xref: /linux/tools/perf/builtin-sched.c (revision b2d0f5d5dc53532e6f07bc546a476a55ebdfe0f3)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 
5 #include "util/util.h"
6 #include "util/evlist.h"
7 #include "util/cache.h"
8 #include "util/evsel.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/tool.h"
14 #include "util/cloexec.h"
15 #include "util/thread_map.h"
16 #include "util/color.h"
17 #include "util/stat.h"
18 #include "util/callchain.h"
19 #include "util/time-utils.h"
20 
21 #include <subcmd/parse-options.h>
22 #include "util/trace-event.h"
23 
24 #include "util/debug.h"
25 
26 #include <linux/kernel.h>
27 #include <linux/log2.h>
28 #include <sys/prctl.h>
29 #include <sys/resource.h>
30 #include <inttypes.h>
31 
32 #include <errno.h>
33 #include <semaphore.h>
34 #include <pthread.h>
35 #include <math.h>
36 #include <api/fs/fs.h>
37 #include <linux/time64.h>
38 
39 #include "sane_ctype.h"
40 
41 #define PR_SET_NAME		15               /* Set process name */
42 #define MAX_CPUS		4096
43 #define COMM_LEN		20
44 #define SYM_LEN			129
45 #define MAX_PID			1024000
46 
47 struct sched_atom;
48 
49 struct task_desc {
50 	unsigned long		nr;
51 	unsigned long		pid;
52 	char			comm[COMM_LEN];
53 
54 	unsigned long		nr_events;
55 	unsigned long		curr_event;
56 	struct sched_atom	**atoms;
57 
58 	pthread_t		thread;
59 	sem_t			sleep_sem;
60 
61 	sem_t			ready_for_work;
62 	sem_t			work_done_sem;
63 
64 	u64			cpu_usage;
65 };
66 
67 enum sched_event_type {
68 	SCHED_EVENT_RUN,
69 	SCHED_EVENT_SLEEP,
70 	SCHED_EVENT_WAKEUP,
71 	SCHED_EVENT_MIGRATION,
72 };
73 
74 struct sched_atom {
75 	enum sched_event_type	type;
76 	int			specific_wait;
77 	u64			timestamp;
78 	u64			duration;
79 	unsigned long		nr;
80 	sem_t			*wait_sem;
81 	struct task_desc	*wakee;
82 };
83 
84 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
85 
86 /* task state bitmask, copied from include/linux/sched.h */
87 #define TASK_RUNNING		0
88 #define TASK_INTERRUPTIBLE	1
89 #define TASK_UNINTERRUPTIBLE	2
90 #define __TASK_STOPPED		4
91 #define __TASK_TRACED		8
92 /* in tsk->exit_state */
93 #define EXIT_DEAD		16
94 #define EXIT_ZOMBIE		32
95 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
96 /* in tsk->state again */
97 #define TASK_DEAD		64
98 #define TASK_WAKEKILL		128
99 #define TASK_WAKING		256
100 #define TASK_PARKED		512
101 
102 enum thread_state {
103 	THREAD_SLEEPING = 0,
104 	THREAD_WAIT_CPU,
105 	THREAD_SCHED_IN,
106 	THREAD_IGNORE
107 };
108 
109 struct work_atom {
110 	struct list_head	list;
111 	enum thread_state	state;
112 	u64			sched_out_time;
113 	u64			wake_up_time;
114 	u64			sched_in_time;
115 	u64			runtime;
116 };
117 
118 struct work_atoms {
119 	struct list_head	work_list;
120 	struct thread		*thread;
121 	struct rb_node		node;
122 	u64			max_lat;
123 	u64			max_lat_at;
124 	u64			total_lat;
125 	u64			nb_atoms;
126 	u64			total_runtime;
127 	int			num_merged;
128 };
129 
130 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
131 
132 struct perf_sched;
133 
134 struct trace_sched_handler {
135 	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
136 			    struct perf_sample *sample, struct machine *machine);
137 
138 	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
139 			     struct perf_sample *sample, struct machine *machine);
140 
141 	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
142 			    struct perf_sample *sample, struct machine *machine);
143 
144 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
145 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
146 			  struct machine *machine);
147 
148 	int (*migrate_task_event)(struct perf_sched *sched,
149 				  struct perf_evsel *evsel,
150 				  struct perf_sample *sample,
151 				  struct machine *machine);
152 };
153 
154 #define COLOR_PIDS PERF_COLOR_BLUE
155 #define COLOR_CPUS PERF_COLOR_BG_RED
156 
157 struct perf_sched_map {
158 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
159 	int			*comp_cpus;
160 	bool			 comp;
161 	struct thread_map	*color_pids;
162 	const char		*color_pids_str;
163 	struct cpu_map		*color_cpus;
164 	const char		*color_cpus_str;
165 	struct cpu_map		*cpus;
166 	const char		*cpus_str;
167 };
168 
169 struct perf_sched {
170 	struct perf_tool tool;
171 	const char	 *sort_order;
172 	unsigned long	 nr_tasks;
173 	struct task_desc **pid_to_task;
174 	struct task_desc **tasks;
175 	const struct trace_sched_handler *tp_handler;
176 	pthread_mutex_t	 start_work_mutex;
177 	pthread_mutex_t	 work_done_wait_mutex;
178 	int		 profile_cpu;
179 /*
180  * Track the current task - that way we can know whether there's any
181  * weird events, such as a task being switched away that is not current.
