Linux Audio

Check our new training course

Loading...
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3#include "perf.h"
 
   4
   5#include "util/util.h"
   6#include "util/evlist.h"
   7#include "util/cache.h"
   8#include "util/evsel.h"
 
   9#include "util/symbol.h"
  10#include "util/thread.h"
  11#include "util/header.h"
  12#include "util/session.h"
  13#include "util/tool.h"
  14#include "util/cloexec.h"
  15#include "util/thread_map.h"
  16#include "util/color.h"
  17#include "util/stat.h"
 
  18#include "util/callchain.h"
  19#include "util/time-utils.h"
  20
 
  21#include <subcmd/parse-options.h>
  22#include "util/trace-event.h"
  23
  24#include "util/debug.h"
 
  25
  26#include <linux/kernel.h>
  27#include <linux/log2.h>
 
  28#include <sys/prctl.h>
  29#include <sys/resource.h>
  30#include <inttypes.h>
  31
  32#include <errno.h>
  33#include <semaphore.h>
  34#include <pthread.h>
  35#include <math.h>
  36#include <api/fs/fs.h>
 
  37#include <linux/time64.h>
 
  38
  39#include "sane_ctype.h"
  40
  41#define PR_SET_NAME		15               /* Set process name */
  42#define MAX_CPUS		4096
  43#define COMM_LEN		20
  44#define SYM_LEN			129
  45#define MAX_PID			1024000
  46
 
 
 
  47struct sched_atom;
  48
  49struct task_desc {
  50	unsigned long		nr;
  51	unsigned long		pid;
  52	char			comm[COMM_LEN];
  53
  54	unsigned long		nr_events;
  55	unsigned long		curr_event;
  56	struct sched_atom	**atoms;
  57
  58	pthread_t		thread;
  59	sem_t			sleep_sem;
  60
  61	sem_t			ready_for_work;
  62	sem_t			work_done_sem;
  63
  64	u64			cpu_usage;
  65};
  66
  67enum sched_event_type {
  68	SCHED_EVENT_RUN,
  69	SCHED_EVENT_SLEEP,
  70	SCHED_EVENT_WAKEUP,
  71	SCHED_EVENT_MIGRATION,
  72};
  73
  74struct sched_atom {
  75	enum sched_event_type	type;
  76	int			specific_wait;
  77	u64			timestamp;
  78	u64			duration;
  79	unsigned long		nr;
  80	sem_t			*wait_sem;
  81	struct task_desc	*wakee;
  82};
  83
  84#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  85
  86/* task state bitmask, copied from include/linux/sched.h */
  87#define TASK_RUNNING		0
  88#define TASK_INTERRUPTIBLE	1
  89#define TASK_UNINTERRUPTIBLE	2
  90#define __TASK_STOPPED		4
  91#define __TASK_TRACED		8
  92/* in tsk->exit_state */
  93#define EXIT_DEAD		16
  94#define EXIT_ZOMBIE		32
  95#define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
  96/* in tsk->state again */
  97#define TASK_DEAD		64
  98#define TASK_WAKEKILL		128
  99#define TASK_WAKING		256
 100#define TASK_PARKED		512
 101
 102enum thread_state {
 103	THREAD_SLEEPING = 0,
 104	THREAD_WAIT_CPU,
 105	THREAD_SCHED_IN,
 106	THREAD_IGNORE
 107};
 108
 109struct work_atom {
 110	struct list_head	list;
 111	enum thread_state	state;
 112	u64			sched_out_time;
 113	u64			wake_up_time;
 114	u64			sched_in_time;
 115	u64			runtime;
 116};
 117
 118struct work_atoms {
 119	struct list_head	work_list;
 120	struct thread		*thread;
 121	struct rb_node		node;
 122	u64			max_lat;
 123	u64			max_lat_at;
 124	u64			total_lat;
 125	u64			nb_atoms;
 126	u64			total_runtime;
 127	int			num_merged;
 128};
 129
 130typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 131
 132struct perf_sched;
 133
 134struct trace_sched_handler {
 135	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 136			    struct perf_sample *sample, struct machine *machine);
 137
 138	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 139			     struct perf_sample *sample, struct machine *machine);
 140
 141	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
 142			    struct perf_sample *sample, struct machine *machine);
 143
 144	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 145	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 146			  struct machine *machine);
 147
 148	int (*migrate_task_event)(struct perf_sched *sched,
 149				  struct perf_evsel *evsel,
 150				  struct perf_sample *sample,
 151				  struct machine *machine);
 152};
 153
 154#define COLOR_PIDS PERF_COLOR_BLUE
 155#define COLOR_CPUS PERF_COLOR_BG_RED
 156
 157struct perf_sched_map {
 158	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 159	int			*comp_cpus;
 160	bool			 comp;
 161	struct thread_map	*color_pids;
 162	const char		*color_pids_str;
 163	struct cpu_map		*color_cpus;
 164	const char		*color_cpus_str;
 165	struct cpu_map		*cpus;
 166	const char		*cpus_str;
 167};
 168
 169struct perf_sched {
 170	struct perf_tool tool;
 171	const char	 *sort_order;
 172	unsigned long	 nr_tasks;
 173	struct task_desc **pid_to_task;
 174	struct task_desc **tasks;
 175	const struct trace_sched_handler *tp_handler;
 176	pthread_mutex_t	 start_work_mutex;
 177	pthread_mutex_t	 work_done_wait_mutex;
 178	int		 profile_cpu;
 179/*
 180 * Track the current task - that way we can know whether there's any
 181 * weird events, such as a task being switched away that is not current.
 182 */
 183	int		 max_cpu;
 184	u32		 curr_pid[MAX_CPUS];
 185	struct thread	 *curr_thread[MAX_CPUS];
 186	char		 next_shortname1;
 187	char		 next_shortname2;
 188	unsigned int	 replay_repeat;
 189	unsigned long	 nr_run_events;
 190	unsigned long	 nr_sleep_events;
 191	unsigned long	 nr_wakeup_events;
 192	unsigned long	 nr_sleep_corrections;
 193	unsigned long	 nr_run_events_optimized;
 194	unsigned long	 targetless_wakeups;
 195	unsigned long	 multitarget_wakeups;
 196	unsigned long	 nr_runs;
 197	unsigned long	 nr_timestamps;
 198	unsigned long	 nr_unordered_timestamps;
 199	unsigned long	 nr_context_switch_bugs;
 200	unsigned long	 nr_events;
 201	unsigned long	 nr_lost_chunks;
 202	unsigned long	 nr_lost_events;
 203	u64		 run_measurement_overhead;
 204	u64		 sleep_measurement_overhead;
 205	u64		 start_time;
 206	u64		 cpu_usage;
 207	u64		 runavg_cpu_usage;
 208	u64		 parent_cpu_usage;
 209	u64		 runavg_parent_cpu_usage;
 210	u64		 sum_runtime;
 211	u64		 sum_fluct;
 212	u64		 run_avg;
 213	u64		 all_runtime;
 214	u64		 all_count;
 215	u64		 cpu_last_switched[MAX_CPUS];
 216	struct rb_root	 atom_root, sorted_atom_root, merged_atom_root;
 217	struct list_head sort_list, cmp_pid;
 218	bool force;
 219	bool skip_merge;
 220	struct perf_sched_map map;
 221
 222	/* options for timehist command */
 223	bool		summary;
 224	bool		summary_only;
 225	bool		idle_hist;
 226	bool		show_callchain;
 227	unsigned int	max_stack;
 228	bool		show_cpu_visual;
 229	bool		show_wakeups;
 230	bool		show_next;
 231	bool		show_migrations;
 232	bool		show_state;
 233	u64		skipped_samples;
 234	const char	*time_str;
 235	struct perf_time_interval ptime;
 236	struct perf_time_interval hist_time;
 237};
 238
 239/* per thread run time data */
 240struct thread_runtime {
 241	u64 last_time;      /* time of previous sched in/out event */
 242	u64 dt_run;         /* run time */
 243	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 244	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 245	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 246	u64 dt_delay;       /* time between wakeup and sched-in */
 247	u64 ready_to_run;   /* time of wakeup */
 248
 249	struct stats run_stats;
 250	u64 total_run_time;
 251	u64 total_sleep_time;
 252	u64 total_iowait_time;
 253	u64 total_preempt_time;
 254	u64 total_delay_time;
 255
 256	int last_state;
 257
 258	char shortname[3];
 259	bool comm_changed;
 260
 261	u64 migrations;
 262};
 263
 264/* per event run time data */
 265struct evsel_runtime {
 266	u64 *last_time; /* time this event was last seen per cpu */
 267	u32 ncpu;       /* highest cpu slot allocated */
 268};
 269
 270/* per cpu idle time data */
 271struct idle_thread_runtime {
 272	struct thread_runtime	tr;
 273	struct thread		*last_thread;
 274	struct rb_root		sorted_root;
 275	struct callchain_root	callchain;
 276	struct callchain_cursor	cursor;
 277};
 278
 279/* track idle times per cpu */
 280static struct thread **idle_threads;
 281static int idle_max_cpu;
 282static char idle_comm[] = "<idle>";
 283
 284static u64 get_nsecs(void)
 285{
 286	struct timespec ts;
 287
 288	clock_gettime(CLOCK_MONOTONIC, &ts);
 289
 290	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 291}
 292
 293static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 294{
 295	u64 T0 = get_nsecs(), T1;
 296
 297	do {
 298		T1 = get_nsecs();
 299	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 300}
 301
 302static void sleep_nsecs(u64 nsecs)
 303{
 304	struct timespec ts;
 305
 306	ts.tv_nsec = nsecs % 999999999;
 307	ts.tv_sec = nsecs / 999999999;
 308
 309	nanosleep(&ts, NULL);
 310}
 311
 312static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 313{
 314	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 315	int i;
 316
 317	for (i = 0; i < 10; i++) {
 318		T0 = get_nsecs();
 319		burn_nsecs(sched, 0);
 320		T1 = get_nsecs();
 321		delta = T1-T0;
 322		min_delta = min(min_delta, delta);
 323	}
 324	sched->run_measurement_overhead = min_delta;
 325
 326	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 327}
 328
 329static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 330{
 331	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 332	int i;
 333
 334	for (i = 0; i < 10; i++) {
 335		T0 = get_nsecs();
 336		sleep_nsecs(10000);
 337		T1 = get_nsecs();
 338		delta = T1-T0;
 339		min_delta = min(min_delta, delta);
 340	}
 341	min_delta -= 10000;
 342	sched->sleep_measurement_overhead = min_delta;
 343
 344	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 345}
 346
 347static struct sched_atom *
 348get_new_event(struct task_desc *task, u64 timestamp)
 349{
 350	struct sched_atom *event = zalloc(sizeof(*event));
 351	unsigned long idx = task->nr_events;
 352	size_t size;
 353
 354	event->timestamp = timestamp;
 355	event->nr = idx;
 356
 357	task->nr_events++;
 358	size = sizeof(struct sched_atom *) * task->nr_events;
 359	task->atoms = realloc(task->atoms, size);
 360	BUG_ON(!task->atoms);
 361
 362	task->atoms[idx] = event;
 363
 364	return event;
 365}
 366
 367static struct sched_atom *last_event(struct task_desc *task)
 368{
 369	if (!task->nr_events)
 370		return NULL;
 371
 372	return task->atoms[task->nr_events - 1];
 373}
 374
 375static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 376				u64 timestamp, u64 duration)
 377{
 378	struct sched_atom *event, *curr_event = last_event(task);
 379
 380	/*
 381	 * optimize an existing RUN event by merging this one
 382	 * to it:
 383	 */
 384	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 385		sched->nr_run_events_optimized++;
 386		curr_event->duration += duration;
 387		return;
 388	}
 389
 390	event = get_new_event(task, timestamp);
 391
 392	event->type = SCHED_EVENT_RUN;
 393	event->duration = duration;
 394
 395	sched->nr_run_events++;
 396}
 397
 398static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 399				   u64 timestamp, struct task_desc *wakee)
 400{
 401	struct sched_atom *event, *wakee_event;
 402
 403	event = get_new_event(task, timestamp);
 404	event->type = SCHED_EVENT_WAKEUP;
 405	event->wakee = wakee;
 406
 407	wakee_event = last_event(wakee);
 408	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 409		sched->targetless_wakeups++;
 410		return;
 411	}
 412	if (wakee_event->wait_sem) {
 413		sched->multitarget_wakeups++;
 414		return;
 415	}
 416
 417	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 418	sem_init(wakee_event->wait_sem, 0, 0);
 419	wakee_event->specific_wait = 1;
 420	event->wait_sem = wakee_event->wait_sem;
 421
 422	sched->nr_wakeup_events++;
 423}
 424
 425static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 426				  u64 timestamp, u64 task_state __maybe_unused)
 427{
 428	struct sched_atom *event = get_new_event(task, timestamp);
 429
 430	event->type = SCHED_EVENT_SLEEP;
 431
 432	sched->nr_sleep_events++;
 433}
 434
 435static struct task_desc *register_pid(struct perf_sched *sched,
 436				      unsigned long pid, const char *comm)
 437{
 438	struct task_desc *task;
 439	static int pid_max;
 440
 441	if (sched->pid_to_task == NULL) {
 442		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 443			pid_max = MAX_PID;
 444		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 445	}
 446	if (pid >= (unsigned long)pid_max) {
 447		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 448			sizeof(struct task_desc *))) == NULL);
 449		while (pid >= (unsigned long)pid_max)
 450			sched->pid_to_task[pid_max++] = NULL;
 451	}
 452
 453	task = sched->pid_to_task[pid];
 454
 455	if (task)
 456		return task;
 457
 458	task = zalloc(sizeof(*task));
 459	task->pid = pid;
 460	task->nr = sched->nr_tasks;
 461	strcpy(task->comm, comm);
 462	/*
 463	 * every task starts in sleeping state - this gets ignored
 464	 * if there's no wakeup pointing to this sleep state:
 465	 */
 466	add_sched_event_sleep(sched, task, 0, 0);
 467
 468	sched->pid_to_task[pid] = task;
 469	sched->nr_tasks++;
 470	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 471	BUG_ON(!sched->tasks);
 472	sched->tasks[task->nr] = task;
 473
 474	if (verbose > 0)
 475		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 476
 477	return task;
 478}
 479
 480
 481static void print_task_traces(struct perf_sched *sched)
 482{
 483	struct task_desc *task;
 484	unsigned long i;
 485
 486	for (i = 0; i < sched->nr_tasks; i++) {
 487		task = sched->tasks[i];
 488		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 489			task->nr, task->comm, task->pid, task->nr_events);
 490	}
 491}
 492
 493static void add_cross_task_wakeups(struct perf_sched *sched)
 494{
 495	struct task_desc *task1, *task2;
 496	unsigned long i, j;
 497
 498	for (i = 0; i < sched->nr_tasks; i++) {
 499		task1 = sched->tasks[i];
 500		j = i + 1;
 501		if (j == sched->nr_tasks)
 502			j = 0;
 503		task2 = sched->tasks[j];
 504		add_sched_event_wakeup(sched, task1, 0, task2);
 505	}
 506}
 507
 508static void perf_sched__process_event(struct perf_sched *sched,
 509				      struct sched_atom *atom)
 510{
 511	int ret = 0;
 512
 513	switch (atom->type) {
 514		case SCHED_EVENT_RUN:
 515			burn_nsecs(sched, atom->duration);
 516			break;
 517		case SCHED_EVENT_SLEEP:
 518			if (atom->wait_sem)
 519				ret = sem_wait(atom->wait_sem);
 520			BUG_ON(ret);
 521			break;
 522		case SCHED_EVENT_WAKEUP:
 523			if (atom->wait_sem)
 524				ret = sem_post(atom->wait_sem);
 525			BUG_ON(ret);
 526			break;
 527		case SCHED_EVENT_MIGRATION:
 528			break;
 529		default:
 530			BUG_ON(1);
 531	}
 532}
 533
 534static u64 get_cpu_usage_nsec_parent(void)
 535{
 536	struct rusage ru;
 537	u64 sum;
 538	int err;
 539
 540	err = getrusage(RUSAGE_SELF, &ru);
 541	BUG_ON(err);
 542
 543	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 544	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 545
 546	return sum;
 547}
 548
 549static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 550{
 551	struct perf_event_attr attr;
 552	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 553	int fd;
 554	struct rlimit limit;
 555	bool need_privilege = false;
 556
 557	memset(&attr, 0, sizeof(attr));
 558
 559	attr.type = PERF_TYPE_SOFTWARE;
 560	attr.config = PERF_COUNT_SW_TASK_CLOCK;
 561
 562force_again:
 563	fd = sys_perf_event_open(&attr, 0, -1, -1,
 564				 perf_event_open_cloexec_flag());
 565
 566	if (fd < 0) {
 567		if (errno == EMFILE) {
 568			if (sched->force) {
 569				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 570				limit.rlim_cur += sched->nr_tasks - cur_task;
 571				if (limit.rlim_cur > limit.rlim_max) {
 572					limit.rlim_max = limit.rlim_cur;
 573					need_privilege = true;
 574				}
 575				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 576					if (need_privilege && errno == EPERM)
 577						strcpy(info, "Need privilege\n");
 578				} else
 579					goto force_again;
 580			} else
 581				strcpy(info, "Have a try with -f option\n");
 582		}
 583		pr_err("Error: sys_perf_event_open() syscall returned "
 584		       "with %d (%s)\n%s", fd,
 585		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
 586		exit(EXIT_FAILURE);
 587	}
 588	return fd;
 589}
 590
 591static u64 get_cpu_usage_nsec_self(int fd)
 592{
 593	u64 runtime;
 594	int ret;
 595
 596	ret = read(fd, &runtime, sizeof(runtime));
 597	BUG_ON(ret != sizeof(runtime));
 598
 599	return runtime;
 600}
 601
 602struct sched_thread_parms {
 603	struct task_desc  *task;
 604	struct perf_sched *sched;
 605	int fd;
 606};
 607
 608static void *thread_func(void *ctx)
 609{
 610	struct sched_thread_parms *parms = ctx;
 611	struct task_desc *this_task = parms->task;
 612	struct perf_sched *sched = parms->sched;
 613	u64 cpu_usage_0, cpu_usage_1;
 614	unsigned long i, ret;
 615	char comm2[22];
 616	int fd = parms->fd;
 617
 618	zfree(&parms);
 619
 620	sprintf(comm2, ":%s", this_task->comm);
 621	prctl(PR_SET_NAME, comm2);
 622	if (fd < 0)
 623		return NULL;
 624again:
 625	ret = sem_post(&this_task->ready_for_work);
 626	BUG_ON(ret);
 627	ret = pthread_mutex_lock(&sched->start_work_mutex);
 628	BUG_ON(ret);
 629	ret = pthread_mutex_unlock(&sched->start_work_mutex);
 630	BUG_ON(ret);
 631
 632	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 633
 634	for (i = 0; i < this_task->nr_events; i++) {
 635		this_task->curr_event = i;
 636		perf_sched__process_event(sched, this_task->atoms[i]);
 637	}
 638
 639	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 640	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 641	ret = sem_post(&this_task->work_done_sem);
 642	BUG_ON(ret);
 643
 644	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 645	BUG_ON(ret);
 646	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 647	BUG_ON(ret);
 648
 649	goto again;
 650}
 651
 652static void create_tasks(struct perf_sched *sched)
 653{
 654	struct task_desc *task;
 655	pthread_attr_t attr;
 656	unsigned long i;
 657	int err;
 658
 659	err = pthread_attr_init(&attr);
 660	BUG_ON(err);
 661	err = pthread_attr_setstacksize(&attr,
 662			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 663	BUG_ON(err);
 664	err = pthread_mutex_lock(&sched->start_work_mutex);
 665	BUG_ON(err);
 666	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 667	BUG_ON(err);
 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->sleep_sem, 0, 0);
 675		sem_init(&task->ready_for_work, 0, 0);
 676		sem_init(&task->work_done_sem, 0, 0);
 677		task->curr_event = 0;
 678		err = pthread_create(&task->thread, &attr, thread_func, parms);
 679		BUG_ON(err);
 680	}
 681}
 682
 683static void wait_for_tasks(struct perf_sched *sched)
 684{
 685	u64 cpu_usage_0, cpu_usage_1;
 686	struct task_desc *task;
 687	unsigned long i, ret;
 688
 689	sched->start_time = get_nsecs();
 690	sched->cpu_usage = 0;
 691	pthread_mutex_unlock(&sched->work_done_wait_mutex);
 692
 693	for (i = 0; i < sched->nr_tasks; i++) {
 694		task = sched->tasks[i];
 695		ret = sem_wait(&task->ready_for_work);
 696		BUG_ON(ret);
 697		sem_init(&task->ready_for_work, 0, 0);
 698	}
 699	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 700	BUG_ON(ret);
 701
 702	cpu_usage_0 = get_cpu_usage_nsec_parent();
 703
 704	pthread_mutex_unlock(&sched->start_work_mutex);
 705
 706	for (i = 0; i < sched->nr_tasks; i++) {
 707		task = sched->tasks[i];
 708		ret = sem_wait(&task->work_done_sem);
 709		BUG_ON(ret);
 710		sem_init(&task->work_done_sem, 0, 0);
 711		sched->cpu_usage += task->cpu_usage;
 712		task->cpu_usage = 0;
 713	}
 714
 715	cpu_usage_1 = get_cpu_usage_nsec_parent();
 716	if (!sched->runavg_cpu_usage)
 717		sched->runavg_cpu_usage = sched->cpu_usage;
 718	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 719
 720	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 721	if (!sched->runavg_parent_cpu_usage)
 722		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 723	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 724					 sched->parent_cpu_usage)/sched->replay_repeat;
 725
 726	ret = pthread_mutex_lock(&sched->start_work_mutex);
 727	BUG_ON(ret);
 728
 729	for (i = 0; i < sched->nr_tasks; i++) {
 730		task = sched->tasks[i];
 731		sem_init(&task->sleep_sem, 0, 0);
 732		task->curr_event = 0;
 733	}
 734}
 735
 736static void run_one_test(struct perf_sched *sched)
 737{
 738	u64 T0, T1, delta, avg_delta, fluct;
 739
 740	T0 = get_nsecs();
 741	wait_for_tasks(sched);
 742	T1 = get_nsecs();
 743
 744	delta = T1 - T0;
 745	sched->sum_runtime += delta;
 746	sched->nr_runs++;
 747
 748	avg_delta = sched->sum_runtime / sched->nr_runs;
 749	if (delta < avg_delta)
 750		fluct = avg_delta - delta;
 751	else
 752		fluct = delta - avg_delta;
 753	sched->sum_fluct += fluct;
 754	if (!sched->run_avg)
 755		sched->run_avg = delta;
 756	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 757
 758	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 759
 760	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 761
 762	printf("cpu: %0.2f / %0.2f",
 763		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 764
 765#if 0
 766	/*
 767	 * rusage statistics done by the parent, these are less
 768	 * accurate than the sched->sum_exec_runtime based statistics:
 769	 */
 770	printf(" [%0.2f / %0.2f]",
 771		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 772		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 773#endif
 774
 775	printf("\n");
 776
 777	if (sched->nr_sleep_corrections)
 778		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 779	sched->nr_sleep_corrections = 0;
 780}
 781
 782static void test_calibrations(struct perf_sched *sched)
 783{
 784	u64 T0, T1;
 785
 786	T0 = get_nsecs();
 787	burn_nsecs(sched, NSEC_PER_MSEC);
 788	T1 = get_nsecs();
 789
 790	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 791
 792	T0 = get_nsecs();
 793	sleep_nsecs(NSEC_PER_MSEC);
 794	T1 = get_nsecs();
 795
 796	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 797}
 798
 799static int
 800replay_wakeup_event(struct perf_sched *sched,
 801		    struct perf_evsel *evsel, struct perf_sample *sample,
 802		    struct machine *machine __maybe_unused)
 803{
 804	const char *comm = perf_evsel__strval(evsel, sample, "comm");
 805	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
 806	struct task_desc *waker, *wakee;
 807
 808	if (verbose > 0) {
 809		printf("sched_wakeup event %p\n", evsel);
 810
 811		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 812	}
 813
 814	waker = register_pid(sched, sample->tid, "<unknown>");
 815	wakee = register_pid(sched, pid, comm);
 816
 817	add_sched_event_wakeup(sched, waker, sample->time, wakee);
 818	return 0;
 819}
 820
 821static int replay_switch_event(struct perf_sched *sched,
 822			       struct perf_evsel *evsel,
 823			       struct perf_sample *sample,
 824			       struct machine *machine __maybe_unused)
 825{
 826	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
 827		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
 828	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
 829		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
 830	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
 831	struct task_desc *prev, __maybe_unused *next;
 832	u64 timestamp0, timestamp = sample->time;
 833	int cpu = sample->cpu;
 834	s64 delta;
 835
 836	if (verbose > 0)
 837		printf("sched_switch event %p\n", evsel);
 838
 839	if (cpu >= MAX_CPUS || cpu < 0)
 840		return 0;
 841
 842	timestamp0 = sched->cpu_last_switched[cpu];
 843	if (timestamp0)
 844		delta = timestamp - timestamp0;
 845	else
 846		delta = 0;
 847
 848	if (delta < 0) {
 849		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 850		return -1;
 851	}
 852
 853	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 854		 prev_comm, prev_pid, next_comm, next_pid, delta);
 855
 856	prev = register_pid(sched, prev_pid, prev_comm);
 857	next = register_pid(sched, next_pid, next_comm);
 858
 859	sched->cpu_last_switched[cpu] = timestamp;
 860
 861	add_sched_event_run(sched, prev, timestamp, delta);
 862	add_sched_event_sleep(sched, prev, timestamp, prev_state);
 863
 864	return 0;
 865}
 866
 867static int replay_fork_event(struct perf_sched *sched,
 868			     union perf_event *event,
 869			     struct machine *machine)
 870{
 871	struct thread *child, *parent;
 872
 873	child = machine__findnew_thread(machine, event->fork.pid,
 874					event->fork.tid);
 875	parent = machine__findnew_thread(machine, event->fork.ppid,
 876					 event->fork.ptid);
 877
 878	if (child == NULL || parent == NULL) {
 879		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 880				 child, parent);
 881		goto out_put;
 882	}
 883
 884	if (verbose > 0) {
 885		printf("fork event\n");
 886		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 887		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 888	}
 889
 890	register_pid(sched, parent->tid, thread__comm_str(parent));
 891	register_pid(sched, child->tid, thread__comm_str(child));
 892out_put:
 893	thread__put(child);
 894	thread__put(parent);
 895	return 0;
 896}
 897
 898struct sort_dimension {
 899	const char		*name;
 900	sort_fn_t		cmp;
 901	struct list_head	list;
 902};
 903
 904/*
 905 * handle runtime stats saved per thread
 906 */
 907static struct thread_runtime *thread__init_runtime(struct thread *thread)
 908{
 909	struct thread_runtime *r;
 910
 911	r = zalloc(sizeof(struct thread_runtime));
 912	if (!r)
 913		return NULL;
 914
 915	init_stats(&r->run_stats);
 916	thread__set_priv(thread, r);
 917
 918	return r;
 919}
 920
 921static struct thread_runtime *thread__get_runtime(struct thread *thread)
 922{
 923	struct thread_runtime *tr;
 924
 925	tr = thread__priv(thread);
 926	if (tr == NULL) {
 927		tr = thread__init_runtime(thread);
 928		if (tr == NULL)
 929			pr_debug("Failed to malloc memory for runtime data.\n");
 930	}
 931
 932	return tr;
 933}
 934
 935static int
 936thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 937{
 938	struct sort_dimension *sort;
 939	int ret = 0;
 940
 941	BUG_ON(list_empty(list));
 942
 943	list_for_each_entry(sort, list, list) {
 944		ret = sort->cmp(l, r);
 945		if (ret)
 946			return ret;
 947	}
 948
 949	return ret;
 950}
 951
 952static struct work_atoms *
 953thread_atoms_search(struct rb_root *root, struct thread *thread,
 954			 struct list_head *sort_list)
 955{
 956	struct rb_node *node = root->rb_node;
 957	struct work_atoms key = { .thread = thread };
 958
 959	while (node) {
 960		struct work_atoms *atoms;
 961		int cmp;
 962
 963		atoms = container_of(node, struct work_atoms, node);
 964
 965		cmp = thread_lat_cmp(sort_list, &key, atoms);
 966		if (cmp > 0)
 967			node = node->rb_left;
 968		else if (cmp < 0)
 969			node = node->rb_right;
 970		else {
 971			BUG_ON(thread != atoms->thread);
 972			return atoms;
 973		}
 974	}
 975	return NULL;
 976}
 977
 978static void
 979__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 980			 struct list_head *sort_list)
 981{
 982	struct rb_node **new = &(root->rb_node), *parent = NULL;
 
