xref: /linux/tools/perf/builtin-sched.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 
47 struct sched_atom;
48 
49 struct 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 
67 enum sched_event_type {
68 	SCHED_EVENT_RUN,
69 	SCHED_EVENT_SLEEP,
70 	SCHED_EVENT_WAKEUP,
71 	SCHED_EVENT_MIGRATION,
72 };
73 
74 struct 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 
102 enum thread_state {
103 	THREAD_SLEEPING = 0,
104 	THREAD_WAIT_CPU,
105 	THREAD_SCHED_IN,
106 	THREAD_IGNORE
107 };
108 
109 struct 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 
118 struct 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 
130 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
131 
132 struct perf_sched;
133 
134 struct 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 
157 struct 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 
169 struct 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_cached 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 */
240 struct 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 */
265 struct 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 */
271 struct idle_thread_runtime {
272 	struct thread_runtime	tr;
273 	struct thread		*last_thread;
274 	struct rb_root_cached	sorted_root;
275 	struct callchain_root	callchain;
276 	struct callchain_cursor	cursor;
277 };
278 
279 /* track idle times per cpu */
280 static struct thread **idle_threads;
281 static int idle_max_cpu;
282 static char idle_comm[] = "<idle>";
283 
284 static 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 
293 static 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 
302 static 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 
312 static 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 
329 static 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 
347 static struct sched_atom *
348 get_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 
367 static 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 
375 static 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 
398 static 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 
425 static 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 
435 static 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 
481 static 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 
493 static 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 
508 static 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 
534 static 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 
549 static 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 
562 force_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 
591 static 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 
602 struct sched_thread_parms {
603 	struct task_desc  *task;
604 	struct perf_sched *sched;
605 	int fd;
606 };
607 
608 static 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;
624 again:
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 
652 static 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 
683 static 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 
736 static 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 
782 static 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 
799 static int
800 replay_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 
821 static 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 
867 static 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));
892 out_put:
893 	thread__put(child);
894 	thread__put(parent);
895 	return 0;
896 }
897 
898 struct 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  */
907 static 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 
921 static 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 
935 static int
936 thread_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 
952 static struct work_atoms *
953 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
954 			 struct list_head *sort_list)
955 {
956 	struct rb_node *node = root->rb_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 
978 static void
979 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
980 			 struct list_head *sort_list)
981 {
982 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
983 	bool leftmost = true;
984 
985 	while (*new) {
986 		struct work_atoms *this;
987 		int cmp;
988 
989 		this = container_of(*new, struct work_atoms, node);
990 		parent = *new;
991 
992 		cmp = thread_lat_cmp(sort_list, data, this);
993 
994 		if (cmp > 0)
995 			new = &((*new)->rb_left);
996 		else {
997 			new = &((*new)->rb_right);
998 			leftmost = false;
999 		}
1000 	}
1001 
1002 	rb_link_node(&data->node, parent, new);
1003 	rb_insert_color_cached(&data->node, root, leftmost);
1004 }
1005 
1006 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1007 {
1008 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1009 	if (!atoms) {
1010 		pr_err("No memory at %s\n", __func__);
1011 		return -1;
1012 	}
1013 
1014 	atoms->thread = thread__get(thread);
1015 	INIT_LIST_HEAD(&atoms->work_list);
1016 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1017 	return 0;
1018 }
1019 
1020 static char sched_out_state(u64 prev_state)
1021 {
1022 	const char *str = TASK_STATE_TO_CHAR_STR;
1023 
1024 	return str[prev_state];
1025 }
1026 
1027 static int
1028 add_sched_out_event(struct work_atoms *atoms,
1029 		    char run_state,
1030 		    u64 timestamp)
1031 {
1032 	struct work_atom *atom = zalloc(sizeof(*atom));
1033 	if (!atom) {
1034 		pr_err("Non memory at %s", __func__);
1035 		return -1;
1036 	}
1037 
1038 	atom->sched_out_time = timestamp;
1039 
1040 	if (run_state == 'R') {
1041 		atom->state = THREAD_WAIT_CPU;
1042 		atom->wake_up_time = atom->sched_out_time;
1043 	}
1044 
1045 	list_add_tail(&atom->list, &atoms->work_list);
1046 	return 0;
1047 }
1048 
1049 static void
1050 add_runtime_event(struct work_atoms *atoms, u64 delta,
1051 		  u64 timestamp __maybe_unused)
1052 {
1053 	struct work_atom *atom;
1054 
1055 	BUG_ON(list_empty(&atoms->work_list));
1056 
1057 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1058 
1059 	atom->runtime += delta;
1060 	atoms->total_runtime += delta;
1061 }
1062 
1063 static void
1064 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1065 {
1066 	struct work_atom *atom;
1067 	u64 delta;
1068 
1069 	if (list_empty(&atoms->work_list))
1070 		return;
1071 
1072 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1073 
1074 	if (atom->state != THREAD_WAIT_CPU)
1075 		return;
1076 
1077 	if (timestamp < atom->wake_up_time) {
1078 		atom->state = THREAD_IGNORE;
1079 		return;
1080 	}
1081 
1082 	atom->state = THREAD_SCHED_IN;
1083 	atom->sched_in_time = timestamp;
1084 
1085 	delta = atom->sched_in_time - atom->wake_up_time;
1086 	atoms->total_lat += delta;
1087 	if (delta > atoms->max_lat) {
1088 		atoms->max_lat = delta;
1089 		atoms->max_lat_at = timestamp;
1090 	}
1091 	atoms->nb_atoms++;
1092 }
1093 
1094 static int latency_switch_event(struct perf_sched *sched,
1095 				struct perf_evsel *evsel,
1096 				struct perf_sample *sample,
1097 				struct machine *machine)
1098 {
1099 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1100 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1101 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1102 	struct work_atoms *out_events, *in_events;
1103 	struct thread *sched_out, *sched_in;
1104 	u64 timestamp0, timestamp = sample->time;
1105 	int cpu = sample->cpu, err = -1;
1106 	s64 delta;
1107 
1108 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1109 
1110 	timestamp0 = sched->cpu_last_switched[cpu];
1111 	sched->cpu_last_switched[cpu] = timestamp;
1112 	if (timestamp0)
1113 		delta = timestamp - timestamp0;
1114 	else
1115 		delta = 0;
1116 
1117 	if (delta < 0) {
1118 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1119 		return -1;
1120 	}
1121 
1122 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1123 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1124 	if (sched_out == NULL || sched_in == NULL)
1125 		goto out_put;
1126 
1127 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1128 	if (!out_events) {
1129 		if (thread_atoms_insert(sched, sched_out))
1130 			goto out_put;
1131 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1132 		if (!out_events) {
1133 			pr_err("out-event: Internal tree error");
1134 			goto out_put;
1135 		}
1136 	}
1137 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1138 		return -1;
1139 
1140 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1141 	if (!in_events) {
1142 		if (thread_atoms_insert(sched, sched_in))
1143 			goto out_put;
1144 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1145 		if (!in_events) {
1146 			pr_err("in-event: Internal tree error");
1147 			goto out_put;
1148 		}
1149 		/*
1150 		 * Take came in we have not heard about yet,
1151 		 * add in an initial atom in runnable state:
1152 		 */
1153 		if (add_sched_out_event(in_events, 'R', timestamp))
1154 			goto out_put;
1155 	}
1156 	add_sched_in_event(in_events, timestamp);
1157 	err = 0;
1158 out_put:
1159 	thread__put(sched_out);
1160 	thread__put(sched_in);
1161 	return err;
1162 }
1163 
1164 static int latency_runtime_event(struct perf_sched *sched,
1165 				 struct perf_evsel *evsel,
1166 				 struct perf_sample *sample,
1167 				 struct machine *machine)
1168 {
1169 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1170 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1171 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1172 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1173 	u64 timestamp = sample->time;
1174 	int cpu = sample->cpu, err = -1;
1175 
1176 	if (thread == NULL)
1177 		return -1;
1178 
1179 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1180 	if (!atoms) {
1181 		if (thread_atoms_insert(sched, thread))
1182 			goto out_put;
1183 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1184 		if (!atoms) {
1185 			pr_err("in-event: Internal tree error");
1186 			goto out_put;
1187 		}
1188 		if (add_sched_out_event(atoms, 'R', timestamp))
1189 			goto out_put;
1190 	}
1191 
1192 	add_runtime_event(atoms, runtime, timestamp);
1193 	err = 0;
1194 out_put:
1195 	thread__put(thread);
1196 	return err;
1197 }
1198 
1199 static int latency_wakeup_event(struct perf_sched *sched,
1200 				struct perf_evsel *evsel,
1201 				struct perf_sample *sample,
1202 				struct machine *machine)
1203 {
1204 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1205 	struct work_atoms *atoms;
1206 	struct work_atom *atom;
1207 	struct thread *wakee;
1208 	u64 timestamp = sample->time;
1209 	int err = -1;
1210 
1211 	wakee = machine__findnew_thread(machine, -1, pid);
1212 	if (wakee == NULL)
1213 		return -1;
1214 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1215 	if (!atoms) {
1216 		if (thread_atoms_insert(sched, wakee))
1217 			goto out_put;
1218 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1219 		if (!atoms) {
1220 			pr_err("wakeup-event: Internal tree error");
1221 			goto out_put;
1222 		}
1223 		if (add_sched_out_event(atoms, 'S', timestamp))
1224 			goto out_put;
1225 	}
1226 
1227 	BUG_ON(list_empty(&atoms->work_list));
1228 
1229 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1230 
1231 	/*
1232 	 * As we do not guarantee the wakeup event happens when
1233 	 * task is out of run queue, also may happen when task is
1234 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1235 	 * then we should not set the ->wake_up_time when wake up a
1236 	 * task which is on run queue.