182  */
183 	int		 max_cpu;
184 	u32		 curr_pid[MAX_CPUS];
185 	struct thread	 *curr_thread[MAX_CPUS];
186 	char		 next_shortname1;
187 	char		 next_shortname2;
188 	unsigned int	 replay_repeat;
189 	unsigned long	 nr_run_events;
190 	unsigned long	 nr_sleep_events;
191 	unsigned long	 nr_wakeup_events;
192 	unsigned long	 nr_sleep_corrections;
193 	unsigned long	 nr_run_events_optimized;
194 	unsigned long	 targetless_wakeups;
195 	unsigned long	 multitarget_wakeups;
196 	unsigned long	 nr_runs;
197 	unsigned long	 nr_timestamps;
198 	unsigned long	 nr_unordered_timestamps;
199 	unsigned long	 nr_context_switch_bugs;
200 	unsigned long	 nr_events;
201 	unsigned long	 nr_lost_chunks;
202 	unsigned long	 nr_lost_events;
203 	u64		 run_measurement_overhead;
204 	u64		 sleep_measurement_overhead;
205 	u64		 start_time;
206 	u64		 cpu_usage;
207 	u64		 runavg_cpu_usage;
208 	u64		 parent_cpu_usage;
209 	u64		 runavg_parent_cpu_usage;
210 	u64		 sum_runtime;
211 	u64		 sum_fluct;
212 	u64		 run_avg;
213 	u64		 all_runtime;
214 	u64		 all_count;
215 	u64		 cpu_last_switched[MAX_CPUS];
216 	struct rb_root	 atom_root, sorted_atom_root, merged_atom_root;
217 	struct list_head sort_list, cmp_pid;
218 	bool force;
219 	bool skip_merge;
220 	struct perf_sched_map map;
221 
222 	/* options for timehist command */
223 	bool		summary;
224 	bool		summary_only;
225 	bool		idle_hist;
226 	bool		show_callchain;
227 	unsigned int	max_stack;
228 	bool		show_cpu_visual;
229 	bool		show_wakeups;
230 	bool		show_next;
231 	bool		show_migrations;
232 	bool		show_state;
233 	u64		skipped_samples;
234 	const char	*time_str;
235 	struct perf_time_interval ptime;
236 	struct perf_time_interval hist_time;
237 };
238 
239 /* per thread run time data */
240 struct thread_runtime {
241 	u64 last_time;      /* time of previous sched in/out event */
242 	u64 dt_run;         /* run time */
243 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
244 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
245 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
246 	u64 dt_delay;       /* time between wakeup and sched-in */
247 	u64 ready_to_run;   /* time of wakeup */
248 
249 	struct stats run_stats;
250 	u64 total_run_time;
251 	u64 total_sleep_time;
252 	u64 total_iowait_time;
253 	u64 total_preempt_time;
254 	u64 total_delay_time;
255 
256 	int last_state;
257 	u64 migrations;
258 };
259 
260 /* per event run time data */
261 struct evsel_runtime {
262 	u64 *last_time; /* time this event was last seen per cpu */
263 	u32 ncpu;       /* highest cpu slot allocated */
264 };
265 
266 /* per cpu idle time data */
267 struct idle_thread_runtime {
268 	struct thread_runtime	tr;
269 	struct thread		*last_thread;
270 	struct rb_root		sorted_root;
271 	struct callchain_root	callchain;
272 	struct callchain_cursor	cursor;
273 };
274 
275 /* track idle times per cpu */
276 static struct thread **idle_threads;
277 static int idle_max_cpu;
278 static char idle_comm[] = "<idle>";
279 
280 static u64 get_nsecs(void)
281 {
282 	struct timespec ts;
283 
284 	clock_gettime(CLOCK_MONOTONIC, &ts);
285 
286 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
287 }
288 
289 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
290 {
291 	u64 T0 = get_nsecs(), T1;
292 
293 	do {
294 		T1 = get_nsecs();
295 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
296 }
297 
298 static void sleep_nsecs(u64 nsecs)
299 {
300 	struct timespec ts;
301 
302 	ts.tv_nsec = nsecs % 999999999;
303 	ts.tv_sec = nsecs / 999999999;
304 
305 	nanosleep(&ts, NULL);
306 }
307 
308 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
309 {
310 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
311 	int i;
312 
313 	for (i = 0; i < 10; i++) {
314 		T0 = get_nsecs();
315 		burn_nsecs(sched, 0);
316 		T1 = get_nsecs();
317 		delta = T1-T0;
318 		min_delta = min(min_delta, delta);
319 	}
320 	sched->run_measurement_overhead = min_delta;
321 
322 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
323 }
324 
325 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
326 {
327 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 	int i;
329 
330 	for (i = 0; i < 10; i++) {
331 		T0 = get_nsecs();
332 		sleep_nsecs(10000);
333 		T1 = get_nsecs();
334 		delta = T1-T0;
335 		min_delta = min(min_delta, delta);
336 	}
337 	min_delta -= 10000;
338 	sched->sleep_measurement_overhead = min_delta;
339 
340 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
341 }
342 
343 static struct sched_atom *
344 get_new_event(struct task_desc *task, u64 timestamp)
345 {
346 	struct sched_atom *event = zalloc(sizeof(*event));
347 	unsigned long idx = task->nr_events;
348 	size_t size;
349 
350 	event->timestamp = timestamp;
351 	event->nr = idx;
352 
353 	task->nr_events++;
354 	size = sizeof(struct sched_atom *) * task->nr_events;
355 	task->atoms = realloc(task->atoms, size);
356 	BUG_ON(!task->atoms);
357 
