Linux Audio

Check our new training course

Linux BSP development engineering services

Need help to port Linux and bootloaders to your hardware?
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}
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3#include "perf-sys.h"
   4
   5#include "util/cpumap.h"
   6#include "util/evlist.h"
 
   7#include "util/evsel.h"
   8#include "util/evsel_fprintf.h"
   9#include "util/mutex.h"
  10#include "util/symbol.h"
  11#include "util/thread.h"
  12#include "util/header.h"
  13#include "util/session.h"
  14#include "util/tool.h"
  15#include "util/cloexec.h"
  16#include "util/thread_map.h"
  17#include "util/color.h"
  18#include "util/stat.h"
  19#include "util/string2.h"
  20#include "util/callchain.h"
  21#include "util/time-utils.h"
  22
  23#include <subcmd/pager.h>
  24#include <subcmd/parse-options.h>
  25#include "util/trace-event.h"
  26
  27#include "util/debug.h"
  28#include "util/event.h"
  29#include "util/util.h"
  30
  31#include <linux/kernel.h>
  32#include <linux/log2.h>
  33#include <linux/zalloc.h>
  34#include <sys/prctl.h>
  35#include <sys/resource.h>
  36#include <inttypes.h>
  37
  38#include <errno.h>
  39#include <semaphore.h>
  40#include <pthread.h>
  41#include <math.h>
  42#include <api/fs/fs.h>
  43#include <perf/cpumap.h>
  44#include <linux/time64.h>
  45#include <linux/err.h>
  46
  47#include <linux/ctype.h>
  48
  49#define PR_SET_NAME		15               /* Set process name */
  50#define MAX_CPUS		4096
  51#define COMM_LEN		20
  52#define SYM_LEN			129
  53#define MAX_PID			1024000
  54
  55static const char *cpu_list;
  56static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  57
  58struct sched_atom;
  59
  60struct task_desc {
  61	unsigned long		nr;
  62	unsigned long		pid;
  63	char			comm[COMM_LEN];
  64
  65	unsigned long		nr_events;
  66	unsigned long		curr_event;
  67	struct sched_atom	**atoms;
  68
  69	pthread_t		thread;
  70	sem_t			sleep_sem;
  71
  72	sem_t			ready_for_work;
  73	sem_t			work_done_sem;
  74
  75	u64			cpu_usage;
  76};
  77
  78enum sched_event_type {
  79	SCHED_EVENT_RUN,
  80	SCHED_EVENT_SLEEP,
  81	SCHED_EVENT_WAKEUP,
  82	SCHED_EVENT_MIGRATION,
  83};
  84
  85struct sched_atom {
  86	enum sched_event_type	type;
  87	int			specific_wait;
  88	u64			timestamp;
  89	u64			duration;
  90	unsigned long		nr;
  91	sem_t			*wait_sem;
  92	struct task_desc	*wakee;
  93};
  94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  95enum thread_state {
  96	THREAD_SLEEPING = 0,
  97	THREAD_WAIT_CPU,
  98	THREAD_SCHED_IN,
  99	THREAD_IGNORE
 100};
 101
 102struct work_atom {
 103	struct list_head	list;
 104	enum thread_state	state;
 105	u64			sched_out_time;
 106	u64			wake_up_time;
 107	u64			sched_in_time;
 108	u64			runtime;
 109};
 110
 111struct work_atoms {
 112	struct list_head	work_list;
 113	struct thread		*thread;
 114	struct rb_node		node;
 115	u64			max_lat;
 116	u64			max_lat_start;
 117	u64			max_lat_end;
 118	u64			total_lat;
 119	u64			nb_atoms;
 120	u64			total_runtime;
 121	int			num_merged;
 122};
 123
 124typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 125
 126struct perf_sched;
 127
 128struct trace_sched_handler {
 129	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
 130			    struct perf_sample *sample, struct machine *machine);
 131
 132	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
 133			     struct perf_sample *sample, struct machine *machine);
 134
 135	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
 136			    struct perf_sample *sample, struct machine *machine);
 137
 138	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
 139	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
 140			  struct machine *machine);
 141
 142	int (*migrate_task_event)(struct perf_sched *sched,
 143				  struct evsel *evsel,
 144				  struct perf_sample *sample,
 145				  struct machine *machine);
 146};
 147
 148#define COLOR_PIDS PERF_COLOR_BLUE
 149#define COLOR_CPUS PERF_COLOR_BG_RED
 150
 151struct perf_sched_map {
 152	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
 153	struct perf_cpu		*comp_cpus;
 154	bool			 comp;
 155	struct perf_thread_map *color_pids;
 156	const char		*color_pids_str;
 157	struct perf_cpu_map	*color_cpus;
 158	const char		*color_cpus_str;
 159	struct perf_cpu_map	*cpus;
 160	const char		*cpus_str;
 161};
 162
 163struct perf_sched {
 164	struct perf_tool tool;
 165	const char	 *sort_order;
 166	unsigned long	 nr_tasks;
 167	struct task_desc **pid_to_task;
 168	struct task_desc **tasks;
 169	const struct trace_sched_handler *tp_handler;
 170	struct mutex	 start_work_mutex;
 171	struct mutex	 work_done_wait_mutex;
 172	int		 profile_cpu;
 173/*
 174 * Track the current task - that way we can know whether there's any
 175 * weird events, such as a task being switched away that is not current.
 176 */
 177	struct perf_cpu	 max_cpu;
 178	u32		 *curr_pid;
 179	struct thread	 **curr_thread;
 180	char		 next_shortname1;
 181	char		 next_shortname2;
 182	unsigned int	 replay_repeat;
 183	unsigned long	 nr_run_events;
 184	unsigned long	 nr_sleep_events;
 185	unsigned long	 nr_wakeup_events;
 186	unsigned long	 nr_sleep_corrections;
 187	unsigned long	 nr_run_events_optimized;
 188	unsigned long	 targetless_wakeups;
 189	unsigned long	 multitarget_wakeups;
 190	unsigned long	 nr_runs;
 191	unsigned long	 nr_timestamps;
 192	unsigned long	 nr_unordered_timestamps;
 193	unsigned long	 nr_context_switch_bugs;
 194	unsigned long	 nr_events;
 195	unsigned long	 nr_lost_chunks;
 196	unsigned long	 nr_lost_events;
 197	u64		 run_measurement_overhead;
 198	u64		 sleep_measurement_overhead;
 199	u64		 start_time;
 200	u64		 cpu_usage;
 201	u64		 runavg_cpu_usage;
 202	u64		 parent_cpu_usage;
 203	u64		 runavg_parent_cpu_usage;
 204	u64		 sum_runtime;
 205	u64		 sum_fluct;
 206	u64		 run_avg;
 207	u64		 all_runtime;
 208	u64		 all_count;
 209	u64		 *cpu_last_switched;
 210	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
 211	struct list_head sort_list, cmp_pid;
 212	bool force;
 213	bool skip_merge;
 214	struct perf_sched_map map;
 215
 216	/* options for timehist command */
 217	bool		summary;
 218	bool		summary_only;
 219	bool		idle_hist;
 220	bool		show_callchain;
 221	unsigned int	max_stack;
 222	bool		show_cpu_visual;
 223	bool		show_wakeups;
 224	bool		show_next;
 225	bool		show_migrations;
 226	bool		show_state;
 227	u64		skipped_samples;
 228	const char	*time_str;
 229	struct perf_time_interval ptime;
 230	struct perf_time_interval hist_time;
 231	volatile bool   thread_funcs_exit;
 232};
 233
 234/* per thread run time data */
 235struct thread_runtime {
 236	u64 last_time;      /* time of previous sched in/out event */
 237	u64 dt_run;         /* run time */
 238	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
 239	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
 240	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
 241	u64 dt_delay;       /* time between wakeup and sched-in */
 242	u64 ready_to_run;   /* time of wakeup */
 243
 244	struct stats run_stats;
 245	u64 total_run_time;
 246	u64 total_sleep_time;
 247	u64 total_iowait_time;
 248	u64 total_preempt_time;
 249	u64 total_delay_time;
 250
 251	char last_state;
 252
 253	char shortname[3];
 254	bool comm_changed;
 255
 256	u64 migrations;
 257};
 258
 259/* per event run time data */
 260struct evsel_runtime {
 261	u64 *last_time; /* time this event was last seen per cpu */
 262	u32 ncpu;       /* highest cpu slot allocated */
 263};
 264
 265/* per cpu idle time data */
 266struct idle_thread_runtime {
 267	struct thread_runtime	tr;
 268	struct thread		*last_thread;
 269	struct rb_root_cached	sorted_root;
 270	struct callchain_root	callchain;
 271	struct callchain_cursor	cursor;
 272};
 273
 274/* track idle times per cpu */
 275static struct thread **idle_threads;
 276static int idle_max_cpu;
 277static char idle_comm[] = "<idle>";
 278
 279static u64 get_nsecs(void)
 280{
 281	struct timespec ts;
 282
 283	clock_gettime(CLOCK_MONOTONIC, &ts);
 284
 285	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
 286}
 287
 288static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
 289{
 290	u64 T0 = get_nsecs(), T1;
 291
 292	do {
 293		T1 = get_nsecs();
 294	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
 295}
 296
 297static void sleep_nsecs(u64 nsecs)
 298{
 299	struct timespec ts;
 300
 301	ts.tv_nsec = nsecs % 999999999;
 302	ts.tv_sec = nsecs / 999999999;
 303
 304	nanosleep(&ts, NULL);
 305}
 306
 307static void calibrate_run_measurement_overhead(struct perf_sched *sched)
 308{
 309	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 310	int i;
 311
 312	for (i = 0; i < 10; i++) {
 313		T0 = get_nsecs();
 314		burn_nsecs(sched, 0);
 315		T1 = get_nsecs();
 316		delta = T1-T0;
 317		min_delta = min(min_delta, delta);
 318	}
 319	sched->run_measurement_overhead = min_delta;
 320
 321	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 322}
 323
 324static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
 325{
 326	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
 327	int i;
 328
 329	for (i = 0; i < 10; i++) {
 330		T0 = get_nsecs();
 331		sleep_nsecs(10000);
 332		T1 = get_nsecs();
 333		delta = T1-T0;
 334		min_delta = min(min_delta, delta);
 335	}
 336	min_delta -= 10000;
 337	sched->sleep_measurement_overhead = min_delta;
 338
 339	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
 340}
 341
 342static struct sched_atom *
 343get_new_event(struct task_desc *task, u64 timestamp)
 344{
 345	struct sched_atom *event = zalloc(sizeof(*event));
 346	unsigned long idx = task->nr_events;
 347	size_t size;
 348
 349	event->timestamp = timestamp;
 350	event->nr = idx;
 351
 352	task->nr_events++;
 353	size = sizeof(struct sched_atom *) * task->nr_events;
 354	task->atoms = realloc(task->atoms, size);
 355	BUG_ON(!task->atoms);
 356
 357	task->atoms[idx] = event;
 358
 359	return event;
 360}
 361
 362static struct sched_atom *last_event(struct task_desc *task)
 363{
 364	if (!task->nr_events)
 365		return NULL;
 366
 367	return task->atoms[task->nr_events - 1];
 368}
 369
 370static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
 371				u64 timestamp, u64 duration)
 372{
 373	struct sched_atom *event, *curr_event = last_event(task);
 374
 375	/*
 376	 * optimize an existing RUN event by merging this one
 377	 * to it:
 378	 */
 379	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
 380		sched->nr_run_events_optimized++;
 381		curr_event->duration += duration;
 382		return;
 383	}
 384
 385	event = get_new_event(task, timestamp);
 386
 387	event->type = SCHED_EVENT_RUN;
 388	event->duration = duration;
 389
 390	sched->nr_run_events++;
 391}
 392
 393static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
 394				   u64 timestamp, struct task_desc *wakee)
 395{
 396	struct sched_atom *event, *wakee_event;
 397
 398	event = get_new_event(task, timestamp);
 399	event->type = SCHED_EVENT_WAKEUP;
 400	event->wakee = wakee;
 401
 402	wakee_event = last_event(wakee);
 403	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
 404		sched->targetless_wakeups++;
 405		return;
 406	}
 407	if (wakee_event->wait_sem) {
 408		sched->multitarget_wakeups++;
 409		return;
 410	}
 411
 412	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
 413	sem_init(wakee_event->wait_sem, 0, 0);
 414	wakee_event->specific_wait = 1;
 415	event->wait_sem = wakee_event->wait_sem;
 416
 417	sched->nr_wakeup_events++;
 418}
 419
 420static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
 421				  u64 timestamp, const char task_state __maybe_unused)
 422{
 423	struct sched_atom *event = get_new_event(task, timestamp);
 424
 425	event->type = SCHED_EVENT_SLEEP;
 426
 427	sched->nr_sleep_events++;
 428}
 429
 430static struct task_desc *register_pid(struct perf_sched *sched,
 431				      unsigned long pid, const char *comm)
 432{
 433	struct task_desc *task;
 434	static int pid_max;
 435
 436	if (sched->pid_to_task == NULL) {
 437		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
 438			pid_max = MAX_PID;
 439		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
 440	}
 441	if (pid >= (unsigned long)pid_max) {
 442		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
 443			sizeof(struct task_desc *))) == NULL);
 444		while (pid >= (unsigned long)pid_max)
 445			sched->pid_to_task[pid_max++] = NULL;
 446	}
 447
 448	task = sched->pid_to_task[pid];
 449
 450	if (task)
 451		return task;
 452
 453	task = zalloc(sizeof(*task));
 454	task->pid = pid;
 455	task->nr = sched->nr_tasks;
 456	strcpy(task->comm, comm);
 457	/*
 458	 * every task starts in sleeping state - this gets ignored
 459	 * if there's no wakeup pointing to this sleep state:
 460	 */
 461	add_sched_event_sleep(sched, task, 0, 0);
 462
 463	sched->pid_to_task[pid] = task;
 464	sched->nr_tasks++;
 465	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
 466	BUG_ON(!sched->tasks);
 467	sched->tasks[task->nr] = task;
 468
 469	if (verbose > 0)
 470		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
 471
 472	return task;
 473}
 474
 475
 476static void print_task_traces(struct perf_sched *sched)
 477{
 478	struct task_desc *task;
 479	unsigned long i;
 480
 481	for (i = 0; i < sched->nr_tasks; i++) {
 482		task = sched->tasks[i];
 483		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
 484			task->nr, task->comm, task->pid, task->nr_events);
 485	}
 486}
 487
 488static void add_cross_task_wakeups(struct perf_sched *sched)
 489{
 490	struct task_desc *task1, *task2;
 491	unsigned long i, j;
 492
 493	for (i = 0; i < sched->nr_tasks; i++) {
 494		task1 = sched->tasks[i];
 495		j = i + 1;
 496		if (j == sched->nr_tasks)
 497			j = 0;
 498		task2 = sched->tasks[j];
 499		add_sched_event_wakeup(sched, task1, 0, task2);
 500	}
 501}
 502
 503static void perf_sched__process_event(struct perf_sched *sched,
 504				      struct sched_atom *atom)
 505{
 506	int ret = 0;
 507
 508	switch (atom->type) {
 509		case SCHED_EVENT_RUN:
 510			burn_nsecs(sched, atom->duration);
 511			break;
 512		case SCHED_EVENT_SLEEP:
 513			if (atom->wait_sem)
 514				ret = sem_wait(atom->wait_sem);
 515			BUG_ON(ret);
 516			break;
 517		case SCHED_EVENT_WAKEUP:
 518			if (atom->wait_sem)
 519				ret = sem_post(atom->wait_sem);
 520			BUG_ON(ret);
 521			break;
 522		case SCHED_EVENT_MIGRATION:
 523			break;
 524		default:
 525			BUG_ON(1);
 526	}
 527}
 528
 529static u64 get_cpu_usage_nsec_parent(void)
 530{
 531	struct rusage ru;
 532	u64 sum;
 533	int err;
 534
 535	err = getrusage(RUSAGE_SELF, &ru);
 536	BUG_ON(err);
 537
 538	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
 539	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
 540
 541	return sum;
 542}
 543
 544static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
 545{
 546	struct perf_event_attr attr;
 547	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
 548	int fd;
 549	struct rlimit limit;
 550	bool need_privilege = false;
 551
 552	memset(&attr, 0, sizeof(attr));
 553
 554	attr.type = PERF_TYPE_SOFTWARE;
 555	attr.config = PERF_COUNT_SW_TASK_CLOCK;
 556
 557force_again:
 558	fd = sys_perf_event_open(&attr, 0, -1, -1,
 559				 perf_event_open_cloexec_flag());
 560
 561	if (fd < 0) {
 562		if (errno == EMFILE) {
 563			if (sched->force) {
 564				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
 565				limit.rlim_cur += sched->nr_tasks - cur_task;
 566				if (limit.rlim_cur > limit.rlim_max) {
 567					limit.rlim_max = limit.rlim_cur;
 568					need_privilege = true;
 569				}
 570				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
 571					if (need_privilege && errno == EPERM)
 572						strcpy(info, "Need privilege\n");
 573				} else
 574					goto force_again;
 575			} else
 576				strcpy(info, "Have a try with -f option\n");
 577		}
 578		pr_err("Error: sys_perf_event_open() syscall returned "
 579		       "with %d (%s)\n%s", fd,
 580		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
 581		exit(EXIT_FAILURE);
 582	}
 583	return fd;
 584}
 585
 586static u64 get_cpu_usage_nsec_self(int fd)
 587{
 588	u64 runtime;
 589	int ret;
 590
 591	ret = read(fd, &runtime, sizeof(runtime));
 592	BUG_ON(ret != sizeof(runtime));
 593
 594	return runtime;
 595}
 596
 597struct sched_thread_parms {
 598	struct task_desc  *task;
 599	struct perf_sched *sched;
 600	int fd;
 601};
 602
 603static void *thread_func(void *ctx)
 604{
 605	struct sched_thread_parms *parms = ctx;
 606	struct task_desc *this_task = parms->task;
 607	struct perf_sched *sched = parms->sched;
 608	u64 cpu_usage_0, cpu_usage_1;
 609	unsigned long i, ret;
 610	char comm2[22];
 611	int fd = parms->fd;
 612
 613	zfree(&parms);
 614
 615	sprintf(comm2, ":%s", this_task->comm);
 616	prctl(PR_SET_NAME, comm2);
 617	if (fd < 0)
 618		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 619
 620	while (!sched->thread_funcs_exit) {
 621		ret = sem_post(&this_task->ready_for_work);
 622		BUG_ON(ret);
 623		mutex_lock(&sched->start_work_mutex);
 624		mutex_unlock(&sched->start_work_mutex);
 625
 626		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
 627
 628		for (i = 0; i < this_task->nr_events; i++) {
 629			this_task->curr_event = i;
 630			perf_sched__process_event(sched, this_task->atoms[i]);
 631		}
 632
 633		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
 634		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
 635		ret = sem_post(&this_task->work_done_sem);
 636		BUG_ON(ret);
 637
 638		mutex_lock(&sched->work_done_wait_mutex);
 639		mutex_unlock(&sched->work_done_wait_mutex);
 640	}
 641	return NULL;
 642}
 643
 644static void create_tasks(struct perf_sched *sched)
 645	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
 646	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
 647{
 648	struct task_desc *task;
 649	pthread_attr_t attr;
 650	unsigned long i;
 651	int err;
 652
 653	err = pthread_attr_init(&attr);
 654	BUG_ON(err);
 655	err = pthread_attr_setstacksize(&attr,
 656			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
 