 983
 984	while (*new) {
 985		struct work_atoms *this;
 986		int cmp;
 987
 988		this = container_of(*new, struct work_atoms, node);
 989		parent = *new;
 990
 991		cmp = thread_lat_cmp(sort_list, data, this);
 992
 993		if (cmp > 0)
 994			new = &((*new)->rb_left);
 995		else
 996			new = &((*new)->rb_right);
 
 
 997	}
 998
 999	rb_link_node(&data->node, parent, new);
1000	rb_insert_color(&data->node, root);
1001}
1002
1003static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1004{
1005	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1006	if (!atoms) {
1007		pr_err("No memory at %s\n", __func__);
1008		return -1;
1009	}
1010
1011	atoms->thread = thread__get(thread);
1012	INIT_LIST_HEAD(&atoms->work_list);
1013	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1014	return 0;
1015}
1016
1017static char sched_out_state(u64 prev_state)
1018{
1019	const char *str = TASK_STATE_TO_CHAR_STR;
1020
1021	return str[prev_state];
1022}
1023
1024static int
1025add_sched_out_event(struct work_atoms *atoms,
1026		    char run_state,
1027		    u64 timestamp)
1028{
1029	struct work_atom *atom = zalloc(sizeof(*atom));
1030	if (!atom) {
1031		pr_err("Non memory at %s", __func__);
1032		return -1;
1033	}
1034
1035	atom->sched_out_time = timestamp;
1036
1037	if (run_state == 'R') {
1038		atom->state = THREAD_WAIT_CPU;
1039		atom->wake_up_time = atom->sched_out_time;
1040	}
1041
1042	list_add_tail(&atom->list, &atoms->work_list);
1043	return 0;
1044}
1045
1046static void
1047add_runtime_event(struct work_atoms *atoms, u64 delta,
1048		  u64 timestamp __maybe_unused)
1049{
1050	struct work_atom *atom;
1051
1052	BUG_ON(list_empty(&atoms->work_list));
1053
1054	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1055
1056	atom->runtime += delta;
1057	atoms->total_runtime += delta;
1058}
1059
1060static void
1061add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1062{
1063	struct work_atom *atom;
1064	u64 delta;
1065
1066	if (list_empty(&atoms->work_list))
1067		return;
1068
1069	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1070
1071	if (atom->state != THREAD_WAIT_CPU)
1072		return;
1073
1074	if (timestamp < atom->wake_up_time) {
1075		atom->state = THREAD_IGNORE;
1076		return;
1077	}
1078
1079	atom->state = THREAD_SCHED_IN;
1080	atom->sched_in_time = timestamp;
1081
1082	delta = atom->sched_in_time - atom->wake_up_time;
1083	atoms->total_lat += delta;
1084	if (delta > atoms->max_lat) {
1085		atoms->max_lat = delta;
1086		atoms->max_lat_at = timestamp;
1087	}
1088	atoms->nb_atoms++;
1089}
1090
1091static int latency_switch_event(struct perf_sched *sched,
1092				struct perf_evsel *evsel,
1093				struct perf_sample *sample,
1094				struct machine *machine)
1095{
1096	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1097		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1098	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1099	struct work_atoms *out_events, *in_events;
1100	struct thread *sched_out, *sched_in;
1101	u64 timestamp0, timestamp = sample->time;
1102	int cpu = sample->cpu, err = -1;
1103	s64 delta;
1104
1105	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1106
1107	timestamp0 = sched->cpu_last_switched[cpu];
1108	sched->cpu_last_switched[cpu] = timestamp;
1109	if (timestamp0)
1110		delta = timestamp - timestamp0;
1111	else
1112		delta = 0;
1113
1114	if (delta < 0) {
1115		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1116		return -1;
1117	}
1118
1119	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1120	sched_in = machine__findnew_thread(machine, -1, next_pid);
1121	if (sched_out == NULL || sched_in == NULL)
1122		goto out_put;
1123
1124	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1125	if (!out_events) {
1126		if (thread_atoms_insert(sched, sched_out))
1127			goto out_put;
1128		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1129		if (!out_events) {
1130			pr_err("out-event: Internal tree error");
1131			goto out_put;
1132		}
1133	}
1134	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1135		return -1;
1136
1137	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1138	if (!in_events) {
1139		if (thread_atoms_insert(sched, sched_in))
1140			goto out_put;
1141		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1142		if (!in_events) {
1143			pr_err("in-event: Internal tree error");
1144			goto out_put;
1145		}
1146		/*
1147		 * Take came in we have not heard about yet,
1148		 * add in an initial atom in runnable state:
1149		 */
1150		if (add_sched_out_event(in_events, 'R', timestamp))
1151			goto out_put;
1152	}
1153	add_sched_in_event(in_events, timestamp);
1154	err = 0;
1155out_put:
1156	thread__put(sched_out);
1157	thread__put(sched_in);
1158	return err;
1159}
1160
1161static int latency_runtime_event(struct perf_sched *sched,
1162				 struct perf_evsel *evsel,
1163				 struct perf_sample *sample,
1164				 struct machine *machine)
1165{
1166	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1167	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1168	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1169	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1170	u64 timestamp = sample->time;
1171	int cpu = sample->cpu, err = -1;
1172
1173	if (thread == NULL)
1174		return -1;
1175
1176	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1177	if (!atoms) {
1178		if (thread_atoms_insert(sched, thread))
1179			goto out_put;
1180		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1181		if (!atoms) {
1182			pr_err("in-event: Internal tree error");
1183			goto out_put;
1184		}
1185		if (add_sched_out_event(atoms, 'R', timestamp))
1186			goto out_put;
1187	}
1188
1189	add_runtime_event(atoms, runtime, timestamp);
1190	err = 0;
1191out_put:
1192	thread__put(thread);
1193	return err;
1194}
1195
1196static int latency_wakeup_event(struct perf_sched *sched,
1197				struct perf_evsel *evsel,
1198				struct perf_sample *sample,
1199				struct machine *machine)
1200{
1201	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1202	struct work_atoms *atoms;
1203	struct work_atom *atom;
1204	struct thread *wakee;
1205	u64 timestamp = sample->time;
1206	int err = -1;
1207
1208	wakee = machine__findnew_thread(machine, -1, pid);
1209	if (wakee == NULL)
1210		return -1;
1211	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1212	if (!atoms) {
1213		if (thread_atoms_insert(sched, wakee))
1214			goto out_put;
1215		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1216		if (!atoms) {
1217			pr_err("wakeup-event: Internal tree error");
1218			goto out_put;
1219		}
1220		if (add_sched_out_event(atoms, 'S', timestamp))
1221			goto out_put;
1222	}
1223
1224	BUG_ON(list_empty(&atoms->work_list));
1225
1226	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1227
1228	/*
1229	 * As we do not guarantee the wakeup event happens when
1230	 * task is out of run queue, also may happen when task is
1231	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1232	 * then we should not set the ->wake_up_time when wake up a
1233	 * task which is on run queue.
1234	 *
1235	 * You WILL be missing events if you've recorded only
1236	 * one CPU, or are only looking at only one, so don't
1237	 * skip in this case.
1238	 */
1239	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1240		goto out_ok;
1241
1242	sched->nr_timestamps++;
1243	if (atom->sched_out_time > timestamp) {
1244		sched->nr_unordered_timestamps++;
1245		goto out_ok;
1246	}
1247
1248	atom->state = THREAD_WAIT_CPU;
1249	atom->wake_up_time = timestamp;
1250out_ok:
1251	err = 0;
1252out_put:
1253	thread__put(wakee);
1254	return err;
1255}
1256
1257static int latency_migrate_task_event(struct perf_sched *sched,
1258				      struct perf_evsel *evsel,
1259				      struct perf_sample *sample,
1260				      struct machine *machine)
1261{
1262	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1263	u64 timestamp = sample->time;
1264	struct work_atoms *atoms;
1265	struct work_atom *atom;
1266	struct thread *migrant;
1267	int err = -1;
1268
1269	/*
1270	 * Only need to worry about migration when profiling one CPU.
1271	 */
1272	if (sched->profile_cpu == -1)
1273		return 0;
1274
1275	migrant = machine__findnew_thread(machine, -1, pid);
1276	if (migrant == NULL)
1277		return -1;
1278	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1279	if (!atoms) {
1280		if (thread_atoms_insert(sched, migrant))
1281			goto out_put;
1282		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1283		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1284		if (!atoms) {
1285			pr_err("migration-event: Internal tree error");
1286			goto out_put;
1287		}
1288		if (add_sched_out_event(atoms, 'R', timestamp))
1289			goto out_put;
1290	}
1291
1292	BUG_ON(list_empty(&atoms->work_list));
1293
1294	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1295	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1296
1297	sched->nr_timestamps++;
1298
1299	if (atom->sched_out_time > timestamp)
1300		sched->nr_unordered_timestamps++;
1301	err = 0;
1302out_put:
1303	thread__put(migrant);
1304	return err;
1305}
1306
1307static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1308{
1309	int i;
1310	int ret;
1311	u64 avg;
1312	char max_lat_at[32];
1313
1314	if (!work_list->nb_atoms)
1315		return;
1316	/*
1317	 * Ignore idle threads:
1318	 */
1319	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1320		return;
1321
1322	sched->all_runtime += work_list->total_runtime;
1323	sched->all_count   += work_list->nb_atoms;
1324
1325	if (work_list->num_merged > 1)
1326		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1327	else
1328		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1329
1330	for (i = 0; i < 24 - ret; i++)
1331		printf(" ");
1332
1333	avg = work_list->total_lat / work_list->nb_atoms;
1334	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1335
1336	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1337	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1338		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1339		 (double)work_list->max_lat / NSEC_PER_MSEC,
1340		 max_lat_at);
1341}
1342
1343static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1344{
1345	if (l->thread == r->thread)
1346		return 0;
1347	if (l->thread->tid < r->thread->tid)
1348		return -1;
1349	if (l->thread->tid > r->thread->tid)
1350		return 1;
1351	return (int)(l->thread - r->thread);
1352}
1353
1354static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1355{
1356	u64 avgl, avgr;
1357
1358	if (!l->nb_atoms)
1359		return -1;
1360
1361	if (!r->nb_atoms)
1362		return 1;
1363
1364	avgl = l->total_lat / l->nb_atoms;
1365	avgr = r->total_lat / r->nb_atoms;
1366
1367	if (avgl < avgr)
1368		return -1;
1369	if (avgl > avgr)
1370		return 1;
1371
1372	return 0;
1373}
1374
1375static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1376{
1377	if (l->max_lat < r->max_lat)
1378		return -1;
1379	if (l->max_lat > r->max_lat)
1380		return 1;
1381
1382	return 0;
1383}
1384
1385static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1386{
1387	if (l->nb_atoms < r->nb_atoms)
1388		return -1;
1389	if (l->nb_atoms > r->nb_atoms)
1390		return 1;
1391
1392	return 0;
1393}
1394
1395static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1396{
1397	if (l->total_runtime < r->total_runtime)
1398		return -1;
1399	if (l->total_runtime > r->total_runtime)
1400		return 1;
1401
1402	return 0;
1403}
1404
1405static int sort_dimension__add(const char *tok, struct list_head *list)
1406{
1407	size_t i;
1408	static struct sort_dimension avg_sort_dimension = {
1409		.name = "avg",
1410		.cmp  = avg_cmp,
1411	};
1412	static struct sort_dimension max_sort_dimension = {
1413		.name = "max",
1414		.cmp  = max_cmp,
1415	};
1416	static struct sort_dimension pid_sort_dimension = {
1417		.name = "pid",
1418		.cmp  = pid_cmp,
1419	};
1420	static struct sort_dimension runtime_sort_dimension = {
1421		.name = "runtime",
1422		.cmp  = runtime_cmp,
1423	};
1424	static struct sort_dimension switch_sort_dimension = {
1425		.name = "switch",
1426		.cmp  = switch_cmp,
1427	};
1428	struct sort_dimension *available_sorts[] = {
1429		&pid_sort_dimension,
1430		&avg_sort_dimension,
1431		&max_sort_dimension,
1432		&switch_sort_dimension,
1433		&runtime_sort_dimension,
1434	};
1435
1436	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1437		if (!strcmp(available_sorts[i]->name, tok)) {
1438			list_add_tail(&available_sorts[i]->list, list);
1439
1440			return 0;
1441		}
1442	}
1443
1444	return -1;
1445}
1446
1447static void perf_sched__sort_lat(struct perf_sched *sched)
1448{
1449	struct rb_node *node;
1450	struct rb_root *root = &sched->atom_root;
1451again:
1452	for (;;) {
1453		struct work_atoms *data;
1454		node = rb_first(root);
1455		if (!node)
1456			break;
1457
1458		rb_erase(node, root);
1459		data = rb_entry(node, struct work_atoms, node);
1460		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1461	}
1462	if (root == &sched->atom_root) {
1463		root = &sched->merged_atom_root;
1464		goto again;
1465	}
1466}
1467
1468static int process_sched_wakeup_event(struct perf_tool *tool,
1469				      struct perf_evsel *evsel,
1470				      struct perf_sample *sample,
1471				      struct machine *machine)
1472{
1473	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1474
1475	if (sched->tp_handler->wakeup_event)
1476		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1477
1478	return 0;
1479}
1480
1481union map_priv {
1482	void	*ptr;
1483	bool	 color;
1484};
1485
1486static bool thread__has_color(struct thread *thread)
1487{
1488	union map_priv priv = {
1489		.ptr = thread__priv(thread),
1490	};
1491
1492	return priv.color;
1493}
1494
1495static struct thread*
1496map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1497{
1498	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1499	union map_priv priv = {
1500		.color = false,
1501	};
1502
1503	if (!sched->map.color_pids || !thread || thread__priv(thread))
1504		return thread;
1505
1506	if (thread_map__has(sched->map.color_pids, tid))
1507		priv.color = true;
1508
1509	thread__set_priv(thread, priv.ptr);
1510	return thread;
1511}
1512
1513static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1514			    struct perf_sample *sample, struct machine *machine)
1515{
1516	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1517	struct thread *sched_in;
1518	struct thread_runtime *tr;
1519	int new_shortname;
1520	u64 timestamp0, timestamp = sample->time;
1521	s64 delta;
1522	int i, this_cpu = sample->cpu;
1523	int cpus_nr;
1524	bool new_cpu = false;
1525	const char *color = PERF_COLOR_NORMAL;
1526	char stimestamp[32];
1527
1528	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1529
1530	if (this_cpu > sched->max_cpu)
1531		sched->max_cpu = this_cpu;
1532
1533	if (sched->map.comp) {
1534		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1535		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1536			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1537			new_cpu = true;
1538		}
1539	} else
1540		cpus_nr = sched->max_cpu;
1541
1542	timestamp0 = sched->cpu_last_switched[this_cpu];
1543	sched->cpu_last_switched[this_cpu] = timestamp;
1544	if (timestamp0)
1545		delta = timestamp - timestamp0;
1546	else
1547		delta = 0;
1548
1549	if (delta < 0) {
1550		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1551		return -1;
1552	}
1553
1554	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1555	if (sched_in == NULL)
1556		return -1;
1557
1558	tr = thread__get_runtime(sched_in);
1559	if (tr == NULL) {
1560		thread__put(sched_in);
1561		return -1;
1562	}
1563
1564	sched->curr_thread[this_cpu] = thread__get(sched_in);
1565
1566	printf("  ");
1567
1568	new_shortname = 0;
1569	if (!tr->shortname[0]) {
1570		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1571			/*
1572			 * Don't allocate a letter-number for swapper:0
1573			 * as a shortname. Instead, we use '.' for it.
1574			 */
1575			tr->shortname[0] = '.';
1576			tr->shortname[1] = ' ';
1577		} else {
1578			tr->shortname[0] = sched->next_shortname1;
1579			tr->shortname[1] = sched->next_shortname2;
1580
1581			if (sched->next_shortname1 < 'Z') {
1582				sched->next_shortname1++;
1583			} else {
1584				sched->next_shortname1 = 'A';
1585				if (sched->next_shortname2 < '9')
1586					sched->next_shortname2++;
1587				else
1588					sched->next_shortname2 = '0';
1589			}
1590		}
1591		new_shortname = 1;
1592	}
1593
1594	for (i = 0; i < cpus_nr; i++) {
1595		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1596		struct thread *curr_thread = sched->curr_thread[cpu];
1597		struct thread_runtime *curr_tr;
1598		const char *pid_color = color;
1599		const char *cpu_color = color;
1600
1601		if (curr_thread && thread__has_color(curr_thread))
1602			pid_color = COLOR_PIDS;
1603
1604		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1605			continue;
1606
1607		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1608			cpu_color = COLOR_CPUS;
1609
1610		if (cpu != this_cpu)
1611			color_fprintf(stdout, color, " ");
1612		else
1613			color_fprintf(stdout, cpu_color, "*");
1614
1615		if (sched->curr_thread[cpu]) {
1616			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1617			if (curr_tr == NULL) {
1618				thread__put(sched_in);
1619				return -1;
1620			}
1621			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1622		} else
1623			color_fprintf(stdout, color, "   ");
1624	}
1625
1626	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1627		goto out;
1628
1629	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1630	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1631	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1632		const char *pid_color = color;
1633
1634		if (thread__has_color(sched_in))
1635			pid_color = COLOR_PIDS;
1636
1637		color_fprintf(stdout, pid_color, "%s => %s:%d",
1638		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1639		tr->comm_changed = false;
1640	}
1641
1642	if (sched->map.comp && new_cpu)
1643		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1644
1645out:
1646	color_fprintf(stdout, color, "\n");
1647
1648	thread__put(sched_in);
1649
1650	return 0;
1651}
1652
1653static int process_sched_switch_event(struct perf_tool *tool,
1654				      struct perf_evsel *evsel,
1655				      struct perf_sample *sample,
1656				      struct machine *machine)
1657{
1658	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1659	int this_cpu = sample->cpu, err = 0;
1660	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1661	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1662
1663	if (sched->curr_pid[this_cpu] != (u32)-1) {
1664		/*
1665		 * Are we trying to switch away a PID that is
1666		 * not current?
1667		 */
1668		if (sched->curr_pid[this_cpu] != prev_pid)
1669			sched->nr_context_switch_bugs++;
1670	}
1671
1672	if (sched->tp_handler->switch_event)
1673		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1674
1675	sched->curr_pid[this_cpu] = next_pid;
1676	return err;
1677}
1678
1679static int process_sched_runtime_event(struct perf_tool *tool,
1680				       struct perf_evsel *evsel,
1681				       struct perf_sample *sample,
1682				       struct machine *machine)
1683{
1684	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1685
1686	if (sched->tp_handler->runtime_event)
1687		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1688
1689	return 0;
1690}
1691
1692static int perf_sched__process_fork_event(struct perf_tool *tool,
1693					  union perf_event *event,
1694					  struct perf_sample *sample,
1695					  struct machine *machine)
1696{
1697	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1698
1699	/* run the fork event through the perf machineruy */
1700	perf_event__process_fork(tool, event, sample, machine);
1701
1702	/* and then run additional processing needed for this command */
1703	if (sched->tp_handler->fork_event)
1704		return sched->tp_handler->fork_event(sched, event, machine);
1705
1706	return 0;
1707}
1708
1709static int process_sched_migrate_task_event(struct perf_tool *tool,
1710					    struct perf_evsel *evsel,
1711					    struct perf_sample *sample,
1712					    struct machine *machine)
1713{
1714	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1715
1716	if (sched->tp_handler->migrate_task_event)
1717		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1718
1719	return 0;
1720}
1721
1722typedef int (*tracepoint_handler)(struct perf_tool *tool,
1723				  struct perf_evsel *evsel,
1724				  struct perf_sample *sample,
1725				  struct machine *machine);
1726
1727static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1728						 union perf_event *event __maybe_unused,
1729						 struct perf_sample *sample,
1730						 struct perf_evsel *evsel,
1731						 struct machine *machine)
1732{
1733	int err = 0;
1734
1735	if (evsel->handler != NULL) {
1736		tracepoint_handler f = evsel->handler;
1737		err = f(tool, evsel, sample, machine);
1738	}
1739
1740	return err;
1741}
1742
1743static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1744				    union perf_event *event,
1745				    struct perf_sample *sample,
1746				    struct machine *machine)
1747{
1748	struct thread *thread;
1749	struct thread_runtime *tr;
1750	int err;
1751
1752	err = perf_event__process_comm(tool, event, sample, machine);
1753	if (err)
1754		return err;
1755
1756	thread = machine__find_thread(machine, sample->pid, sample->tid);
1757	if (!thread) {
1758		pr_err("Internal error: can't find thread\n");
1759		return -1;
1760	}
1761
1762	tr = thread__get_runtime(thread);
1763	if (tr == NULL) {
1764		thread__put(thread);
1765		return -1;
1766	}
1767
1768	tr->comm_changed = true;
1769	thread__put(thread);
1770
1771	return 0;
1772}
1773
1774static int perf_sched__read_events(struct perf_sched *sched)
1775{
1776	const struct perf_evsel_str_handler handlers[] = {
1777		{ "sched:sched_switch",	      process_sched_switch_event, },
1778		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1779		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1780		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1781		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1782	};
1783	struct perf_session *session;
1784	struct perf_data data = {
1785		.file      = {
1786			.path = input_name,
1787		},
1788		.mode      = PERF_DATA_MODE_READ,
1789		.force     = sched->force,
1790	};
1791	int rc = -1;
1792
1793	session = perf_session__new(&data, false, &sched->tool);
1794	if (session == NULL) {
1795		pr_debug("No Memory for session\n");
1796		return -1;
1797	}
1798
1799	symbol__init(&session->header.env);
1800
1801	if (perf_session__set_tracepoints_handlers(session, handlers))
1802		goto out_delete;
1803
1804	if (perf_session__has_traces(session, "record -R")) {
1805		int err = perf_session__process_events(session);
1806		if (err) {
1807			pr_err("Failed to process events, error %d", err);
1808			goto out_delete;
1809		}
1810
1811		sched->nr_events      = session->evlist->stats.nr_events[0];
1812		sched->nr_lost_events = session->evlist->stats.total_lost;
1813		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1814	}
1815
1816	rc = 0;
1817out_delete:
1818	perf_session__delete(session);
1819	return rc;
1820}
1821
1822/*
1823 * scheduling times are printed as msec.usec
1824 */
1825static inline void print_sched_time(unsigned long long nsecs, int width)
1826{
1827	unsigned long msecs;
1828	unsigned long usecs;
1829
1830	msecs  = nsecs / NSEC_PER_MSEC;
1831	nsecs -= msecs * NSEC_PER_MSEC;
1832	usecs  = nsecs / NSEC_PER_USEC;
1833	printf("%*lu.%03lu ", width, msecs, usecs);
1834}
1835
1836/*
1837 * returns runtime data for event, allocating memory for it the
1838 * first time it is used.
1839 */
1840static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1841{
1842	struct evsel_runtime *r = evsel->priv;
1843
1844	if (r == NULL) {
1845		r = zalloc(sizeof(struct evsel_runtime));
1846		evsel->priv = r;
1847	}
1848
1849	return r;
1850}
1851
1852/*
1853 * save last time event was seen per cpu
1854 */
1855static void perf_evsel__save_time(struct perf_evsel *evsel,
1856				  u64 timestamp, u32 cpu)
1857{
1858	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1859
1860	if (r == NULL)
1861		return;
1862
1863	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1864		int i, n = __roundup_pow_of_two(cpu+1);
1865		void *p = r->last_time;
1866
1867		p = realloc(r->last_time, n * sizeof(u64));
1868		if (!p)
1869			return;
1870
1871		r->last_time = p;
1872		for (i = r->ncpu; i < n; ++i)
1873			r->last_time[i] = (u64) 0;
1874
1875		r->ncpu = n;
1876	}
1877
1878	r->last_time[cpu] = timestamp;
1879}
1880
1881/* returns last time this event was seen on the given cpu */
1882static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1883{
1884	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1885
1886	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1887		return 0;
1888
1889	return r->last_time[cpu];
1890}
1891
1892static int comm_width = 30;
1893
1894static char *timehist_get_commstr(struct thread *thread)
1895{
1896	static char str[32];
1897	const char *comm = thread__comm_str(thread);
1898	pid_t tid = thread->tid;
1899	pid_t pid = thread->pid_;
1900	int n;
1901
1902	if (pid == 0)
1903		n = scnprintf(str, sizeof(str), "%s", comm);
1904
1905	else if (tid != pid)
1906		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1907
1908	else
1909		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1910
1911	if (n > comm_width)
1912		comm_width = n;
1913
1914	return str;
1915}
1916
1917static void timehist_header(struct perf_sched *sched)
1918{
1919	u32 ncpus = sched->max_cpu + 1;
1920	u32 i, j;
1921
1922	printf("%15s %6s ", "time", "cpu");
1923
1924	if (sched->show_cpu_visual) {
1925		printf(" ");
1926		for (i = 0, j = 0; i < ncpus; ++i) {
1927			printf("%x", j++);
1928			if (j > 15)
1929				j = 0;
1930		}
1931		printf(" ");
1932	}
1933
1934	printf(" %-*s  %9s  %9s  %9s", comm_width,
1935		"task name", "wait time", "sch delay", "run time");
1936
1937	if (sched->show_state)
1938		printf("  %s", "state");
1939
1940	printf("\n");
1941
1942	/*
1943	 * units row
1944	 */
1945	printf("%15s %-6s ", "", "");
1946
1947	if (sched->show_cpu_visual)
1948		printf(" %*s ", ncpus, "");
1949
1950	printf(" %-*s  %9s  %9s  %9s", comm_width,
1951	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1952
1953	if (sched->show_state)
1954		printf("  %5s", "");
1955
1956	printf("\n");
1957
1958	/*
1959	 * separator
1960	 */
1961	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1962
1963	if (sched->show_cpu_visual)
1964		printf(" %.*s ", ncpus, graph_dotted_line);
1965
1966	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1967		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1968		graph_dotted_line);
1969
1970	if (sched->show_state)
1971		printf("  %.5s", graph_dotted_line);
1972
1973	printf("\n");
1974}
1975
1976static char task_state_char(struct thread *thread, int state)
1977{
1978	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1979	unsigned bit = state ? ffs(state) : 0;
1980
1981	/* 'I' for idle */
1982	if (thread->tid == 0)
1983		return 'I';
1984
1985	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1986}
1987
1988static void timehist_print_sample(struct perf_sched *sched,
1989				  struct perf_evsel *evsel,
1990				  struct perf_sample *sample,
1991				  struct addr_location *al,
1992				  struct thread *thread,
1993				  u64 t, int state)
1994{
1995	struct thread_runtime *tr = thread__priv(thread);
1996	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1997	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1998	u32 max_cpus = sched->max_cpu + 1;
1999	char tstr[64];
2000	char nstr[30];
2001	u64 wait_time;
2002
 
