059
2060         return prio_str;
2061 }
2062
2063 static void timehist_header(struct perf_sched *sched)
2064 {
2065         u32 ncpus = sched->max_cpu.cpu + 1;
2066         u32 i, j;
2067
2068         printf("%15s %6s ", "time", "cpu");
2069
2070         if (sched->show_cpu_visual) {
2071                 printf(" ");
2072                 for (i = 0, j = 0; i < ncpus; ++i) {
2073                         printf("%x", j++);
2074                         if (j > 15)
2075                                 j = 0;
2076                 }
2077                 printf(" ");
2078         }
2079
2080         printf(" %-*s", comm_width, "task name");
2081
2082         if (sched->show_prio)
2083                 printf("  %-*s", MAX_PRIO_STR_LEN, "prio");
2084
2085         printf("  %9s  %9s  %9s", "wait time", "sch delay", "run time");
2086
2087         if (sched->pre_migrations)
2088                 printf("  %9s", "pre-mig time");
2089
2090         if (sched->show_state)
2091                 printf("  %s", "state");
2092
2093         printf("\n");
2094
2095         /*
2096          * units row
2097          */
2098         printf("%15s %-6s ", "", "");
2099
2100         if (sched->show_cpu_visual)
2101                 printf(" %*s ", ncpus, "");
2102
2103         printf(" %-*s", comm_width, "[tid/pid]");
2104
2105         if (sched->show_prio)
2106                 printf("  %-*s", MAX_PRIO_STR_LEN, "");
2107
2108         printf("  %9s  %9s  %9s", "(msec)", "(msec)", "(msec)");
2109
2110         if (sched->pre_migrations)
2111                 printf("  %9s", "(msec)");
2112
2113         printf("\n");
2114
2115         /*
2116          * separator
2117          */
2118         printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2119
2120         if (sched->show_cpu_visual)
2121                 printf(" %.*s ", ncpus, graph_dotted_line);
2122
2123         printf(" %.*s", comm_width, graph_dotted_line);
2124
2125         if (sched->show_prio)
2126                 printf("  %.*s", MAX_PRIO_STR_LEN, graph_dotted_line);
2127
2128         printf("  %.9s  %.9s  %.9s", graph_dotted_line, graph_dotted_line, graph_dotted_line);
2129
2130         if (sched->pre_migrations)
2131                 printf("  %.9s", graph_dotted_line);
2132
2133         if (sched->show_state)
2134                 printf("  %.5s", graph_dotted_line);
2135
2136         printf("\n");
2137 }
2138
2139 static void timehist_print_sample(struct perf_sched *sched,
2140                                   struct evsel *evsel,
2141                                   struct perf_sample *sample,
2142                                   struct addr_location *al,
2143                                   struct thread *thread,
2144                                   u64 t, const char state)
2145 {
2146         struct thread_runtime *tr = thread__priv(thread);
2147         const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2148         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2149         u32 max_cpus = sched->max_cpu.cpu + 1;
2150         char tstr[64];
2151         char nstr[30];
2152         u64 wait_time;
2153
2154         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2155                 return;
2156
2157         timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2158         printf("%15s [%04d] ", tstr, sample->cpu);
2159
2160         if (sched->show_cpu_visual) {
2161                 u32 i;
2162                 char c;
2163
2164                 printf(" ");
2165                 for (i = 0; i < max_cpus; ++i) {
2166                         /* flag idle times with 'i'; others are sched events */
2167                         if (i == sample->cpu)
2168                                 c = (thread__tid(thread) == 0) ? 'i' : 's';
2169                         else
2170                                 c = ' ';
2171                         printf("%c", c);
2172                 }
2173                 printf(" ");
2174         }
2175
2176         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2177
2178         if (sched->show_prio)
2179                 printf(" %-*s ", MAX_PRIO_STR_LEN, timehist_get_priostr(evsel, thread, sample));
2180
2181         wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2182         print_sched_time(wait_time, 6);
2183
2184         print_sched_time(tr->dt_delay, 6);
2185         print_sched_time(tr->dt_run, 6);
2186         if (sched->pre_migrations)
2187                 print_sched_time(tr->dt_pre_mig, 6);
2188
2189         if (sched->show_state)
2190                 printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2191
2192         if (sched->show_next) {
2193                 snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2194                 printf(" %-*s", comm_width, nstr);
2195         }
2196
2197         if (sched->show_wakeups && !sched->show_next)
2198                 printf("  %-*s", comm_width, "");
2199
2200         if (thread__tid(thread) == 0)
2201                 goto out;
2202
2203         if (sched->show_callchain)
2204                 printf("  ");
2205
2206         sample__fprintf_sym(sample, al, 0,
2207                             EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2208                             EVSEL__PRINT_CALLCHAIN_ARROW |
2209                             EVSEL__PRINT_SKIP_IGNORED,
2210                             get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2211
2212 out:
2213         printf("\n");
2214 }
2215
2216 /*
2217  * Explanation of delta-time stats:
2218  *
2219  *            t = time of current schedule out event
2220  *        tprev = time of previous sched out event
2221  *                also time of schedule-in event for current task
2222  *    last_time = time of last sched change event for current task
2223  *                (i.e, time process was last scheduled out)
2224  * ready_to_run = time of wakeup for current task
2225  *     migrated = time of task migration to another CPU
2226  *
2227  * -----|-------------|-------------|-------------|-------------|-----
2228  *    last         ready         migrated       tprev           t
2229  *    time         to run
2230  *
2231  *      |---------------- dt_wait ----------------|
2232  *                   |--------- dt_delay ---------|-- dt_run --|
2233  *                   |- dt_pre_mig -|
2234  *
2235  *     dt_run = run time of current task
2236  *    dt_wait = time between last schedule out event for task and tprev
2237  *              represents time spent off the cpu
2238  *   dt_delay = time between wakeup and schedule-in of task
2239  * dt_pre_mig = time between wakeup and migration to another CPU
2240  */
2241
2242 static void timehist_update_runtime_stats(struct thread_runtime *r,
2243                                          u64 t, u64 tprev)
2244 {
2245         r->dt_delay   = 0;
2246         r->dt_sleep   = 0;
2247         r->dt_iowait  = 0;
2248         r->dt_preempt = 0;
2249         r->dt_run     = 0;
2250         r->dt_pre_mig = 0;
2251
2252         if (tprev) {
2253                 r->dt_run = t - tprev;
2254                 if (r->ready_to_run) {
2255                         if (r->ready_to_run > tprev)
2256                                 pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2257                         else
2258                                 r->dt_delay = tprev - r->ready_to_run;
2259
2260                         if ((r->migrated > r->ready_to_run) && (r->migrated < tprev))
2261                                 r->dt_pre_mig = r->migrated - r->ready_to_run;
2262                 }
2263
2264                 if (r->last_time > tprev)
2265                         pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2266                 else if (r->last_time) {
2267                         u64 dt_wait = tprev - r->last_time;
2268
2269                         if (r->last_state == 'R')
2270                                 r->dt_preempt = dt_wait;
2271                         else if (r->last_state == 'D')
2272                                 r->dt_iowait = dt_wait;
2273                         else
2274                                 r->dt_sleep = dt_wait;
2275                 }
2276         }
2277
2278         update_stats(&r->run_stats, r->dt_run);
2279
2280         r->total_run_time     += r->dt_run;
2281         r->total_delay_time   += r->dt_delay;
2282         r->total_sleep_time   += r->dt_sleep;
2283         r->total_iowait_time  += r->dt_iowait;
2284         r->total_preempt_time += r->dt_preempt;
2285         r->total_pre_mig_time += r->dt_pre_mig;
2286 }
2287
2288 static bool is_idle_sample(struct perf_sample *sample,