 
 
 
 657	BUG_ON(err);
 658	mutex_lock(&sched->start_work_mutex);
 659	mutex_lock(&sched->work_done_wait_mutex);
 660	for (i = 0; i < sched->nr_tasks; i++) {
 661		struct sched_thread_parms *parms = malloc(sizeof(*parms));
 662		BUG_ON(parms == NULL);
 663		parms->task = task = sched->tasks[i];
 664		parms->sched = sched;
 665		parms->fd = self_open_counters(sched, i);
 666		sem_init(&task->sleep_sem, 0, 0);
 667		sem_init(&task->ready_for_work, 0, 0);
 668		sem_init(&task->work_done_sem, 0, 0);
 669		task->curr_event = 0;
 670		err = pthread_create(&task->thread, &attr, thread_func, parms);
 671		BUG_ON(err);
 672	}
 673}
 674
 675static void destroy_tasks(struct perf_sched *sched)
 676	UNLOCK_FUNCTION(sched->start_work_mutex)
 677	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
 678{
 679	struct task_desc *task;
 680	unsigned long i;
 681	int err;
 682
 683	mutex_unlock(&sched->start_work_mutex);
 684	mutex_unlock(&sched->work_done_wait_mutex);
 685	/* Get rid of threads so they won't be upset by mutex destrunction */
 686	for (i = 0; i < sched->nr_tasks; i++) {
 687		task = sched->tasks[i];
 688		err = pthread_join(task->thread, NULL);
 689		BUG_ON(err);
 690		sem_destroy(&task->sleep_sem);
 691		sem_destroy(&task->ready_for_work);
 692		sem_destroy(&task->work_done_sem);
 693	}
 694}
 695
 696static void wait_for_tasks(struct perf_sched *sched)
 697	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 698	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 699{
 700	u64 cpu_usage_0, cpu_usage_1;
 701	struct task_desc *task;
 702	unsigned long i, ret;
 703
 704	sched->start_time = get_nsecs();
 705	sched->cpu_usage = 0;
 706	mutex_unlock(&sched->work_done_wait_mutex);
 707
 708	for (i = 0; i < sched->nr_tasks; i++) {
 709		task = sched->tasks[i];
 710		ret = sem_wait(&task->ready_for_work);
 711		BUG_ON(ret);
 712		sem_init(&task->ready_for_work, 0, 0);
 713	}
 714	mutex_lock(&sched->work_done_wait_mutex);
 
 715
 716	cpu_usage_0 = get_cpu_usage_nsec_parent();
 717
 718	mutex_unlock(&sched->start_work_mutex);
 719
 720	for (i = 0; i < sched->nr_tasks; i++) {
 721		task = sched->tasks[i];
 722		ret = sem_wait(&task->work_done_sem);
 723		BUG_ON(ret);
 724		sem_init(&task->work_done_sem, 0, 0);
 725		sched->cpu_usage += task->cpu_usage;
 726		task->cpu_usage = 0;
 727	}
 728
 729	cpu_usage_1 = get_cpu_usage_nsec_parent();
 730	if (!sched->runavg_cpu_usage)
 731		sched->runavg_cpu_usage = sched->cpu_usage;
 732	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
 733
 734	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
 735	if (!sched->runavg_parent_cpu_usage)
 736		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
 737	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
 738					 sched->parent_cpu_usage)/sched->replay_repeat;
 739
 740	mutex_lock(&sched->start_work_mutex);
 