 
 
2003	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2004	printf("%15s [%04d] ", tstr, sample->cpu);
2005
2006	if (sched->show_cpu_visual) {
2007		u32 i;
2008		char c;
2009
2010		printf(" ");
2011		for (i = 0; i < max_cpus; ++i) {
2012			/* flag idle times with 'i'; others are sched events */
2013			if (i == sample->cpu)
2014				c = (thread->tid == 0) ? 'i' : 's';
2015			else
2016				c = ' ';
2017			printf("%c", c);
2018		}
2019		printf(" ");
2020	}
2021
2022	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2023
2024	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2025	print_sched_time(wait_time, 6);
2026
2027	print_sched_time(tr->dt_delay, 6);
2028	print_sched_time(tr->dt_run, 6);
2029
2030	if (sched->show_state)
2031		printf(" %5c ", task_state_char(thread, state));
2032
2033	if (sched->show_next) {
2034		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2035		printf(" %-*s", comm_width, nstr);
2036	}
2037
2038	if (sched->show_wakeups && !sched->show_next)
2039		printf("  %-*s", comm_width, "");
2040
2041	if (thread->tid == 0)
2042		goto out;
2043
2044	if (sched->show_callchain)
2045		printf("  ");
2046
2047	sample__fprintf_sym(sample, al, 0,
2048			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2049			    EVSEL__PRINT_CALLCHAIN_ARROW |
2050			    EVSEL__PRINT_SKIP_IGNORED,
2051			    &callchain_cursor, stdout);
2052
2053out:
2054	printf("\n");
2055}
2056
2057/*
2058 * Explanation of delta-time stats:
2059 *
2060 *            t = time of current schedule out event
2061 *        tprev = time of previous sched out event
2062 *                also time of schedule-in event for current task
2063 *    last_time = time of last sched change event for current task
2064 *                (i.e, time process was last scheduled out)
2065 * ready_to_run = time of wakeup for current task
2066 *
2067 * -----|------------|------------|------------|------
2068 *    last         ready        tprev          t
2069 *    time         to run
2070 *
2071 *      |-------- dt_wait --------|
2072 *                   |- dt_delay -|-- dt_run --|
2073 *
2074 *   dt_run = run time of current task
2075 *  dt_wait = time between last schedule out event for task and tprev
2076 *            represents time spent off the cpu
2077 * dt_delay = time between wakeup and schedule-in of task
2078 */
2079
2080static void timehist_update_runtime_stats(struct thread_runtime *r,
2081					 u64 t, u64 tprev)
2082{
2083	r->dt_delay   = 0;
2084	r->dt_sleep   = 0;
2085	r->dt_iowait  = 0;
2086	r->dt_preempt = 0;
2087	r->dt_run     = 0;
2088
2089	if (tprev) {
2090		r->dt_run = t - tprev;
2091		if (r->ready_to_run) {
2092			if (r->ready_to_run > tprev)
2093				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2094			else
2095				r->dt_delay = tprev - r->ready_to_run;
2096		}
2097
2098		if (r->last_time > tprev)
2099			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2100		else if (r->last_time) {
2101			u64 dt_wait = tprev - r->last_time;
2102
2103			if (r->last_state == TASK_RUNNING)
2104				r->dt_preempt = dt_wait;
2105			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2106				r->dt_iowait = dt_wait;
2107			else
2108				r->dt_sleep = dt_wait;
2109		}
2110	}
2111
2112	update_stats(&r->run_stats, r->dt_run);
2113
2114	r->total_run_time     += r->dt_run;
2115	r->total_delay_time   += r->dt_delay;
2116	r->total_sleep_time   += r->dt_sleep;
2117	r->total_iowait_time  += r->dt_iowait;
2118	r->total_preempt_time += r->dt_preempt;
2119}
2120
2121static bool is_idle_sample(struct perf_sample *sample,
2122			   struct perf_evsel *evsel)
2123{
2124	/* pid 0 == swapper == idle task */
2125	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2126		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2127
2128	return sample->pid == 0;
2129}
2130
2131static void save_task_callchain(struct perf_sched *sched,
2132				struct perf_sample *sample,
2133				struct perf_evsel *evsel,
2134				struct machine *machine)
2135{
2136	struct callchain_cursor *cursor = &callchain_cursor;
2137	struct thread *thread;
2138
2139	/* want main thread for process - has maps */
2140	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2141	if (thread == NULL) {
2142		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2143		return;
2144	}
2145
2146	if (!symbol_conf.use_callchain || sample->callchain == NULL)
2147		return;
2148
2149	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2150				      NULL, NULL, sched->max_stack + 2) != 0) {
2151		if (verbose > 0)
2152			pr_err("Failed to resolve callchain. Skipping\n");
2153
2154		return;
2155	}
2156
2157	callchain_cursor_commit(cursor);
2158
2159	while (true) {
2160		struct callchain_cursor_node *node;
2161		struct symbol *sym;
2162
2163		node = callchain_cursor_current(cursor);
2164		if (node == NULL)
2165			break;
2166
2167		sym = node->sym;
2168		if (sym) {
2169			if (!strcmp(sym->name, "schedule") ||
2170			    !strcmp(sym->name, "__schedule") ||
2171			    !strcmp(sym->name, "preempt_schedule"))
2172				sym->ignore = 1;
2173		}
2174
2175		callchain_cursor_advance(cursor);
2176	}
2177}
2178
2179static int init_idle_thread(struct thread *thread)
2180{
2181	struct idle_thread_runtime *itr;
2182
2183	thread__set_comm(thread, idle_comm, 0);
2184
2185	itr = zalloc(sizeof(*itr));
2186	if (itr == NULL)
2187		return -ENOMEM;
2188
2189	init_stats(&itr->tr.run_stats);
2190	callchain_init(&itr->callchain);
2191	callchain_cursor_reset(&itr->cursor);
2192	thread__set_priv(thread, itr);
2193
2194	return 0;
2195}
2196
2197/*
2198 * Track idle stats per cpu by maintaining a local thread
2199 * struct for the idle task on each cpu.
2200 */
2201static int init_idle_threads(int ncpu)
2202{
2203	int i, ret;
2204
2205	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2206	if (!idle_threads)
2207		return -ENOMEM;
2208
2209	idle_max_cpu = ncpu;
2210
2211	/* allocate the actual thread struct if needed */
2212	for (i = 0; i < ncpu; ++i) {
2213		idle_threads[i] = thread__new(0, 0);
2214		if (idle_threads[i] == NULL)
2215			return -ENOMEM;
2216
2217		ret = init_idle_thread(idle_threads[i]);
2218		if (ret < 0)
2219			return ret;
2220	}
2221
2222	return 0;
2223}
2224
2225static void free_idle_threads(void)
2226{
2227	int i;
2228
2229	if (idle_threads == NULL)
2230		return;
2231
2232	for (i = 0; i < idle_max_cpu; ++i) {
2233		if ((idle_threads[i]))
2234			thread__delete(idle_threads[i]);
2235	}
2236
2237	free(idle_threads);
2238}
2239
2240static struct thread *get_idle_thread(int cpu)
2241{
2242	/*
2243	 * expand/allocate array of pointers to local thread
2244	 * structs if needed
2245	 */
2246	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2247		int i, j = __roundup_pow_of_two(cpu+1);
2248		void *p;
2249
2250		p = realloc(idle_threads, j * sizeof(struct thread *));
2251		if (!p)
2252			return NULL;
2253
2254		idle_threads = (struct thread **) p;
2255		for (i = idle_max_cpu; i < j; ++i)
2256			idle_threads[i] = NULL;
2257
2258		idle_max_cpu = j;
2259	}
2260
2261	/* allocate a new thread struct if needed */
2262	if (idle_threads[cpu] == NULL) {
2263		idle_threads[cpu] = thread__new(0, 0);
2264
2265		if (idle_threads[cpu]) {
2266			if (init_idle_thread(idle_threads[cpu]) < 0)
2267				return NULL;
2268		}
2269	}
2270
2271	return idle_threads[cpu];
2272}
2273
2274static void save_idle_callchain(struct idle_thread_runtime *itr,
 
2275				struct perf_sample *sample)
2276{
2277	if (!symbol_conf.use_callchain || sample->callchain == NULL)
2278		return;
2279
2280	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2281}
2282
2283static struct thread *timehist_get_thread(struct perf_sched *sched,
2284					  struct perf_sample *sample,
2285					  struct machine *machine,
2286					  struct perf_evsel *evsel)
2287{
2288	struct thread *thread;
2289
2290	if (is_idle_sample(sample, evsel)) {
2291		thread = get_idle_thread(sample->cpu);
2292		if (thread == NULL)
2293			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2294
2295	} else {
2296		/* there were samples with tid 0 but non-zero pid */
2297		thread = machine__findnew_thread(machine, sample->pid,
2298						 sample->tid ?: sample->pid);
2299		if (thread == NULL) {
2300			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2301				 sample->tid);
2302		}
2303
2304		save_task_callchain(sched, sample, evsel, machine);
2305		if (sched->idle_hist) {
2306			struct thread *idle;
2307			struct idle_thread_runtime *itr;
2308
2309			idle = get_idle_thread(sample->cpu);
2310			if (idle == NULL) {
2311				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2312				return NULL;
2313			}
2314
2315			itr = thread__priv(idle);
2316			if (itr == NULL)
2317				return NULL;
2318
2319			itr->last_thread = thread;
2320
2321			/* copy task callchain when entering to idle */
2322			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2323				save_idle_callchain(itr, sample);
2324		}
2325	}
2326
2327	return thread;
2328}
2329
2330static bool timehist_skip_sample(struct perf_sched *sched,
2331				 struct thread *thread,
2332				 struct perf_evsel *evsel,
2333				 struct perf_sample *sample)
2334{
2335	bool rc = false;
2336
2337	if (thread__is_filtered(thread)) {
2338		rc = true;
2339		sched->skipped_samples++;
2340	}
2341
2342	if (sched->idle_hist) {
2343		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2344			rc = true;
2345		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2346			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2347			rc = true;
2348	}
2349
2350	return rc;
2351}
2352
2353static void timehist_print_wakeup_event(struct perf_sched *sched,
2354					struct perf_evsel *evsel,
2355					struct perf_sample *sample,
2356					struct machine *machine,
2357					struct thread *awakened)
2358{
2359	struct thread *thread;
2360	char tstr[64];
2361
2362	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2363	if (thread == NULL)
2364		return;
2365
2366	/* show wakeup unless both awakee and awaker are filtered */
2367	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2368	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2369		return;
2370	}
2371
2372	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2373	printf("%15s [%04d] ", tstr, sample->cpu);
2374	if (sched->show_cpu_visual)
2375		printf(" %*s ", sched->max_cpu + 1, "");
2376
2377	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2378
2379	/* dt spacer */
2380	printf("  %9s  %9s  %9s ", "", "", "");
2381
2382	printf("awakened: %s", timehist_get_commstr(awakened));
2383
2384	printf("\n");
2385}
2386
 
 
 
 
 
 
 
 
 
2387static int timehist_sched_wakeup_event(struct perf_tool *tool,
2388				       union perf_event *event __maybe_unused,
2389				       struct perf_evsel *evsel,
2390				       struct perf_sample *sample,
2391				       struct machine *machine)
2392{
2393	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2394	struct thread *thread;
2395	struct thread_runtime *tr = NULL;
2396	/* want pid of awakened task not pid in sample */
2397	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2398
2399	thread = machine__findnew_thread(machine, 0, pid);
2400	if (thread == NULL)
2401		return -1;
2402
2403	tr = thread__get_runtime(thread);
2404	if (tr == NULL)
2405		return -1;
2406
2407	if (tr->ready_to_run == 0)
2408		tr->ready_to_run = sample->time;
2409
2410	/* show wakeups if requested */
2411	if (sched->show_wakeups &&
2412	    !perf_time__skip_sample(&sched->ptime, sample->time))
2413		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2414
2415	return 0;
2416}
2417
2418static void timehist_print_migration_event(struct perf_sched *sched,
2419					struct perf_evsel *evsel,
2420					struct perf_sample *sample,
2421					struct machine *machine,
2422					struct thread *migrated)
2423{
2424	struct thread *thread;
2425	char tstr[64];
2426	u32 max_cpus = sched->max_cpu + 1;
2427	u32 ocpu, dcpu;
2428
2429	if (sched->summary_only)
2430		return;
2431
2432	max_cpus = sched->max_cpu + 1;
2433	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2434	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2435
2436	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2437	if (thread == NULL)
2438		return;
2439
2440	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2441	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2442		return;
2443	}
2444
2445	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2446	printf("%15s [%04d] ", tstr, sample->cpu);
2447
2448	if (sched->show_cpu_visual) {
2449		u32 i;
2450		char c;
2451
2452		printf("  ");
2453		for (i = 0; i < max_cpus; ++i) {
2454			c = (i == sample->cpu) ? 'm' : ' ';
2455			printf("%c", c);
2456		}
2457		printf("  ");
2458	}
2459
2460	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2461
2462	/* dt spacer */
2463	printf("  %9s  %9s  %9s ", "", "", "");
2464
2465	printf("migrated: %s", timehist_get_commstr(migrated));
2466	printf(" cpu %d => %d", ocpu, dcpu);
2467
2468	printf("\n");
2469}
2470
2471static int timehist_migrate_task_event(struct perf_tool *tool,
2472				       union perf_event *event __maybe_unused,
2473				       struct perf_evsel *evsel,
2474				       struct perf_sample *sample,
2475				       struct machine *machine)
2476{
2477	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2478	struct thread *thread;
2479	struct thread_runtime *tr = NULL;
2480	/* want pid of migrated task not pid in sample */
2481	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2482
2483	thread = machine__findnew_thread(machine, 0, pid);
2484	if (thread == NULL)
2485		return -1;
2486
2487	tr = thread__get_runtime(thread);
2488	if (tr == NULL)
2489		return -1;
2490
2491	tr->migrations++;
2492
2493	/* show migrations if requested */
2494	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2495
2496	return 0;
2497}
2498
2499static int timehist_sched_change_event(struct perf_tool *tool,
2500				       union perf_event *event,
2501				       struct perf_evsel *evsel,
2502				       struct perf_sample *sample,
2503				       struct machine *machine)
2504{
2505	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2506	struct perf_time_interval *ptime = &sched->ptime;
2507	struct addr_location al;
2508	struct thread *thread;
2509	struct thread_runtime *tr = NULL;
2510	u64 tprev, t = sample->time;
2511	int rc = 0;
2512	int state = perf_evsel__intval(evsel, sample, "prev_state");
2513
2514
2515	if (machine__resolve(machine, &al, sample) < 0) {
2516		pr_err("problem processing %d event. skipping it\n",
2517		       event->header.type);
2518		rc = -1;
2519		goto out;
2520	}
2521
2522	thread = timehist_get_thread(sched, sample, machine, evsel);
2523	if (thread == NULL) {
2524		rc = -1;
2525		goto out;
2526	}
2527
2528	if (timehist_skip_sample(sched, thread, evsel, sample))
2529		goto out;
2530
2531	tr = thread__get_runtime(thread);
2532	if (tr == NULL) {
2533		rc = -1;
2534		goto out;
2535	}
2536
2537	tprev = perf_evsel__get_time(evsel, sample->cpu);
2538
2539	/*
2540	 * If start time given:
2541	 * - sample time is under window user cares about - skip sample
2542	 * - tprev is under window user cares about  - reset to start of window
2543	 */
2544	if (ptime->start && ptime->start > t)
2545		goto out;
2546
2547	if (tprev && ptime->start > tprev)
2548		tprev = ptime->start;
2549
2550	/*
2551	 * If end time given:
2552	 * - previous sched event is out of window - we are done
2553	 * - sample time is beyond window user cares about - reset it
2554	 *   to close out stats for time window interest
2555	 */
2556	if (ptime->end) {
2557		if (tprev > ptime->end)
2558			goto out;
2559
2560		if (t > ptime->end)
2561			t = ptime->end;
2562	}
2563
2564	if (!sched->idle_hist || thread->tid == 0) {
2565		timehist_update_runtime_stats(tr, t, tprev);
 
2566
2567		if (sched->idle_hist) {
2568			struct idle_thread_runtime *itr = (void *)tr;
2569			struct thread_runtime *last_tr;
2570
2571			BUG_ON(thread->tid != 0);
2572
2573			if (itr->last_thread == NULL)
2574				goto out;
2575
2576			/* add current idle time as last thread's runtime */
2577			last_tr = thread__get_runtime(itr->last_thread);
2578			if (last_tr == NULL)
2579				goto out;
2580
2581			timehist_update_runtime_stats(last_tr, t, tprev);
2582			/*
2583			 * remove delta time of last thread as it's not updated
2584			 * and otherwise it will show an invalid value next
2585			 * time.  we only care total run time and run stat.
2586			 */
2587			last_tr->dt_run = 0;
2588			last_tr->dt_delay = 0;
2589			last_tr->dt_sleep = 0;
2590			last_tr->dt_iowait = 0;
2591			last_tr->dt_preempt = 0;
2592
2593			if (itr->cursor.nr)
2594				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2595
2596			itr->last_thread = NULL;
2597		}
2598	}
2599
2600	if (!sched->summary_only)
2601		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2602
2603out:
2604	if (sched->hist_time.start == 0 && t >= ptime->start)
2605		sched->hist_time.start = t;
2606	if (ptime->end == 0 || t <= ptime->end)
2607		sched->hist_time.end = t;
2608
2609	if (tr) {
2610		/* time of this sched_switch event becomes last time task seen */
2611		tr->last_time = sample->time;
2612
2613		/* last state is used to determine where to account wait time */
2614		tr->last_state = state;
2615
2616		/* sched out event for task so reset ready to run time */
2617		tr->ready_to_run = 0;
2618	}
2619
2620	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2621
2622	return rc;
2623}
2624
2625static int timehist_sched_switch_event(struct perf_tool *tool,
2626			     union perf_event *event,
2627			     struct perf_evsel *evsel,
2628			     struct perf_sample *sample,
2629			     struct machine *machine __maybe_unused)
2630{
2631	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2632}
2633
2634static int process_lost(struct perf_tool *tool __maybe_unused,
2635			union perf_event *event,
2636			struct perf_sample *sample,
2637			struct machine *machine __maybe_unused)
2638{
2639	char tstr[64];
2640
2641	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2642	printf("%15s ", tstr);
2643	printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2644
2645	return 0;
2646}
2647
2648
2649static void print_thread_runtime(struct thread *t,
2650				 struct thread_runtime *r)
2651{
2652	double mean = avg_stats(&r->run_stats);
2653	float stddev;
2654
2655	printf("%*s   %5d  %9" PRIu64 " ",
2656	       comm_width, timehist_get_commstr(t), t->ppid,
2657	       (u64) r->run_stats.n);
2658
2659	print_sched_time(r->total_run_time, 8);
2660	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2661	print_sched_time(r->run_stats.min, 6);
2662	printf(" ");
2663	print_sched_time((u64) mean, 6);
2664	printf(" ");
2665	print_sched_time(r->run_stats.max, 6);
2666	printf("  ");
2667	printf("%5.2f", stddev);
2668	printf("   %5" PRIu64, r->migrations);
2669	printf("\n");
2670}
2671
2672static void print_thread_waittime(struct thread *t,
2673				  struct thread_runtime *r)
2674{
2675	printf("%*s   %5d  %9" PRIu64 " ",
2676	       comm_width, timehist_get_commstr(t), t->ppid,
2677	       (u64) r->run_stats.n);
2678
2679	print_sched_time(r->total_run_time, 8);
2680	print_sched_time(r->total_sleep_time, 6);
2681	printf(" ");
2682	print_sched_time(r->total_iowait_time, 6);
2683	printf(" ");
2684	print_sched_time(r->total_preempt_time, 6);
2685	printf(" ");
2686	print_sched_time(r->total_delay_time, 6);
2687	printf("\n");
2688}
2689
2690struct total_run_stats {
2691	struct perf_sched *sched;
2692	u64  sched_count;
2693	u64  task_count;
2694	u64  total_run_time;
2695};
2696
2697static int __show_thread_runtime(struct thread *t, void *priv)
2698{
2699	struct total_run_stats *stats = priv;
2700	struct thread_runtime *r;
2701
2702	if (thread__is_filtered(t))
2703		return 0;
2704
2705	r = thread__priv(t);
2706	if (r && r->run_stats.n) {
2707		stats->task_count++;
2708		stats->sched_count += r->run_stats.n;
2709		stats->total_run_time += r->total_run_time;
2710
2711		if (stats->sched->show_state)
2712			print_thread_waittime(t, r);
2713		else
2714			print_thread_runtime(t, r);
2715	}
2716
2717	return 0;
2718}
2719
2720static int show_thread_runtime(struct thread *t, void *priv)
2721{
2722	if (t->dead)
2723		return 0;
2724
2725	return __show_thread_runtime(t, priv);
2726}
2727
2728static int show_deadthread_runtime(struct thread *t, void *priv)
2729{
2730	if (!t->dead)
2731		return 0;
2732
2733	return __show_thread_runtime(t, priv);
2734}
2735
2736static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2737{
2738	const char *sep = " <- ";
2739	struct callchain_list *chain;
2740	size_t ret = 0;
2741	char bf[1024];
2742	bool first;
2743
2744	if (node == NULL)
2745		return 0;
2746
2747	ret = callchain__fprintf_folded(fp, node->parent);
2748	first = (ret == 0);
2749
2750	list_for_each_entry(chain, &node->val, list) {
2751		if (chain->ip >= PERF_CONTEXT_MAX)
2752			continue;
2753		if (chain->ms.sym && chain->ms.sym->ignore)
2754			continue;
2755		ret += fprintf(fp, "%s%s", first ? "" : sep,
2756			       callchain_list__sym_name(chain, bf, sizeof(bf),
2757							false));
2758		first = false;
2759	}
2760
2761	return ret;
2762}
2763
2764static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2765{
2766	size_t ret = 0;
2767	FILE *fp = stdout;
2768	struct callchain_node *chain;
2769	struct rb_node *rb_node = rb_first(root);
2770
2771	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2772	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2773	       graph_dotted_line);
2774
2775	while (rb_node) {
2776		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2777		rb_node = rb_next(rb_node);
2778
2779		ret += fprintf(fp, "  ");
2780		print_sched_time(chain->hit, 12);
2781		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2782		ret += fprintf(fp, " %8d  ", chain->count);
2783		ret += callchain__fprintf_folded(fp, chain);
2784		ret += fprintf(fp, "\n");
2785	}
2786
2787	return ret;
2788}
2789
2790static void timehist_print_summary(struct perf_sched *sched,
2791				   struct perf_session *session)
2792{
2793	struct machine *m = &session->machines.host;
2794	struct total_run_stats totals;
2795	u64 task_count;
2796	struct thread *t;
2797	struct thread_runtime *r;
2798	int i;
2799	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2800
2801	memset(&totals, 0, sizeof(totals));
2802	totals.sched = sched;
2803
2804	if (sched->idle_hist) {
2805		printf("\nIdle-time summary\n");
2806		printf("%*s  parent  sched-out  ", comm_width, "comm");
2807		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2808	} else if (sched->show_state) {
2809		printf("\nWait-time summary\n");
2810		printf("%*s  parent   sched-in  ", comm_width, "comm");
2811		printf("   run-time      sleep      iowait     preempt       delay\n");
2812	} else {
2813		printf("\nRuntime summary\n");
2814		printf("%*s  parent   sched-in  ", comm_width, "comm");
2815		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2816	}
2817	printf("%*s            (count)  ", comm_width, "");
2818	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2819	       sched->show_state ? "(msec)" : "%");
2820	printf("%.117s\n", graph_dotted_line);
2821
2822	machine__for_each_thread(m, show_thread_runtime, &totals);
2823	task_count = totals.task_count;
2824	if (!task_count)
2825		printf("<no still running tasks>\n");
2826
2827	printf("\nTerminated tasks:\n");
2828	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2829	if (task_count == totals.task_count)
2830		printf("<no terminated tasks>\n");
2831
2832	/* CPU idle stats not tracked when samples were skipped */
2833	if (sched->skipped_samples && !sched->idle_hist)
2834		return;
2835
2836	printf("\nIdle stats:\n");
2837	for (i = 0; i < idle_max_cpu; ++i) {
 