358 	task->atoms[idx] = event;
359 
360 	return event;
361 }
362 
363 static struct sched_atom *last_event(struct task_desc *task)
364 {
365 	if (!task->nr_events)
366 		return NULL;
367 
368 	return task->atoms[task->nr_events - 1];
369 }
370 
371 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
372 				u64 timestamp, u64 duration)
373 {
374 	struct sched_atom *event, *curr_event = last_event(task);
375 
376 	/*
377 	 * optimize an existing RUN event by merging this one
378 	 * to it:
379 	 */
380 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
381 		sched->nr_run_events_optimized++;
382 		curr_event->duration += duration;
383 		return;
384 	}
385 
386 	event = get_new_event(task, timestamp);
387 
388 	event->type = SCHED_EVENT_RUN;
389 	event->duration = duration;
390 
391 	sched->nr_run_events++;
392 }
393 
394 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
395 				   u64 timestamp, struct task_desc *wakee)
396 {
397 	struct sched_atom *event, *wakee_event;
398 
399 	event = get_new_event(task, timestamp);
400 	event->type = SCHED_EVENT_WAKEUP;
401 	event->wakee = wakee;
402 
403 	wakee_event = last_event(wakee);
404 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
405 		sched->targetless_wakeups++;
406 		return;
407 	}
408 	if (wakee_event->wait_sem) {
409 		sched->multitarget_wakeups++;
410 		return;
411 	}
412 
413 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
414 	sem_init(wakee_event->wait_sem, 0, 0);
415 	wakee_event->specific_wait = 1;
416 	event->wait_sem = wakee_event->wait_sem;
417 
418 	sched->nr_wakeup_events++;
419 }
420 
421 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
422 				  u64 timestamp, u64 task_state __maybe_unused)
423 {
424 	struct sched_atom *event = get_new_event(task, timestamp);
425 
426 	event->type = SCHED_EVENT_SLEEP;
427 
428 	sched->nr_sleep_events++;
429 }
430 
431 static struct task_desc *register_pid(struct perf_sched *sched,
432 				      unsigned long pid, const char *comm)
433 {
434 	struct task_desc *task;
435 	static int pid_max;
436 
437 	if (sched->pid_to_task == NULL) {
438 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
439 			pid_max = MAX_PID;
440 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
441 	}
442 	if (pid >= (unsigned long)pid_max) {
443 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
444 			sizeof(struct task_desc *))) == NULL);
445 		while (pid >= (unsigned long)pid_max)
446 			sched->pid_to_task[pid_max++] = NULL;
447 	}
448 
449 	task = sched->pid_to_task[pid];
450 
451 	if (task)
452 		return task;
453 
454 	task = zalloc(sizeof(*task));
455 	task->pid = pid;
456 	task->nr = sched->nr_tasks;
457 	strcpy(task->comm, comm);
458 	/*
459 	 * every task starts in sleeping state - this gets ignored
460 	 * if there's no wakeup pointing to this sleep state:
461 	 */
462 	add_sched_event_sleep(sched, task, 0, 0);
463 
464 	sched->pid_to_task[pid] = task;
465 	sched->nr_tasks++;
466 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
467 	BUG_ON(!sched->tasks);
468 	sched->tasks[task->nr] = task;
469 
470 	if (verbose > 0)
471 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
472 
473 	return task;
474 }
475 
476 
477 static void print_task_traces(struct perf_sched *sched)
478 {
479 	struct task_desc *task;
480 	unsigned long i;
481 
482 	for (i = 0; i < sched->nr_tasks; i++) {
483 		task = sched->tasks[i];
484 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
485 			task->nr, task->comm, task->pid, task->nr_events);
486 	}
487 }
488 
489 static void add_cross_task_wakeups(struct perf_sched *sched)
490 {
491 	struct task_desc *task1, *task2;
492 	unsigned long i, j;
493 
494 	for (i = 0; i < sched->nr_tasks; i++) {
495 		task1 = sched->tasks[i];
496 		j = i + 1;
497 		if (j == sched->nr_tasks)
498 			j = 0;
499 		task2 = sched->tasks[j];
500 		add_sched_event_wakeup(sched, task1, 0, task2);
501 	}
502 }
503 
504 static void perf_sched__process_event(struct perf_sched *sched,
505 				      struct sched_atom *atom)
506 {
507 	int ret = 0;
508 
509 	switch (atom->type) {
510 		case SCHED_EVENT_RUN:
511 			burn_nsecs(sched, atom->duration);
512 			break;
513 		case SCHED_EVENT_SLEEP:
514 			if (atom->wait_sem)
515 				ret = sem_wait(atom->wait_sem);
516 			BUG_ON(ret);
517 			break;
518 		case SCHED_EVENT_WAKEUP:
519 			if (atom->wait_sem)
520 				ret = sem_post(atom->wait_sem);
521 			BUG_ON(ret);
522 			break;
523 		case SCHED_EVENT_MIGRATION:
524 			break;
525 		default:
526 			BUG_ON(1);
527 	}
528 }
529 
530 static u64 get_cpu_usage_nsec_parent(void)
531 {
532 	struct rusage ru;
533 	u64 sum;
534 	int err;
535 
536 	err = getrusage(RUSAGE_SELF, &ru);
537 	BUG_ON(err);
538 
539 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
540 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
541 
542 	return sum;
543 }
544 
545 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
546 {
547 	struct perf_event_attr attr;