 741
 742	for (i = 0; i < sched->nr_tasks; i++) {
 743		task = sched->tasks[i];
 744		sem_init(&task->sleep_sem, 0, 0);
 745		task->curr_event = 0;
 746	}
 747}
 748
 749static void run_one_test(struct perf_sched *sched)
 750	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
 751	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
 752{
 753	u64 T0, T1, delta, avg_delta, fluct;
 754
 755	T0 = get_nsecs();
 756	wait_for_tasks(sched);
 757	T1 = get_nsecs();
 758
 759	delta = T1 - T0;
 760	sched->sum_runtime += delta;
 761	sched->nr_runs++;
 762
 763	avg_delta = sched->sum_runtime / sched->nr_runs;
 764	if (delta < avg_delta)
 765		fluct = avg_delta - delta;
 766	else
 767		fluct = delta - avg_delta;
 768	sched->sum_fluct += fluct;
 769	if (!sched->run_avg)
 770		sched->run_avg = delta;
 771	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
 772
 773	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
 774
 775	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
 776
 777	printf("cpu: %0.2f / %0.2f",
 778		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
 779
 780#if 0
 781	/*
 782	 * rusage statistics done by the parent, these are less
 783	 * accurate than the sched->sum_exec_runtime based statistics:
 784	 */
 785	printf(" [%0.2f / %0.2f]",
 786		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
 787		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
 788#endif
 789
 790	printf("\n");
 791
 792	if (sched->nr_sleep_corrections)
 793		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
 794	sched->nr_sleep_corrections = 0;
 795}
 796
 797static void test_calibrations(struct perf_sched *sched)
 798{
 799	u64 T0, T1;
 800
 801	T0 = get_nsecs();
 802	burn_nsecs(sched, NSEC_PER_MSEC);
 803	T1 = get_nsecs();
 804
 805	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
 806
 807	T0 = get_nsecs();
 808	sleep_nsecs(NSEC_PER_MSEC);
 809	T1 = get_nsecs();
 810
 811	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
 812}
 813
 814static int
 815replay_wakeup_event(struct perf_sched *sched,
 816		    struct evsel *evsel, struct perf_sample *sample,
 817		    struct machine *machine __maybe_unused)
 818{
 819	const char *comm = evsel__strval(evsel, sample, "comm");
 820	const u32 pid	 = evsel__intval(evsel, sample, "pid");
 821	struct task_desc *waker, *wakee;
 822
 823	if (verbose > 0) {
 824		printf("sched_wakeup event %p\n", evsel);
 825
 826		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
 827	}
 828
 829	waker = register_pid(sched, sample->tid, "<unknown>");
 830	wakee = register_pid(sched, pid, comm);
 831
 832	add_sched_event_wakeup(sched, waker, sample->time, wakee);
 833	return 0;
 834}
 835
 836static int replay_switch_event(struct perf_sched *sched,
 837			       struct evsel *evsel,
 838			       struct perf_sample *sample,
 839			       struct machine *machine __maybe_unused)
 840{
 841	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
 842		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
 843	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
 844		  next_pid = evsel__intval(evsel, sample, "next_pid");
 845	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
 846	struct task_desc *prev, __maybe_unused *next;
 847	u64 timestamp0, timestamp = sample->time;
 848	int cpu = sample->cpu;
 849	s64 delta;
 850
 851	if (verbose > 0)
 852		printf("sched_switch event %p\n", evsel);
 853
 854	if (cpu >= MAX_CPUS || cpu < 0)
 855		return 0;
 856
 857	timestamp0 = sched->cpu_last_switched[cpu];
 858	if (timestamp0)
 859		delta = timestamp - timestamp0;
 860	else
 861		delta = 0;
 862
 863	if (delta < 0) {
 864		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 865		return -1;
 866	}
 867
 868	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
 869		 prev_comm, prev_pid, next_comm, next_pid, delta);
 870
 871	prev = register_pid(sched, prev_pid, prev_comm);
 872	next = register_pid(sched, next_pid, next_comm);
 873
 874	sched->cpu_last_switched[cpu] = timestamp;
 875
 876	add_sched_event_run(sched, prev, timestamp, delta);
 877	add_sched_event_sleep(sched, prev, timestamp, prev_state);
 878
 879	return 0;
 880}
 881
 882static int replay_fork_event(struct perf_sched *sched,
 883			     union perf_event *event,
 884			     struct machine *machine)
 885{
 886	struct thread *child, *parent;
 887
 888	child = machine__findnew_thread(machine, event->fork.pid,
 889					event->fork.tid);
 890	parent = machine__findnew_thread(machine, event->fork.ppid,
 891					 event->fork.ptid);
 892
 893	if (child == NULL || parent == NULL) {
 894		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
 895				 child, parent);
 896		goto out_put;
 897	}
 898
 899	if (verbose > 0) {
 900		printf("fork event\n");
 901		printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
 902		printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
 903	}
 904
 905	register_pid(sched, thread__tid(parent), thread__comm_str(parent));
 906	register_pid(sched, thread__tid(child), thread__comm_str(child));
 907out_put:
 908	thread__put(child);
 909	thread__put(parent);
 910	return 0;
 911}
 912
 913struct sort_dimension {
 914	const char		*name;
 915	sort_fn_t		cmp;
 916	struct list_head	list;
 917};
 918
 919/*
 920 * handle runtime stats saved per thread
 921 */
 922static struct thread_runtime *thread__init_runtime(struct thread *thread)
 923{
 924	struct thread_runtime *r;
 925
 926	r = zalloc(sizeof(struct thread_runtime));
 927	if (!r)
 928		return NULL;
 929
 930	init_stats(&r->run_stats);
 931	thread__set_priv(thread, r);
 932
 933	return r;
 934}
 935
 936static struct thread_runtime *thread__get_runtime(struct thread *thread)
 937{
 938	struct thread_runtime *tr;
 939
 940	tr = thread__priv(thread);
 941	if (tr == NULL) {
 942		tr = thread__init_runtime(thread);
 943		if (tr == NULL)
 944			pr_debug("Failed to malloc memory for runtime data.\n");
 945	}
 946
 947	return tr;
 948}
 949
 950static int
 951thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 952{
 953	struct sort_dimension *sort;
 954	int ret = 0;
 955
 956	BUG_ON(list_empty(list));
 957
 958	list_for_each_entry(sort, list, list) {
 959		ret = sort->cmp(l, r);
 960		if (ret)
 961			return ret;
 962	}
 963
 964	return ret;
 965}
 966
 967static struct work_atoms *
 968thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
 969			 struct list_head *sort_list)
 970{
 971	struct rb_node *node = root->rb_root.rb_node;
 972	struct work_atoms key = { .thread = thread };
 973
 974	while (node) {
 975		struct work_atoms *atoms;
 976		int cmp;
 977
 978		atoms = container_of(node, struct work_atoms, node);
 979
 980		cmp = thread_lat_cmp(sort_list, &key, atoms);
 981		if (cmp > 0)
 982			node = node->rb_left;
 983		else if (cmp < 0)
 984			node = node->rb_right;
 985		else {
 986			BUG_ON(thread != atoms->thread);
 987			return atoms;
 988		}
 989	}
 990	return NULL;
 991}
 992
 993static void
 994__thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
 995			 struct list_head *sort_list)
 996{
 997	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 998	bool leftmost = true;
 999
1000	while (*new) {
1001		struct work_atoms *this;
1002		int cmp;
1003
1004		this = container_of(*new, struct work_atoms, node);
1005		parent = *new;
1006
1007		cmp = thread_lat_cmp(sort_list, data, this);
1008
1009		if (cmp > 0)
1010			new = &((*new)->rb_left);
1011		else {
1012			new = &((*new)->rb_right);
1013			leftmost = false;
1014		}
1015	}
1016
1017	rb_link_node(&data->node, parent, new);
1018	rb_insert_color_cached(&data->node, root, leftmost);
1019}
1020
1021static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1022{
1023	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1024	if (!atoms) {
1025		pr_err("No memory at %s\n", __func__);
1026		return -1;
1027	}
1028
1029	atoms->thread = thread__get(thread);
1030	INIT_LIST_HEAD(&atoms->work_list);
1031	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1032	return 0;
1033}
1034
 
 
 
 
 
 
 