 
 
2838		t = idle_threads[i];
2839		if (!t)
2840			continue;
2841
2842		r = thread__priv(t);
2843		if (r && r->run_stats.n) {
2844			totals.sched_count += r->run_stats.n;
2845			printf("    CPU %2d idle for ", i);
2846			print_sched_time(r->total_run_time, 6);
2847			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2848		} else
2849			printf("    CPU %2d idle entire time window\n", i);
2850	}
2851
2852	if (sched->idle_hist && symbol_conf.use_callchain) {
2853		callchain_param.mode  = CHAIN_FOLDED;
2854		callchain_param.value = CCVAL_PERIOD;
2855
2856		callchain_register_param(&callchain_param);
2857
2858		printf("\nIdle stats by callchain:\n");
2859		for (i = 0; i < idle_max_cpu; ++i) {
2860			struct idle_thread_runtime *itr;
2861
2862			t = idle_threads[i];
2863			if (!t)
2864				continue;
2865
2866			itr = thread__priv(t);
2867			if (itr == NULL)
2868				continue;
2869
2870			callchain_param.sort(&itr->sorted_root, &itr->callchain,
2871					     0, &callchain_param);
2872
2873			printf("  CPU %2d:", i);
2874			print_sched_time(itr->tr.total_run_time, 6);
2875			printf(" msec\n");
2876			timehist_print_idlehist_callchain(&itr->sorted_root);
2877			printf("\n");
2878		}
2879	}
2880
2881	printf("\n"
2882	       "    Total number of unique tasks: %" PRIu64 "\n"
2883	       "Total number of context switches: %" PRIu64 "\n",
2884	       totals.task_count, totals.sched_count);
2885
2886	printf("           Total run time (msec): ");
2887	print_sched_time(totals.total_run_time, 2);
2888	printf("\n");
2889
2890	printf("    Total scheduling time (msec): ");
2891	print_sched_time(hist_time, 2);
2892	printf(" (x %d)\n", sched->max_cpu);
2893}
2894
2895typedef int (*sched_handler)(struct perf_tool *tool,
2896			  union perf_event *event,
2897			  struct perf_evsel *evsel,
2898			  struct perf_sample *sample,
2899			  struct machine *machine);
2900
2901static int perf_timehist__process_sample(struct perf_tool *tool,
2902					 union perf_event *event,
2903					 struct perf_sample *sample,
2904					 struct perf_evsel *evsel,
2905					 struct machine *machine)
2906{
2907	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2908	int err = 0;
2909	int this_cpu = sample->cpu;
2910
2911	if (this_cpu > sched->max_cpu)
2912		sched->max_cpu = this_cpu;
2913
2914	if (evsel->handler != NULL) {
2915		sched_handler f = evsel->handler;
2916
2917		err = f(tool, event, evsel, sample, machine);
2918	}
2919
2920	return err;
2921}
2922
2923static int timehist_check_attr(struct perf_sched *sched,
2924			       struct perf_evlist *evlist)
2925{
2926	struct perf_evsel *evsel;
2927	struct evsel_runtime *er;
2928
2929	list_for_each_entry(evsel, &evlist->entries, node) {
2930		er = perf_evsel__get_runtime(evsel);
2931		if (er == NULL) {
2932			pr_err("Failed to allocate memory for evsel runtime data\n");
2933			return -1;
2934		}
2935
2936		if (sched->show_callchain &&
2937		    !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2938			pr_info("Samples do not have callchains.\n");
2939			sched->show_callchain = 0;
2940			symbol_conf.use_callchain = 0;
2941		}
2942	}
2943
2944	return 0;
2945}
2946
2947static int perf_sched__timehist(struct perf_sched *sched)
2948{
2949	const struct perf_evsel_str_handler handlers[] = {
2950		{ "sched:sched_switch",       timehist_sched_switch_event, },
2951		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
 
2952		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2953	};
2954	const struct perf_evsel_str_handler migrate_handlers[] = {
2955		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2956	};
2957	struct perf_data data = {
2958		.file      = {
2959			.path = input_name,
2960		},
2961		.mode      = PERF_DATA_MODE_READ,
2962		.force     = sched->force,
2963	};
2964
2965	struct perf_session *session;
2966	struct perf_evlist *evlist;
2967	int err = -1;
2968
2969	/*
2970	 * event handlers for timehist option
2971	 */
2972	sched->tool.sample	 = perf_timehist__process_sample;
2973	sched->tool.mmap	 = perf_event__process_mmap;
2974	sched->tool.comm	 = perf_event__process_comm;
2975	sched->tool.exit	 = perf_event__process_exit;
2976	sched->tool.fork	 = perf_event__process_fork;
2977	sched->tool.lost	 = process_lost;
2978	sched->tool.attr	 = perf_event__process_attr;
2979	sched->tool.tracing_data = perf_event__process_tracing_data;
2980	sched->tool.build_id	 = perf_event__process_build_id;
2981
2982	sched->tool.ordered_events = true;
2983	sched->tool.ordering_requires_timestamps = true;
2984
2985	symbol_conf.use_callchain = sched->show_callchain;
2986
2987	session = perf_session__new(&data, false, &sched->tool);
2988	if (session == NULL)
2989		return -ENOMEM;
 
 
 
 
 
 
2990
2991	evlist = session->evlist;
2992
2993	symbol__init(&session->header.env);
2994
2995	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2996		pr_err("Invalid time string\n");
2997		return -EINVAL;
2998	}
2999
3000	if (timehist_check_attr(sched, evlist) != 0)
3001		goto out;
3002
3003	setup_pager();
3004
 
 
 
 
 
3005	/* setup per-evsel handlers */
3006	if (perf_session__set_tracepoints_handlers(session, handlers))
3007		goto out;
3008
3009	/* sched_switch event at a minimum needs to exist */
3010	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3011						  "sched:sched_switch")) {
3012		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3013		goto out;
3014	}
3015
3016	if (sched->show_migrations &&
3017	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3018		goto out;
3019
3020	/* pre-allocate struct for per-CPU idle stats */
3021	sched->max_cpu = session->header.env.nr_cpus_online;
3022	if (sched->max_cpu == 0)
3023		sched->max_cpu = 4;
3024	if (init_idle_threads(sched->max_cpu))
3025		goto out;
3026
3027	/* summary_only implies summary option, but don't overwrite summary if set */
3028	if (sched->summary_only)
3029		sched->summary = sched->summary_only;
3030
3031	if (!sched->summary_only)
3032		timehist_header(sched);
3033
3034	err = perf_session__process_events(session);
3035	if (err) {
3036		pr_err("Failed to process events, error %d", err);
3037		goto out;
3038	}
3039
3040	sched->nr_events      = evlist->stats.nr_events[0];
3041	sched->nr_lost_events = evlist->stats.total_lost;
3042	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3043
3044	if (sched->summary)
3045		timehist_print_summary(sched, session);
3046
3047out:
3048	free_idle_threads();
3049	perf_session__delete(session);
3050
3051	return err;
3052}
3053
3054
3055static void print_bad_events(struct perf_sched *sched)
3056{
3057	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3058		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3059			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3060			sched->nr_unordered_timestamps, sched->nr_timestamps);
3061	}
3062	if (sched->nr_lost_events && sched->nr_events) {
3063		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3064			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3065			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3066	}
3067	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3068		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3069			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3070			sched->nr_context_switch_bugs, sched->nr_timestamps);
3071		if (sched->nr_lost_events)
3072			printf(" (due to lost events?)");
3073		printf("\n");
3074	}
3075}
3076
3077static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3078{
3079	struct rb_node **new = &(root->rb_node), *parent = NULL;
3080	struct work_atoms *this;
3081	const char *comm = thread__comm_str(data->thread), *this_comm;
 
3082
3083	while (*new) {
3084		int cmp;
3085
3086		this = container_of(*new, struct work_atoms, node);
3087		parent = *new;
3088
3089		this_comm = thread__comm_str(this->thread);
3090		cmp = strcmp(comm, this_comm);
3091		if (cmp > 0) {
3092			new = &((*new)->rb_left);
3093		} else if (cmp < 0) {
3094			new = &((*new)->rb_right);
 
3095		} else {
3096			this->num_merged++;
3097			this->total_runtime += data->total_runtime;
3098			this->nb_atoms += data->nb_atoms;
3099			this->total_lat += data->total_lat;
3100			list_splice(&data->work_list, &this->work_list);
3101			if (this->max_lat < data->max_lat) {
3102				this->max_lat = data->max_lat;
3103				this->max_lat_at = data->max_lat_at;
3104			}
3105			zfree(&data);
3106			return;
3107		}
3108	}
3109
3110	data->num_merged++;
3111	rb_link_node(&data->node, parent, new);
3112	rb_insert_color(&data->node, root);
3113}
3114
3115static void perf_sched__merge_lat(struct perf_sched *sched)
3116{
3117	struct work_atoms *data;
3118	struct rb_node *node;
3119
3120	if (sched->skip_merge)
3121		return;
3122
3123	while ((node = rb_first(&sched->atom_root))) {
3124		rb_erase(node, &sched->atom_root);
3125		data = rb_entry(node, struct work_atoms, node);
3126		__merge_work_atoms(&sched->merged_atom_root, data);
3127	}
3128}
3129
3130static int perf_sched__lat(struct perf_sched *sched)
3131{
3132	struct rb_node *next;
3133
3134	setup_pager();
3135
3136	if (perf_sched__read_events(sched))
3137		return -1;
3138
3139	perf_sched__merge_lat(sched);
3140	perf_sched__sort_lat(sched);
3141
3142	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3143	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3144	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3145
3146	next = rb_first(&sched->sorted_atom_root);
3147
3148	while (next) {
3149		struct work_atoms *work_list;
3150
3151		work_list = rb_entry(next, struct work_atoms, node);
3152		output_lat_thread(sched, work_list);
3153		next = rb_next(next);
3154		thread__zput(work_list->thread);
3155	}
3156
3157	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3158	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3159		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3160
3161	printf(" ---------------------------------------------------\n");
3162
3163	print_bad_events(sched);
3164	printf("\n");
3165
3166	return 0;
3167}
3168
3169static int setup_map_cpus(struct perf_sched *sched)
3170{
3171	struct cpu_map *map;
3172
3173	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3174
3175	if (sched->map.comp) {
3176		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3177		if (!sched->map.comp_cpus)
3178			return -1;
3179	}
3180
3181	if (!sched->map.cpus_str)
3182		return 0;
3183
3184	map = cpu_map__new(sched->map.cpus_str);
3185	if (!map) {
3186		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3187		return -1;
3188	}
3189
3190	sched->map.cpus = map;
3191	return 0;
3192}
3193
3194static int setup_color_pids(struct perf_sched *sched)
3195{
3196	struct thread_map *map;
3197
3198	if (!sched->map.color_pids_str)
3199		return 0;
3200
3201	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3202	if (!map) {
3203		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3204		return -1;
3205	}
3206
3207	sched->map.color_pids = map;
3208	return 0;
3209}
3210
3211static int setup_color_cpus(struct perf_sched *sched)
3212{
3213	struct cpu_map *map;
3214
3215	if (!sched->map.color_cpus_str)
3216		return 0;
3217
3218	map = cpu_map__new(sched->map.color_cpus_str);
3219	if (!map) {
3220		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3221		return -1;
3222	}
3223
3224	sched->map.color_cpus = map;
3225	return 0;
3226}
3227
3228static int perf_sched__map(struct perf_sched *sched)
3229{
3230	if (setup_map_cpus(sched))
3231		return -1;
3232
3233	if (setup_color_pids(sched))
3234		return -1;
3235
3236	if (setup_color_cpus(sched))
3237		return -1;
3238
3239	setup_pager();
3240	if (perf_sched__read_events(sched))
3241		return -1;
3242	print_bad_events(sched);
3243	return 0;
3244}
3245
3246static int perf_sched__replay(struct perf_sched *sched)
3247{
3248	unsigned long i;
3249
3250	calibrate_run_measurement_overhead(sched);
3251	calibrate_sleep_measurement_overhead(sched);
3252
3253	test_calibrations(sched);
3254
3255	if (perf_sched__read_events(sched))
3256		return -1;
3257
3258	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3259	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3260	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3261
3262	if (sched->targetless_wakeups)
3263		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3264	if (sched->multitarget_wakeups)
3265		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3266	if (sched->nr_run_events_optimized)
3267		printf("run atoms optimized: %ld\n",
3268			sched->nr_run_events_optimized);
3269
3270	print_task_traces(sched);
3271	add_cross_task_wakeups(sched);
3272
3273	create_tasks(sched);
3274	printf("------------------------------------------------------------\n");
3275	for (i = 0; i < sched->replay_repeat; i++)
3276		run_one_test(sched);
3277
3278	return 0;
3279}
3280
3281static void setup_sorting(struct perf_sched *sched, const struct option *options,
3282			  const char * const usage_msg[])
3283{
3284	char *tmp, *tok, *str = strdup(sched->sort_order);
3285
3286	for (tok = strtok_r(str, ", ", &tmp);
3287			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3288		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3289			usage_with_options_msg(usage_msg, options,
3290					"Unknown --sort key: `%s'", tok);
3291		}
3292	}
3293
3294	free(str);
3295
3296	sort_dimension__add("pid", &sched->cmp_pid);
3297}
3298
3299static int __cmd_record(int argc, const char **argv)
3300{
3301	unsigned int rec_argc, i, j;
3302	const char **rec_argv;
3303	const char * const record_args[] = {
3304		"record",
3305		"-a",
3306		"-R",
3307		"-m", "1024",
3308		"-c", "1",
3309		"-e", "sched:sched_switch",
3310		"-e", "sched:sched_stat_wait",
3311		"-e", "sched:sched_stat_sleep",
3312		"-e", "sched:sched_stat_iowait",
3313		"-e", "sched:sched_stat_runtime",
3314		"-e", "sched:sched_process_fork",
3315		"-e", "sched:sched_wakeup",
3316		"-e", "sched:sched_wakeup_new",
3317		"-e", "sched:sched_migrate_task",
3318	};
 
3319
3320	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 
 
 
 
3321	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3322
3323	if (rec_argv == NULL)
3324		return -ENOMEM;
3325
3326	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3327		rec_argv[i] = strdup(record_args[i]);
3328
 
 
 
 
 
 
 
3329	for (j = 1; j < (unsigned int)argc; j++, i++)
3330		rec_argv[i] = argv[j];
3331
3332	BUG_ON(i != rec_argc);
3333
3334	return cmd_record(i, rec_argv);
3335}
3336
3337int cmd_sched(int argc, const char **argv)
3338{
3339	const char default_sort_order[] = "avg, max, switch, runtime";
3340	struct perf_sched sched = {
3341		.tool = {
3342			.sample		 = perf_sched__process_tracepoint_sample,
3343			.comm		 = perf_sched__process_comm,
3344			.namespaces	 = perf_event__process_namespaces,
3345			.lost		 = perf_event__process_lost,
3346			.fork		 = perf_sched__process_fork_event,
3347			.ordered_events = true,
3348		},
3349		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3350		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3351		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3352		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3353		.sort_order	      = default_sort_order,
3354		.replay_repeat	      = 10,
3355		.profile_cpu	      = -1,
3356		.next_shortname1      = 'A',
3357		.next_shortname2      = '0',
3358		.skip_merge           = 0,
3359		.show_callchain	      = 1,
3360		.max_stack            = 5,
3361	};
3362	const struct option sched_options[] = {
3363	OPT_STRING('i', "input", &input_name, "file",
3364		    "input file name"),
3365	OPT_INCR('v', "verbose", &verbose,
3366		    "be more verbose (show symbol address, etc)"),
3367	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3368		    "dump raw trace in ASCII"),
3369	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3370	OPT_END()
3371	};
3372	const struct option latency_options[] = {
3373	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3374		   "sort by key(s): runtime, switch, avg, max"),
3375	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3376		    "CPU to profile on"),
3377	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3378		    "latency stats per pid instead of per comm"),
3379	OPT_PARENT(sched_options)
3380	};
3381	const struct option replay_options[] = {
3382	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3383		     "repeat the workload replay N times (-1: infinite)"),
3384	OPT_PARENT(sched_options)
3385	};
3386	const struct option map_options[] = {
3387	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3388		    "map output in compact mode"),
3389	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3390		   "highlight given pids in map"),
3391	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3392                    "highlight given CPUs in map"),
3393	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3394                    "display given CPUs in map"),
3395	OPT_PARENT(sched_options)
3396	};
3397	const struct option timehist_options[] = {
3398	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3399		   "file", "vmlinux pathname"),
3400	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3401		   "file", "kallsyms pathname"),
3402	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3403		    "Display call chains if present (default on)"),
3404	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3405		   "Maximum number of functions to display backtrace."),
3406	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3407		    "Look for files with symbols relative to this directory"),
3408	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3409		    "Show only syscall summary with statistics"),
3410	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3411		    "Show all syscalls and summary with statistics"),
3412	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3413	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3414	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3415	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3416	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3417	OPT_STRING(0, "time", &sched.time_str, "str",
3418		   "Time span for analysis (start,stop)"),
3419	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3420	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3421		   "analyze events only for given process id(s)"),
3422	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3423		   "analyze events only for given thread id(s)"),
 
3424	OPT_PARENT(sched_options)
3425	};
3426
3427	const char * const latency_usage[] = {
3428		"perf sched latency [<options>]",
3429		NULL
3430	};
3431	const char * const replay_usage[] = {
3432		"perf sched replay [<options>]",
3433		NULL
3434	};
3435	const char * const map_usage[] = {
3436		"perf sched map [<options>]",
3437		NULL
3438	};
3439	const char * const timehist_usage[] = {
3440		"perf sched timehist [<options>]",
3441		NULL
3442	};
3443	const char *const sched_subcommands[] = { "record", "latency", "map",
3444						  "replay", "script",
3445						  "timehist", NULL };
3446	const char *sched_usage[] = {
3447		NULL,
3448		NULL
3449	};
3450	struct trace_sched_handler lat_ops  = {
3451		.wakeup_event	    = latency_wakeup_event,
3452		.switch_event	    = latency_switch_event,
3453		.runtime_event	    = latency_runtime_event,
3454		.migrate_task_event = latency_migrate_task_event,
3455	};
3456	struct trace_sched_handler map_ops  = {
3457		.switch_event	    = map_switch_event,
3458	};
3459	struct trace_sched_handler replay_ops  = {
3460		.wakeup_event	    = replay_wakeup_event,
3461		.switch_event	    = replay_switch_event,
3462		.fork_event	    = replay_fork_event,
3463	};
3464	unsigned int i;
3465
3466	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3467		sched.curr_pid[i] = -1;
3468
3469	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3470					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3471	if (!argc)
3472		usage_with_options(sched_usage, sched_options);
3473
3474	/*
3475	 * Aliased to 'perf script' for now:
3476	 */
3477	if (!strcmp(argv[0], "script"))
3478		return cmd_script(argc, argv);
3479
3480	if (!strncmp(argv[0], "rec", 3)) {
3481		return __cmd_record(argc, argv);
3482	} else if (!strncmp(argv[0], "lat", 3)) {
3483		sched.tp_handler = &lat_ops;
3484		if (argc > 1) {
3485			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3486			if (argc)
3487				usage_with_options(latency_usage, latency_options);
3488		}
3489		setup_sorting(&sched, latency_options, latency_usage);
3490		return perf_sched__lat(&sched);
3491	} else if (!strcmp(argv[0], "map")) {
3492		if (argc) {
3493			argc = parse_options(argc, argv, map_options, map_usage, 0);
3494			if (argc)
3495				usage_with_options(map_usage, map_options);
3496		}
3497		sched.tp_handler = &map_ops;
3498		setup_sorting(&sched, latency_options, latency_usage);
3499		return perf_sched__map(&sched);
3500	} else if (!strncmp(argv[0], "rep", 3)) {
3501		sched.tp_handler = &replay_ops;
3502		if (argc) {
3503			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3504			if (argc)
3505				usage_with_options(replay_usage, replay_options);
3506		}
3507		return perf_sched__replay(&sched);
3508	} else if (!strcmp(argv[0], "timehist")) {
3509		if (argc) {
3510			argc = parse_options(argc, argv, timehist_options,
3511					     timehist_usage, 0);
3512			if (argc)
3513				usage_with_options(timehist_usage, timehist_options);
3514		}
3515		if ((sched.show_wakeups || sched.show_next) &&
3516		    sched.summary_only) {
3517			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3518			parse_options_usage(timehist_usage, timehist_options, "s", true);
3519			if (sched.show_wakeups)
3520				parse_options_usage(NULL, timehist_options, "w", true);
3521			if (sched.show_next)
3522				parse_options_usage(NULL, timehist_options, "n", true);
3523			return -EINVAL;
3524		}
3525
3526		return perf_sched__timehist(&sched);
3527	} else {
3528		usage_with_options(sched_usage, sched_options);
3529	}
3530
3531	return 0;
3532}
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3#include "perf.h"
   4#include "perf-sys.h"
   5
   6#include "util/cpumap.h"
   7#include "util/evlist.h"
 