1237 	 *
1238 	 * You WILL be missing events if you've recorded only
1239 	 * one CPU, or are only looking at only one, so don't
1240 	 * skip in this case.
1241 	 */
1242 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1243 		goto out_ok;
1244 
1245 	sched->nr_timestamps++;
1246 	if (atom->sched_out_time > timestamp) {
1247 		sched->nr_unordered_timestamps++;
1248 		goto out_ok;
1249 	}
1250 
1251 	atom->state = THREAD_WAIT_CPU;
1252 	atom->wake_up_time = timestamp;
1253 out_ok:
1254 	err = 0;
1255 out_put:
1256 	thread__put(wakee);
1257 	return err;
1258 }
1259 
1260 static int latency_migrate_task_event(struct perf_sched *sched,
1261 				      struct perf_evsel *evsel,
1262 				      struct perf_sample *sample,
1263 				      struct machine *machine)
1264 {
1265 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1266 	u64 timestamp = sample->time;
1267 	struct work_atoms *atoms;
1268 	struct work_atom *atom;
1269 	struct thread *migrant;
1270 	int err = -1;
1271 
1272 	/*
1273 	 * Only need to worry about migration when profiling one CPU.
1274 	 */
1275 	if (sched->profile_cpu == -1)
1276 		return 0;
1277 
1278 	migrant = machine__findnew_thread(machine, -1, pid);
1279 	if (migrant == NULL)
1280 		return -1;
1281 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1282 	if (!atoms) {
1283 		if (thread_atoms_insert(sched, migrant))
1284 			goto out_put;
1285 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1286 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1287 		if (!atoms) {
1288 			pr_err("migration-event: Internal tree error");
1289 			goto out_put;
1290 		}
1291 		if (add_sched_out_event(atoms, 'R', timestamp))
1292 			goto out_put;
1293 	}
1294 
1295 	BUG_ON(list_empty(&atoms->work_list));
1296 
1297 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1298 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1299 
1300 	sched->nr_timestamps++;
1301 
1302 	if (atom->sched_out_time > timestamp)
1303 		sched->nr_unordered_timestamps++;
1304 	err = 0;
1305 out_put:
1306 	thread__put(migrant);
1307 	return err;
1308 }
1309 
1310 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1311 {
1312 	int i;
1313 	int ret;
1314 	u64 avg;
1315 	char max_lat_at[32];
1316 
1317 	if (!work_list->nb_atoms)
1318 		return;
1319 	/*
1320 	 * Ignore idle threads:
1321 	 */
1322 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1323 		return;
1324 
1325 	sched->all_runtime += work_list->total_runtime;
1326 	sched->all_count   += work_list->nb_atoms;
1327 
1328 	if (work_list->num_merged > 1)
1329 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1330 	else
1331 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1332 
1333 	for (i = 0; i < 24 - ret; i++)
1334 		printf(" ");
1335 
1336 	avg = work_list->total_lat / work_list->nb_atoms;
1337 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1338 
1339 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1340 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1341 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1342 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1343 		 max_lat_at);
1344 }
1345 
1346 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1347 {
1348 	if (l->thread == r->thread)
1349 		return 0;
1350 	if (l->thread->tid < r->thread->tid)
1351 		return -1;
1352 	if (l->thread->tid > r->thread->tid)
1353 		return 1;
1354 	return (int)(l->thread - r->thread);
1355 }
1356 
1357 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1358 {
1359 	u64 avgl, avgr;
1360 
1361 	if (!l->nb_atoms)
1362 		return -1;
1363 
1364 	if (!r->nb_atoms)
1365 		return 1;
1366 
1367 	avgl = l->total_lat / l->nb_atoms;
1368 	avgr = r->total_lat / r->nb_atoms;
1369 
1370 	if (avgl < avgr)
1371 		return -1;
1372 	if (avgl > avgr)
1373 		return 1;
1374 
1375 	return 0;
1376 }
1377 
1378 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1379 {
1380 	if (l->max_lat < r->max_lat)
1381 		return -1;
1382 	if (l->max_lat > r->max_lat)
1383 		return 1;
1384 
1385 	return 0;
1386 }
1387 
1388 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1389 {
1390 	if (l->nb_atoms < r->nb_atoms)
1391 		return -1;
1392 	if (l->nb_atoms > r->nb_atoms)
1393 		return 1;
1394 
1395 	return 0;
1396 }
1397 
1398 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1399 {
1400 	if (l->total_runtime < r->total_runtime)
1401 		return -1;
1402 	if (l->total_runtime > r->total_runtime)
1403 		return 1;
1404 
1405 	return 0;
1406 }
1407 
1408 static int sort_dimension__add(const char *tok, struct list_head *list)
1409 {
1410 	size_t i;
1411 	static struct sort_dimension avg_sort_dimension = {
1412 		.name = "avg",
1413 		.cmp  = avg_cmp,
1414 	};
1415 	static struct sort_dimension max_sort_dimension = {
1416 		.name = "max",
1417 		.cmp  = max_cmp,
1418 	};
1419 	static struct sort_dimension pid_sort_dimension = {
1420 		.name = "pid",
1421 		.cmp  = pid_cmp,
1422 	};
1423 	static struct sort_dimension runtime_sort_dimension = {
1424 		.name = "runtime",
1425 		.cmp  = runtime_cmp,
1426 	};
1427 	static struct sort_dimension switch_sort_dimension = {
1428 		.name = "switch",
1429 		.cmp  = switch_cmp,
1430 	};
1431 	struct sort_dimension *available_sorts[] = {
1432 		&pid_sort_dimension,
1433 		&avg_sort_dimension,
1434 		&max_sort_dimension,
1435 		&switch_sort_dimension,
1436 		&runtime_sort_dimension,
1437 	};
1438 
1439 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1440 		if (!strcmp(available_sorts[i]->name, tok)) {
1441 			list_add_tail(&available_sorts[i]->list, list);
1442 
1443 			return 0;
1444 		}
1445 	}
1446 
1447 	return -1;
1448 }
1449 
1450 static void perf_sched__sort_lat(struct perf_sched *sched)
1451 {
1452 	struct rb_node *node;
1453 	struct rb_root_cached *root = &sched->atom_root;
1454 again:
1455 	for (;;) {
1456 		struct work_atoms *data;
1457 		node = rb_first_cached(root);
1458 		if (!node)
1459 			break;
1460 
1461 		rb_erase_cached(node, root);
1462 		data = rb_entry(node, struct work_atoms, node);
1463 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1464 	}
1465 	if (root == &sched->atom_root) {
1466 		root = &sched->merged_atom_root;
1467 		goto again;
1468 	}
1469 }
1470 
1471 static int process_sched_wakeup_event(struct perf_tool *tool,
1472 				      struct perf_evsel *evsel,
1473 				      struct perf_sample *sample,
1474 				      struct machine *machine)
1475 {
1476 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1477 
1478 	if (sched->tp_handler->wakeup_event)
1479 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1480 
1481 	return 0;
1482 }
1483 
1484 union map_priv {
1485 	void	*ptr;
1486 	bool	 color;
1487 };
1488 
1489 static bool thread__has_color(struct thread *thread)
1490 {
1491 	union map_priv priv = {
1492 		.ptr = thread__priv(thread),
1493 	};
1494 
1495 	return priv.color;
1496 }
1497 
1498 static struct thread*
1499 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1500 {
1501 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1502 	union map_priv priv = {
1503 		.color = false,
1504 	};
1505 
1506 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1507 		return thread;
1508 
1509 	if (thread_map__has(sched->map.color_pids, tid))
1510 		priv.color = true;
1511 
1512 	thread__set_priv(thread, priv.ptr);
1513 	return thread;
1514 }
1515 
1516 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1517 			    struct perf_sample *sample, struct machine *machine)
1518 {
1519 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1520 	struct thread *sched_in;
1521 	struct thread_runtime *tr;
1522 	int new_shortname;
1523 	u64 timestamp0, timestamp = sample->time;
1524 	s64 delta;
1525 	int i, this_cpu = sample->cpu;
1526 	int cpus_nr;
1527 	bool new_cpu = false;
1528 	const char *color = PERF_COLOR_NORMAL;
1529 	char stimestamp[32];
1530 
1531 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1532 
1533 	if (this_cpu > sched->max_cpu)
1534 		sched->max_cpu = this_cpu;
1535 
1536 	if (sched->map.comp) {
1537 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1538 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1539 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1540 			new_cpu = true;
1541 		}
1542 	} else
1543 		cpus_nr = sched->max_cpu;
1544 
1545 	timestamp0 = sched->cpu_last_switched[this_cpu];
1546 	sched->cpu_last_switched[this_cpu] = timestamp;
1547 	if (timestamp0)
1548 		delta = timestamp - timestamp0;
1549 	else
1550 		delta = 0;
1551 
1552 	if (delta < 0) {
1553 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1554 		return -1;
1555 	}
1556 
1557 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1558 	if (sched_in == NULL)
1559 		return -1;
1560 
1561 	tr = thread__get_runtime(sched_in);
1562 	if (tr == NULL) {
1563 		thread__put(sched_in);
1564 		return -1;
1565 	}
1566 
1567 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1568 
1569 	printf("  ");
1570 
1571 	new_shortname = 0;
1572 	if (!tr->shortname[0]) {
1573 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1574 			/*
1575 			 * Don't allocate a letter-number for swapper:0
1576 			 * as a shortname. Instead, we use '.' for it.