1035static int
1036add_sched_out_event(struct work_atoms *atoms,
1037		    char run_state,
1038		    u64 timestamp)
1039{
1040	struct work_atom *atom = zalloc(sizeof(*atom));
1041	if (!atom) {
1042		pr_err("Non memory at %s", __func__);
1043		return -1;
1044	}
1045
1046	atom->sched_out_time = timestamp;
1047
1048	if (run_state == 'R') {
1049		atom->state = THREAD_WAIT_CPU;
1050		atom->wake_up_time = atom->sched_out_time;
1051	}
1052
1053	list_add_tail(&atom->list, &atoms->work_list);
1054	return 0;
1055}
1056
1057static void
1058add_runtime_event(struct work_atoms *atoms, u64 delta,
1059		  u64 timestamp __maybe_unused)
1060{
1061	struct work_atom *atom;
1062
1063	BUG_ON(list_empty(&atoms->work_list));
1064
1065	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1066
1067	atom->runtime += delta;
1068	atoms->total_runtime += delta;
1069}
1070
1071static void
1072add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1073{
1074	struct work_atom *atom;
1075	u64 delta;
1076
1077	if (list_empty(&atoms->work_list))
1078		return;
1079
1080	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1081
1082	if (atom->state != THREAD_WAIT_CPU)
1083		return;
1084
1085	if (timestamp < atom->wake_up_time) {
1086		atom->state = THREAD_IGNORE;
1087		return;
1088	}
1089
1090	atom->state = THREAD_SCHED_IN;
1091	atom->sched_in_time = timestamp;
1092
1093	delta = atom->sched_in_time - atom->wake_up_time;
1094	atoms->total_lat += delta;
1095	if (delta > atoms->max_lat) {
1096		atoms->max_lat = delta;
1097		atoms->max_lat_start = atom->wake_up_time;
1098		atoms->max_lat_end = timestamp;
1099	}
1100	atoms->nb_atoms++;
1101}
1102
1103static int latency_switch_event(struct perf_sched *sched,
1104				struct evsel *evsel,
1105				struct perf_sample *sample,
1106				struct machine *machine)
1107{
1108	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1109		  next_pid = evsel__intval(evsel, sample, "next_pid");
1110	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1111	struct work_atoms *out_events, *in_events;
1112	struct thread *sched_out, *sched_in;
1113	u64 timestamp0, timestamp = sample->time;
1114	int cpu = sample->cpu, err = -1;
1115	s64 delta;
1116
1117	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1118
1119	timestamp0 = sched->cpu_last_switched[cpu];
1120	sched->cpu_last_switched[cpu] = timestamp;
1121	if (timestamp0)
1122		delta = timestamp - timestamp0;
1123	else
1124		delta = 0;
1125
1126	if (delta < 0) {
1127		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1128		return -1;
1129	}
1130
1131	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1132	sched_in = machine__findnew_thread(machine, -1, next_pid);
1133	if (sched_out == NULL || sched_in == NULL)
1134		goto out_put;
1135
1136	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1137	if (!out_events) {
1138		if (thread_atoms_insert(sched, sched_out))
1139			goto out_put;
1140		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1141		if (!out_events) {
1142			pr_err("out-event: Internal tree error");
1143			goto out_put;
1144		}
1145	}
1146	if (add_sched_out_event(out_events, prev_state, timestamp))
1147		return -1;
1148
1149	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1150	if (!in_events) {
1151		if (thread_atoms_insert(sched, sched_in))
1152			goto out_put;
1153		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1154		if (!in_events) {
1155			pr_err("in-event: Internal tree error");
1156			goto out_put;
1157		}
1158		/*
1159		 * Take came in we have not heard about yet,
1160		 * add in an initial atom in runnable state:
1161		 */
1162		if (add_sched_out_event(in_events, 'R', timestamp))
1163			goto out_put;
1164	}
1165	add_sched_in_event(in_events, timestamp);
1166	err = 0;
1167out_put:
1168	thread__put(sched_out);
1169	thread__put(sched_in);
1170	return err;
1171}
1172
1173static int latency_runtime_event(struct perf_sched *sched,
1174				 struct evsel *evsel,
1175				 struct perf_sample *sample,
1176				 struct machine *machine)
1177{
1178	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1179	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1180	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1181	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1182	u64 timestamp = sample->time;
1183	int cpu = sample->cpu, err = -1;
1184
1185	if (thread == NULL)
1186		return -1;
1187
1188	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1189	if (!atoms) {
1190		if (thread_atoms_insert(sched, thread))
1191			goto out_put;
1192		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1193		if (!atoms) {
1194			pr_err("in-event: Internal tree error");
1195			goto out_put;
1196		}
1197		if (add_sched_out_event(atoms, 'R', timestamp))
1198			goto out_put;
1199	}
1200
1201	add_runtime_event(atoms, runtime, timestamp);
1202	err = 0;
1203out_put:
1204	thread__put(thread);
1205	return err;
1206}
1207
1208static int latency_wakeup_event(struct perf_sched *sched,
1209				struct evsel *evsel,
1210				struct perf_sample *sample,
1211				struct machine *machine)
1212{
1213	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1214	struct work_atoms *atoms;
1215	struct work_atom *atom;
1216	struct thread *wakee;
1217	u64 timestamp = sample->time;
1218	int err = -1;
1219
1220	wakee = machine__findnew_thread(machine, -1, pid);
1221	if (wakee == NULL)
1222		return -1;
1223	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1224	if (!atoms) {
1225		if (thread_atoms_insert(sched, wakee))
1226			goto out_put;
1227		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1228		if (!atoms) {
1229			pr_err("wakeup-event: Internal tree error");
1230			goto out_put;
1231		}
1232		if (add_sched_out_event(atoms, 'S', timestamp))
1233			goto out_put;
1234	}
1235
1236	BUG_ON(list_empty(&atoms->work_list));
1237
1238	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1239
1240	/*
1241	 * As we do not guarantee the wakeup event happens when
1242	 * task is out of run queue, also may happen when task is
1243	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1244	 * then we should not set the ->wake_up_time when wake up a
1245	 * task which is on run queue.
1246	 *
1247	 * You WILL be missing events if you've recorded only
1248	 * one CPU, or are only looking at only one, so don't
1249	 * skip in this case.
1250	 */
1251	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1252		goto out_ok;
1253
1254	sched->nr_timestamps++;
1255	if (atom->sched_out_time > timestamp) {
1256		sched->nr_unordered_timestamps++;
1257		goto out_ok;
1258	}
1259
1260	atom->state = THREAD_WAIT_CPU;
1261	atom->wake_up_time = timestamp;
1262out_ok:
1263	err = 0;
1264out_put:
1265	thread__put(wakee);
1266	return err;
1267}
1268
1269static int latency_migrate_task_event(struct perf_sched *sched,
1270				      struct evsel *evsel,
1271				      struct perf_sample *sample,
1272				      struct machine *machine)
1273{
1274	const u32 pid = evsel__intval(evsel, sample, "pid");
1275	u64 timestamp = sample->time;
1276	struct work_atoms *atoms;
1277	struct work_atom *atom;
1278	struct thread *migrant;
1279	int err = -1;
1280
1281	/*
1282	 * Only need to worry about migration when profiling one CPU.
1283	 */
1284	if (sched->profile_cpu == -1)
1285		return 0;
1286
1287	migrant = machine__findnew_thread(machine, -1, pid);
1288	if (migrant == NULL)
1289		return -1;
1290	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1291	if (!atoms) {
1292		if (thread_atoms_insert(sched, migrant))
1293			goto out_put;
1294		register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1295		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1296		if (!atoms) {
1297			pr_err("migration-event: Internal tree error");
1298			goto out_put;
1299		}
1300		if (add_sched_out_event(atoms, 'R', timestamp))
1301			goto out_put;
1302	}
1303
1304	BUG_ON(list_empty(&atoms->work_list));
1305
1306	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1307	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1308
1309	sched->nr_timestamps++;
1310
1311	if (atom->sched_out_time > timestamp)
1312		sched->nr_unordered_timestamps++;
1313	err = 0;
1314out_put:
1315	thread__put(migrant);
1316	return err;
1317}
1318
1319static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1320{
1321	int i;
1322	int ret;
1323	u64 avg;
1324	char max_lat_start[32], max_lat_end[32];
1325
1326	if (!work_list->nb_atoms)
1327		return;
1328	/*
1329	 * Ignore idle threads:
1330	 */
1331	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1332		return;
1333
1334	sched->all_runtime += work_list->total_runtime;
1335	sched->all_count   += work_list->nb_atoms;
1336
1337	if (work_list->num_merged > 1) {
1338		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1339			     work_list->num_merged);
1340	} else {
1341		ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1342			     thread__tid(work_list->thread));
1343	}
1344
1345	for (i = 0; i < 24 - ret; i++)
1346		printf(" ");
1347
1348	avg = work_list->total_lat / work_list->nb_atoms;
1349	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1351
1352	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1354		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355		 (double)work_list->max_lat / NSEC_PER_MSEC,
1356		 max_lat_start, max_lat_end);
1357}
1358
1359static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1360{
1361	pid_t l_tid, r_tid;
1362
1363	if (RC_CHK_EQUAL(l->thread, r->thread))
1364		return 0;
1365	l_tid = thread__tid(l->thread);
1366	r_tid = thread__tid(r->thread);
1367	if (l_tid < r_tid)
1368		return -1;
1369	if (l_tid > r_tid)
1370		return 1;
1371	return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1372}
1373
1374static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1375{
1376	u64 avgl, avgr;
1377
1378	if (!l->nb_atoms)
1379		return -1;
1380
1381	if (!r->nb_atoms)
1382		return 1;
1383
1384	avgl = l->total_lat / l->nb_atoms;
1385	avgr = r->total_lat / r->nb_atoms;
1386
1387	if (avgl < avgr)
1388		return -1;
1389	if (avgl > avgr)
1390		return 1;
1391
1392	return 0;
1393}
1394
1395static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1396{
1397	if (l->max_lat < r->max_lat)
1398		return -1;
1399	if (l->max_lat > r->max_lat)
1400		return 1;
1401
1402	return 0;
1403}
1404
1405static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1406{
1407	if (l->nb_atoms < r->nb_atoms)
1408		return -1;
1409	if (l->nb_atoms > r->nb_atoms)
1410		return 1;
1411
1412	return 0;
1413}
1414
1415static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1416{
1417	if (l->total_runtime < r->total_runtime)
1418		return -1;
1419	if (l->total_runtime > r->total_runtime)
1420		return 1;
1421
1422	return 0;
1423}
1424
1425static int sort_dimension__add(const char *tok, struct list_head *list)
1426{
1427	size_t i;
1428	static struct sort_dimension avg_sort_dimension = {
1429		.name = "avg",
1430		.cmp  = avg_cmp,
1431	};
1432	static struct sort_dimension max_sort_dimension = {
1433		.name = "max",
1434		.cmp  = max_cmp,
1435	};
1436	static struct sort_dimension pid_sort_dimension = {
1437		.name = "pid",
1438		.cmp  = pid_cmp,
1439	};
1440	static struct sort_dimension runtime_sort_dimension = {
1441		.name = "runtime",
1442		.cmp  = runtime_cmp,
1443	};
1444	static struct sort_dimension switch_sort_dimension = {
1445		.name = "switch",
1446		.cmp  = switch_cmp,
1447	};
1448	struct sort_dimension *available_sorts[] = {
1449		&pid_sort_dimension,
1450		&avg_sort_dimension,
1451		&max_sort_dimension,
1452		&switch_sort_dimension,
1453		&runtime_sort_dimension,
1454	};
1455
1456	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1457		if (!strcmp(available_sorts[i]->name, tok)) {
1458			list_add_tail(&available_sorts[i]->list, list);
1459
1460			return 0;
1461		}
1462	}
1463
1464	return -1;
1465}
1466
1467static void perf_sched__sort_lat(struct perf_sched *sched)
1468{
1469	struct rb_node *node;
1470	struct rb_root_cached *root = &sched->atom_root;
1471again:
1472	for (;;) {
1473		struct work_atoms *data;
1474		node = rb_first_cached(root);
1475		if (!node)
1476			break;
1477
1478		rb_erase_cached(node, root);
1479		data = rb_entry(node, struct work_atoms, node);
1480		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1481	}
1482	if (root == &sched->atom_root) {
1483		root = &sched->merged_atom_root;
1484		goto again;
1485	}
1486}
1487
1488static int process_sched_wakeup_event(struct perf_tool *tool,
1489				      struct evsel *evsel,
1490				      struct perf_sample *sample,
1491				      struct machine *machine)
1492{
1493	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1494
1495	if (sched->tp_handler->wakeup_event)
1496		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1497
1498	return 0;
1499}
1500
1501static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1502				      struct evsel *evsel __maybe_unused,
1503				      struct perf_sample *sample __maybe_unused,
1504				      struct machine *machine __maybe_unused)
1505{
1506	return 0;
1507}
1508
1509union map_priv {
1510	void	*ptr;
1511	bool	 color;
1512};
1513
1514static bool thread__has_color(struct thread *thread)
1515{
1516	union map_priv priv = {
1517		.ptr = thread__priv(thread),
1518	};
1519
1520	return priv.color;
1521}
1522
1523static struct thread*
1524map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1525{
1526	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1527	union map_priv priv = {
1528		.color = false,
1529	};
1530
1531	if (!sched->map.color_pids || !thread || thread__priv(thread))
1532		return thread;
1533
1534	if (thread_map__has(sched->map.color_pids, tid))
1535		priv.color = true;
1536
1537	thread__set_priv(thread, priv.ptr);
1538	return thread;
1539}
1540
1541static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1542			    struct perf_sample *sample, struct machine *machine)
1543{
1544	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1545	struct thread *sched_in;
1546	struct thread_runtime *tr;
1547	int new_shortname;
1548	u64 timestamp0, timestamp = sample->time;
1549	s64 delta;
1550	int i;
1551	struct perf_cpu this_cpu = {
1552		.cpu = sample->cpu,
1553	};
1554	int cpus_nr;
1555	bool new_cpu = false;
1556	const char *color = PERF_COLOR_NORMAL;
1557	char stimestamp[32];
1558
1559	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1560
1561	if (this_cpu.cpu > sched->max_cpu.cpu)
1562		sched->max_cpu = this_cpu;
1563
1564	if (sched->map.comp) {
1565		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1566		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1567			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1568			new_cpu = true;
1569		}
1570	} else
1571		cpus_nr = sched->max_cpu.cpu;
1572
1573	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1574	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1575	if (timestamp0)
1576		delta = timestamp - timestamp0;
1577	else
1578		delta = 0;
1579
1580	if (delta < 0) {
1581		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1582		return -1;
1583	}
1584
1585	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1586	if (sched_in == NULL)
1587		return -1;
1588
1589	tr = thread__get_runtime(sched_in);
1590	if (tr == NULL) {
1591		thread__put(sched_in);
1592		return -1;
1593	}
1594
1595	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1596
1597	printf("  ");
1598
1599	new_shortname = 0;
1600	if (!tr->shortname[0]) {
1601		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1602			/*
1603			 * Don't allocate a letter-number for swapper:0
1604			 * as a shortname. Instead, we use '.' for it.
1605			 */
1606			tr->shortname[0] = '.';
1607			tr->shortname[1] = ' ';
1608		} else {
1609			tr->shortname[0] = sched->next_shortname1;
1610			tr->shortname[1] = sched->next_shortname2;
1611
1612			if (sched->next_shortname1 < 'Z') {
1613				sched->next_shortname1++;
1614			} else {
1615				sched->next_shortname1 = 'A';
1616				if (sched->next_shortname2 < '9')
1617					sched->next_shortname2++;
1618				else
1619					sched->next_shortname2 = '0';
1620			}
1621		}
1622		new_shortname = 1;
1623	}
1624
1625	for (i = 0; i < cpus_nr; i++) {
1626		struct perf_cpu cpu = {
1627			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1628		};
1629		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1630		struct thread_runtime *curr_tr;
1631		const char *pid_color = color;
1632		const char *cpu_color = color;
1633
1634		if (curr_thread && thread__has_color(curr_thread))
1635			pid_color = COLOR_PIDS;
1636
1637		if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1638			continue;
1639
1640		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1641			cpu_color = COLOR_CPUS;
1642
1643		if (cpu.cpu != this_cpu.cpu)
1644			color_fprintf(stdout, color, " ");
1645		else
1646			color_fprintf(stdout, cpu_color, "*");
1647
1648		if (sched->curr_thread[cpu.cpu]) {
1649			curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1650			if (curr_tr == NULL) {
1651				thread__put(sched_in);
1652				return -1;
1653			}
1654			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1655		} else
1656			color_fprintf(stdout, color, "   ");
1657	}
1658
1659	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1660		goto out;
1661
1662	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1663	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1664	if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1665		const char *pid_color = color;
1666
1667		if (thread__has_color(sched_in))
1668			pid_color = COLOR_PIDS;
1669
1670		color_fprintf(stdout, pid_color, "%s => %s:%d",
1671			tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1672		tr->comm_changed = false;
1673	}
1674
1675	if (sched->map.comp && new_cpu)
1676		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1677
1678out:
1679	color_fprintf(stdout, color, "\n");
1680
1681	thread__put(sched_in);
1682
1683	return 0;
1684}
1685
1686static int process_sched_switch_event(struct perf_tool *tool,
1687				      struct evsel *evsel,
1688				      struct perf_sample *sample,
1689				      struct machine *machine)
1690{
1691	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1692	int this_cpu = sample->cpu, err = 0;
1693	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1694	    next_pid = evsel__intval(evsel, sample, "next_pid");
1695
1696	if (sched->curr_pid[this_cpu] != (u32)-1) {
1697		/*
1698		 * Are we trying to switch away a PID that is
1699		 * not current?
1700		 */
1701		if (sched->curr_pid[this_cpu] != prev_pid)
1702			sched->nr_context_switch_bugs++;
1703	}
1704
1705	if (sched->tp_handler->switch_event)
1706		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1707
1708	sched->curr_pid[this_cpu] = next_pid;
1709	return err;
1710}
1711
1712static int process_sched_runtime_event(struct perf_tool *tool,
1713				       struct evsel *evsel,
1714				       struct perf_sample *sample,
1715				       struct machine *machine)
1716{
1717	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1718
1719	if (sched->tp_handler->runtime_event)
1720		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1721
1722	return 0;
1723}
1724
1725static int perf_sched__process_fork_event(struct perf_tool *tool,
1726					  union perf_event *event,
1727					  struct perf_sample *sample,
1728					  struct machine *machine)
1729{
1730	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731
1732	/* run the fork event through the perf machinery */
1733	perf_event__process_fork(tool, event, sample, machine);
1734
1735	/* and then run additional processing needed for this command */
1736	if (sched->tp_handler->fork_event)
1737		return sched->tp_handler->fork_event(sched, event, machine);
1738
1739	return 0;
1740}
1741
1742static int process_sched_migrate_task_event(struct perf_tool *tool,
1743					    struct evsel *evsel,
1744					    struct perf_sample *sample,
1745					    struct machine *machine)
1746{
1747	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1748
1749	if (sched->tp_handler->migrate_task_event)
1750		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1751
1752	return 0;
1753}
1754
1755typedef int (*tracepoint_handler)(struct perf_tool *tool,
1756				  struct evsel *evsel,
1757				  struct perf_sample *sample,
1758				  struct machine *machine);
1759
1760static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1761						 union perf_event *event __maybe_unused,
1762						 struct perf_sample *sample,
1763						 struct evsel *evsel,
1764						 struct machine *machine)
1765{
1766	int err = 0;
1767
1768	if (evsel->handler != NULL) {
1769		tracepoint_handler f = evsel->handler;
1770		err = f(tool, evsel, sample, machine);
1771	}
1772
1773	return err;
1774}
1775
1776static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1777				    union perf_event *event,
1778				    struct perf_sample *sample,
1779				    struct machine *machine)
1780{
1781	struct thread *thread;
1782	struct thread_runtime *tr;
1783	int err;
1784
1785	err = perf_event__process_comm(tool, event, sample, machine);
1786	if (err)
1787		return err;
1788
1789	thread = machine__find_thread(machine, sample->pid, sample->tid);
1790	if (!thread) {
1791		pr_err("Internal error: can't find thread\n");
1792		return -1;
1793	}
1794
1795	tr = thread__get_runtime(thread);
1796	if (tr == NULL) {
1797		thread__put(thread);
1798		return -1;
1799	}
1800
1801	tr->comm_changed = true;
1802	thread__put(thread);
1803
1804	return 0;
1805}
1806
1807static int perf_sched__read_events(struct perf_sched *sched)
1808{
1809	struct evsel_str_handler handlers[] = {
1810		{ "sched:sched_switch",	      process_sched_switch_event, },
1811		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1812		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1813		{ "sched:sched_waking",	      process_sched_wakeup_event, },
1814		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1815		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1816	};
1817	struct perf_session *session;
1818	struct perf_data data = {
1819		.path  = input_name,
1820		.mode  = PERF_DATA_MODE_READ,
1821		.force = sched->force,
 
 
1822	};
1823	int rc = -1;
1824
1825	session = perf_session__new(&data, &sched->tool);
1826	if (IS_ERR(session)) {
1827		pr_debug("Error creating perf session");
1828		return PTR_ERR(session);
1829	}
1830
1831	symbol__init(&session->header.env);
1832
1833	/* prefer sched_waking if it is captured */
1834	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1835		handlers[2].handler = process_sched_wakeup_ignore;
1836
1837	if (perf_session__set_tracepoints_handlers(session, handlers))
1838		goto out_delete;
1839
1840	if (perf_session__has_traces(session, "record -R")) {
1841		int err = perf_session__process_events(session);
1842		if (err) {
1843			pr_err("Failed to process events, error %d", err);
1844			goto out_delete;
1845		}
1846
1847		sched->nr_events      = session->evlist->stats.nr_events[0];
1848		sched->nr_lost_events = session->evlist->stats.total_lost;
1849		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1850	}
1851
1852	rc = 0;
1853out_delete:
1854	perf_session__delete(session);
1855	return rc;
1856}
1857
1858/*
1859 * scheduling times are printed as msec.usec
1860 */
1861static inline void print_sched_time(unsigned long long nsecs, int width)
1862{
1863	unsigned long msecs;
1864	unsigned long usecs;
1865
1866	msecs  = nsecs / NSEC_PER_MSEC;
1867	nsecs -= msecs * NSEC_PER_MSEC;
1868	usecs  = nsecs / NSEC_PER_USEC;
1869	printf("%*lu.%03lu ", width, msecs, usecs);
1870}
1871
1872/*
1873 * returns runtime data for event, allocating memory for it the
1874 * first time it is used.
1875 */
1876static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1877{
1878	struct evsel_runtime *r = evsel->priv;
1879
1880	if (r == NULL) {
1881		r = zalloc(sizeof(struct evsel_runtime));
1882		evsel->priv = r;
1883	}
1884
1885	return r;
1886}
1887
1888/*
1889 * save last time event was seen per cpu
1890 */
1891static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
 
