xref: /openbmc/linux/tools/perf/builtin-sched.c (revision 08283d30)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 #include "perf-sys.h"
5 
6 #include "util/cpumap.h"
7 #include "util/evlist.h"
8 #include "util/evsel.h"
9 #include "util/evsel_fprintf.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
22 
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
26 
27 #include "util/debug.h"
28 #include "util/event.h"
29 
30 #include <linux/kernel.h>
31 #include <linux/log2.h>
32 #include <linux/zalloc.h>
33 #include <sys/prctl.h>
34 #include <sys/resource.h>
35 #include <inttypes.h>
36 
37 #include <errno.h>
38 #include <semaphore.h>
39 #include <pthread.h>
40 #include <math.h>
41 #include <api/fs/fs.h>
42 #include <perf/cpumap.h>
43 #include <linux/time64.h>
44 #include <linux/err.h>
45 
46 #include <linux/ctype.h>
47 
48 #define PR_SET_NAME		15               /* Set process name */
49 #define MAX_CPUS		4096
50 #define COMM_LEN		20
51 #define SYM_LEN			129
52 #define MAX_PID			1024000
53 
54 struct sched_atom;
55 
56 struct task_desc {
57 	unsigned long		nr;
58 	unsigned long		pid;
59 	char			comm[COMM_LEN];
60 
61 	unsigned long		nr_events;
62 	unsigned long		curr_event;
63 	struct sched_atom	**atoms;
64 
65 	pthread_t		thread;
66 	sem_t			sleep_sem;
67 
68 	sem_t			ready_for_work;
69 	sem_t			work_done_sem;
70 
71 	u64			cpu_usage;
72 };
73 
74 enum sched_event_type {
75 	SCHED_EVENT_RUN,
76 	SCHED_EVENT_SLEEP,
77 	SCHED_EVENT_WAKEUP,
78 	SCHED_EVENT_MIGRATION,
79 };
80 
81 struct sched_atom {
82 	enum sched_event_type	type;
83 	int			specific_wait;
84 	u64			timestamp;
85 	u64			duration;
86 	unsigned long		nr;
87 	sem_t			*wait_sem;
88 	struct task_desc	*wakee;
89 };
90 
91 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
92 
93 /* task state bitmask, copied from include/linux/sched.h */
94 #define TASK_RUNNING		0
95 #define TASK_INTERRUPTIBLE	1
96 #define TASK_UNINTERRUPTIBLE	2
97 #define __TASK_STOPPED		4
98 #define __TASK_TRACED		8
99 /* in tsk->exit_state */
100 #define EXIT_DEAD		16
101 #define EXIT_ZOMBIE		32
102 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
103 /* in tsk->state again */
104 #define TASK_DEAD		64
105 #define TASK_WAKEKILL		128
106 #define TASK_WAKING		256
107 #define TASK_PARKED		512
108 
109 enum thread_state {
110 	THREAD_SLEEPING = 0,
111 	THREAD_WAIT_CPU,
112 	THREAD_SCHED_IN,
113 	THREAD_IGNORE
114 };
115 
116 struct work_atom {
117 	struct list_head	list;
118 	enum thread_state	state;
119 	u64			sched_out_time;
120 	u64			wake_up_time;
121 	u64			sched_in_time;
122 	u64			runtime;
123 };
124 
125 struct work_atoms {
126 	struct list_head	work_list;
127 	struct thread		*thread;
128 	struct rb_node		node;
129 	u64			max_lat;
130 	u64			max_lat_at;
131 	u64			total_lat;
132 	u64			nb_atoms;
133 	u64			total_runtime;
134 	int			num_merged;
135 };
136 
137 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
138 
139 struct perf_sched;
140 
141 struct trace_sched_handler {
142 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
143 			    struct perf_sample *sample, struct machine *machine);
144 
145 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
146 			     struct perf_sample *sample, struct machine *machine);
147 
148 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
149 			    struct perf_sample *sample, struct machine *machine);
150 
151 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
152 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
153 			  struct machine *machine);
154 
155 	int (*migrate_task_event)(struct perf_sched *sched,
156 				  struct evsel *evsel,
157 				  struct perf_sample *sample,
158 				  struct machine *machine);
159 };
160 
161 #define COLOR_PIDS PERF_COLOR_BLUE
162 #define COLOR_CPUS PERF_COLOR_BG_RED
163 
164 struct perf_sched_map {
165 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
166 	int			*comp_cpus;
167 	bool			 comp;
168 	struct perf_thread_map *color_pids;
169 	const char		*color_pids_str;
170 	struct perf_cpu_map	*color_cpus;
171 	const char		*color_cpus_str;
172 	struct perf_cpu_map	*cpus;
173 	const char		*cpus_str;
174 };
175 
176 struct perf_sched {
177 	struct perf_tool tool;
178 	const char	 *sort_order;
179 	unsigned long	 nr_tasks;
180 	struct task_desc **pid_to_task;
181 	struct task_desc **tasks;
182 	const struct trace_sched_handler *tp_handler;
183 	pthread_mutex_t	 start_work_mutex;
184 	pthread_mutex_t	 work_done_wait_mutex;
185 	int		 profile_cpu;
186 /*
187  * Track the current task - that way we can know whether there's any
188  * weird events, such as a task being switched away that is not current.
189  */
190 	int		 max_cpu;
191 	u32		 curr_pid[MAX_CPUS];
192 	struct thread	 *curr_thread[MAX_CPUS];
193 	char		 next_shortname1;
194 	char		 next_shortname2;
195 	unsigned int	 replay_repeat;
196 	unsigned long	 nr_run_events;
197 	unsigned long	 nr_sleep_events;
198 	unsigned long	 nr_wakeup_events;
199 	unsigned long	 nr_sleep_corrections;
200 	unsigned long	 nr_run_events_optimized;
201 	unsigned long	 targetless_wakeups;
202 	unsigned long	 multitarget_wakeups;
203 	unsigned long	 nr_runs;
204 	unsigned long	 nr_timestamps;
205 	unsigned long	 nr_unordered_timestamps;
206 	unsigned long	 nr_context_switch_bugs;
207 	unsigned long	 nr_events;
208 	unsigned long	 nr_lost_chunks;
209 	unsigned long	 nr_lost_events;
210 	u64		 run_measurement_overhead;
211 	u64		 sleep_measurement_overhead;
212 	u64		 start_time;
213 	u64		 cpu_usage;
214 	u64		 runavg_cpu_usage;
215 	u64		 parent_cpu_usage;
216 	u64		 runavg_parent_cpu_usage;
217 	u64		 sum_runtime;
218 	u64		 sum_fluct;
219 	u64		 run_avg;
220 	u64		 all_runtime;
221 	u64		 all_count;
222 	u64		 cpu_last_switched[MAX_CPUS];
223 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
224 	struct list_head sort_list, cmp_pid;
225 	bool force;
226 	bool skip_merge;
227 	struct perf_sched_map map;
228 
229 	/* options for timehist command */
230 	bool		summary;
231 	bool		summary_only;
232 	bool		idle_hist;
233 	bool		show_callchain;
234 	unsigned int	max_stack;
235 	bool		show_cpu_visual;
236 	bool		show_wakeups;
237 	bool		show_next;
238 	bool		show_migrations;
239 	bool		show_state;
240 	u64		skipped_samples;
241 	const char	*time_str;
242 	struct perf_time_interval ptime;
243 	struct perf_time_interval hist_time;
244 };
245 
246 /* per thread run time data */
247 struct thread_runtime {
248 	u64 last_time;      /* time of previous sched in/out event */
249 	u64 dt_run;         /* run time */
250 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
251 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
252 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
253 	u64 dt_delay;       /* time between wakeup and sched-in */
254 	u64 ready_to_run;   /* time of wakeup */
255 
256 	struct stats run_stats;
257 	u64 total_run_time;
258 	u64 total_sleep_time;
259 	u64 total_iowait_time;
260 	u64 total_preempt_time;
261 	u64 total_delay_time;
262 
263 	int last_state;
264 
265 	char shortname[3];
266 	bool comm_changed;
267 
268 	u64 migrations;
269 };
270 
271 /* per event run time data */
272 struct evsel_runtime {
273 	u64 *last_time; /* time this event was last seen per cpu */
274 	u32 ncpu;       /* highest cpu slot allocated */
275 };
276 
277 /* per cpu idle time data */
278 struct idle_thread_runtime {
279 	struct thread_runtime	tr;
280 	struct thread		*last_thread;
281 	struct rb_root_cached	sorted_root;
282 	struct callchain_root	callchain;
283 	struct callchain_cursor	cursor;
284 };
285 
286 /* track idle times per cpu */
287 static struct thread **idle_threads;
288 static int idle_max_cpu;
289 static char idle_comm[] = "<idle>";
290 
291 static u64 get_nsecs(void)
292 {
293 	struct timespec ts;
294 
295 	clock_gettime(CLOCK_MONOTONIC, &ts);
296 
297 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
298 }
299 
300 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
301 {
302 	u64 T0 = get_nsecs(), T1;
303 
304 	do {
305 		T1 = get_nsecs();
306 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
307 }
308 
309 static void sleep_nsecs(u64 nsecs)
310 {
311 	struct timespec ts;
312 
313 	ts.tv_nsec = nsecs % 999999999;
314 	ts.tv_sec = nsecs / 999999999;
315 
316 	nanosleep(&ts, NULL);
317 }
318 
319 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
320 {
321 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
322 	int i;
323 
324 	for (i = 0; i < 10; i++) {
325 		T0 = get_nsecs();
326 		burn_nsecs(sched, 0);
327 		T1 = get_nsecs();
328 		delta = T1-T0;
329 		min_delta = min(min_delta, delta);
330 	}
331 	sched->run_measurement_overhead = min_delta;
332 
333 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
334 }
335 
336 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
337 {
338 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
339 	int i;
340 
341 	for (i = 0; i < 10; i++) {
342 		T0 = get_nsecs();
343 		sleep_nsecs(10000);
344 		T1 = get_nsecs();
345 		delta = T1-T0;
346 		min_delta = min(min_delta, delta);
347 	}
348 	min_delta -= 10000;
349 	sched->sleep_measurement_overhead = min_delta;
350 
351 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
352 }
353 
354 static struct sched_atom *
355 get_new_event(struct task_desc *task, u64 timestamp)
356 {
357 	struct sched_atom *event = zalloc(sizeof(*event));
358 	unsigned long idx = task->nr_events;
359 	size_t size;
360 
361 	event->timestamp = timestamp;
362 	event->nr = idx;
363 
364 	task->nr_events++;
365 	size = sizeof(struct sched_atom *) * task->nr_events;
366 	task->atoms = realloc(task->atoms, size);
367 	BUG_ON(!task->atoms);
368 
369 	task->atoms[idx] = event;
370 
371 	return event;
372 }
373 
374 static struct sched_atom *last_event(struct task_desc *task)
375 {
376 	if (!task->nr_events)
377 		return NULL;
378 
379 	return task->atoms[task->nr_events - 1];
380 }
381 
382 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
383 				u64 timestamp, u64 duration)
384 {
385 	struct sched_atom *event, *curr_event = last_event(task);
386 
387 	/*
388 	 * optimize an existing RUN event by merging this one
389 	 * to it:
390 	 */
391 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
392 		sched->nr_run_events_optimized++;
393 		curr_event->duration += duration;
394 		return;
395 	}
396 
397 	event = get_new_event(task, timestamp);
398 
399 	event->type = SCHED_EVENT_RUN;
400 	event->duration = duration;
401 
402 	sched->nr_run_events++;
403 }
404 
405 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
406 				   u64 timestamp, struct task_desc *wakee)
407 {
408 	struct sched_atom *event, *wakee_event;
409 
410 	event = get_new_event(task, timestamp);
411 	event->type = SCHED_EVENT_WAKEUP;
412 	event->wakee = wakee;
413 
414 	wakee_event = last_event(wakee);
415 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
416 		sched->targetless_wakeups++;
417 		return;
418 	}
419 	if (wakee_event->wait_sem) {
420 		sched->multitarget_wakeups++;
421 		return;
422 	}
423 
424 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
425 	sem_init(wakee_event->wait_sem, 0, 0);
426 	wakee_event->specific_wait = 1;
427 	event->wait_sem = wakee_event->wait_sem;
428 
429 	sched->nr_wakeup_events++;
430 }
431 
432 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
433 				  u64 timestamp, u64 task_state __maybe_unused)
434 {
435 	struct sched_atom *event = get_new_event(task, timestamp);
436 
437 	event->type = SCHED_EVENT_SLEEP;
438 
439 	sched->nr_sleep_events++;
440 }
441 
442 static struct task_desc *register_pid(struct perf_sched *sched,
443 				      unsigned long pid, const char *comm)
444 {
445 	struct task_desc *task;
446 	static int pid_max;
447 
448 	if (sched->pid_to_task == NULL) {