1577 			 */
1578 			tr->shortname[0] = '.';
1579 			tr->shortname[1] = ' ';
1580 		} else {
1581 			tr->shortname[0] = sched->next_shortname1;
1582 			tr->shortname[1] = sched->next_shortname2;
1583 
1584 			if (sched->next_shortname1 < 'Z') {
1585 				sched->next_shortname1++;
1586 			} else {
1587 				sched->next_shortname1 = 'A';
1588 				if (sched->next_shortname2 < '9')
1589 					sched->next_shortname2++;
1590 				else
1591 					sched->next_shortname2 = '0';
1592 			}
1593 		}
1594 		new_shortname = 1;
1595 	}
1596 
1597 	for (i = 0; i < cpus_nr; i++) {
1598 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1599 		struct thread *curr_thread = sched->curr_thread[cpu];
1600 		struct thread_runtime *curr_tr;
1601 		const char *pid_color = color;
1602 		const char *cpu_color = color;
1603 
1604 		if (curr_thread && thread__has_color(curr_thread))
1605 			pid_color = COLOR_PIDS;
1606 
1607 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1608 			continue;
1609 
1610 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1611 			cpu_color = COLOR_CPUS;
1612 
1613 		if (cpu != this_cpu)
1614 			color_fprintf(stdout, color, " ");
1615 		else
1616 			color_fprintf(stdout, cpu_color, "*");
1617 
1618 		if (sched->curr_thread[cpu]) {
1619 			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1620 			if (curr_tr == NULL) {
1621 				thread__put(sched_in);
1622 				return -1;
1623 			}
1624 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1625 		} else
1626 			color_fprintf(stdout, color, "   ");
1627 	}
1628 
1629 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1630 		goto out;
1631 
1632 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1633 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1634 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1635 		const char *pid_color = color;
1636 
1637 		if (thread__has_color(sched_in))
1638 			pid_color = COLOR_PIDS;
1639 
1640 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1641 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1642 		tr->comm_changed = false;
1643 	}
1644 
1645 	if (sched->map.comp && new_cpu)
1646 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1647 
1648 out:
1649 	color_fprintf(stdout, color, "\n");
1650 
1651 	thread__put(sched_in);
1652 
1653 	return 0;
1654 }
1655 
1656 static int process_sched_switch_event(struct perf_tool *tool,
1657 				      struct perf_evsel *evsel,
1658 				      struct perf_sample *sample,
1659 				      struct machine *machine)
1660 {
1661 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1662 	int this_cpu = sample->cpu, err = 0;
1663 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1664 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1665 
1666 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1667 		/*
1668 		 * Are we trying to switch away a PID that is
1669 		 * not current?
1670 		 */
1671 		if (sched->curr_pid[this_cpu] != prev_pid)
1672 			sched->nr_context_switch_bugs++;
1673 	}
1674 
1675 	if (sched->tp_handler->switch_event)
1676 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1677 
1678 	sched->curr_pid[this_cpu] = next_pid;
1679 	return err;
1680 }
1681 
1682 static int process_sched_runtime_event(struct perf_tool *tool,
1683 				       struct perf_evsel *evsel,
1684 				       struct perf_sample *sample,
1685 				       struct machine *machine)
1686 {
1687 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1688 
1689 	if (sched->tp_handler->runtime_event)
1690 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1691 
1692 	return 0;
1693 }
1694 
1695 static int perf_sched__process_fork_event(struct perf_tool *tool,
1696 					  union perf_event *event,
1697 					  struct perf_sample *sample,
1698 					  struct machine *machine)
1699 {
1700 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1701 
1702 	/* run the fork event through the perf machineruy */
1703 	perf_event__process_fork(tool, event, sample, machine);
1704 
1705 	/* and then run additional processing needed for this command */
1706 	if (sched->tp_handler->fork_event)
1707 		return sched->tp_handler->fork_event(sched, event, machine);
1708 
1709 	return 0;
1710 }
1711 
1712 static int process_sched_migrate_task_event(struct perf_tool *tool,
1713 					    struct perf_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->migrate_task_event)
1720 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1721 
1722 	return 0;
1723 }
1724 
1725 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1726 				  struct perf_evsel *evsel,
1727 				  struct perf_sample *sample,
1728 				  struct machine *machine);
1729 
1730 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1731 						 union perf_event *event __maybe_unused,
1732 						 struct perf_sample *sample,
1733 						 struct perf_evsel *evsel,
1734 						 struct machine *machine)
1735 {
1736 	int err = 0;
1737 
1738 	if (evsel->handler != NULL) {
1739 		tracepoint_handler f = evsel->handler;
1740 		err = f(tool, evsel, sample, machine);
1741 	}
1742 
1743 	return err;
1744 }
1745 
1746 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1747 				    union perf_event *event,
1748 				    struct perf_sample *sample,
1749 				    struct machine *machine)
1750 {
1751 	struct thread *thread;
1752 	struct thread_runtime *tr;
1753 	int err;
1754 
1755 	err = perf_event__process_comm(tool, event, sample, machine);
1756 	if (err)
1757 		return err;
1758 
1759 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1760 	if (!thread) {
1761 		pr_err("Internal error: can't find thread\n");
1762 		return -1;
1763 	}
1764 
1765 	tr = thread__get_runtime(thread);
1766 	if (tr == NULL) {
1767 		thread__put(thread);
1768 		return -1;
1769 	}
1770 
1771 	tr->comm_changed = true;
1772 	thread__put(thread);
1773 
1774 	return 0;
1775 }
1776 
1777 static int perf_sched__read_events(struct perf_sched *sched)
1778 {
1779 	const struct perf_evsel_str_handler handlers[] = {
1780 		{ "sched:sched_switch",	      process_sched_switch_event, },
1781 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1782 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1783 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1784 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1785 	};
1786 	struct perf_session *session;
1787 	struct perf_data data = {
1788 		.path  = input_name,
1789 		.mode  = PERF_DATA_MODE_READ,
1790 		.force = sched->force,
1791 	};
1792 	int rc = -1;
1793 
1794 	session = perf_session__new(&data, false, &sched->tool);
1795 	if (session == NULL) {
1796 		pr_debug("No Memory for session\n");
1797 		return -1;
1798 	}
1799 
1800 	symbol__init(&session->header.env);
1801 
1802 	if (perf_session__set_tracepoints_handlers(session, handlers))
1803 		goto out_delete;
1804 
1805 	if (perf_session__has_traces(session, "record -R")) {
1806 		int err = perf_session__process_events(session);
1807 		if (err) {
1808 			pr_err("Failed to process events, error %d", err);
1809 			goto out_delete;
1810 		}
1811 
1812 		sched->nr_events      = session->evlist->stats.nr_events[0];
1813 		sched->nr_lost_events = session->evlist->stats.total_lost;
1814 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1815 	}
1816 
1817 	rc = 0;
1818 out_delete:
1819 	perf_session__delete(session);
1820 	return rc;
1821 }
1822 
1823 /*
1824  * scheduling times are printed as msec.usec
1825  */
1826 static inline void print_sched_time(unsigned long long nsecs, int width)
1827 {
1828 	unsigned long msecs;
1829 	unsigned long usecs;
1830 
1831 	msecs  = nsecs / NSEC_PER_MSEC;
1832 	nsecs -= msecs * NSEC_PER_MSEC;
1833 	usecs  = nsecs / NSEC_PER_USEC;
1834 	printf("%*lu.%03lu ", width, msecs, usecs);
1835 }
1836 
1837 /*
1838  * returns runtime data for event, allocating memory for it the
1839  * first time it is used.