1892{
1893	struct evsel_runtime *r = evsel__get_runtime(evsel);
1894
1895	if (r == NULL)
1896		return;
1897
1898	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1899		int i, n = __roundup_pow_of_two(cpu+1);
1900		void *p = r->last_time;
1901
1902		p = realloc(r->last_time, n * sizeof(u64));
1903		if (!p)
1904			return;
1905
1906		r->last_time = p;
1907		for (i = r->ncpu; i < n; ++i)
1908			r->last_time[i] = (u64) 0;
1909
1910		r->ncpu = n;
1911	}
1912
1913	r->last_time[cpu] = timestamp;
1914}
1915
1916/* returns last time this event was seen on the given cpu */
1917static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1918{
1919	struct evsel_runtime *r = evsel__get_runtime(evsel);
1920
1921	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1922		return 0;
1923
1924	return r->last_time[cpu];
1925}
1926
1927static int comm_width = 30;
1928
1929static char *timehist_get_commstr(struct thread *thread)
1930{
1931	static char str[32];
1932	const char *comm = thread__comm_str(thread);
1933	pid_t tid = thread__tid(thread);
1934	pid_t pid = thread__pid(thread);
1935	int n;
1936
1937	if (pid == 0)
1938		n = scnprintf(str, sizeof(str), "%s", comm);
1939
1940	else if (tid != pid)
1941		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1942
1943	else
1944		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1945
1946	if (n > comm_width)
1947		comm_width = n;
1948
1949	return str;
1950}
1951
1952static void timehist_header(struct perf_sched *sched)
1953{
1954	u32 ncpus = sched->max_cpu.cpu + 1;
1955	u32 i, j;
1956
1957	printf("%15s %6s ", "time", "cpu");
1958
1959	if (sched->show_cpu_visual) {
1960		printf(" ");
1961		for (i = 0, j = 0; i < ncpus; ++i) {
1962			printf("%x", j++);
1963			if (j > 15)
1964				j = 0;
1965		}
1966		printf(" ");
1967	}
1968
1969	printf(" %-*s  %9s  %9s  %9s", comm_width,
1970		"task name", "wait time", "sch delay", "run time");
1971
1972	if (sched->show_state)
1973		printf("  %s", "state");
1974
1975	printf("\n");
1976
1977	/*
1978	 * units row
1979	 */
1980	printf("%15s %-6s ", "", "");
1981
1982	if (sched->show_cpu_visual)
1983		printf(" %*s ", ncpus, "");
1984
1985	printf(" %-*s  %9s  %9s  %9s", comm_width,
1986	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1987
1988	if (sched->show_state)
1989		printf("  %5s", "");
1990
1991	printf("\n");
1992
1993	/*
1994	 * separator
1995	 */
1996	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1997
1998	if (sched->show_cpu_visual)
1999		printf(" %.*s ", ncpus, graph_dotted_line);
2000
2001	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2002		graph_dotted_line, graph_dotted_line, graph_dotted_line,
2003		graph_dotted_line);
2004
2005	if (sched->show_state)
2006		printf("  %.5s", graph_dotted_line);
2007
2008	printf("\n");
2009}
2010
 
 
 
 
 
 
 
 
 
 
 
 
2011static void timehist_print_sample(struct perf_sched *sched,
2012				  struct evsel *evsel,
2013				  struct perf_sample *sample,
2014				  struct addr_location *al,
2015				  struct thread *thread,
2016				  u64 t, const char state)
2017{
2018	struct thread_runtime *tr = thread__priv(thread);
2019	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2020	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2021	u32 max_cpus = sched->max_cpu.cpu + 1;
2022	char tstr[64];
2023	char nstr[30];
2024	u64 wait_time;
2025
2026	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2027		return;
2028
2029	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2030	printf("%15s [%04d] ", tstr, sample->cpu);
2031
2032	if (sched->show_cpu_visual) {
2033		u32 i;
2034		char c;
2035
2036		printf(" ");
2037		for (i = 0; i < max_cpus; ++i) {
2038			/* flag idle times with 'i'; others are sched events */
2039			if (i == sample->cpu)
2040				c = (thread__tid(thread) == 0) ? 'i' : 's';
2041			else
2042				c = ' ';
2043			printf("%c", c);
2044		}
2045		printf(" ");
2046	}
2047
2048	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2049
2050	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2051	print_sched_time(wait_time, 6);
2052
2053	print_sched_time(tr->dt_delay, 6);
2054	print_sched_time(tr->dt_run, 6);
2055
2056	if (sched->show_state)
2057		printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2058
2059	if (sched->show_next) {
2060		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2061		printf(" %-*s", comm_width, nstr);
2062	}
2063
2064	if (sched->show_wakeups && !sched->show_next)
2065		printf("  %-*s", comm_width, "");
2066
2067	if (thread__tid(thread) == 0)
2068		goto out;
2069
2070	if (sched->show_callchain)
2071		printf("  ");
2072
2073	sample__fprintf_sym(sample, al, 0,
2074			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2075			    EVSEL__PRINT_CALLCHAIN_ARROW |
2076			    EVSEL__PRINT_SKIP_IGNORED,
2077			    get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2078
2079out:
2080	printf("\n");
2081}
2082
2083/*
2084 * Explanation of delta-time stats:
2085 *
2086 *            t = time of current schedule out event
2087 *        tprev = time of previous sched out event
2088 *                also time of schedule-in event for current task
2089 *    last_time = time of last sched change event for current task
2090 *                (i.e, time process was last scheduled out)
2091 * ready_to_run = time of wakeup for current task
2092 *
2093 * -----|------------|------------|------------|------
2094 *    last         ready        tprev          t
2095 *    time         to run
2096 *
2097 *      |-------- dt_wait --------|
2098 *                   |- dt_delay -|-- dt_run --|
2099 *
2100 *   dt_run = run time of current task
2101 *  dt_wait = time between last schedule out event for task and tprev
2102 *            represents time spent off the cpu
2103 * dt_delay = time between wakeup and schedule-in of task
2104 */
2105
2106static void timehist_update_runtime_stats(struct thread_runtime *r,
2107					 u64 t, u64 tprev)
2108{
2109	r->dt_delay   = 0;
2110	r->dt_sleep   = 0;
2111	r->dt_iowait  = 0;
2112	r->dt_preempt = 0;
2113	r->dt_run     = 0;
2114
2115	if (tprev) {
2116		r->dt_run = t - tprev;
2117		if (r->ready_to_run) {
2118			if (r->ready_to_run > tprev)
2119				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2120			else
2121				r->dt_delay = tprev - r->ready_to_run;
2122		}
2123
2124		if (r->last_time > tprev)
2125			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2126		else if (r->last_time) {
2127			u64 dt_wait = tprev - r->last_time;
2128
2129			if (r->last_state == 'R')
2130				r->dt_preempt = dt_wait;
2131			else if (r->last_state == 'D')
2132				r->dt_iowait = dt_wait;
2133			else
2134				r->dt_sleep = dt_wait;
2135		}
2136	}
2137
2138	update_stats(&r->run_stats, r->dt_run);
2139
2140	r->total_run_time     += r->dt_run;
2141	r->total_delay_time   += r->dt_delay;
2142	r->total_sleep_time   += r->dt_sleep;
2143	r->total_iowait_time  += r->dt_iowait;
2144	r->total_preempt_time += r->dt_preempt;
2145}
2146
2147static bool is_idle_sample(struct perf_sample *sample,
2148			   struct evsel *evsel)
2149{
2150	/* pid 0 == swapper == idle task */
2151	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2152		return evsel__intval(evsel, sample, "prev_pid") == 0;
2153
2154	return sample->pid == 0;
2155}
2156
2157static void save_task_callchain(struct perf_sched *sched,
2158				struct perf_sample *sample,
2159				struct evsel *evsel,
2160				struct machine *machine)
2161{
2162	struct callchain_cursor *cursor;
2163	struct thread *thread;
2164
2165	/* want main thread for process - has maps */
2166	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2167	if (thread == NULL) {
2168		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2169		return;
2170	}
2171
2172	if (!sched->show_callchain || sample->callchain == NULL)
2173		return;
2174
2175	cursor = get_tls_callchain_cursor();
2176
2177	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2178				      NULL, NULL, sched->max_stack + 2) != 0) {
2179		if (verbose > 0)
2180			pr_err("Failed to resolve callchain. Skipping\n");
2181
2182		return;
2183	}
2184
2185	callchain_cursor_commit(cursor);
2186
2187	while (true) {
2188		struct callchain_cursor_node *node;
2189		struct symbol *sym;
2190
2191		node = callchain_cursor_current(cursor);
2192		if (node == NULL)
2193			break;
2194
2195		sym = node->ms.sym;
2196		if (sym) {
2197			if (!strcmp(sym->name, "schedule") ||
2198			    !strcmp(sym->name, "__schedule") ||
2199			    !strcmp(sym->name, "preempt_schedule"))
2200				sym->ignore = 1;
2201		}
2202
2203		callchain_cursor_advance(cursor);
2204	}
2205}
2206
2207static int init_idle_thread(struct thread *thread)
2208{
2209	struct idle_thread_runtime *itr;
2210
2211	thread__set_comm(thread, idle_comm, 0);
2212
2213	itr = zalloc(sizeof(*itr));
2214	if (itr == NULL)
2215		return -ENOMEM;
2216
2217	init_stats(&itr->tr.run_stats);
2218	callchain_init(&itr->callchain);
2219	callchain_cursor_reset(&itr->cursor);
2220	thread__set_priv(thread, itr);
2221
2222	return 0;
2223}
2224
2225/*
2226 * Track idle stats per cpu by maintaining a local thread
2227 * struct for the idle task on each cpu.
2228 */
2229static int init_idle_threads(int ncpu)
2230{
2231	int i, ret;
2232
2233	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2234	if (!idle_threads)
2235		return -ENOMEM;
2236
2237	idle_max_cpu = ncpu;
2238
2239	/* allocate the actual thread struct if needed */
2240	for (i = 0; i < ncpu; ++i) {
2241		idle_threads[i] = thread__new(0, 0);
2242		if (idle_threads[i] == NULL)
2243			return -ENOMEM;
2244
2245		ret = init_idle_thread(idle_threads[i]);
2246		if (ret < 0)
2247			return ret;
2248	}
2249
2250	return 0;
2251}
2252
2253static void free_idle_threads(void)
2254{
2255	int i;
2256
2257	if (idle_threads == NULL)
2258		return;
2259
2260	for (i = 0; i < idle_max_cpu; ++i) {
2261		if ((idle_threads[i]))
2262			thread__delete(idle_threads[i]);
2263	}
2264
2265	free(idle_threads);
2266}
2267
2268static struct thread *get_idle_thread(int cpu)
2269{
2270	/*
2271	 * expand/allocate array of pointers to local thread
2272	 * structs if needed
2273	 */
2274	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2275		int i, j = __roundup_pow_of_two(cpu+1);
2276		void *p;
2277
2278		p = realloc(idle_threads, j * sizeof(struct thread *));
2279		if (!p)
2280			return NULL;
2281
2282		idle_threads = (struct thread **) p;
2283		for (i = idle_max_cpu; i < j; ++i)
2284			idle_threads[i] = NULL;
2285
2286		idle_max_cpu = j;
2287	}
2288
2289	/* allocate a new thread struct if needed */
2290	if (idle_threads[cpu] == NULL) {
2291		idle_threads[cpu] = thread__new(0, 0);
2292
2293		if (idle_threads[cpu]) {
2294			if (init_idle_thread(idle_threads[cpu]) < 0)
2295				return NULL;
2296		}
2297	}
2298
2299	return idle_threads[cpu];
2300}
2301
2302static void save_idle_callchain(struct perf_sched *sched,
2303				struct idle_thread_runtime *itr,
2304				struct perf_sample *sample)
2305{
2306	struct callchain_cursor *cursor;
2307
2308	if (!sched->show_callchain || sample->callchain == NULL)
2309		return;
2310
2311	cursor = get_tls_callchain_cursor();
2312	if (cursor == NULL)
2313		return;
2314
2315	callchain_cursor__copy(&itr->cursor, cursor);
2316}
2317
2318static struct thread *timehist_get_thread(struct perf_sched *sched,
2319					  struct perf_sample *sample,
2320					  struct machine *machine,
2321					  struct evsel *evsel)
2322{
2323	struct thread *thread;
2324
2325	if (is_idle_sample(sample, evsel)) {
2326		thread = get_idle_thread(sample->cpu);
2327		if (thread == NULL)
2328			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329
2330	} else {
2331		/* there were samples with tid 0 but non-zero pid */
2332		thread = machine__findnew_thread(machine, sample->pid,
2333						 sample->tid ?: sample->pid);
2334		if (thread == NULL) {
2335			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2336				 sample->tid);
2337		}
2338
2339		save_task_callchain(sched, sample, evsel, machine);
2340		if (sched->idle_hist) {
2341			struct thread *idle;
2342			struct idle_thread_runtime *itr;
2343
2344			idle = get_idle_thread(sample->cpu);
2345			if (idle == NULL) {
2346				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2347				return NULL;
2348			}
2349
2350			itr = thread__priv(idle);
2351			if (itr == NULL)
2352				return NULL;
2353
2354			itr->last_thread = thread;
2355
2356			/* copy task callchain when entering to idle */
2357			if (evsel__intval(evsel, sample, "next_pid") == 0)
2358				save_idle_callchain(sched, itr, sample);
2359		}
2360	}
2361
2362	return thread;
2363}
2364
2365static bool timehist_skip_sample(struct perf_sched *sched,
2366				 struct thread *thread,
2367				 struct evsel *evsel,
2368				 struct perf_sample *sample)
2369{
2370	bool rc = false;
2371
2372	if (thread__is_filtered(thread)) {
2373		rc = true;
2374		sched->skipped_samples++;
2375	}
2376
2377	if (sched->idle_hist) {
2378		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2379			rc = true;
2380		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2381			 evsel__intval(evsel, sample, "next_pid") != 0)
2382			rc = true;
2383	}
2384
2385	return rc;
2386}
2387
2388static void timehist_print_wakeup_event(struct perf_sched *sched,
2389					struct evsel *evsel,
2390					struct perf_sample *sample,
2391					struct machine *machine,
2392					struct thread *awakened)
2393{
2394	struct thread *thread;
2395	char tstr[64];
2396
2397	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2398	if (thread == NULL)
2399		return;
2400
2401	/* show wakeup unless both awakee and awaker are filtered */
2402	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2403	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2404		return;
2405	}
2406
2407	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2408	printf("%15s [%04d] ", tstr, sample->cpu);
2409	if (sched->show_cpu_visual)
2410		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2411
2412	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2413
2414	/* dt spacer */
2415	printf("  %9s  %9s  %9s ", "", "", "");
2416
2417	printf("awakened: %s", timehist_get_commstr(awakened));
2418
2419	printf("\n");
2420}
2421
2422static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2423					union perf_event *event __maybe_unused,
2424					struct evsel *evsel __maybe_unused,
2425					struct perf_sample *sample __maybe_unused,
2426					struct machine *machine __maybe_unused)
2427{
2428	return 0;
2429}
2430
2431static int timehist_sched_wakeup_event(struct perf_tool *tool,
2432				       union perf_event *event __maybe_unused,
2433				       struct evsel *evsel,
2434				       struct perf_sample *sample,
2435				       struct machine *machine)
2436{
2437	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2438	struct thread *thread;
2439	struct thread_runtime *tr = NULL;
2440	/* want pid of awakened task not pid in sample */
2441	const u32 pid = evsel__intval(evsel, sample, "pid");
2442
2443	thread = machine__findnew_thread(machine, 0, pid);
2444	if (thread == NULL)
2445		return -1;
2446
2447	tr = thread__get_runtime(thread);
2448	if (tr == NULL)
2449		return -1;
2450
2451	if (tr->ready_to_run == 0)
2452		tr->ready_to_run = sample->time;
2453
2454	/* show wakeups if requested */
2455	if (sched->show_wakeups &&
2456	    !perf_time__skip_sample(&sched->ptime, sample->time))
2457		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2458
2459	return 0;
2460}
2461
2462static void timehist_print_migration_event(struct perf_sched *sched,
2463					struct evsel *evsel,
2464					struct perf_sample *sample,
2465					struct machine *machine,
2466					struct thread *migrated)
2467{
2468	struct thread *thread;
2469	char tstr[64];
2470	u32 max_cpus;
2471	u32 ocpu, dcpu;
2472
2473	if (sched->summary_only)
2474		return;
2475
2476	max_cpus = sched->max_cpu.cpu + 1;
2477	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2478	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2479
2480	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2481	if (thread == NULL)
2482		return;
2483
2484	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2485	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2486		return;
2487	}
2488
2489	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2490	printf("%15s [%04d] ", tstr, sample->cpu);
2491
2492	if (sched->show_cpu_visual) {
2493		u32 i;
2494		char c;
2495
2496		printf("  ");
2497		for (i = 0; i < max_cpus; ++i) {
2498			c = (i == sample->cpu) ? 'm' : ' ';
2499			printf("%c", c);
2500		}
2501		printf("  ");
2502	}
2503
2504	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2505
2506	/* dt spacer */
2507	printf("  %9s  %9s  %9s ", "", "", "");
2508
2509	printf("migrated: %s", timehist_get_commstr(migrated));
2510	printf(" cpu %d => %d", ocpu, dcpu);
2511
2512	printf("\n");
2513}
2514
2515static int timehist_migrate_task_event(struct perf_tool *tool,
2516				       union perf_event *event __maybe_unused,
2517				       struct evsel *evsel,
2518				       struct perf_sample *sample,
2519				       struct machine *machine)
2520{
2521	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2522	struct thread *thread;
2523	struct thread_runtime *tr = NULL;
2524	/* want pid of migrated task not pid in sample */
2525	const u32 pid = evsel__intval(evsel, sample, "pid");
2526
2527	thread = machine__findnew_thread(machine, 0, pid);
2528	if (thread == NULL)
2529		return -1;
2530
2531	tr = thread__get_runtime(thread);
2532	if (tr == NULL)
2533		return -1;
2534
2535	tr->migrations++;
2536
2537	/* show migrations if requested */
2538	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2539
2540	return 0;
2541}
2542
2543static int timehist_sched_change_event(struct perf_tool *tool,
2544				       union perf_event *event,
2545				       struct evsel *evsel,
2546				       struct perf_sample *sample,
2547				       struct machine *machine)
2548{
2549	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2550	struct perf_time_interval *ptime = &sched->ptime;
2551	struct addr_location al;
2552	struct thread *thread;
2553	struct thread_runtime *tr = NULL;
2554	u64 tprev, t = sample->time;
2555	int rc = 0;
2556	const char state = evsel__taskstate(evsel, sample, "prev_state");
 