2289                            struct evsel *evsel)
2290 {
2291         /* pid 0 == swapper == idle task */
2292         if (evsel__name_is(evsel, "sched:sched_switch"))
2293                 return evsel__intval(evsel, sample, "prev_pid") == 0;
2294
2295         return sample->pid == 0;
2296 }
2297
2298 static void save_task_callchain(struct perf_sched *sched,
2299                                 struct perf_sample *sample,
2300                                 struct evsel *evsel,
2301                                 struct machine *machine)
2302 {
2303         struct callchain_cursor *cursor;
2304         struct thread *thread;
2305
2306         /* want main thread for process - has maps */
2307         thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2308         if (thread == NULL) {
2309                 pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2310                 return;
2311         }
2312
2313         if (!sched->show_callchain || sample->callchain == NULL)
2314                 return;
2315
2316         cursor = get_tls_callchain_cursor();
2317
2318         if (thread__resolve_callchain(thread, cursor, evsel, sample,
2319                                       NULL, NULL, sched->max_stack + 2) != 0) {
2320                 if (verbose > 0)
2321                         pr_err("Failed to resolve callchain. Skipping\n");
2322
2323                 return;
2324         }
2325
2326         callchain_cursor_commit(cursor);
2327
2328         while (true) {
2329                 struct callchain_cursor_node *node;
2330                 struct symbol *sym;
2331
2332                 node = callchain_cursor_current(cursor);
2333                 if (node == NULL)
2334                         break;
2335
2336                 sym = node->ms.sym;
2337                 if (sym) {
2338                         if (!strcmp(sym->name, "schedule") ||
2339                             !strcmp(sym->name, "__schedule") ||
2340                             !strcmp(sym->name, "preempt_schedule"))
2341                                 sym->ignore = 1;
2342                 }
2343
2344                 callchain_cursor_advance(cursor);
2345         }
2346 }
2347
2348 static int init_idle_thread(struct thread *thread)
2349 {
2350         struct idle_thread_runtime *itr;
2351
2352         thread__set_comm(thread, idle_comm, 0);
2353
2354         itr = zalloc(sizeof(*itr));
2355         if (itr == NULL)
2356                 return -ENOMEM;
2357
2358         init_prio(&itr->tr);
2359         init_stats(&itr->tr.run_stats);
2360         callchain_init(&itr->callchain);
2361         callchain_cursor_reset(&itr->cursor);
2362         thread__set_priv(thread, itr);
2363
2364         return 0;
2365 }
2366
2367 /*
2368  * Track idle stats per cpu by maintaining a local thread
2369  * struct for the idle task on each cpu.
2370  */
2371 static int init_idle_threads(int ncpu)
2372 {
2373         int i, ret;
2374
2375         idle_threads = zalloc(ncpu * sizeof(struct thread *));
2376         if (!idle_threads)
2377                 return -ENOMEM;
2378
2379         idle_max_cpu = ncpu;
2380
2381         /* allocate the actual thread struct if needed */
2382         for (i = 0; i < ncpu; ++i) {
2383                 idle_threads[i] = thread__new(0, 0);
2384                 if (idle_threads[i] == NULL)
2385                         return -ENOMEM;
2386
2387                 ret = init_idle_thread(idle_threads[i]);
2388                 if (ret < 0)
2389                         return ret;
2390         }
2391
2392         return 0;
2393 }
2394
2395 static void free_idle_threads(void)
2396 {
2397         int i;
2398
2399         if (idle_threads == NULL)
2400                 return;
2401
2402         for (i = 0; i < idle_max_cpu; ++i) {
2403                 if ((idle_threads[i]))
2404                         thread__delete(idle_threads[i]);
2405         }
2406
2407         free(idle_threads);
2408 }
2409
2410 static struct thread *get_idle_thread(int cpu)
2411 {
2412         /*
2413          * expand/allocate array of pointers to local thread
2414          * structs if needed
2415          */
2416         if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2417                 int i, j = __roundup_pow_of_two(cpu+1);
2418                 void *p;
2419
2420                 p = realloc(idle_threads, j * sizeof(struct thread *));
2421                 if (!p)
2422                         return NULL;
2423
2424                 idle_threads = (struct thread **) p;
2425                 for (i = idle_max_cpu; i < j; ++i)
2426                         idle_threads[i] = NULL;
2427
2428                 idle_max_cpu = j;
2429         }
2430
2431         /* allocate a new thread struct if needed */
2432         if (idle_threads[cpu] == NULL) {
2433                 idle_threads[cpu] = thread__new(0, 0);
2434
2435                 if (idle_threads[cpu]) {
2436                         if (init_idle_thread(idle_threads[cpu]) < 0)
2437                                 return NULL;
2438                 }
2439         }
2440
2441         return idle_threads[cpu];
2442 }
2443
2444 static void save_idle_callchain(struct perf_sched *sched,
2445                                 struct idle_thread_runtime *itr,
2446                                 struct perf_sample *sample)
2447 {
2448         struct callchain_cursor *cursor;
2449
2450         if (!sched->show_callchain || sample->callchain == NULL)
2451                 return;
2452
2453         cursor = get_tls_callchain_cursor();
2454         if (cursor == NULL)
2455                 return;
2456
2457         callchain_cursor__copy(&itr->cursor, cursor);
2458 }
2459
2460 static struct thread *timehist_get_thread(struct perf_sched *sched,
2461                                           struct perf_sample *sample,
2462                                           struct machine *machine,
2463                                           struct evsel *evsel)
2464 {
2465         struct thread *thread;
2466
2467         if (is_idle_sample(sample, evsel)) {
2468                 thread = get_idle_thread(sample->cpu);
2469                 if (thread == NULL)
2470                         pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2471
2472         } else {
2473                 /* there were samples with tid 0 but non-zero pid */
2474                 thread = machine__findnew_thread(machine, sample->pid,
2475                                                  sample->tid ?: sample->pid);
2476                 if (thread == NULL) {
2477                         pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2478                                  sample->tid);
2479                 }
2480
2481                 save_task_callchain(sched, sample, evsel, machine);
2482                 if (sched->idle_hist) {
2483                         struct thread *idle;
2484                         struct idle_thread_runtime *itr;
2485
2486                         idle = get_idle_thread(sample->cpu);
2487                         if (idle == NULL) {
2488                                 pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2489                                 return NULL;
2490                         }
2491
2492                         itr = thread__priv(idle);
2493                         if (itr == NULL)
2494                                 return NULL;
2495
2496                         itr->last_thread = thread;
2497
2498                         /* copy task callchain when entering to idle */
2499                         if (evsel__intval(evsel, sample, "next_pid") == 0)
2500                                 save_idle_callchain(sched, itr, sample);
2501                 }
2502         }
2503
2504         return thread;
2505 }
2506
2507 static bool timehist_skip_sample(struct perf_sched *sched,
2508                                  struct thread *thread,
2509                                  struct evsel *evsel,
2510                                  struct perf_sample *sample)
2511 {
2512         bool rc = false;
2513         int prio = -1;
2514         struct thread_runtime *tr = NULL;
2515
2516         if (thread__is_filtered(thread)) {
2517                 rc = true;
2518                 sched->skipped_samples++;
2519         }
2520
2521         if (sched->prio_str) {
2522                 /*
2523                  * Because priority may be changed during task execution,
2524                  * first read priority from prev sched_in event for current task.
2525                  * If prev sched_in event is not saved, then read priority from
2526                  * current task sched_out event.