1840  */
1841 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1842 {
1843 	struct evsel_runtime *r = evsel->priv;
1844 
1845 	if (r == NULL) {
1846 		r = zalloc(sizeof(struct evsel_runtime));
1847 		evsel->priv = r;
1848 	}
1849 
1850 	return r;
1851 }
1852 
1853 /*
1854  * save last time event was seen per cpu
1855  */
1856 static void perf_evsel__save_time(struct perf_evsel *evsel,
1857 				  u64 timestamp, u32 cpu)
1858 {
1859 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1860 
1861 	if (r == NULL)
1862 		return;
1863 
1864 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1865 		int i, n = __roundup_pow_of_two(cpu+1);
1866 		void *p = r->last_time;
1867 
1868 		p = realloc(r->last_time, n * sizeof(u64));
1869 		if (!p)
1870 			return;
1871 
1872 		r->last_time = p;
1873 		for (i = r->ncpu; i < n; ++i)
1874 			r->last_time[i] = (u64) 0;
1875 
1876 		r->ncpu = n;
1877 	}
1878 
1879 	r->last_time[cpu] = timestamp;
1880 }
1881 
1882 /* returns last time this event was seen on the given cpu */
1883 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1884 {
1885 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1886 
1887 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1888 		return 0;
1889 
1890 	return r->last_time[cpu];
1891 }
1892 
1893 static int comm_width = 30;
1894 
1895 static char *timehist_get_commstr(struct thread *thread)
1896 {
1897 	static char str[32];
1898 	const char *comm = thread__comm_str(thread);
1899 	pid_t tid = thread->tid;
1900 	pid_t pid = thread->pid_;
1901 	int n;
1902 
1903 	if (pid == 0)
1904 		n = scnprintf(str, sizeof(str), "%s", comm);
1905 
1906 	else if (tid != pid)
1907 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1908 
1909 	else
1910 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1911 
1912 	if (n > comm_width)
1913 		comm_width = n;
1914 
1915 	return str;
1916 }
1917 
1918 static void timehist_header(struct perf_sched *sched)
1919 {
1920 	u32 ncpus = sched->max_cpu + 1;
1921 	u32 i, j;
1922 
1923 	printf("%15s %6s ", "time", "cpu");
1924 
1925 	if (sched->show_cpu_visual) {
1926 		printf(" ");
1927 		for (i = 0, j = 0; i < ncpus; ++i) {
1928 			printf("%x", j++);
1929 			if (j > 15)
1930 				j = 0;
1931 		}
1932 		printf(" ");
1933 	}
1934 
1935 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1936 		"task name", "wait time", "sch delay", "run time");
1937 
1938 	if (sched->show_state)
1939 		printf("  %s", "state");
1940 
1941 	printf("\n");
1942 
1943 	/*
1944 	 * units row
1945 	 */
1946 	printf("%15s %-6s ", "", "");
1947 
1948 	if (sched->show_cpu_visual)
1949 		printf(" %*s ", ncpus, "");
1950 
1951 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1952 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1953 
1954 	if (sched->show_state)
1955 		printf("  %5s", "");
1956 
1957 	printf("\n");
1958 
1959 	/*
1960 	 * separator
1961 	 */
1962 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1963 
1964 	if (sched->show_cpu_visual)
1965 		printf(" %.*s ", ncpus, graph_dotted_line);
1966 
1967 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1968 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1969 		graph_dotted_line);
1970 
1971 	if (sched->show_state)
1972 		printf("  %.5s", graph_dotted_line);
1973 
1974 	printf("\n");
1975 }
1976 
1977 static char task_state_char(struct thread *thread, int state)
1978 {
1979 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1980 	unsigned bit = state ? ffs(state) : 0;
1981 
1982 	/* 'I' for idle */
1983 	if (thread->tid == 0)
1984 		return 'I';
1985 
1986 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1987 }
1988 
1989 static void timehist_print_sample(struct perf_sched *sched,
1990 				  struct perf_evsel *evsel,
1991 				  struct perf_sample *sample,
1992 				  struct addr_location *al,
1993 				  struct thread *thread,
1994 				  u64 t, int state)
1995 {
1996 	struct thread_runtime *tr = thread__priv(thread);
1997 	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1998 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1999 	u32 max_cpus = sched->max_cpu + 1;
2000 	char tstr[64];
2001 	char nstr[30];
2002 	u64 wait_time;
2003 
2004 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2005 	printf("%15s [%04d] ", tstr, sample->cpu);
2006 
2007 	if (sched->show_cpu_visual) {
2008 		u32 i;
2009 		char c;
2010 
2011 		printf(" ");
2012 		for (i = 0; i < max_cpus; ++i) {
2013 			/* flag idle times with 'i'; others are sched events */
2014 			if (i == sample->cpu)
2015 				c = (thread->tid == 0) ? 'i' : 's';
2016 			else
2017 				c = ' ';
2018 			printf("%c", c);
2019 		}
2020 		printf(" ");
2021 	}
2022 
2023 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2024 
2025 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2026 	print_sched_time(wait_time, 6);
2027 
2028 	print_sched_time(tr->dt_delay, 6);
2029 	print_sched_time(tr->dt_run, 6);
2030 
2031 	if (sched->show_state)
2032 		printf(" %5c ", task_state_char(thread, state));
2033 
2034 	if (sched->show_next) {
2035 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2036 		printf(" %-*s", comm_width, nstr);
2037 	}
2038 
2039 	if (sched->show_wakeups && !sched->show_next)
2040 		printf("  %-*s", comm_width, "");
2041 
2042 	if (thread->tid == 0)
2043 		goto out;
2044 
2045 	if (sched->show_callchain)
2046 		printf("  ");
2047 
2048 	sample__fprintf_sym(sample, al, 0,
2049 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2050 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2051 			    EVSEL__PRINT_SKIP_IGNORED,
2052 			    &callchain_cursor, stdout);
2053 
2054 out:
2055 	printf("\n");
2056 }
2057 
2058 /*
2059  * Explanation of delta-time stats:
2060  *
2061  *            t = time of current schedule out event
2062  *        tprev = time of previous sched out event
2063  *                also time of schedule-in event for current task
2064  *    last_time = time of last sched change event for current task
2065  *                (i.e, time process was last scheduled out)
2066  * ready_to_run = time of wakeup for current task
2067  *
2068  * -----|------------|------------|------------|------
2069  *    last         ready        tprev          t
2070  *    time         to run
2071  *
2072  *      |-------- dt_wait --------|
2073  *                   |- dt_delay -|-- dt_run --|
2074  *
2075  *   dt_run = run time of current task
2076  *  dt_wait = time between last schedule out event for task and tprev
2077  *            represents time spent off the cpu
2078  * dt_delay = time between wakeup and schedule-in of task
2079  */
2080 
2081 static void timehist_update_runtime_stats(struct thread_runtime *r,
2082 					 u64 t, u64 tprev)
2083 {
2084 	r->dt_delay   = 0;
2085 	r->dt_sleep   = 0;
2086 	r->dt_iowait  = 0;
2087 	r->dt_preempt = 0;
2088 	r->dt_run     = 0;
2089 
2090 	if (tprev) {
2091 		r->dt_run = t - tprev;
2092 		if (r->ready_to_run) {
2093 			if (r->ready_to_run > tprev)
2094 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2095 			else
2096 				r->dt_delay = tprev - r->ready_to_run;
2097 		}
2098 
2099 		if (r->last_time > tprev)
2100 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2101 		else if (r->last_time) {
2102 			u64 dt_wait = tprev - r->last_time;
2103 
2104 			if (r->last_state == TASK_RUNNING)
2105 				r->dt_preempt = dt_wait;
2106 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2107 				r->dt_iowait = dt_wait;
2108 			else
2109 				r->dt_sleep = dt_wait;
2110 		}
2111 	}
2112 
2113 	update_stats(&r->run_stats, r->dt_run);
2114 
2115 	r->total_run_time     += r->dt_run;
2116 	r->total_delay_time   += r->dt_delay;
2117 	r->total_sleep_time   += r->dt_sleep;
2118 	r->total_iowait_time  += r->dt_iowait;
2119 	r->total_preempt_time += r->dt_preempt;
2120 }
2121 
2122 static bool is_idle_sample(struct perf_sample *sample,
2123 			   struct perf_evsel *evsel)
2124 {
2125 	/* pid 0 == swapper == idle task */
2126 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2127 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2128 
2129 	return sample->pid == 0;
2130 }
2131 
2132 static void save_task_callchain(struct perf_sched *sched,
2133 				struct perf_sample *sample,
2134 				struct perf_evsel *evsel,
2135 				struct machine *machine)
2136 {
2137 	struct callchain_cursor *cursor = &callchain_cursor;
2138 	struct thread *thread;
2139 
2140 	/* want main thread for process - has maps */
2141 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2142 	if (thread == NULL) {
2143 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2144 		return;
2145 	}
2146 
2147 	if (!sched->show_callchain || sample->callchain == NULL)
2148 		return;
2149 
2150 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2151 				      NULL, NULL, sched->max_stack + 2) != 0) {
2152 		if (verbose > 0)
2153 			pr_err("Failed to resolve callchain. Skipping\n");
2154 
2155 		return;
2156 	}
2157 
2158 	callchain_cursor_commit(cursor);
2159 
2160 	while (true) {
2161 		struct callchain_cursor_node *node;
2162 		struct symbol *sym;
2163 
2164 		node = callchain_cursor_current(cursor);
2165 		if (node == NULL)
2166 			break;
2167 
2168 		sym = node->sym;
2169 		if (sym) {
2170 			if (!strcmp(sym->name, "schedule") ||
2171 			    !strcmp(sym->name, "__schedule") ||
2172 			    !strcmp(sym->name, "preempt_schedule"))
2173 				sym->ignore = 1;
2174 		}
2175 
2176 		callchain_cursor_advance(cursor);
2177 	}
2178 }
2179 
2180 static int init_idle_thread(struct thread *thread)
2181 {
2182 	struct idle_thread_runtime *itr;
2183 
2184 	thread__set_comm(thread, idle_comm, 0);
2185 
2186 	itr = zalloc(sizeof(*itr));
2187 	if (itr == NULL)
2188 		return -ENOMEM;
2189 
2190 	init_stats(&itr->tr.run_stats);
2191 	callchain_init(&itr->callchain);
2192 	callchain_cursor_reset(&itr->cursor);
2193 	thread__set_priv(thread, itr);
2194 
2195 	return 0;
2196 }
2197 
2198 /*
2199  * Track idle stats per cpu by maintaining a local thread
2200  * struct for the idle task on each cpu.