   8#include "util/evsel.h"
   9#include "util/evsel_fprintf.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
  30#include <linux/kernel.h>
  31#include <linux/log2.h>
  32#include <linux/zalloc.h>
  33#include <sys/prctl.h>
  34#include <sys/resource.h>
  35#include <inttypes.h>
  36
  37#include <errno.h>
  38#include <semaphore.h>
  39#include <pthread.h>
  40#include <math.h>
  41#include <api/fs/fs.h>
  42#include <perf/cpumap.h>
  43#include <linux/time64.h>
  44#include <linux/err.h>
  45
  46#include <linux/ctype.h>
  47
  48#define PR_SET_NAME		15               /* Set process name */
  49#define MAX_CPUS		4096
  50#define COMM_LEN		20
  51#define SYM_LEN			129
  52#define MAX_PID			1024000
  53
  54static const char *cpu_list;
  55static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  56
  57struct sched_atom;
  58
  59struct task_desc {
  60	unsigned long		nr;
  61	unsigned long		pid;
  62	char			comm[COMM_LEN];
  63
  64	unsigned long		nr_events;
  65	unsigned long		curr_event;
  66	struct sched_atom	**atoms;
  67
  68	pthread_t		thread;
  69	sem_t			sleep_sem;
  70
  71	sem_t			ready_for_work;
  72	sem_t			work_done_sem;
  73
  74	u64			cpu_usage;
  75};
  76
  77enum sched_event_type {
  78	SCHED_EVENT_RUN,
  79	SCHED_EVENT_SLEEP,
  80	SCHED_EVENT_WAKEUP,
  81	SCHED_EVENT_MIGRATION,
  82};
  83
  84struct sched_atom {
  85	enum sched_event_type	type;
  86	int			specific_wait;
  87	u64			timestamp;
  88	u64			duration;
  89	unsigned long		nr;
  90	sem_t			*wait_sem;
  91	struct task_desc	*wakee;
  92};
  93
  94#define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
  95
  96/* task state bitmask, copied from include/linux/sched.h */
  97#define TASK_RUNNING		0
  98#define TASK_INTERRUPTIBLE	1
  99#define TASK_UNINTERRUPTIBLE	2
 100#define __TASK_STOPPED		4
 101#define __TASK_TRACED		8
 102/* in tsk->exit_state */
 103#define EXIT_DEAD		16
 104#define EXIT_ZOMBIE		32
 105#define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
 106/* in tsk->state again */
 107#define TASK_DEAD		64
 108#define TASK_WAKEKILL		128
 109#define TASK_WAKING		256
 110#define TASK_PARKED		512
 111
 112enum thread_state {
 113	THREAD_SLEEPING = 0,
 114	THREAD_WAIT_CPU,
 115	THREAD_SCHED_IN,
 116	THREAD_IGNORE
 117};
 118
 119struct work_atom {
 120	struct list_head	list;
 121	enum thread_state	state;
 122	u64			sched_out_time;
 123	u64			wake_up_time;
 124	u64			sched_in_time;
 125	u64			runtime;
 126};
 127
 128struct work_atoms {
 129	struct list_head	work_list;
 130	struct thread		*thread;
 131	struct rb_node		node;
 132	u64			max_lat;
 133	u64			max_lat_at;
 134	u64			total_lat;
 135	u64			nb_atoms;
 136	u64			total_runtime;
 137	int			num_merged;
 138};
 139
 140typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 141
 142struct perf_sched;
 143
 144struct trace_sched_handler {
 145	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 146			    struct perf_sample *sample, struct machine *machine);
 147
 148	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 149			     struct perf_sample *sample, struct machine *machine);
 150
 151	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 152			    struct perf_sample *sample, struct machine *machine);
 153
 154	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 155	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 156			  struct machine *machine);
 157
 158	int (*migrate_task_event)(struct perf_sched *sched,
 159				  struct evsel *evsel,
 160				  struct perf_sample *sample,
 161				  struct machine *machine);
 162};
 163
 164#define COLOR_PIDS PERF_COLOR_BLUE
 165#define COLOR_CPUS PERF_COLOR_BG_RED
 166
 167struct perf_sched_map {
 168	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 169	int			*comp_cpus;
 170	bool			 comp;
 171	struct perf_thread_map *color_pids;
 172	const char		*color_pids_str;
 173	struct perf_cpu_map	*color_cpus;
 174	const char		*color_cpus_str;
 175	struct perf_cpu_map	*cpus;
 176	const char		*cpus_str;
 177};
 178
 179struct perf_sched {
 180	struct perf_tool tool;
 181	const char	 *sort_order;
 182	unsigned long	 nr_tasks;
 183	struct task_desc **pid_to_task;
 184	struct task_desc **tasks;
 185	const struct trace_sched_handler *tp_handler;
 186	pthread_mutex_t	 start_work_mutex;
 187	pthread_mutex_t	 work_done_wait_mutex;
 188	int		 profile_cpu;
 189/*
 190 * Track the current task - that way we can know whether there's any
 191 * weird events, such as a task being switched away that is not current.
 192 */
 193	int		 max_cpu;
 194	u32		 curr_pid[MAX_CPUS];
 195	struct thread	 *curr_thread[MAX_CPUS];
 196	char		 next_shortname1;
 197	char		 next_shortname2;
 198	unsigned int	 replay_repeat;
 199	unsigned long	 nr_run_events;
 200	unsigned long	 nr_sleep_events;
 201	unsigned long	 nr_wakeup_events;
 202	unsigned long	 nr_sleep_corrections;
 203	unsigned long	 nr_run_events_optimized;
 204	unsigned long	 targetless_wakeups;
 205	unsigned long	 multitarget_wakeups;
 206	unsigned long	 nr_runs;
 207	unsigned long	 nr_timestamps;
 208	unsigned long	 nr_unordered_timestamps;
 209	unsigned long	 nr_context_switch_bugs;
 210	unsigned long	 nr_events;
 211	unsigned long	 nr_lost_chunks;
 212	unsigned long	 nr_lost_events;
 213	u64		 run_measurement_overhead;
 214	u64		 sleep_measurement_overhead;
 215	u64		 start_time;
 216	u64		 cpu_usage;
 217	u64		 runavg_cpu_usage;
 218	u64		 parent_cpu_usage;
 219	u64		 runavg_parent_cpu_usage;
 220	u64		 sum_runtime;
 221	u64		 sum_fluct;
 222	u64		 run_avg;
 223	u64		 all_runtime;
 224	u64		 all_count;
 225	u64		 cpu_last_switched[MAX_CPUS];
 226	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 227	struct list_head sort_list, cmp_pid;
 228	bool force;
 229	bool skip_merge;
 230	struct perf_sched_map map;
 231
 232	/* options for timehist command */
 233	bool		summary;
 234	bool		summary_only;
 235	bool		idle_hist;
 236	bool		show_callchain;
 237	unsigned int	max_stack;
 238	bool		show_cpu_visual;
 239	bool		show_wakeups;
 240	bool		show_next;
 241	bool		show_migrations;
 242	bool		show_state;
 243	u64		skipped_samples;
 244	const char	*time_str;
 245	struct perf_time_interval ptime;
 246	struct perf_time_interval hist_time;
 247};
 248
 249/* per thread run time data */
 250struct thread_runtime {
 251	u64 last_time;      /* time of previous sched in/out event */
 252	u64 dt_run;         /* run time */
 253	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 254	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 255	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 256	u64 dt_delay;       /* time between wakeup and sched-in */
 257	u64 ready_to_run;   /* time of wakeup */
 258
 259	struct stats run_stats;
 260	u64 total_run_time;
 261	u64 total_sleep_time;
 262	u64 total_iowait_time;
 263	u64 total_preempt_time;
 264	u64 total_delay_time;
 265
 266	int last_state;
 267
 268	char shortname[3];
 269	bool comm_changed;
 270
 271	u64 migrations;
 272};
 273
 274/* per event run time data */
 275struct evsel_runtime {
 276	u64 *last_time; /* time this event was last seen per cpu */
 277	u32 ncpu;       /* highest cpu slot allocated */
 278};
 279
 280/* per cpu idle time data */
 281struct idle_thread_runtime {
 282	struct thread_runtime	tr;
 283	struct thread		*last_thread;
 284	struct rb_root_cached	sorted_root;
 285	struct callchain_root	callchain;
 286	struct callchain_cursor	cursor;
 287};
 288
 289/* track idle times per cpu */
 290static struct thread **idle_threads;
 291static int idle_max_cpu;
 292static char idle_comm[] = "<idle>";
 293
 294static u64 get_nsecs(void)
 295{
 296	struct timespec ts;
 297
 298	clock_gettime(CLOCK_MONOTONIC, &ts);
 299
 300	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 301}
 302
 303static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 304{
 305	u64 T0 = get_nsecs(), T1;
 306
 307	do {
 308		T1 = get_nsecs();
 309	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 310}
 311
 312static void sleep_nsecs(u64 nsecs)
 313{
 314	struct timespec ts;
 315
 316	ts.tv_nsec = nsecs % 999999999;
 317	ts.tv_sec = nsecs / 999999999;
 318
 319	nanosleep(&ts, NULL);
 320}
 321
 322static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 323{
 324	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 325	int i;
 326
 327	for (i = 0; i < 10; i++) {
 328		T0 = get_nsecs();
 329		burn_nsecs(sched, 0);
 330		T1 = get_nsecs();
 331		delta = T1-T0;
 332		min_delta = min(min_delta, delta);
 333	}
 334	sched->run_measurement_overhead = min_delta;
 335
 336	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 337}
 338
 339static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 340{
 341	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 342	int i;
 343
 344	for (i = 0; i < 10; i++) {
 345		T0 = get_nsecs();
 346		sleep_nsecs(10000);
 347		T1 = get_nsecs();
 348		delta = T1-T0;
 349		min_delta = min(min_delta, delta);
 350	}
 351	min_delta -= 10000;
 352	sched->sleep_measurement_overhead = min_delta;
 353
 354	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 355}
 356
 357static struct sched_atom *
 358get_new_event(struct task_desc *task, u64 timestamp)
 359{
 360	struct sched_atom *event = zalloc(sizeof(*event));
 361	unsigned long idx = task->nr_events;
 362	size_t size;
 363
 364	event->timestamp = timestamp;
 365	event->nr = idx;
 366
 367	task->nr_events++;
 368	size = sizeof(struct sched_atom *) * task->nr_events;
 369	task->atoms = realloc(task->atoms, size);
 370	BUG_ON(!task->atoms);
 371
 372	task->atoms[idx] = event;
 373
 374	return event;
 375}
 376
 377static struct sched_atom *last_event(struct task_desc *task)
 378{
 379	if (!task->nr_events)
 380		return NULL;
 381
 382	return task->atoms[task->nr_events - 1];
 383}
 384
 385static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 386				u64 timestamp, u64 duration)
 387{
 388	struct sched_atom *event, *curr_event = last_event(task);
 389
 390	/*
 391	 * optimize an existing RUN event by merging this one
 392	 * to it:
 393	 */
 394	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 395		sched->nr_run_events_optimized++;
 396		curr_event->duration += duration;
 397		return;
 398	}
 399
 400	event = get_new_event(task, timestamp);
 401
 402	event->type = SCHED_EVENT_RUN;
 403	event->duration = duration;
 404
 405	sched->nr_run_events++;
 406}
 407
 408static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 409				   u64 timestamp, struct task_desc *wakee)
 410{
 411	struct sched_atom *event, *wakee_event;
 412
 413	event = get_new_event(task, timestamp);
 414	event->type = SCHED_EVENT_WAKEUP;
 415	event->wakee = wakee;
 416
 417	wakee_event = last_event(wakee);
 418	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 419		sched->targetless_wakeups++;
 420		return;
 421	}
 422	if (wakee_event->wait_sem) {
 423		sched->multitarget_wakeups++;
 424		return;
 425	}
 426
 427	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 428	sem_init(wakee_event->wait_sem, 0, 0);
 429	wakee_event->specific_wait = 1;
 430	event->wait_sem = wakee_event->wait_sem;
 431
 432	sched->nr_wakeup_events++;
 433}
 434
 435static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 436				  u64 timestamp, u64 task_state __maybe_unused)
 437{
 438	struct sched_atom *event = get_new_event(task, timestamp);
 439
 440	event->type = SCHED_EVENT_SLEEP;
 441
 442	sched->nr_sleep_events++;
 443}
 444
 445static struct task_desc *register_pid(struct perf_sched *sched,
 446				      unsigned long pid, const char *comm)
 447{
 448	struct task_desc *task;
 449	static int pid_max;
 450
 451	if (sched->pid_to_task == NULL) {
 452		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 453			pid_max = MAX_PID;
 454		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 455	}
 456	if (pid >= (unsigned long)pid_max) {
 457		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 458			sizeof(struct task_desc *))) == NULL);
 459		while (pid >= (unsigned long)pid_max)
 460			sched->pid_to_task[pid_max++] = NULL;
 461	}
 462
 463	task = sched->pid_to_task[pid];
 464
 465	if (task)
 466		return task;
 467
 468	task = zalloc(sizeof(*task));
 469	task->pid = pid;
 470	task->nr = sched->nr_tasks;
 471	strcpy(task->comm, comm);
 472	/*
 473	 * every task starts in sleeping state - this gets ignored
 474	 * if there's no wakeup pointing to this sleep state:
 475	 */
 476	add_sched_event_sleep(sched, task, 0, 0);
 477
 478	sched->pid_to_task[pid] = task;
 479	sched->nr_tasks++;
 480	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 481	BUG_ON(!sched->tasks);
 482	sched->tasks[task->nr] = task;
 483
 484	if (verbose > 0)
 485		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 486
 487	return task;
 488}
 489
 490
 491static void print_task_traces(struct perf_sched *sched)
 492{
 493	struct task_desc *task;
 494	unsigned long i;
 495
 496	for (i = 0; i < sched->nr_tasks; i++) {
 497		task = sched->tasks[i];
 498		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 499			task->nr, task->comm, task->pid, task->nr_events);
 500	}
 501}
 502
 503static void add_cross_task_wakeups(struct perf_sched *sched)
 504{
 505	struct task_desc *task1, *task2;
 506	unsigned long i, j;
 507
 508	for (i = 0; i < sched->nr_tasks; i++) {
 509		task1 = sched->tasks[i];
 510		j = i + 1;
 511		if (j == sched->nr_tasks)
 512			j = 0;
 513		task2 = sched->tasks[j];
 514		add_sched_event_wakeup(sched, task1, 0, task2);
 515	}
 516}
 517
 518static void perf_sched__process_event(struct perf_sched *sched,
 519				      struct sched_atom *atom)
 520{
 521	int ret = 0;
 522
 523	switch (atom->type) {
 524		case SCHED_EVENT_RUN:
 525			burn_nsecs(sched, atom->duration);
 526			break;
 527		case SCHED_EVENT_SLEEP:
 528			if (atom->wait_sem)
 529				ret = sem_wait(atom->wait_sem);
 530			BUG_ON(ret);
 531			break;
 532		case SCHED_EVENT_WAKEUP:
 533			if (atom->wait_sem)
 534				ret = sem_post(atom->wait_sem);
 535			BUG_ON(ret);
 536			break;
 537		case SCHED_EVENT_MIGRATION:
 538			break;
 539		default:
 540			BUG_ON(1);
 541	}
 542}
 543
 544static u64 get_cpu_usage_nsec_parent(void)
 545{
 546	struct rusage ru;
 547	u64 sum;
 548	int err;
 549
 550	err = getrusage(RUSAGE_SELF, &ru);
 551	BUG_ON(err);
 552
 553	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 554	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 555
 556	return sum;
 557}
 558
 559static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 560{
 561	struct perf_event_attr attr;
 562	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 563	int fd;
 564	struct rlimit limit;
 565	bool need_privilege = false;
 566
 567	memset(&attr, 0, sizeof(attr));
 568
 569	attr.type = PERF_TYPE_SOFTWARE;
 570	attr.config = PERF_COUNT_SW_TASK_CLOCK;
 571
 572force_again:
 573	fd = sys_perf_event_open(&attr, 0, -1, -1,
 574				 perf_event_open_cloexec_flag());
 575
 576	if (fd < 0) {
 577		if (errno == EMFILE) {
 578			if (sched->force) {
 579				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 580				limit.rlim_cur += sched->nr_tasks - cur_task;
 581				if (limit.rlim_cur > limit.rlim_max) {
 582					limit.rlim_max = limit.rlim_cur;
 583					need_privilege = true;
 584				}
 585				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 586					if (need_privilege && errno == EPERM)
 587						strcpy(info, "Need privilege\n");
 588				} else
 589					goto force_again;
 590			} else
 591				strcpy(info, "Have a try with -f option\n");
 592		}
 593		pr_err("Error: sys_perf_event_open() syscall returned "
 594		       "with %d (%s)\n%s", fd,
 595		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
 596		exit(EXIT_FAILURE);
 597	}
 598	return fd;
 599}
 600
 601static u64 get_cpu_usage_nsec_self(int fd)
 602{
 603	u64 runtime;
 604	int ret;
 605
 606	ret = read(fd, &runtime, sizeof(runtime));
 607	BUG_ON(ret != sizeof(runtime));
 608
 609	return runtime;
 610}
 611
 612struct sched_thread_parms {
 613	struct task_desc  *task;
 614	struct perf_sched *sched;
 615	int fd;
 616};
 617
 618static void *thread_func(void *ctx)
 619{
 620	struct sched_thread_parms *parms = ctx;
 621	struct task_desc *this_task = parms->task;
 622	struct perf_sched *sched = parms->sched;
 623	u64 cpu_usage_0, cpu_usage_1;
 624	unsigned long i, ret;
 625	char comm2[22];
 626	int fd = parms->fd;
 627
 628	zfree(&parms);
 629
 630	sprintf(comm2, ":%s", this_task->comm);
 631	prctl(PR_SET_NAME, comm2);
 632	if (fd < 0)
 633		return NULL;
 634again:
 635	ret = sem_post(&this_task->ready_for_work);
 636	BUG_ON(ret);
 637	ret = pthread_mutex_lock(&sched->start_work_mutex);
 638	BUG_ON(ret);
 639	ret = pthread_mutex_unlock(&sched->start_work_mutex);
 640	BUG_ON(ret);
 641
 642	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 643
 644	for (i = 0; i < this_task->nr_events; i++) {
 645		this_task->curr_event = i;
 646		perf_sched__process_event(sched, this_task->atoms[i]);
 647	}
 648
 649	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 650	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 651	ret = sem_post(&this_task->work_done_sem);
 652	BUG_ON(ret);
 653
 654	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 655	BUG_ON(ret);
 656	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
 657	BUG_ON(ret);
 658
 659	goto again;
 660}
 661
 662static void create_tasks(struct perf_sched *sched)
 663{
 664	struct task_desc *task;
 665	pthread_attr_t attr;
 666	unsigned long i;
 667	int err;
 668
 669	err = pthread_attr_init(&attr);
 670	BUG_ON(err);
 671	err = pthread_attr_setstacksize(&attr,
 672			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
 673	BUG_ON(err);
 674	err = pthread_mutex_lock(&sched->start_work_mutex);
 675	BUG_ON(err);
 676	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
 677	BUG_ON(err);
 678	for (i = 0; i < sched->nr_tasks; i++) {
 679		struct sched_thread_parms *parms = malloc(sizeof(*parms));
 680		BUG_ON(parms == NULL);
 681		parms->task = task = sched->tasks[i];
 682		parms->sched = sched;
 683		parms->fd = self_open_counters(sched, i);
 684		sem_init(&task->sleep_sem, 0, 0);
 685		sem_init(&task->ready_for_work, 0, 0);
 686		sem_init(&task->work_done_sem, 0, 0);
 687		task->curr_event = 0;
 688		err = pthread_create(&task->thread, &attr, thread_func, parms);
 689		BUG_ON(err);
 690	}
 691}
 692
 693static void wait_for_tasks(struct perf_sched *sched)
 694{
 695	u64 cpu_usage_0, cpu_usage_1;
 696	struct task_desc *task;
 697	unsigned long i, ret;
 698
 699	sched->start_time = get_nsecs();
 700	sched->cpu_usage = 0;
 701	pthread_mutex_unlock(&sched->work_done_wait_mutex);
 702
 703	for (i = 0; i < sched->nr_tasks; i++) {
 704		task = sched->tasks[i];
 705		ret = sem_wait(&task->ready_for_work);
 706		BUG_ON(ret);
 707		sem_init(&task->ready_for_work, 0, 0);
 708	}
 709	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
 710	BUG_ON(ret);
 711
 712	cpu_usage_0 = get_cpu_usage_nsec_parent();
 713
 714	pthread_mutex_unlock(&sched->start_work_mutex);
 715
 716	for (i = 0; i < sched->nr_tasks; i++) {
 717		task = sched->tasks[i];
 718		ret = sem_wait(&task->work_done_sem);
 719		BUG_ON(ret);
 720		sem_init(&task->work_done_sem, 0, 0);
 721		sched->cpu_usage += task->cpu_usage;
 722		task->cpu_usage = 0;
 723	}
 724
 725	cpu_usage_1 = get_cpu_usage_nsec_parent();
 726	if (!sched->runavg_cpu_usage)
 727		sched->runavg_cpu_usage = sched->cpu_usage;
 728	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 729
 730	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 731	if (!sched->runavg_parent_cpu_usage)
 732		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 733	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 734					 sched->parent_cpu_usage)/sched->replay_repeat;
 735
 736	ret = pthread_mutex_lock(&sched->start_work_mutex);
 737	BUG_ON(ret);
 738
 739	for (i = 0; i < sched->nr_tasks; i++) {
 740		task = sched->tasks[i];
 741		sem_init(&task->sleep_sem, 0, 0);
 742		task->curr_event = 0;
 743	}
 744}
 745
 746static void run_one_test(struct perf_sched *sched)
 747{
 748	u64 T0, T1, delta, avg_delta, fluct;
 749
 750	T0 = get_nsecs();
 751	wait_for_tasks(sched);
 752	T1 = get_nsecs();
 753
 754	delta = T1 - T0;
 755	sched->sum_runtime += delta;
 756	sched->nr_runs++;
 757
 758	avg_delta = sched->sum_runtime / sched->nr_runs;
 759	if (delta < avg_delta)
 760		fluct = avg_delta - delta;
 761	else
 762		fluct = delta - avg_delta;
 763	sched->sum_fluct += fluct;
 764	if (!sched->run_avg)
 765		sched->run_avg = delta;
 766	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 767
 768	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 769
 770	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 771
 772	printf("cpu: %0.2f / %0.2f",
 773		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 774
 775#if 0
 776	/*
 777	 * rusage statistics done by the parent, these are less
 778	 * accurate than the sched->sum_exec_runtime based statistics:
 779	 */
 780	printf(" [%0.2f / %0.2f]",
 781		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 782		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 783#endif
 784
 785	printf("\n");
 786
 787	if (sched->nr_sleep_corrections)
 788		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 789	sched->nr_sleep_corrections = 0;
 790}
 791
 792static void test_calibrations(struct perf_sched *sched)
 793{
 794	u64 T0, T1;
 795
 796	T0 = get_nsecs();
 797	burn_nsecs(sched, NSEC_PER_MSEC);
 798	T1 = get_nsecs();
 799
 800	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 801
 802	T0 = get_nsecs();
 803	sleep_nsecs(NSEC_PER_MSEC);
 804	T1 = get_nsecs();
 805
 806	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 807}
 808
 809static int
 810replay_wakeup_event(struct perf_sched *sched,
 811		    struct evsel *evsel, struct perf_sample *sample,
 812		    struct machine *machine __maybe_unused)
 813{
 814	const char *comm = evsel__strval(evsel, sample, "comm");
 815	const u32 pid	 = evsel__intval(evsel, sample, "pid");
 816	struct task_desc *waker, *wakee;
 817
 818	if (verbose > 0) {
 819		printf("sched_wakeup event %p\n", evsel);
 820
 821		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 822	}
 823
 824	waker = register_pid(sched, sample->tid, "<unknown>");
 825	wakee = register_pid(sched, pid, comm);
 826
 827	add_sched_event_wakeup(sched, waker, sample->time, wakee);
 828	return 0;
 829}
 830
 831static int replay_switch_event(struct perf_sched *sched,
 832			       struct evsel *evsel,
 833			       struct perf_sample *sample,
 834			       struct machine *machine __maybe_unused)
 835{
 836	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 837		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
 838	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 839		  next_pid = evsel__intval(evsel, sample, "next_pid");
 840	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
 841	struct task_desc *prev, __maybe_unused *next;
 842	u64 timestamp0, timestamp = sample->time;
 843	int cpu = sample->cpu;
 844	s64 delta;
 845
 846	if (verbose > 0)
 847		printf("sched_switch event %p\n", evsel);
 848
 849	if (cpu >= MAX_CPUS || cpu < 0)
 850		return 0;
 851
 852	timestamp0 = sched->cpu_last_switched[cpu];
 853	if (timestamp0)
 854		delta = timestamp - timestamp0;
 855	else
 856		delta = 0;
 857
 858	if (delta < 0) {
 859		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 860		return -1;
 861	}
 862
 863	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 864		 prev_comm, prev_pid, next_comm, next_pid, delta);
 865
 866	prev = register_pid(sched, prev_pid, prev_comm);
 867	next = register_pid(sched, next_pid, next_comm);
 868
 869	sched->cpu_last_switched[cpu] = timestamp;
 870
 871	add_sched_event_run(sched, prev, timestamp, delta);
 872	add_sched_event_sleep(sched, prev, timestamp, prev_state);
 873
 874	return 0;
 875}
 876
 877static int replay_fork_event(struct perf_sched *sched,
 878			     union perf_event *event,
 879			     struct machine *machine)
 880{
 881	struct thread *child, *parent;
 882
 883	child = machine__findnew_thread(machine, event->fork.pid,
 884					event->fork.tid);
 885	parent = machine__findnew_thread(machine, event->fork.ppid,
 886					 event->fork.ptid);
 887
 888	if (child == NULL || parent == NULL) {
 889		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 890				 child, parent);
 891		goto out_put;
 892	}
 893
 894	if (verbose > 0) {
 895		printf("fork event\n");
 896		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
 897		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
 898	}
 899
 900	register_pid(sched, parent->tid, thread__comm_str(parent));
 901	register_pid(sched, child->tid, thread__comm_str(child));
 902out_put:
 903	thread__put(child);
 904	thread__put(parent);
 905	return 0;
 906}
 907
 908struct sort_dimension {
 909	const char		*name;
 910	sort_fn_t		cmp;
 911	struct list_head	list;
 912};
 913
 914/*
 915 * handle runtime stats saved per thread
 916 */
 917static struct thread_runtime *thread__init_runtime(struct thread *thread)
 918{
 919	struct thread_runtime *r;
 920
 921	r = zalloc(sizeof(struct thread_runtime));
 922	if (!r)
 923		return NULL;
 924
 925	init_stats(&r->run_stats);
 926	thread__set_priv(thread, r);
 927
 928	return r;
 929}
 930
 931static struct thread_runtime *thread__get_runtime(struct thread *thread)
 932{
 933	struct thread_runtime *tr;
 934
 935	tr = thread__priv(thread);
 936	if (tr == NULL) {
 937		tr = thread__init_runtime(thread);
 938		if (tr == NULL)
 939			pr_debug("Failed to malloc memory for runtime data.\n");
 940	}
 941
 942	return tr;
 943}
 944
 945static int
 946thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 947{
 948	struct sort_dimension *sort;
 949	int ret = 0;
 950
 951	BUG_ON(list_empty(list));
 952
 953	list_for_each_entry(sort, list, list) {
 954		ret = sort->cmp(l, r);
 955		if (ret)
 956			return ret;
 957	}
 958
 959	return ret;
 960}
 961
 962static struct work_atoms *
 963thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 964			 struct list_head *sort_list)
 965{
 966	struct rb_node *node = root->rb_root.rb_node;
 967	struct work_atoms key = { .thread = thread };
 968
 969	while (node) {
 970		struct work_atoms *atoms;
 971		int cmp;
 972
 973		atoms = container_of(node, struct work_atoms, node);
 974
 975		cmp = thread_lat_cmp(sort_list, &key, atoms);
 976		if (cmp > 0)
 977			node = node->rb_left;
 978		else if (cmp < 0)
 979			node = node->rb_right;
 980		else {
 981			BUG_ON(thread != atoms->thread);
 982			return atoms;
 983		}
 984	}
 985	return NULL;
 986}
 987
 988static void
 989__thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 990			 struct list_head *sort_list)
 991{
 992	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 993	bool leftmost = true;
 994
 995	while (*new) {
 996		struct work_atoms *this;
 997		int cmp;
 998
 999		this = container_of(*new, struct work_atoms, node);
1000		parent = *new;
1001
1002		cmp = thread_lat_cmp(sort_list, data, this);
1003
1004		if (cmp > 0)
1005			new = &((*new)->rb_left);
1006		else {
1007			new = &((*new)->rb_right);
1008			leftmost = false;
1009		}
1010	}
1011
1012	rb_link_node(&data->node, parent, new);
1013	rb_insert_color_cached(&data->node, root, leftmost);
1014}
1015
1016static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1017{
1018	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1019	if (!atoms) {
1020		pr_err("No memory at %s\n", __func__);
1021		return -1;
1022	}
1023
1024	atoms->thread = thread__get(thread);
1025	INIT_LIST_HEAD(&atoms->work_list);
1026	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1027	return 0;
1028}
1029
1030static char sched_out_state(u64 prev_state)
1031{
1032	const char *str = TASK_STATE_TO_CHAR_STR;
1033
1034	return str[prev_state];
1035}
1036
1037static int
1038add_sched_out_event(struct work_atoms *atoms,
1039		    char run_state,
1040		    u64 timestamp)
1041{
1042	struct work_atom *atom = zalloc(sizeof(*atom));
1043	if (!atom) {
1044		pr_err("Non memory at %s", __func__);
1045		return -1;
1046	}
1047
1048	atom->sched_out_time = timestamp;
1049
1050	if (run_state == 'R') {
1051		atom->state = THREAD_WAIT_CPU;
1052		atom->wake_up_time = atom->sched_out_time;
1053	}
1054
1055	list_add_tail(&atom->list, &atoms->work_list);
1056	return 0;
1057}
1058
1059static void
1060add_runtime_event(struct work_atoms *atoms, u64 delta,
1061		  u64 timestamp __maybe_unused)
1062{
1063	struct work_atom *atom;
1064
1065	BUG_ON(list_empty(&atoms->work_list));
1066
1067	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1068
1069	atom->runtime += delta;
1070	atoms->total_runtime += delta;
1071}
1072
1073static void
1074add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1075{
1076	struct work_atom *atom;
1077	u64 delta;
1078
1079	if (list_empty(&atoms->work_list))
1080		return;
1081
1082	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1083
1084	if (atom->state != THREAD_WAIT_CPU)
1085		return;
1086
1087	if (timestamp < atom->wake_up_time) {
1088		atom->state = THREAD_IGNORE;
1089		return;
1090	}
1091
1092	atom->state = THREAD_SCHED_IN;
1093	atom->sched_in_time = timestamp;
1094
1095	delta = atom->sched_in_time - atom->wake_up_time;
1096	atoms->total_lat += delta;
1097	if (delta > atoms->max_lat) {
1098		atoms->max_lat = delta;
1099		atoms->max_lat_at = timestamp;
1100	}
1101	atoms->nb_atoms++;
1102}
1103
1104static int latency_switch_event(struct perf_sched *sched,
1105				struct evsel *evsel,
1106				struct perf_sample *sample,
1107				struct machine *machine)
1108{
1109	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1110		  next_pid = evsel__intval(evsel, sample, "next_pid");
1111	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1112	struct work_atoms *out_events, *in_events;
1113	struct thread *sched_out, *sched_in;
1114	u64 timestamp0, timestamp = sample->time;
1115	int cpu = sample->cpu, err = -1;
1116	s64 delta;
1117
1118	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1119
1120	timestamp0 = sched->cpu_last_switched[cpu];
1121	sched->cpu_last_switched[cpu] = timestamp;
1122	if (timestamp0)
1123		delta = timestamp - timestamp0;
1124	else
1125		delta = 0;
1126
1127	if (delta < 0) {
1128		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1129		return -1;
1130	}
1131
1132	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1133	sched_in = machine__findnew_thread(machine, -1, next_pid);
1134	if (sched_out == NULL || sched_in == NULL)
1135		goto out_put;
1136
1137	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1138	if (!out_events) {
1139		if (thread_atoms_insert(sched, sched_out))
1140			goto out_put;
1141		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1142		if (!out_events) {
1143			pr_err("out-event: Internal tree error");
1144			goto out_put;
1145		}
1146	}
1147	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1148		return -1;
1149
1150	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1151	if (!in_events) {
1152		if (thread_atoms_insert(sched, sched_in))
1153			goto out_put;
1154		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1155		if (!in_events) {
1156			pr_err("in-event: Internal tree error");
1157			goto out_put;
1158		}
1159		/*
1160		 * Take came in we have not heard about yet,
1161		 * add in an initial atom in runnable state:
1162		 */
1163		if (add_sched_out_event(in_events, 'R', timestamp))
1164			goto out_put;
1165	}
1166	add_sched_in_event(in_events, timestamp);
1167	err = 0;
1168out_put:
1169	thread__put(sched_out);
1170	thread__put(sched_in);
1171	return err;
1172}
1173
1174static int latency_runtime_event(struct perf_sched *sched,
1175				 struct evsel *evsel,
1176				 struct perf_sample *sample,
1177				 struct machine *machine)
1178{
1179	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1180	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1181	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1182	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1183	u64 timestamp = sample->time;
1184	int cpu = sample->cpu, err = -1;
1185
1186	if (thread == NULL)
1187		return -1;
1188
1189	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1190	if (!atoms) {
1191		if (thread_atoms_insert(sched, thread))
1192			goto out_put;
1193		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1194		if (!atoms) {
1195			pr_err("in-event: Internal tree error");
1196			goto out_put;
1197		}
1198		if (add_sched_out_event(atoms, 'R', timestamp))
1199			goto out_put;
1200	}
1201
1202	add_runtime_event(atoms, runtime, timestamp);
1203	err = 0;
1204out_put:
1205	thread__put(thread);
1206	return err;
1207}
1208
1209static int latency_wakeup_event(struct perf_sched *sched,
1210				struct evsel *evsel,
1211				struct perf_sample *sample,
1212				struct machine *machine)
1213{
1214	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1215	struct work_atoms *atoms;
1216	struct work_atom *atom;
1217	struct thread *wakee;
1218	u64 timestamp = sample->time;
1219	int err = -1;
1220
1221	wakee = machine__findnew_thread(machine, -1, pid);
1222	if (wakee == NULL)
1223		return -1;
1224	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1225	if (!atoms) {
1226		if (thread_atoms_insert(sched, wakee))
1227			goto out_put;
1228		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1229		if (!atoms) {
1230			pr_err("wakeup-event: Internal tree error");
1231			goto out_put;
1232		}
1233		if (add_sched_out_event(atoms, 'S', timestamp))
1234			goto out_put;
1235	}
1236
1237	BUG_ON(list_empty(&atoms->work_list));
1238
1239	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1240
1241	/*
1242	 * As we do not guarantee the wakeup event happens when
1243	 * task is out of run queue, also may happen when task is
1244	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1245	 * then we should not set the ->wake_up_time when wake up a
1246	 * task which is on run queue.
1247	 *
1248	 * You WILL be missing events if you've recorded only
1249	 * one CPU, or are only looking at only one, so don't
1250	 * skip in this case.
1251	 */
1252	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1253		goto out_ok;
1254
1255	sched->nr_timestamps++;
1256	if (atom->sched_out_time > timestamp) {
1257		sched->nr_unordered_timestamps++;
1258		goto out_ok;
1259	}
1260
1261	atom->state = THREAD_WAIT_CPU;
1262	atom->wake_up_time = timestamp;
1263out_ok:
1264	err = 0;
1265out_put:
1266	thread__put(wakee);
1267	return err;
1268}
1269
1270static int latency_migrate_task_event(struct perf_sched *sched,
1271				      struct evsel *evsel,
1272				      struct perf_sample *sample,
1273				      struct machine *machine)
1274{
1275	const u32 pid = evsel__intval(evsel, sample, "pid");
1276	u64 timestamp = sample->time;
1277	struct work_atoms *atoms;
1278	struct work_atom *atom;
1279	struct thread *migrant;
1280	int err = -1;
1281
1282	/*
1283	 * Only need to worry about migration when profiling one CPU.
1284	 */
1285	if (sched->profile_cpu == -1)
1286		return 0;
1287
1288	migrant = machine__findnew_thread(machine, -1, pid);
1289	if (migrant == NULL)
1290		return -1;
1291	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1292	if (!atoms) {
1293		if (thread_atoms_insert(sched, migrant))
1294			goto out_put;
1295		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1296		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1297		if (!atoms) {
1298			pr_err("migration-event: Internal tree error");
1299			goto out_put;
1300		}
1301		if (add_sched_out_event(atoms, 'R', timestamp))
1302			goto out_put;
1303	}
1304
1305	BUG_ON(list_empty(&atoms->work_list));
1306
1307	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1308	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1309
1310	sched->nr_timestamps++;
1311
1312	if (atom->sched_out_time > timestamp)
1313		sched->nr_unordered_timestamps++;
1314	err = 0;
1315out_put:
1316	thread__put(migrant);
1317	return err;
1318}
1319
1320static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1321{
1322	int i;
1323	int ret;
1324	u64 avg;
1325	char max_lat_at[32];
1326
1327	if (!work_list->nb_atoms)
1328		return;
1329	/*
1330	 * Ignore idle threads:
1331	 */
1332	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1333		return;
1334
1335	sched->all_runtime += work_list->total_runtime;
1336	sched->all_count   += work_list->nb_atoms;
1337
1338	if (work_list->num_merged > 1)
1339		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1340	else
1341		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1342
1343	for (i = 0; i < 24 - ret; i++)
1344		printf(" ");
1345
1346	avg = work_list->total_lat / work_list->nb_atoms;
1347	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1348
1349	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1350	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1351		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1352		 (double)work_list->max_lat / NSEC_PER_MSEC,
1353		 max_lat_at);
1354}
1355
1356static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1357{
1358	if (l->thread == r->thread)
1359		return 0;
1360	if (l->thread->tid < r->thread->tid)
1361		return -1;
1362	if (l->thread->tid > r->thread->tid)
1363		return 1;
1364	return (int)(l->thread - r->thread);
1365}
1366
1367static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1368{
1369	u64 avgl, avgr;
1370
1371	if (!l->nb_atoms)
1372		return -1;
1373
1374	if (!r->nb_atoms)
1375		return 1;
1376
1377	avgl = l->total_lat / l->nb_atoms;
1378	avgr = r->total_lat / r->nb_atoms;
1379
1380	if (avgl < avgr)
1381		return -1;
1382	if (avgl > avgr)
1383		return 1;
1384
1385	return 0;
1386}
1387
1388static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1389{
1390	if (l->max_lat < r->max_lat)
1391		return -1;
1392	if (l->max_lat > r->max_lat)
1393		return 1;
1394
1395	return 0;
1396}
1397
1398static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1399{
1400	if (l->nb_atoms < r->nb_atoms)
1401		return -1;
1402	if (l->nb_atoms > r->nb_atoms)
1403		return 1;
1404
1405	return 0;
1406}
1407
1408static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1409{
1410	if (l->total_runtime < r->total_runtime)
1411		return -1;
1412	if (l->total_runtime > r->total_runtime)
1413		return 1;
1414
1415	return 0;
1416}
1417
1418static int sort_dimension__add(const char *tok, struct list_head *list)
1419{
1420	size_t i;
1421	static struct sort_dimension avg_sort_dimension = {
1422		.name = "avg",
1423		.cmp  = avg_cmp,
1424	};
1425	static struct sort_dimension max_sort_dimension = {
1426		.name = "max",
1427		.cmp  = max_cmp,
1428	};
1429	static struct sort_dimension pid_sort_dimension = {
1430		.name = "pid",
1431		.cmp  = pid_cmp,
1432	};
1433	static struct sort_dimension runtime_sort_dimension = {
1434		.name = "runtime",
1435		.cmp  = runtime_cmp,
1436	};
1437	static struct sort_dimension switch_sort_dimension = {
1438		.name = "switch",
1439		.cmp  = switch_cmp,
1440	};
1441	struct sort_dimension *available_sorts[] = {
1442		&pid_sort_dimension,
1443		&avg_sort_dimension,
1444		&max_sort_dimension,
1445		&switch_sort_dimension,
1446		&runtime_sort_dimension,
1447	};
1448
1449	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1450		if (!strcmp(available_sorts[i]->name, tok)) {
1451			list_add_tail(&available_sorts[i]->list, list);
1452
1453			return 0;
1454		}
1455	}
1456
1457	return -1;
1458}
1459
1460static void perf_sched__sort_lat(struct perf_sched *sched)
1461{
1462	struct rb_node *node;
1463	struct rb_root_cached *root = &sched->atom_root;
1464again:
1465	for (;;) {
1466		struct work_atoms *data;
1467		node = rb_first_cached(root);
1468		if (!node)
1469			break;
1470
1471		rb_erase_cached(node, root);
1472		data = rb_entry(node, struct work_atoms, node);
1473		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1474	}
1475	if (root == &sched->atom_root) {
1476		root = &sched->merged_atom_root;
1477		goto again;
1478	}
1479}
1480
1481static int process_sched_wakeup_event(struct perf_tool *tool,
1482				      struct evsel *evsel,
1483				      struct perf_sample *sample,
1484				      struct machine *machine)
1485{
1486	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1487
1488	if (sched->tp_handler->wakeup_event)
1489		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1490
1491	return 0;
1492}
1493
1494union map_priv {
1495	void	*ptr;
1496	bool	 color;
1497};
1498
1499static bool thread__has_color(struct thread *thread)
1500{
1501	union map_priv priv = {
1502		.ptr = thread__priv(thread),
1503	};
1504
1505	return priv.color;
1506}
1507
1508static struct thread*
1509map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1510{
1511	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1512	union map_priv priv = {
1513		.color = false,
1514	};
1515
1516	if (!sched->map.color_pids || !thread || thread__priv(thread))
1517		return thread;
1518
1519	if (thread_map__has(sched->map.color_pids, tid))
1520		priv.color = true;
1521
1522	thread__set_priv(thread, priv.ptr);
1523	return thread;
1524}
1525
1526static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1527			    struct perf_sample *sample, struct machine *machine)
1528{
1529	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1530	struct thread *sched_in;
1531	struct thread_runtime *tr;
1532	int new_shortname;
1533	u64 timestamp0, timestamp = sample->time;
1534	s64 delta;
1535	int i, this_cpu = sample->cpu;
1536	int cpus_nr;
1537	bool new_cpu = false;
1538	const char *color = PERF_COLOR_NORMAL;
1539	char stimestamp[32];
1540
1541	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1542
1543	if (this_cpu > sched->max_cpu)
1544		sched->max_cpu = this_cpu;
1545
1546	if (sched->map.comp) {
1547		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1548		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1549			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1550			new_cpu = true;
1551		}
1552	} else
1553		cpus_nr = sched->max_cpu;
1554
1555	timestamp0 = sched->cpu_last_switched[this_cpu];
1556	sched->cpu_last_switched[this_cpu] = timestamp;
1557	if (timestamp0)
1558		delta = timestamp - timestamp0;
1559	else
1560		delta = 0;
1561
1562	if (delta < 0) {
1563		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1564		return -1;
1565	}
1566
1567	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1568	if (sched_in == NULL)
1569		return -1;
1570
1571	tr = thread__get_runtime(sched_in);
1572	if (tr == NULL) {
1573		thread__put(sched_in);
1574		return -1;
1575	}
1576
1577	sched->curr_thread[this_cpu] = thread__get(sched_in);
1578
1579	printf("  ");
1580
1581	new_shortname = 0;
1582	if (!tr->shortname[0]) {
1583		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1584			/*
1585			 * Don't allocate a letter-number for swapper:0
1586			 * as a shortname. Instead, we use '.' for it.
1587			 */
1588			tr->shortname[0] = '.';
1589			tr->shortname[1] = ' ';
1590		} else {
1591			tr->shortname[0] = sched->next_shortname1;
1592			tr->shortname[1] = sched->next_shortname2;
1593
1594			if (sched->next_shortname1 < 'Z') {
1595				sched->next_shortname1++;
1596			} else {
1597				sched->next_shortname1 = 'A';
1598				if (sched->next_shortname2 < '9')
1599					sched->next_shortname2++;
1600				else
1601					sched->next_shortname2 = '0';
1602			}
1603		}
1604		new_shortname = 1;
1605	}
1606
1607	for (i = 0; i < cpus_nr; i++) {
1608		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1609		struct thread *curr_thread = sched->curr_thread[cpu];
1610		struct thread_runtime *curr_tr;
1611		const char *pid_color = color;
1612		const char *cpu_color = color;
1613
1614		if (curr_thread && thread__has_color(curr_thread))
1615			pid_color = COLOR_PIDS;
1616
1617		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1618			continue;
1619
1620		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1621			cpu_color = COLOR_CPUS;
1622
1623		if (cpu != this_cpu)
1624			color_fprintf(stdout, color, " ");
1625		else
1626			color_fprintf(stdout, cpu_color, "*");
1627
1628		if (sched->curr_thread[cpu]) {
1629			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1630			if (curr_tr == NULL) {
1631				thread__put(sched_in);
1632				return -1;
1633			}
1634			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1635		} else
1636			color_fprintf(stdout, color, "   ");
1637	}
1638
1639	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1640		goto out;
1641
1642	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1643	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1644	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1645		const char *pid_color = color;
1646
1647		if (thread__has_color(sched_in))
1648			pid_color = COLOR_PIDS;
1649
1650		color_fprintf(stdout, pid_color, "%s => %s:%d",
1651		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1652		tr->comm_changed = false;
1653	}
1654
1655	if (sched->map.comp && new_cpu)
1656		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1657
1658out:
1659	color_fprintf(stdout, color, "\n");
1660
1661	thread__put(sched_in);
1662
1663	return 0;
1664}
1665
1666static int process_sched_switch_event(struct perf_tool *tool,
1667				      struct evsel *evsel,
1668				      struct perf_sample *sample,
1669				      struct machine *machine)
1670{
1671	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1672	int this_cpu = sample->cpu, err = 0;
1673	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1674	    next_pid = evsel__intval(evsel, sample, "next_pid");
1675
1676	if (sched->curr_pid[this_cpu] != (u32)-1) {
1677		/*
1678		 * Are we trying to switch away a PID that is
1679		 * not current?
1680		 */
1681		if (sched->curr_pid[this_cpu] != prev_pid)
1682			sched->nr_context_switch_bugs++;
1683	}
1684
1685	if (sched->tp_handler->switch_event)
1686		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1687
1688	sched->curr_pid[this_cpu] = next_pid;
1689	return err;
1690}
1691
1692static int process_sched_runtime_event(struct perf_tool *tool,
1693				       struct evsel *evsel,
1694				       struct perf_sample *sample,
1695				       struct machine *machine)
1696{
1697	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1698
1699	if (sched->tp_handler->runtime_event)
1700		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1701
1702	return 0;
1703}
1704
1705static int perf_sched__process_fork_event(struct perf_tool *tool,
1706					  union perf_event *event,
1707					  struct perf_sample *sample,
1708					  struct machine *machine)
1709{
1710	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1711
1712	/* run the fork event through the perf machineruy */
1713	perf_event__process_fork(tool, event, sample, machine);
1714
1715	/* and then run additional processing needed for this command */
1716	if (sched->tp_handler->fork_event)
1717		return sched->tp_handler->fork_event(sched, event, machine);
1718
1719	return 0;
1720}
1721
1722static int process_sched_migrate_task_event(struct perf_tool *tool,
1723					    struct evsel *evsel,
1724					    struct perf_sample *sample,
1725					    struct machine *machine)
1726{
1727	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1728
1729	if (sched->tp_handler->migrate_task_event)
1730		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1731
1732	return 0;
1733}
1734
1735typedef int (*tracepoint_handler)(struct perf_tool *tool,
1736				  struct evsel *evsel,
1737				  struct perf_sample *sample,
1738				  struct machine *machine);
1739
1740static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1741						 union perf_event *event __maybe_unused,
1742						 struct perf_sample *sample,
1743						 struct evsel *evsel,
1744						 struct machine *machine)
1745{
1746	int err = 0;
1747
1748	if (evsel->handler != NULL) {
1749		tracepoint_handler f = evsel->handler;
1750		err = f(tool, evsel, sample, machine);
1751	}
1752
1753	return err;
1754}
1755
1756static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1757				    union perf_event *event,
1758				    struct perf_sample *sample,
1759				    struct machine *machine)
1760{
1761	struct thread *thread;
1762	struct thread_runtime *tr;
1763	int err;
1764
1765	err = perf_event__process_comm(tool, event, sample, machine);
1766	if (err)
1767		return err;
1768
1769	thread = machine__find_thread(machine, sample->pid, sample->tid);
1770	if (!thread) {
1771		pr_err("Internal error: can't find thread\n");
1772		return -1;
1773	}
1774
1775	tr = thread__get_runtime(thread);
1776	if (tr == NULL) {
1777		thread__put(thread);
1778		return -1;
1779	}
1780
1781	tr->comm_changed = true;
1782	thread__put(thread);
1783
1784	return 0;
1785}
1786
1787static int perf_sched__read_events(struct perf_sched *sched)
1788{
1789	const struct evsel_str_handler handlers[] = {
1790		{ "sched:sched_switch",	      process_sched_switch_event, },
1791		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1792		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1793		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1794		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1795	};
1796	struct perf_session *session;
1797	struct perf_data data = {
1798		.path  = input_name,
1799		.mode  = PERF_DATA_MODE_READ,
1800		.force = sched->force,
 