2527                  */
2528                 tr = thread__get_runtime(thread);
2529                 if (tr && tr->prio != -1)
2530                         prio = tr->prio;
2531                 else if (evsel__name_is(evsel, "sched:sched_switch"))
2532                         prio = evsel__intval(evsel, sample, "prev_prio");
2533
2534                 if (prio != -1 && !test_bit(prio, sched->prio_bitmap)) {
2535                         rc = true;
2536                         sched->skipped_samples++;
2537                 }
2538         }
2539
2540         if (sched->idle_hist) {
2541                 if (!evsel__name_is(evsel, "sched:sched_switch"))
2542                         rc = true;
2543                 else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2544                          evsel__intval(evsel, sample, "next_pid") != 0)
2545                         rc = true;
2546         }
2547
2548         return rc;
2549 }
2550
2551 static void timehist_print_wakeup_event(struct perf_sched *sched,
2552                                         struct evsel *evsel,
2553                                         struct perf_sample *sample,
2554                                         struct machine *machine,
2555                                         struct thread *awakened)
2556 {
2557         struct thread *thread;
2558         char tstr[64];
2559
2560         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2561         if (thread == NULL)
2562                 return;
2563
2564         /* show wakeup unless both awakee and awaker are filtered */
2565         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2566             timehist_skip_sample(sched, awakened, evsel, sample)) {
2567                 return;
2568         }
2569
2570         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2571         printf("%15s [%04d] ", tstr, sample->cpu);
2572         if (sched->show_cpu_visual)
2573                 printf(" %*s ", sched->max_cpu.cpu + 1, "");
2574
2575         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2576
2577         /* dt spacer */
2578         printf("  %9s  %9s  %9s ", "", "", "");
2579
2580         printf("awakened: %s", timehist_get_commstr(awakened));
2581
2582         printf("\n");
2583 }
2584
2585 static int timehist_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
2586                                         union perf_event *event __maybe_unused,
2587                                         struct evsel *evsel __maybe_unused,
2588                                         struct perf_sample *sample __maybe_unused,
2589                                         struct machine *machine __maybe_unused)
2590 {
2591         return 0;
2592 }
2593
2594 static int timehist_sched_wakeup_event(const struct perf_tool *tool,
2595                                        union perf_event *event __maybe_unused,
2596                                        struct evsel *evsel,
2597                                        struct perf_sample *sample,
2598                                        struct machine *machine)
2599 {
2600         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2601         struct thread *thread;
2602         struct thread_runtime *tr = NULL;
2603         /* want pid of awakened task not pid in sample */
2604         const u32 pid = evsel__intval(evsel, sample, "pid");
2605
2606         thread = machine__findnew_thread(machine, 0, pid);
2607         if (thread == NULL)
2608                 return -1;
2609
2610         tr = thread__get_runtime(thread);
2611         if (tr == NULL)
2612                 return -1;
2613
2614         if (tr->ready_to_run == 0)
2615                 tr->ready_to_run = sample->time;
2616
2617         /* show wakeups if requested */
2618         if (sched->show_wakeups &&
2619             !perf_time__skip_sample(&sched->ptime, sample->time))
2620                 timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2621
2622         return 0;
2623 }
2624
2625 static void timehist_print_migration_event(struct perf_sched *sched,
2626                                         struct evsel *evsel,
2627                                         struct perf_sample *sample,
2628                                         struct machine *machine,
2629                                         struct thread *migrated)
2630 {
2631         struct thread *thread;
2632         char tstr[64];
2633         u32 max_cpus;
2634         u32 ocpu, dcpu;
2635
2636         if (sched->summary_only)
2637                 return;
2638
2639         max_cpus = sched->max_cpu.cpu + 1;
2640         ocpu = evsel__intval(evsel, sample, "orig_cpu");
2641         dcpu = evsel__intval(evsel, sample, "dest_cpu");
2642
2643         thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2644         if (thread == NULL)
2645                 return;
2646
2647         if (timehist_skip_sample(sched, thread, evsel, sample) &&
2648             timehist_skip_sample(sched, migrated, evsel, sample)) {
2649                 return;
2650         }
2651
2652         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2653         printf("%15s [%04d] ", tstr, sample->cpu);
2654
2655         if (sched->show_cpu_visual) {
2656                 u32 i;
2657                 char c;
2658
2659                 printf("  ");
2660                 for (i = 0; i < max_cpus; ++i) {
2661                         c = (i == sample->cpu) ? 'm' : ' ';
2662                         printf("%c", c);
2663                 }
2664                 printf("  ");
2665         }
2666
2667         printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2668
2669         /* dt spacer */
2670         printf("  %9s  %9s  %9s ", "", "", "");
2671
2672         printf("migrated: %s", timehist_get_commstr(migrated));
2673         printf(" cpu %d => %d", ocpu, dcpu);
2674
2675         printf("\n");
2676 }
2677
2678 static int timehist_migrate_task_event(const struct perf_tool *tool,
2679                                        union perf_event *event __maybe_unused,
2680                                        struct evsel *evsel,
2681                                        struct perf_sample *sample,
2682                                        struct machine *machine)
2683 {
2684         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2685         struct thread *thread;
2686         struct thread_runtime *tr = NULL;
2687         /* want pid of migrated task not pid in sample */
2688         const u32 pid = evsel__intval(evsel, sample, "pid");
2689
2690         thread = machine__findnew_thread(machine, 0, pid);
2691         if (thread == NULL)
2692                 return -1;
2693
2694         tr = thread__get_runtime(thread);
2695         if (tr == NULL)
2696                 return -1;
2697
2698         tr->migrations++;
2699         tr->migrated = sample->time;
2700
2701         /* show migrations if requested */
2702         if (sched->show_migrations) {
2703                 timehist_print_migration_event(sched, evsel, sample,
2704                                                         machine, thread);
2705         }
2706
2707         return 0;
2708 }
2709
2710 static void timehist_update_task_prio(struct evsel *evsel,
2711                                       struct perf_sample *sample,
2712                                       struct machine *machine)
2713 {
2714         struct thread *thread;
2715         struct thread_runtime *tr = NULL;
2716         const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2717         const u32 next_prio = evsel__intval(evsel, sample, "next_prio");
2718
2719         if (next_pid == 0)
2720                 thread = get_idle_thread(sample->cpu);
2721         else
2722                 thread = machine__findnew_thread(machine, -1, next_pid);
2723
2724         if (thread == NULL)
2725                 return;
2726
2727         tr = thread__get_runtime(thread);
2728         if (tr == NULL)
2729                 return;
2730
2731         tr->prio = next_prio;
2732 }
2733
2734 static int timehist_sched_change_event(const struct perf_tool *tool,
2735                                        union perf_event *event,
2736                                        struct evsel *evsel,
2737                                        struct perf_sample *sample,
2738                                        struct machine *machine)
2739 {
2740         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2741         struct perf_time_interval *ptime = &sched->ptime;
2742         struct addr_location al;
2743         struct thread *thread;
2744         struct thread_runtime *tr = NULL;
2745         u64 tprev, t = sample->time;
2746         int rc = 0;
2747         const char state = evsel__taskstate(evsel, sample, "prev_state");
2748
2749         addr_location__init(&al);
2750         if (machine__resolve(machine, &al, sample) < 0) {
2751                 pr_err("problem processing %d event. skipping it\n",
2752                        event->header.type);
2753                 rc = -1;
2754                 goto out;
2755         }
2756
2757         if (sched->show_prio || sched->prio_str)
2758                 timehist_update_task_prio(evsel, sample, machine);
2759
2760         thread = timehist_get_thread(sched, sample, machine, evsel);
2761         if (thread == NULL) {
2762                 rc = -1;
2763                 goto out;
2764         }
2765
2766         if (timehist_skip_sample(sched, thread, evsel, sample))
2767                 goto out;
2768
2769         tr = thread__get_runtime(thread);
2770         if (tr == NULL) {
2771                 rc = -1;
2772                 goto out;
2773         }
2774
2775         tprev = evsel__get_time(evsel, sample->cpu);
2776
2777         /*
2778          * If start time given:
2779          * - sample time is under window user cares about - skip sample
2780          * - tprev is under window user cares about  - reset to start of window
2781          */
2782         if (ptime->start && ptime->start > t)
2783                 goto out;
2784
2785         if (tprev && ptime->start > tprev)
2786                 tprev = ptime->start;
2787
2788         /*
2789          * If end time given:
2790          * - previous sched event is out of window - we are done
2791          * - sample time is beyond window user cares about - reset it
2792          *   to close out stats for time window interest
2793          * - If tprev is 0, that is, sched_in event for current task is
2794          *   not recorded, cannot determine whether sched_in event is
2795          *   within time window interest - ignore it
2796          */
2797         if (ptime->end) {
2798                 if (!tprev || tprev > ptime->end)
2799                         goto out;
2800
2801                 if (t > ptime->end)
2802                         t = ptime->end;
2803         }
2804
2805         if (!sched->idle_hist || thread__tid(thread) == 0) {
2806                 if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2807                         timehist_update_runtime_stats(tr, t, tprev);
2808
2809                 if (sched->idle_hist) {
2810                         struct idle_thread_runtime *itr = (void *)tr;
2811                         struct thread_runtime *last_tr;
2812
2813                         if (itr->last_thread == NULL)
2814                                 goto out;
2815
2816                         /* add current idle time as last thread's runtime */
2817                         last_tr = thread__get_runtime(itr->last_thread);
2818                         if (last_tr == NULL)
2819                                 goto out;
2820
2821                         timehist_update_runtime_stats(last_tr, t, tprev);
2822                         /*
2823                          * remove delta time of last thread as it's not updated
2824                          * and otherwise it will show an invalid value next
2825                          * time.  we only care total run time and run stat.
