xref: /openbmc/linux/tools/perf/builtin-sched.c (revision b593bce5)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 
5 #include "util/evlist.h"
6 #include "util/cache.h"
7 #include "util/evsel.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/tool.h"
13 #include "util/cloexec.h"
14 #include "util/thread_map.h"
15 #include "util/color.h"
16 #include "util/stat.h"
17 #include "util/string2.h"
18 #include "util/callchain.h"
19 #include "util/time-utils.h"
20 
21 #include <subcmd/parse-options.h>
22 #include "util/trace-event.h"
23 
24 #include "util/debug.h"
25 
26 #include <linux/kernel.h>
27 #include <linux/log2.h>
28 #include <linux/zalloc.h>
29 #include <sys/prctl.h>
30 #include <sys/resource.h>
31 #include <inttypes.h>
32 
33 #include <errno.h>
34 #include <semaphore.h>
35 #include <pthread.h>
36 #include <math.h>
37 #include <api/fs/fs.h>
38 #include <linux/time64.h>
39 
40 #include <linux/ctype.h>
41 
42 #define PR_SET_NAME		15               /* Set process name */
43 #define MAX_CPUS		4096
44 #define COMM_LEN		20
45 #define SYM_LEN			129
46 #define MAX_PID			1024000
47 
48 struct sched_atom;
49 
50 struct task_desc {
51 	unsigned long		nr;
52 	unsigned long		pid;
53 	char			comm[COMM_LEN];
54 
55 	unsigned long		nr_events;
56 	unsigned long		curr_event;
57 	struct sched_atom	**atoms;
58 
59 	pthread_t		thread;
60 	sem_t			sleep_sem;
61 
62 	sem_t			ready_for_work;
63 	sem_t			work_done_sem;
64 
65 	u64			cpu_usage;
66 };
67 
68 enum sched_event_type {
69 	SCHED_EVENT_RUN,
70 	SCHED_EVENT_SLEEP,
71 	SCHED_EVENT_WAKEUP,
72 	SCHED_EVENT_MIGRATION,
73 };
74 
75 struct sched_atom {
76 	enum sched_event_type	type;
77 	int			specific_wait;
78 	u64			timestamp;
79 	u64			duration;
80 	unsigned long		nr;
81 	sem_t			*wait_sem;
82 	struct task_desc	*wakee;
83 };
84 
85 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
86 
87 /* task state bitmask, copied from include/linux/sched.h */
88 #define TASK_RUNNING		0
89 #define TASK_INTERRUPTIBLE	1
90 #define TASK_UNINTERRUPTIBLE	2
91 #define __TASK_STOPPED		4
92 #define __TASK_TRACED		8
93 /* in tsk->exit_state */
94 #define EXIT_DEAD		16
95 #define EXIT_ZOMBIE		32
96 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
97 /* in tsk->state again */
98 #define TASK_DEAD		64
99 #define TASK_WAKEKILL		128
100 #define TASK_WAKING		256
101 #define TASK_PARKED		512
102 
103 enum thread_state {
104 	THREAD_SLEEPING = 0,
105 	THREAD_WAIT_CPU,
106 	THREAD_SCHED_IN,
107 	THREAD_IGNORE
108 };
109 
110 struct work_atom {
111 	struct list_head	list;
112 	enum thread_state	state;
113 	u64			sched_out_time;
114 	u64			wake_up_time;
115 	u64			sched_in_time;
116 	u64			runtime;
117 };
118 
119 struct work_atoms {
120 	struct list_head	work_list;
121 	struct thread		*thread;
122 	struct rb_node		node;
123 	u64			max_lat;
124 	u64			max_lat_at;
125 	u64			total_lat;
126 	u64			nb_atoms;
127 	u64			total_runtime;
128 	int			num_merged;
129 };
130 
131 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
132 
133 struct perf_sched;
134 
135 struct trace_sched_handler {
136 	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
137 			    struct perf_sample *sample, struct machine *machine);
138 
139 	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
140 			     struct perf_sample *sample, struct machine *machine);
141 
142 	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
143 			    struct perf_sample *sample, struct machine *machine);
144 
145 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
146 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
147 			  struct machine *machine);
148 
149 	int (*migrate_task_event)(struct perf_sched *sched,
150 				  struct perf_evsel *evsel,
151 				  struct perf_sample *sample,
152 				  struct machine *machine);
153 };
154 
155 #define COLOR_PIDS PERF_COLOR_BLUE
156 #define COLOR_CPUS PERF_COLOR_BG_RED
157 
158 struct perf_sched_map {
159 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
160 	int			*comp_cpus;
161 	bool			 comp;
162 	struct thread_map	*color_pids;
163 	const char		*color_pids_str;
164 	struct cpu_map		*color_cpus;
165 	const char		*color_cpus_str;
166 	struct cpu_map		*cpus;
167 	const char		*cpus_str;
168 };
169 
170 struct perf_sched {
171 	struct perf_tool tool;
172 	const char	 *sort_order;
173 	unsigned long	 nr_tasks;
174 	struct task_desc **pid_to_task;
175 	struct task_desc **tasks;
176 	const struct trace_sched_handler *tp_handler;
177 	pthread_mutex_t	 start_work_mutex;
178 	pthread_mutex_t	 work_done_wait_mutex;
179 	int		 profile_cpu;
180 /*
181  * Track the current task - that way we can know whether there's any
182  * weird events, such as a task being switched away that is not current.
183  */
184 	int		 max_cpu;
185 	u32		 curr_pid[MAX_CPUS];
186 	struct thread	 *curr_thread[MAX_CPUS];
187 	char		 next_shortname1;
188 	char		 next_shortname2;
189 	unsigned int	 replay_repeat;
190 	unsigned long	 nr_run_events;
191 	unsigned long	 nr_sleep_events;
192 	unsigned long	 nr_wakeup_events;
193 	unsigned long	 nr_sleep_corrections;
194 	unsigned long	 nr_run_events_optimized;
195 	unsigned long	 targetless_wakeups;
196 	unsigned long	 multitarget_wakeups;
197 	unsigned long	 nr_runs;
198 	unsigned long	 nr_timestamps;
199 	unsigned long	 nr_unordered_timestamps;
200 	unsigned long	 nr_context_switch_bugs;
201 	unsigned long	 nr_events;
202 	unsigned long	 nr_lost_chunks;
203 	unsigned long	 nr_lost_events;
204 	u64		 run_measurement_overhead;
205 	u64		 sleep_measurement_overhead;
206 	u64		 start_time;
207 	u64		 cpu_usage;
208 	u64		 runavg_cpu_usage;
209 	u64		 parent_cpu_usage;
210 	u64		 runavg_parent_cpu_usage;
211 	u64		 sum_runtime;
212 	u64		 sum_fluct;
213 	u64		 run_avg;
214 	u64		 all_runtime;
215 	u64		 all_count;
216 	u64		 cpu_last_switched[MAX_CPUS];
217 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
218 	struct list_head sort_list, cmp_pid;
219 	bool force;
220 	bool skip_merge;
221 	struct perf_sched_map map;
222 
223 	/* options for timehist command */
224 	bool		summary;
225 	bool		summary_only;
226 	bool		idle_hist;
227 	bool		show_callchain;
228 	unsigned int	max_stack;
229 	bool		show_cpu_visual;
230 	bool		show_wakeups;
231 	bool		show_next;
232 	bool		show_migrations;
233 	bool		show_state;
234 	u64		skipped_samples;
235 	const char	*time_str;
236 	struct perf_time_interval ptime;
237 	struct perf_time_interval hist_time;
238 };
239 
240 /* per thread run time data */
241 struct thread_runtime {
242 	u64 last_time;      /* time of previous sched in/out event */
243 	u64 dt_run;         /* run time */
244 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
245 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
246 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
247 	u64 dt_delay;       /* time between wakeup and sched-in */
248 	u64 ready_to_run;   /* time of wakeup */
249 
250 	struct stats run_stats;
251 	u64 total_run_time;
252 	u64 total_sleep_time;
253 	u64 total_iowait_time;
254 	u64 total_preempt_time;
255 	u64 total_delay_time;
256 
257 	int last_state;
258 
259 	char shortname[3];
260 	bool comm_changed;
261 
262 	u64 migrations;
263 };
264 
265 /* per event run time data */
266 struct evsel_runtime {
267 	u64 *last_time; /* time this event was last seen per cpu */
268 	u32 ncpu;       /* highest cpu slot allocated */
269 };
270 
271 /* per cpu idle time data */
272 struct idle_thread_runtime {
273 	struct thread_runtime	tr;
274 	struct thread		*last_thread;
275 	struct rb_root_cached	sorted_root;
276 	struct callchain_root	callchain;
277 	struct callchain_cursor	cursor;
278 };
279 
280 /* track idle times per cpu */
281 static struct thread **idle_threads;
282 static int idle_max_cpu;
283 static char idle_comm[] = "<idle>";
284 
285 static u64 get_nsecs(void)
286 {
287 	struct timespec ts;
288 
289 	clock_gettime(CLOCK_MONOTONIC, &ts);
290 
291 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
292 }
293 
294 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
295 {
296 	u64 T0 = get_nsecs(), T1;
297 
298 	do {
299 		T1 = get_nsecs();
300 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
301 }
302 
303 static void sleep_nsecs(u64 nsecs)
304 {
305 	struct timespec ts;
306 
307 	ts.tv_nsec = nsecs % 999999999;
308 	ts.tv_sec = nsecs / 999999999;
309 
310 	nanosleep(&ts, NULL);
311 }
312 
313 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
314 {
315 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
316 	int i;
317 
318 	for (i = 0; i < 10; i++) {
319 		T0 = get_nsecs();
320 		burn_nsecs(sched, 0);
321 		T1 = get_nsecs();
322 		delta = T1-T0;
323 		min_delta = min(min_delta, delta);
324 	}
325 	sched->run_measurement_overhead = min_delta;
326 
327 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
328 }
329 
330 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
331 {
332 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
333 	int i;
334 
335 	for (i = 0; i < 10; i++) {
336 		T0 = get_nsecs();
337 		sleep_nsecs(10000);
338 		T1 = get_nsecs();
339 		delta = T1-T0;
340 		min_delta = min(min_delta, delta);
341 	}
342 	min_delta -= 10000;
343 	sched->sleep_measurement_overhead = min_delta;
344 
345 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
346 }
347 
348 static struct sched_atom *
349 get_new_event(struct task_desc *task, u64 timestamp)
350 {
351 	struct sched_atom *event = zalloc(sizeof(*event));
352 	unsigned long idx = task->nr_events;
353 	size_t size;
354 
355 	event->timestamp = timestamp;
356 	event->nr = idx;
357 
358 	task->nr_events++;
359 	size = sizeof(struct sched_atom *) * task->nr_events;
360 	task->atoms = realloc(task->atoms, size);
361 	BUG_ON(!task->atoms);
362 
363 	task->atoms[idx] = event;
364 
365 	return event;
366 }
367 
368 static struct sched_atom *last_event(struct task_desc *task)
369 {
370 	if (!task->nr_events)
371 		return NULL;
372 
373 	return task->atoms[task->nr_events - 1];
374 }
375 
376 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
377 				u64 timestamp, u64 duration)
378 {
379 	struct sched_atom *event, *curr_event = last_event(task);
380 
381 	/*
382 	 * optimize an existing RUN event by merging this one
383 	 * to it:
384 	 */
385 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
386 		sched->nr_run_events_optimized++;
387 		curr_event->duration += duration;
388 		return;
389 	}
390 
391 	event = get_new_event(task, timestamp);
392 
393 	event->type = SCHED_EVENT_RUN;
394 	event->duration = duration;
395 
396 	sched->nr_run_events++;
397 }
398 
399 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
400 				   u64 timestamp, struct task_desc *wakee)
401 {
402 	struct sched_atom *event, *wakee_event;
403 
404 	event = get_new_event(task, timestamp);
405 	event->type = SCHED_EVENT_WAKEUP;
406 	event->wakee = wakee;
407 
408 	wakee_event = last_event(wakee);
409 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
410 		sched->targetless_wakeups++;
411 		return;
412 	}
413 	if (wakee_event->wait_sem) {
414 		sched->multitarget_wakeups++;
415 		return;
416 	}
417 
418 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
419 	sem_init(wakee_event->wait_sem, 0, 0);
420 	wakee_event->specific_wait = 1;
421 	event->wait_sem = wakee_event->wait_sem;
422 
423 	sched->nr_wakeup_events++;
424 }
425 
426 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
427 				  u64 timestamp, u64 task_state __maybe_unused)
428 {
429 	struct sched_atom *event = get_new_event(task, timestamp);
430 
431 	event->type = SCHED_EVENT_SLEEP;
432 
433 	sched->nr_sleep_events++;
434 }
435 
436 static struct task_desc *register_pid(struct perf_sched *sched,
437 				      unsigned long pid, const char *comm)
438 {
439 	struct task_desc *task;