449 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
450 			pid_max = MAX_PID;
451 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
452 	}
453 	if (pid >= (unsigned long)pid_max) {
454 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
455 			sizeof(struct task_desc *))) == NULL);
456 		while (pid >= (unsigned long)pid_max)
457 			sched->pid_to_task[pid_max++] = NULL;
458 	}
459 
460 	task = sched->pid_to_task[pid];
461 
462 	if (task)
463 		return task;
464 
465 	task = zalloc(sizeof(*task));
466 	task->pid = pid;
467 	task->nr = sched->nr_tasks;
468 	strcpy(task->comm, comm);
469 	/*
470 	 * every task starts in sleeping state - this gets ignored
471 	 * if there's no wakeup pointing to this sleep state:
472 	 */
473 	add_sched_event_sleep(sched, task, 0, 0);
474 
475 	sched->pid_to_task[pid] = task;
476 	sched->nr_tasks++;
477 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
478 	BUG_ON(!sched->tasks);
479 	sched->tasks[task->nr] = task;
480 
481 	if (verbose > 0)
482 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
483 
484 	return task;
485 }
486 
487 
488 static void print_task_traces(struct perf_sched *sched)
489 {
490 	struct task_desc *task;
491 	unsigned long i;
492 
493 	for (i = 0; i < sched->nr_tasks; i++) {
494 		task = sched->tasks[i];
495 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
496 			task->nr, task->comm, task->pid, task->nr_events);
497 	}
498 }
499 
500 static void add_cross_task_wakeups(struct perf_sched *sched)
501 {
502 	struct task_desc *task1, *task2;
503 	unsigned long i, j;
504 
505 	for (i = 0; i < sched->nr_tasks; i++) {
506 		task1 = sched->tasks[i];
507 		j = i + 1;
508 		if (j == sched->nr_tasks)
509 			j = 0;
510 		task2 = sched->tasks[j];
511 		add_sched_event_wakeup(sched, task1, 0, task2);
512 	}
513 }
514 
515 static void perf_sched__process_event(struct perf_sched *sched,
516 				      struct sched_atom *atom)
517 {
518 	int ret = 0;
519 
520 	switch (atom->type) {
521 		case SCHED_EVENT_RUN:
522 			burn_nsecs(sched, atom->duration);
523 			break;
524 		case SCHED_EVENT_SLEEP:
525 			if (atom->wait_sem)
526 				ret = sem_wait(atom->wait_sem);
527 			BUG_ON(ret);
528 			break;
529 		case SCHED_EVENT_WAKEUP:
530 			if (atom->wait_sem)
531 				ret = sem_post(atom->wait_sem);
532 			BUG_ON(ret);
533 			break;
534 		case SCHED_EVENT_MIGRATION:
535 			break;
536 		default:
537 			BUG_ON(1);
538 	}
539 }
540 
541 static u64 get_cpu_usage_nsec_parent(void)
542 {
543 	struct rusage ru;
544 	u64 sum;
545 	int err;
546 
547 	err = getrusage(RUSAGE_SELF, &ru);
548 	BUG_ON(err);
549 
550 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
551 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
552 
553 	return sum;
554 }
555 
556 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
557 {
558 	struct perf_event_attr attr;
559 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
560 	int fd;
561 	struct rlimit limit;
562 	bool need_privilege = false;
563 
564 	memset(&attr, 0, sizeof(attr));
565 
566 	attr.type = PERF_TYPE_SOFTWARE;
567 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
568 
569 force_again:
570 	fd = sys_perf_event_open(&attr, 0, -1, -1,
571 				 perf_event_open_cloexec_flag());
572 
573 	if (fd < 0) {
574 		if (errno == EMFILE) {
575 			if (sched->force) {
576 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
577 				limit.rlim_cur += sched->nr_tasks - cur_task;
578 				if (limit.rlim_cur > limit.rlim_max) {
579 					limit.rlim_max = limit.rlim_cur;
580 					need_privilege = true;
581 				}
582 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
583 					if (need_privilege && errno == EPERM)
584 						strcpy(info, "Need privilege\n");
585 				} else
586 					goto force_again;
587 			} else
588 				strcpy(info, "Have a try with -f option\n");
589 		}
590 		pr_err("Error: sys_perf_event_open() syscall returned "
591 		       "with %d (%s)\n%s", fd,
592 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
593 		exit(EXIT_FAILURE);
594 	}
595 	return fd;
596 }
597 
598 static u64 get_cpu_usage_nsec_self(int fd)
599 {
600 	u64 runtime;
601 	int ret;
602 
603 	ret = read(fd, &runtime, sizeof(runtime));
604 	BUG_ON(ret != sizeof(runtime));
605 
606 	return runtime;
607 }
608 
609 struct sched_thread_parms {
610 	struct task_desc  *task;
611 	struct perf_sched *sched;
612 	int fd;
613 };
614 
615 static void *thread_func(void *ctx)
616 {
617 	struct sched_thread_parms *parms = ctx;
618 	struct task_desc *this_task = parms->task;
619 	struct perf_sched *sched = parms->sched;
620 	u64 cpu_usage_0, cpu_usage_1;
621 	unsigned long i, ret;
622 	char comm2[22];
623 	int fd = parms->fd;
624 
625 	zfree(&parms);
626 
627 	sprintf(comm2, ":%s", this_task->comm);
628 	prctl(PR_SET_NAME, comm2);
629 	if (fd < 0)
630 		return NULL;
631 again:
632 	ret = sem_post(&this_task->ready_for_work);
633 	BUG_ON(ret);
634 	ret = pthread_mutex_lock(&sched->start_work_mutex);
635 	BUG_ON(ret);
636 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
637 	BUG_ON(ret);
638 
639 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
640 
641 	for (i = 0; i < this_task->nr_events; i++) {
642 		this_task->curr_event = i;
643 		perf_sched__process_event(sched, this_task->atoms[i]);
644 	}
645 
646 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
647 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
648 	ret = sem_post(&this_task->work_done_sem);
649 	BUG_ON(ret);
650 
651 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
652 	BUG_ON(ret);
653 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
654 	BUG_ON(ret);
655 
656 	goto again;
657 }
658 
659 static void create_tasks(struct perf_sched *sched)
660 {
661 	struct task_desc *task;
662 	pthread_attr_t attr;
663 	unsigned long i;
664 	int err;
665 
666 	err = pthread_attr_init(&attr);
667 	BUG_ON(err);
668 	err = pthread_attr_setstacksize(&attr,
669 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
670 	BUG_ON(err);
671 	err = pthread_mutex_lock(&sched->start_work_mutex);
672 	BUG_ON(err);
673 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
674 	BUG_ON(err);
675 	for (i = 0; i < sched->nr_tasks; i++) {
676 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
677 		BUG_ON(parms == NULL);
678 		parms->task = task = sched->tasks[i];
679 		parms->sched = sched;
680 		parms->fd = self_open_counters(sched, i);
681 		sem_init(&task->sleep_sem, 0, 0);
682 		sem_init(&task->ready_for_work, 0, 0);
683 		sem_init(&task->work_done_sem, 0, 0);
684 		task->curr_event = 0;
685 		err = pthread_create(&task->thread, &attr, thread_func, parms);
686 		BUG_ON(err);
687 	}
688 }
689 
690 static void wait_for_tasks(struct perf_sched *sched)
691 {
692 	u64 cpu_usage_0, cpu_usage_1;
693 	struct task_desc *task;
694 	unsigned long i, ret;
695 
696 	sched->start_time = get_nsecs();
697 	sched->cpu_usage = 0;
698 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
699 
700 	for (i = 0; i < sched->nr_tasks; i++) {
701 		task = sched->tasks[i];
702 		ret = sem_wait(&task->ready_for_work);
703 		BUG_ON(ret);
704 		sem_init(&task->ready_for_work, 0, 0);
705 	}
706 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
707 	BUG_ON(ret);
708 
709 	cpu_usage_0 = get_cpu_usage_nsec_parent();
710 
711 	pthread_mutex_unlock(&sched->start_work_mutex);
712 
713 	for (i = 0; i < sched->nr_tasks; i++) {
714 		task = sched->tasks[i];
715 		ret = sem_wait(&task->work_done_sem);
716 		BUG_ON(ret);
717 		sem_init(&task->work_done_sem, 0, 0);
718 		sched->cpu_usage += task->cpu_usage;
719 		task->cpu_usage = 0;
720 	}
721 
722 	cpu_usage_1 = get_cpu_usage_nsec_parent();
723 	if (!sched->runavg_cpu_usage)
724 		sched->runavg_cpu_usage = sched->cpu_usage;
725 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
726 
727 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
728 	if (!sched->runavg_parent_cpu_usage)
729 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
730 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
731 					 sched->parent_cpu_usage)/sched->replay_repeat;
732 
733 	ret = pthread_mutex_lock(&sched->start_work_mutex);
734 	BUG_ON(ret);
735 
736 	for (i = 0; i < sched->nr_tasks; i++) {
737 		task = sched->tasks[i];
738 		sem_init(&task->sleep_sem, 0, 0);
739 		task->curr_event = 0;
740 	}
741 }
742 
743 static void run_one_test(struct perf_sched *sched)
744 {
745 	u64 T0, T1, delta, avg_delta, fluct;
746 
747 	T0 = get_nsecs();
748 	wait_for_tasks(sched);
749 	T1 = get_nsecs();
750 
751 	delta = T1 - T0;
752 	sched->sum_runtime += delta;
753 	sched->nr_runs++;
754 
755 	avg_delta = sched->sum_runtime / sched->nr_runs;
756 	if (delta < avg_delta)
757 		fluct = avg_delta - delta;
758 	else
759 		fluct = delta - avg_delta;
760 	sched->sum_fluct += fluct;
761 	if (!sched->run_avg)
762 		sched->run_avg = delta;
763 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
764 
765 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
766 
767 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
768 
769 	printf("cpu: %0.2f / %0.2f",
770 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
771 
772 #if 0
773 	/*
774 	 * rusage statistics done by the parent, these are less
775 	 * accurate than the sched->sum_exec_runtime based statistics:
776 	 */
777 	printf(" [%0.2f / %0.2f]",
778 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
779 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
780 #endif
781 
782 	printf("\n");
783 
784 	if (sched->nr_sleep_corrections)
785 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
786 	sched->nr_sleep_corrections = 0;
787 }
788 
789 static void test_calibrations(struct perf_sched *sched)
790 {
791 	u64 T0, T1;
792 
793 	T0 = get_nsecs();
794 	burn_nsecs(sched, NSEC_PER_MSEC);
795 	T1 = get_nsecs();
796 
797 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
798 
799 	T0 = get_nsecs();
800 	sleep_nsecs(NSEC_PER_MSEC);
801 	T1 = get_nsecs();
802 
803 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
804 }
805 
806 static int
807 replay_wakeup_event(struct perf_sched *sched,
808 		    struct evsel *evsel, struct perf_sample *sample,
809 		    struct machine *machine __maybe_unused)
810 {
811 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
812 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
813 	struct task_desc *waker, *wakee;
814 
815 	if (verbose > 0) {
816 		printf("sched_wakeup event %p\n", evsel);
817 
818 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
819 	}
820 
821 	waker = register_pid(sched, sample->tid, "<unknown>");
822 	wakee = register_pid(sched, pid, comm);
823 
824 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
825 	return 0;
826 }
827 
828 static int replay_switch_event(struct perf_sched *sched,
829 			       struct evsel *evsel,
830 			       struct perf_sample *sample,
831 			       struct machine *machine __maybe_unused)
832 {
833 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
834 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
835 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
836 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
837 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
838 	struct task_desc *prev, __maybe_unused *next;
839 	u64 timestamp0, timestamp = sample->time;
840 	int cpu = sample->cpu;
841 	s64 delta;
842 
843 	if (verbose > 0)
844 		printf("sched_switch event %p\n", evsel);
845 
846 	if (cpu >= MAX_CPUS || cpu < 0)
847 		return 0;
848 
849 	timestamp0 = sched->cpu_last_switched[cpu];
850 	if (timestamp0)
851 		delta = timestamp - timestamp0;
852 	else
853 		delta = 0;
854 
855 	if (delta < 0) {
856 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
857 		return -1;
858 	}
859 