548 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
549 	int fd;
550 	struct rlimit limit;
551 	bool need_privilege = false;
552 
553 	memset(&attr, 0, sizeof(attr));
554 
555 	attr.type = PERF_TYPE_SOFTWARE;
556 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
557 
558 force_again:
559 	fd = sys_perf_event_open(&attr, 0, -1, -1,
560 				 perf_event_open_cloexec_flag());
561 
562 	if (fd < 0) {
563 		if (errno == EMFILE) {
564 			if (sched->force) {
565 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
566 				limit.rlim_cur += sched->nr_tasks - cur_task;
567 				if (limit.rlim_cur > limit.rlim_max) {
568 					limit.rlim_max = limit.rlim_cur;
569 					need_privilege = true;
570 				}
571 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
572 					if (need_privilege && errno == EPERM)
573 						strcpy(info, "Need privilege\n");
574 				} else
575 					goto force_again;
576 			} else
577 				strcpy(info, "Have a try with -f option\n");
578 		}
579 		pr_err("Error: sys_perf_event_open() syscall returned "
580 		       "with %d (%s)\n%s", fd,
581 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
582 		exit(EXIT_FAILURE);
583 	}
584 	return fd;
585 }
586 
587 static u64 get_cpu_usage_nsec_self(int fd)
588 {
589 	u64 runtime;
590 	int ret;
591 
592 	ret = read(fd, &runtime, sizeof(runtime));
593 	BUG_ON(ret != sizeof(runtime));
594 
595 	return runtime;
596 }
597 
598 struct sched_thread_parms {
599 	struct task_desc  *task;
600 	struct perf_sched *sched;
601 	int fd;
602 };
603 
604 static void *thread_func(void *ctx)
605 {
606 	struct sched_thread_parms *parms = ctx;
607 	struct task_desc *this_task = parms->task;
608 	struct perf_sched *sched = parms->sched;
609 	u64 cpu_usage_0, cpu_usage_1;
610 	unsigned long i, ret;
611 	char comm2[22];
612 	int fd = parms->fd;
613 
614 	zfree(&parms);
615 
616 	sprintf(comm2, ":%s", this_task->comm);
617 	prctl(PR_SET_NAME, comm2);
618 	if (fd < 0)
619 		return NULL;
620 again:
621 	ret = sem_post(&this_task->ready_for_work);
622 	BUG_ON(ret);
623 	ret = pthread_mutex_lock(&sched->start_work_mutex);
624 	BUG_ON(ret);
625 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
626 	BUG_ON(ret);
627 
628 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
629 
630 	for (i = 0; i < this_task->nr_events; i++) {
631 		this_task->curr_event = i;
632 		perf_sched__process_event(sched, this_task->atoms[i]);
633 	}
634 
635 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
636 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
637 	ret = sem_post(&this_task->work_done_sem);
638 	BUG_ON(ret);
639 
640 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
641 	BUG_ON(ret);
642 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
643 	BUG_ON(ret);
644 
645 	goto again;
646 }
647 
648 static void create_tasks(struct perf_sched *sched)
649 {
650 	struct task_desc *task;
651 	pthread_attr_t attr;
652 	unsigned long i;
653 	int err;
654 
655 	err = pthread_attr_init(&attr);
656 	BUG_ON(err);
657 	err = pthread_attr_setstacksize(&attr,
658 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
659 	BUG_ON(err);
660 	err = pthread_mutex_lock(&sched->start_work_mutex);
661 	BUG_ON(err);
662 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
663 	BUG_ON(err);
664 	for (i = 0; i < sched->nr_tasks; i++) {
665 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
666 		BUG_ON(parms == NULL);
667 		parms->task = task = sched->tasks[i];
668 		parms->sched = sched;
669 		parms->fd = self_open_counters(sched, i);
670 		sem_init(&task->sleep_sem, 0, 0);
671 		sem_init(&task->ready_for_work, 0, 0);
672 		sem_init(&task->work_done_sem, 0, 0);
673 		task->curr_event = 0;
674 		err = pthread_create(&task->thread, &attr, thread_func, parms);
675 		BUG_ON(err);
676 	}
677 }
678 
679 static void wait_for_tasks(struct perf_sched *sched)
680 {
681 	u64 cpu_usage_0, cpu_usage_1;
682 	struct task_desc *task;
683 	unsigned long i, ret;
684 
685 	sched->start_time = get_nsecs();
686 	sched->cpu_usage = 0;
687 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
688 
689 	for (i = 0; i < sched->nr_tasks; i++) {
690 		task = sched->tasks[i];
691 		ret = sem_wait(&task->ready_for_work);
692 		BUG_ON(ret);
693 		sem_init(&task->ready_for_work, 0, 0);
694 	}
695 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
696 	BUG_ON(ret);
697 
698 	cpu_usage_0 = get_cpu_usage_nsec_parent();
699 
700 	pthread_mutex_unlock(&sched->start_work_mutex);
701 
702 	for (i = 0; i < sched->nr_tasks; i++) {
703 		task = sched->tasks[i];
704 		ret = sem_wait(&task->work_done_sem);
705 		BUG_ON(ret);
706 		sem_init(&task->work_done_sem, 0, 0);
707 		sched->cpu_usage += task->cpu_usage;
708 		task->cpu_usage = 0;
709 	}
710 
711 	cpu_usage_1 = get_cpu_usage_nsec_parent();
712 	if (!sched->runavg_cpu_usage)
713 		sched->runavg_cpu_usage = sched->cpu_usage;
714 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
715 