 
1801	};
1802	int rc = -1;
1803
1804	session = perf_session__new(&data, false, &sched->tool);
1805	if (IS_ERR(session)) {
1806		pr_debug("Error creating perf session");
1807		return PTR_ERR(session);
1808	}
1809
1810	symbol__init(&session->header.env);
1811
1812	if (perf_session__set_tracepoints_handlers(session, handlers))
1813		goto out_delete;
1814
1815	if (perf_session__has_traces(session, "record -R")) {
1816		int err = perf_session__process_events(session);
1817		if (err) {
1818			pr_err("Failed to process events, error %d", err);
1819			goto out_delete;
1820		}
1821
1822		sched->nr_events      = session->evlist->stats.nr_events[0];
1823		sched->nr_lost_events = session->evlist->stats.total_lost;
1824		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1825	}
1826
1827	rc = 0;
1828out_delete:
1829	perf_session__delete(session);
1830	return rc;
1831}
1832
1833/*
1834 * scheduling times are printed as msec.usec
1835 */
1836static inline void print_sched_time(unsigned long long nsecs, int width)
1837{
1838	unsigned long msecs;
1839	unsigned long usecs;
1840
1841	msecs  = nsecs / NSEC_PER_MSEC;
1842	nsecs -= msecs * NSEC_PER_MSEC;
1843	usecs  = nsecs / NSEC_PER_USEC;
1844	printf("%*lu.%03lu ", width, msecs, usecs);
1845}
1846
1847/*
1848 * returns runtime data for event, allocating memory for it the
1849 * first time it is used.
1850 */
1851static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1852{
1853	struct evsel_runtime *r = evsel->priv;
1854
1855	if (r == NULL) {
1856		r = zalloc(sizeof(struct evsel_runtime));
1857		evsel->priv = r;
1858	}
1859
1860	return r;
1861}
1862
1863/*
1864 * save last time event was seen per cpu
1865 */
1866static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
 