860 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
861 		 prev_comm, prev_pid, next_comm, next_pid, delta);
862 
863 	prev = register_pid(sched, prev_pid, prev_comm);
864 	next = register_pid(sched, next_pid, next_comm);
865 
866 	sched->cpu_last_switched[cpu] = timestamp;
867 
868 	add_sched_event_run(sched, prev, timestamp, delta);
869 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
870 
871 	return 0;
872 }
873 
874 static int replay_fork_event(struct perf_sched *sched,
875 			     union perf_event *event,
876 			     struct machine *machine)
877 {
878 	struct thread *child, *parent;
879 
880 	child = machine__findnew_thread(machine, event->fork.pid,
881 					event->fork.tid);
882 	parent = machine__findnew_thread(machine, event->fork.ppid,
883 					 event->fork.ptid);
884 
885 	if (child == NULL || parent == NULL) {
886 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
887 				 child, parent);
888 		goto out_put;
889 	}
890 
891 	if (verbose > 0) {
892 		printf("fork event\n");
893 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
894 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
895 	}
896 
897 	register_pid(sched, parent->tid, thread__comm_str(parent));
898 	register_pid(sched, child->tid, thread__comm_str(child));
899 out_put:
900 	thread__put(child);
901 	thread__put(parent);
902 	return 0;
903 }
904 
905 struct sort_dimension {
906 	const char		*name;
907 	sort_fn_t		cmp;
908 	struct list_head	list;
909 };
910 
911 /*
912  * handle runtime stats saved per thread
913  */
914 static struct thread_runtime *thread__init_runtime(struct thread *thread)
915 {
916 	struct thread_runtime *r;
917 
918 	r = zalloc(sizeof(struct thread_runtime));
919 	if (!r)
920 		return NULL;
921 
922 	init_stats(&r->run_stats);
923 	thread__set_priv(thread, r);
924 
925 	return r;
926 }
927 
928 static struct thread_runtime *thread__get_runtime(struct thread *thread)
929 {
930 	struct thread_runtime *tr;
931 
932 	tr = thread__priv(thread);
933 	if (tr == NULL) {
934 		tr = thread__init_runtime(thread);
935 		if (tr == NULL)
936 			pr_debug("Failed to malloc memory for runtime data.\n");
937 	}
938 
939 	return tr;
940 }
941 
942 static int
943 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
944 {
945 	struct sort_dimension *sort;
946 	int ret = 0;
947 
948 	BUG_ON(list_empty(list));
949 
950 	list_for_each_entry(sort, list, list) {
951 		ret = sort->cmp(l, r);
952 		if (ret)
953 			return ret;
954 	}
955 
956 	return ret;
957 }
958 
959 static struct work_atoms *
960 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
961 			 struct list_head *sort_list)
962 {
963 	struct rb_node *node = root->rb_root.rb_node;
964 	struct work_atoms key = { .thread = thread };
965 
966 	while (node) {
967 		struct work_atoms *atoms;
968 		int cmp;
969 
970 		atoms = container_of(node, struct work_atoms, node);
971 
972 		cmp = thread_lat_cmp(sort_list, &key, atoms);
973 		if (cmp > 0)
974 			node = node->rb_left;
975 		else if (cmp < 0)
976 			node = node->rb_right;
977 		else {
978 			BUG_ON(thread != atoms->thread);
979 			return atoms;
980 		}
981 	}
982 	return NULL;
983 }
984 
985 static void
986 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
987 			 struct list_head *sort_list)
988 {
989 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
990 	bool leftmost = true;
991 
992 	while (*new) {
993 		struct work_atoms *this;
994 		int cmp;
995 
996 		this = container_of(*new, struct work_atoms, node);
997 		parent = *new;
998 
999 		cmp = thread_lat_cmp(sort_list, data, this);
1000 
1001 		if (cmp > 0)
1002 			new = &((*new)->rb_left);
1003 		else {
1004 			new = &((*new)->rb_right);
1005 			leftmost = false;
1006 		}
1007 	}
1008 
1009 	rb_link_node(&data->node, parent, new);
1010 	rb_insert_color_cached(&data->node, root, leftmost);
1011 }
1012 
1013 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1014 {
1015 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1016 	if (!atoms) {
1017 		pr_err("No memory at %s\n", __func__);
1018 		return -1;
1019 	}
1020 
1021 	atoms->thread = thread__get(thread);
1022 	INIT_LIST_HEAD(&atoms->work_list);
1023 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1024 	return 0;
1025 }
1026 
1027 static char sched_out_state(u64 prev_state)
1028 {
1029 	const char *str = TASK_STATE_TO_CHAR_STR;
1030 
1031 	return str[prev_state];
1032 }
1033 
1034 static int
1035 add_sched_out_event(struct work_atoms *atoms,
1036 		    char run_state,
1037 		    u64 timestamp)
1038 {
1039 	struct work_atom *atom = zalloc(sizeof(*atom));
1040 	if (!atom) {
1041 		pr_err("Non memory at %s", __func__);
1042 		return -1;
1043 	}
1044 
1045 	atom->sched_out_time = timestamp;
1046 
1047 	if (run_state == 'R') {
1048 		atom->state = THREAD_WAIT_CPU;
1049 		atom->wake_up_time = atom->sched_out_time;
1050 	}
1051 
1052 	list_add_tail(&atom->list, &atoms->work_list);
1053 	return 0;
1054 }
1055 
1056 static void
1057 add_runtime_event(struct work_atoms *atoms, u64 delta,
1058 		  u64 timestamp __maybe_unused)
1059 {
1060 	struct work_atom *atom;
1061 
1062 	BUG_ON(list_empty(&atoms->work_list));
1063 
1064 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1065 
1066 	atom->runtime += delta;
1067 	atoms->total_runtime += delta;
1068 }
1069 
1070 static void
1071 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1072 {
1073 	struct work_atom *atom;
1074 	u64 delta;
1075 
1076 	if (list_empty(&atoms->work_list))
1077 		return;
1078 
1079 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1080 
1081 	if (atom->state != THREAD_WAIT_CPU)
1082 		return;
1083 
1084 	if (timestamp < atom->wake_up_time) {
1085 		atom->state = THREAD_IGNORE;
1086 		return;
1087 	}
1088 
1089 	atom->state = THREAD_SCHED_IN;
1090 	atom->sched_in_time = timestamp;
1091 
1092 	delta = atom->sched_in_time - atom->wake_up_time;
1093 	atoms->total_lat += delta;
1094 	if (delta > atoms->max_lat) {
1095 		atoms->max_lat = delta;
1096 		atoms->max_lat_at = timestamp;
1097 	}
1098 	atoms->nb_atoms++;
1099 }
1100 
1101 static int latency_switch_event(struct perf_sched *sched,
1102 				struct evsel *evsel,
1103 				struct perf_sample *sample,
1104 				struct machine *machine)
1105 {
1106 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1107 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1108 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1109 	struct work_atoms *out_events, *in_events;
1110 	struct thread *sched_out, *sched_in;
1111 	u64 timestamp0, timestamp = sample->time;
1112 	int cpu = sample->cpu, err = -1;
1113 	s64 delta;
1114 
1115 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1116 
1117 	timestamp0 = sched->cpu_last_switched[cpu];
1118 	sched->cpu_last_switched[cpu] = timestamp;
1119 	if (timestamp0)
1120 		delta = timestamp - timestamp0;
1121 	else
1122 		delta = 0;
1123 
1124 	if (delta < 0) {
1125 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1126 		return -1;
1127 	}
1128 
1129 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1130 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1131 	if (sched_out == NULL || sched_in == NULL)
1132 		goto out_put;
1133 
1134 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1135 	if (!out_events) {
1136 		if (thread_atoms_insert(sched, sched_out))
1137 			goto out_put;
1138 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1139 		if (!out_events) {
1140 			pr_err("out-event: Internal tree error");
1141 			goto out_put;
1142 		}
1143 	}
1144 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1145 		return -1;
1146 
1147 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1148 	if (!in_events) {
1149 		if (thread_atoms_insert(sched, sched_in))
1150 			goto out_put;
1151 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1152 		if (!in_events) {
1153 			pr_err("in-event: Internal tree error");
1154 			goto out_put;
1155 		}
1156 		/*
1157 		 * Take came in we have not heard about yet,
1158 		 * add in an initial atom in runnable state:
1159 		 */
1160 		if (add_sched_out_event(in_events, 'R', timestamp))
1161 			goto out_put;
1162 	}
1163 	add_sched_in_event(in_events, timestamp);
1164 	err = 0;
1165 out_put:
1166 	thread__put(sched_out);
1167 	thread__put(sched_in);
1168 	return err;
1169 }
1170 
1171 static int latency_runtime_event(struct perf_sched *sched,
1172 				 struct evsel *evsel,
1173 				 struct perf_sample *sample,
1174 				 struct machine *machine)
1175 {
1176 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1177 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1178 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1179 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1180 	u64 timestamp = sample->time;
1181 	int cpu = sample->cpu, err = -1;
1182 
1183 	if (thread == NULL)
1184 		return -1;
1185 
1186 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1187 	if (!atoms) {
1188 		if (thread_atoms_insert(sched, thread))
1189 			goto out_put;
1190 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1191 		if (!atoms) {
1192 			pr_err("in-event: Internal tree error");
1193 			goto out_put;
1194 		}
1195 		if (add_sched_out_event(atoms, 'R', timestamp))
1196 			goto out_put;
1197 	}
1198 
1199 	add_runtime_event(atoms, runtime, timestamp);
1200 	err = 0;
1201 out_put:
1202 	thread__put(thread);
1203 	return err;
1204 }
1205 
1206 static int latency_wakeup_event(struct perf_sched *sched,
1207 				struct evsel *evsel,
1208 				struct perf_sample *sample,
1209 				struct machine *machine)
1210 {
1211 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1212 	struct work_atoms *atoms;
1213 	struct work_atom *atom;
1214 	struct thread *wakee;
1215 	u64 timestamp = sample->time;
1216 	int err = -1;
1217 
1218 	wakee = machine__findnew_thread(machine, -1, pid);
1219 	if (wakee == NULL)
1220 		return -1;
1221 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1222 	if (!atoms) {
1223 		if (thread_atoms_insert(sched, wakee))
1224 			goto out_put;
1225 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1226 		if (!atoms) {
1227 			pr_err("wakeup-event: Internal tree error");
1228 			goto out_put;
1229 		}
1230 		if (add_sched_out_event(atoms, 'S', timestamp))
1231 			goto out_put;
1232 	}
1233 
1234 	BUG_ON(list_empty(&atoms->work_list));
1235 
1236 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1237 
1238 	/*
1239 	 * As we do not guarantee the wakeup event happens when
1240 	 * task is out of run queue, also may happen when task is
1241 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1242 	 * then we should not set the ->wake_up_time when wake up a
1243 	 * task which is on run queue.
1244 	 *
1245 	 * You WILL be missing events if you've recorded only
1246 	 * one CPU, or are only looking at only one, so don't
1247 	 * skip in this case.
1248 	 */
1249 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1250 		goto out_ok;
1251 
1252 	sched->nr_timestamps++;
1253 	if (atom->sched_out_time > timestamp) {
1254 		sched->nr_unordered_timestamps++;
1255 		goto out_ok;
1256 	}
1257 
1258 	atom->state = THREAD_WAIT_CPU;
1259 	atom->wake_up_time = timestamp;
1260 out_ok:
1261 	err = 0;
1262 out_put:
1263 	thread__put(wakee);
1264 	return err;
1265 }
1266 
1267 static int latency_migrate_task_event(struct perf_sched *sched,
1268 				      struct evsel *evsel,
1269 				      struct perf_sample *sample,
1270 				      struct machine *machine)
1271 {
1272 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1273 	u64 timestamp = sample->time;
1274 	struct work_atoms *atoms;
1275 	struct work_atom *atom;
1276 	struct thread *migrant;
1277 	int err = -1;
1278 
1279 	/*
1280 	 * Only need to worry about migration when profiling one CPU.