2826                          */
2827                         last_tr->dt_run = 0;
2828                         last_tr->dt_delay = 0;
2829                         last_tr->dt_sleep = 0;
2830                         last_tr->dt_iowait = 0;
2831                         last_tr->dt_preempt = 0;
2832
2833                         if (itr->cursor.nr)
2834                                 callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2835
2836                         itr->last_thread = NULL;
2837                 }
2838
2839                 if (!sched->summary_only)
2840                         timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2841         }
2842
2843 out:
2844         if (sched->hist_time.start == 0 && t >= ptime->start)
2845                 sched->hist_time.start = t;
2846         if (ptime->end == 0 || t <= ptime->end)
2847                 sched->hist_time.end = t;
2848
2849         if (tr) {
2850                 /* time of this sched_switch event becomes last time task seen */
2851                 tr->last_time = sample->time;
2852
2853                 /* last state is used to determine where to account wait time */
2854                 tr->last_state = state;
2855
2856                 /* sched out event for task so reset ready to run time and migrated time */
2857                 if (state == 'R')
2858                         tr->ready_to_run = t;
2859                 else
2860                         tr->ready_to_run = 0;
2861
2862                 tr->migrated = 0;
2863         }
2864
2865         evsel__save_time(evsel, sample->time, sample->cpu);
2866
2867         addr_location__exit(&al);
2868         return rc;
2869 }
2870
2871 static int timehist_sched_switch_event(const struct perf_tool *tool,
2872                              union perf_event *event,
2873                              struct evsel *evsel,
2874                              struct perf_sample *sample,
2875                              struct machine *machine __maybe_unused)
2876 {
2877         return timehist_sched_change_event(tool, event, evsel, sample, machine);
2878 }
2879
2880 static int process_lost(const struct perf_tool *tool __maybe_unused,
2881                         union perf_event *event,
2882                         struct perf_sample *sample,
2883                         struct machine *machine __maybe_unused)
2884 {
2885         char tstr[64];
2886
2887         timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2888         printf("%15s ", tstr);
2889         printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2890
2891         return 0;
2892 }
2893
2894
2895 static void print_thread_runtime(struct thread *t,
2896                                  struct thread_runtime *r)
2897 {
2898         double mean = avg_stats(&r->run_stats);
2899         float stddev;
2900
2901         printf("%*s   %5d  %9" PRIu64 " ",
2902                comm_width, timehist_get_commstr(t), thread__ppid(t),
2903                (u64) r->run_stats.n);
2904
2905         print_sched_time(r->total_run_time, 8);
2906         stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2907         print_sched_time(r->run_stats.min, 6);
2908         printf(" ");
2909         print_sched_time((u64) mean, 6);
2910         printf(" ");
2911         print_sched_time(r->run_stats.max, 6);
2912         printf("  ");
2913         printf("%5.2f", stddev);
2914         printf("   %5" PRIu64, r->migrations);
2915         printf("\n");
2916 }
2917
2918 static void print_thread_waittime(struct thread *t,
2919                                   struct thread_runtime *r)
2920 {
2921         printf("%*s   %5d  %9" PRIu64 " ",
2922                comm_width, timehist_get_commstr(t), thread__ppid(t),
2923                (u64) r->run_stats.n);
2924
2925         print_sched_time(r->total_run_time, 8);
2926         print_sched_time(r->total_sleep_time, 6);
2927         printf(" ");
2928         print_sched_time(r->total_iowait_time, 6);
2929         printf(" ");
2930         print_sched_time(r->total_preempt_time, 6);
2931         printf(" ");
2932         print_sched_time(r->total_delay_time, 6);
2933         printf("\n");
2934 }
2935
2936 struct total_run_stats {
2937         struct perf_sched *sched;
2938         u64  sched_count;
2939         u64  task_count;
2940         u64  total_run_time;
2941 };
2942
2943 static int show_thread_runtime(struct thread *t, void *priv)
2944 {
2945         struct total_run_stats *stats = priv;
2946         struct thread_runtime *r;
2947
2948         if (thread__is_filtered(t))
2949                 return 0;
2950
2951         r = thread__priv(t);
2952         if (r && r->run_stats.n) {
2953                 stats->task_count++;
2954                 stats->sched_count += r->run_stats.n;
2955                 stats->total_run_time += r->total_run_time;
2956
2957                 if (stats->sched->show_state)
2958                         print_thread_waittime(t, r);
2959                 else
2960                         print_thread_runtime(t, r);
2961         }
2962
2963         return 0;
2964 }
2965
2966 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2967 {
2968         const char *sep = " <- ";
2969         struct callchain_list *chain;
2970         size_t ret = 0;
2971         char bf[1024];
2972         bool first;
2973
2974         if (node == NULL)
2975                 return 0;
2976
2977         ret = callchain__fprintf_folded(fp, node->parent);
2978         first = (ret == 0);
2979
2980         list_for_each_entry(chain, &node->val, list) {
2981                 if (chain->ip >= PERF_CONTEXT_MAX)
2982                         continue;
2983                 if (chain->ms.sym && chain->ms.sym->ignore)
2984                         continue;
2985                 ret += fprintf(fp, "%s%s", first ? "" : sep,
2986                                callchain_list__sym_name(chain, bf, sizeof(bf),
2987                                                         false));
2988                 first = false;
2989         }
2990
2991         return ret;
2992 }
2993
2994 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2995 {
2996         size_t ret = 0;
2997         FILE *fp = stdout;
2998         struct callchain_node *chain;
2999         struct rb_node *rb_node = rb_first_cached(root);
3000
3001         printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
3002         printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
3003                graph_dotted_line);
3004
3005         while (rb_node) {
3006                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
3007                 rb_node = rb_next(rb_node);
3008
3009                 ret += fprintf(fp, "  ");
3010                 print_sched_time(chain->hit, 12);
3011                 ret += 16;  /* print_sched_time returns 2nd arg + 4 */
3012                 ret += fprintf(fp, " %8d  ", chain->count);
3013                 ret += callchain__fprintf_folded(fp, chain);
3014                 ret += fprintf(fp, "\n");
3015         }
3016
3017         return ret;
3018 }
3019
3020 static void timehist_print_summary(struct perf_sched *sched,
3021                                    struct perf_session *session)
3022 {
3023         struct machine *m = &session->machines.host;
3024         struct total_run_stats totals;
3025         u64 task_count;
3026         struct thread *t;
3027         struct thread_runtime *r;
3028         int i;
3029         u64 hist_time = sched->hist_time.end - sched->hist_time.start;
3030
3031         memset(&totals, 0, sizeof(totals));
3032         totals.sched = sched;
3033
3034         if (sched->idle_hist) {
3035                 printf("\nIdle-time summary\n");
3036                 printf("%*s  parent  sched-out  ", comm_width, "comm");
3037                 printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
3038         } else if (sched->show_state) {
3039                 printf("\nWait-time summary\n");
3040                 printf("%*s  parent   sched-in  ", comm_width, "comm");
3041                 printf("   run-time      sleep      iowait     preempt       delay\n");
3042         } else {
3043                 printf("\nRuntime summary\n");
3044                 printf("%*s  parent   sched-in  ", comm_width, "comm");
3045                 printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
3046         }