1867{
1868	struct evsel_runtime *r = evsel__get_runtime(evsel);
1869
1870	if (r == NULL)
1871		return;
1872
1873	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1874		int i, n = __roundup_pow_of_two(cpu+1);
1875		void *p = r->last_time;
1876
1877		p = realloc(r->last_time, n * sizeof(u64));
1878		if (!p)
1879			return;
1880
1881		r->last_time = p;
1882		for (i = r->ncpu; i < n; ++i)
1883			r->last_time[i] = (u64) 0;
1884
1885		r->ncpu = n;
1886	}
1887
1888	r->last_time[cpu] = timestamp;
1889}
1890
1891/* returns last time this event was seen on the given cpu */
1892static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1893{
1894	struct evsel_runtime *r = evsel__get_runtime(evsel);
1895
1896	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1897		return 0;
1898
1899	return r->last_time[cpu];
1900}
1901
1902static int comm_width = 30;
1903
1904static char *timehist_get_commstr(struct thread *thread)
1905{
1906	static char str[32];
1907	const char *comm = thread__comm_str(thread);
1908	pid_t tid = thread->tid;
1909	pid_t pid = thread->pid_;
1910	int n;
1911
1912	if (pid == 0)
1913		n = scnprintf(str, sizeof(str), "%s", comm);
1914
1915	else if (tid != pid)
1916		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1917
1918	else
1919		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1920
1921	if (n > comm_width)
1922		comm_width = n;
1923
1924	return str;
1925}
1926
1927static void timehist_header(struct perf_sched *sched)
1928{
1929	u32 ncpus = sched->max_cpu + 1;
1930	u32 i, j;
1931
1932	printf("%15s %6s ", "time", "cpu");
1933
1934	if (sched->show_cpu_visual) {
1935		printf(" ");
1936		for (i = 0, j = 0; i < ncpus; ++i) {
1937			printf("%x", j++);
1938			if (j > 15)
1939				j = 0;
1940		}
1941		printf(" ");
1942	}
1943
1944	printf(" %-*s  %9s  %9s  %9s", comm_width,
1945		"task name", "wait time", "sch delay", "run time");
1946
1947	if (sched->show_state)
1948		printf("  %s", "state");
1949
1950	printf("\n");
1951
1952	/*
1953	 * units row
1954	 */
1955	printf("%15s %-6s ", "", "");
1956
1957	if (sched->show_cpu_visual)
1958		printf(" %*s ", ncpus, "");
1959
1960	printf(" %-*s  %9s  %9s  %9s", comm_width,
1961	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1962
1963	if (sched->show_state)
1964		printf("  %5s", "");
1965
1966	printf("\n");
1967
1968	/*
1969	 * separator
1970	 */
1971	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1972
1973	if (sched->show_cpu_visual)
1974		printf(" %.*s ", ncpus, graph_dotted_line);
1975
1976	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1977		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1978		graph_dotted_line);
1979
1980	if (sched->show_state)
1981		printf("  %.5s", graph_dotted_line);
1982
1983	printf("\n");
1984}
1985
1986static char task_state_char(struct thread *thread, int state)
1987{
1988	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1989	unsigned bit = state ? ffs(state) : 0;
1990
1991	/* 'I' for idle */
1992	if (thread->tid == 0)
1993		return 'I';
1994
1995	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1996}
1997
1998static void timehist_print_sample(struct perf_sched *sched,
1999				  struct evsel *evsel,
2000				  struct perf_sample *sample,
2001				  struct addr_location *al,
2002				  struct thread *thread,
2003				  u64 t, int state)
2004{
2005	struct thread_runtime *tr = thread__priv(thread);
2006	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2007	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2008	u32 max_cpus = sched->max_cpu + 1;
2009	char tstr[64];
2010	char nstr[30];
2011	u64 wait_time;
2012
2013	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2014		return;
2015
2016	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2017	printf("%15s [%04d] ", tstr, sample->cpu);
2018
2019	if (sched->show_cpu_visual) {
2020		u32 i;
2021		char c;
2022
2023		printf(" ");
2024		for (i = 0; i < max_cpus; ++i) {
2025			/* flag idle times with 'i'; others are sched events */
2026			if (i == sample->cpu)
2027				c = (thread->tid == 0) ? 'i' : 's';
2028			else
2029				c = ' ';
2030			printf("%c", c);
2031		}
2032		printf(" ");
2033	}
2034
2035	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2036
2037	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2038	print_sched_time(wait_time, 6);
2039
2040	print_sched_time(tr->dt_delay, 6);
2041	print_sched_time(tr->dt_run, 6);
2042
2043	if (sched->show_state)
2044		printf(" %5c ", task_state_char(thread, state));
2045
2046	if (sched->show_next) {
2047		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2048		printf(" %-*s", comm_width, nstr);
2049	}
2050
2051	if (sched->show_wakeups && !sched->show_next)
2052		printf("  %-*s", comm_width, "");
2053
2054	if (thread->tid == 0)
2055		goto out;
2056
2057	if (sched->show_callchain)
2058		printf("  ");
2059
2060	sample__fprintf_sym(sample, al, 0,
2061			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2062			    EVSEL__PRINT_CALLCHAIN_ARROW |
2063			    EVSEL__PRINT_SKIP_IGNORED,
2064			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2065
2066out:
2067	printf("\n");
2068}
2069
2070/*
2071 * Explanation of delta-time stats:
2072 *
2073 *            t = time of current schedule out event
2074 *        tprev = time of previous sched out event
2075 *                also time of schedule-in event for current task
2076 *    last_time = time of last sched change event for current task
2077 *                (i.e, time process was last scheduled out)
2078 * ready_to_run = time of wakeup for current task
2079 *
2080 * -----|------------|------------|------------|------
2081 *    last         ready        tprev          t
2082 *    time         to run
2083 *
2084 *      |-------- dt_wait --------|
2085 *                   |- dt_delay -|-- dt_run --|
2086 *
2087 *   dt_run = run time of current task
2088 *  dt_wait = time between last schedule out event for task and tprev
2089 *            represents time spent off the cpu
2090 * dt_delay = time between wakeup and schedule-in of task
2091 */
2092
2093static void timehist_update_runtime_stats(struct thread_runtime *r,
2094					 u64 t, u64 tprev)
2095{
2096	r->dt_delay   = 0;
2097	r->dt_sleep   = 0;
2098	r->dt_iowait  = 0;
2099	r->dt_preempt = 0;
2100	r->dt_run     = 0;
2101
2102	if (tprev) {
2103		r->dt_run = t - tprev;
2104		if (r->ready_to_run) {
2105			if (r->ready_to_run > tprev)
2106				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2107			else
2108				r->dt_delay = tprev - r->ready_to_run;
2109		}
2110
2111		if (r->last_time > tprev)
2112			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2113		else if (r->last_time) {
2114			u64 dt_wait = tprev - r->last_time;
2115
2116			if (r->last_state == TASK_RUNNING)
2117				r->dt_preempt = dt_wait;
2118			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2119				r->dt_iowait = dt_wait;
2120			else
2121				r->dt_sleep = dt_wait;
2122		}
2123	}
2124
2125	update_stats(&r->run_stats, r->dt_run);
2126
2127	r->total_run_time     += r->dt_run;
2128	r->total_delay_time   += r->dt_delay;
2129	r->total_sleep_time   += r->dt_sleep;
2130	r->total_iowait_time  += r->dt_iowait;
2131	r->total_preempt_time += r->dt_preempt;
2132}
2133
2134static bool is_idle_sample(struct perf_sample *sample,
2135			   struct evsel *evsel)
2136{
2137	/* pid 0 == swapper == idle task */
2138	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2139		return evsel__intval(evsel, sample, "prev_pid") == 0;
2140
2141	return sample->pid == 0;
2142}
2143
2144static void save_task_callchain(struct perf_sched *sched,
2145				struct perf_sample *sample,
2146				struct evsel *evsel,
2147				struct machine *machine)
2148{
2149	struct callchain_cursor *cursor = &callchain_cursor;
2150	struct thread *thread;
2151
2152	/* want main thread for process - has maps */
2153	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2154	if (thread == NULL) {
2155		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2156		return;
2157	}
2158
2159	if (!sched->show_callchain || sample->callchain == NULL)
2160		return;
2161
2162	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2163				      NULL, NULL, sched->max_stack + 2) != 0) {
2164		if (verbose > 0)
2165			pr_err("Failed to resolve callchain. Skipping\n");
2166
2167		return;
2168	}
2169
2170	callchain_cursor_commit(cursor);
2171
2172	while (true) {
2173		struct callchain_cursor_node *node;
2174		struct symbol *sym;
2175
2176		node = callchain_cursor_current(cursor);
2177		if (node == NULL)
2178			break;
2179
2180		sym = node->ms.sym;
2181		if (sym) {
2182			if (!strcmp(sym->name, "schedule") ||
2183			    !strcmp(sym->name, "__schedule") ||
2184			    !strcmp(sym->name, "preempt_schedule"))
2185				sym->ignore = 1;
2186		}
2187
2188		callchain_cursor_advance(cursor);
2189	}
2190}
2191
2192static int init_idle_thread(struct thread *thread)
2193{
2194	struct idle_thread_runtime *itr;
2195
2196	thread__set_comm(thread, idle_comm, 0);
2197
2198	itr = zalloc(sizeof(*itr));
2199	if (itr == NULL)
2200		return -ENOMEM;
2201
2202	init_stats(&itr->tr.run_stats);
2203	callchain_init(&itr->callchain);
2204	callchain_cursor_reset(&itr->cursor);
2205	thread__set_priv(thread, itr);
2206
2207	return 0;
2208}
2209
2210/*
2211 * Track idle stats per cpu by maintaining a local thread
2212 * struct for the idle task on each cpu.
2213 */
2214static int init_idle_threads(int ncpu)
2215{
2216	int i, ret;
2217
2218	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2219	if (!idle_threads)
2220		return -ENOMEM;
2221
2222	idle_max_cpu = ncpu;
2223
2224	/* allocate the actual thread struct if needed */
2225	for (i = 0; i < ncpu; ++i) {
2226		idle_threads[i] = thread__new(0, 0);
2227		if (idle_threads[i] == NULL)
2228			return -ENOMEM;
2229
2230		ret = init_idle_thread(idle_threads[i]);
2231		if (ret < 0)
2232			return ret;
2233	}
2234
2235	return 0;
2236}
2237
2238static void free_idle_threads(void)
2239{
2240	int i;
2241
2242	if (idle_threads == NULL)
2243		return;
2244
2245	for (i = 0; i < idle_max_cpu; ++i) {
2246		if ((idle_threads[i]))
2247			thread__delete(idle_threads[i]);
2248	}
2249
2250	free(idle_threads);
2251}
2252
2253static struct thread *get_idle_thread(int cpu)
2254{
2255	/*
2256	 * expand/allocate array of pointers to local thread
2257	 * structs if needed
2258	 */
2259	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2260		int i, j = __roundup_pow_of_two(cpu+1);
2261		void *p;
2262
2263		p = realloc(idle_threads, j * sizeof(struct thread *));
2264		if (!p)
2265			return NULL;
2266
2267		idle_threads = (struct thread **) p;
2268		for (i = idle_max_cpu; i < j; ++i)
2269			idle_threads[i] = NULL;
2270
2271		idle_max_cpu = j;
2272	}
2273
2274	/* allocate a new thread struct if needed */
2275	if (idle_threads[cpu] == NULL) {
2276		idle_threads[cpu] = thread__new(0, 0);
2277
2278		if (idle_threads[cpu]) {
2279			if (init_idle_thread(idle_threads[cpu]) < 0)
2280				return NULL;
2281		}
2282	}
2283
2284	return idle_threads[cpu];
2285}
2286
2287static void save_idle_callchain(struct perf_sched *sched,
2288				struct idle_thread_runtime *itr,
2289				struct perf_sample *sample)
2290{
2291	if (!sched->show_callchain || sample->callchain == NULL)
2292		return;
2293
2294	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2295}
2296
2297static struct thread *timehist_get_thread(struct perf_sched *sched,
2298					  struct perf_sample *sample,
2299					  struct machine *machine,
2300					  struct evsel *evsel)
2301{
2302	struct thread *thread;
2303
2304	if (is_idle_sample(sample, evsel)) {
2305		thread = get_idle_thread(sample->cpu);
2306		if (thread == NULL)
2307			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2308
2309	} else {
2310		/* there were samples with tid 0 but non-zero pid */
2311		thread = machine__findnew_thread(machine, sample->pid,
2312						 sample->tid ?: sample->pid);
2313		if (thread == NULL) {
2314			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2315				 sample->tid);
2316		}
2317
2318		save_task_callchain(sched, sample, evsel, machine);
2319		if (sched->idle_hist) {
2320			struct thread *idle;
2321			struct idle_thread_runtime *itr;
2322
2323			idle = get_idle_thread(sample->cpu);
2324			if (idle == NULL) {
2325				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2326				return NULL;
2327			}
2328
2329			itr = thread__priv(idle);
2330			if (itr == NULL)
2331				return NULL;
2332
2333			itr->last_thread = thread;
2334
2335			/* copy task callchain when entering to idle */
2336			if (evsel__intval(evsel, sample, "next_pid") == 0)
2337				save_idle_callchain(sched, itr, sample);
2338		}
2339	}
2340
2341	return thread;
2342}
2343
2344static bool timehist_skip_sample(struct perf_sched *sched,
2345				 struct thread *thread,
2346				 struct evsel *evsel,
2347				 struct perf_sample *sample)
2348{
2349	bool rc = false;
2350
2351	if (thread__is_filtered(thread)) {
2352		rc = true;
2353		sched->skipped_samples++;
2354	}
2355
2356	if (sched->idle_hist) {
2357		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2358			rc = true;
2359		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2360			 evsel__intval(evsel, sample, "next_pid") != 0)
2361			rc = true;
2362	}
2363
2364	return rc;
2365}
2366
2367static void timehist_print_wakeup_event(struct perf_sched *sched,
2368					struct evsel *evsel,
2369					struct perf_sample *sample,
2370					struct machine *machine,
2371					struct thread *awakened)
2372{
2373	struct thread *thread;
2374	char tstr[64];
2375
2376	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2377	if (thread == NULL)
2378		return;
2379
2380	/* show wakeup unless both awakee and awaker are filtered */
2381	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2382	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2383		return;
2384	}
2385
2386	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2387	printf("%15s [%04d] ", tstr, sample->cpu);
2388	if (sched->show_cpu_visual)
2389		printf(" %*s ", sched->max_cpu + 1, "");
2390
2391	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2392
2393	/* dt spacer */
2394	printf("  %9s  %9s  %9s ", "", "", "");
2395
2396	printf("awakened: %s", timehist_get_commstr(awakened));
2397
2398	printf("\n");
2399}
2400
2401static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2402					union perf_event *event __maybe_unused,
2403					struct evsel *evsel __maybe_unused,
2404					struct perf_sample *sample __maybe_unused,
2405					struct machine *machine __maybe_unused)
2406{
2407	return 0;
2408}
2409
2410static int timehist_sched_wakeup_event(struct perf_tool *tool,
2411				       union perf_event *event __maybe_unused,
2412				       struct evsel *evsel,
2413				       struct perf_sample *sample,
2414				       struct machine *machine)
2415{
2416	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2417	struct thread *thread;
2418	struct thread_runtime *tr = NULL;
2419	/* want pid of awakened task not pid in sample */
2420	const u32 pid = evsel__intval(evsel, sample, "pid");
2421
2422	thread = machine__findnew_thread(machine, 0, pid);
2423	if (thread == NULL)
2424		return -1;
2425
2426	tr = thread__get_runtime(thread);
2427	if (tr == NULL)
2428		return -1;
2429
2430	if (tr->ready_to_run == 0)
2431		tr->ready_to_run = sample->time;
2432
2433	/* show wakeups if requested */
2434	if (sched->show_wakeups &&
2435	    !perf_time__skip_sample(&sched->ptime, sample->time))
2436		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2437
2438	return 0;
2439}
2440
2441static void timehist_print_migration_event(struct perf_sched *sched,
2442					struct evsel *evsel,
2443					struct perf_sample *sample,
2444					struct machine *machine,
2445					struct thread *migrated)
2446{
2447	struct thread *thread;
2448	char tstr[64];
2449	u32 max_cpus = sched->max_cpu + 1;
2450	u32 ocpu, dcpu;
2451
2452	if (sched->summary_only)
2453		return;
2454
2455	max_cpus = sched->max_cpu + 1;
2456	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2457	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2458
2459	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2460	if (thread == NULL)
2461		return;
2462
2463	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2464	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2465		return;
2466	}
2467
2468	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2469	printf("%15s [%04d] ", tstr, sample->cpu);
2470
2471	if (sched->show_cpu_visual) {
2472		u32 i;
2473		char c;
2474
2475		printf("  ");
2476		for (i = 0; i < max_cpus; ++i) {
2477			c = (i == sample->cpu) ? 'm' : ' ';
2478			printf("%c", c);
2479		}
2480		printf("  ");
2481	}
2482
2483	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2484
2485	/* dt spacer */
2486	printf("  %9s  %9s  %9s ", "", "", "");
2487
2488	printf("migrated: %s", timehist_get_commstr(migrated));
2489	printf(" cpu %d => %d", ocpu, dcpu);
2490
2491	printf("\n");
2492}
2493
2494static int timehist_migrate_task_event(struct perf_tool *tool,
2495				       union perf_event *event __maybe_unused,
2496				       struct evsel *evsel,
2497				       struct perf_sample *sample,
2498				       struct machine *machine)
2499{
2500	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2501	struct thread *thread;
2502	struct thread_runtime *tr = NULL;
2503	/* want pid of migrated task not pid in sample */
2504	const u32 pid = evsel__intval(evsel, sample, "pid");
2505
2506	thread = machine__findnew_thread(machine, 0, pid);
2507	if (thread == NULL)
2508		return -1;
2509
2510	tr = thread__get_runtime(thread);
2511	if (tr == NULL)
2512		return -1;
2513
2514	tr->migrations++;
2515
2516	/* show migrations if requested */
2517	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2518
2519	return 0;
2520}
2521
2522static int timehist_sched_change_event(struct perf_tool *tool,
2523				       union perf_event *event,
2524				       struct evsel *evsel,
2525				       struct perf_sample *sample,
2526				       struct machine *machine)
2527{
2528	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2529	struct perf_time_interval *ptime = &sched->ptime;
2530	struct addr_location al;
2531	struct thread *thread;
2532	struct thread_runtime *tr = NULL;
2533	u64 tprev, t = sample->time;
2534	int rc = 0;
2535	int state = evsel__intval(evsel, sample, "prev_state");
 