1281 	 */
1282 	if (sched->profile_cpu == -1)
1283 		return 0;
1284 
1285 	migrant = machine__findnew_thread(machine, -1, pid);
1286 	if (migrant == NULL)
1287 		return -1;
1288 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1289 	if (!atoms) {
1290 		if (thread_atoms_insert(sched, migrant))
1291 			goto out_put;
1292 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1293 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1294 		if (!atoms) {
1295 			pr_err("migration-event: Internal tree error");
1296 			goto out_put;
1297 		}
1298 		if (add_sched_out_event(atoms, 'R', timestamp))
1299 			goto out_put;
1300 	}
1301 
1302 	BUG_ON(list_empty(&atoms->work_list));
1303 
1304 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1305 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1306 
1307 	sched->nr_timestamps++;
1308 
1309 	if (atom->sched_out_time > timestamp)
1310 		sched->nr_unordered_timestamps++;
1311 	err = 0;
1312 out_put:
1313 	thread__put(migrant);
1314 	return err;
1315 }
1316 
1317 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1318 {
1319 	int i;
1320 	int ret;
1321 	u64 avg;
1322 	char max_lat_at[32];
1323 
1324 	if (!work_list->nb_atoms)
1325 		return;
1326 	/*
1327 	 * Ignore idle threads:
1328 	 */
1329 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1330 		return;
1331 
1332 	sched->all_runtime += work_list->total_runtime;
1333 	sched->all_count   += work_list->nb_atoms;
1334 
1335 	if (work_list->num_merged > 1)
1336 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1337 	else
1338 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1339 
1340 	for (i = 0; i < 24 - ret; i++)
1341 		printf(" ");
1342 
1343 	avg = work_list->total_lat / work_list->nb_atoms;
1344 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1345 
1346 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1347 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1348 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1349 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1350 		 max_lat_at);
1351 }
1352 
1353 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1354 {
1355 	if (l->thread == r->thread)
1356 		return 0;
1357 	if (l->thread->tid < r->thread->tid)
1358 		return -1;
1359 	if (l->thread->tid > r->thread->tid)
1360 		return 1;
1361 	return (int)(l->thread - r->thread);
1362 }
1363 
1364 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1365 {
1366 	u64 avgl, avgr;
1367 
1368 	if (!l->nb_atoms)
1369 		return -1;
1370 
1371 	if (!r->nb_atoms)
1372 		return 1;
1373 
1374 	avgl = l->total_lat / l->nb_atoms;
1375 	avgr = r->total_lat / r->nb_atoms;
1376 
1377 	if (avgl < avgr)
1378 		return -1;
1379 	if (avgl > avgr)
1380 		return 1;
1381 
1382 	return 0;
1383 }
1384 
1385 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1386 {
1387 	if (l->max_lat < r->max_lat)
1388 		return -1;
1389 	if (l->max_lat > r->max_lat)
1390 		return 1;
1391 
1392 	return 0;
1393 }
1394 
1395 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1396 {
1397 	if (l->nb_atoms < r->nb_atoms)
1398 		return -1;
1399 	if (l->nb_atoms > r->nb_atoms)
1400 		return 1;
1401 
1402 	return 0;
1403 }
1404 
1405 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1406 {
1407 	if (l->total_runtime < r->total_runtime)
1408 		return -1;
1409 	if (l->total_runtime > r->total_runtime)
1410 		return 1;
1411 
1412 	return 0;
1413 }
1414 
1415 static int sort_dimension__add(const char *tok, struct list_head *list)
1416 {
1417 	size_t i;
1418 	static struct sort_dimension avg_sort_dimension = {
1419 		.name = "avg",
1420 		.cmp  = avg_cmp,
1421 	};
1422 	static struct sort_dimension max_sort_dimension = {
1423 		.name = "max",
1424 		.cmp  = max_cmp,
1425 	};
1426 	static struct sort_dimension pid_sort_dimension = {
1427 		.name = "pid",
1428 		.cmp  = pid_cmp,
1429 	};
1430 	static struct sort_dimension runtime_sort_dimension = {
1431 		.name = "runtime",
1432 		.cmp  = runtime_cmp,
1433 	};
1434 	static struct sort_dimension switch_sort_dimension = {
1435 		.name = "switch",
1436 		.cmp  = switch_cmp,
1437 	};
1438 	struct sort_dimension *available_sorts[] = {
1439 		&pid_sort_dimension,
1440 		&avg_sort_dimension,
1441 		&max_sort_dimension,
1442 		&switch_sort_dimension,
1443 		&runtime_sort_dimension,
1444 	};
1445 
1446 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1447 		if (!strcmp(available_sorts[i]->name, tok)) {
1448 			list_add_tail(&available_sorts[i]->list, list);
1449 
1450 			return 0;
1451 		}
1452 	}
1453 
1454 	return -1;
1455 }
1456 
1457 static void perf_sched__sort_lat(struct perf_sched *sched)
1458 {
1459 	struct rb_node *node;
1460 	struct rb_root_cached *root = &sched->atom_root;
1461 again:
1462 	for (;;) {
1463 		struct work_atoms *data;
1464 		node = rb_first_cached(root);
1465 		if (!node)
1466 			break;
1467 
1468 		rb_erase_cached(node, root);
1469 		data = rb_entry(node, struct work_atoms, node);
1470 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1471 	}
1472 	if (root == &sched->atom_root) {
1473 		root = &sched->merged_atom_root;
1474 		goto again;
1475 	}
1476 }
1477 
1478 static int process_sched_wakeup_event(struct perf_tool *tool,
1479 				      struct evsel *evsel,
1480 				      struct perf_sample *sample,
1481 				      struct machine *machine)
1482 {
1483 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1484 
1485 	if (sched->tp_handler->wakeup_event)
1486 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1487 
1488 	return 0;
1489 }
1490 
1491 union map_priv {
1492 	void	*ptr;
1493 	bool	 color;
1494 };
1495 
1496 static bool thread__has_color(struct thread *thread)
1497 {
1498 	union map_priv priv = {
1499 		.ptr = thread__priv(thread),
1500 	};
1501 
1502 	return priv.color;
1503 }
1504 
1505 static struct thread*
1506 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1507 {
1508 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1509 	union map_priv priv = {
1510 		.color = false,
1511 	};
1512 
1513 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1514 		return thread;
1515 
1516 	if (thread_map__has(sched->map.color_pids, tid))
1517 		priv.color = true;
1518 
1519 	thread__set_priv(thread, priv.ptr);
1520 	return thread;
1521 }
1522 
1523 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1524 			    struct perf_sample *sample, struct machine *machine)
1525 {
1526 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1527 	struct thread *sched_in;
1528 	struct thread_runtime *tr;
1529 	int new_shortname;
1530 	u64 timestamp0, timestamp = sample->time;
1531 	s64 delta;
1532 	int i, this_cpu = sample->cpu;
1533 	int cpus_nr;
1534 	bool new_cpu = false;
1535 	const char *color = PERF_COLOR_NORMAL;
1536 	char stimestamp[32];
1537 
1538 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1539 
1540 	if (this_cpu > sched->max_cpu)
1541 		sched->max_cpu = this_cpu;
1542 
1543 	if (sched->map.comp) {
1544 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1545 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1546 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1547 			new_cpu = true;
1548 		}
1549 	} else
1550 		cpus_nr = sched->max_cpu;
1551 
1552 	timestamp0 = sched->cpu_last_switched[this_cpu];
1553 	sched->cpu_last_switched[this_cpu] = timestamp;
1554 	if (timestamp0)
1555 		delta = timestamp - timestamp0;
1556 	else
1557 		delta = 0;
1558 
1559 	if (delta < 0) {
1560 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1561 		return -1;
1562 	}
1563 
1564 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1565 	if (sched_in == NULL)
1566 		return -1;
1567 
1568 	tr = thread__get_runtime(sched_in);
1569 	if (tr == NULL) {
1570 		thread__put(sched_in);
1571 		return -1;
1572 	}
1573 
1574 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1575 
1576 	printf("  ");
1577 
1578 	new_shortname = 0;
1579 	if (!tr->shortname[0]) {
1580 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1581 			/*
1582 			 * Don't allocate a letter-number for swapper:0
1583 			 * as a shortname. Instead, we use '.' for it.
1584 			 */
1585 			tr->shortname[0] = '.';
1586 			tr->shortname[1] = ' ';
1587 		} else {
1588 			tr->shortname[0] = sched->next_shortname1;
1589 			tr->shortname[1] = sched->next_shortname2;
1590 
1591 			if (sched->next_shortname1 < 'Z') {
1592 				sched->next_shortname1++;
1593 			} else {
1594 				sched->next_shortname1 = 'A';
1595 				if (sched->next_shortname2 < '9')
1596 					sched->next_shortname2++;
1597 				else
1598 					sched->next_shortname2 = '0';
1599 			}
1600 		}
1601 		new_shortname = 1;
1602 	}
1603 
1604 	for (i = 0; i < cpus_nr; i++) {
1605 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1606 		struct thread *curr_thread = sched->curr_thread[cpu];
1607 		struct thread_runtime *curr_tr;
1608 		const char *pid_color = color;
1609 		const char *cpu_color = color;
1610 
1611 		if (curr_thread && thread__has_color(curr_thread))
1612 			pid_color = COLOR_PIDS;
1613 
1614 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1615 			continue;
1616 
1617 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1618 			cpu_color = COLOR_CPUS;
1619 
1620 		if (cpu != this_cpu)
1621 			color_fprintf(stdout, color, " ");
1622 		else
1623 			color_fprintf(stdout, cpu_color, "*");
1624 
1625 		if (sched->curr_thread[cpu]) {
1626 			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1627 			if (curr_tr == NULL) {
1628 				thread__put(sched_in);
1629 				return -1;
1630 			}
1631 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1632 		} else
1633 			color_fprintf(stdout, color, "   ");
1634 	}
1635 
1636 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1637 		goto out;
1638 
1639 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1640 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1641 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1642 		const char *pid_color = color;
1643 
1644 		if (thread__has_color(sched_in))
1645 			pid_color = COLOR_PIDS;
1646 
1647 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1648 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1649 		tr->comm_changed = false;
1650 	}
1651 
1652 	if (sched->map.comp && new_cpu)
1653 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1654 
1655 out:
1656 	color_fprintf(stdout, color, "\n");
1657 
1658 	thread__put(sched_in);
1659 
1660 	return 0;
1661 }
1662 
1663 static int process_sched_switch_event(struct perf_tool *tool,
1664 				      struct evsel *evsel,
1665 				      struct perf_sample *sample,
1666 				      struct machine *machine)
1667 {
1668 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1669 	int this_cpu = sample->cpu, err = 0;
1670 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1671 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1672 
1673 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1674 		/*
1675 		 * Are we trying to switch away a PID that is
1676 		 * not current?
1677 		 */
1678 		if (sched->curr_pid[this_cpu] != prev_pid)
1679 			sched->nr_context_switch_bugs++;
1680 	}
1681 
1682 	if (sched->tp_handler->switch_event)
1683 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1684 
1685 	sched->curr_pid[this_cpu] = next_pid;
1686 	return err;
1687 }
1688 
1689 static int process_sched_runtime_event(struct perf_tool *tool,
1690 				       struct evsel *evsel,
1691 				       struct perf_sample *sample,
1692 				       struct machine *machine)
1693 {
1694 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1695 
1696 	if (sched->tp_handler->runtime_event)
1697 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1698 
1699 	return 0;
1700 }
1701 
1702 static int perf_sched__process_fork_event(struct perf_tool *tool,
1703 					  union perf_event *event,
1704 					  struct perf_sample *sample,
1705 					  struct machine *machine)
1706 {
1707 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1708 
1709 	/* run the fork event through the perf machineruy */
1710 	perf_event__process_fork(tool, event, sample, machine);
1711 
1712 	/* and then run additional processing needed for this command */
1713 	if (sched->tp_handler->fork_event)
1714 		return sched->tp_handler->fork_event(sched, event, machine);
1715 
1716 	return 0;
1717 }
1718 
1719 static int process_sched_migrate_task_event(struct perf_tool *tool,
1720 					    struct evsel *evsel,
1721 					    struct perf_sample *sample,
1722 					    struct machine *machine)
1723 {
1724 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1725 
1726 	if (sched->tp_handler->migrate_task_event)
1727 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1728 
1729 	return 0;
1730 }
1731 
1732 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1733 				  struct evsel *evsel,
1734 				  struct perf_sample *sample,
1735 				  struct machine *machine);
1736 
1737 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1738 						 union perf_event *event __maybe_unused,
1739 						 struct perf_sample *sample,
1740 						 struct evsel *evsel,
1741 						 struct machine *machine)
1742 {
1743 	int err = 0;
1744 
1745 	if (evsel->handler != NULL) {
1746 		tracepoint_handler f = evsel->handler;
1747 		err = f(tool, evsel, sample, machine);
1748 	}
1749 
1750 	return err;
1751 }
1752 
1753 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1754 				    union perf_event *event,
1755 				    struct perf_sample *sample,
1756 				    struct machine *machine)
1757 {
1758 	struct thread *thread;
1759 	struct thread_runtime *tr;
1760 	int err;
1761 
1762 	err = perf_event__process_comm(tool, event, sample, machine);
1763 	if (err)
1764 		return err;
1765 
1766 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1767 	if (!thread) {
1768 		pr_err("Internal error: can't find thread\n");
1769 		return -1;
1770 	}
1771 
1772 	tr = thread__get_runtime(thread);
1773 	if (tr == NULL) {
1774 		thread__put(thread);
1775 		return -1;
1776 	}
1777 
1778 	tr->comm_changed = true;
1779 	thread__put(thread);
1780 
1781 	return 0;
1782 }
1783 
1784 static int perf_sched__read_events(struct perf_sched *sched)
1785 {
1786 	const struct evsel_str_handler handlers[] = {
1787 		{ "sched:sched_switch",	      process_sched_switch_event, },
1788 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1789 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1790 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1791 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1792 	};
1793 	struct perf_session *session;
1794 	struct perf_data data = {
1795 		.path  = input_name,
1796 		.mode  = PERF_DATA_MODE_READ,
1797 		.force = sched->force,
1798 	};
1799 	int rc = -1;
1800 
1801 	session = perf_session__new(&data, false, &sched->tool);
1802 	if (IS_ERR(session)) {
1803 		pr_debug("Error creating perf session");
1804 		return PTR_ERR(session);
1805 	}
1806 
1807 	symbol__init(&session->header.env);
1808 
1809 	if (perf_session__set_tracepoints_handlers(session, handlers))
1810 		goto out_delete;
1811 
1812 	if (perf_session__has_traces(session, "record -R")) {
1813 		int err = perf_session__process_events(session);
1814 		if (err) {
1815 			pr_err("Failed to process events, error %d", err);
1816 			goto out_delete;
1817 		}
1818 
1819 		sched->nr_events      = session->evlist->stats.nr_events[0];
1820 		sched->nr_lost_events = session->evlist->stats.total_lost;
1821 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1822 	}
1823 
1824 	rc = 0;
1825 out_delete:
1826 	perf_session__delete(session);
1827 	return rc;
1828 }
1829 
1830 /*
1831  * scheduling times are printed as msec.usec
1832  */
1833 static inline void print_sched_time(unsigned long long nsecs, int width)
1834 {
1835 	unsigned long msecs;
1836 	unsigned long usecs;
1837 
1838 	msecs  = nsecs / NSEC_PER_MSEC;
1839 	nsecs -= msecs * NSEC_PER_MSEC;
1840 	usecs  = nsecs / NSEC_PER_USEC;
1841 	printf("%*lu.%03lu ", width, msecs, usecs);
1842 }
1843 
1844 /*
1845  * returns runtime data for event, allocating memory for it the
1846  * first time it is used.