3047         printf("%*s            (count)  ", comm_width, "");
3048         printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
3049                sched->show_state ? "(msec)" : "%");
3050         printf("%.117s\n", graph_dotted_line);
3051
3052         machine__for_each_thread(m, show_thread_runtime, &totals);
3053         task_count = totals.task_count;
3054         if (!task_count)
3055                 printf("<no still running tasks>\n");
3056
3057         /* CPU idle stats not tracked when samples were skipped */
3058         if (sched->skipped_samples && !sched->idle_hist)
3059                 return;
3060
3061         printf("\nIdle stats:\n");
3062         for (i = 0; i < idle_max_cpu; ++i) {
3063                 if (cpu_list && !test_bit(i, cpu_bitmap))
3064                         continue;
3065
3066                 t = idle_threads[i];
3067                 if (!t)
3068                         continue;
3069
3070                 r = thread__priv(t);
3071                 if (r && r->run_stats.n) {
3072                         totals.sched_count += r->run_stats.n;
3073                         printf("    CPU %2d idle for ", i);
3074                         print_sched_time(r->total_run_time, 6);
3075                         printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
3076                 } else
3077                         printf("    CPU %2d idle entire time window\n", i);
3078         }
3079
3080         if (sched->idle_hist && sched->show_callchain) {
3081                 callchain_param.mode  = CHAIN_FOLDED;
3082                 callchain_param.value = CCVAL_PERIOD;
3083
3084                 callchain_register_param(&callchain_param);
3085
3086                 printf("\nIdle stats by callchain:\n");
3087                 for (i = 0; i < idle_max_cpu; ++i) {
3088                         struct idle_thread_runtime *itr;
3089
3090                         t = idle_threads[i];
3091                         if (!t)
3092                                 continue;
3093
3094                         itr = thread__priv(t);
3095                         if (itr == NULL)
3096                                 continue;
3097
3098                         callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
3099                                              0, &callchain_param);
3100
3101                         printf("  CPU %2d:", i);
3102                         print_sched_time(itr->tr.total_run_time, 6);
3103                         printf(" msec\n");
3104                         timehist_print_idlehist_callchain(&itr->sorted_root);
3105                         printf("\n");
3106                 }
3107         }
3108
3109         printf("\n"
3110                "    Total number of unique tasks: %" PRIu64 "\n"
3111                "Total number of context switches: %" PRIu64 "\n",
3112                totals.task_count, totals.sched_count);
3113
3114         printf("           Total run time (msec): ");
3115         print_sched_time(totals.total_run_time, 2);
3116         printf("\n");
3117
3118         printf("    Total scheduling time (msec): ");
3119         print_sched_time(hist_time, 2);
3120         printf(" (x %d)\n", sched->max_cpu.cpu);
3121 }
3122
3123 typedef int (*sched_handler)(const struct perf_tool *tool,
3124                           union perf_event *event,
3125                           struct evsel *evsel,
3126                           struct perf_sample *sample,
3127                           struct machine *machine);
3128
3129 static int perf_timehist__process_sample(const struct perf_tool *tool,
3130                                          union perf_event *event,
3131                                          struct perf_sample *sample,
3132                                          struct evsel *evsel,
3133                                          struct machine *machine)
3134 {
3135         struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
3136         int err = 0;
3137         struct perf_cpu this_cpu = {
3138                 .cpu = sample->cpu,
3139         };
3140
3141         if (this_cpu.cpu > sched->max_cpu.cpu)
3142                 sched->max_cpu = this_cpu;
3143
3144         if (evsel->handler != NULL) {
3145                 sched_handler f = evsel->handler;
3146
3147                 err = f(tool, event, evsel, sample, machine);
3148         }
3149
3150         return err;
3151 }
3152
3153 static int timehist_check_attr(struct perf_sched *sched,
3154                                struct evlist *evlist)
3155 {
3156         struct evsel *evsel;
3157         struct evsel_runtime *er;
3158
3159         list_for_each_entry(evsel, &evlist->core.entries, core.node) {
3160                 er = evsel__get_runtime(evsel);
3161                 if (er == NULL) {
3162                         pr_err("Failed to allocate memory for evsel runtime data\n");
3163                         return -1;
3164                 }
3165
3166                 /* only need to save callchain related to sched_switch event */
3167                 if (sched->show_callchain &&
3168                     evsel__name_is(evsel, "sched:sched_switch") &&
3169                     !evsel__has_callchain(evsel)) {
3170                         pr_info("Samples of sched_switch event do not have callchains.\n");
3171                         sched->show_callchain = 0;
3172                         symbol_conf.use_callchain = 0;
3173                 }
3174         }
3175
3176         return 0;
3177 }
3178
3179 static int timehist_parse_prio_str(struct perf_sched *sched)
3180 {
3181         char *p;
3182         unsigned long start_prio, end_prio;
3183         const char *str = sched->prio_str;
3184
3185         if (!str)
3186                 return 0;
3187
3188         while (isdigit(*str)) {
3189                 p = NULL;
3190                 start_prio = strtoul(str, &p, 0);
3191                 if (start_prio >= MAX_PRIO || (*p != '\0' && *p != ',' && *p != '-'))
3192                         return -1;
3193
3194                 if (*p == '-') {
3195                         str = ++p;
3196                         p = NULL;
3197                         end_prio = strtoul(str, &p, 0);
3198
3199                         if (end_prio >= MAX_PRIO || (*p != '\0' && *p != ','))
3200                                 return -1;
3201
3202                         if (end_prio < start_prio)
3203                                 return -1;
3204                 } else {
3205                         end_prio = start_prio;
3206                 }
3207
3208                 for (; start_prio <= end_prio; start_prio++)
3209                         __set_bit(start_prio, sched->prio_bitmap);
3210
3211                 if (*p)
3212                         ++p;
3213
3214                 str = p;
3215         }
3216
3217         return 0;
3218 }
3219
3220 static int perf_sched__timehist(struct perf_sched *sched)
3221 {
3222         struct evsel_str_handler handlers[] = {
3223                 { "sched:sched_switch",       timehist_sched_switch_event, },
3224                 { "sched:sched_wakeup",       timehist_sched_wakeup_event, },
3225                 { "sched:sched_waking",       timehist_sched_wakeup_event, },
3226                 { "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
3227         };
3228         const struct evsel_str_handler migrate_handlers[] = {
3229                 { "sched:sched_migrate_task", timehist_migrate_task_event, },
3230         };
3231         struct perf_data data = {
3232                 .path  = input_name,
3233                 .mode  = PERF_DATA_MODE_READ,
3234                 .force = sched->force,
3235         };
3236
3237         struct perf_session *session;
3238         struct evlist *evlist;
3239         int err = -1;
3240
3241         /*
3242          * event handlers for timehist option
3243          */
3244         sched->tool.sample       = perf_timehist__process_sample;
3245         sched->tool.mmap         = perf_event__process_mmap;
3246         sched->tool.comm         = perf_event__process_comm;
3247         sched->tool.exit         = perf_event__process_exit;
3248         sched->tool.fork         = perf_event__process_fork;
3249         sched->tool.lost         = process_lost;
3250         sched->tool.attr         = perf_event__process_attr;
3251         sched->tool.tracing_data = perf_event__process_tracing_data;