440 	static int pid_max;
441 
442 	if (sched->pid_to_task == NULL) {
443 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
444 			pid_max = MAX_PID;
445 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
446 	}
447 	if (pid >= (unsigned long)pid_max) {
448 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
449 			sizeof(struct task_desc *))) == NULL);
450 		while (pid >= (unsigned long)pid_max)
451 			sched->pid_to_task[pid_max++] = NULL;
452 	}
453 
454 	task = sched->pid_to_task[pid];
455 
456 	if (task)
457 		return task;
458 
459 	task = zalloc(sizeof(*task));
460 	task->pid = pid;
461 	task->nr = sched->nr_tasks;
462 	strcpy(task->comm, comm);
463 	/*
464 	 * every task starts in sleeping state - this gets ignored
465 	 * if there's no wakeup pointing to this sleep state:
466 	 */
467 	add_sched_event_sleep(sched, task, 0, 0);
468 
469 	sched->pid_to_task[pid] = task;
470 	sched->nr_tasks++;
471 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
472 	BUG_ON(!sched->tasks);
473 	sched->tasks[task->nr] = task;
474 
475 	if (verbose > 0)
476 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
477 
478 	return task;
479 }
480 
481 
482 static void print_task_traces(struct perf_sched *sched)
483 {
484 	struct task_desc *task;
485 	unsigned long i;
486 
487 	for (i = 0; i < sched->nr_tasks; i++) {
488 		task = sched->tasks[i];
489 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
490 			task->nr, task->comm, task->pid, task->nr_events);
491 	}
492 }
493 
494 static void add_cross_task_wakeups(struct perf_sched *sched)
495 {
496 	struct task_desc *task1, *task2;
497 	unsigned long i, j;
498 
499 	for (i = 0; i < sched->nr_tasks; i++) {
500 		task1 = sched->tasks[i];
501 		j = i + 1;
502 		if (j == sched->nr_tasks)
503 			j = 0;
504 		task2 = sched->tasks[j];
505 		add_sched_event_wakeup(sched, task1, 0, task2);
506 	}
507 }
508 
509 static void perf_sched__process_event(struct perf_sched *sched,
510 				      struct sched_atom *atom)
511 {
512 	int ret = 0;
513 
514 	switch (atom->type) {
515 		case SCHED_EVENT_RUN:
516 			burn_nsecs(sched, atom->duration);
517 			break;
518 		case SCHED_EVENT_SLEEP:
519 			if (atom->wait_sem)
520 				ret = sem_wait(atom->wait_sem);
521 			BUG_ON(ret);
522 			break;
523 		case SCHED_EVENT_WAKEUP:
524 			if (atom->wait_sem)
525 				ret = sem_post(atom->wait_sem);
526 			BUG_ON(ret);
527 			break;
528 		case SCHED_EVENT_MIGRATION:
529 			break;
530 		default:
531 			BUG_ON(1);
532 	}
533 }
534 
535 static u64 get_cpu_usage_nsec_parent(void)
536 {
537 	struct rusage ru;
538 	u64 sum;
539 	int err;
540 
541 	err = getrusage(RUSAGE_SELF, &ru);
542 	BUG_ON(err);
543 
544 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
545 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
546 
547 	return sum;
548 }
549 
550 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
551 {
552 	struct perf_event_attr attr;
553 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
554 	int fd;
555 	struct rlimit limit;
556 	bool need_privilege = false;
557 
558 	memset(&attr, 0, sizeof(attr));
559 
560 	attr.type = PERF_TYPE_SOFTWARE;
561 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
562 
563 force_again:
564 	fd = sys_perf_event_open(&attr, 0, -1, -1,
565 				 perf_event_open_cloexec_flag());
566 
567 	if (fd < 0) {
568 		if (errno == EMFILE) {
569 			if (sched->force) {
570 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
571 				limit.rlim_cur += sched->nr_tasks - cur_task;
572 				if (limit.rlim_cur > limit.rlim_max) {
573 					limit.rlim_max = limit.rlim_cur;
574 					need_privilege = true;
575 				}
576 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
577 					if (need_privilege && errno == EPERM)
578 						strcpy(info, "Need privilege\n");
579 				} else
580 					goto force_again;
581 			} else
582 				strcpy(info, "Have a try with -f option\n");
583 		}
584 		pr_err("Error: sys_perf_event_open() syscall returned "
585 		       "with %d (%s)\n%s", fd,
586 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
587 		exit(EXIT_FAILURE);
588 	}
589 	return fd;
590 }
591 
592 static u64 get_cpu_usage_nsec_self(int fd)
593 {
594 	u64 runtime;
595 	int ret;
596 
597 	ret = read(fd, &runtime, sizeof(runtime));
598 	BUG_ON(ret != sizeof(runtime));
599 
600 	return runtime;
601 }
602 
603 struct sched_thread_parms {
604 	struct task_desc  *task;
605 	struct perf_sched *sched;
606 	int fd;
607 };
608 
609 static void *thread_func(void *ctx)
610 {
611 	struct sched_thread_parms *parms = ctx;
612 	struct task_desc *this_task = parms->task;
613 	struct perf_sched *sched = parms->sched;
614 	u64 cpu_usage_0, cpu_usage_1;
615 	unsigned long i, ret;
616 	char comm2[22];
617 	int fd = parms->fd;
618 
619 	zfree(&parms);
620 
621 	sprintf(comm2, ":%s", this_task->comm);
622 	prctl(PR_SET_NAME, comm2);
623 	if (fd < 0)
624 		return NULL;
625 again:
626 	ret = sem_post(&this_task->ready_for_work);
627 	BUG_ON(ret);
628 	ret = pthread_mutex_lock(&sched->start_work_mutex);
629 	BUG_ON(ret);
630 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
631 	BUG_ON(ret);
632 
633 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
634 
635 	for (i = 0; i < this_task->nr_events; i++) {
636 		this_task->curr_event = i;
637 		perf_sched__process_event(sched, this_task->atoms[i]);
638 	}
639 
640 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
641 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
642 	ret = sem_post(&this_task->work_done_sem);
643 	BUG_ON(ret);
644 
645 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
646 	BUG_ON(ret);
647 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
648 	BUG_ON(ret);
649 
650 	goto again;
651 }
652 
653 static void create_tasks(struct perf_sched *sched)
654 {
655 	struct task_desc *task;
656 	pthread_attr_t attr;
657 	unsigned long i;
658 	int err;
659 
660 	err = pthread_attr_init(&attr);
661 	BUG_ON(err);
662 	err = pthread_attr_setstacksize(&attr,
663 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
664 	BUG_ON(err);
665 	err = pthread_mutex_lock(&sched->start_work_mutex);
666 	BUG_ON(err);
667 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
668 	BUG_ON(err);
669 	for (i = 0; i < sched->nr_tasks; i++) {
670 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
671 		BUG_ON(parms == NULL);
672 		parms->task = task = sched->tasks[i];
673 		parms->sched = sched;
674 		parms->fd = self_open_counters(sched, i);
675 		sem_init(&task->sleep_sem, 0, 0);
676 		sem_init(&task->ready_for_work, 0, 0);
677 		sem_init(&task->work_done_sem, 0, 0);
678 		task->curr_event = 0;
679 		err = pthread_create(&task->thread, &attr, thread_func, parms);
680 		BUG_ON(err);
681 	}
682 }
683 
684 static void wait_for_tasks(struct perf_sched *sched)
685 {
686 	u64 cpu_usage_0, cpu_usage_1;
687 	struct task_desc *task;
688 	unsigned long i, ret;
689 
690 	sched->start_time = get_nsecs();
691 	sched->cpu_usage = 0;
692 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
693 
694 	for (i = 0; i < sched->nr_tasks; i++) {
695 		task = sched->tasks[i];
696 		ret = sem_wait(&task->ready_for_work);
697 		BUG_ON(ret);
698 		sem_init(&task->ready_for_work, 0, 0);
699 	}
700 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
701 	BUG_ON(ret);
702 
703 	cpu_usage_0 = get_cpu_usage_nsec_parent();
704 
705 	pthread_mutex_unlock(&sched->start_work_mutex);
706 
707 	for (i = 0; i < sched->nr_tasks; i++) {
708 		task = sched->tasks[i];
709 		ret = sem_wait(&task->work_done_sem);
710 		BUG_ON(ret);
711 		sem_init(&task->work_done_sem, 0, 0);
712 		sched->cpu_usage += task->cpu_usage;
713 		task->cpu_usage = 0;
714 	}
715 
716 	cpu_usage_1 = get_cpu_usage_nsec_parent();
717 	if (!sched->runavg_cpu_usage)
718 		sched->runavg_cpu_usage = sched->cpu_usage;
719 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
720 
721 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
722 	if (!sched->runavg_parent_cpu_usage)
723 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
724 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
725 					 sched->parent_cpu_usage)/sched->replay_repeat;
726 
727 	ret = pthread_mutex_lock(&sched->start_work_mutex);
728 	BUG_ON(ret);
729 
730 	for (i = 0; i < sched->nr_tasks; i++) {
731 		task = sched->tasks[i];
732 		sem_init(&task->sleep_sem, 0, 0);
733 		task->curr_event = 0;
734 	}
735 }
736 
737 static void run_one_test(struct perf_sched *sched)
738 {
739 	u64 T0, T1, delta, avg_delta, fluct;
740 
741 	T0 = get_nsecs();
742 	wait_for_tasks(sched);
743 	T1 = get_nsecs();
744 
745 	delta = T1 - T0;
746 	sched->sum_runtime += delta;
747 	sched->nr_runs++;
748 
749 	avg_delta = sched->sum_runtime / sched->nr_runs;
750 	if (delta < avg_delta)
751 		fluct = avg_delta - delta;
752 	else
753 		fluct = delta - avg_delta;
754 	sched->sum_fluct += fluct;
755 	if (!sched->run_avg)
756 		sched->run_avg = delta;
757 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
758 
759 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
760 
761 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
762 
763 	printf("cpu: %0.2f / %0.2f",
764 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
765 
766 #if 0
767 	/*
768 	 * rusage statistics done by the parent, these are less
769 	 * accurate than the sched->sum_exec_runtime based statistics:
770 	 */
771 	printf(" [%0.2f / %0.2f]",
772 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
773 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
774 #endif
775 
776 	printf("\n");
777 
778 	if (sched->nr_sleep_corrections)
779 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
780 	sched->nr_sleep_corrections = 0;
781 }
782 
783 static void test_calibrations(struct perf_sched *sched)
784 {
785 	u64 T0, T1;
786 
787 	T0 = get_nsecs();
788 	burn_nsecs(sched, NSEC_PER_MSEC);
789 	T1 = get_nsecs();
790 
791 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
792 
793 	T0 = get_nsecs();
794 	sleep_nsecs(NSEC_PER_MSEC);
795 	T1 = get_nsecs();
796 
797 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
798 }
799 
800 static int
801 replay_wakeup_event(struct perf_sched *sched,
802 		    struct perf_evsel *evsel, struct perf_sample *sample,
803 		    struct machine *machine __maybe_unused)
804 {
805 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
806 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
807 	struct task_desc *waker, *wakee;
808 
809 	if (verbose > 0) {
810 		printf("sched_wakeup event %p\n", evsel);
811 
812 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
813 	}
814 
815 	waker = register_pid(sched, sample->tid, "<unknown>");
816 	wakee = register_pid(sched, pid, comm);
817 
818 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
819 	return 0;
820 }
821 
822 static int replay_switch_event(struct perf_sched *sched,
823 			       struct perf_evsel *evsel,
824 			       struct perf_sample *sample,
825 			       struct machine *machine __maybe_unused)
826 {
827 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
828 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
829 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
830 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
831 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
832 	struct task_desc *prev, __maybe_unused *next;
833 	u64 timestamp0, timestamp = sample->time;
834 	int cpu = sample->cpu;
835 	s64 delta;
836 
837 	if (verbose > 0)
838 		printf("sched_switch event %p\n", evsel);
839 
840 	if (cpu >= MAX_CPUS || cpu < 0)
841 		return 0;
842 
843 	timestamp0 = sched->cpu_last_switched[cpu];
844 	if (timestamp0)
845 		delta = timestamp - timestamp0;
846 	else
847 		delta = 0;
848 
849 	if (delta < 0) {
850 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
851 		return -1;