1847  */
1848 static struct evsel_runtime *perf_evsel__get_runtime(struct evsel *evsel)
1849 {
1850 	struct evsel_runtime *r = evsel->priv;
1851 
1852 	if (r == NULL) {
1853 		r = zalloc(sizeof(struct evsel_runtime));
1854 		evsel->priv = r;
1855 	}
1856 
1857 	return r;
1858 }
1859 
1860 /*
1861  * save last time event was seen per cpu
1862  */
1863 static void perf_evsel__save_time(struct evsel *evsel,
1864 				  u64 timestamp, u32 cpu)
1865 {
1866 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1867 
1868 	if (r == NULL)
1869 		return;
1870 
1871 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1872 		int i, n = __roundup_pow_of_two(cpu+1);
1873 		void *p = r->last_time;
1874 
1875 		p = realloc(r->last_time, n * sizeof(u64));
1876 		if (!p)
1877 			return;
1878 
1879 		r->last_time = p;
1880 		for (i = r->ncpu; i < n; ++i)
1881 			r->last_time[i] = (u64) 0;
1882 
1883 		r->ncpu = n;
1884 	}
1885 
1886 	r->last_time[cpu] = timestamp;
1887 }
1888 
1889 /* returns last time this event was seen on the given cpu */
1890 static u64 perf_evsel__get_time(struct evsel *evsel, u32 cpu)
1891 {
1892 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1893 
1894 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1895 		return 0;
1896 
1897 	return r->last_time[cpu];
1898 }
1899 
1900 static int comm_width = 30;
1901 
1902 static char *timehist_get_commstr(struct thread *thread)
1903 {
1904 	static char str[32];
1905 	const char *comm = thread__comm_str(thread);
1906 	pid_t tid = thread->tid;
1907 	pid_t pid = thread->pid_;
1908 	int n;
1909 
1910 	if (pid == 0)
1911 		n = scnprintf(str, sizeof(str), "%s", comm);
1912 
1913 	else if (tid != pid)
1914 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1915 
1916 	else
1917 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1918 
1919 	if (n > comm_width)
1920 		comm_width = n;
1921 
1922 	return str;
1923 }
1924 
1925 static void timehist_header(struct perf_sched *sched)
1926 {
1927 	u32 ncpus = sched->max_cpu + 1;
1928 	u32 i, j;
1929 
1930 	printf("%15s %6s ", "time", "cpu");
1931 
1932 	if (sched->show_cpu_visual) {
1933 		printf(" ");
1934 		for (i = 0, j = 0; i < ncpus; ++i) {
1935 			printf("%x", j++);
1936 			if (j > 15)
1937 				j = 0;
1938 		}
1939 		printf(" ");
1940 	}
1941 
1942 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1943 		"task name", "wait time", "sch delay", "run time");
1944 
1945 	if (sched->show_state)
1946 		printf("  %s", "state");
1947 
1948 	printf("\n");
1949 
1950 	/*
1951 	 * units row
1952 	 */
1953 	printf("%15s %-6s ", "", "");
1954 
1955 	if (sched->show_cpu_visual)
1956 		printf(" %*s ", ncpus, "");
1957 
1958 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1959 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1960 
1961 	if (sched->show_state)
1962 		printf("  %5s", "");
1963 
1964 	printf("\n");
1965 
1966 	/*
1967 	 * separator
1968 	 */
1969 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1970 
1971 	if (sched->show_cpu_visual)
1972 		printf(" %.*s ", ncpus, graph_dotted_line);
1973 
1974 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1975 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1976 		graph_dotted_line);
1977 
1978 	if (sched->show_state)
1979 		printf("  %.5s", graph_dotted_line);
1980 
1981 	printf("\n");
1982 }
1983 
1984 static char task_state_char(struct thread *thread, int state)
1985 {
1986 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1987 	unsigned bit = state ? ffs(state) : 0;
1988 
1989 	/* 'I' for idle */
1990 	if (thread->tid == 0)
1991 		return 'I';
1992 
1993 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1994 }
1995 
1996 static void timehist_print_sample(struct perf_sched *sched,
1997 				  struct evsel *evsel,
1998 				  struct perf_sample *sample,
1999 				  struct addr_location *al,
2000 				  struct thread *thread,
2001 				  u64 t, int state)
2002 {
2003 	struct thread_runtime *tr = thread__priv(thread);
2004 	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
2005 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
2006 	u32 max_cpus = sched->max_cpu + 1;
2007 	char tstr[64];
2008 	char nstr[30];
2009 	u64 wait_time;
2010 
2011 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2012 	printf("%15s [%04d] ", tstr, sample->cpu);
2013 
2014 	if (sched->show_cpu_visual) {
2015 		u32 i;
2016 		char c;
2017 
2018 		printf(" ");
2019 		for (i = 0; i < max_cpus; ++i) {
2020 			/* flag idle times with 'i'; others are sched events */
2021 			if (i == sample->cpu)
2022 				c = (thread->tid == 0) ? 'i' : 's';
2023 			else
2024 				c = ' ';
2025 			printf("%c", c);
2026 		}
2027 		printf(" ");
2028 	}
2029 
2030 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2031 
2032 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2033 	print_sched_time(wait_time, 6);
2034 
2035 	print_sched_time(tr->dt_delay, 6);
2036 	print_sched_time(tr->dt_run, 6);
2037 
2038 	if (sched->show_state)
2039 		printf(" %5c ", task_state_char(thread, state));
2040 
2041 	if (sched->show_next) {
2042 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2043 		printf(" %-*s", comm_width, nstr);
2044 	}
2045 
2046 	if (sched->show_wakeups && !sched->show_next)
2047 		printf("  %-*s", comm_width, "");
2048 
2049 	if (thread->tid == 0)
2050 		goto out;
2051 
2052 	if (sched->show_callchain)
2053 		printf("  ");
2054 
2055 	sample__fprintf_sym(sample, al, 0,
2056 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2057 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2058 			    EVSEL__PRINT_SKIP_IGNORED,
2059 			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2060 
2061 out:
2062 	printf("\n");
2063 }
2064 
2065 /*
2066  * Explanation of delta-time stats:
2067  *
2068  *            t = time of current schedule out event
2069  *        tprev = time of previous sched out event
2070  *                also time of schedule-in event for current task
2071  *    last_time = time of last sched change event for current task
2072  *                (i.e, time process was last scheduled out)
2073  * ready_to_run = time of wakeup for current task
2074  *
2075  * -----|------------|------------|------------|------
2076  *    last         ready        tprev          t
2077  *    time         to run
2078  *
2079  *      |-------- dt_wait --------|
2080  *                   |- dt_delay -|-- dt_run --|
2081  *
2082  *   dt_run = run time of current task
2083  *  dt_wait = time between last schedule out event for task and tprev
2084  *            represents time spent off the cpu
2085  * dt_delay = time between wakeup and schedule-in of task
2086  */
2087 
2088 static void timehist_update_runtime_stats(struct thread_runtime *r,
2089 					 u64 t, u64 tprev)
2090 {
2091 	r->dt_delay   = 0;
2092 	r->dt_sleep   = 0;
2093 	r->dt_iowait  = 0;
2094 	r->dt_preempt = 0;
2095 	r->dt_run     = 0;
2096 
2097 	if (tprev) {
2098 		r->dt_run = t - tprev;
2099 		if (r->ready_to_run) {
2100 			if (r->ready_to_run > tprev)
2101 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2102 			else
2103 				r->dt_delay = tprev - r->ready_to_run;
2104 		}
2105 
2106 		if (r->last_time > tprev)
2107 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2108 		else if (r->last_time) {
2109 			u64 dt_wait = tprev - r->last_time;
2110 
2111 			if (r->last_state == TASK_RUNNING)
2112 				r->dt_preempt = dt_wait;
2113 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2114 				r->dt_iowait = dt_wait;
2115 			else
2116 				r->dt_sleep = dt_wait;
2117 		}
2118 	}
2119 
2120 	update_stats(&r->run_stats, r->dt_run);
2121 
2122 	r->total_run_time     += r->dt_run;
2123 	r->total_delay_time   += r->dt_delay;
2124 	r->total_sleep_time   += r->dt_sleep;
2125 	r->total_iowait_time  += r->dt_iowait;
2126 	r->total_preempt_time += r->dt_preempt;
2127 }
2128 
2129 static bool is_idle_sample(struct perf_sample *sample,
2130 			   struct evsel *evsel)
2131 {
2132 	/* pid 0 == swapper == idle task */
2133 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2134 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2135 
2136 	return sample->pid == 0;
2137 }
2138 
2139 static void save_task_callchain(struct perf_sched *sched,
2140 				struct perf_sample *sample,
2141 				struct evsel *evsel,
2142 				struct machine *machine)
2143 {
2144 	struct callchain_cursor *cursor = &callchain_cursor;
2145 	struct thread *thread;
2146 
2147 	/* want main thread for process - has maps */
2148 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2149 	if (thread == NULL) {
2150 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2151 		return;
2152 	}
2153 
2154 	if (!sched->show_callchain || sample->callchain == NULL)
2155 		return;
2156 
2157 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2158 				      NULL, NULL, sched->max_stack + 2) != 0) {
2159 		if (verbose > 0)
2160 			pr_err("Failed to resolve callchain. Skipping\n");
2161 
2162 		return;
2163 	}
2164 
2165 	callchain_cursor_commit(cursor);
2166 
2167 	while (true) {
2168 		struct callchain_cursor_node *node;
2169 		struct symbol *sym;
2170 
2171 		node = callchain_cursor_current(cursor);
2172 		if (node == NULL)
2173 			break;
2174 
2175 		sym = node->sym;
2176 		if (sym) {
2177 			if (!strcmp(sym->name, "schedule") ||
2178 			    !strcmp(sym->name, "__schedule") ||
2179 			    !strcmp(sym->name, "preempt_schedule"))
2180 				sym->ignore = 1;
2181 		}
2182 
2183 		callchain_cursor_advance(cursor);
2184 	}
2185 }
2186 
2187 static int init_idle_thread(struct thread *thread)
2188 {
2189 	struct idle_thread_runtime *itr;
2190 
2191 	thread__set_comm(thread, idle_comm, 0);
2192 
2193 	itr = zalloc(sizeof(*itr));
2194 	if (itr == NULL)
2195 		return -ENOMEM;
2196 
2197 	init_stats(&itr->tr.run_stats);
2198 	callchain_init(&itr->callchain);
2199 	callchain_cursor_reset(&itr->cursor);
2200 	thread__set_priv(thread, itr);
2201 
2202 	return 0;
2203 }
2204 
2205 /*
2206  * Track idle stats per cpu by maintaining a local thread
2207  * struct for the idle task on each cpu.