2536
2537	if (machine__resolve(machine, &al, sample) < 0) {
2538		pr_err("problem processing %d event. skipping it\n",
2539		       event->header.type);
2540		rc = -1;
2541		goto out;
2542	}
2543
2544	thread = timehist_get_thread(sched, sample, machine, evsel);
2545	if (thread == NULL) {
2546		rc = -1;
2547		goto out;
2548	}
2549
2550	if (timehist_skip_sample(sched, thread, evsel, sample))
2551		goto out;
2552
2553	tr = thread__get_runtime(thread);
2554	if (tr == NULL) {
2555		rc = -1;
2556		goto out;
2557	}
2558
2559	tprev = evsel__get_time(evsel, sample->cpu);
2560
2561	/*
2562	 * If start time given:
2563	 * - sample time is under window user cares about - skip sample
2564	 * - tprev is under window user cares about  - reset to start of window
2565	 */
2566	if (ptime->start && ptime->start > t)
2567		goto out;
2568
2569	if (tprev && ptime->start > tprev)
2570		tprev = ptime->start;
2571
2572	/*
2573	 * If end time given:
2574	 * - previous sched event is out of window - we are done
2575	 * - sample time is beyond window user cares about - reset it
2576	 *   to close out stats for time window interest
2577	 */
2578	if (ptime->end) {
2579		if (tprev > ptime->end)
2580			goto out;
2581
2582		if (t > ptime->end)
2583			t = ptime->end;
2584	}
2585
2586	if (!sched->idle_hist || thread->tid == 0) {
2587		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2588			timehist_update_runtime_stats(tr, t, tprev);
2589
2590		if (sched->idle_hist) {
2591			struct idle_thread_runtime *itr = (void *)tr;
2592			struct thread_runtime *last_tr;
2593
2594			BUG_ON(thread->tid != 0);
2595
2596			if (itr->last_thread == NULL)
2597				goto out;
2598
2599			/* add current idle time as last thread's runtime */
2600			last_tr = thread__get_runtime(itr->last_thread);
2601			if (last_tr == NULL)
2602				goto out;
2603
2604			timehist_update_runtime_stats(last_tr, t, tprev);
2605			/*
2606			 * remove delta time of last thread as it's not updated
2607			 * and otherwise it will show an invalid value next
2608			 * time.  we only care total run time and run stat.
2609			 */
2610			last_tr->dt_run = 0;
2611			last_tr->dt_delay = 0;
2612			last_tr->dt_sleep = 0;
2613			last_tr->dt_iowait = 0;
2614			last_tr->dt_preempt = 0;
2615
2616			if (itr->cursor.nr)
2617				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2618
2619			itr->last_thread = NULL;
2620		}
2621	}
2622
2623	if (!sched->summary_only)
2624		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2625
2626out:
2627	if (sched->hist_time.start == 0 && t >= ptime->start)
2628		sched->hist_time.start = t;
2629	if (ptime->end == 0 || t <= ptime->end)
2630		sched->hist_time.end = t;
2631
2632	if (tr) {
2633		/* time of this sched_switch event becomes last time task seen */
2634		tr->last_time = sample->time;
2635
2636		/* last state is used to determine where to account wait time */
2637		tr->last_state = state;
2638
2639		/* sched out event for task so reset ready to run time */
2640		tr->ready_to_run = 0;
2641	}
2642
2643	evsel__save_time(evsel, sample->time, sample->cpu);
2644
2645	return rc;
2646}
2647
2648static int timehist_sched_switch_event(struct perf_tool *tool,
2649			     union perf_event *event,
2650			     struct evsel *evsel,
2651			     struct perf_sample *sample,
2652			     struct machine *machine __maybe_unused)
2653{
2654	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2655}
2656
2657static int process_lost(struct perf_tool *tool __maybe_unused,
2658			union perf_event *event,
2659			struct perf_sample *sample,
2660			struct machine *machine __maybe_unused)
2661{
2662	char tstr[64];
2663
2664	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2665	printf("%15s ", tstr);
2666	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2667
2668	return 0;
2669}
2670
2671
2672static void print_thread_runtime(struct thread *t,
2673				 struct thread_runtime *r)
2674{
2675	double mean = avg_stats(&r->run_stats);
2676	float stddev;
2677
2678	printf("%*s   %5d  %9" PRIu64 " ",
2679	       comm_width, timehist_get_commstr(t), t->ppid,
2680	       (u64) r->run_stats.n);
2681
2682	print_sched_time(r->total_run_time, 8);
2683	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2684	print_sched_time(r->run_stats.min, 6);
2685	printf(" ");
2686	print_sched_time((u64) mean, 6);
2687	printf(" ");
2688	print_sched_time(r->run_stats.max, 6);
2689	printf("  ");
2690	printf("%5.2f", stddev);
2691	printf("   %5" PRIu64, r->migrations);
2692	printf("\n");
2693}
2694
2695static void print_thread_waittime(struct thread *t,
2696				  struct thread_runtime *r)
2697{
2698	printf("%*s   %5d  %9" PRIu64 " ",
2699	       comm_width, timehist_get_commstr(t), t->ppid,
2700	       (u64) r->run_stats.n);
2701
2702	print_sched_time(r->total_run_time, 8);
2703	print_sched_time(r->total_sleep_time, 6);
2704	printf(" ");
2705	print_sched_time(r->total_iowait_time, 6);
2706	printf(" ");
2707	print_sched_time(r->total_preempt_time, 6);
2708	printf(" ");
2709	print_sched_time(r->total_delay_time, 6);
2710	printf("\n");
2711}
2712
2713struct total_run_stats {
2714	struct perf_sched *sched;
2715	u64  sched_count;
2716	u64  task_count;
2717	u64  total_run_time;
2718};
2719
2720static int __show_thread_runtime(struct thread *t, void *priv)
2721{
2722	struct total_run_stats *stats = priv;
2723	struct thread_runtime *r;
2724
2725	if (thread__is_filtered(t))
2726		return 0;
2727
2728	r = thread__priv(t);
2729	if (r && r->run_stats.n) {
2730		stats->task_count++;
2731		stats->sched_count += r->run_stats.n;
2732		stats->total_run_time += r->total_run_time;
2733
2734		if (stats->sched->show_state)
2735			print_thread_waittime(t, r);
2736		else
2737			print_thread_runtime(t, r);
2738	}
2739
2740	return 0;
2741}
2742
2743static int show_thread_runtime(struct thread *t, void *priv)
2744{
2745	if (t->dead)
2746		return 0;
2747
2748	return __show_thread_runtime(t, priv);
2749}
2750
2751static int show_deadthread_runtime(struct thread *t, void *priv)
2752{
2753	if (!t->dead)
2754		return 0;
2755
2756	return __show_thread_runtime(t, priv);
2757}
2758
2759static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2760{
2761	const char *sep = " <- ";
2762	struct callchain_list *chain;
2763	size_t ret = 0;
2764	char bf[1024];
2765	bool first;
2766
2767	if (node == NULL)
2768		return 0;
2769
2770	ret = callchain__fprintf_folded(fp, node->parent);
2771	first = (ret == 0);
2772
2773	list_for_each_entry(chain, &node->val, list) {
2774		if (chain->ip >= PERF_CONTEXT_MAX)
2775			continue;
2776		if (chain->ms.sym && chain->ms.sym->ignore)
2777			continue;
2778		ret += fprintf(fp, "%s%s", first ? "" : sep,
2779			       callchain_list__sym_name(chain, bf, sizeof(bf),
2780							false));
2781		first = false;
2782	}
2783
2784	return ret;
2785}
2786
2787static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2788{
2789	size_t ret = 0;
2790	FILE *fp = stdout;
2791	struct callchain_node *chain;
2792	struct rb_node *rb_node = rb_first_cached(root);
2793
2794	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2795	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2796	       graph_dotted_line);
2797
2798	while (rb_node) {
2799		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2800		rb_node = rb_next(rb_node);
2801
2802		ret += fprintf(fp, "  ");
2803		print_sched_time(chain->hit, 12);
2804		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2805		ret += fprintf(fp, " %8d  ", chain->count);
2806		ret += callchain__fprintf_folded(fp, chain);
2807		ret += fprintf(fp, "\n");
2808	}
2809
2810	return ret;
2811}
2812
2813static void timehist_print_summary(struct perf_sched *sched,
2814				   struct perf_session *session)
2815{
2816	struct machine *m = &session->machines.host;
2817	struct total_run_stats totals;
2818	u64 task_count;
2819	struct thread *t;
2820	struct thread_runtime *r;
2821	int i;
2822	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2823
2824	memset(&totals, 0, sizeof(totals));
2825	totals.sched = sched;
2826
2827	if (sched->idle_hist) {
2828		printf("\nIdle-time summary\n");
2829		printf("%*s  parent  sched-out  ", comm_width, "comm");
2830		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2831	} else if (sched->show_state) {
2832		printf("\nWait-time summary\n");
2833		printf("%*s  parent   sched-in  ", comm_width, "comm");
2834		printf("   run-time      sleep      iowait     preempt       delay\n");
2835	} else {
2836		printf("\nRuntime summary\n");
2837		printf("%*s  parent   sched-in  ", comm_width, "comm");
2838		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2839	}
2840	printf("%*s            (count)  ", comm_width, "");
2841	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2842	       sched->show_state ? "(msec)" : "%");
2843	printf("%.117s\n", graph_dotted_line);
2844
2845	machine__for_each_thread(m, show_thread_runtime, &totals);
2846	task_count = totals.task_count;
2847	if (!task_count)
2848		printf("<no still running tasks>\n");
2849
2850	printf("\nTerminated tasks:\n");
2851	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2852	if (task_count == totals.task_count)
2853		printf("<no terminated tasks>\n");
2854
2855	/* CPU idle stats not tracked when samples were skipped */
2856	if (sched->skipped_samples && !sched->idle_hist)
2857		return;
2858
2859	printf("\nIdle stats:\n");
2860	for (i = 0; i < idle_max_cpu; ++i) {
2861		if (cpu_list && !test_bit(i, cpu_bitmap))
2862			continue;
2863
2864		t = idle_threads[i];
2865		if (!t)
2866			continue;
2867
2868		r = thread__priv(t);
2869		if (r && r->run_stats.n) {
2870			totals.sched_count += r->run_stats.n;
2871			printf("    CPU %2d idle for ", i);
2872			print_sched_time(r->total_run_time, 6);
2873			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2874		} else
2875			printf("    CPU %2d idle entire time window\n", i);
2876	}
2877
2878	if (sched->idle_hist && sched->show_callchain) {
2879		callchain_param.mode  = CHAIN_FOLDED;
2880		callchain_param.value = CCVAL_PERIOD;
2881
2882		callchain_register_param(&callchain_param);
2883
2884		printf("\nIdle stats by callchain:\n");
2885		for (i = 0; i < idle_max_cpu; ++i) {
2886			struct idle_thread_runtime *itr;
2887
2888			t = idle_threads[i];
2889			if (!t)
2890				continue;
2891
2892			itr = thread__priv(t);
2893			if (itr == NULL)
2894				continue;
2895
2896			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2897					     0, &callchain_param);
2898
2899			printf("  CPU %2d:", i);
2900			print_sched_time(itr->tr.total_run_time, 6);
2901			printf(" msec\n");
2902			timehist_print_idlehist_callchain(&itr->sorted_root);
2903			printf("\n");
2904		}
2905	}
2906
2907	printf("\n"
2908	       "    Total number of unique tasks: %" PRIu64 "\n"
2909	       "Total number of context switches: %" PRIu64 "\n",
2910	       totals.task_count, totals.sched_count);
2911
2912	printf("           Total run time (msec): ");
2913	print_sched_time(totals.total_run_time, 2);
2914	printf("\n");
2915
2916	printf("    Total scheduling time (msec): ");
2917	print_sched_time(hist_time, 2);
2918	printf(" (x %d)\n", sched->max_cpu);
2919}
2920
2921typedef int (*sched_handler)(struct perf_tool *tool,
2922			  union perf_event *event,
2923			  struct evsel *evsel,
2924			  struct perf_sample *sample,
2925			  struct machine *machine);
2926
2927static int perf_timehist__process_sample(struct perf_tool *tool,
2928					 union perf_event *event,
2929					 struct perf_sample *sample,
2930					 struct evsel *evsel,
2931					 struct machine *machine)
2932{
2933	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2934	int err = 0;
2935	int this_cpu = sample->cpu;
2936
2937	if (this_cpu > sched->max_cpu)
2938		sched->max_cpu = this_cpu;
2939
2940	if (evsel->handler != NULL) {
2941		sched_handler f = evsel->handler;
2942
2943		err = f(tool, event, evsel, sample, machine);
2944	}
2945
2946	return err;
2947}
2948
2949static int timehist_check_attr(struct perf_sched *sched,
2950			       struct evlist *evlist)
2951{
2952	struct evsel *evsel;
2953	struct evsel_runtime *er;
2954
2955	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2956		er = evsel__get_runtime(evsel);
2957		if (er == NULL) {
2958			pr_err("Failed to allocate memory for evsel runtime data\n");
2959			return -1;
2960		}
2961
2962		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
 
2963			pr_info("Samples do not have callchains.\n");
2964			sched->show_callchain = 0;
2965			symbol_conf.use_callchain = 0;
2966		}
2967	}
2968
2969	return 0;
2970}
2971
2972static int perf_sched__timehist(struct perf_sched *sched)
2973{
2974	struct evsel_str_handler handlers[] = {
2975		{ "sched:sched_switch",       timehist_sched_switch_event, },
2976		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2977		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
2978		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2979	};
2980	const struct evsel_str_handler migrate_handlers[] = {
2981		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2982	};
2983	struct perf_data data = {
2984		.path  = input_name,
2985		.mode  = PERF_DATA_MODE_READ,
2986		.force = sched->force,
 
 
2987	};
2988
2989	struct perf_session *session;
2990	struct evlist *evlist;
2991	int err = -1;
2992
2993	/*
2994	 * event handlers for timehist option
2995	 */
2996	sched->tool.sample	 = perf_timehist__process_sample;
2997	sched->tool.mmap	 = perf_event__process_mmap;
2998	sched->tool.comm	 = perf_event__process_comm;
2999	sched->tool.exit	 = perf_event__process_exit;
3000	sched->tool.fork	 = perf_event__process_fork;
3001	sched->tool.lost	 = process_lost;
3002	sched->tool.attr	 = perf_event__process_attr;
3003	sched->tool.tracing_data = perf_event__process_tracing_data;
3004	sched->tool.build_id	 = perf_event__process_build_id;
3005
3006	sched->tool.ordered_events = true;
3007	sched->tool.ordering_requires_timestamps = true;
3008
3009	symbol_conf.use_callchain = sched->show_callchain;
3010
3011	session = perf_session__new(&data, false, &sched->tool);
3012	if (IS_ERR(session))
3013		return PTR_ERR(session);
3014
3015	if (cpu_list) {
3016		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3017		if (err < 0)
3018			goto out;
3019	}
3020
3021	evlist = session->evlist;
3022
3023	symbol__init(&session->header.env);
3024
3025	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3026		pr_err("Invalid time string\n");
3027		return -EINVAL;
3028	}
3029
3030	if (timehist_check_attr(sched, evlist) != 0)
3031		goto out;
3032
3033	setup_pager();
3034
3035	/* prefer sched_waking if it is captured */
3036	if (perf_evlist__find_tracepoint_by_name(session->evlist,
3037						  "sched:sched_waking"))
3038		handlers[1].handler = timehist_sched_wakeup_ignore;
3039
3040	/* setup per-evsel handlers */
3041	if (perf_session__set_tracepoints_handlers(session, handlers))
3042		goto out;
3043
3044	/* sched_switch event at a minimum needs to exist */
3045	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3046						  "sched:sched_switch")) {
3047		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3048		goto out;
3049	}
3050
3051	if (sched->show_migrations &&
3052	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3053		goto out;
3054
3055	/* pre-allocate struct for per-CPU idle stats */
3056	sched->max_cpu = session->header.env.nr_cpus_online;
3057	if (sched->max_cpu == 0)
3058		sched->max_cpu = 4;
3059	if (init_idle_threads(sched->max_cpu))
3060		goto out;
3061
3062	/* summary_only implies summary option, but don't overwrite summary if set */
3063	if (sched->summary_only)
3064		sched->summary = sched->summary_only;
3065
3066	if (!sched->summary_only)
3067		timehist_header(sched);
3068
3069	err = perf_session__process_events(session);
3070	if (err) {
3071		pr_err("Failed to process events, error %d", err);
3072		goto out;
3073	}
3074
3075	sched->nr_events      = evlist->stats.nr_events[0];
3076	sched->nr_lost_events = evlist->stats.total_lost;
3077	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3078
3079	if (sched->summary)
3080		timehist_print_summary(sched, session);
3081
3082out:
3083	free_idle_threads();
3084	perf_session__delete(session);
3085
3086	return err;
3087}
3088
3089
3090static void print_bad_events(struct perf_sched *sched)
3091{
3092	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3093		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3094			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3095			sched->nr_unordered_timestamps, sched->nr_timestamps);
3096	}
3097	if (sched->nr_lost_events && sched->nr_events) {
3098		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3099			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3100			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3101	}
3102	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3103		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3104			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3105			sched->nr_context_switch_bugs, sched->nr_timestamps);
3106		if (sched->nr_lost_events)
3107			printf(" (due to lost events?)");
3108		printf("\n");
3109	}
3110}
3111
3112static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3113{
3114	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3115	struct work_atoms *this;
3116	const char *comm = thread__comm_str(data->thread), *this_comm;
3117	bool leftmost = true;
3118
3119	while (*new) {
3120		int cmp;
3121
3122		this = container_of(*new, struct work_atoms, node);
3123		parent = *new;
3124
3125		this_comm = thread__comm_str(this->thread);
3126		cmp = strcmp(comm, this_comm);
3127		if (cmp > 0) {
3128			new = &((*new)->rb_left);
3129		} else if (cmp < 0) {
3130			new = &((*new)->rb_right);
3131			leftmost = false;
3132		} else {
3133			this->num_merged++;
3134			this->total_runtime += data->total_runtime;
3135			this->nb_atoms += data->nb_atoms;
3136			this->total_lat += data->total_lat;
3137			list_splice(&data->work_list, &this->work_list);
3138			if (this->max_lat < data->max_lat) {
3139				this->max_lat = data->max_lat;
3140				this->max_lat_at = data->max_lat_at;
3141			}
3142			zfree(&data);
3143			return;
3144		}
3145	}
3146
3147	data->num_merged++;
3148	rb_link_node(&data->node, parent, new);
3149	rb_insert_color_cached(&data->node, root, leftmost);
3150}
3151
3152static void perf_sched__merge_lat(struct perf_sched *sched)
3153{
3154	struct work_atoms *data;
3155	struct rb_node *node;
3156
3157	if (sched->skip_merge)
3158		return;
3159
3160	while ((node = rb_first_cached(&sched->atom_root))) {
3161		rb_erase_cached(node, &sched->atom_root);
3162		data = rb_entry(node, struct work_atoms, node);
3163		__merge_work_atoms(&sched->merged_atom_root, data);
3164	}
3165}
3166
3167static int perf_sched__lat(struct perf_sched *sched)
3168{
3169	struct rb_node *next;
3170
3171	setup_pager();
3172
3173	if (perf_sched__read_events(sched))
3174		return -1;
3175
3176	perf_sched__merge_lat(sched);
3177	perf_sched__sort_lat(sched);
3178
3179	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3180	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3181	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3182
3183	next = rb_first_cached(&sched->sorted_atom_root);
3184
3185	while (next) {
3186		struct work_atoms *work_list;
3187
3188		work_list = rb_entry(next, struct work_atoms, node);
3189		output_lat_thread(sched, work_list);
3190		next = rb_next(next);
3191		thread__zput(work_list->thread);
3192	}
3193
3194	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3195	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3196		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3197
3198	printf(" ---------------------------------------------------\n");
3199
3200	print_bad_events(sched);
3201	printf("\n");
3202
3203	return 0;
3204}
3205
3206static int setup_map_cpus(struct perf_sched *sched)
3207{
3208	struct perf_cpu_map *map;
3209
3210	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3211
3212	if (sched->map.comp) {
3213		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3214		if (!sched->map.comp_cpus)
3215			return -1;
3216	}
3217
3218	if (!sched->map.cpus_str)
3219		return 0;
3220
3221	map = perf_cpu_map__new(sched->map.cpus_str);
3222	if (!map) {
3223		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3224		return -1;
3225	}
3226
3227	sched->map.cpus = map;
3228	return 0;
3229}
3230
3231static int setup_color_pids(struct perf_sched *sched)
3232{
3233	struct perf_thread_map *map;
3234
3235	if (!sched->map.color_pids_str)
3236		return 0;
3237
3238	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3239	if (!map) {
3240		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3241		return -1;
3242	}
3243
3244	sched->map.color_pids = map;
3245	return 0;
3246}
3247
3248static int setup_color_cpus(struct perf_sched *sched)
3249{
3250	struct perf_cpu_map *map;
3251
3252	if (!sched->map.color_cpus_str)
3253		return 0;
3254
3255	map = perf_cpu_map__new(sched->map.color_cpus_str);
3256	if (!map) {
3257		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3258		return -1;
3259	}
3260
3261	sched->map.color_cpus = map;
3262	return 0;
3263}
3264
3265static int perf_sched__map(struct perf_sched *sched)
3266{
3267	if (setup_map_cpus(sched))
3268		return -1;
3269
3270	if (setup_color_pids(sched))
3271		return -1;
3272
3273	if (setup_color_cpus(sched))
3274		return -1;
3275
3276	setup_pager();
3277	if (perf_sched__read_events(sched))
3278		return -1;
3279	print_bad_events(sched);
3280	return 0;
3281}
3282
3283static int perf_sched__replay(struct perf_sched *sched)
3284{
3285	unsigned long i;
3286
3287	calibrate_run_measurement_overhead(sched);
3288	calibrate_sleep_measurement_overhead(sched);
3289
3290	test_calibrations(sched);
3291
3292	if (perf_sched__read_events(sched))
3293		return -1;
3294
3295	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3296	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3297	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3298
3299	if (sched->targetless_wakeups)
3300		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3301	if (sched->multitarget_wakeups)
3302		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3303	if (sched->nr_run_events_optimized)
3304		printf("run atoms optimized: %ld\n",
3305			sched->nr_run_events_optimized);
3306
3307	print_task_traces(sched);
3308	add_cross_task_wakeups(sched);
3309
3310	create_tasks(sched);
3311	printf("------------------------------------------------------------\n");
3312	for (i = 0; i < sched->replay_repeat; i++)
3313		run_one_test(sched);
3314
3315	return 0;
3316}
3317
3318static void setup_sorting(struct perf_sched *sched, const struct option *options,
3319			  const char * const usage_msg[])
3320{
3321	char *tmp, *tok, *str = strdup(sched->sort_order);
3322
3323	for (tok = strtok_r(str, ", ", &tmp);
3324			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3325		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3326			usage_with_options_msg(usage_msg, options,
3327					"Unknown --sort key: `%s'", tok);
3328		}
3329	}
3330
3331	free(str);
3332
3333	sort_dimension__add("pid", &sched->cmp_pid);
3334}
3335
3336static int __cmd_record(int argc, const char **argv)
3337{
3338	unsigned int rec_argc, i, j;
3339	const char **rec_argv;
3340	const char * const record_args[] = {
3341		"record",
3342		"-a",
3343		"-R",
3344		"-m", "1024",
3345		"-c", "1",
3346		"-e", "sched:sched_switch",
3347		"-e", "sched:sched_stat_wait",
3348		"-e", "sched:sched_stat_sleep",
3349		"-e", "sched:sched_stat_iowait",
3350		"-e", "sched:sched_stat_runtime",
3351		"-e", "sched:sched_process_fork",
 
3352		"-e", "sched:sched_wakeup_new",
3353		"-e", "sched:sched_migrate_task",
3354	};
3355	struct tep_event *waking_event;
3356
3357	/*
3358	 * +2 for either "-e", "sched:sched_wakeup" or
3359	 * "-e", "sched:sched_waking"
3360	 */
3361	rec_argc = ARRAY_SIZE(record_args) + 2 + argc - 1;
3362	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3363
3364	if (rec_argv == NULL)
3365		return -ENOMEM;
3366
3367	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3368		rec_argv[i] = strdup(record_args[i]);
3369
3370	rec_argv[i++] = "-e";
3371	waking_event = trace_event__tp_format("sched", "sched_waking");
3372	if (!IS_ERR(waking_event))
3373		rec_argv[i++] = strdup("sched:sched_waking");
3374	else
3375		rec_argv[i++] = strdup("sched:sched_wakeup");
3376
3377	for (j = 1; j < (unsigned int)argc; j++, i++)
3378		rec_argv[i] = argv[j];
3379
3380	BUG_ON(i != rec_argc);
3381
3382	return cmd_record(i, rec_argv);
3383}
3384
3385int cmd_sched(int argc, const char **argv)
3386{
3387	static const char default_sort_order[] = "avg, max, switch, runtime";
3388	struct perf_sched sched = {
3389		.tool = {
3390			.sample		 = perf_sched__process_tracepoint_sample,
3391			.comm		 = perf_sched__process_comm,
3392			.namespaces	 = perf_event__process_namespaces,
3393			.lost		 = perf_event__process_lost,
3394			.fork		 = perf_sched__process_fork_event,
3395			.ordered_events = true,
3396		},
3397		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3398		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3399		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3400		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3401		.sort_order	      = default_sort_order,
3402		.replay_repeat	      = 10,
3403		.profile_cpu	      = -1,
3404		.next_shortname1      = 'A',
3405		.next_shortname2      = '0',
3406		.skip_merge           = 0,
3407		.show_callchain	      = 1,
3408		.max_stack            = 5,
3409	};
3410	const struct option sched_options[] = {
3411	OPT_STRING('i', "input", &input_name, "file",
3412		    "input file name"),
3413	OPT_INCR('v', "verbose", &verbose,
3414		    "be more verbose (show symbol address, etc)"),
3415	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3416		    "dump raw trace in ASCII"),
3417	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3418	OPT_END()
3419	};
3420	const struct option latency_options[] = {
3421	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3422		   "sort by key(s): runtime, switch, avg, max"),
3423	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3424		    "CPU to profile on"),
3425	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3426		    "latency stats per pid instead of per comm"),
3427	OPT_PARENT(sched_options)
3428	};
3429	const struct option replay_options[] = {
3430	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3431		     "repeat the workload replay N times (-1: infinite)"),
3432	OPT_PARENT(sched_options)
3433	};
3434	const struct option map_options[] = {
3435	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3436		    "map output in compact mode"),
3437	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3438		   "highlight given pids in map"),
3439	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3440                    "highlight given CPUs in map"),
3441	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3442                    "display given CPUs in map"),
3443	OPT_PARENT(sched_options)
3444	};
3445	const struct option timehist_options[] = {
3446	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3447		   "file", "vmlinux pathname"),
3448	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3449		   "file", "kallsyms pathname"),
3450	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3451		    "Display call chains if present (default on)"),
3452	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3453		   "Maximum number of functions to display backtrace."),
3454	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3455		    "Look for files with symbols relative to this directory"),
3456	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3457		    "Show only syscall summary with statistics"),
3458	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3459		    "Show all syscalls and summary with statistics"),
3460	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3461	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3462	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3463	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3464	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3465	OPT_STRING(0, "time", &sched.time_str, "str",
3466		   "Time span for analysis (start,stop)"),
3467	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3468	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3469		   "analyze events only for given process id(s)"),
3470	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3471		   "analyze events only for given thread id(s)"),
3472	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3473	OPT_PARENT(sched_options)
3474	};
3475
3476	const char * const latency_usage[] = {
3477		"perf sched latency [<options>]",
3478		NULL
3479	};
3480	const char * const replay_usage[] = {
3481		"perf sched replay [<options>]",
3482		NULL
3483	};
3484	const char * const map_usage[] = {
3485		"perf sched map [<options>]",
3486		NULL
3487	};
3488	const char * const timehist_usage[] = {
3489		"perf sched timehist [<options>]",
3490		NULL
3491	};
3492	const char *const sched_subcommands[] = { "record", "latency", "map",
3493						  "replay", "script",
3494						  "timehist", NULL };
3495	const char *sched_usage[] = {
3496		NULL,
3497		NULL
3498	};
3499	struct trace_sched_handler lat_ops  = {
3500		.wakeup_event	    = latency_wakeup_event,
3501		.switch_event	    = latency_switch_event,
3502		.runtime_event	    = latency_runtime_event,
3503		.migrate_task_event = latency_migrate_task_event,
3504	};
3505	struct trace_sched_handler map_ops  = {
3506		.switch_event	    = map_switch_event,
3507	};
3508	struct trace_sched_handler replay_ops  = {
3509		.wakeup_event	    = replay_wakeup_event,
3510		.switch_event	    = replay_switch_event,
3511		.fork_event	    = replay_fork_event,
3512	};
3513	unsigned int i;
3514
3515	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3516		sched.curr_pid[i] = -1;
3517
3518	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3519					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3520	if (!argc)
3521		usage_with_options(sched_usage, sched_options);
3522
3523	/*
3524	 * Aliased to 'perf script' for now:
3525	 */
3526	if (!strcmp(argv[0], "script"))
3527		return cmd_script(argc, argv);
3528
3529	if (!strncmp(argv[0], "rec", 3)) {
3530		return __cmd_record(argc, argv);
3531	} else if (!strncmp(argv[0], "lat", 3)) {
3532		sched.tp_handler = &lat_ops;
3533		if (argc > 1) {
3534			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3535			if (argc)
3536				usage_with_options(latency_usage, latency_options);
3537		}
3538		setup_sorting(&sched, latency_options, latency_usage);
3539		return perf_sched__lat(&sched);
3540	} else if (!strcmp(argv[0], "map")) {
3541		if (argc) {
3542			argc = parse_options(argc, argv, map_options, map_usage, 0);
3543			if (argc)
3544				usage_with_options(map_usage, map_options);
3545		}
3546		sched.tp_handler = &map_ops;
3547		setup_sorting(&sched, latency_options, latency_usage);
3548		return perf_sched__map(&sched);
3549	} else if (!strncmp(argv[0], "rep", 3)) {
3550		sched.tp_handler = &replay_ops;
3551		if (argc) {
3552			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3553			if (argc)
3554				usage_with_options(replay_usage, replay_options);
3555		}
3556		return perf_sched__replay(&sched);
3557	} else if (!strcmp(argv[0], "timehist")) {
3558		if (argc) {
3559			argc = parse_options(argc, argv, timehist_options,
3560					     timehist_usage, 0);
3561			if (argc)
3562				usage_with_options(timehist_usage, timehist_options);
3563		}
3564		if ((sched.show_wakeups || sched.show_next) &&
3565		    sched.summary_only) {
3566			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3567			parse_options_usage(timehist_usage, timehist_options, "s", true);
3568			if (sched.show_wakeups)
3569				parse_options_usage(NULL, timehist_options, "w", true);
3570			if (sched.show_next)
3571				parse_options_usage(NULL, timehist_options, "n", true);
3572			return -EINVAL;
3573		}
3574
3575		return perf_sched__timehist(&sched);
3576	} else {
3577		usage_with_options(sched_usage, sched_options);
3578	}
3579
3580	return 0;
3581}