852 	}
853 
854 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
855 		 prev_comm, prev_pid, next_comm, next_pid, delta);
856 
857 	prev = register_pid(sched, prev_pid, prev_comm);
858 	next = register_pid(sched, next_pid, next_comm);
859 
860 	sched->cpu_last_switched[cpu] = timestamp;
861 
862 	add_sched_event_run(sched, prev, timestamp, delta);
863 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
864 
865 	return 0;
866 }
867 
868 static int replay_fork_event(struct perf_sched *sched,
869 			     union perf_event *event,
870 			     struct machine *machine)
871 {
872 	struct thread *child, *parent;
873 
874 	child = machine__findnew_thread(machine, event->fork.pid,
875 					event->fork.tid);
876 	parent = machine__findnew_thread(machine, event->fork.ppid,
877 					 event->fork.ptid);
878 
879 	if (child == NULL || parent == NULL) {
880 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
881 				 child, parent);
882 		goto out_put;
883 	}
884 
885 	if (verbose > 0) {
886 		printf("fork event\n");
887 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
888 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
889 	}
890 
891 	register_pid(sched, parent->tid, thread__comm_str(parent));
892 	register_pid(sched, child->tid, thread__comm_str(child));
893 out_put:
894 	thread__put(child);
895 	thread__put(parent);
896 	return 0;
897 }
898 
899 struct sort_dimension {
900 	const char		*name;
901 	sort_fn_t		cmp;
902 	struct list_head	list;
903 };
904 
905 /*
906  * handle runtime stats saved per thread
907  */
908 static struct thread_runtime *thread__init_runtime(struct thread *thread)
909 {
910 	struct thread_runtime *r;
911 
912 	r = zalloc(sizeof(struct thread_runtime));
913 	if (!r)
914 		return NULL;
915 
916 	init_stats(&r->run_stats);
917 	thread__set_priv(thread, r);
918 
919 	return r;
920 }
921 
922 static struct thread_runtime *thread__get_runtime(struct thread *thread)
923 {
924 	struct thread_runtime *tr;
925 
926 	tr = thread__priv(thread);
927 	if (tr == NULL) {
928 		tr = thread__init_runtime(thread);
929 		if (tr == NULL)
930 			pr_debug("Failed to malloc memory for runtime data.\n");
931 	}
932 
933 	return tr;
934 }
935 
936 static int
937 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
938 {
939 	struct sort_dimension *sort;
940 	int ret = 0;
941 
942 	BUG_ON(list_empty(list));
943 
944 	list_for_each_entry(sort, list, list) {
945 		ret = sort->cmp(l, r);
946 		if (ret)
947 			return ret;
948 	}
949 
950 	return ret;
951 }
952 
953 static struct work_atoms *
954 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
955 			 struct list_head *sort_list)
956 {
957 	struct rb_node *node = root->rb_root.rb_node;
958 	struct work_atoms key = { .thread = thread };
959 
960 	while (node) {
961 		struct work_atoms *atoms;
962 		int cmp;
963 
964 		atoms = container_of(node, struct work_atoms, node);
965 
966 		cmp = thread_lat_cmp(sort_list, &key, atoms);
967 		if (cmp > 0)
968 			node = node->rb_left;
969 		else if (cmp < 0)
970 			node = node->rb_right;
971 		else {
972 			BUG_ON(thread != atoms->thread);
973 			return atoms;
974 		}
975 	}
976 	return NULL;
977 }
978 
979 static void
980 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
981 			 struct list_head *sort_list)
982 {
983 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
984 	bool leftmost = true;
985 
986 	while (*new) {
987 		struct work_atoms *this;
988 		int cmp;
989 
990 		this = container_of(*new, struct work_atoms, node);
991 		parent = *new;
992 
993 		cmp = thread_lat_cmp(sort_list, data, this);
994 
995 		if (cmp > 0)
996 			new = &((*new)->rb_left);
997 		else {
998 			new = &((*new)->rb_right);
999 			leftmost = false;
1000 		}
1001 	}
1002 
1003 	rb_link_node(&data->node, parent, new);
1004 	rb_insert_color_cached(&data->node, root, leftmost);
1005 }
1006 
1007 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1008 {
1009 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1010 	if (!atoms) {
1011 		pr_err("No memory at %s\n", __func__);
1012 		return -1;
1013 	}
1014 
1015 	atoms->thread = thread__get(thread);
1016 	INIT_LIST_HEAD(&atoms->work_list);
1017 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1018 	return 0;
1019 }
1020 
1021 static char sched_out_state(u64 prev_state)
1022 {
1023 	const char *str = TASK_STATE_TO_CHAR_STR;
1024 
1025 	return str[prev_state];
1026 }
1027 
1028 static int
1029 add_sched_out_event(struct work_atoms *atoms,
1030 		    char run_state,
1031 		    u64 timestamp)
1032 {
1033 	struct work_atom *atom = zalloc(sizeof(*atom));
1034 	if (!atom) {
1035 		pr_err("Non memory at %s", __func__);
1036 		return -1;
1037 	}
1038 
1039 	atom->sched_out_time = timestamp;
1040 
1041 	if (run_state == 'R') {
1042 		atom->state = THREAD_WAIT_CPU;
1043 		atom->wake_up_time = atom->sched_out_time;
1044 	}
1045 
1046 	list_add_tail(&atom->list, &atoms->work_list);
1047 	return 0;
1048 }
1049 
1050 static void
1051 add_runtime_event(struct work_atoms *atoms, u64 delta,
1052 		  u64 timestamp __maybe_unused)
1053 {
1054 	struct work_atom *atom;
1055 
1056 	BUG_ON(list_empty(&atoms->work_list));
1057 
1058 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1059 
1060 	atom->runtime += delta;
1061 	atoms->total_runtime += delta;
1062 }
1063 
1064 static void
1065 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1066 {
1067 	struct work_atom *atom;
1068 	u64 delta;
1069 
1070 	if (list_empty(&atoms->work_list))
1071 		return;
1072 
1073 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1074 
1075 	if (atom->state != THREAD_WAIT_CPU)
1076 		return;
1077 
1078 	if (timestamp < atom->wake_up_time) {
1079 		atom->state = THREAD_IGNORE;
1080 		return;
1081 	}
1082 
1083 	atom->state = THREAD_SCHED_IN;
1084 	atom->sched_in_time = timestamp;
1085 
1086 	delta = atom->sched_in_time - atom->wake_up_time;
1087 	atoms->total_lat += delta;
1088 	if (delta > atoms->max_lat) {
1089 		atoms->max_lat = delta;
1090 		atoms->max_lat_at = timestamp;
1091 	}
1092 	atoms->nb_atoms++;
1093 }
1094 
1095 static int latency_switch_event(struct perf_sched *sched,
1096 				struct perf_evsel *evsel,
1097 				struct perf_sample *sample,
1098 				struct machine *machine)
1099 {
1100 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1101 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1102 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1103 	struct work_atoms *out_events, *in_events;
1104 	struct thread *sched_out, *sched_in;
1105 	u64 timestamp0, timestamp = sample->time;
1106 	int cpu = sample->cpu, err = -1;
1107 	s64 delta;
1108 
1109 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1110 
1111 	timestamp0 = sched->cpu_last_switched[cpu];
1112 	sched->cpu_last_switched[cpu] = timestamp;
1113 	if (timestamp0)
1114 		delta = timestamp - timestamp0;
1115 	else
1116 		delta = 0;
1117 
1118 	if (delta < 0) {
1119 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1120 		return -1;
1121 	}
1122 
1123 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1124 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1125 	if (sched_out == NULL || sched_in == NULL)
1126 		goto out_put;
1127 
1128 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1129 	if (!out_events) {
1130 		if (thread_atoms_insert(sched, sched_out))
1131 			goto out_put;
1132 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1133 		if (!out_events) {
1134 			pr_err("out-event: Internal tree error");
1135 			goto out_put;
1136 		}
1137 	}
1138 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1139 		return -1;
1140 
1141 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1142 	if (!in_events) {
1143 		if (thread_atoms_insert(sched, sched_in))
1144 			goto out_put;
1145 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1146 		if (!in_events) {
1147 			pr_err("in-event: Internal tree error");
1148 			goto out_put;
1149 		}
1150 		/*
1151 		 * Take came in we have not heard about yet,
1152 		 * add in an initial atom in runnable state:
1153 		 */
1154 		if (add_sched_out_event(in_events, 'R', timestamp))
1155 			goto out_put;
1156 	}
1157 	add_sched_in_event(in_events, timestamp);
1158 	err = 0;
1159 out_put:
1160 	thread__put(sched_out);
1161 	thread__put(sched_in);
1162 	return err;
1163 }
1164 
1165 static int latency_runtime_event(struct perf_sched *sched,
1166 				 struct perf_evsel *evsel,
1167 				 struct perf_sample *sample,
1168 				 struct machine *machine)
1169 {
1170 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1171 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1172 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1173 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1174 	u64 timestamp = sample->time;
1175 	int cpu = sample->cpu, err = -1;
1176 
1177 	if (thread == NULL)
1178 		return -1;
1179 
1180 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1181 	if (!atoms) {
1182 		if (thread_atoms_insert(sched, thread))
1183 			goto out_put;
1184 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1185 		if (!atoms) {
1186 			pr_err("in-event: Internal tree error");
1187 			goto out_put;
1188 		}
1189 		if (add_sched_out_event(atoms, 'R', timestamp))
1190 			goto out_put;
1191 	}
1192 
1193 	add_runtime_event(atoms, runtime, timestamp);
1194 	err = 0;
1195 out_put:
1196 	thread__put(thread);
1197 	return err;
1198 }
1199 
1200 static int latency_wakeup_event(struct perf_sched *sched,
1201 				struct perf_evsel *evsel,
1202 				struct perf_sample *sample,
1203 				struct machine *machine)
1204 {
1205 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1206 	struct work_atoms *atoms;
1207 	struct work_atom *atom;
1208 	struct thread *wakee;
1209 	u64 timestamp = sample->time;
1210 	int err = -1;
1211 
1212 	wakee = machine__findnew_thread(machine, -1, pid);
1213 	if (wakee == NULL)
1214 		return -1;
1215 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1216 	if (!atoms) {
1217 		if (thread_atoms_insert(sched, wakee))
1218 			goto out_put;
1219 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1220 		if (!atoms) {
1221 			pr_err("wakeup-event: Internal tree error");
1222 			goto out_put;
1223 		}
1224 		if (add_sched_out_event(atoms, 'S', timestamp))
1225 			goto out_put;
1226 	}
1227 
1228 	BUG_ON(list_empty(&atoms->work_list));
1229 
1230 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1231 
1232 	/*
1233 	 * As we do not guarantee the wakeup event happens when
1234 	 * task is out of run queue, also may happen when task is
1235 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1236 	 * then we should not set the ->wake_up_time when wake up a
1237 	 * task which is on run queue.
1238 	 *
1239 	 * You WILL be missing events if you've recorded only
1240 	 * one CPU, or are only looking at only one, so don't
1241 	 * skip in this case.
1242 	 */
1243 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1244 		goto out_ok;
1245 
1246 	sched->nr_timestamps++;
1247 	if (atom->sched_out_time > timestamp) {
1248 		sched->nr_unordered_timestamps++;
1249 		goto out_ok;
1250 	}
1251 
1252 	atom->state = THREAD_WAIT_CPU;
1253 	atom->wake_up_time = timestamp;
1254 out_ok:
1255 	err = 0;
1256 out_put:
1257 	thread__put(wakee);
1258 	return err;
1259 }
1260 
1261 static int latency_migrate_task_event(struct perf_sched *sched,
1262 				      struct perf_evsel *evsel,
1263 				      struct perf_sample *sample,
1264 				      struct machine *machine)
1265 {
1266 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1267 	u64 timestamp = sample->time;
1268 	struct work_atoms *atoms;
1269 	struct work_atom *atom;
1270 	struct thread *migrant;
1271 	int err = -1;
1272 
1273 	/*
1274 	 * Only need to worry about migration when profiling one CPU.