716 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
717 	if (!sched->runavg_parent_cpu_usage)
718 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
719 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
720 					 sched->parent_cpu_usage)/sched->replay_repeat;
721 
722 	ret = pthread_mutex_lock(&sched->start_work_mutex);
723 	BUG_ON(ret);
724 
725 	for (i = 0; i < sched->nr_tasks; i++) {
726 		task = sched->tasks[i];
727 		sem_init(&task->sleep_sem, 0, 0);
728 		task->curr_event = 0;
729 	}
730 }
731 
732 static void run_one_test(struct perf_sched *sched)
733 {
734 	u64 T0, T1, delta, avg_delta, fluct;
735 
736 	T0 = get_nsecs();
737 	wait_for_tasks(sched);
738 	T1 = get_nsecs();
739 
740 	delta = T1 - T0;
741 	sched->sum_runtime += delta;
742 	sched->nr_runs++;
743 
744 	avg_delta = sched->sum_runtime / sched->nr_runs;
745 	if (delta < avg_delta)
746 		fluct = avg_delta - delta;
747 	else
748 		fluct = delta - avg_delta;
749 	sched->sum_fluct += fluct;
750 	if (!sched->run_avg)
751 		sched->run_avg = delta;
752 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
753 
754 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
755 
756 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
757 
758 	printf("cpu: %0.2f / %0.2f",
759 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
760 
761 #if 0
762 	/*
763 	 * rusage statistics done by the parent, these are less
764 	 * accurate than the sched->sum_exec_runtime based statistics:
765 	 */
766 	printf(" [%0.2f / %0.2f]",
767 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
768 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
769 #endif
770 
771 	printf("\n");
772 
773 	if (sched->nr_sleep_corrections)
774 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
775 	sched->nr_sleep_corrections = 0;
776 }
777 
778 static void test_calibrations(struct perf_sched *sched)
779 {
780 	u64 T0, T1;
781 
782 	T0 = get_nsecs();
783 	burn_nsecs(sched, NSEC_PER_MSEC);
784 	T1 = get_nsecs();
785 
786 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
787 
788 	T0 = get_nsecs();
789 	sleep_nsecs(NSEC_PER_MSEC);
790 	T1 = get_nsecs();
791 
792 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
793 }
794 
795 static int
796 replay_wakeup_event(struct perf_sched *sched,
797 		    struct perf_evsel *evsel, struct perf_sample *sample,
798 		    struct machine *machine __maybe_unused)
799 {
800 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
801 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
802 	struct task_desc *waker, *wakee;
803 
804 	if (verbose > 0) {
805 		printf("sched_wakeup event %p\n", evsel);
806 
807 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
808 	}
809 
810 	waker = register_pid(sched, sample->tid, "<unknown>");
811 	wakee = register_pid(sched, pid, comm);
812 
813 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
814 	return 0;
815 }
816 
817 static int replay_switch_event(struct perf_sched *sched,
818 			       struct perf_evsel *evsel,
819 			       struct perf_sample *sample,
820 			       struct machine *machine __maybe_unused)
821 {
822 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
823 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
824 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
825 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
826 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
827 	struct task_desc *prev, __maybe_unused *next;
828 	u64 timestamp0, timestamp = sample->time;
829 	int cpu = sample->cpu;
830 	s64 delta;
831 
832 	if (verbose > 0)
833 		printf("sched_switch event %p\n", evsel);
834 
835 	if (cpu >= MAX_CPUS || cpu < 0)
836 		return 0;
837 
838 	timestamp0 = sched->cpu_last_switched[cpu];
839 	if (timestamp0)
840 		delta = timestamp - timestamp0;
841 	else
842 		delta = 0;
843 
844 	if (delta < 0) {
845 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
846 		return -1;
847 	}
848 
849 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
850 		 prev_comm, prev_pid, next_comm, next_pid, delta);
851 
852 	prev = register_pid(sched, prev_pid, prev_comm);
853 	next = register_pid(sched, next_pid, next_comm);
854 
855 	sched->cpu_last_switched[cpu] = timestamp;
856 
857 	add_sched_event_run(sched, prev, timestamp, delta);
858 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
859 
860 	return 0;
861 }
862 
863 static int replay_fork_event(struct perf_sched *sched,
864 			     union perf_event *event,
865 			     struct machine *machine)
866 {
867 	struct thread *child, *parent;
868 
869 	child = machine__findnew_thread(machine, event->fork.pid,
870 					event->fork.tid);
871 	parent = machine__findnew_thread(machine, event->fork.ppid,
872 					 event->fork.ptid);
873 
874 	if (child == NULL || parent == NULL) {
875 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
876 				 child, parent);
877 		goto out_put;
878 	}
879 
880 	if (verbose > 0) {
881 		printf("fork event\n");