3252         sched->tool.build_id     = perf_event__process_build_id;
3253
3254         sched->tool.ordering_requires_timestamps = true;
3255
3256         symbol_conf.use_callchain = sched->show_callchain;
3257
3258         session = perf_session__new(&data, &sched->tool);
3259         if (IS_ERR(session))
3260                 return PTR_ERR(session);
3261
3262         if (cpu_list) {
3263                 err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3264                 if (err < 0)
3265                         goto out;
3266         }
3267
3268         evlist = session->evlist;
3269
3270         symbol__init(&session->header.env);
3271
3272         if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3273                 pr_err("Invalid time string\n");
3274                 err = -EINVAL;
3275                 goto out;
3276         }
3277
3278         if (timehist_check_attr(sched, evlist) != 0)
3279                 goto out;
3280
3281         if (timehist_parse_prio_str(sched) != 0) {
3282                 pr_err("Invalid prio string\n");
3283                 goto out;
3284         }
3285
3286         setup_pager();
3287
3288         /* prefer sched_waking if it is captured */
3289         if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3290                 handlers[1].handler = timehist_sched_wakeup_ignore;
3291
3292         /* setup per-evsel handlers */
3293         if (perf_session__set_tracepoints_handlers(session, handlers))
3294                 goto out;
3295
3296         /* sched_switch event at a minimum needs to exist */
3297         if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3298                 pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3299                 goto out;
3300         }
3301
3302         if ((sched->show_migrations || sched->pre_migrations) &&
3303                 perf_session__set_tracepoints_handlers(session, migrate_handlers))
3304                 goto out;
3305
3306         /* pre-allocate struct for per-CPU idle stats */
3307         sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3308         if (sched->max_cpu.cpu == 0)
3309                 sched->max_cpu.cpu = 4;
3310         if (init_idle_threads(sched->max_cpu.cpu))
3311                 goto out;
3312
3313         /* summary_only implies summary option, but don't overwrite summary if set */
3314         if (sched->summary_only)
3315                 sched->summary = sched->summary_only;
3316
3317         if (!sched->summary_only)
3318                 timehist_header(sched);
3319
3320         err = perf_session__process_events(session);
3321         if (err) {
3322                 pr_err("Failed to process events, error %d", err);
3323                 goto out;
3324         }
3325
3326         sched->nr_events      = evlist->stats.nr_events[0];
3327         sched->nr_lost_events = evlist->stats.total_lost;
3328         sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3329
3330         if (sched->summary)
3331                 timehist_print_summary(sched, session);
3332
3333 out:
3334         free_idle_threads();
3335         perf_session__delete(session);
3336
3337         return err;
3338 }
3339
3340
3341 static void print_bad_events(struct perf_sched *sched)
3342 {
3343         if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3344                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3345                         (double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3346                         sched->nr_unordered_timestamps, sched->nr_timestamps);
3347         }
3348         if (sched->nr_lost_events && sched->nr_events) {
3349                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3350                         (double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3351                         sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3352         }
3353         if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3354                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3355                         (double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3356                         sched->nr_context_switch_bugs, sched->nr_timestamps);
3357                 if (sched->nr_lost_events)
3358                         printf(" (due to lost events?)");
3359                 printf("\n");
3360         }
3361 }
3362
3363 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3364 {
3365         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3366         struct work_atoms *this;
3367         const char *comm = thread__comm_str(data->thread), *this_comm;
3368         bool leftmost = true;
3369
3370         while (*new) {
3371                 int cmp;
3372
3373                 this = container_of(*new, struct work_atoms, node);
3374                 parent = *new;
3375
3376                 this_comm = thread__comm_str(this->thread);
3377                 cmp = strcmp(comm, this_comm);
3378                 if (cmp > 0) {
3379                         new = &((*new)->rb_left);
3380                 } else if (cmp < 0) {
3381                         new = &((*new)->rb_right);
3382                         leftmost = false;
3383                 } else {
3384                         this->num_merged++;
3385                         this->total_runtime += data->total_runtime;
3386                         this->nb_atoms += data->nb_atoms;
3387                         this->total_lat += data->total_lat;
3388                         list_splice(&data->work_list, &this->work_list);
3389                         if (this->max_lat < data->max_lat) {
3390                                 this->max_lat = data->max_lat;
3391                                 this->max_lat_start = data->max_lat_start;
3392                                 this->max_lat_end = data->max_lat_end;
3393                         }
3394                         zfree(&data);
3395                         return;
3396                 }
3397         }
3398
3399         data->num_merged++;
3400         rb_link_node(&data->node, parent, new);
3401         rb_insert_color_cached(&data->node, root, leftmost);
3402 }
3403
3404 static void perf_sched__merge_lat(struct perf_sched *sched)
3405 {
3406         struct work_atoms *data;
3407         struct rb_node *node;
3408
3409         if (sched->skip_merge)
3410                 return;
3411
3412         while ((node = rb_first_cached(&sched->atom_root))) {
3413                 rb_erase_cached(node, &sched->atom_root);
3414                 data = rb_entry(node, struct work_atoms, node);
3415                 __merge_work_atoms(&sched->merged_atom_root, data);
3416         }
3417 }
3418
3419 static int setup_cpus_switch_event(struct perf_sched *sched)
3420 {
3421         unsigned int i;
3422
3423         sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3424         if (!sched->cpu_last_switched)
3425                 return -1;
3426
3427         sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3428         if (!sched->curr_pid) {
3429                 zfree(&sched->cpu_last_switched);
3430                 return -1;
3431         }
3432
3433         for (i = 0; i < MAX_CPUS; i++)
3434                 sched->curr_pid[i] = -1;
3435
3436         return 0;
3437 }
3438
3439 static void free_cpus_switch_event(struct perf_sched *sched)
3440 {
3441         zfree(&sched->curr_pid);
3442         zfree(&sched->cpu_last_switched);
3443 }
3444
3445 static int perf_sched__lat(struct perf_sched *sched)
3446 {
3447         int rc = -1;
3448         struct rb_node *next;
3449
3450         setup_pager();
3451
3452         if (setup_cpus_switch_event(sched))
3453                 return rc;
3454
3455         if (perf_sched__read_events(sched))
3456                 goto out_free_cpus_switch_event;
3457
3458         perf_sched__merge_lat(sched);
3459         perf_sched__sort_lat(sched);
3460
3461         printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3462         printf("  Task                  |   Runtime ms  |  Count   | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3463         printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3464
3465         next = rb_first_cached(&sched->sorted_atom_root);
3466
3467         while (next) {
3468                 struct work_atoms *work_list;
3469
3470                 work_list = rb_entry(next, struct work_atoms, node);