2201  */
2202 static int init_idle_threads(int ncpu)
2203 {
2204 	int i, ret;
2205 
2206 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2207 	if (!idle_threads)
2208 		return -ENOMEM;
2209 
2210 	idle_max_cpu = ncpu;
2211 
2212 	/* allocate the actual thread struct if needed */
2213 	for (i = 0; i < ncpu; ++i) {
2214 		idle_threads[i] = thread__new(0, 0);
2215 		if (idle_threads[i] == NULL)
2216 			return -ENOMEM;
2217 
2218 		ret = init_idle_thread(idle_threads[i]);
2219 		if (ret < 0)
2220 			return ret;
2221 	}
2222 
2223 	return 0;
2224 }
2225 
2226 static void free_idle_threads(void)
2227 {
2228 	int i;
2229 
2230 	if (idle_threads == NULL)
2231 		return;
2232 
2233 	for (i = 0; i < idle_max_cpu; ++i) {
2234 		if ((idle_threads[i]))
2235 			thread__delete(idle_threads[i]);
2236 	}
2237 
2238 	free(idle_threads);
2239 }
2240 
2241 static struct thread *get_idle_thread(int cpu)
2242 {
2243 	/*
2244 	 * expand/allocate array of pointers to local thread
2245 	 * structs if needed
2246 	 */
2247 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2248 		int i, j = __roundup_pow_of_two(cpu+1);
2249 		void *p;
2250 
2251 		p = realloc(idle_threads, j * sizeof(struct thread *));
2252 		if (!p)
2253 			return NULL;
2254 
2255 		idle_threads = (struct thread **) p;
2256 		for (i = idle_max_cpu; i < j; ++i)
2257 			idle_threads[i] = NULL;
2258 
2259 		idle_max_cpu = j;
2260 	}
2261 
2262 	/* allocate a new thread struct if needed */
2263 	if (idle_threads[cpu] == NULL) {
2264 		idle_threads[cpu] = thread__new(0, 0);
2265 
2266 		if (idle_threads[cpu]) {
2267 			if (init_idle_thread(idle_threads[cpu]) < 0)
2268 				return NULL;
2269 		}
2270 	}
2271 
2272 	return idle_threads[cpu];
2273 }
2274 
2275 static void save_idle_callchain(struct perf_sched *sched,
2276 				struct idle_thread_runtime *itr,
2277 				struct perf_sample *sample)
2278 {
2279 	if (!sched->show_callchain || sample->callchain == NULL)
2280 		return;
2281 
2282 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2283 }
2284 
2285 static struct thread *timehist_get_thread(struct perf_sched *sched,
2286 					  struct perf_sample *sample,
2287 					  struct machine *machine,
2288 					  struct perf_evsel *evsel)
2289 {
2290 	struct thread *thread;
2291 
2292 	if (is_idle_sample(sample, evsel)) {
2293 		thread = get_idle_thread(sample->cpu);
2294 		if (thread == NULL)
2295 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2296 
2297 	} else {
2298 		/* there were samples with tid 0 but non-zero pid */
2299 		thread = machine__findnew_thread(machine, sample->pid,
2300 						 sample->tid ?: sample->pid);
2301 		if (thread == NULL) {
2302 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2303 				 sample->tid);
2304 		}
2305 
2306 		save_task_callchain(sched, sample, evsel, machine);
2307 		if (sched->idle_hist) {
2308 			struct thread *idle;
2309 			struct idle_thread_runtime *itr;
2310 
2311 			idle = get_idle_thread(sample->cpu);
2312 			if (idle == NULL) {
2313 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2314 				return NULL;
2315 			}
2316 
2317 			itr = thread__priv(idle);
2318 			if (itr == NULL)
2319 				return NULL;
2320 
2321 			itr->last_thread = thread;
2322 
2323 			/* copy task callchain when entering to idle */
2324 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2325 				save_idle_callchain(sched, itr, sample);
2326 		}
2327 	}
2328 
2329 	return thread;
2330 }
2331 
2332 static bool timehist_skip_sample(struct perf_sched *sched,
2333 				 struct thread *thread,
2334 				 struct perf_evsel *evsel,
2335 				 struct perf_sample *sample)
2336 {
2337 	bool rc = false;
2338 
2339 	if (thread__is_filtered(thread)) {
2340 		rc = true;
2341 		sched->skipped_samples++;
2342 	}
2343 
2344 	if (sched->idle_hist) {
2345 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2346 			rc = true;
2347 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2348 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2349 			rc = true;
2350 	}
2351 
2352 	return rc;
2353 }
2354 
2355 static void timehist_print_wakeup_event(struct perf_sched *sched,
2356 					struct perf_evsel *evsel,
2357 					struct perf_sample *sample,
2358 					struct machine *machine,
2359 					struct thread *awakened)
2360 {
2361 	struct thread *thread;
2362 	char tstr[64];
2363 
2364 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2365 	if (thread == NULL)
2366 		return;
2367 
2368 	/* show wakeup unless both awakee and awaker are filtered */
2369 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2370 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2371 		return;
2372 	}
2373 
2374 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2375 	printf("%15s [%04d] ", tstr, sample->cpu);
2376 	if (sched->show_cpu_visual)
2377 		printf(" %*s ", sched->max_cpu + 1, "");
2378 
2379 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2380 
2381 	/* dt spacer */
2382 	printf("  %9s  %9s  %9s ", "", "", "");
2383 
2384 	printf("awakened: %s", timehist_get_commstr(awakened));
2385 
2386 	printf("\n");
2387 }
2388 
2389 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2390 				       union perf_event *event __maybe_unused,
2391 				       struct perf_evsel *evsel,
2392 				       struct perf_sample *sample,
2393 				       struct machine *machine)
2394 {
2395 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2396 	struct thread *thread;
2397 	struct thread_runtime *tr = NULL;
2398 	/* want pid of awakened task not pid in sample */
2399 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2400 
2401 	thread = machine__findnew_thread(machine, 0, pid);
2402 	if (thread == NULL)
2403 		return -1;
2404 
2405 	tr = thread__get_runtime(thread);
2406 	if (tr == NULL)
2407 		return -1;
2408 
2409 	if (tr->ready_to_run == 0)
2410 		tr->ready_to_run = sample->time;
2411 
2412 	/* show wakeups if requested */
2413 	if (sched->show_wakeups &&
2414 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2415 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2416 
2417 	return 0;
2418 }
2419 
2420 static void timehist_print_migration_event(struct perf_sched *sched,
2421 					struct perf_evsel *evsel,
2422 					struct perf_sample *sample,
2423 					struct machine *machine,
2424 					struct thread *migrated)
2425 {
2426 	struct thread *thread;
2427 	char tstr[64];
2428 	u32 max_cpus = sched->max_cpu + 1;
2429 	u32 ocpu, dcpu;
2430 
2431 	if (sched->summary_only)
2432 		return;
2433 
2434 	max_cpus = sched->max_cpu + 1;
2435 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2436 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2437 
2438 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2439 	if (thread == NULL)
2440 		return;
2441 
2442 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2443 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2444 		return;
2445 	}
2446 
2447 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2448 	printf("%15s [%04d] ", tstr, sample->cpu);
2449 
2450 	if (sched->show_cpu_visual) {
2451 		u32 i;
2452 		char c;
2453 
2454 		printf("  ");
2455 		for (i = 0; i < max_cpus; ++i) {
2456 			c = (i == sample->cpu) ? 'm' : ' ';
2457 			printf("%c", c);
2458 		}
2459 		printf("  ");
2460 	}
2461 
2462 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2463 
2464 	/* dt spacer */
2465 	printf("  %9s  %9s  %9s ", "", "", "");
2466 
2467 	printf("migrated: %s", timehist_get_commstr(migrated));
2468 	printf(" cpu %d => %d", ocpu, dcpu);
2469 
2470 	printf("\n");
2471 }
2472 
2473 static int timehist_migrate_task_event(struct perf_tool *tool,
2474 				       union perf_event *event __maybe_unused,
2475 				       struct perf_evsel *evsel,
2476 				       struct perf_sample *sample,
2477 				       struct machine *machine)
2478 {
2479 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2480 	struct thread *thread;
2481 	struct thread_runtime *tr = NULL;
2482 	/* want pid of migrated task not pid in sample */
2483 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2484 
2485 	thread = machine__findnew_thread(machine, 0, pid);
2486 	if (thread == NULL)
2487 		return -1;
2488 
2489 	tr = thread__get_runtime(thread);
2490 	if (tr == NULL)
2491 		return -1;
2492 
2493 	tr->migrations++;
2494 
2495 	/* show migrations if requested */
2496 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2497 
2498 	return 0;
2499 }
2500 
2501 static int timehist_sched_change_event(struct perf_tool *tool,
2502 				       union perf_event *event,
2503 				       struct perf_evsel *evsel,
2504 				       struct perf_sample *sample,
2505 				       struct machine *machine)
2506 {
2507 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2508 	struct perf_time_interval *ptime = &sched->ptime;
2509 	struct addr_location al;
2510 	struct thread *thread;
2511 	struct thread_runtime *tr = NULL;
2512 	u64 tprev, t = sample->time;
2513 	int rc = 0;
2514 	int state = perf_evsel__intval(evsel, sample, "prev_state");
2515 
2516 
2517 	if (machine__resolve(machine, &al, sample) < 0) {
2518 		pr_err("problem processing %d event. skipping it\n",
2519 		       event->header.type);
2520 		rc = -1;
2521 		goto out;
2522 	}
2523 
2524 	thread = timehist_get_thread(sched, sample, machine, evsel);
2525 	if (thread == NULL) {
2526 		rc = -1;
2527 		goto out;
2528 	}
2529 
2530 	if (timehist_skip_sample(sched, thread, evsel, sample))
2531 		goto out;
2532 
2533 	tr = thread__get_runtime(thread);
2534 	if (tr == NULL) {
2535 		rc = -1;
2536 		goto out;
2537 	}
2538 
2539 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2540 
2541 	/*
2542 	 * If start time given:
2543 	 * - sample time is under window user cares about - skip sample
2544 	 * - tprev is under window user cares about  - reset to start of window
2545 	 */
2546 	if (ptime->start && ptime->start > t)
2547 		goto out;
2548 
2549 	if (tprev && ptime->start > tprev)
2550 		tprev = ptime->start;
2551 
2552 	/*
2553 	 * If end time given:
2554 	 * - previous sched event is out of window - we are done
2555 	 * - sample time is beyond window user cares about - reset it
2556 	 *   to close out stats for time window interest
2557 	 */
2558 	if (ptime->end) {
2559 		if (tprev > ptime->end)
2560 			goto out;
2561 
2562 		if (t > ptime->end)
2563 			t = ptime->end;
2564 	}
2565 
2566 	if (!sched->idle_hist || thread->tid == 0) {
2567 		timehist_update_runtime_stats(tr, t, tprev);
2568 
2569 		if (sched->idle_hist) {
2570 			struct idle_thread_runtime *itr = (void *)tr;
2571 			struct thread_runtime *last_tr;
2572 
2573 			BUG_ON(thread->tid != 0);
2574 
2575 			if (itr->last_thread == NULL)
2576 				goto out;
2577 
2578 			/* add current idle time as last thread's runtime */
2579 			last_tr = thread__get_runtime(itr->last_thread);
2580 			if (last_tr == NULL)
2581 				goto out;
2582 
2583 			timehist_update_runtime_stats(last_tr, t, tprev);
2584 			/*
2585 			 * remove delta time of last thread as it's not updated
2586 			 * and otherwise it will show an invalid value next
2587 			 * time.  we only care total run time and run stat.