882 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
883 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
884 	}
885 
886 	register_pid(sched, parent->tid, thread__comm_str(parent));
887 	register_pid(sched, child->tid, thread__comm_str(child));
888 out_put:
889 	thread__put(child);
890 	thread__put(parent);
891 	return 0;
892 }
893 
894 struct sort_dimension {
895 	const char		*name;
896 	sort_fn_t		cmp;
897 	struct list_head	list;
898 };
899 
900 static int
901 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
902 {
903 	struct sort_dimension *sort;
904 	int ret = 0;
905 
906 	BUG_ON(list_empty(list));
907 
908 	list_for_each_entry(sort, list, list) {
909 		ret = sort->cmp(l, r);
910 		if (ret)
911 			return ret;
912 	}
913 
914 	return ret;
915 }
916 
917 static struct work_atoms *
918 thread_atoms_search(struct rb_root *root, struct thread *thread,
919 			 struct list_head *sort_list)
920 {
921 	struct rb_node *node = root->rb_node;
922 	struct work_atoms key = { .thread = thread };
923 
924 	while (node) {
925 		struct work_atoms *atoms;
926 		int cmp;
927 
928 		atoms = container_of(node, struct work_atoms, node);
929 
930 		cmp = thread_lat_cmp(sort_list, &key, atoms);
931 		if (cmp > 0)
932 			node = node->rb_left;
933 		else if (cmp < 0)
934 			node = node->rb_right;
935 		else {
936 			BUG_ON(thread != atoms->thread);
937 			return atoms;
938 		}
939 	}
940 	return NULL;
941 }
942 
943 static void
944 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
945 			 struct list_head *sort_list)
946 {
947 	struct rb_node **new = &(root->rb_node), *parent = NULL;
948 
949 	while (*new) {
950 		struct work_atoms *this;
951 		int cmp;
952 
953 		this = container_of(*new, struct work_atoms, node);
954 		parent = *new;
955 
956 		cmp = thread_lat_cmp(sort_list, data, this);
957 
958 		if (cmp > 0)
959 			new = &((*new)->rb_left);
960 		else
961 			new = &((*new)->rb_right);
962 	}
963 
964 	rb_link_node(&data->node, parent, new);
965 	rb_insert_color(&data->node, root);
966 }
967 
968 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
969 {
970 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
971 	if (!atoms) {
972 		pr_err("No memory at %s\n", __func__);
973 		return -1;
974 	}
975 
976 	atoms->thread = thread__get(thread);
977 	INIT_LIST_HEAD(&atoms->work_list);
978 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
979 	return 0;
980 }
981 
982 static char sched_out_state(u64 prev_state)
983 {
984 	const char *str = TASK_STATE_TO_CHAR_STR;
985 
986 	return str[prev_state];
987 }
988 
989 static int
990 add_sched_out_event(struct work_atoms *atoms,
991 		    char run_state,
992 		    u64 timestamp)
993 {
994 	struct work_atom *atom = zalloc(sizeof(*atom));
995 	if (!atom) {
996 		pr_err("Non memory at %s", __func__);
997 		return -1;
998 	}
999 
1000 	atom->sched_out_time = timestamp;
1001 
1002 	if (run_state == 'R') {
1003 		atom->state = THREAD_WAIT_CPU;
1004 		atom->wake_up_time = atom->sched_out_time;
1005 	}
1006 
1007 	list_add_tail(&atom->list, &atoms->work_list);
1008 	return 0;
1009 }
1010 
1011 static void
1012 add_runtime_event(struct work_atoms *atoms, u64 delta,
1013 		  u64 timestamp __maybe_unused)
1014 {
1015 	struct work_atom *atom;
1016 
1017 	BUG_ON(list_empty(&atoms->work_list));
1018 
1019 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1020 
1021 	atom->runtime += delta;
1022 	atoms->total_runtime += delta;
1023 }
1024 
1025 static void
1026 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1027 {
1028 	struct work_atom *atom;
1029 	u64 delta;
1030 
1031 	if (list_empty(&atoms->work_list))
1032 		return;
1033 
1034 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1035 
1036 	if (atom->state != THREAD_WAIT_CPU)
1037 		return;
1038 
1039 	if (timestamp < atom->wake_up_time) {
1040 		atom->state = THREAD_IGNORE;
1041 		return;
1042 	}
1043 
1044 	atom->state = THREAD_SCHED_IN;
1045 	atom->sched_in_time = timestamp;
1046 
1047 	delta = atom->sched_in_time - atom->wake_up_time;
1048 	atoms->total_lat += delta;
1049 	if (delta > atoms->max_lat) {
1050 		atoms->max_lat = delta;
1051 		atoms->max_lat_at = timestamp;
1052 	}
1053 	atoms->nb_atoms++;
1054 }
1055 
1056 static int latency_switch_event(struct perf_sched *sched,
1057 				struct perf_evsel *evsel,
1058 				struct perf_sample *sample,
1059 				struct machine *machine)
1060 {
1061 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1062 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1063 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1064 	struct work_atoms *out_events, *in_events;
1065 	struct thread *sched_out, *sched_in;
1066 	u64 timestamp0, timestamp = sample->time;
1067 	int cpu = sample->cpu, err = -1;
1068 	s64 delta;
1069 
1070 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1071 
1072 	timestamp0 = sched->cpu_last_switched[cpu];
1073 	sched->cpu_last_switched[cpu] = timestamp;
1074 	if (timestamp0)