1275 	 */
1276 	if (sched->profile_cpu == -1)
1277 		return 0;
1278 
1279 	migrant = machine__findnew_thread(machine, -1, pid);
1280 	if (migrant == NULL)
1281 		return -1;
1282 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1283 	if (!atoms) {
1284 		if (thread_atoms_insert(sched, migrant))
1285 			goto out_put;
1286 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1287 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1288 		if (!atoms) {
1289 			pr_err("migration-event: Internal tree error");
1290 			goto out_put;
1291 		}
1292 		if (add_sched_out_event(atoms, 'R', timestamp))
1293 			goto out_put;
1294 	}
1295 
1296 	BUG_ON(list_empty(&atoms->work_list));
1297 
1298 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1299 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1300 
1301 	sched->nr_timestamps++;
1302 
1303 	if (atom->sched_out_time > timestamp)
1304 		sched->nr_unordered_timestamps++;
1305 	err = 0;
1306 out_put:
1307 	thread__put(migrant);
1308 	return err;
1309 }
1310 
1311 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1312 {
1313 	int i;
1314 	int ret;
1315 	u64 avg;
1316 	char max_lat_at[32];
1317 
1318 	if (!work_list->nb_atoms)
1319 		return;
1320 	/*
1321 	 * Ignore idle threads:
1322 	 */
1323 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1324 		return;
1325 
1326 	sched->all_runtime += work_list->total_runtime;
1327 	sched->all_count   += work_list->nb_atoms;
1328 
1329 	if (work_list->num_merged > 1)
1330 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1331 	else
1332 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1333 
1334 	for (i = 0; i < 24 - ret; i++)
1335 		printf(" ");
1336 
1337 	avg = work_list->total_lat / work_list->nb_atoms;
1338 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1339 
1340 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1341 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1342 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1343 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1344 		 max_lat_at);
1345 }
1346 
1347 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1348 {
1349 	if (l->thread == r->thread)
1350 		return 0;
1351 	if (l->thread->tid < r->thread->tid)
1352 		return -1;
1353 	if (l->thread->tid > r->thread->tid)
1354 		return 1;
1355 	return (int)(l->thread - r->thread);
1356 }
1357 
1358 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1359 {
1360 	u64 avgl, avgr;
1361 
1362 	if (!l->nb_atoms)
1363 		return -1;
1364 
1365 	if (!r->nb_atoms)
1366 		return 1;
1367 
1368 	avgl = l->total_lat / l->nb_atoms;
1369 	avgr = r->total_lat / r->nb_atoms;
1370 
1371 	if (avgl < avgr)
1372 		return -1;
1373 	if (avgl > avgr)
1374 		return 1;
1375 
1376 	return 0;
1377 }
1378 
1379 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1380 {
1381 	if (l->max_lat < r->max_lat)
1382 		return -1;
1383 	if (l->max_lat > r->max_lat)
1384 		return 1;
1385 
1386 	return 0;
1387 }
1388 
1389 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1390 {
1391 	if (l->nb_atoms < r->nb_atoms)
1392 		return -1;
1393 	if (l->nb_atoms > r->nb_atoms)
1394 		return 1;
1395 
1396 	return 0;
1397 }
1398 
1399 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1400 {
1401 	if (l->total_runtime < r->total_runtime)
1402 		return -1;
1403 	if (l->total_runtime > r->total_runtime)
1404 		return 1;
1405 
1406 	return 0;
1407 }
1408 
1409 static int sort_dimension__add(const char *tok, struct list_head *list)
1410 {
1411 	size_t i;
1412 	static struct sort_dimension avg_sort_dimension = {
1413 		.name = "avg",
1414 		.cmp  = avg_cmp,
1415 	};
1416 	static struct sort_dimension max_sort_dimension = {
1417 		.name = "max",
1418 		.cmp  = max_cmp,
1419 	};
1420 	static struct sort_dimension pid_sort_dimension = {
1421 		.name = "pid",
1422 		.cmp  = pid_cmp,
1423 	};
1424 	static struct sort_dimension runtime_sort_dimension = {
1425 		.name = "runtime",
1426 		.cmp  = runtime_cmp,
1427 	};
1428 	static struct sort_dimension switch_sort_dimension = {
1429 		.name = "switch",
1430 		.cmp  = switch_cmp,
1431 	};
1432 	struct sort_dimension *available_sorts[] = {
1433 		&pid_sort_dimension,
1434 		&avg_sort_dimension,
1435 		&max_sort_dimension,
1436 		&switch_sort_dimension,
1437 		&runtime_sort_dimension,
1438 	};
1439 
1440 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1441 		if (!strcmp(available_sorts[i]->name, tok)) {
1442 			list_add_tail(&available_sorts[i]->list, list);
1443 
1444 			return 0;
1445 		}
1446 	}
1447 
1448 	return -1;
1449 }
1450 
1451 static void perf_sched__sort_lat(struct perf_sched *sched)
1452 {
1453 	struct rb_node *node;
1454 	struct rb_root_cached *root = &sched->atom_root;
1455 again:
1456 	for (;;) {
1457 		struct work_atoms *data;
1458 		node = rb_first_cached(root);
1459 		if (!node)
1460 			break;
1461 
1462 		rb_erase_cached(node, root);
1463 		data = rb_entry(node, struct work_atoms, node);
1464 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1465 	}
1466 	if (root == &sched->atom_root) {
1467 		root = &sched->merged_atom_root;
1468 		goto again;
1469 	}
1470 }
1471 
1472 static int process_sched_wakeup_event(struct perf_tool *tool,
1473 				      struct perf_evsel *evsel,
1474 				      struct perf_sample *sample,
1475 				      struct machine *machine)
1476 {
1477 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1478 
1479 	if (sched->tp_handler->wakeup_event)
1480 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1481 
1482 	return 0;
1483 }
1484 
1485 union map_priv {
1486 	void	*ptr;
1487 	bool	 color;
1488 };
1489 
1490 static bool thread__has_color(struct thread *thread)
1491 {
1492 	union map_priv priv = {
1493 		.ptr = thread__priv(thread),
1494 	};
1495 
1496 	return priv.color;
1497 }
1498 
1499 static struct thread*
1500 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1501 {
1502 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1503 	union map_priv priv = {
1504 		.color = false,
1505 	};
1506 
1507 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1508 		return thread;
1509 
1510 	if (thread_map__has(sched->map.color_pids, tid))
1511 		priv.color = true;
1512 
1513 	thread__set_priv(thread, priv.ptr);
1514 	return thread;
1515 }
1516 
1517 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1518 			    struct perf_sample *sample, struct machine *machine)
1519 {
1520 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1521 	struct thread *sched_in;
1522 	struct thread_runtime *tr;
1523 	int new_shortname;
1524 	u64 timestamp0, timestamp = sample->time;
1525 	s64 delta;
1526 	int i, this_cpu = sample->cpu;
1527 	int cpus_nr;
1528 	bool new_cpu = false;
1529 	const char *color = PERF_COLOR_NORMAL;
1530 	char stimestamp[32];
1531 
1532 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1533 
1534 	if (this_cpu > sched->max_cpu)
1535 		sched->max_cpu = this_cpu;
1536 
1537 	if (sched->map.comp) {
1538 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1539 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1540 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1541 			new_cpu = true;
1542 		}
1543 	} else
1544 		cpus_nr = sched->max_cpu;
1545 
1546 	timestamp0 = sched->cpu_last_switched[this_cpu];
1547 	sched->cpu_last_switched[this_cpu] = timestamp;
1548 	if (timestamp0)
1549 		delta = timestamp - timestamp0;
1550 	else
1551 		delta = 0;
1552 
1553 	if (delta < 0) {
1554 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1555 		return -1;
1556 	}
1557 
1558 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1559 	if (sched_in == NULL)
1560 		return -1;
1561 
1562 	tr = thread__get_runtime(sched_in);
1563 	if (tr == NULL) {
1564 		thread__put(sched_in);
1565 		return -1;
1566 	}
1567 
1568 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1569 
1570 	printf("  ");
1571 
1572 	new_shortname = 0;
1573 	if (!tr->shortname[0]) {
1574 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1575 			/*
1576 			 * Don't allocate a letter-number for swapper:0
1577 			 * as a shortname. Instead, we use '.' for it.
1578 			 */
1579 			tr->shortname[0] = '.';
1580 			tr->shortname[1] = ' ';
1581 		} else {
1582 			tr->shortname[0] = sched->next_shortname1;
1583 			tr->shortname[1] = sched->next_shortname2;
1584 
1585 			if (sched->next_shortname1 < 'Z') {
1586 				sched->next_shortname1++;
1587 			} else {
1588 				sched->next_shortname1 = 'A';
1589 				if (sched->next_shortname2 < '9')
1590 					sched->next_shortname2++;
1591 				else
1592 					sched->next_shortname2 = '0';
1593 			}
1594 		}
1595 		new_shortname = 1;
1596 	}
1597 
1598 	for (i = 0; i < cpus_nr; i++) {
1599 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1600 		struct thread *curr_thread = sched->curr_thread[cpu];
1601 		struct thread_runtime *curr_tr;
1602 		const char *pid_color = color;
1603 		const char *cpu_color = color;
1604 
1605 		if (curr_thread && thread__has_color(curr_thread))
1606 			pid_color = COLOR_PIDS;
1607 
1608 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1609 			continue;
1610 
1611 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1612 			cpu_color = COLOR_CPUS;
1613 
1614 		if (cpu != this_cpu)
1615 			color_fprintf(stdout, color, " ");
1616 		else
1617 			color_fprintf(stdout, cpu_color, "*");
1618 
1619 		if (sched->curr_thread[cpu]) {
1620 			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1621 			if (curr_tr == NULL) {
1622 				thread__put(sched_in);
1623 				return -1;
1624 			}
1625 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1626 		} else
1627 			color_fprintf(stdout, color, "   ");
1628 	}
1629 
1630 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1631 		goto out;
1632 
1633 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1634 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1635 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1636 		const char *pid_color = color;
1637 
1638 		if (thread__has_color(sched_in))
1639 			pid_color = COLOR_PIDS;
1640 
1641 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1642 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1643 		tr->comm_changed = false;
1644 	}
1645 
1646 	if (sched->map.comp && new_cpu)
1647 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1648 
1649 out:
1650 	color_fprintf(stdout, color, "\n");
1651 
1652 	thread__put(sched_in);
1653 
1654 	return 0;
1655 }
1656 
1657 static int process_sched_switch_event(struct perf_tool *tool,
1658 				      struct perf_evsel *evsel,
1659 				      struct perf_sample *sample,
1660 				      struct machine *machine)
1661 {
1662 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1663 	int this_cpu = sample->cpu, err = 0;
1664 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1665 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1666 
1667 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1668 		/*
1669 		 * Are we trying to switch away a PID that is
1670 		 * not current?
1671 		 */
1672 		if (sched->curr_pid[this_cpu] != prev_pid)
1673 			sched->nr_context_switch_bugs++;
1674 	}
1675 
1676 	if (sched->tp_handler->switch_event)
1677 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1678 
1679 	sched->curr_pid[this_cpu] = next_pid;
1680 	return err;
1681 }
1682 
1683 static int process_sched_runtime_event(struct perf_tool *tool,
1684 				       struct perf_evsel *evsel,
1685 				       struct perf_sample *sample,
1686 				       struct machine *machine)
1687 {
1688 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1689 
1690 	if (sched->tp_handler->runtime_event)
1691 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1692 
1693 	return 0;
1694 }
1695 
1696 static int perf_sched__process_fork_event(struct perf_tool *tool,
1697 					  union perf_event *event,
1698 					  struct perf_sample *sample,
1699 					  struct machine *machine)
1700 {
1701 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1702 
1703 	/* run the fork event through the perf machineruy */
1704 	perf_event__process_fork(tool, event, sample, machine);
1705 
1706 	/* and then run additional processing needed for this command */
1707 	if (sched->tp_handler->fork_event)
1708 		return sched->tp_handler->fork_event(sched, event, machine);
1709 
1710 	return 0;
1711 }
1712 
1713 static int process_sched_migrate_task_event(struct perf_tool *tool,
1714 					    struct perf_evsel *evsel,
1715 					    struct perf_sample *sample,
1716 					    struct machine *machine)
1717 {
1718 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1719 
1720 	if (sched->tp_handler->migrate_task_event)
1721 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1722 
1723 	return 0;
1724 }
1725 
1726 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1727 				  struct perf_evsel *evsel,
1728 				  struct perf_sample *sample,
1729 				  struct machine *machine);
1730 
1731 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1732 						 union perf_event *event __maybe_unused,
1733 						 struct perf_sample *sample,
1734 						 struct perf_evsel *evsel,
1735 						 struct machine *machine)
1736 {
1737 	int err = 0;
1738 
1739 	if (evsel->handler != NULL) {
1740 		tracepoint_handler f = evsel->handler;
1741 		err = f(tool, evsel, sample, machine);
1742 	}
1743 
1744 	return err;
1745 }
1746 
1747 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1748 				    union perf_event *event,
1749 				    struct perf_sample *sample,
1750 				    struct machine *machine)
1751 {
1752 	struct thread *thread;
1753 	struct thread_runtime *tr;
1754 	int err;
1755 
1756 	err = perf_event__process_comm(tool, event, sample, machine);
1757 	if (err)
1758 		return err;
1759 
1760 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1761 	if (!thread) {
1762 		pr_err("Internal error: can't find thread\n");
1763 		return -1;
1764 	}
1765 
1766 	tr = thread__get_runtime(thread);
1767 	if (tr == NULL) {
1768 		thread__put(thread);
1769 		return -1;
1770 	}
1771 
1772 	tr->comm_changed = true;
1773 	thread__put(thread);
1774 
1775 	return 0;
1776 }
1777 
1778 static int perf_sched__read_events(struct perf_sched *sched)
1779 {
1780 	const struct perf_evsel_str_handler handlers[] = {
1781 		{ "sched:sched_switch",	      process_sched_switch_event, },
1782 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1783 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1784 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1785 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1786 	};
1787 	struct perf_session *session;
1788 	struct perf_data data = {
1789 		.path  = input_name,
1790 		.mode  = PERF_DATA_MODE_READ,
1791 		.force = sched->force,
1792 	};
1793 	int rc = -1;
1794 
1795 	session = perf_session__new(&data, false, &sched->tool);
1796 	if (session == NULL) {
1797 		pr_debug("No Memory for session\n");
1798 		return -1;
1799 	}
1800 
1801 	symbol__init(&session->header.env);
1802 
1803 	if (perf_session__set_tracepoints_handlers(session, handlers))
1804 		goto out_delete;
1805 
1806 	if (perf_session__has_traces(session, "record -R")) {
1807 		int err = perf_session__process_events(session);
1808 		if (err) {
1809 			pr_err("Failed to process events, error %d", err);
1810 			goto out_delete;
1811 		}
1812 
1813 		sched->nr_events      = session->evlist->stats.nr_events[0];
1814 		sched->nr_lost_events = session->evlist->stats.total_lost;
1815 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1816 	}
1817 
1818 	rc = 0;
1819 out_delete:
1820 	perf_session__delete(session);
1821 	return rc;
1822 }
1823 
1824 /*
1825  * scheduling times are printed as msec.usec
1826  */
1827 static inline void print_sched_time(unsigned long long nsecs, int width)
1828 {
1829 	unsigned long msecs;
1830 	unsigned long usecs;
1831 
1832 	msecs  = nsecs / NSEC_PER_MSEC;
1833 	nsecs -= msecs * NSEC_PER_MSEC;
1834 	usecs  = nsecs / NSEC_PER_USEC;
1835 	printf("%*lu.%03lu ", width, msecs, usecs);
1836 }
1837 
1838 /*
1839  * returns runtime data for event, allocating memory for it the
1840  * first time it is used.