2208  */
2209 static int init_idle_threads(int ncpu)
2210 {
2211 	int i, ret;
2212 
2213 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2214 	if (!idle_threads)
2215 		return -ENOMEM;
2216 
2217 	idle_max_cpu = ncpu;
2218 
2219 	/* allocate the actual thread struct if needed */
2220 	for (i = 0; i < ncpu; ++i) {
2221 		idle_threads[i] = thread__new(0, 0);
2222 		if (idle_threads[i] == NULL)
2223 			return -ENOMEM;
2224 
2225 		ret = init_idle_thread(idle_threads[i]);
2226 		if (ret < 0)
2227 			return ret;
2228 	}
2229 
2230 	return 0;
2231 }
2232 
2233 static void free_idle_threads(void)
2234 {
2235 	int i;
2236 
2237 	if (idle_threads == NULL)
2238 		return;
2239 
2240 	for (i = 0; i < idle_max_cpu; ++i) {
2241 		if ((idle_threads[i]))
2242 			thread__delete(idle_threads[i]);
2243 	}
2244 
2245 	free(idle_threads);
2246 }
2247 
2248 static struct thread *get_idle_thread(int cpu)
2249 {
2250 	/*
2251 	 * expand/allocate array of pointers to local thread
2252 	 * structs if needed
2253 	 */
2254 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2255 		int i, j = __roundup_pow_of_two(cpu+1);
2256 		void *p;
2257 
2258 		p = realloc(idle_threads, j * sizeof(struct thread *));
2259 		if (!p)
2260 			return NULL;
2261 
2262 		idle_threads = (struct thread **) p;
2263 		for (i = idle_max_cpu; i < j; ++i)
2264 			idle_threads[i] = NULL;
2265 
2266 		idle_max_cpu = j;
2267 	}
2268 
2269 	/* allocate a new thread struct if needed */
2270 	if (idle_threads[cpu] == NULL) {
2271 		idle_threads[cpu] = thread__new(0, 0);
2272 
2273 		if (idle_threads[cpu]) {
2274 			if (init_idle_thread(idle_threads[cpu]) < 0)
2275 				return NULL;
2276 		}
2277 	}
2278 
2279 	return idle_threads[cpu];
2280 }
2281 
2282 static void save_idle_callchain(struct perf_sched *sched,
2283 				struct idle_thread_runtime *itr,
2284 				struct perf_sample *sample)
2285 {
2286 	if (!sched->show_callchain || sample->callchain == NULL)
2287 		return;
2288 
2289 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2290 }
2291 
2292 static struct thread *timehist_get_thread(struct perf_sched *sched,
2293 					  struct perf_sample *sample,
2294 					  struct machine *machine,
2295 					  struct evsel *evsel)
2296 {
2297 	struct thread *thread;
2298 
2299 	if (is_idle_sample(sample, evsel)) {
2300 		thread = get_idle_thread(sample->cpu);
2301 		if (thread == NULL)
2302 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2303 
2304 	} else {
2305 		/* there were samples with tid 0 but non-zero pid */
2306 		thread = machine__findnew_thread(machine, sample->pid,
2307 						 sample->tid ?: sample->pid);
2308 		if (thread == NULL) {
2309 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2310 				 sample->tid);
2311 		}
2312 
2313 		save_task_callchain(sched, sample, evsel, machine);
2314 		if (sched->idle_hist) {
2315 			struct thread *idle;
2316 			struct idle_thread_runtime *itr;
2317 
2318 			idle = get_idle_thread(sample->cpu);
2319 			if (idle == NULL) {
2320 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2321 				return NULL;
2322 			}
2323 
2324 			itr = thread__priv(idle);
2325 			if (itr == NULL)
2326 				return NULL;
2327 
2328 			itr->last_thread = thread;
2329 
2330 			/* copy task callchain when entering to idle */
2331 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2332 				save_idle_callchain(sched, itr, sample);
2333 		}
2334 	}
2335 
2336 	return thread;
2337 }
2338 
2339 static bool timehist_skip_sample(struct perf_sched *sched,
2340 				 struct thread *thread,
2341 				 struct evsel *evsel,
2342 				 struct perf_sample *sample)
2343 {
2344 	bool rc = false;
2345 
2346 	if (thread__is_filtered(thread)) {
2347 		rc = true;
2348 		sched->skipped_samples++;
2349 	}
2350 
2351 	if (sched->idle_hist) {
2352 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2353 			rc = true;
2354 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2355 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2356 			rc = true;
2357 	}
2358 
2359 	return rc;
2360 }
2361 
2362 static void timehist_print_wakeup_event(struct perf_sched *sched,
2363 					struct evsel *evsel,
2364 					struct perf_sample *sample,
2365 					struct machine *machine,
2366 					struct thread *awakened)
2367 {
2368 	struct thread *thread;
2369 	char tstr[64];
2370 
2371 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2372 	if (thread == NULL)
2373 		return;
2374 
2375 	/* show wakeup unless both awakee and awaker are filtered */
2376 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2377 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2378 		return;
2379 	}
2380 
2381 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2382 	printf("%15s [%04d] ", tstr, sample->cpu);
2383 	if (sched->show_cpu_visual)
2384 		printf(" %*s ", sched->max_cpu + 1, "");
2385 
2386 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2387 
2388 	/* dt spacer */
2389 	printf("  %9s  %9s  %9s ", "", "", "");
2390 
2391 	printf("awakened: %s", timehist_get_commstr(awakened));
2392 
2393 	printf("\n");
2394 }
2395 
2396 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2397 				       union perf_event *event __maybe_unused,
2398 				       struct evsel *evsel,
2399 				       struct perf_sample *sample,
2400 				       struct machine *machine)
2401 {
2402 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2403 	struct thread *thread;
2404 	struct thread_runtime *tr = NULL;
2405 	/* want pid of awakened task not pid in sample */
2406 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2407 
2408 	thread = machine__findnew_thread(machine, 0, pid);
2409 	if (thread == NULL)
2410 		return -1;
2411 
2412 	tr = thread__get_runtime(thread);
2413 	if (tr == NULL)
2414 		return -1;
2415 
2416 	if (tr->ready_to_run == 0)
2417 		tr->ready_to_run = sample->time;
2418 
2419 	/* show wakeups if requested */
2420 	if (sched->show_wakeups &&
2421 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2422 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2423 
2424 	return 0;
2425 }
2426 
2427 static void timehist_print_migration_event(struct perf_sched *sched,
2428 					struct evsel *evsel,
2429 					struct perf_sample *sample,
2430 					struct machine *machine,
2431 					struct thread *migrated)
2432 {
2433 	struct thread *thread;
2434 	char tstr[64];
2435 	u32 max_cpus = sched->max_cpu + 1;
2436 	u32 ocpu, dcpu;
2437 
2438 	if (sched->summary_only)
2439 		return;
2440 
2441 	max_cpus = sched->max_cpu + 1;
2442 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2443 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2444 
2445 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2446 	if (thread == NULL)
2447 		return;
2448 
2449 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2450 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2451 		return;
2452 	}
2453 
2454 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2455 	printf("%15s [%04d] ", tstr, sample->cpu);
2456 
2457 	if (sched->show_cpu_visual) {
2458 		u32 i;
2459 		char c;
2460 
2461 		printf("  ");
2462 		for (i = 0; i < max_cpus; ++i) {
2463 			c = (i == sample->cpu) ? 'm' : ' ';
2464 			printf("%c", c);
2465 		}
2466 		printf("  ");
2467 	}
2468 
2469 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2470 
2471 	/* dt spacer */
2472 	printf("  %9s  %9s  %9s ", "", "", "");
2473 
2474 	printf("migrated: %s", timehist_get_commstr(migrated));
2475 	printf(" cpu %d => %d", ocpu, dcpu);
2476 
2477 	printf("\n");
2478 }
2479 
2480 static int timehist_migrate_task_event(struct perf_tool *tool,
2481 				       union perf_event *event __maybe_unused,
2482 				       struct evsel *evsel,
2483 				       struct perf_sample *sample,
2484 				       struct machine *machine)
2485 {
2486 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2487 	struct thread *thread;
2488 	struct thread_runtime *tr = NULL;
2489 	/* want pid of migrated task not pid in sample */
2490 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2491 
2492 	thread = machine__findnew_thread(machine, 0, pid);
2493 	if (thread == NULL)
2494 		return -1;
2495 
2496 	tr = thread__get_runtime(thread);
2497 	if (tr == NULL)
2498 		return -1;
2499 
2500 	tr->migrations++;
2501 
2502 	/* show migrations if requested */
2503 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2504 
2505 	return 0;
2506 }
2507 
2508 static int timehist_sched_change_event(struct perf_tool *tool,
2509 				       union perf_event *event,
2510 				       struct evsel *evsel,
2511 				       struct perf_sample *sample,
2512 				       struct machine *machine)
2513 {
2514 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2515 	struct perf_time_interval *ptime = &sched->ptime;
2516 	struct addr_location al;
2517 	struct thread *thread;
2518 	struct thread_runtime *tr = NULL;
2519 	u64 tprev, t = sample->time;
2520 	int rc = 0;
2521 	int state = perf_evsel__intval(evsel, sample, "prev_state");
2522 
2523 
2524 	if (machine__resolve(machine, &al, sample) < 0) {
2525 		pr_err("problem processing %d event. skipping it\n",
2526 		       event->header.type);
2527 		rc = -1;
2528 		goto out;
2529 	}
2530 
2531 	thread = timehist_get_thread(sched, sample, machine, evsel);
2532 	if (thread == NULL) {
2533 		rc = -1;
2534 		goto out;
2535 	}
2536 
2537 	if (timehist_skip_sample(sched, thread, evsel, sample))
2538 		goto out;
2539 
2540 	tr = thread__get_runtime(thread);
2541 	if (tr == NULL) {
2542 		rc = -1;
2543 		goto out;
2544 	}
2545 
2546 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2547 
2548 	/*
2549 	 * If start time given:
2550 	 * - sample time is under window user cares about - skip sample
2551 	 * - tprev is under window user cares about  - reset to start of window
2552 	 */
2553 	if (ptime->start && ptime->start > t)
2554 		goto out;
2555 
2556 	if (tprev && ptime->start > tprev)
2557 		tprev = ptime->start;
2558 
2559 	/*
2560 	 * If end time given:
2561 	 * - previous sched event is out of window - we are done
2562 	 * - sample time is beyond window user cares about - reset it
2563 	 *   to close out stats for time window interest
2564 	 */
2565 	if (ptime->end) {
2566 		if (tprev > ptime->end)
2567 			goto out;
2568 
2569 		if (t > ptime->end)
2570 			t = ptime->end;
2571 	}
2572 
2573 	if (!sched->idle_hist || thread->tid == 0) {
2574 		timehist_update_runtime_stats(tr, t, tprev);
2575 
2576 		if (sched->idle_hist) {
2577 			struct idle_thread_runtime *itr = (void *)tr;
2578 			struct thread_runtime *last_tr;
2579 
2580 			BUG_ON(thread->tid != 0);
2581 
2582 			if (itr->last_thread == NULL)
2583 				goto out;
2584 
2585 			/* add current idle time as last thread's runtime */
2586 			last_tr = thread__get_runtime(itr->last_thread);
2587 			if (last_tr == NULL)
2588 				goto out;
2589 
2590 			timehist_update_runtime_stats(last_tr, t, tprev);
2591 			/*
2592 			 * remove delta time of last thread as it's not updated
2593 			 * and otherwise it will show an invalid value next
2594 			 * time.  we only care total run time and run stat.