1075 		delta = timestamp - timestamp0;
1076 	else
1077 		delta = 0;
1078 
1079 	if (delta < 0) {
1080 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1081 		return -1;
1082 	}
1083 
1084 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1085 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1086 	if (sched_out == NULL || sched_in == NULL)
1087 		goto out_put;
1088 
1089 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1090 	if (!out_events) {
1091 		if (thread_atoms_insert(sched, sched_out))
1092 			goto out_put;
1093 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1094 		if (!out_events) {
1095 			pr_err("out-event: Internal tree error");
1096 			goto out_put;
1097 		}
1098 	}
1099 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1100 		return -1;
1101 
1102 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1103 	if (!in_events) {
1104 		if (thread_atoms_insert(sched, sched_in))
1105 			goto out_put;
1106 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1107 		if (!in_events) {
1108 			pr_err("in-event: Internal tree error");
1109 			goto out_put;
1110 		}
1111 		/*
1112 		 * Take came in we have not heard about yet,
1113 		 * add in an initial atom in runnable state:
1114 		 */
1115 		if (add_sched_out_event(in_events, 'R', timestamp))
1116 			goto out_put;
1117 	}
1118 	add_sched_in_event(in_events, timestamp);
1119 	err = 0;
1120 out_put:
1121 	thread__put(sched_out);
1122 	thread__put(sched_in);
1123 	return err;
1124 }
1125 
1126 static int latency_runtime_event(struct perf_sched *sched,
1127 				 struct perf_evsel *evsel,
1128 				 struct perf_sample *sample,
1129 				 struct machine *machine)
1130 {
1131 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1132 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1133 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1134 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1135 	u64 timestamp = sample->time;
1136 	int cpu = sample->cpu, err = -1;
1137 
1138 	if (thread == NULL)
1139 		return -1;
1140 
1141 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1142 	if (!atoms) {
1143 		if (thread_atoms_insert(sched, thread))
1144 			goto out_put;
1145 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1146 		if (!atoms) {
1147 			pr_err("in-event: Internal tree error");
1148 			goto out_put;
1149 		}
1150 		if (add_sched_out_event(atoms, 'R', timestamp))
1151 			goto out_put;
1152 	}
1153 
1154 	add_runtime_event(atoms, runtime, timestamp);
1155 	err = 0;
1156 out_put:
1157 	thread__put(thread);
1158 	return err;
1159 }
1160 
1161 static int latency_wakeup_event(struct perf_sched *sched,
1162 				struct perf_evsel *evsel,
1163 				struct perf_sample *sample,
1164 				struct machine *machine)
1165 {
1166 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1167 	struct work_atoms *atoms;
1168 	struct work_atom *atom;
1169 	struct thread *wakee;
1170 	u64 timestamp = sample->time;
1171 	int err = -1;
1172 
1173 	wakee = machine__findnew_thread(machine, -1, pid);
1174 	if (wakee == NULL)
1175 		return -1;
1176 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1177 	if (!atoms) {
1178 		if (thread_atoms_insert(sched, wakee))
1179 			goto out_put;
1180 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1181 		if (!atoms) {
1182 			pr_err("wakeup-event: Internal tree error");
1183 			goto out_put;
1184 		}
1185 		if (add_sched_out_event(atoms, 'S', timestamp))
1186 			goto out_put;
1187 	}
1188 
1189 	BUG_ON(list_empty(&atoms->work_list));
1190 
1191 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1192 
1193 	/*
1194 	 * As we do not guarantee the wakeup event happens when
1195 	 * task is out of run queue, also may happen when task is
1196 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1197 	 * then we should not set the ->wake_up_time when wake up a
1198 	 * task which is on run queue.
1199 	 *
1200 	 * You WILL be missing events if you've recorded only
1201 	 * one CPU, or are only looking at only one, so don't
1202 	 * skip in this case.
1203 	 */
1204 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1205 		goto out_ok;
1206 
1207 	sched->nr_timestamps++;
1208 	if (atom->sched_out_time > timestamp) {
1209 		sched->nr_unordered_timestamps++;
1210 		goto out_ok;
1211 	}
1212 
1213 	atom->state = THREAD_WAIT_CPU;
1214 	atom->wake_up_time = timestamp;
1215 out_ok:
1216 	err = 0;
1217 out_put:
1218 	thread__put(wakee);
1219 	return err;
1220 }
1221 
1222 static int latency_migrate_task_event(struct perf_sched *sched,
1223 				      struct perf_evsel *evsel,
1224 				      struct perf_sample *sample,
1225 				      struct machine *machine)
1226 {
1227 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1228 	u64 timestamp = sample->time;
1229 	struct work_atoms *atoms;
1230 	struct work_atom *atom;
1231 	struct thread *migrant;
1232 	int err = -1;
1233 
1234 	/*
1235 	 * Only need to worry about migration when profiling one CPU.