2557
2558	addr_location__init(&al);
2559	if (machine__resolve(machine, &al, sample) < 0) {
2560		pr_err("problem processing %d event. skipping it\n",
2561		       event->header.type);
2562		rc = -1;
2563		goto out;
2564	}
2565
2566	thread = timehist_get_thread(sched, sample, machine, evsel);
2567	if (thread == NULL) {
2568		rc = -1;
2569		goto out;
2570	}
2571
2572	if (timehist_skip_sample(sched, thread, evsel, sample))
2573		goto out;
2574
2575	tr = thread__get_runtime(thread);
2576	if (tr == NULL) {
2577		rc = -1;
2578		goto out;
2579	}
2580
2581	tprev = evsel__get_time(evsel, sample->cpu);
2582
2583	/*
2584	 * If start time given:
2585	 * - sample time is under window user cares about - skip sample
2586	 * - tprev is under window user cares about  - reset to start of window
2587	 */
2588	if (ptime->start && ptime->start > t)
2589		goto out;
2590
2591	if (tprev && ptime->start > tprev)
2592		tprev = ptime->start;
2593
2594	/*
2595	 * If end time given:
2596	 * - previous sched event is out of window - we are done
2597	 * - sample time is beyond window user cares about - reset it
2598	 *   to close out stats for time window interest
2599	 */
2600	if (ptime->end) {
2601		if (tprev > ptime->end)
2602			goto out;
2603
2604		if (t > ptime->end)
2605			t = ptime->end;
2606	}
2607
2608	if (!sched->idle_hist || thread__tid(thread) == 0) {
2609		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2610			timehist_update_runtime_stats(tr, t, tprev);
2611
2612		if (sched->idle_hist) {
2613			struct idle_thread_runtime *itr = (void *)tr;
2614			struct thread_runtime *last_tr;
2615
2616			BUG_ON(thread__tid(thread) != 0);
2617
2618			if (itr->last_thread == NULL)
2619				goto out;
2620
2621			/* add current idle time as last thread's runtime */
2622			last_tr = thread__get_runtime(itr->last_thread);
2623			if (last_tr == NULL)
2624				goto out;
2625
2626			timehist_update_runtime_stats(last_tr, t, tprev);
2627			/*
2628			 * remove delta time of last thread as it's not updated
2629			 * and otherwise it will show an invalid value next
2630			 * time.  we only care total run time and run stat.
2631			 */
2632			last_tr->dt_run = 0;
2633			last_tr->dt_delay = 0;
2634			last_tr->dt_sleep = 0;
2635			last_tr->dt_iowait = 0;
2636			last_tr->dt_preempt = 0;
2637
2638			if (itr->cursor.nr)
2639				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2640
2641			itr->last_thread = NULL;
2642		}
2643	}
2644
2645	if (!sched->summary_only)
2646		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2647
2648out:
2649	if (sched->hist_time.start == 0 && t >= ptime->start)
2650		sched->hist_time.start = t;
2651	if (ptime->end == 0 || t <= ptime->end)
2652		sched->hist_time.end = t;
2653
2654	if (tr) {
2655		/* time of this sched_switch event becomes last time task seen */
2656		tr->last_time = sample->time;
2657
2658		/* last state is used to determine where to account wait time */
2659		tr->last_state = state;
2660
2661		/* sched out event for task so reset ready to run time */
2662		tr->ready_to_run = 0;
2663	}
2664
2665	evsel__save_time(evsel, sample->time, sample->cpu);
2666
2667	addr_location__exit(&al);
2668	return rc;
2669}
2670
2671static int timehist_sched_switch_event(struct perf_tool *tool,
2672			     union perf_event *event,
2673			     struct evsel *evsel,
2674			     struct perf_sample *sample,
2675			     struct machine *machine __maybe_unused)
2676{
2677	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2678}
2679
2680static int process_lost(struct perf_tool *tool __maybe_unused,
2681			union perf_event *event,
2682			struct perf_sample *sample,
2683			struct machine *machine __maybe_unused)
2684{
2685	char tstr[64];
2686
2687	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2688	printf("%15s ", tstr);
2689	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2690
2691	return 0;
2692}
2693
2694
2695static void print_thread_runtime(struct thread *t,
2696				 struct thread_runtime *r)
2697{
2698	double mean = avg_stats(&r->run_stats);
2699	float stddev;
2700
2701	printf("%*s   %5d  %9" PRIu64 " ",
2702	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2703	       (u64) r->run_stats.n);
2704
2705	print_sched_time(r->total_run_time, 8);
2706	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2707	print_sched_time(r->run_stats.min, 6);
2708	printf(" ");
2709	print_sched_time((u64) mean, 6);
2710	printf(" ");
2711	print_sched_time(r->run_stats.max, 6);
2712	printf("  ");
2713	printf("%5.2f", stddev);
2714	printf("   %5" PRIu64, r->migrations);
2715	printf("\n");
2716}
2717
2718static void print_thread_waittime(struct thread *t,
2719				  struct thread_runtime *r)
2720{
2721	printf("%*s   %5d  %9" PRIu64 " ",
2722	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2723	       (u64) r->run_stats.n);
2724
2725	print_sched_time(r->total_run_time, 8);
2726	print_sched_time(r->total_sleep_time, 6);
2727	printf(" ");
2728	print_sched_time(r->total_iowait_time, 6);
2729	printf(" ");
2730	print_sched_time(r->total_preempt_time, 6);
2731	printf(" ");
2732	print_sched_time(r->total_delay_time, 6);
2733	printf("\n");
2734}
2735
2736struct total_run_stats {
2737	struct perf_sched *sched;
2738	u64  sched_count;
2739	u64  task_count;
2740	u64  total_run_time;
2741};
2742
2743static int show_thread_runtime(struct thread *t, void *priv)
2744{
2745	struct total_run_stats *stats = priv;
2746	struct thread_runtime *r;
2747
2748	if (thread__is_filtered(t))
2749		return 0;
2750
2751	r = thread__priv(t);
2752	if (r && r->run_stats.n) {
2753		stats->task_count++;
2754		stats->sched_count += r->run_stats.n;
2755		stats->total_run_time += r->total_run_time;
2756
2757		if (stats->sched->show_state)
2758			print_thread_waittime(t, r);
2759		else
2760			print_thread_runtime(t, r);
2761	}
2762
2763	return 0;
2764}
2765
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2766static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2767{
2768	const char *sep = " <- ";
2769	struct callchain_list *chain;
2770	size_t ret = 0;
2771	char bf[1024];
2772	bool first;
2773
2774	if (node == NULL)
2775		return 0;
2776
2777	ret = callchain__fprintf_folded(fp, node->parent);
2778	first = (ret == 0);
2779
2780	list_for_each_entry(chain, &node->val, list) {
2781		if (chain->ip >= PERF_CONTEXT_MAX)
2782			continue;
2783		if (chain->ms.sym && chain->ms.sym->ignore)
2784			continue;
2785		ret += fprintf(fp, "%s%s", first ? "" : sep,
2786			       callchain_list__sym_name(chain, bf, sizeof(bf),
2787							false));
2788		first = false;
2789	}
2790
2791	return ret;
2792}
2793
2794static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2795{
2796	size_t ret = 0;
2797	FILE *fp = stdout;
2798	struct callchain_node *chain;
2799	struct rb_node *rb_node = rb_first_cached(root);
2800
2801	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2802	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2803	       graph_dotted_line);
2804
2805	while (rb_node) {
2806		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2807		rb_node = rb_next(rb_node);
2808
2809		ret += fprintf(fp, "  ");
2810		print_sched_time(chain->hit, 12);
2811		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2812		ret += fprintf(fp, " %8d  ", chain->count);
2813		ret += callchain__fprintf_folded(fp, chain);
2814		ret += fprintf(fp, "\n");
2815	}
2816
2817	return ret;
2818}
2819
2820static void timehist_print_summary(struct perf_sched *sched,
2821				   struct perf_session *session)
2822{
2823	struct machine *m = &session->machines.host;
2824	struct total_run_stats totals;
2825	u64 task_count;
2826	struct thread *t;
2827	struct thread_runtime *r;
2828	int i;
2829	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2830
2831	memset(&totals, 0, sizeof(totals));
2832	totals.sched = sched;
2833
2834	if (sched->idle_hist) {
2835		printf("\nIdle-time summary\n");
2836		printf("%*s  parent  sched-out  ", comm_width, "comm");
2837		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2838	} else if (sched->show_state) {
2839		printf("\nWait-time summary\n");
2840		printf("%*s  parent   sched-in  ", comm_width, "comm");
2841		printf("   run-time      sleep      iowait     preempt       delay\n");
2842	} else {
2843		printf("\nRuntime summary\n");
2844		printf("%*s  parent   sched-in  ", comm_width, "comm");
2845		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2846	}
2847	printf("%*s            (count)  ", comm_width, "");
2848	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2849	       sched->show_state ? "(msec)" : "%");
2850	printf("%.117s\n", graph_dotted_line);
2851
2852	machine__for_each_thread(m, show_thread_runtime, &totals);
2853	task_count = totals.task_count;
2854	if (!task_count)
2855		printf("<no still running tasks>\n");
2856
 