2595 			 */
2596 			last_tr->dt_run = 0;
2597 			last_tr->dt_delay = 0;
2598 			last_tr->dt_sleep = 0;
2599 			last_tr->dt_iowait = 0;
2600 			last_tr->dt_preempt = 0;
2601 
2602 			if (itr->cursor.nr)
2603 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2604 
2605 			itr->last_thread = NULL;
2606 		}
2607 	}
2608 
2609 	if (!sched->summary_only)
2610 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2611 
2612 out:
2613 	if (sched->hist_time.start == 0 && t >= ptime->start)
2614 		sched->hist_time.start = t;
2615 	if (ptime->end == 0 || t <= ptime->end)
2616 		sched->hist_time.end = t;
2617 
2618 	if (tr) {
2619 		/* time of this sched_switch event becomes last time task seen */
2620 		tr->last_time = sample->time;
2621 
2622 		/* last state is used to determine where to account wait time */
2623 		tr->last_state = state;
2624 
2625 		/* sched out event for task so reset ready to run time */
2626 		tr->ready_to_run = 0;
2627 	}
2628 
2629 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2630 
2631 	return rc;
2632 }
2633 
2634 static int timehist_sched_switch_event(struct perf_tool *tool,
2635 			     union perf_event *event,
2636 			     struct evsel *evsel,
2637 			     struct perf_sample *sample,
2638 			     struct machine *machine __maybe_unused)
2639 {
2640 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2641 }
2642 
2643 static int process_lost(struct perf_tool *tool __maybe_unused,
2644 			union perf_event *event,
2645 			struct perf_sample *sample,
2646 			struct machine *machine __maybe_unused)
2647 {
2648 	char tstr[64];
2649 
2650 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2651 	printf("%15s ", tstr);
2652 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2653 
2654 	return 0;
2655 }
2656 
2657 
2658 static void print_thread_runtime(struct thread *t,
2659 				 struct thread_runtime *r)
2660 {
2661 	double mean = avg_stats(&r->run_stats);
2662 	float stddev;
2663 
2664 	printf("%*s   %5d  %9" PRIu64 " ",
2665 	       comm_width, timehist_get_commstr(t), t->ppid,
2666 	       (u64) r->run_stats.n);
2667 
2668 	print_sched_time(r->total_run_time, 8);
2669 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2670 	print_sched_time(r->run_stats.min, 6);
2671 	printf(" ");
2672 	print_sched_time((u64) mean, 6);
2673 	printf(" ");
2674 	print_sched_time(r->run_stats.max, 6);
2675 	printf("  ");
2676 	printf("%5.2f", stddev);
2677 	printf("   %5" PRIu64, r->migrations);
2678 	printf("\n");
2679 }
2680 
2681 static void print_thread_waittime(struct thread *t,
2682 				  struct thread_runtime *r)
2683 {
2684 	printf("%*s   %5d  %9" PRIu64 " ",
2685 	       comm_width, timehist_get_commstr(t), t->ppid,
2686 	       (u64) r->run_stats.n);
2687 
2688 	print_sched_time(r->total_run_time, 8);
2689 	print_sched_time(r->total_sleep_time, 6);
2690 	printf(" ");
2691 	print_sched_time(r->total_iowait_time, 6);
2692 	printf(" ");
2693 	print_sched_time(r->total_preempt_time, 6);
2694 	printf(" ");
2695 	print_sched_time(r->total_delay_time, 6);
2696 	printf("\n");
2697 }
2698 
2699 struct total_run_stats {
2700 	struct perf_sched *sched;
2701 	u64  sched_count;
2702 	u64  task_count;
2703 	u64  total_run_time;
2704 };
2705 
2706 static int __show_thread_runtime(struct thread *t, void *priv)
2707 {
2708 	struct total_run_stats *stats = priv;
2709 	struct thread_runtime *r;
2710 
2711 	if (thread__is_filtered(t))
2712 		return 0;
2713 
2714 	r = thread__priv(t);
2715 	if (r && r->run_stats.n) {
2716 		stats->task_count++;
2717 		stats->sched_count += r->run_stats.n;
2718 		stats->total_run_time += r->total_run_time;
2719 
2720 		if (stats->sched->show_state)
2721 			print_thread_waittime(t, r);
2722 		else
2723 			print_thread_runtime(t, r);
2724 	}
2725 
2726 	return 0;
2727 }
2728 
2729 static int show_thread_runtime(struct thread *t, void *priv)
2730 {
2731 	if (t->dead)
2732 		return 0;
2733 
2734 	return __show_thread_runtime(t, priv);
2735 }
2736 
2737 static int show_deadthread_runtime(struct thread *t, void *priv)
2738 {
2739 	if (!t->dead)
2740 		return 0;
2741 
2742 	return __show_thread_runtime(t, priv);
2743 }
2744 
2745 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2746 {
2747 	const char *sep = " <- ";
2748 	struct callchain_list *chain;
2749 	size_t ret = 0;
2750 	char bf[1024];
2751 	bool first;
2752 
2753 	if (node == NULL)
2754 		return 0;
2755 
2756 	ret = callchain__fprintf_folded(fp, node->parent);
2757 	first = (ret == 0);
2758 
2759 	list_for_each_entry(chain, &node->val, list) {
2760 		if (chain->ip >= PERF_CONTEXT_MAX)
2761 			continue;
2762 		if (chain->ms.sym && chain->ms.sym->ignore)
2763 			continue;
2764 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2765 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2766 							false));
2767 		first = false;
2768 	}
2769 
2770 	return ret;
2771 }
2772 
2773 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2774 {
2775 	size_t ret = 0;
2776 	FILE *fp = stdout;
2777 	struct callchain_node *chain;
2778 	struct rb_node *rb_node = rb_first_cached(root);
2779 
2780 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2781 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2782 	       graph_dotted_line);
2783 
2784 	while (rb_node) {
2785 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2786 		rb_node = rb_next(rb_node);
2787 
2788 		ret += fprintf(fp, "  ");
2789 		print_sched_time(chain->hit, 12);
2790 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2791 		ret += fprintf(fp, " %8d  ", chain->count);
2792 		ret += callchain__fprintf_folded(fp, chain);
2793 		ret += fprintf(fp, "\n");
2794 	}
2795 
2796 	return ret;
2797 }
2798 
2799 static void timehist_print_summary(struct perf_sched *sched,
2800 				   struct perf_session *session)
2801 {
2802 	struct machine *m = &session->machines.host;
2803 	struct total_run_stats totals;
2804 	u64 task_count;
2805 	struct thread *t;
2806 	struct thread_runtime *r;
2807 	int i;
2808 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2809 
2810 	memset(&totals, 0, sizeof(totals));
2811 	totals.sched = sched;
2812 
2813 	if (sched->idle_hist) {
2814 		printf("\nIdle-time summary\n");
2815 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2816 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2817 	} else if (sched->show_state) {
2818 		printf("\nWait-time summary\n");
2819 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2820 		printf("   run-time      sleep      iowait     preempt       delay\n");
2821 	} else {
2822 		printf("\nRuntime summary\n");
2823 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2824 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2825 	}
2826 	printf("%*s            (count)  ", comm_width, "");
2827 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2828 	       sched->show_state ? "(msec)" : "%");
2829 	printf("%.117s\n", graph_dotted_line);
2830 
2831 	machine__for_each_thread(m, show_thread_runtime, &totals);
2832 	task_count = totals.task_count;
2833 	if (!task_count)
2834 		printf("<no still running tasks>\n");
2835 
2836 	printf("\nTerminated tasks:\n");
2837 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2838 	if (task_count == totals.task_count)
2839 		printf("<no terminated tasks>\n");
2840 
2841 	/* CPU idle stats not tracked when samples were skipped */
2842 	if (sched->skipped_samples && !sched->idle_hist)
2843 		return;
2844 
2845 	printf("\nIdle stats:\n");
2846 	for (i = 0; i < idle_max_cpu; ++i) {
2847 		t = idle_threads[i];
2848 		if (!t)
2849 			continue;
2850 
2851 		r = thread__priv(t);
2852 		if (r && r->run_stats.n) {
2853 			totals.sched_count += r->run_stats.n;
2854 			printf("    CPU %2d idle for ", i);
2855 			print_sched_time(r->total_run_time, 6);
2856 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2857 		} else
2858 			printf("    CPU %2d idle entire time window\n", i);
2859 	}
2860 
2861 	if (sched->idle_hist && sched->show_callchain) {
2862 		callchain_param.mode  = CHAIN_FOLDED;
2863 		callchain_param.value = CCVAL_PERIOD;
2864 
2865 		callchain_register_param(&callchain_param);
2866 
2867 		printf("\nIdle stats by callchain:\n");
2868 		for (i = 0; i < idle_max_cpu; ++i) {
2869 			struct idle_thread_runtime *itr;
2870 
2871 			t = idle_threads[i];
2872 			if (!t)
2873 				continue;
2874 
2875 			itr = thread__priv(t);
2876 			if (itr == NULL)
2877 				continue;
2878 
2879 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2880 					     0, &callchain_param);
2881 
2882 			printf("  CPU %2d:", i);
2883 			print_sched_time(itr->tr.total_run_time, 6);
2884 			printf(" msec\n");
2885 			timehist_print_idlehist_callchain(&itr->sorted_root);
2886 			printf("\n");
2887 		}
2888 	}
2889 
2890 	printf("\n"
2891 	       "    Total number of unique tasks: %" PRIu64 "\n"
2892 	       "Total number of context switches: %" PRIu64 "\n",
2893 	       totals.task_count, totals.sched_count);
2894 
2895 	printf("           Total run time (msec): ");
2896 	print_sched_time(totals.total_run_time, 2);
2897 	printf("\n");
2898 
2899 	printf("    Total scheduling time (msec): ");
2900 	print_sched_time(hist_time, 2);
2901 	printf(" (x %d)\n", sched->max_cpu);
2902 }
2903 
2904 typedef int (*sched_handler)(struct perf_tool *tool,
2905 			  union perf_event *event,
2906 			  struct evsel *evsel,
2907 			  struct perf_sample *sample,
2908 			  struct machine *machine);
2909 
2910 static int perf_timehist__process_sample(struct perf_tool *tool,
2911 					 union perf_event *event,
2912 					 struct perf_sample *sample,
2913 					 struct evsel *evsel,
2914 					 struct machine *machine)
2915 {
2916 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2917 	int err = 0;
2918 	int this_cpu = sample->cpu;
2919 
2920 	if (this_cpu > sched->max_cpu)
2921 		sched->max_cpu = this_cpu;
2922 
2923 	if (evsel->handler != NULL) {
2924 		sched_handler f = evsel->handler;
2925 
2926 		err = f(tool, event, evsel, sample, machine);
2927 	}
2928 
2929 	return err;
2930 }
2931 
2932 static int timehist_check_attr(struct perf_sched *sched,
2933 			       struct evlist *evlist)
2934 {
2935 	struct evsel *evsel;
2936 	struct evsel_runtime *er;
2937 
2938 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2939 		er = perf_evsel__get_runtime(evsel);
2940 		if (er == NULL) {
2941 			pr_err("Failed to allocate memory for evsel runtime data\n");
2942 			return -1;
2943 		}
2944 
2945 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2946 			pr_info("Samples do not have callchains.\n");
2947 			sched->show_callchain = 0;
2948 			symbol_conf.use_callchain = 0;
2949 		}
2950 	}
2951 
2952 	return 0;
2953 }
2954 
2955 static int perf_sched__timehist(struct perf_sched *sched)
2956 {
2957 	const struct evsel_str_handler handlers[] = {
2958 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2959 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2960 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2961 	};
2962 	const struct evsel_str_handler migrate_handlers[] = {
2963 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2964 	};
2965 	struct perf_data data = {
2966 		.path  = input_name,
2967 		.mode  = PERF_DATA_MODE_READ,
2968 		.force = sched->force,
2969 	};
2970 
2971 	struct perf_session *session;
2972 	struct evlist *evlist;
2973 	int err = -1;
2974 
2975 	/*
2976 	 * event handlers for timehist option
2977 	 */
2978 	sched->tool.sample	 = perf_timehist__process_sample;
2979 	sched->tool.mmap	 = perf_event__process_mmap;
2980 	sched->tool.comm	 = perf_event__process_comm;
2981 	sched->tool.exit	 = perf_event__process_exit;
2982 	sched->tool.fork	 = perf_event__process_fork;
2983 	sched->tool.lost	 = process_lost;
2984 	sched->tool.attr	 = perf_event__process_attr;
2985 	sched->tool.tracing_data = perf_event__process_tracing_data;
2986 	sched->tool.build_id	 = perf_event__process_build_id;
2987 
2988 	sched->tool.ordered_events = true;
2989 	sched->tool.ordering_requires_timestamps = true;
2990 
2991 	symbol_conf.use_callchain = sched->show_callchain;
2992 
2993 	session = perf_session__new(&data, false, &sched->tool);
2994 	if (IS_ERR(session))
2995 		return PTR_ERR(session);
2996 
2997 	evlist = session->evlist;
2998 
2999 	symbol__init(&session->header.env);
3000 
3001 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3002 		pr_err("Invalid time string\n");
3003 		return -EINVAL;
3004 	}
3005 
3006 	if (timehist_check_attr(sched, evlist) != 0)
3007 		goto out;
3008 
3009 	setup_pager();
3010 
3011 	/* setup per-evsel handlers */
3012 	if (perf_session__set_tracepoints_handlers(session, handlers))
3013 		goto out;
3014 
3015 	/* sched_switch event at a minimum needs to exist */
3016 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3017 						  "sched:sched_switch")) {
3018 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3019 		goto out;
3020 	}
3021 
3022 	if (sched->show_migrations &&
3023 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3024 		goto out;
3025 
3026 	/* pre-allocate struct for per-CPU idle stats */
3027 	sched->max_cpu = session->header.env.nr_cpus_online;
3028 	if (sched->max_cpu == 0)
3029 		sched->max_cpu = 4;
3030 	if (init_idle_threads(sched->max_cpu))
3031 		goto out;
3032 
3033 	/* summary_only implies summary option, but don't overwrite summary if set */
3034 	if (sched->summary_only)
3035 		sched->summary = sched->summary_only;
3036 
3037 	if (!sched->summary_only)
3038 		timehist_header(sched);
3039 
3040 	err = perf_session__process_events(session);
3041 	if (err) {
3042 		pr_err("Failed to process events, error %d", err);
3043 		goto out;
3044 	}
3045 
3046 	sched->nr_events      = evlist->stats.nr_events[0];
3047 	sched->nr_lost_events = evlist->stats.total_lost;
3048 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3049 
3050 	if (sched->summary)
3051 		timehist_print_summary(sched, session);
3052 
3053 out:
3054 	free_idle_threads();
3055 	perf_session__delete(session);
3056 
3057 	return err;
3058 }
3059 
3060 
3061 static void print_bad_events(struct perf_sched *sched)
3062 {
3063 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3064 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3065 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3066 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3067 	}
3068 	if (sched->nr_lost_events && sched->nr_events) {
3069 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3070 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3071 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3072 	}
3073 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3074 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3075 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3076 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3077 		if (sched->nr_lost_events)
3078 			printf(" (due to lost events?)");
3079 		printf("\n");
3080 	}
3081 }
3082 
3083 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3084 {