2588 			 */
2589 			last_tr->dt_run = 0;
2590 			last_tr->dt_delay = 0;
2591 			last_tr->dt_sleep = 0;
2592 			last_tr->dt_iowait = 0;
2593 			last_tr->dt_preempt = 0;
2594 
2595 			if (itr->cursor.nr)
2596 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2597 
2598 			itr->last_thread = NULL;
2599 		}
2600 	}
2601 
2602 	if (!sched->summary_only)
2603 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2604 
2605 out:
2606 	if (sched->hist_time.start == 0 && t >= ptime->start)
2607 		sched->hist_time.start = t;
2608 	if (ptime->end == 0 || t <= ptime->end)
2609 		sched->hist_time.end = t;
2610 
2611 	if (tr) {
2612 		/* time of this sched_switch event becomes last time task seen */
2613 		tr->last_time = sample->time;
2614 
2615 		/* last state is used to determine where to account wait time */
2616 		tr->last_state = state;
2617 
2618 		/* sched out event for task so reset ready to run time */
2619 		tr->ready_to_run = 0;
2620 	}
2621 
2622 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2623 
2624 	return rc;
2625 }
2626 
2627 static int timehist_sched_switch_event(struct perf_tool *tool,
2628 			     union perf_event *event,
2629 			     struct perf_evsel *evsel,
2630 			     struct perf_sample *sample,
2631 			     struct machine *machine __maybe_unused)
2632 {
2633 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2634 }
2635 
2636 static int process_lost(struct perf_tool *tool __maybe_unused,
2637 			union perf_event *event,
2638 			struct perf_sample *sample,
2639 			struct machine *machine __maybe_unused)
2640 {
2641 	char tstr[64];
2642 
2643 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2644 	printf("%15s ", tstr);
2645 	printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2646 
2647 	return 0;
2648 }
2649 
2650 
2651 static void print_thread_runtime(struct thread *t,
2652 				 struct thread_runtime *r)
2653 {
2654 	double mean = avg_stats(&r->run_stats);
2655 	float stddev;
2656 
2657 	printf("%*s   %5d  %9" PRIu64 " ",
2658 	       comm_width, timehist_get_commstr(t), t->ppid,
2659 	       (u64) r->run_stats.n);
2660 
2661 	print_sched_time(r->total_run_time, 8);
2662 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2663 	print_sched_time(r->run_stats.min, 6);
2664 	printf(" ");
2665 	print_sched_time((u64) mean, 6);
2666 	printf(" ");
2667 	print_sched_time(r->run_stats.max, 6);
2668 	printf("  ");
2669 	printf("%5.2f", stddev);
2670 	printf("   %5" PRIu64, r->migrations);
2671 	printf("\n");
2672 }
2673 
2674 static void print_thread_waittime(struct thread *t,
2675 				  struct thread_runtime *r)
2676 {
2677 	printf("%*s   %5d  %9" PRIu64 " ",
2678 	       comm_width, timehist_get_commstr(t), t->ppid,
2679 	       (u64) r->run_stats.n);
2680 
2681 	print_sched_time(r->total_run_time, 8);
2682 	print_sched_time(r->total_sleep_time, 6);
2683 	printf(" ");
2684 	print_sched_time(r->total_iowait_time, 6);
2685 	printf(" ");
2686 	print_sched_time(r->total_preempt_time, 6);
2687 	printf(" ");
2688 	print_sched_time(r->total_delay_time, 6);
2689 	printf("\n");
2690 }
2691 
2692 struct total_run_stats {
2693 	struct perf_sched *sched;
2694 	u64  sched_count;
2695 	u64  task_count;
2696 	u64  total_run_time;
2697 };
2698 
2699 static int __show_thread_runtime(struct thread *t, void *priv)
2700 {
2701 	struct total_run_stats *stats = priv;
2702 	struct thread_runtime *r;
2703 
2704 	if (thread__is_filtered(t))
2705 		return 0;
2706 
2707 	r = thread__priv(t);
2708 	if (r && r->run_stats.n) {
2709 		stats->task_count++;
2710 		stats->sched_count += r->run_stats.n;
2711 		stats->total_run_time += r->total_run_time;
2712 
2713 		if (stats->sched->show_state)
2714 			print_thread_waittime(t, r);
2715 		else
2716 			print_thread_runtime(t, r);
2717 	}
2718 
2719 	return 0;
2720 }
2721 
2722 static int show_thread_runtime(struct thread *t, void *priv)
2723 {
2724 	if (t->dead)
2725 		return 0;
2726 
2727 	return __show_thread_runtime(t, priv);
2728 }
2729 
2730 static int show_deadthread_runtime(struct thread *t, void *priv)
2731 {
2732 	if (!t->dead)
2733 		return 0;
2734 
2735 	return __show_thread_runtime(t, priv);
2736 }
2737 
2738 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2739 {
2740 	const char *sep = " <- ";
2741 	struct callchain_list *chain;
2742 	size_t ret = 0;
2743 	char bf[1024];
2744 	bool first;
2745 
2746 	if (node == NULL)
2747 		return 0;
2748 
2749 	ret = callchain__fprintf_folded(fp, node->parent);
2750 	first = (ret == 0);
2751 
2752 	list_for_each_entry(chain, &node->val, list) {
2753 		if (chain->ip >= PERF_CONTEXT_MAX)
2754 			continue;
2755 		if (chain->ms.sym && chain->ms.sym->ignore)
2756 			continue;
2757 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2758 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2759 							false));
2760 		first = false;
2761 	}
2762 
2763 	return ret;
2764 }
2765 
2766 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2767 {
2768 	size_t ret = 0;
2769 	FILE *fp = stdout;
2770 	struct callchain_node *chain;
2771 	struct rb_node *rb_node = rb_first_cached(root);
2772 
2773 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2774 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2775 	       graph_dotted_line);
2776 
2777 	while (rb_node) {
2778 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2779 		rb_node = rb_next(rb_node);
2780 
2781 		ret += fprintf(fp, "  ");
2782 		print_sched_time(chain->hit, 12);
2783 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2784 		ret += fprintf(fp, " %8d  ", chain->count);
2785 		ret += callchain__fprintf_folded(fp, chain);
2786 		ret += fprintf(fp, "\n");
2787 	}
2788 
2789 	return ret;
2790 }
2791 
2792 static void timehist_print_summary(struct perf_sched *sched,
2793 				   struct perf_session *session)
2794 {
2795 	struct machine *m = &session->machines.host;
2796 	struct total_run_stats totals;
2797 	u64 task_count;
2798 	struct thread *t;
2799 	struct thread_runtime *r;
2800 	int i;
2801 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2802 
2803 	memset(&totals, 0, sizeof(totals));
2804 	totals.sched = sched;
2805 
2806 	if (sched->idle_hist) {
2807 		printf("\nIdle-time summary\n");
2808 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2809 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2810 	} else if (sched->show_state) {
2811 		printf("\nWait-time summary\n");
2812 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2813 		printf("   run-time      sleep      iowait     preempt       delay\n");
2814 	} else {
2815 		printf("\nRuntime summary\n");
2816 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2817 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2818 	}
2819 	printf("%*s            (count)  ", comm_width, "");
2820 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2821 	       sched->show_state ? "(msec)" : "%");
2822 	printf("%.117s\n", graph_dotted_line);
2823 
2824 	machine__for_each_thread(m, show_thread_runtime, &totals);
2825 	task_count = totals.task_count;
2826 	if (!task_count)
2827 		printf("<no still running tasks>\n");
2828 
2829 	printf("\nTerminated tasks:\n");
2830 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2831 	if (task_count == totals.task_count)
2832 		printf("<no terminated tasks>\n");
2833 
2834 	/* CPU idle stats not tracked when samples were skipped */
2835 	if (sched->skipped_samples && !sched->idle_hist)
2836 		return;
2837 
2838 	printf("\nIdle stats:\n");
2839 	for (i = 0; i < idle_max_cpu; ++i) {
2840 		t = idle_threads[i];
2841 		if (!t)
2842 			continue;
2843 
2844 		r = thread__priv(t);
2845 		if (r && r->run_stats.n) {
2846 			totals.sched_count += r->run_stats.n;
2847 			printf("    CPU %2d idle for ", i);
2848 			print_sched_time(r->total_run_time, 6);
2849 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2850 		} else
2851 			printf("    CPU %2d idle entire time window\n", i);
2852 	}
2853 
2854 	if (sched->idle_hist && sched->show_callchain) {
2855 		callchain_param.mode  = CHAIN_FOLDED;
2856 		callchain_param.value = CCVAL_PERIOD;
2857 
2858 		callchain_register_param(&callchain_param);
2859 
2860 		printf("\nIdle stats by callchain:\n");
2861 		for (i = 0; i < idle_max_cpu; ++i) {
2862 			struct idle_thread_runtime *itr;
2863 
2864 			t = idle_threads[i];
2865 			if (!t)
2866 				continue;
2867 
2868 			itr = thread__priv(t);
2869 			if (itr == NULL)
2870 				continue;
2871 
2872 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2873 					     0, &callchain_param);
2874 
2875 			printf("  CPU %2d:", i);
2876 			print_sched_time(itr->tr.total_run_time, 6);
2877 			printf(" msec\n");
2878 			timehist_print_idlehist_callchain(&itr->sorted_root);
2879 			printf("\n");
2880 		}
2881 	}
2882 
2883 	printf("\n"
2884 	       "    Total number of unique tasks: %" PRIu64 "\n"
2885 	       "Total number of context switches: %" PRIu64 "\n",
2886 	       totals.task_count, totals.sched_count);
2887 
2888 	printf("           Total run time (msec): ");
2889 	print_sched_time(totals.total_run_time, 2);
2890 	printf("\n");
2891 
2892 	printf("    Total scheduling time (msec): ");
2893 	print_sched_time(hist_time, 2);
2894 	printf(" (x %d)\n", sched->max_cpu);
2895 }
2896 
2897 typedef int (*sched_handler)(struct perf_tool *tool,
2898 			  union perf_event *event,
2899 			  struct perf_evsel *evsel,
2900 			  struct perf_sample *sample,
2901 			  struct machine *machine);
2902 
2903 static int perf_timehist__process_sample(struct perf_tool *tool,
2904 					 union perf_event *event,
2905 					 struct perf_sample *sample,
2906 					 struct perf_evsel *evsel,
2907 					 struct machine *machine)
2908 {
2909 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2910 	int err = 0;
2911 	int this_cpu = sample->cpu;
2912 
2913 	if (this_cpu > sched->max_cpu)
2914 		sched->max_cpu = this_cpu;
2915 