3471                 output_lat_thread(sched, work_list);
3472                 next = rb_next(next);
3473                 thread__zput(work_list->thread);
3474         }
3475
3476         printf(" -----------------------------------------------------------------------------------------------------------------\n");
3477         printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3478                 (double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3479
3480         printf(" ---------------------------------------------------\n");
3481
3482         print_bad_events(sched);
3483         printf("\n");
3484
3485         rc = 0;
3486
3487 out_free_cpus_switch_event:
3488         free_cpus_switch_event(sched);
3489         return rc;
3490 }
3491
3492 static int setup_map_cpus(struct perf_sched *sched)
3493 {
3494         sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3495
3496         if (sched->map.comp) {
3497                 sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3498                 if (!sched->map.comp_cpus)
3499                         return -1;
3500         }
3501
3502         if (sched->map.cpus_str) {
3503                 sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3504                 if (!sched->map.cpus) {
3505                         pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3506                         zfree(&sched->map.comp_cpus);
3507                         return -1;
3508                 }
3509         }
3510
3511         return 0;
3512 }
3513
3514 static int setup_color_pids(struct perf_sched *sched)
3515 {
3516         struct perf_thread_map *map;
3517
3518         if (!sched->map.color_pids_str)
3519                 return 0;
3520
3521         map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3522         if (!map) {
3523                 pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3524                 return -1;
3525         }
3526
3527         sched->map.color_pids = map;
3528         return 0;
3529 }
3530
3531 static int setup_color_cpus(struct perf_sched *sched)
3532 {
3533         struct perf_cpu_map *map;
3534
3535         if (!sched->map.color_cpus_str)
3536                 return 0;
3537
3538         map = perf_cpu_map__new(sched->map.color_cpus_str);
3539         if (!map) {
3540                 pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3541                 return -1;
3542         }
3543
3544         sched->map.color_cpus = map;
3545         return 0;
3546 }
3547
3548 static int perf_sched__map(struct perf_sched *sched)
3549 {
3550         int rc = -1;
3551
3552         sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3553         if (!sched->curr_thread)
3554                 return rc;
3555
3556         sched->curr_out_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_out_thread)));
3557         if (!sched->curr_out_thread)
3558                 return rc;
3559
3560         if (setup_cpus_switch_event(sched))
3561                 goto out_free_curr_thread;
3562
3563         if (setup_map_cpus(sched))
3564                 goto out_free_cpus_switch_event;
3565
3566         if (setup_color_pids(sched))
3567                 goto out_put_map_cpus;
3568
3569         if (setup_color_cpus(sched))
3570                 goto out_put_color_pids;
3571
3572         setup_pager();
3573         if (perf_sched__read_events(sched))
3574                 goto out_put_color_cpus;
3575
3576         rc = 0;
3577         print_bad_events(sched);
3578
3579 out_put_color_cpus:
3580         perf_cpu_map__put(sched->map.color_cpus);
3581
3582 out_put_color_pids:
3583         perf_thread_map__put(sched->map.color_pids);
3584
3585 out_put_map_cpus:
3586         zfree(&sched->map.comp_cpus);
3587         perf_cpu_map__put(sched->map.cpus);
3588
3589 out_free_cpus_switch_event:
3590         free_cpus_switch_event(sched);
3591
3592 out_free_curr_thread:
3593         zfree(&sched->curr_thread);
3594         return rc;
3595 }
3596
3597 static int perf_sched__replay(struct perf_sched *sched)
3598 {
3599         int ret;
3600         unsigned long i;
3601
3602         mutex_init(&sched->start_work_mutex);
3603         mutex_init(&sched->work_done_wait_mutex);
3604
3605         ret = setup_cpus_switch_event(sched);
3606         if (ret)
3607                 goto out_mutex_destroy;
3608
3609         calibrate_run_measurement_overhead(sched);
3610         calibrate_sleep_measurement_overhead(sched);
3611
3612         test_calibrations(sched);
3613
3614         ret = perf_sched__read_events(sched);
3615         if (ret)
3616                 goto out_free_cpus_switch_event;
3617
3618         printf("nr_run_events:        %ld\n", sched->nr_run_events);
3619         printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3620         printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3621
3622         if (sched->targetless_wakeups)
3623                 printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3624         if (sched->multitarget_wakeups)
3625                 printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3626         if (sched->nr_run_events_optimized)
3627                 printf("run atoms optimized: %ld\n",
3628                         sched->nr_run_events_optimized);
3629
3630         print_task_traces(sched);
3631         add_cross_task_wakeups(sched);
3632
3633         sched->thread_funcs_exit = false;
3634         create_tasks(sched);
3635         printf("------------------------------------------------------------\n");
3636         if (sched->replay_repeat == 0)
3637                 sched->replay_repeat = UINT_MAX;
3638
3639         for (i = 0; i < sched->replay_repeat; i++)
3640                 run_one_test(sched);
3641
3642         sched->thread_funcs_exit = true;
3643         destroy_tasks(sched);
3644
3645 out_free_cpus_switch_event:
3646         free_cpus_switch_event(sched);
3647
3648 out_mutex_destroy:
3649         mutex_destroy(&sched->start_work_mutex);
3650         mutex_destroy(&sched->work_done_wait_mutex);
3651         return ret;
3652 }
3653
3654 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3655                           const char * const usage_msg[])
3656 {
3657         char *tmp, *tok, *str = strdup(sched->sort_order);
3658
3659         for (tok = strtok_r(str, ", ", &tmp);
3660                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
3661                 if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3662                         usage_with_options_msg(usage_msg, options,
3663                                         "Unknown --sort key: `%s'", tok);
3664                 }
3665         }
3666
3667         free(str);
3668
3669         sort_dimension__add("pid", &sched->cmp_pid);
3670 }
3671
3672 static bool schedstat_events_exposed(void)
3673 {
3674         /*
3675          * Select "sched:sched_stat_wait" event to check
3676          * whether schedstat tracepoints are exposed.
3677          */
3678         return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3679                 false : true;
3680 }
3681
3682 static int __cmd_record(int argc, const char **argv)
3683 {
3684         unsigned int rec_argc, i, j;
3685         char **rec_argv;
3686         const char **rec_argv_copy;
3687         const char * const record_args[] = {
3688                 "record",
3689                 "-a",
3690                 "-R",
3691                 "-m", "1024",
3692                 "-c", "1",
3693                 "-e", "sched:sched_switch",
3694                 "-e", "sched:sched_stat_runtime",
3695                 "-e", "sched:sched_process_fork",
3696                 "-e", "sched:sched_wakeup_new",
3697                 "-e", "sched:sched_migrate_task",
3698         };
3699
3700         /*
3701          * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3702          * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3703          * to prevent "perf sched record" execution failure, determine
3704          * whether to record schedstat events according to actual situation.
3705          */
3706         const char * const schedstat_args[] = {
3707                 "-e", "sched:sched_stat_wait",
3708                 "-e", "sched:sched_stat_sleep",
3709                 "-e", "sched:sched_stat_iowait",
3710         };
3711         unsigned int schedstat_argc = schedstat_events_exposed() ?