3085 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3086 	struct work_atoms *this;
3087 	const char *comm = thread__comm_str(data->thread), *this_comm;
3088 	bool leftmost = true;
3089 
3090 	while (*new) {
3091 		int cmp;
3092 
3093 		this = container_of(*new, struct work_atoms, node);
3094 		parent = *new;
3095 
3096 		this_comm = thread__comm_str(this->thread);
3097 		cmp = strcmp(comm, this_comm);
3098 		if (cmp > 0) {
3099 			new = &((*new)->rb_left);
3100 		} else if (cmp < 0) {
3101 			new = &((*new)->rb_right);
3102 			leftmost = false;
3103 		} else {
3104 			this->num_merged++;
3105 			this->total_runtime += data->total_runtime;
3106 			this->nb_atoms += data->nb_atoms;
3107 			this->total_lat += data->total_lat;
3108 			list_splice(&data->work_list, &this->work_list);
3109 			if (this->max_lat < data->max_lat) {
3110 				this->max_lat = data->max_lat;
3111 				this->max_lat_at = data->max_lat_at;
3112 			}
3113 			zfree(&data);
3114 			return;
3115 		}
3116 	}
3117 
3118 	data->num_merged++;
3119 	rb_link_node(&data->node, parent, new);
3120 	rb_insert_color_cached(&data->node, root, leftmost);
3121 }
3122 
3123 static void perf_sched__merge_lat(struct perf_sched *sched)
3124 {
3125 	struct work_atoms *data;
3126 	struct rb_node *node;
3127 
3128 	if (sched->skip_merge)
3129 		return;
3130 
3131 	while ((node = rb_first_cached(&sched->atom_root))) {
3132 		rb_erase_cached(node, &sched->atom_root);
3133 		data = rb_entry(node, struct work_atoms, node);
3134 		__merge_work_atoms(&sched->merged_atom_root, data);
3135 	}
3136 }
3137 
3138 static int perf_sched__lat(struct perf_sched *sched)
3139 {
3140 	struct rb_node *next;
3141 
3142 	setup_pager();
3143 
3144 	if (perf_sched__read_events(sched))
3145 		return -1;
3146 
3147 	perf_sched__merge_lat(sched);
3148 	perf_sched__sort_lat(sched);
3149 
3150 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3151 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3152 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3153 
3154 	next = rb_first_cached(&sched->sorted_atom_root);
3155 
3156 	while (next) {
3157 		struct work_atoms *work_list;
3158 
3159 		work_list = rb_entry(next, struct work_atoms, node);
3160 		output_lat_thread(sched, work_list);
3161 		next = rb_next(next);
3162 		thread__zput(work_list->thread);
3163 	}
3164 
3165 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3166 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3167 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3168 
3169 	printf(" ---------------------------------------------------\n");
3170 
3171 	print_bad_events(sched);
3172 	printf("\n");
3173 
3174 	return 0;
3175 }
3176 
3177 static int setup_map_cpus(struct perf_sched *sched)
3178 {
3179 	struct perf_cpu_map *map;
3180 
3181 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3182 
3183 	if (sched->map.comp) {
3184 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3185 		if (!sched->map.comp_cpus)
3186 			return -1;
3187 	}
3188 
3189 	if (!sched->map.cpus_str)
3190 		return 0;
3191 
3192 	map = perf_cpu_map__new(sched->map.cpus_str);
3193 	if (!map) {
3194 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3195 		return -1;
3196 	}
3197 
3198 	sched->map.cpus = map;
3199 	return 0;
3200 }
3201 
3202 static int setup_color_pids(struct perf_sched *sched)
3203 {
3204 	struct perf_thread_map *map;
3205 
3206 	if (!sched->map.color_pids_str)
3207 		return 0;
3208 
3209 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3210 	if (!map) {
3211 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3212 		return -1;
3213 	}
3214 
3215 	sched->map.color_pids = map;
3216 	return 0;
3217 }
3218 
3219 static int setup_color_cpus(struct perf_sched *sched)
3220 {
3221 	struct perf_cpu_map *map;
3222 
3223 	if (!sched->map.color_cpus_str)
3224 		return 0;
3225 
3226 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3227 	if (!map) {
3228 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3229 		return -1;
3230 	}
3231 
3232 	sched->map.color_cpus = map;
3233 	return 0;
3234 }
3235 
3236 static int perf_sched__map(struct perf_sched *sched)
3237 {
3238 	if (setup_map_cpus(sched))
3239 		return -1;
3240 
3241 	if (setup_color_pids(sched))
3242 		return -1;
3243 
3244 	if (setup_color_cpus(sched))
3245 		return -1;
3246 
3247 	setup_pager();
3248 	if (perf_sched__read_events(sched))
3249 		return -1;
3250 	print_bad_events(sched);
3251 	return 0;
3252 }
3253 
3254 static int perf_sched__replay(struct perf_sched *sched)
3255 {
3256 	unsigned long i;
3257 
3258 	calibrate_run_measurement_overhead(sched);
3259 	calibrate_sleep_measurement_overhead(sched);
3260 
3261 	test_calibrations(sched);
3262 
3263 	if (perf_sched__read_events(sched))
3264 		return -1;
3265 
3266 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3267 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3268 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3269 
3270 	if (sched->targetless_wakeups)
3271 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3272 	if (sched->multitarget_wakeups)
3273 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3274 	if (sched->nr_run_events_optimized)
3275 		printf("run atoms optimized: %ld\n",
3276 			sched->nr_run_events_optimized);
3277 
3278 	print_task_traces(sched);
3279 	add_cross_task_wakeups(sched);
3280 
3281 	create_tasks(sched);
3282 	printf("------------------------------------------------------------\n");
3283 	for (i = 0; i < sched->replay_repeat; i++)
3284 		run_one_test(sched);
3285 
3286 	return 0;
3287 }
3288 
3289 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3290 			  const char * const usage_msg[])
3291 {
3292 	char *tmp, *tok, *str = strdup(sched->sort_order);
3293 
3294 	for (tok = strtok_r(str, ", ", &tmp);
3295 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3296 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3297 			usage_with_options_msg(usage_msg, options,
3298 					"Unknown --sort key: `%s'", tok);
3299 		}
3300 	}
3301 
3302 	free(str);
3303 
3304 	sort_dimension__add("pid", &sched->cmp_pid);
3305 }
3306 
3307 static int __cmd_record(int argc, const char **argv)
3308 {
3309 	unsigned int rec_argc, i, j;
3310 	const char **rec_argv;
3311 	const char * const record_args[] = {
3312 		"record",
3313 		"-a",
3314 		"-R",
3315 		"-m", "1024",
3316 		"-c", "1",
3317 		"-e", "sched:sched_switch",
3318 		"-e", "sched:sched_stat_wait",
3319 		"-e", "sched:sched_stat_sleep",
3320 		"-e", "sched:sched_stat_iowait",
3321 		"-e", "sched:sched_stat_runtime",
3322 		"-e", "sched:sched_process_fork",
3323 		"-e", "sched:sched_wakeup",
3324 		"-e", "sched:sched_wakeup_new",
3325 		"-e", "sched:sched_migrate_task",
3326 	};
3327 
3328 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3329 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3330 
3331 	if (rec_argv == NULL)
3332 		return -ENOMEM;
3333 
3334 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3335 		rec_argv[i] = strdup(record_args[i]);
3336 
3337 	for (j = 1; j < (unsigned int)argc; j++, i++)
3338 		rec_argv[i] = argv[j];
3339 
3340 	BUG_ON(i != rec_argc);
3341 
3342 	return cmd_record(i, rec_argv);
3343 }
3344 
3345 int cmd_sched(int argc, const char **argv)
3346 {
3347 	static const char default_sort_order[] = "avg, max, switch, runtime";
3348 	struct perf_sched sched = {
3349 		.tool = {
3350 			.sample		 = perf_sched__process_tracepoint_sample,
3351 			.comm		 = perf_sched__process_comm,
3352 			.namespaces	 = perf_event__process_namespaces,
3353 			.lost		 = perf_event__process_lost,
3354 			.fork		 = perf_sched__process_fork_event,
3355 			.ordered_events = true,
3356 		},
3357 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3358 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3359 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3360 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3361 		.sort_order	      = default_sort_order,
3362 		.replay_repeat	      = 10,
3363 		.profile_cpu	      = -1,
3364 		.next_shortname1      = 'A',
3365 		.next_shortname2      = '0',
3366 		.skip_merge           = 0,
3367 		.show_callchain	      = 1,
3368 		.max_stack            = 5,
3369 	};
3370 	const struct option sched_options[] = {
3371 	OPT_STRING('i', "input", &input_name, "file",
3372 		    "input file name"),
3373 	OPT_INCR('v', "verbose", &verbose,
3374 		    "be more verbose (show symbol address, etc)"),
3375 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3376 		    "dump raw trace in ASCII"),
3377 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3378 	OPT_END()
3379 	};
3380 	const struct option latency_options[] = {
3381 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3382 		   "sort by key(s): runtime, switch, avg, max"),
3383 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3384 		    "CPU to profile on"),
3385 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3386 		    "latency stats per pid instead of per comm"),
3387 	OPT_PARENT(sched_options)
3388 	};
3389 	const struct option replay_options[] = {
3390 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3391 		     "repeat the workload replay N times (-1: infinite)"),
3392 	OPT_PARENT(sched_options)
3393 	};
3394 	const struct option map_options[] = {
3395 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3396 		    "map output in compact mode"),
3397 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3398 		   "highlight given pids in map"),
3399 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3400                     "highlight given CPUs in map"),
3401 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3402                     "display given CPUs in map"),
3403 	OPT_PARENT(sched_options)
3404 	};
3405 	const struct option timehist_options[] = {
3406 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3407 		   "file", "vmlinux pathname"),
3408 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3409 		   "file", "kallsyms pathname"),
3410 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3411 		    "Display call chains if present (default on)"),
3412 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3413 		   "Maximum number of functions to display backtrace."),
3414 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3415 		    "Look for files with symbols relative to this directory"),
3416 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3417 		    "Show only syscall summary with statistics"),
3418 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3419 		    "Show all syscalls and summary with statistics"),
3420 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3421 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3422 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3423 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3424 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3425 	OPT_STRING(0, "time", &sched.time_str, "str",
3426 		   "Time span for analysis (start,stop)"),
3427 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3428 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3429 		   "analyze events only for given process id(s)"),
3430 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3431 		   "analyze events only for given thread id(s)"),
3432 	OPT_PARENT(sched_options)
3433 	};
3434 
3435 	const char * const latency_usage[] = {
3436 		"perf sched latency [<options>]",
3437 		NULL
3438 	};
3439 	const char * const replay_usage[] = {
3440 		"perf sched replay [<options>]",
3441 		NULL
3442 	};
3443 	const char * const map_usage[] = {
3444 		"perf sched map [<options>]",
3445 		NULL
3446 	};
3447 	const char * const timehist_usage[] = {
3448 		"perf sched timehist [<options>]",
3449 		NULL
3450 	};
3451 	const char *const sched_subcommands[] = { "record", "latency", "map",
3452 						  "replay", "script",
3453 						  "timehist", NULL };
3454 	const char *sched_usage[] = {
3455 		NULL,
3456 		NULL
3457 	};
3458 	struct trace_sched_handler lat_ops  = {
3459 		.wakeup_event	    = latency_wakeup_event,
3460 		.switch_event	    = latency_switch_event,
3461 		.runtime_event	    = latency_runtime_event,
3462 		.migrate_task_event = latency_migrate_task_event,
3463 	};
3464 	struct trace_sched_handler map_ops  = {
3465 		.switch_event	    = map_switch_event,
3466 	};
3467 	struct trace_sched_handler replay_ops  = {
3468 		.wakeup_event	    = replay_wakeup_event,
3469 		.switch_event	    = replay_switch_event,
3470 		.fork_event	    = replay_fork_event,
3471 	};
3472 	unsigned int i;
3473 
3474 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3475 		sched.curr_pid[i] = -1;
3476 
3477 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3478 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3479 	if (!argc)
3480 		usage_with_options(sched_usage, sched_options);
3481 
3482 	/*
3483 	 * Aliased to 'perf script' for now:
3484 	 */
3485 	if (!strcmp(argv[0], "script"))
3486 		return cmd_script(argc, argv);
3487 
3488 	if (!strncmp(argv[0], "rec", 3)) {
3489 		return __cmd_record(argc, argv);
3490 	} else if (!strncmp(argv[0], "lat", 3)) {
3491 		sched.tp_handler = &lat_ops;
3492 		if (argc > 1) {
3493 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3494 			if (argc)
3495 				usage_with_options(latency_usage, latency_options);
3496 		}
3497 		setup_sorting(&sched, latency_options, latency_usage);
3498 		return perf_sched__lat(&sched);
3499 	} else if (!strcmp(argv[0], "map")) {
3500 		if (argc) {
3501 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3502 			if (argc)
3503 				usage_with_options(map_usage, map_options);
3504 		}
3505 		sched.tp_handler = &map_ops;
3506 		setup_sorting(&sched, latency_options, latency_usage);
3507 		return perf_sched__map(&sched);
3508 	} else if (!strncmp(argv[0], "rep", 3)) {
3509 		sched.tp_handler = &replay_ops;
3510 		if (argc) {
3511 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3512 			if (argc)
3513 				usage_with_options(replay_usage, replay_options);
3514 		}
3515 		return perf_sched__replay(&sched);
3516 	} else if (!strcmp(argv[0], "timehist")) {
3517 		if (argc) {
3518 			argc = parse_options(argc, argv, timehist_options,
3519 					     timehist_usage, 0);
3520 			if (argc)
3521 				usage_with_options(timehist_usage, timehist_options);
3522 		}
3523 		if ((sched.show_wakeups || sched.show_next) &&
3524 		    sched.summary_only) {
3525 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3526 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3527 			if (sched.show_wakeups)
3528 				parse_options_usage(NULL, timehist_options, "w", true);
3529 			if (sched.show_next)
3530 				parse_options_usage(NULL, timehist_options, "n", true);
3531 			return -EINVAL;
3532 		}
3533 
3534 		return perf_sched__timehist(&sched);
3535 	} else {
3536 		usage_with_options(sched_usage, sched_options);
3537 	}
3538 
3539 	return 0;
3540 }
3541