1841  */
1842 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1843 {
1844 	struct evsel_runtime *r = evsel->priv;
1845 
1846 	if (r == NULL) {
1847 		r = zalloc(sizeof(struct evsel_runtime));
1848 		evsel->priv = r;
1849 	}
1850 
1851 	return r;
1852 }
1853 
1854 /*
1855  * save last time event was seen per cpu
1856  */
1857 static void perf_evsel__save_time(struct perf_evsel *evsel,
1858 				  u64 timestamp, u32 cpu)
1859 {
1860 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1861 
1862 	if (r == NULL)
1863 		return;
1864 
1865 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1866 		int i, n = __roundup_pow_of_two(cpu+1);
1867 		void *p = r->last_time;
1868 
1869 		p = realloc(r->last_time, n * sizeof(u64));
1870 		if (!p)
1871 			return;
1872 
1873 		r->last_time = p;
1874 		for (i = r->ncpu; i < n; ++i)
1875 			r->last_time[i] = (u64) 0;
1876 
1877 		r->ncpu = n;
1878 	}
1879 
1880 	r->last_time[cpu] = timestamp;
1881 }
1882 
1883 /* returns last time this event was seen on the given cpu */
1884 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1885 {
1886 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1887 
1888 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1889 		return 0;
1890 
1891 	return r->last_time[cpu];
1892 }
1893 
1894 static int comm_width = 30;
1895 
1896 static char *timehist_get_commstr(struct thread *thread)
1897 {
1898 	static char str[32];
1899 	const char *comm = thread__comm_str(thread);
1900 	pid_t tid = thread->tid;
1901 	pid_t pid = thread->pid_;
1902 	int n;
1903 
1904 	if (pid == 0)
1905 		n = scnprintf(str, sizeof(str), "%s", comm);
1906 
1907 	else if (tid != pid)
1908 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1909 
1910 	else
1911 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1912 
1913 	if (n > comm_width)
1914 		comm_width = n;
1915 
1916 	return str;
1917 }
1918 
1919 static void timehist_header(struct perf_sched *sched)
1920 {
1921 	u32 ncpus = sched->max_cpu + 1;
1922 	u32 i, j;
1923 
1924 	printf("%15s %6s ", "time", "cpu");
1925 
1926 	if (sched->show_cpu_visual) {
1927 		printf(" ");
1928 		for (i = 0, j = 0; i < ncpus; ++i) {
1929 			printf("%x", j++);
1930 			if (j > 15)
1931 				j = 0;
1932 		}
1933 		printf(" ");
1934 	}
1935 
1936 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1937 		"task name", "wait time", "sch delay", "run time");
1938 
1939 	if (sched->show_state)
1940 		printf("  %s", "state");
1941 
1942 	printf("\n");
1943 
1944 	/*
1945 	 * units row
1946 	 */
1947 	printf("%15s %-6s ", "", "");
1948 
1949 	if (sched->show_cpu_visual)
1950 		printf(" %*s ", ncpus, "");
1951 
1952 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1953 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1954 
1955 	if (sched->show_state)
1956 		printf("  %5s", "");
1957 
1958 	printf("\n");
1959 
1960 	/*
1961 	 * separator
1962 	 */
1963 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1964 
1965 	if (sched->show_cpu_visual)
1966 		printf(" %.*s ", ncpus, graph_dotted_line);
1967 
1968 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1969 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1970 		graph_dotted_line);
1971 
1972 	if (sched->show_state)
1973 		printf("  %.5s", graph_dotted_line);
1974 
1975 	printf("\n");
1976 }
1977 
1978 static char task_state_char(struct thread *thread, int state)
1979 {
1980 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1981 	unsigned bit = state ? ffs(state) : 0;
1982 
1983 	/* 'I' for idle */
1984 	if (thread->tid == 0)
1985 		return 'I';
1986 
1987 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1988 }
1989 
1990 static void timehist_print_sample(struct perf_sched *sched,
1991 				  struct perf_evsel *evsel,
1992 				  struct perf_sample *sample,
1993 				  struct addr_location *al,
1994 				  struct thread *thread,
1995 				  u64 t, int state)
1996 {
1997 	struct thread_runtime *tr = thread__priv(thread);
1998 	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1999 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
2000 	u32 max_cpus = sched->max_cpu + 1;
2001 	char tstr[64];
2002 	char nstr[30];
2003 	u64 wait_time;
2004 
2005 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2006 	printf("%15s [%04d] ", tstr, sample->cpu);
2007 
2008 	if (sched->show_cpu_visual) {
2009 		u32 i;
2010 		char c;
2011 
2012 		printf(" ");
2013 		for (i = 0; i < max_cpus; ++i) {
2014 			/* flag idle times with 'i'; others are sched events */
2015 			if (i == sample->cpu)
2016 				c = (thread->tid == 0) ? 'i' : 's';
2017 			else
2018 				c = ' ';
2019 			printf("%c", c);
2020 		}
2021 		printf(" ");
2022 	}
2023 
2024 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2025 
2026 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2027 	print_sched_time(wait_time, 6);
2028 
2029 	print_sched_time(tr->dt_delay, 6);
2030 	print_sched_time(tr->dt_run, 6);
2031 
2032 	if (sched->show_state)
2033 		printf(" %5c ", task_state_char(thread, state));
2034 
2035 	if (sched->show_next) {
2036 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2037 		printf(" %-*s", comm_width, nstr);
2038 	}
2039 
2040 	if (sched->show_wakeups && !sched->show_next)
2041 		printf("  %-*s", comm_width, "");
2042 
2043 	if (thread->tid == 0)
2044 		goto out;
2045 
2046 	if (sched->show_callchain)
2047 		printf("  ");
2048 
2049 	sample__fprintf_sym(sample, al, 0,
2050 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2051 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2052 			    EVSEL__PRINT_SKIP_IGNORED,
2053 			    &callchain_cursor, stdout);
2054 
2055 out:
2056 	printf("\n");
2057 }
2058 
2059 /*
2060  * Explanation of delta-time stats:
2061  *
2062  *            t = time of current schedule out event
2063  *        tprev = time of previous sched out event
2064  *                also time of schedule-in event for current task
2065  *    last_time = time of last sched change event for current task
2066  *                (i.e, time process was last scheduled out)
2067  * ready_to_run = time of wakeup for current task
2068  *
2069  * -----|------------|------------|------------|------
2070  *    last         ready        tprev          t
2071  *    time         to run
2072  *
2073  *      |-------- dt_wait --------|
2074  *                   |- dt_delay -|-- dt_run --|
2075  *
2076  *   dt_run = run time of current task
2077  *  dt_wait = time between last schedule out event for task and tprev
2078  *            represents time spent off the cpu
2079  * dt_delay = time between wakeup and schedule-in of task
2080  */
2081 
2082 static void timehist_update_runtime_stats(struct thread_runtime *r,
2083 					 u64 t, u64 tprev)
2084 {
2085 	r->dt_delay   = 0;
2086 	r->dt_sleep   = 0;
2087 	r->dt_iowait  = 0;
2088 	r->dt_preempt = 0;
2089 	r->dt_run     = 0;
2090 
2091 	if (tprev) {
2092 		r->dt_run = t - tprev;
2093 		if (r->ready_to_run) {
2094 			if (r->ready_to_run > tprev)
2095 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2096 			else
2097 				r->dt_delay = tprev - r->ready_to_run;
2098 		}
2099 
2100 		if (r->last_time > tprev)
2101 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2102 		else if (r->last_time) {
2103 			u64 dt_wait = tprev - r->last_time;
2104 
2105 			if (r->last_state == TASK_RUNNING)
2106 				r->dt_preempt = dt_wait;
2107 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2108 				r->dt_iowait = dt_wait;
2109 			else
2110 				r->dt_sleep = dt_wait;
2111 		}
2112 	}
2113 
2114 	update_stats(&r->run_stats, r->dt_run);
2115 
2116 	r->total_run_time     += r->dt_run;
2117 	r->total_delay_time   += r->dt_delay;
2118 	r->total_sleep_time   += r->dt_sleep;
2119 	r->total_iowait_time  += r->dt_iowait;
2120 	r->total_preempt_time += r->dt_preempt;
2121 }
2122 
2123 static bool is_idle_sample(struct perf_sample *sample,
2124 			   struct perf_evsel *evsel)
2125 {
2126 	/* pid 0 == swapper == idle task */
2127 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2128 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2129 
2130 	return sample->pid == 0;
2131 }
2132 
2133 static void save_task_callchain(struct perf_sched *sched,
2134 				struct perf_sample *sample,
2135 				struct perf_evsel *evsel,
2136 				struct machine *machine)
2137 {
2138 	struct callchain_cursor *cursor = &callchain_cursor;
2139 	struct thread *thread;
2140 
2141 	/* want main thread for process - has maps */
2142 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2143 	if (thread == NULL) {
2144 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2145 		return;
2146 	}
2147 
2148 	if (!sched->show_callchain || sample->callchain == NULL)
2149 		return;
2150 
2151 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2152 				      NULL, NULL, sched->max_stack + 2) != 0) {
2153 		if (verbose > 0)
2154 			pr_err("Failed to resolve callchain. Skipping\n");
2155 
2156 		return;
2157 	}
2158 
2159 	callchain_cursor_commit(cursor);
2160 
2161 	while (true) {
2162 		struct callchain_cursor_node *node;
2163 		struct symbol *sym;
2164 
2165 		node = callchain_cursor_current(cursor);
2166 		if (node == NULL)
2167 			break;
2168 
2169 		sym = node->sym;
2170 		if (sym) {
2171 			if (!strcmp(sym->name, "schedule") ||
2172 			    !strcmp(sym->name, "__schedule") ||
2173 			    !strcmp(sym->name, "preempt_schedule"))
2174 				sym->ignore = 1;
2175 		}
2176 
2177 		callchain_cursor_advance(cursor);
2178 	}
2179 }
2180 
2181 static int init_idle_thread(struct thread *thread)
2182 {
2183 	struct idle_thread_runtime *itr;
2184 
2185 	thread__set_comm(thread, idle_comm, 0);
2186 
2187 	itr = zalloc(sizeof(*itr));
2188 	if (itr == NULL)
2189 		return -ENOMEM;
2190 
2191 	init_stats(&itr->tr.run_stats);
2192 	callchain_init(&itr->callchain);
2193 	callchain_cursor_reset(&itr->cursor);
2194 	thread__set_priv(thread, itr);
2195 
2196 	return 0;
2197 }
2198 
2199 /*
2200  * Track idle stats per cpu by maintaining a local thread
2201  * struct for the idle task on each cpu.