 
 
 
 
2857	/* CPU idle stats not tracked when samples were skipped */
2858	if (sched->skipped_samples && !sched->idle_hist)
2859		return;
2860
2861	printf("\nIdle stats:\n");
2862	for (i = 0; i < idle_max_cpu; ++i) {
2863		if (cpu_list && !test_bit(i, cpu_bitmap))
2864			continue;
2865
2866		t = idle_threads[i];
2867		if (!t)
2868			continue;
2869
2870		r = thread__priv(t);
2871		if (r && r->run_stats.n) {
2872			totals.sched_count += r->run_stats.n;
2873			printf("    CPU %2d idle for ", i);
2874			print_sched_time(r->total_run_time, 6);
2875			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2876		} else
2877			printf("    CPU %2d idle entire time window\n", i);
2878	}
2879
2880	if (sched->idle_hist && sched->show_callchain) {
2881		callchain_param.mode  = CHAIN_FOLDED;
2882		callchain_param.value = CCVAL_PERIOD;
2883
2884		callchain_register_param(&callchain_param);
2885
2886		printf("\nIdle stats by callchain:\n");
2887		for (i = 0; i < idle_max_cpu; ++i) {
2888			struct idle_thread_runtime *itr;
2889
2890			t = idle_threads[i];
2891			if (!t)
2892				continue;
2893
2894			itr = thread__priv(t);
2895			if (itr == NULL)
2896				continue;
2897
2898			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2899					     0, &callchain_param);
2900
2901			printf("  CPU %2d:", i);
2902			print_sched_time(itr->tr.total_run_time, 6);
2903			printf(" msec\n");
2904			timehist_print_idlehist_callchain(&itr->sorted_root);
2905			printf("\n");
2906		}
2907	}
2908
2909	printf("\n"
2910	       "    Total number of unique tasks: %" PRIu64 "\n"
2911	       "Total number of context switches: %" PRIu64 "\n",
2912	       totals.task_count, totals.sched_count);
2913
2914	printf("           Total run time (msec): ");
2915	print_sched_time(totals.total_run_time, 2);
2916	printf("\n");
2917
2918	printf("    Total scheduling time (msec): ");
2919	print_sched_time(hist_time, 2);
2920	printf(" (x %d)\n", sched->max_cpu.cpu);
2921}
2922
2923typedef int (*sched_handler)(struct perf_tool *tool,
2924			  union perf_event *event,
2925			  struct evsel *evsel,
2926			  struct perf_sample *sample,
2927			  struct machine *machine);
2928
2929static int perf_timehist__process_sample(struct perf_tool *tool,
2930					 union perf_event *event,
2931					 struct perf_sample *sample,
2932					 struct evsel *evsel,
2933					 struct machine *machine)
2934{
2935	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2936	int err = 0;
2937	struct perf_cpu this_cpu = {
2938		.cpu = sample->cpu,
2939	};
2940
2941	if (this_cpu.cpu > sched->max_cpu.cpu)
2942		sched->max_cpu = this_cpu;
2943
2944	if (evsel->handler != NULL) {
2945		sched_handler f = evsel->handler;
2946
2947		err = f(tool, event, evsel, sample, machine);
2948	}
2949
2950	return err;
2951}
2952
2953static int timehist_check_attr(struct perf_sched *sched,
2954			       struct evlist *evlist)
2955{
2956	struct evsel *evsel;
2957	struct evsel_runtime *er;
2958
2959	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2960		er = evsel__get_runtime(evsel);
2961		if (er == NULL) {
2962			pr_err("Failed to allocate memory for evsel runtime data\n");
2963			return -1;
2964		}
2965
2966		/* only need to save callchain related to sched_switch event */
2967		if (sched->show_callchain &&
2968		    evsel__name_is(evsel, "sched:sched_switch") &&
2969		    !evsel__has_callchain(evsel)) {
2970			pr_info("Samples of sched_switch event do not have callchains.\n");
2971			sched->show_callchain = 0;
2972			symbol_conf.use_callchain = 0;
2973		}
2974	}
2975
2976	return 0;
2977}
2978
2979static int perf_sched__timehist(struct perf_sched *sched)
2980{
2981	struct evsel_str_handler handlers[] = {
2982		{ "sched:sched_switch",       timehist_sched_switch_event, },
2983		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2984		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
2985		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2986	};
2987	const struct evsel_str_handler migrate_handlers[] = {
2988		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2989	};
2990	struct perf_data data = {
2991		.path  = input_name,
2992		.mode  = PERF_DATA_MODE_READ,
2993		.force = sched->force,
 
 
2994	};
2995
2996	struct perf_session *session;
2997	struct evlist *evlist;
2998	int err = -1;
2999
3000	/*
3001	 * event handlers for timehist option
3002	 */
3003	sched->tool.sample	 = perf_timehist__process_sample;
3004	sched->tool.mmap	 = perf_event__process_mmap;
3005	sched->tool.comm	 = perf_event__process_comm;
3006	sched->tool.exit	 = perf_event__process_exit;
3007	sched->tool.fork	 = perf_event__process_fork;
3008	sched->tool.lost	 = process_lost;
3009	sched->tool.attr	 = perf_event__process_attr;
3010	sched->tool.tracing_data = perf_event__process_tracing_data;
3011	sched->tool.build_id	 = perf_event__process_build_id;
3012
3013	sched->tool.ordered_events = true;
3014	sched->tool.ordering_requires_timestamps = true;
3015
3016	symbol_conf.use_callchain = sched->show_callchain;
3017
3018	session = perf_session__new(&data, &sched->tool);
3019	if (IS_ERR(session))
3020		return PTR_ERR(session);
3021
3022	if (cpu_list) {
3023		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3024		if (err < 0)
3025			goto out;
3026	}
3027
3028	evlist = session->evlist;
3029
3030	symbol__init(&session->header.env);
3031
3032	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3033		pr_err("Invalid time string\n");
3034		return -EINVAL;
3035	}
3036
3037	if (timehist_check_attr(sched, evlist) != 0)
3038		goto out;
3039
3040	setup_pager();
3041
3042	/* prefer sched_waking if it is captured */
3043	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3044		handlers[1].handler = timehist_sched_wakeup_ignore;
3045
3046	/* setup per-evsel handlers */
3047	if (perf_session__set_tracepoints_handlers(session, handlers))
3048		goto out;
3049
3050	/* sched_switch event at a minimum needs to exist */
3051	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
 
3052		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3053		goto out;
3054	}
3055
3056	if (sched->show_migrations &&
3057	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3058		goto out;
3059
3060	/* pre-allocate struct for per-CPU idle stats */
3061	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3062	if (sched->max_cpu.cpu == 0)
3063		sched->max_cpu.cpu = 4;
3064	if (init_idle_threads(sched->max_cpu.cpu))
3065		goto out;
3066
3067	/* summary_only implies summary option, but don't overwrite summary if set */
3068	if (sched->summary_only)
3069		sched->summary = sched->summary_only;
3070
3071	if (!sched->summary_only)
3072		timehist_header(sched);
3073
3074	err = perf_session__process_events(session);
3075	if (err) {
3076		pr_err("Failed to process events, error %d", err);
3077		goto out;
3078	}
3079
3080	sched->nr_events      = evlist->stats.nr_events[0];
3081	sched->nr_lost_events = evlist->stats.total_lost;
3082	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3083
3084	if (sched->summary)
3085		timehist_print_summary(sched, session);
3086
3087out:
3088	free_idle_threads();
3089	perf_session__delete(session);
3090
3091	return err;
3092}
3093
3094
3095static void print_bad_events(struct perf_sched *sched)
3096{
3097	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3098		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3099			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3100			sched->nr_unordered_timestamps, sched->nr_timestamps);
3101	}
3102	if (sched->nr_lost_events && sched->nr_events) {
3103		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3104			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3105			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3106	}
3107	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3108		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3109			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3110			sched->nr_context_switch_bugs, sched->nr_timestamps);
3111		if (sched->nr_lost_events)
3112			printf(" (due to lost events?)");
3113		printf("\n");
3114	}
3115}
3116
3117static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3118{
3119	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3120	struct work_atoms *this;
3121	const char *comm = thread__comm_str(data->thread), *this_comm;
3122	bool leftmost = true;
3123
3124	while (*new) {
3125		int cmp;
3126
3127		this = container_of(*new, struct work_atoms, node);
3128		parent = *new;
3129
3130		this_comm = thread__comm_str(this->thread);
3131		cmp = strcmp(comm, this_comm);
3132		if (cmp > 0) {
3133			new = &((*new)->rb_left);
3134		} else if (cmp < 0) {
3135			new = &((*new)->rb_right);
3136			leftmost = false;
3137		} else {
3138			this->num_merged++;
3139			this->total_runtime += data->total_runtime;
3140			this->nb_atoms += data->nb_atoms;
3141			this->total_lat += data->total_lat;
3142			list_splice(&data->work_list, &this->work_list);
3143			if (this->max_lat < data->max_lat) {
3144				this->max_lat = data->max_lat;
3145				this->max_lat_start = data->max_lat_start;
3146				this->max_lat_end = data->max_lat_end;
3147			}
3148			zfree(&data);
3149			return;
3150		}
3151	}
3152
3153	data->num_merged++;
3154	rb_link_node(&data->node, parent, new);
3155	rb_insert_color_cached(&data->node, root, leftmost);
3156}
3157
3158static void perf_sched__merge_lat(struct perf_sched *sched)
3159{
3160	struct work_atoms *data;
3161	struct rb_node *node;
3162
3163	if (sched->skip_merge)
3164		return;
3165
3166	while ((node = rb_first_cached(&sched->atom_root))) {
3167		rb_erase_cached(node, &sched->atom_root);
3168		data = rb_entry(node, struct work_atoms, node);
3169		__merge_work_atoms(&sched->merged_atom_root, data);
3170	}
3171}
3172
3173static int setup_cpus_switch_event(struct perf_sched *sched)
3174{
3175	unsigned int i;
3176
3177	sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3178	if (!sched->cpu_last_switched)
3179		return -1;
3180
3181	sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3182	if (!sched->curr_pid) {
3183		zfree(&sched->cpu_last_switched);
3184		return -1;
3185	}
3186
3187	for (i = 0; i < MAX_CPUS; i++)
3188		sched->curr_pid[i] = -1;
3189
3190	return 0;
3191}
3192
3193static void free_cpus_switch_event(struct perf_sched *sched)
3194{
3195	zfree(&sched->curr_pid);
3196	zfree(&sched->cpu_last_switched);
3197}
3198
3199static int perf_sched__lat(struct perf_sched *sched)
3200{
3201	int rc = -1;
3202	struct rb_node *next;
3203
3204	setup_pager();
3205
3206	if (setup_cpus_switch_event(sched))
3207		return rc;
3208
3209	if (perf_sched__read_events(sched))
3210		goto out_free_cpus_switch_event;
3211
3212	perf_sched__merge_lat(sched);
3213	perf_sched__sort_lat(sched);
3214
3215	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3216	printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3217	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3218
3219	next = rb_first_cached(&sched->sorted_atom_root);
3220
3221	while (next) {
3222		struct work_atoms *work_list;
3223
3224		work_list = rb_entry(next, struct work_atoms, node);
3225		output_lat_thread(sched, work_list);
3226		next = rb_next(next);
3227		thread__zput(work_list->thread);
3228	}
3229
3230	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3231	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3232		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3233
3234	printf(" ---------------------------------------------------\n");
3235
3236	print_bad_events(sched);
3237	printf("\n");
3238
3239	rc = 0;
3240
3241out_free_cpus_switch_event:
3242	free_cpus_switch_event(sched);
3243	return rc;
3244}
3245
3246static int setup_map_cpus(struct perf_sched *sched)
3247{
3248	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
 
 
3249
3250	if (sched->map.comp) {
3251		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3252		if (!sched->map.comp_cpus)
3253			return -1;
3254	}
3255
3256	if (sched->map.cpus_str) {
3257		sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3258		if (!sched->map.cpus) {
3259			pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3260			zfree(&sched->map.comp_cpus);
3261			return -1;
3262		}
3263	}
3264
 