2916 	if (evsel->handler != NULL) {
2917 		sched_handler f = evsel->handler;
2918 
2919 		err = f(tool, event, evsel, sample, machine);
2920 	}
2921 
2922 	return err;
2923 }
2924 
2925 static int timehist_check_attr(struct perf_sched *sched,
2926 			       struct perf_evlist *evlist)
2927 {
2928 	struct perf_evsel *evsel;
2929 	struct evsel_runtime *er;
2930 
2931 	list_for_each_entry(evsel, &evlist->entries, node) {
2932 		er = perf_evsel__get_runtime(evsel);
2933 		if (er == NULL) {
2934 			pr_err("Failed to allocate memory for evsel runtime data\n");
2935 			return -1;
2936 		}
2937 
2938 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2939 			pr_info("Samples do not have callchains.\n");
2940 			sched->show_callchain = 0;
2941 			symbol_conf.use_callchain = 0;
2942 		}
2943 	}
2944 
2945 	return 0;
2946 }
2947 
2948 static int perf_sched__timehist(struct perf_sched *sched)
2949 {
2950 	const struct perf_evsel_str_handler handlers[] = {
2951 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2952 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2953 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2954 	};
2955 	const struct perf_evsel_str_handler migrate_handlers[] = {
2956 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2957 	};
2958 	struct perf_data data = {
2959 		.path  = input_name,
2960 		.mode  = PERF_DATA_MODE_READ,
2961 		.force = sched->force,
2962 	};
2963 
2964 	struct perf_session *session;
2965 	struct perf_evlist *evlist;
2966 	int err = -1;
2967 
2968 	/*
2969 	 * event handlers for timehist option
2970 	 */
2971 	sched->tool.sample	 = perf_timehist__process_sample;
2972 	sched->tool.mmap	 = perf_event__process_mmap;
2973 	sched->tool.comm	 = perf_event__process_comm;
2974 	sched->tool.exit	 = perf_event__process_exit;
2975 	sched->tool.fork	 = perf_event__process_fork;
2976 	sched->tool.lost	 = process_lost;
2977 	sched->tool.attr	 = perf_event__process_attr;
2978 	sched->tool.tracing_data = perf_event__process_tracing_data;
2979 	sched->tool.build_id	 = perf_event__process_build_id;
2980 
2981 	sched->tool.ordered_events = true;
2982 	sched->tool.ordering_requires_timestamps = true;
2983 
2984 	symbol_conf.use_callchain = sched->show_callchain;
2985 
2986 	session = perf_session__new(&data, false, &sched->tool);
2987 	if (session == NULL)
2988 		return -ENOMEM;
2989 
2990 	evlist = session->evlist;
2991 
2992 	symbol__init(&session->header.env);
2993 
2994 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2995 		pr_err("Invalid time string\n");
2996 		return -EINVAL;
2997 	}
2998 
2999 	if (timehist_check_attr(sched, evlist) != 0)
3000 		goto out;
3001 
3002 	setup_pager();
3003 
3004 	/* setup per-evsel handlers */
3005 	if (perf_session__set_tracepoints_handlers(session, handlers))
3006 		goto out;
3007 
3008 	/* sched_switch event at a minimum needs to exist */
3009 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3010 						  "sched:sched_switch")) {
3011 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3012 		goto out;
3013 	}
3014 
3015 	if (sched->show_migrations &&
3016 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3017 		goto out;
3018 
3019 	/* pre-allocate struct for per-CPU idle stats */
3020 	sched->max_cpu = session->header.env.nr_cpus_online;
3021 	if (sched->max_cpu == 0)
3022 		sched->max_cpu = 4;
3023 	if (init_idle_threads(sched->max_cpu))
3024 		goto out;
3025 
3026 	/* summary_only implies summary option, but don't overwrite summary if set */
3027 	if (sched->summary_only)
3028 		sched->summary = sched->summary_only;
3029 
3030 	if (!sched->summary_only)
3031 		timehist_header(sched);
3032 
3033 	err = perf_session__process_events(session);
3034 	if (err) {
3035 		pr_err("Failed to process events, error %d", err);
3036 		goto out;
3037 	}
3038 
3039 	sched->nr_events      = evlist->stats.nr_events[0];
3040 	sched->nr_lost_events = evlist->stats.total_lost;
3041 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3042 
3043 	if (sched->summary)
3044 		timehist_print_summary(sched, session);
3045 
3046 out:
3047 	free_idle_threads();
3048 	perf_session__delete(session);
3049 
3050 	return err;
3051 }
3052 
3053 
3054 static void print_bad_events(struct perf_sched *sched)
3055 {
3056 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3057 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3058 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3059 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3060 	}
3061 	if (sched->nr_lost_events && sched->nr_events) {
3062 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3063 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3064 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3065 	}
3066 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3067 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3068 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3069 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3070 		if (sched->nr_lost_events)
3071 			printf(" (due to lost events?)");
3072 		printf("\n");
3073 	}
3074 }
3075 
3076 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3077 {
3078 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3079 	struct work_atoms *this;
3080 	const char *comm = thread__comm_str(data->thread), *this_comm;
3081 	bool leftmost = true;
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 			leftmost = false;
3096 		} else {
3097 			this->num_merged++;
3098 			this->total_runtime += data->total_runtime;
3099 			this->nb_atoms += data->nb_atoms;
3100 			this->total_lat += data->total_lat;
3101 			list_splice(&data->work_list, &this->work_list);
3102 			if (this->max_lat < data->max_lat) {
3103 				this->max_lat = data->max_lat;
3104 				this->max_lat_at = data->max_lat_at;
3105 			}
3106 			zfree(&data);
3107 			return;
3108 		}
3109 	}
3110 
3111 	data->num_merged++;
3112 	rb_link_node(&data->node, parent, new);
3113 	rb_insert_color_cached(&data->node, root, leftmost);
3114 }
3115 
3116 static void perf_sched__merge_lat(struct perf_sched *sched)
3117 {
3118 	struct work_atoms *data;
3119 	struct rb_node *node;
3120 
3121 	if (sched->skip_merge)
3122 		return;
3123 
3124 	while ((node = rb_first_cached(&sched->atom_root))) {
3125 		rb_erase_cached(node, &sched->atom_root);
3126 		data = rb_entry(node, struct work_atoms, node);
3127 		__merge_work_atoms(&sched->merged_atom_root, data);
3128 	}
3129 }
3130 
3131 static int perf_sched__lat(struct perf_sched *sched)
3132 {
3133 	struct rb_node *next;
3134 
3135 	setup_pager();
3136 
3137 	if (perf_sched__read_events(sched))
3138 		return -1;
3139 
3140 	perf_sched__merge_lat(sched);
3141 	perf_sched__sort_lat(sched);
3142 
3143 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3144 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3145 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3146 
3147 	next = rb_first_cached(&sched->sorted_atom_root);
3148 
3149 	while (next) {
3150 		struct work_atoms *work_list;
3151 
3152 		work_list = rb_entry(next, struct work_atoms, node);
3153 		output_lat_thread(sched, work_list);
3154 		next = rb_next(next);
3155 		thread__zput(work_list->thread);
3156 	}
3157 
3158 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3159 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3160 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3161 
3162 	printf(" ---------------------------------------------------\n");
3163 
3164 	print_bad_events(sched);
3165 	printf("\n");
3166 
3167 	return 0;
3168 }
3169 
3170 static int setup_map_cpus(struct perf_sched *sched)
3171 {
3172 	struct cpu_map *map;
3173 
3174 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3175 
3176 	if (sched->map.comp) {
3177 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3178 		if (!sched->map.comp_cpus)
3179 			return -1;
3180 	}
3181 
3182 	if (!sched->map.cpus_str)
3183 		return 0;
3184 
3185 	map = cpu_map__new(sched->map.cpus_str);
3186 	if (!map) {
3187 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3188 		return -1;
3189 	}
3190 
3191 	sched->map.cpus = map;
3192 	return 0;
3193 }
3194 
3195 static int setup_color_pids(struct perf_sched *sched)
3196 {
3197 	struct thread_map *map;
3198 
3199 	if (!sched->map.color_pids_str)
3200 		return 0;
3201 
3202 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3203 	if (!map) {
3204 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3205 		return -1;
3206 	}
3207 
3208 	sched->map.color_pids = map;
3209 	return 0;
3210 }
3211 
3212 static int setup_color_cpus(struct perf_sched *sched)
3213 {
3214 	struct cpu_map *map;
3215 
3216 	if (!sched->map.color_cpus_str)
3217 		return 0;
3218 
3219 	map = cpu_map__new(sched->map.color_cpus_str);
3220 	if (!map) {
3221 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3222 		return -1;
3223 	}
3224 
3225 	sched->map.color_cpus = map;
3226 	return 0;
3227 }
3228 
3229 static int perf_sched__map(struct perf_sched *sched)
3230 {
3231 	if (setup_map_cpus(sched))
3232 		return -1;
3233 
3234 	if (setup_color_pids(sched))
3235 		return -1;
3236 
3237 	if (setup_color_cpus(sched))
3238 		return -1;
3239 
3240 	setup_pager();
3241 	if (perf_sched__read_events(sched))
3242 		return -1;
3243 	print_bad_events(sched);
3244 	return 0;
3245 }
3246 
3247 static int perf_sched__replay(struct perf_sched *sched)
3248 {
3249 	unsigned long i;
3250 
3251 	calibrate_run_measurement_overhead(sched);
3252 	calibrate_sleep_measurement_overhead(sched);
3253 
3254 	test_calibrations(sched);
3255 
3256 	if (perf_sched__read_events(sched))
3257 		return -1;
3258 
3259 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3260 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3261 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3262 
3263 	if (sched->targetless_wakeups)
3264 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3265 	if (sched->multitarget_wakeups)