2202  */
2203 static int init_idle_threads(int ncpu)
2204 {
2205 	int i, ret;
2206 
2207 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2208 	if (!idle_threads)
2209 		return -ENOMEM;
2210 
2211 	idle_max_cpu = ncpu;
2212 
2213 	/* allocate the actual thread struct if needed */
2214 	for (i = 0; i < ncpu; ++i) {
2215 		idle_threads[i] = thread__new(0, 0);
2216 		if (idle_threads[i] == NULL)
2217 			return -ENOMEM;
2218 
2219 		ret = init_idle_thread(idle_threads[i]);
2220 		if (ret < 0)
2221 			return ret;
2222 	}
2223 
2224 	return 0;
2225 }
2226 
2227 static void free_idle_threads(void)
2228 {
2229 	int i;
2230 
2231 	if (idle_threads == NULL)
2232 		return;
2233 
2234 	for (i = 0; i < idle_max_cpu; ++i) {
2235 		if ((idle_threads[i]))
2236 			thread__delete(idle_threads[i]);
2237 	}
2238 
2239 	free(idle_threads);
2240 }
2241 
2242 static struct thread *get_idle_thread(int cpu)
2243 {
2244 	/*
2245 	 * expand/allocate array of pointers to local thread
2246 	 * structs if needed
2247 	 */
2248 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2249 		int i, j = __roundup_pow_of_two(cpu+1);
2250 		void *p;
2251 
2252 		p = realloc(idle_threads, j * sizeof(struct thread *));
2253 		if (!p)
2254 			return NULL;
2255 
2256 		idle_threads = (struct thread **) p;
2257 		for (i = idle_max_cpu; i < j; ++i)
2258 			idle_threads[i] = NULL;
2259 
2260 		idle_max_cpu = j;
2261 	}
2262 
2263 	/* allocate a new thread struct if needed */
2264 	if (idle_threads[cpu] == NULL) {
2265 		idle_threads[cpu] = thread__new(0, 0);
2266 
2267 		if (idle_threads[cpu]) {
2268 			if (init_idle_thread(idle_threads[cpu]) < 0)
2269 				return NULL;
2270 		}
2271 	}
2272 
2273 	return idle_threads[cpu];
2274 }
2275 
2276 static void save_idle_callchain(struct perf_sched *sched,
2277 				struct idle_thread_runtime *itr,
2278 				struct perf_sample *sample)
2279 {
2280 	if (!sched->show_callchain || sample->callchain == NULL)
2281 		return;
2282 
2283 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2284 }
2285 
2286 static struct thread *timehist_get_thread(struct perf_sched *sched,
2287 					  struct perf_sample *sample,
2288 					  struct machine *machine,
2289 					  struct perf_evsel *evsel)
2290 {
2291 	struct thread *thread;
2292 
2293 	if (is_idle_sample(sample, evsel)) {
2294 		thread = get_idle_thread(sample->cpu);
2295 		if (thread == NULL)
2296 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2297 
2298 	} else {
2299 		/* there were samples with tid 0 but non-zero pid */
2300 		thread = machine__findnew_thread(machine, sample->pid,
2301 						 sample->tid ?: sample->pid);
2302 		if (thread == NULL) {
2303 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2304 				 sample->tid);
2305 		}
2306 
2307 		save_task_callchain(sched, sample, evsel, machine);
2308 		if (sched->idle_hist) {
2309 			struct thread *idle;
2310 			struct idle_thread_runtime *itr;
2311 
2312 			idle = get_idle_thread(sample->cpu);
2313 			if (idle == NULL) {
2314 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2315 				return NULL;
2316 			}
2317 
2318 			itr = thread__priv(idle);
2319 			if (itr == NULL)
2320 				return NULL;
2321 
2322 			itr->last_thread = thread;
2323 
2324 			/* copy task callchain when entering to idle */
2325 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2326 				save_idle_callchain(sched, itr, sample);
2327 		}
2328 	}
2329 
2330 	return thread;
2331 }
2332 
2333 static bool timehist_skip_sample(struct perf_sched *sched,
2334 				 struct thread *thread,
2335 				 struct perf_evsel *evsel,
2336 				 struct perf_sample *sample)
2337 {
2338 	bool rc = false;
2339 
2340 	if (thread__is_filtered(thread)) {
2341 		rc = true;
2342 		sched->skipped_samples++;
2343 	}
2344 
2345 	if (sched->idle_hist) {
2346 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2347 			rc = true;
2348 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2349 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2350 			rc = true;
2351 	}
2352 
2353 	return rc;
2354 }
2355 
2356 static void timehist_print_wakeup_event(struct perf_sched *sched,
2357 					struct perf_evsel *evsel,
2358 					struct perf_sample *sample,
2359 					struct machine *machine,
2360 					struct thread *awakened)
2361 {
2362 	struct thread *thread;
2363 	char tstr[64];
2364 
2365 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2366 	if (thread == NULL)
2367 		return;
2368 
2369 	/* show wakeup unless both awakee and awaker are filtered */
2370 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2371 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2372 		return;
2373 	}
2374 
2375 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2376 	printf("%15s [%04d] ", tstr, sample->cpu);
2377 	if (sched->show_cpu_visual)
2378 		printf(" %*s ", sched->max_cpu + 1, "");
2379 
2380 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2381 
2382 	/* dt spacer */
2383 	printf("  %9s  %9s  %9s ", "", "", "");
2384 
2385 	printf("awakened: %s", timehist_get_commstr(awakened));
2386 
2387 	printf("\n");
2388 }
2389 
2390 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2391 				       union perf_event *event __maybe_unused,
2392 				       struct perf_evsel *evsel,
2393 				       struct perf_sample *sample,
2394 				       struct machine *machine)
2395 {
2396 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2397 	struct thread *thread;
2398 	struct thread_runtime *tr = NULL;
2399 	/* want pid of awakened task not pid in sample */
2400 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2401 
2402 	thread = machine__findnew_thread(machine, 0, pid);
2403 	if (thread == NULL)
2404 		return -1;
2405 
2406 	tr = thread__get_runtime(thread);
2407 	if (tr == NULL)
2408 		return -1;
2409 
2410 	if (tr->ready_to_run == 0)
2411 		tr->ready_to_run = sample->time;
2412 
2413 	/* show wakeups if requested */
2414 	if (sched->show_wakeups &&
2415 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2416 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2417 
2418 	return 0;
2419 }
2420 
2421 static void timehist_print_migration_event(struct perf_sched *sched,
2422 					struct perf_evsel *evsel,
2423 					struct perf_sample *sample,
2424 					struct machine *machine,
2425 					struct thread *migrated)
2426 {
2427 	struct thread *thread;
2428 	char tstr[64];
2429 	u32 max_cpus = sched->max_cpu + 1;
2430 	u32 ocpu, dcpu;
2431 
2432 	if (sched->summary_only)
2433 		return;
2434 
2435 	max_cpus = sched->max_cpu + 1;
2436 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2437 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2438 
2439 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2440 	if (thread == NULL)
2441 		return;
2442 
2443 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2444 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2445 		return;
2446 	}
2447 
2448 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2449 	printf("%15s [%04d] ", tstr, sample->cpu);
2450 
2451 	if (sched->show_cpu_visual) {
2452 		u32 i;
2453 		char c;
2454 
2455 		printf("  ");
2456 		for (i = 0; i < max_cpus; ++i) {
2457 			c = (i == sample->cpu) ? 'm' : ' ';
2458 			printf("%c", c);
2459 		}
2460 		printf("  ");
2461 	}
2462 
2463 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2464 
2465 	/* dt spacer */
2466 	printf("  %9s  %9s  %9s ", "", "", "");
2467 
2468 	printf("migrated: %s", timehist_get_commstr(migrated));
2469 	printf(" cpu %d => %d", ocpu, dcpu);
2470 
2471 	printf("\n");
2472 }
2473 
2474 static int timehist_migrate_task_event(struct perf_tool *tool,
2475 				       union perf_event *event __maybe_unused,
2476 				       struct perf_evsel *evsel,
2477 				       struct perf_sample *sample,
2478 				       struct machine *machine)
2479 {
2480 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2481 	struct thread *thread;
2482 	struct thread_runtime *tr = NULL;
2483 	/* want pid of migrated task not pid in sample */
2484 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2485 
2486 	thread = machine__findnew_thread(machine, 0, pid);
2487 	if (thread == NULL)
2488 		return -1;
2489 
2490 	tr = thread__get_runtime(thread);
2491 	if (tr == NULL)
2492 		return -1;
2493 
2494 	tr->migrations++;
2495 
2496 	/* show migrations if requested */
2497 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2498 
2499 	return 0;
2500 }
2501 
2502 static int timehist_sched_change_event(struct perf_tool *tool,
2503 				       union perf_event *event,
2504 				       struct perf_evsel *evsel,
2505 				       struct perf_sample *sample,
2506 				       struct machine *machine)
2507 {
2508 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2509 	struct perf_time_interval *ptime = &sched->ptime;
2510 	struct addr_location al;
2511 	struct thread *thread;
2512 	struct thread_runtime *tr = NULL;
2513 	u64 tprev, t = sample->time;
2514 	int rc = 0;
2515 	int state = perf_evsel__intval(evsel, sample, "prev_state");
2516 
2517 
2518 	if (machine__resolve(machine, &al, sample) < 0) {
2519 		pr_err("problem processing %d event. skipping it\n",
2520 		       event->header.type);
2521 		rc = -1;
2522 		goto out;
2523 	}
2524 
2525 	thread = timehist_get_thread(sched, sample, machine, evsel);
2526 	if (thread == NULL) {
2527 		rc = -1;
2528 		goto out;
2529 	}
2530 
2531 	if (timehist_skip_sample(sched, thread, evsel, sample))
2532 		goto out;
2533 
2534 	tr = thread__get_runtime(thread);
2535 	if (tr == NULL) {
2536 		rc = -1;
2537 		goto out;
2538 	}
2539 
2540 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2541 
2542 	/*
2543 	 * If start time given:
2544 	 * - sample time is under window user cares about - skip sample
2545 	 * - tprev is under window user cares about  - reset to start of window
2546 	 */
2547 	if (ptime->start && ptime->start > t)
2548 		goto out;
2549 
2550 	if (tprev && ptime->start > tprev)
2551 		tprev = ptime->start;
2552 
2553 	/*
2554 	 * If end time given:
2555 	 * - previous sched event is out of window - we are done
2556 	 * - sample time is beyond window user cares about - reset it
2557 	 *   to close out stats for time window interest
2558 	 */
2559 	if (ptime->end) {
2560 		if (tprev > ptime->end)
2561 			goto out;
2562 
2563 		if (t > ptime->end)
2564 			t = ptime->end;
2565 	}
2566 
2567 	if (!sched->idle_hist || thread->tid == 0) {
2568 		timehist_update_runtime_stats(tr, t, tprev);
2569 
2570 		if (sched->idle_hist) {
2571 			struct idle_thread_runtime *itr = (void *)tr;
2572 			struct thread_runtime *last_tr;
2573 
2574 			BUG_ON(thread->tid != 0);
2575 
2576 			if (itr->last_thread == NULL)
2577 				goto out;
2578 
2579 			/* add current idle time as last thread's runtime */
2580 			last_tr = thread__get_runtime(itr->last_thread);
2581 			if (last_tr == NULL)
2582 				goto out;
2583 
2584 			timehist_update_runtime_stats(last_tr, t, tprev);
2585 			/*
2586 			 * remove delta time of last thread as it's not updated
2587 			 * and otherwise it will show an invalid value next
2588 			 * time.  we only care total run time and run stat.