3265	return 0;
3266}
3267
3268static int setup_color_pids(struct perf_sched *sched)
3269{
3270	struct perf_thread_map *map;
3271
3272	if (!sched->map.color_pids_str)
3273		return 0;
3274
3275	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3276	if (!map) {
3277		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3278		return -1;
3279	}
3280
3281	sched->map.color_pids = map;
3282	return 0;
3283}
3284
3285static int setup_color_cpus(struct perf_sched *sched)
3286{
3287	struct perf_cpu_map *map;
3288
3289	if (!sched->map.color_cpus_str)
3290		return 0;
3291
3292	map = perf_cpu_map__new(sched->map.color_cpus_str);
3293	if (!map) {
3294		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3295		return -1;
3296	}
3297
3298	sched->map.color_cpus = map;
3299	return 0;
3300}
3301
3302static int perf_sched__map(struct perf_sched *sched)
3303{
3304	int rc = -1;
3305
3306	sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3307	if (!sched->curr_thread)
3308		return rc;
3309
3310	if (setup_cpus_switch_event(sched))
3311		goto out_free_curr_thread;
3312
3313	if (setup_map_cpus(sched))
3314		goto out_free_cpus_switch_event;
3315
3316	if (setup_color_pids(sched))
3317		goto out_put_map_cpus;
3318
3319	if (setup_color_cpus(sched))
3320		goto out_put_color_pids;
3321
3322	setup_pager();
3323	if (perf_sched__read_events(sched))
3324		goto out_put_color_cpus;
3325
3326	rc = 0;
3327	print_bad_events(sched);
3328
3329out_put_color_cpus:
3330	perf_cpu_map__put(sched->map.color_cpus);
3331
3332out_put_color_pids:
3333	perf_thread_map__put(sched->map.color_pids);
3334
3335out_put_map_cpus:
3336	zfree(&sched->map.comp_cpus);
3337	perf_cpu_map__put(sched->map.cpus);
3338
3339out_free_cpus_switch_event:
3340	free_cpus_switch_event(sched);
3341
3342out_free_curr_thread:
3343	zfree(&sched->curr_thread);
3344	return rc;
3345}
3346
3347static int perf_sched__replay(struct perf_sched *sched)
3348{
3349	int ret;
3350	unsigned long i;
3351
3352	mutex_init(&sched->start_work_mutex);
3353	mutex_init(&sched->work_done_wait_mutex);
3354
3355	ret = setup_cpus_switch_event(sched);
3356	if (ret)
3357		goto out_mutex_destroy;
3358
3359	calibrate_run_measurement_overhead(sched);
3360	calibrate_sleep_measurement_overhead(sched);
3361
3362	test_calibrations(sched);
3363
3364	ret = perf_sched__read_events(sched);
3365	if (ret)
3366		goto out_free_cpus_switch_event;
3367
3368	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3369	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3370	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3371
3372	if (sched->targetless_wakeups)
3373		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3374	if (sched->multitarget_wakeups)
3375		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3376	if (sched->nr_run_events_optimized)
3377		printf("run atoms optimized: %ld\n",
3378			sched->nr_run_events_optimized);
3379
3380	print_task_traces(sched);
3381	add_cross_task_wakeups(sched);
3382
3383	sched->thread_funcs_exit = false;
3384	create_tasks(sched);
3385	printf("------------------------------------------------------------\n");
3386	for (i = 0; i < sched->replay_repeat; i++)
3387		run_one_test(sched);
3388
3389	sched->thread_funcs_exit = true;
3390	destroy_tasks(sched);
3391
3392out_free_cpus_switch_event:
3393	free_cpus_switch_event(sched);
3394
3395out_mutex_destroy:
3396	mutex_destroy(&sched->start_work_mutex);
3397	mutex_destroy(&sched->work_done_wait_mutex);
3398	return ret;
3399}
3400
3401static void setup_sorting(struct perf_sched *sched, const struct option *options,
3402			  const char * const usage_msg[])
3403{
3404	char *tmp, *tok, *str = strdup(sched->sort_order);
3405
3406	for (tok = strtok_r(str, ", ", &tmp);
3407			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3408		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3409			usage_with_options_msg(usage_msg, options,
3410					"Unknown --sort key: `%s'", tok);
3411		}
3412	}
3413
3414	free(str);
3415
3416	sort_dimension__add("pid", &sched->cmp_pid);
3417}
3418
3419static bool schedstat_events_exposed(void)
3420{
3421	/*
3422	 * Select "sched:sched_stat_wait" event to check
3423	 * whether schedstat tracepoints are exposed.
3424	 */
3425	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3426		false : true;
3427}
3428
3429static int __cmd_record(int argc, const char **argv)
3430{
3431	unsigned int rec_argc, i, j;
3432	char **rec_argv;
3433	const char **rec_argv_copy;
3434	const char * const record_args[] = {
3435		"record",
3436		"-a",
3437		"-R",
3438		"-m", "1024",
3439		"-c", "1",
3440		"-e", "sched:sched_switch",
 
 
 
3441		"-e", "sched:sched_stat_runtime",
3442		"-e", "sched:sched_process_fork",
 
3443		"-e", "sched:sched_wakeup_new",
3444		"-e", "sched:sched_migrate_task",
3445	};
3446
3447	/*
3448	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3449	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3450	 * to prevent "perf sched record" execution failure, determine
3451	 * whether to record schedstat events according to actual situation.
3452	 */
3453	const char * const schedstat_args[] = {
3454		"-e", "sched:sched_stat_wait",
3455		"-e", "sched:sched_stat_sleep",
3456		"-e", "sched:sched_stat_iowait",
3457	};
3458	unsigned int schedstat_argc = schedstat_events_exposed() ?
3459		ARRAY_SIZE(schedstat_args) : 0;
3460
3461	struct tep_event *waking_event;
3462	int ret;
3463
3464	/*
3465	 * +2 for either "-e", "sched:sched_wakeup" or
3466	 * "-e", "sched:sched_waking"
3467	 */
3468	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3469	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3470	if (rec_argv == NULL)
3471		return -ENOMEM;
3472	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3473	if (rec_argv_copy == NULL) {
3474		free(rec_argv);
3475		return -ENOMEM;
3476	}
3477
3478	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3479		rec_argv[i] = strdup(record_args[i]);
3480
3481	rec_argv[i++] = strdup("-e");
3482	waking_event = trace_event__tp_format("sched", "sched_waking");
3483	if (!IS_ERR(waking_event))
3484		rec_argv[i++] = strdup("sched:sched_waking");
3485	else
3486		rec_argv[i++] = strdup("sched:sched_wakeup");
3487
3488	for (j = 0; j < schedstat_argc; j++)
3489		rec_argv[i++] = strdup(schedstat_args[j]);
3490
3491	for (j = 1; j < (unsigned int)argc; j++, i++)
3492		rec_argv[i] = strdup(argv[j]);
3493
3494	BUG_ON(i != rec_argc);
3495
3496	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3497	ret = cmd_record(rec_argc, rec_argv_copy);
3498
3499	for (i = 0; i < rec_argc; i++)
3500		free(rec_argv[i]);
3501	free(rec_argv);
3502	free(rec_argv_copy);
3503
3504	return ret;
3505}
3506
3507int cmd_sched(int argc, const char **argv)
3508{
3509	static const char default_sort_order[] = "avg, max, switch, runtime";
3510	struct perf_sched sched = {
3511		.tool = {
3512			.sample		 = perf_sched__process_tracepoint_sample,
3513			.comm		 = perf_sched__process_comm,
3514			.namespaces	 = perf_event__process_namespaces,
3515			.lost		 = perf_event__process_lost,
3516			.fork		 = perf_sched__process_fork_event,
3517			.ordered_events = true,
3518		},
3519		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3520		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
 
 
3521		.sort_order	      = default_sort_order,
3522		.replay_repeat	      = 10,
3523		.profile_cpu	      = -1,
3524		.next_shortname1      = 'A',
3525		.next_shortname2      = '0',
3526		.skip_merge           = 0,
3527		.show_callchain	      = 1,
3528		.max_stack            = 5,
3529	};
3530	const struct option sched_options[] = {
3531	OPT_STRING('i', "input", &input_name, "file",
3532		    "input file name"),
3533	OPT_INCR('v', "verbose", &verbose,
3534		    "be more verbose (show symbol address, etc)"),
3535	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3536		    "dump raw trace in ASCII"),
3537	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3538	OPT_END()
3539	};
3540	const struct option latency_options[] = {
3541	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3542		   "sort by key(s): runtime, switch, avg, max"),
3543	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3544		    "CPU to profile on"),
3545	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3546		    "latency stats per pid instead of per comm"),
3547	OPT_PARENT(sched_options)
3548	};
3549	const struct option replay_options[] = {
3550	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3551		     "repeat the workload replay N times (-1: infinite)"),
3552	OPT_PARENT(sched_options)
3553	};
3554	const struct option map_options[] = {
3555	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3556		    "map output in compact mode"),
3557	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3558		   "highlight given pids in map"),
3559	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3560                    "highlight given CPUs in map"),
3561	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3562                    "display given CPUs in map"),
3563	OPT_PARENT(sched_options)
3564	};
3565	const struct option timehist_options[] = {
3566	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3567		   "file", "vmlinux pathname"),
3568	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3569		   "file", "kallsyms pathname"),
3570	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3571		    "Display call chains if present (default on)"),
3572	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3573		   "Maximum number of functions to display backtrace."),
3574	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3575		    "Look for files with symbols relative to this directory"),
3576	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3577		    "Show only syscall summary with statistics"),
3578	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3579		    "Show all syscalls and summary with statistics"),
3580	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3581	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3582	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3583	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3584	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3585	OPT_STRING(0, "time", &sched.time_str, "str",
3586		   "Time span for analysis (start,stop)"),
3587	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3588	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3589		   "analyze events only for given process id(s)"),
3590	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3591		   "analyze events only for given thread id(s)"),
3592	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3593	OPT_PARENT(sched_options)
3594	};
3595
3596	const char * const latency_usage[] = {
3597		"perf sched latency [<options>]",
3598		NULL
3599	};
3600	const char * const replay_usage[] = {
3601		"perf sched replay [<options>]",
3602		NULL
3603	};
3604	const char * const map_usage[] = {
3605		"perf sched map [<options>]",
3606		NULL
3607	};
3608	const char * const timehist_usage[] = {
3609		"perf sched timehist [<options>]",
3610		NULL
3611	};
3612	const char *const sched_subcommands[] = { "record", "latency", "map",
3613						  "replay", "script",
3614						  "timehist", NULL };
3615	const char *sched_usage[] = {
3616		NULL,
3617		NULL
3618	};
3619	struct trace_sched_handler lat_ops  = {
3620		.wakeup_event	    = latency_wakeup_event,
3621		.switch_event	    = latency_switch_event,
3622		.runtime_event	    = latency_runtime_event,
3623		.migrate_task_event = latency_migrate_task_event,
3624	};
3625	struct trace_sched_handler map_ops  = {
3626		.switch_event	    = map_switch_event,
3627	};
3628	struct trace_sched_handler replay_ops  = {
3629		.wakeup_event	    = replay_wakeup_event,
3630		.switch_event	    = replay_switch_event,
3631		.fork_event	    = replay_fork_event,
3632	};
3633	int ret;
 
 
 
3634
3635	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3636					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3637	if (!argc)
3638		usage_with_options(sched_usage, sched_options);
3639
3640	/*
3641	 * Aliased to 'perf script' for now:
3642	 */
3643	if (!strcmp(argv[0], "script")) {
3644		return cmd_script(argc, argv);
3645	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
 
3646		return __cmd_record(argc, argv);
3647	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3648		sched.tp_handler = &lat_ops;
3649		if (argc > 1) {
3650			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3651			if (argc)
3652				usage_with_options(latency_usage, latency_options);
3653		}
3654		setup_sorting(&sched, latency_options, latency_usage);
3655		return perf_sched__lat(&sched);
3656	} else if (!strcmp(argv[0], "map")) {
3657		if (argc) {
3658			argc = parse_options(argc, argv, map_options, map_usage, 0);
3659			if (argc)
3660				usage_with_options(map_usage, map_options);
3661		}
3662		sched.tp_handler = &map_ops;
3663		setup_sorting(&sched, latency_options, latency_usage);
3664		return perf_sched__map(&sched);
3665	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3666		sched.tp_handler = &replay_ops;
3667		if (argc) {
3668			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3669			if (argc)
3670				usage_with_options(replay_usage, replay_options);
3671		}
3672		return perf_sched__replay(&sched);
3673	} else if (!strcmp(argv[0], "timehist")) {
3674		if (argc) {
3675			argc = parse_options(argc, argv, timehist_options,
3676					     timehist_usage, 0);
3677			if (argc)
3678				usage_with_options(timehist_usage, timehist_options);
3679		}
3680		if ((sched.show_wakeups || sched.show_next) &&
3681		    sched.summary_only) {
3682			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3683			parse_options_usage(timehist_usage, timehist_options, "s", true);
3684			if (sched.show_wakeups)
3685				parse_options_usage(NULL, timehist_options, "w", true);
3686			if (sched.show_next)
3687				parse_options_usage(NULL, timehist_options, "n", true);
3688			return -EINVAL;
3689		}
3690		ret = symbol__validate_sym_arguments();
3691		if (ret)
3692			return ret;
3693
3694		return perf_sched__timehist(&sched);
3695	} else {
3696		usage_with_options(sched_usage, sched_options);
3697	}
3698
3699	return 0;
3700}