3266 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3267 	if (sched->nr_run_events_optimized)
3268 		printf("run atoms optimized: %ld\n",
3269 			sched->nr_run_events_optimized);
3270 
3271 	print_task_traces(sched);
3272 	add_cross_task_wakeups(sched);
3273 
3274 	create_tasks(sched);
3275 	printf("------------------------------------------------------------\n");
3276 	for (i = 0; i < sched->replay_repeat; i++)
3277 		run_one_test(sched);
3278 
3279 	return 0;
3280 }
3281 
3282 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3283 			  const char * const usage_msg[])
3284 {
3285 	char *tmp, *tok, *str = strdup(sched->sort_order);
3286 
3287 	for (tok = strtok_r(str, ", ", &tmp);
3288 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3289 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3290 			usage_with_options_msg(usage_msg, options,
3291 					"Unknown --sort key: `%s'", tok);
3292 		}
3293 	}
3294 
3295 	free(str);
3296 
3297 	sort_dimension__add("pid", &sched->cmp_pid);
3298 }
3299 
3300 static int __cmd_record(int argc, const char **argv)
3301 {
3302 	unsigned int rec_argc, i, j;
3303 	const char **rec_argv;
3304 	const char * const record_args[] = {
3305 		"record",
3306 		"-a",
3307 		"-R",
3308 		"-m", "1024",
3309 		"-c", "1",
3310 		"-e", "sched:sched_switch",
3311 		"-e", "sched:sched_stat_wait",
3312 		"-e", "sched:sched_stat_sleep",
3313 		"-e", "sched:sched_stat_iowait",
3314 		"-e", "sched:sched_stat_runtime",
3315 		"-e", "sched:sched_process_fork",
3316 		"-e", "sched:sched_wakeup",
3317 		"-e", "sched:sched_wakeup_new",
3318 		"-e", "sched:sched_migrate_task",
3319 	};
3320 
3321 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3322 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3323 
3324 	if (rec_argv == NULL)
3325 		return -ENOMEM;
3326 
3327 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3328 		rec_argv[i] = strdup(record_args[i]);
3329 
3330 	for (j = 1; j < (unsigned int)argc; j++, i++)
3331 		rec_argv[i] = argv[j];
3332 
3333 	BUG_ON(i != rec_argc);
3334 
3335 	return cmd_record(i, rec_argv);
3336 }
3337 
3338 int cmd_sched(int argc, const char **argv)
3339 {
3340 	static const char default_sort_order[] = "avg, max, switch, runtime";
3341 	struct perf_sched sched = {
3342 		.tool = {
3343 			.sample		 = perf_sched__process_tracepoint_sample,
3344 			.comm		 = perf_sched__process_comm,
3345 			.namespaces	 = perf_event__process_namespaces,
3346 			.lost		 = perf_event__process_lost,
3347 			.fork		 = perf_sched__process_fork_event,
3348 			.ordered_events = true,
3349 		},
3350 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3351 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3352 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3353 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3354 		.sort_order	      = default_sort_order,
3355 		.replay_repeat	      = 10,
3356 		.profile_cpu	      = -1,
3357 		.next_shortname1      = 'A',
3358 		.next_shortname2      = '0',
3359 		.skip_merge           = 0,
3360 		.show_callchain	      = 1,
3361 		.max_stack            = 5,
3362 	};
3363 	const struct option sched_options[] = {
3364 	OPT_STRING('i', "input", &input_name, "file",
3365 		    "input file name"),
3366 	OPT_INCR('v', "verbose", &verbose,
3367 		    "be more verbose (show symbol address, etc)"),
3368 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3369 		    "dump raw trace in ASCII"),
3370 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3371 	OPT_END()
3372 	};
3373 	const struct option latency_options[] = {
3374 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3375 		   "sort by key(s): runtime, switch, avg, max"),
3376 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3377 		    "CPU to profile on"),
3378 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3379 		    "latency stats per pid instead of per comm"),
3380 	OPT_PARENT(sched_options)
3381 	};
3382 	const struct option replay_options[] = {
3383 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3384 		     "repeat the workload replay N times (-1: infinite)"),
3385 	OPT_PARENT(sched_options)
3386 	};
3387 	const struct option map_options[] = {
3388 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3389 		    "map output in compact mode"),
3390 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3391 		   "highlight given pids in map"),
3392 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3393                     "highlight given CPUs in map"),
3394 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3395                     "display given CPUs in map"),
3396 	OPT_PARENT(sched_options)
3397 	};
3398 	const struct option timehist_options[] = {
3399 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3400 		   "file", "vmlinux pathname"),
3401 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3402 		   "file", "kallsyms pathname"),
3403 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3404 		    "Display call chains if present (default on)"),
3405 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3406 		   "Maximum number of functions to display backtrace."),
3407 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3408 		    "Look for files with symbols relative to this directory"),
3409 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3410 		    "Show only syscall summary with statistics"),
3411 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3412 		    "Show all syscalls and summary with statistics"),
3413 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3414 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3415 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3416 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3417 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3418 	OPT_STRING(0, "time", &sched.time_str, "str",
3419 		   "Time span for analysis (start,stop)"),
3420 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3421 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3422 		   "analyze events only for given process id(s)"),
3423 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3424 		   "analyze events only for given thread id(s)"),
3425 	OPT_PARENT(sched_options)
3426 	};
3427 
3428 	const char * const latency_usage[] = {
3429 		"perf sched latency [<options>]",
3430 		NULL
3431 	};
3432 	const char * const replay_usage[] = {
3433 		"perf sched replay [<options>]",
3434 		NULL
3435 	};
3436 	const char * const map_usage[] = {
3437 		"perf sched map [<options>]",
3438 		NULL
3439 	};
3440 	const char * const timehist_usage[] = {
3441 		"perf sched timehist [<options>]",
3442 		NULL
3443 	};
3444 	const char *const sched_subcommands[] = { "record", "latency", "map",
3445 						  "replay", "script",
3446 						  "timehist", NULL };
3447 	const char *sched_usage[] = {
3448 		NULL,
3449 		NULL
3450 	};
3451 	struct trace_sched_handler lat_ops  = {
3452 		.wakeup_event	    = latency_wakeup_event,
3453 		.switch_event	    = latency_switch_event,
3454 		.runtime_event	    = latency_runtime_event,
3455 		.migrate_task_event = latency_migrate_task_event,
3456 	};
3457 	struct trace_sched_handler map_ops  = {
3458 		.switch_event	    = map_switch_event,
3459 	};
3460 	struct trace_sched_handler replay_ops  = {
3461 		.wakeup_event	    = replay_wakeup_event,
3462 		.switch_event	    = replay_switch_event,
3463 		.fork_event	    = replay_fork_event,
3464 	};
3465 	unsigned int i;
3466 
3467 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3468 		sched.curr_pid[i] = -1;
3469 
3470 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3471 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3472 	if (!argc)
3473 		usage_with_options(sched_usage, sched_options);
3474 
3475 	/*
3476 	 * Aliased to 'perf script' for now:
3477 	 */
3478 	if (!strcmp(argv[0], "script"))
3479 		return cmd_script(argc, argv);
3480 
3481 	if (!strncmp(argv[0], "rec", 3)) {
3482 		return __cmd_record(argc, argv);
3483 	} else if (!strncmp(argv[0], "lat", 3)) {
3484 		sched.tp_handler = &lat_ops;
3485 		if (argc > 1) {
3486 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3487 			if (argc)
3488 				usage_with_options(latency_usage, latency_options);
3489 		}
3490 		setup_sorting(&sched, latency_options, latency_usage);
3491 		return perf_sched__lat(&sched);
3492 	} else if (!strcmp(argv[0], "map")) {
3493 		if (argc) {
3494 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3495 			if (argc)
3496 				usage_with_options(map_usage, map_options);
3497 		}
3498 		sched.tp_handler = &map_ops;
3499 		setup_sorting(&sched, latency_options, latency_usage);
3500 		return perf_sched__map(&sched);
3501 	} else if (!strncmp(argv[0], "rep", 3)) {
3502 		sched.tp_handler = &replay_ops;
3503 		if (argc) {
3504 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3505 			if (argc)
3506 				usage_with_options(replay_usage, replay_options);
3507 		}
3508 		return perf_sched__replay(&sched);
3509 	} else if (!strcmp(argv[0], "timehist")) {
3510 		if (argc) {
3511 			argc = parse_options(argc, argv, timehist_options,
3512 					     timehist_usage, 0);
3513 			if (argc)
3514 				usage_with_options(timehist_usage, timehist_options);
3515 		}
3516 		if ((sched.show_wakeups || sched.show_next) &&
3517 		    sched.summary_only) {
3518 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3519 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3520 			if (sched.show_wakeups)
3521 				parse_options_usage(NULL, timehist_options, "w", true);
3522 			if (sched.show_next)
3523 				parse_options_usage(NULL, timehist_options, "n", true);
3524 			return -EINVAL;
3525 		}
3526 
3527 		return perf_sched__timehist(&sched);
3528 	} else {
3529 		usage_with_options(sched_usage, sched_options);
3530 	}
3531 
3532 	return 0;
3533 }
3534