3712                 ARRAY_SIZE(schedstat_args) : 0;
3713
3714         struct tep_event *waking_event;
3715         int ret;
3716
3717         /*
3718          * +2 for either "-e", "sched:sched_wakeup" or
3719          * "-e", "sched:sched_waking"
3720          */
3721         rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3722         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3723         if (rec_argv == NULL)
3724                 return -ENOMEM;
3725         rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3726         if (rec_argv_copy == NULL) {
3727                 free(rec_argv);
3728                 return -ENOMEM;
3729         }
3730
3731         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3732                 rec_argv[i] = strdup(record_args[i]);
3733
3734         rec_argv[i++] = strdup("-e");
3735         waking_event = trace_event__tp_format("sched", "sched_waking");
3736         if (!IS_ERR(waking_event))
3737                 rec_argv[i++] = strdup("sched:sched_waking");
3738         else
3739                 rec_argv[i++] = strdup("sched:sched_wakeup");
3740
3741         for (j = 0; j < schedstat_argc; j++)
3742                 rec_argv[i++] = strdup(schedstat_args[j]);
3743
3744         for (j = 1; j < (unsigned int)argc; j++, i++)
3745                 rec_argv[i] = strdup(argv[j]);
3746
3747         BUG_ON(i != rec_argc);
3748
3749         memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3750         ret = cmd_record(rec_argc, rec_argv_copy);
3751
3752         for (i = 0; i < rec_argc; i++)
3753                 free(rec_argv[i]);
3754         free(rec_argv);
3755         free(rec_argv_copy);
3756
3757         return ret;
3758 }
3759
3760 int cmd_sched(int argc, const char **argv)
3761 {
3762         static const char default_sort_order[] = "avg, max, switch, runtime";
3763         struct perf_sched sched = {
3764                 .cmp_pid              = LIST_HEAD_INIT(sched.cmp_pid),
3765                 .sort_list            = LIST_HEAD_INIT(sched.sort_list),
3766                 .sort_order           = default_sort_order,
3767                 .replay_repeat        = 10,
3768                 .profile_cpu          = -1,
3769                 .next_shortname1      = 'A',
3770                 .next_shortname2      = '0',
3771                 .skip_merge           = 0,
3772                 .show_callchain       = 1,
3773                 .max_stack            = 5,
3774         };
3775         const struct option sched_options[] = {
3776         OPT_STRING('i', "input", &input_name, "file",
3777                     "input file name"),
3778         OPT_INCR('v', "verbose", &verbose,
3779                     "be more verbose (show symbol address, etc)"),
3780         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3781                     "dump raw trace in ASCII"),
3782         OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3783         OPT_END()
3784         };
3785         const struct option latency_options[] = {
3786         OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3787                    "sort by key(s): runtime, switch, avg, max"),
3788         OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3789                     "CPU to profile on"),
3790         OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3791                     "latency stats per pid instead of per comm"),
3792         OPT_PARENT(sched_options)
3793         };
3794         const struct option replay_options[] = {
3795         OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3796                      "repeat the workload replay N times (0: infinite)"),
3797         OPT_PARENT(sched_options)
3798         };
3799         const struct option map_options[] = {
3800         OPT_BOOLEAN(0, "compact", &sched.map.comp,
3801                     "map output in compact mode"),
3802         OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3803                    "highlight given pids in map"),
3804         OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3805                     "highlight given CPUs in map"),
3806         OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3807                     "display given CPUs in map"),
3808         OPT_STRING(0, "task-name", &sched.map.task_name, "task",
3809                 "map output only for the given task name(s)."),
3810         OPT_BOOLEAN(0, "fuzzy-name", &sched.map.fuzzy,
3811                 "given command name can be partially matched (fuzzy matching)"),
3812         OPT_PARENT(sched_options)
3813         };
3814         const struct option timehist_options[] = {
3815         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3816                    "file", "vmlinux pathname"),
3817         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3818                    "file", "kallsyms pathname"),
3819         OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3820                     "Display call chains if present (default on)"),
3821         OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3822                    "Maximum number of functions to display backtrace."),
3823         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3824                     "Look for files with symbols relative to this directory"),
3825         OPT_BOOLEAN('s', "summary", &sched.summary_only,
3826                     "Show only syscall summary with statistics"),
3827         OPT_BOOLEAN('S', "with-summary", &sched.summary,
3828                     "Show all syscalls and summary with statistics"),
3829         OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3830         OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3831         OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3832         OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3833         OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3834         OPT_STRING(0, "time", &sched.time_str, "str",
3835                    "Time span for analysis (start,stop)"),
3836         OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3837         OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3838                    "analyze events only for given process id(s)"),
3839         OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3840                    "analyze events only for given thread id(s)"),
3841         OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3842         OPT_BOOLEAN(0, "show-prio", &sched.show_prio, "Show task priority"),
3843         OPT_STRING(0, "prio", &sched.prio_str, "prio",
3844                    "analyze events only for given task priority(ies)"),
3845         OPT_BOOLEAN('P', "pre-migrations", &sched.pre_migrations, "Show pre-migration wait time"),
3846         OPT_PARENT(sched_options)
3847         };
3848
3849         const char * const latency_usage[] = {
3850                 "perf sched latency [<options>]",
3851                 NULL
3852         };
3853         const char * const replay_usage[] = {
3854                 "perf sched replay [<options>]",
3855                 NULL
3856         };
3857         const char * const map_usage[] = {
3858                 "perf sched map [<options>]",
3859                 NULL
3860         };
3861         const char * const timehist_usage[] = {
3862                 "perf sched timehist [<options>]",
3863                 NULL
3864         };
3865         const char *const sched_subcommands[] = { "record", "latency", "map",
3866                                                   "replay", "script",
3867                                                   "timehist", NULL };
3868         const char *sched_usage[] = {
3869                 NULL,
3870                 NULL
3871         };
3872         struct trace_sched_handler lat_ops  = {
3873                 .wakeup_event       = latency_wakeup_event,
3874                 .switch_event       = latency_switch_event,
3875                 .runtime_event      = latency_runtime_event,
3876                 .migrate_task_event = latency_migrate_task_event,
3877         };
3878         struct trace_sched_handler map_ops  = {
3879                 .switch_event       = map_switch_event,
3880         };
3881         struct trace_sched_handler replay_ops  = {
3882                 .wakeup_event       = replay_wakeup_event,
3883                 .switch_event       = replay_switch_event,
3884                 .fork_event         = replay_fork_event,
3885         };
3886         int ret;
3887
3888         perf_tool__init(&sched.tool, /*ordered_events=*/true);
3889         sched.tool.sample        = perf_sched__process_tracepoint_sample;
3890         sched.tool.comm          = perf_sched__process_comm;
3891         sched.tool.namespaces    = perf_event__process_namespaces;
3892         sched.tool.lost          = perf_event__process_lost;
3893         sched.tool.fork          = perf_sched__process_fork_event;
3894
3895         argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3896                                         sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3897         if (!argc)
3898                 usage_with_options(sched_usage, sched_options);
3899
3900         /*
3901          * Aliased to 'perf script' for now:
3902          */
3903         if (!strcmp(argv[0], "script")) {
3904                 return cmd_script(argc, argv);
3905         } else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3906                 return __cmd_record(argc, argv);
3907         } else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3908                 sched.tp_handler = &lat_ops;
3909                 if (argc > 1) {
3910                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3911                         if (argc)
3912                                 usage_with_options(latency_usage, latency_options);
3913                 }
3914                 setup_sorting(&sched, latency_options, latency_usage);
3915                 return perf_sched__lat(&sched);
3916         } else if (!strcmp(argv[0], "map")) {
3917                 if (argc) {
3918                         argc = parse_options(argc, argv, map_options, map_usage, 0);
3919                         if (argc)
3920                                 usage_with_options(map_usage, map_options);
3921
3922                         if (sched.map.task_name) {
3923                                 sched.map.task_names = strlist__new(sched.map.task_name, NULL);
3924                                 if (sched.map.task_names == NULL) {
3925                                         fprintf(stderr, "Failed to parse task names\n");
3926                                         return -1;
3927                                 }
3928                         }
3929                 }
3930                 sched.tp_handler = &map_ops;
3931                 setup_sorting(&sched, latency_options, latency_usage);
3932                 return perf_sched__map(&sched);
3933         } else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3934                 sched.tp_handler = &replay_ops;
3935                 if (argc) {
3936                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3937                         if (argc)
3938                                 usage_with_options(replay_usage, replay_options);
3939                 }
3940                 return perf_sched__replay(&sched);
3941         } else if (!strcmp(argv[0], "timehist")) {
3942                 if (argc) {
3943                         argc = parse_options(argc, argv, timehist_options,
3944                                              timehist_usage, 0);
3945                         if (argc)
3946                                 usage_with_options(timehist_usage, timehist_options);
3947                 }
3948                 if ((sched.show_wakeups || sched.show_next) &&
3949                     sched.summary_only) {
3950                         pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3951                         parse_options_usage(timehist_usage, timehist_options, "s", true);
3952                         if (sched.show_wakeups)
3953                                 parse_options_usage(NULL, timehist_options, "w", true);
3954                         if (sched.show_next)
3955                                 parse_options_usage(NULL, timehist_options, "n", true);
3956                         return -EINVAL;
3957                 }
3958                 ret = symbol__validate_sym_arguments();
3959                 if (ret)
3960                         return ret;
3961
3962                 return perf_sched__timehist(&sched);
3963         } else {
3964                 usage_with_options(sched_usage, sched_options);
3965         }
3966
3967         /* free usage string allocated by parse_options_subcommand */
3968         free((void *)sched_usage[0]);
3969
3970         return 0;
3971 }