2589 			 */
2590 			last_tr->dt_run = 0;
2591 			last_tr->dt_delay = 0;
2592 			last_tr->dt_sleep = 0;
2593 			last_tr->dt_iowait = 0;
2594 			last_tr->dt_preempt = 0;
2595 
2596 			if (itr->cursor.nr)
2597 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2598 
2599 			itr->last_thread = NULL;
2600 		}
2601 	}
2602 
2603 	if (!sched->summary_only)
2604 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2605 
2606 out:
2607 	if (sched->hist_time.start == 0 && t >= ptime->start)
2608 		sched->hist_time.start = t;
2609 	if (ptime->end == 0 || t <= ptime->end)
2610 		sched->hist_time.end = t;
2611 
2612 	if (tr) {
2613 		/* time of this sched_switch event becomes last time task seen */
2614 		tr->last_time = sample->time;
2615 
2616 		/* last state is used to determine where to account wait time */
2617 		tr->last_state = state;
2618 
2619 		/* sched out event for task so reset ready to run time */
2620 		tr->ready_to_run = 0;
2621 	}
2622 
2623 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2624 
2625 	return rc;
2626 }
2627 
2628 static int timehist_sched_switch_event(struct perf_tool *tool,
2629 			     union perf_event *event,
2630 			     struct perf_evsel *evsel,
2631 			     struct perf_sample *sample,
2632 			     struct machine *machine __maybe_unused)
2633 {
2634 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2635 }
2636 
2637 static int process_lost(struct perf_tool *tool __maybe_unused,
2638 			union perf_event *event,
2639 			struct perf_sample *sample,
2640 			struct machine *machine __maybe_unused)
2641 {
2642 	char tstr[64];
2643 
2644 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2645 	printf("%15s ", tstr);
2646 	printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2647 
2648 	return 0;
2649 }
2650 
2651 
2652 static void print_thread_runtime(struct thread *t,
2653 				 struct thread_runtime *r)
2654 {
2655 	double mean = avg_stats(&r->run_stats);
2656 	float stddev;
2657 
2658 	printf("%*s   %5d  %9" PRIu64 " ",
2659 	       comm_width, timehist_get_commstr(t), t->ppid,
2660 	       (u64) r->run_stats.n);
2661 
2662 	print_sched_time(r->total_run_time, 8);
2663 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2664 	print_sched_time(r->run_stats.min, 6);
2665 	printf(" ");
2666 	print_sched_time((u64) mean, 6);
2667 	printf(" ");
2668 	print_sched_time(r->run_stats.max, 6);
2669 	printf("  ");
2670 	printf("%5.2f", stddev);
2671 	printf("   %5" PRIu64, r->migrations);
2672 	printf("\n");
2673 }
2674 
2675 static void print_thread_waittime(struct thread *t,
2676 				  struct thread_runtime *r)
2677 {
2678 	printf("%*s   %5d  %9" PRIu64 " ",
2679 	       comm_width, timehist_get_commstr(t), t->ppid,
2680 	       (u64) r->run_stats.n);
2681 
2682 	print_sched_time(r->total_run_time, 8);
2683 	print_sched_time(r->total_sleep_time, 6);
2684 	printf(" ");
2685 	print_sched_time(r->total_iowait_time, 6);
2686 	printf(" ");
2687 	print_sched_time(r->total_preempt_time, 6);
2688 	printf(" ");
2689 	print_sched_time(r->total_delay_time, 6);
2690 	printf("\n");
2691 }
2692 
2693 struct total_run_stats {
2694 	struct perf_sched *sched;
2695 	u64  sched_count;
2696 	u64  task_count;
2697 	u64  total_run_time;
2698 };
2699 
2700 static int __show_thread_runtime(struct thread *t, void *priv)
2701 {
2702 	struct total_run_stats *stats = priv;
2703 	struct thread_runtime *r;
2704 
2705 	if (thread__is_filtered(t))
2706 		return 0;
2707 
2708 	r = thread__priv(t);
2709 	if (r && r->run_stats.n) {
2710 		stats->task_count++;
2711 		stats->sched_count += r->run_stats.n;
2712 		stats->total_run_time += r->total_run_time;
2713 
2714 		if (stats->sched->show_state)
2715 			print_thread_waittime(t, r);
2716 		else
2717 			print_thread_runtime(t, r);
2718 	}
2719 
2720 	return 0;
2721 }
2722 
2723 static int show_thread_runtime(struct thread *t, void *priv)
2724 {
2725 	if (t->dead)
2726 		return 0;
2727 
2728 	return __show_thread_runtime(t, priv);
2729 }
2730 
2731 static int show_deadthread_runtime(struct thread *t, void *priv)
2732 {
2733 	if (!t->dead)
2734 		return 0;
2735 
2736 	return __show_thread_runtime(t, priv);
2737 }
2738 
2739 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2740 {
2741 	const char *sep = " <- ";
2742 	struct callchain_list *chain;
2743 	size_t ret = 0;
2744 	char bf[1024];
2745 	bool first;
2746 
2747 	if (node == NULL)
2748 		return 0;
2749 
2750 	ret = callchain__fprintf_folded(fp, node->parent);
2751 	first = (ret == 0);
2752 
2753 	list_for_each_entry(chain, &node->val, list) {
2754 		if (chain->ip >= PERF_CONTEXT_MAX)
2755 			continue;
2756 		if (chain->ms.sym && chain->ms.sym->ignore)
2757 			continue;
2758 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2759 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2760 							false));
2761 		first = false;
2762 	}
2763 
2764 	return ret;
2765 }
2766 
2767 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2768 {
2769 	size_t ret = 0;
2770 	FILE *fp = stdout;
2771 	struct callchain_node *chain;
2772 	struct rb_node *rb_node = rb_first_cached(root);
2773 
2774 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2775 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2776 	       graph_dotted_line);
2777 
2778 	while (rb_node) {
2779 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2780 		rb_node = rb_next(rb_node);
2781 
2782 		ret += fprintf(fp, "  ");
2783 		print_sched_time(chain->hit, 12);
2784 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2785 		ret += fprintf(fp, " %8d  ", chain->count);
2786 		ret += callchain__fprintf_folded(fp, chain);
2787 		ret += fprintf(fp, "\n");
2788 	}
2789 
2790 	return ret;
2791 }
2792 
2793 static void timehist_print_summary(struct perf_sched *sched,
2794 				   struct perf_session *session)
2795 {
2796 	struct machine *m = &session->machines.host;
2797 	struct total_run_stats totals;
2798 	u64 task_count;
2799 	struct thread *t;
2800 	struct thread_runtime *r;
2801 	int i;
2802 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2803 
2804 	memset(&totals, 0, sizeof(totals));
2805 	totals.sched = sched;
2806 
2807 	if (sched->idle_hist) {
2808 		printf("\nIdle-time summary\n");
2809 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2810 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2811 	} else if (sched->show_state) {
2812 		printf("\nWait-time summary\n");
2813 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2814 		printf("   run-time      sleep      iowait     preempt       delay\n");
2815 	} else {
2816 		printf("\nRuntime summary\n");
2817 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2818 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2819 	}
2820 	printf("%*s            (count)  ", comm_width, "");
2821 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2822 	       sched->show_state ? "(msec)" : "%");
2823 	printf("%.117s\n", graph_dotted_line);
2824 
2825 	machine__for_each_thread(m, show_thread_runtime, &totals);
2826 	task_count = totals.task_count;
2827 	if (!task_count)
2828 		printf("<no still running tasks>\n");
2829 
2830 	printf("\nTerminated tasks:\n");
2831 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2832 	if (task_count == totals.task_count)
2833 		printf("<no terminated tasks>\n");
2834 
2835 	/* CPU idle stats not tracked when samples were skipped */
2836 	if (sched->skipped_samples && !sched->idle_hist)
2837 		return;
2838 
2839 	printf("\nIdle stats:\n");
2840 	for (i = 0; i < idle_max_cpu; ++i) {
2841 		t = idle_threads[i];
2842 		if (!t)
2843 			continue;
2844 
2845 		r = thread__priv(t);
2846 		if (r && r->run_stats.n) {
2847 			totals.sched_count += r->run_stats.n;
2848 			printf("    CPU %2d idle for ", i);
2849 			print_sched_time(r->total_run_time, 6);
2850 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2851 		} else
2852 			printf("    CPU %2d idle entire time window\n", i);
2853 	}
2854 
2855 	if (sched->idle_hist && sched->show_callchain) {
2856 		callchain_param.mode  = CHAIN_FOLDED;
2857 		callchain_param.value = CCVAL_PERIOD;
2858 
2859 		callchain_register_param(&callchain_param);
2860 
2861 		printf("\nIdle stats by callchain:\n");
2862 		for (i = 0; i < idle_max_cpu; ++i) {
2863 			struct idle_thread_runtime *itr;
2864 
2865 			t = idle_threads[i];
2866 			if (!t)
2867 				continue;
2868 
2869 			itr = thread__priv(t);
2870 			if (itr == NULL)
2871 				continue;
2872 
2873 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2874 					     0, &callchain_param);
2875 
2876 			printf("  CPU %2d:", i);
2877 			print_sched_time(itr->tr.total_run_time, 6);
2878 			printf(" msec\n");
2879 			timehist_print_idlehist_callchain(&itr->sorted_root);
2880 			printf("\n");
2881 		}
2882 	}
2883 
2884 	printf("\n"
2885 	       "    Total number of unique tasks: %" PRIu64 "\n"
2886 	       "Total number of context switches: %" PRIu64 "\n",
2887 	       totals.task_count, totals.sched_count);
2888 
2889 	printf("           Total run time (msec): ");
2890 	print_sched_time(totals.total_run_time, 2);
2891 	printf("\n");
2892 
2893 	printf("    Total scheduling time (msec): ");
2894 	print_sched_time(hist_time, 2);
2895 	printf(" (x %d)\n", sched->max_cpu);
2896 }
2897 
2898 typedef int (*sched_handler)(struct perf_tool *tool,
2899 			  union perf_event *event,
2900 			  struct perf_evsel *evsel,
2901 			  struct perf_sample *sample,
2902 			  struct machine *machine);
2903 
2904 static int perf_timehist__process_sample(struct perf_tool *tool,
2905 					 union perf_event *event,
2906 					 struct perf_sample *sample,
2907 					 struct perf_evsel *evsel,
2908 					 struct machine *machine)
2909 {
2910 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2911 	int err = 0;
2912 	int this_cpu = sample->cpu;
2913 
2914 	if (this_cpu > sched->max_cpu)
2915 		sched->max_cpu = this_cpu;
2916 
2917 	if (evsel->handler != NULL) {
2918 		sched_handler f = evsel->handler;
2919 
2920 		err = f(tool, event, evsel, sample, machine);
2921 	}
2922 
2923 	return err;
2924 }
2925 
2926 static int timehist_check_attr(struct perf_sched *sched,
2927 			       struct perf_evlist *evlist)
2928 {
2929 	struct perf_evsel *evsel;
2930 	struct evsel_runtime *er;
2931 
2932 	list_for_each_entry(evsel, &evlist->entries, node) {
2933 		er = perf_evsel__get_runtime(evsel);
2934 		if (er == NULL) {
2935 			pr_err("Failed to allocate memory for evsel runtime data\n");
2936 			return -1;
2937 		}
2938 
2939 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2940 			pr_info("Samples do not have callchains.\n");
2941 			sched->show_callchain = 0;
2942 			symbol_conf.use_callchain = 0;
2943 		}
2944 	}
2945 
2946 	return 0;
2947 }
2948 
2949 static int perf_sched__timehist(struct perf_sched *sched)
2950 {
2951 	const struct perf_evsel_str_handler handlers[] = {
2952 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2953 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2954 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2955 	};
2956 	const struct perf_evsel_str_handler migrate_handlers[] = {
2957 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2958 	};
2959 	struct perf_data data = {
2960 		.path  = input_name,
2961 		.mode  = PERF_DATA_MODE_READ,
2962 		.force = sched->force,
2963 	};
2964 
2965 	struct perf_session *session;
2966 	struct perf_evlist *evlist;
2967 	int err = -1;
2968 
2969 	/*
2970 	 * event handlers for timehist option
2971 	 */
2972 	sched->tool.sample	 = perf_timehist__process_sample;
2973 	sched->tool.mmap	 = perf_event__process_mmap;
2974 	sched->tool.comm	 = perf_event__process_comm;
2975 	sched->tool.exit	 = perf_event__process_exit;
2976 	sched->tool.fork	 = perf_event__process_fork;
2977 	sched->tool.lost	 = process_lost;
2978 	sched->tool.attr	 = perf_event__process_attr;
2979 	sched->tool.tracing_data = perf_event__process_tracing_data;
2980 	sched->tool.build_id	 = perf_event__process_build_id;
2981 
2982 	sched->tool.ordered_events = true;
2983 	sched->tool.ordering_requires_timestamps = true;
2984 
2985 	symbol_conf.use_callchain = sched->show_callchain;
2986 
2987 	session = perf_session__new(&data, false, &sched->tool);
2988 	if (session == NULL)
2989 		return -ENOMEM;
2990 
2991 	evlist = session->evlist;
2992 
2993 	symbol__init(&session->header.env);
2994 
2995 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2996 		pr_err("Invalid time string\n");
2997 		return -EINVAL;
2998 	}
2999 
3000 	if (timehist_check_attr(sched, evlist) != 0)
3001 		goto out;
3002 
3003 	setup_pager();
3004 
3005 	/* setup per-evsel handlers */
3006 	if (perf_session__set_tracepoints_handlers(session, handlers))
3007 		goto out;
3008 
3009 	/* sched_switch event at a minimum needs to exist */
3010 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3011 						  "sched:sched_switch")) {
3012 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3013 		goto out;
3014 	}
3015 
3016 	if (sched->show_migrations &&
3017 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3018 		goto out;
3019 
3020 	/* pre-allocate struct for per-CPU idle stats */
3021 	sched->max_cpu = session->header.env.nr_cpus_online;
3022 	if (sched->max_cpu == 0)
3023 		sched->max_cpu = 4;
3024 	if (init_idle_threads(sched->max_cpu))
3025 		goto out;
3026 
3027 	/* summary_only implies summary option, but don't overwrite summary if set */
3028 	if (sched->summary_only)
3029 		sched->summary = sched->summary_only;
3030 
3031 	if (!sched->summary_only)
3032 		timehist_header(sched);
3033 
3034 	err = perf_session__process_events(session);
3035 	if (err) {
3036 		pr_err("Failed to process events, error %d", err);
3037 		goto out;
3038 	}
3039 
3040 	sched->nr_events      = evlist->stats.nr_events[0];
3041 	sched->nr_lost_events = evlist->stats.total_lost;
3042 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3043 
3044 	if (sched->summary)
3045 		timehist_print_summary(sched, session);
3046 
3047 out:
3048 	free_idle_threads();
3049 	perf_session__delete(session);
3050 
3051 	return err;
3052 }
3053 
3054 
3055 static void print_bad_events(struct perf_sched *sched)
3056 {
3057 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3058 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3059 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3060 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3061 	}
3062 	if (sched->nr_lost_events && sched->nr_events) {
3063 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3064 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3065 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3066 	}
3067 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3068 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3069 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3070 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3071 		if (sched->nr_lost_events)
3072 			printf(" (due to lost events?)");
3073 		printf("\n");
3074 	}
3075 }
3076 
3077 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3078 {
3079 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3080 	struct work_atoms *this;
3081 	const char *comm = thread__comm_str(data->thread), *this_comm;
3082 	bool leftmost = true;
3083 
3084 	while (*new) {
3085 		int cmp;
3086 
3087 		this = container_of(*new, struct work_atoms, node);
3088 		parent = *new;
3089 
3090 		this_comm = thread__comm_str(this->thread);
3091 		cmp = strcmp(comm, this_comm);
3092 		if (cmp > 0) {
3093 			new = &((*new)->rb_left);
3094 		} else if (cmp < 0) {
3095 			new = &((*new)->rb_right);
3096 			leftmost = false;
3097 		} else {
3098 			this->num_merged++;
3099 			this->total_runtime += data->total_runtime;
3100 			this->nb_atoms += data->nb_atoms;
3101 			this->total_lat += data->total_lat;
3102 			list_splice(&data->work_list, &this->work_list);
3103 			if (this->max_lat < data->max_lat) {
3104 				this->max_lat = data->max_lat;
3105 				this->max_lat_at = data->max_lat_at;
3106 			}
3107 			zfree(&data);
3108 			return;
3109 		}
3110 	}
3111 
3112 	data->num_merged++;
3113 	rb_link_node(&data->node, parent, new);
3114 	rb_insert_color_cached(&data->node, root, leftmost);
3115 }
3116 
3117 static void perf_sched__merge_lat(struct perf_sched *sched)
3118 {
3119 	struct work_atoms *data;
3120 	struct rb_node *node;
3121 
3122 	if (sched->skip_merge)
3123 		return;
3124 
3125 	while ((node = rb_first_cached(&sched->atom_root))) {
3126 		rb_erase_cached(node, &sched->atom_root);
3127 		data = rb_entry(node, struct work_atoms, node);
3128 		__merge_work_atoms(&sched->merged_atom_root, data);
3129 	}
3130 }
3131 
3132 static int perf_sched__lat(struct perf_sched *sched)
3133 {
3134 	struct rb_node *next;
3135 
3136 	setup_pager();
3137 
3138 	if (perf_sched__read_events(sched))
3139 		return -1;
3140 
3141 	perf_sched__merge_lat(sched);
3142 	perf_sched__sort_lat(sched);
3143 
3144 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3145 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3146 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3147 
3148 	next = rb_first_cached(&sched->sorted_atom_root);
3149 
3150 	while (next) {
3151 		struct work_atoms *work_list;
3152 
3153 		work_list = rb_entry(next, struct work_atoms, node);
3154 		output_lat_thread(sched, work_list);
3155 		next = rb_next(next);
3156 		thread__zput(work_list->thread);
3157 	}
3158 
3159 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3160 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3161 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3162 
3163 	printf(" ---------------------------------------------------\n");
3164 
3165 	print_bad_events(sched);
3166 	printf("\n");
3167 
3168 	return 0;
3169 }
3170 
3171 static int setup_map_cpus(struct perf_sched *sched)
3172 {
3173 	struct cpu_map *map;
3174 
3175 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3176 
3177 	if (sched->map.comp) {
3178 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3179 		if (!sched->map.comp_cpus)
3180 			return -1;
3181 	}
3182 
3183 	if (!sched->map.cpus_str)
3184 		return 0;
3185 
3186 	map = cpu_map__new(sched->map.cpus_str);
3187 	if (!map) {
3188 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3189 		return -1;
3190 	}
3191 
3192 	sched->map.cpus = map;
3193 	return 0;
3194 }
3195 
3196 static int setup_color_pids(struct perf_sched *sched)
3197 {
3198 	struct thread_map *map;
3199 
3200 	if (!sched->map.color_pids_str)
3201 		return 0;
3202 
3203 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3204 	if (!map) {
3205 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3206 		return -1;
3207 	}
3208 
3209 	sched->map.color_pids = map;
3210 	return 0;
3211 }
3212 
3213 static int setup_color_cpus(struct perf_sched *sched)
3214 {
3215 	struct cpu_map *map;
3216 
3217 	if (!sched->map.color_cpus_str)
3218 		return 0;
3219 
3220 	map = cpu_map__new(sched->map.color_cpus_str);
3221 	if (!map) {
3222 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3223 		return -1;
3224 	}
3225 
3226 	sched->map.color_cpus = map;
3227 	return 0;
3228 }
3229 
3230 static int perf_sched__map(struct perf_sched *sched)
3231 {
3232 	if (setup_map_cpus(sched))
3233 		return -1;
3234 
3235 	if (setup_color_pids(sched))
3236 		return -1;
3237 
3238 	if (setup_color_cpus(sched))
3239 		return -1;
3240 
3241 	setup_pager();
3242 	if (perf_sched__read_events(sched))
3243 		return -1;
3244 	print_bad_events(sched);
3245 	return 0;
3246 }
3247 
3248 static int perf_sched__replay(struct perf_sched *sched)
3249 {
3250 	unsigned long i;
3251 
3252 	calibrate_run_measurement_overhead(sched);
3253 	calibrate_sleep_measurement_overhead(sched);
3254 
3255 	test_calibrations(sched);
3256 
3257 	if (perf_sched__read_events(sched))
3258 		return -1;
3259 
3260 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3261 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3262 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3263 
3264 	if (sched->targetless_wakeups)
3265 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3266 	if (sched->multitarget_wakeups)
3267 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3268 	if (sched->nr_run_events_optimized)
3269 		printf("run atoms optimized: %ld\n",
3270 			sched->nr_run_events_optimized);
3271 
3272 	print_task_traces(sched);
3273 	add_cross_task_wakeups(sched);
3274 
3275 	create_tasks(sched);
3276 	printf("------------------------------------------------------------\n");
3277 	for (i = 0; i < sched->replay_repeat; i++)
3278 		run_one_test(sched);
3279 
3280 	return 0;
3281 }
3282 
3283 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3284 			  const char * const usage_msg[])
3285 {
3286 	char *tmp, *tok, *str = strdup(sched->sort_order);
3287 
3288 	for (tok = strtok_r(str, ", ", &tmp);
3289 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3290 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3291 			usage_with_options_msg(usage_msg, options,
3292 					"Unknown --sort key: `%s'", tok);
3293 		}
3294 	}
3295 
3296 	free(str);
3297 
3298 	sort_dimension__add("pid", &sched->cmp_pid);
3299 }
3300 
3301 static int __cmd_record(int argc, const char **argv)
3302 {
3303 	unsigned int rec_argc, i, j;
3304 	const char **rec_argv;
3305 	const char * const record_args[] = {
3306 		"record",
3307 		"-a",
3308 		"-R",
3309 		"-m", "1024",
3310 		"-c", "1",
3311 		"-e", "sched:sched_switch",
3312 		"-e", "sched:sched_stat_wait",
3313 		"-e", "sched:sched_stat_sleep",
3314 		"-e", "sched:sched_stat_iowait",
3315 		"-e", "sched:sched_stat_runtime",
3316 		"-e", "sched:sched_process_fork",
3317 		"-e", "sched:sched_wakeup",
3318 		"-e", "sched:sched_wakeup_new",
3319 		"-e", "sched:sched_migrate_task",
3320 	};
3321 
3322 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3323 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3324 
3325 	if (rec_argv == NULL)
3326 		return -ENOMEM;
3327 
3328 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3329 		rec_argv[i] = strdup(record_args[i]);
3330 
3331 	for (j = 1; j < (unsigned int)argc; j++, i++)
3332 		rec_argv[i] = argv[j];
3333 
3334 	BUG_ON(i != rec_argc);
3335 
3336 	return cmd_record(i, rec_argv);
3337 }
3338 
3339 int cmd_sched(int argc, const char **argv)
3340 {
3341 	static const char default_sort_order[] = "avg, max, switch, runtime";
3342 	struct perf_sched sched = {
3343 		.tool = {
3344 			.sample		 = perf_sched__process_tracepoint_sample,
3345 			.comm		 = perf_sched__process_comm,
3346 			.namespaces	 = perf_event__process_namespaces,
3347 			.lost		 = perf_event__process_lost,
3348 			.fork		 = perf_sched__process_fork_event,
3349 			.ordered_events = true,
3350 		},
3351 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3352 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3353 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3354 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3355 		.sort_order	      = default_sort_order,
3356 		.replay_repeat	      = 10,
3357 		.profile_cpu	      = -1,
3358 		.next_shortname1      = 'A',
3359 		.next_shortname2      = '0',
3360 		.skip_merge           = 0,
3361 		.show_callchain	      = 1,
3362 		.max_stack            = 5,
3363 	};
3364 	const struct option sched_options[] = {
3365 	OPT_STRING('i', "input", &input_name, "file",
3366 		    "input file name"),
3367 	OPT_INCR('v', "verbose", &verbose,
3368 		    "be more verbose (show symbol address, etc)"),
3369 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3370 		    "dump raw trace in ASCII"),
3371 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3372 	OPT_END()
3373 	};
3374 	const struct option latency_options[] = {
3375 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3376 		   "sort by key(s): runtime, switch, avg, max"),
3377 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3378 		    "CPU to profile on"),
3379 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3380 		    "latency stats per pid instead of per comm"),
3381 	OPT_PARENT(sched_options)
3382 	};
3383 	const struct option replay_options[] = {
3384 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3385 		     "repeat the workload replay N times (-1: infinite)"),
3386 	OPT_PARENT(sched_options)
3387 	};
3388 	const struct option map_options[] = {
3389 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3390 		    "map output in compact mode"),
3391 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3392 		   "highlight given pids in map"),
3393 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3394                     "highlight given CPUs in map"),
3395 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3396                     "display given CPUs in map"),
3397 	OPT_PARENT(sched_options)
3398 	};
3399 	const struct option timehist_options[] = {
3400 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3401 		   "file", "vmlinux pathname"),
3402 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3403 		   "file", "kallsyms pathname"),
3404 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3405 		    "Display call chains if present (default on)"),
3406 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3407 		   "Maximum number of functions to display backtrace."),
3408 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3409 		    "Look for files with symbols relative to this directory"),
3410 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3411 		    "Show only syscall summary with statistics"),
3412 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3413 		    "Show all syscalls and summary with statistics"),
3414 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3415 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3416 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3417 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3418 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3419 	OPT_STRING(0, "time", &sched.time_str, "str",
3420 		   "Time span for analysis (start,stop)"),
3421 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3422 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3423 		   "analyze events only for given process id(s)"),
3424 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3425 		   "analyze events only for given thread id(s)"),
3426 	OPT_PARENT(sched_options)
3427 	};
3428 
3429 	const char * const latency_usage[] = {
3430 		"perf sched latency [<options>]",
3431 		NULL
3432 	};
3433 	const char * const replay_usage[] = {
3434 		"perf sched replay [<options>]",
3435 		NULL
3436 	};
3437 	const char * const map_usage[] = {
3438 		"perf sched map [<options>]",
3439 		NULL
3440 	};
3441 	const char * const timehist_usage[] = {
3442 		"perf sched timehist [<options>]",
3443 		NULL
3444 	};
3445 	const char *const sched_subcommands[] = { "record", "latency", "map",
3446 						  "replay", "script",
3447 						  "timehist", NULL };
3448 	const char *sched_usage[] = {
3449 		NULL,
3450 		NULL
3451 	};
3452 	struct trace_sched_handler lat_ops  = {
3453 		.wakeup_event	    = latency_wakeup_event,
3454 		.switch_event	    = latency_switch_event,
3455 		.runtime_event	    = latency_runtime_event,
3456 		.migrate_task_event = latency_migrate_task_event,
3457 	};
3458 	struct trace_sched_handler map_ops  = {
3459 		.switch_event	    = map_switch_event,
3460 	};
3461 	struct trace_sched_handler replay_ops  = {
3462 		.wakeup_event	    = replay_wakeup_event,
3463 		.switch_event	    = replay_switch_event,
3464 		.fork_event	    = replay_fork_event,
3465 	};
3466 	unsigned int i;
3467 
3468 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3469 		sched.curr_pid[i] = -1;
3470 
3471 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3472 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3473 	if (!argc)
3474 		usage_with_options(sched_usage, sched_options);
3475 
3476 	/*
3477 	 * Aliased to 'perf script' for now:
3478 	 */
3479 	if (!strcmp(argv[0], "script"))
3480 		return cmd_script(argc, argv);
3481 
3482 	if (!strncmp(argv[0], "rec", 3)) {
3483 		return __cmd_record(argc, argv);
3484 	} else if (!strncmp(argv[0], "lat", 3)) {
3485 		sched.tp_handler = &lat_ops;
3486 		if (argc > 1) {
3487 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3488 			if (argc)
3489 				usage_with_options(latency_usage, latency_options);
3490 		}
3491 		setup_sorting(&sched, latency_options, latency_usage);
3492 		return perf_sched__lat(&sched);
3493 	} else if (!strcmp(argv[0], "map")) {
3494 		if (argc) {
3495 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3496 			if (argc)
3497 				usage_with_options(map_usage, map_options);
3498 		}
3499 		sched.tp_handler = &map_ops;
3500 		setup_sorting(&sched, latency_options, latency_usage);
3501 		return perf_sched__map(&sched);
3502 	} else if (!strncmp(argv[0], "rep", 3)) {
3503 		sched.tp_handler = &replay_ops;
3504 		if (argc) {
3505 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3506 			if (argc)
3507 				usage_with_options(replay_usage, replay_options);
3508 		}
3509 		return perf_sched__replay(&sched);
3510 	} else if (!strcmp(argv[0], "timehist")) {
3511 		if (argc) {
3512 			argc = parse_options(argc, argv, timehist_options,
3513 					     timehist_usage, 0);
3514 			if (argc)
3515 				usage_with_options(timehist_usage, timehist_options);
3516 		}
3517 		if ((sched.show_wakeups || sched.show_next) &&
3518 		    sched.summary_only) {
3519 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3520 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3521 			if (sched.show_wakeups)
3522 				parse_options_usage(NULL, timehist_options, "w", true);
3523 			if (sched.show_next)
3524 				parse_options_usage(NULL, timehist_options, "n", true);
3525 			return -EINVAL;
3526 		}
3527 
3528 		return perf_sched__timehist(&sched);
3529 	} else {
3530 		usage_with_options(sched_usage, sched_options);
3531 	}
3532 
3533 	return 0;
3534 }
3535