1236 	 */
1237 	if (sched->profile_cpu == -1)
1238 		return 0;
1239 
1240 	migrant = machine__findnew_thread(machine, -1, pid);
1241 	if (migrant == NULL)
1242 		return -1;
1243 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1244 	if (!atoms) {
1245 		if (thread_atoms_insert(sched, migrant))
1246 			goto out_put;
1247 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1248 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1249 		if (!atoms) {
1250 			pr_err("migration-event: Internal tree error");
1251 			goto out_put;
1252 		}
1253 		if (add_sched_out_event(atoms, 'R', timestamp))
1254 			goto out_put;
1255 	}
1256 
1257 	BUG_ON(list_empty(&atoms->work_list));
1258 
1259 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1260 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1261 
1262 	sched->nr_timestamps++;
1263 
1264 	if (atom->sched_out_time > timestamp)
1265 		sched->nr_unordered_timestamps++;
1266 	err = 0;
1267 out_put:
1268 	thread__put(migrant);
1269 	return err;
1270 }
1271 
1272 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1273 {
1274 	int i;
1275 	int ret;
1276 	u64 avg;
1277 	char max_lat_at[32];
1278 
1279 	if (!work_list->nb_atoms)
1280 		return;
1281 	/*
1282 	 * Ignore idle threads:
1283 	 */
1284 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1285 		return;
1286 
1287 	sched->all_runtime += work_list->total_runtime;
1288 	sched->all_count   += work_list->nb_atoms;
1289 
1290 	if (work_list->num_merged > 1)
1291 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1292 	else
1293 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1294 
1295 	for (i = 0; i < 24 - ret; i++)
1296 		printf(" ");
1297 
1298 	avg = work_list->total_lat / work_list->nb_atoms;
1299 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1300 
1301 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1302 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1303 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1304 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1305 		 max_lat_at);
1306 }
1307 
1308 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1309 {
1310 	if (l->thread == r->thread)
1311 		return 0;
1312 	if (l->thread->tid < r->thread->tid)
1313 		return -1;
1314 	if (l->thread->tid > r->thread->tid)
1315 		return 1;
1316 	return (int)(l->thread - r->thread);
1317 }
1318 
1319 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1320 {
1321 	u64 avgl, avgr;
1322 
1323 	if (!l->nb_atoms)
1324 		return -1;
1325 
1326 	if (!r->nb_atoms)
1327 		return 1;
1328 
1329 	avgl = l->total_lat / l->nb_atoms;
1330 	avgr = r->total_lat / r->nb_atoms;
1331 
1332 	if (avgl < avgr)
1333 		return -1;
1334 	if (avgl > avgr)
1335 		return 1;
1336 
1337 	return 0;
1338 }
1339 
1340 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1341 {
1342 	if (l->max_lat < r->max_lat)
1343 		return -1;
1344 	if (l->max_lat > r->max_lat)
1345 		return 1;
1346 
1347 	return 0;
1348 }
1349 
1350 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1351 {
1352 	if (l->nb_atoms < r->nb_atoms)
1353 		return -1;
1354 	if (l->nb_atoms > r->nb_atoms)
1355 		return 1;
1356 
1357 	return 0;
1358 }
1359 
1360 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1361 {
1362 	if (l->total_runtime < r->total_runtime)
1363 		return -1;
1364 	if (l->total_runtime > r->total_runtime)
1365 		return 1;
1366 
1367 	return 0;
1368 }
1369 
1370 static int sort_dimension__add(const char *tok, struct list_head *list)
1371 {
1372 	size_t i;
1373 	static struct sort_dimension avg_sort_dimension = {
1374 		.name = "avg",
1375 		.cmp  = avg_cmp,
1376 	};
1377 	static struct sort_dimension max_sort_dimension = {
1378 		.name = "max",
1379 		.cmp  = max_cmp,
1380 	};
1381 	static struct sort_dimension pid_sort_dimension = {
1382 		.name = "pid",
1383 		.cmp  = pid_cmp,
1384 	};
1385 	static struct sort_dimension runtime_sort_dimension = {
1386 		.name = "runtime",
1387 		.cmp  = runtime_cmp,
1388 	};
1389 	static struct sort_dimension switch_sort_dimension = {
1390 		.name = "switch",
1391 		.cmp  = switch_cmp,
1392 	};
1393 	struct sort_dimension *available_sorts[] = {
1394 		&pid_sort_dimension,
1395 		&avg_sort_dimension,
1396 		&max_sort_dimension,
1397 		&switch_sort_dimension,
1398 		&runtime_sort_dimension,
1399 	};
1400 
1401 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1402 		if (!strcmp(available_sorts[i]->name, tok)) {
1403 			list_add_tail(&available_sorts[i]->list, list);
1404 
1405 			return 0;
1406 		}
1407 	}
1408 
1409 	return -1;
1410 }
1411 
1412 static void perf_sched__sort_lat(struct perf_sched *sched)
1413 {
1414 	struct rb_node *node;
1415 	struct rb_root *root = &sched->atom_root;
1416 again:
1417 	for (;;) {
1418 		struct work_atoms *data;
1419 		node = rb_first(root);
1420 		if (!node)
1421 			break;
1422 
1423 		rb_erase(node, root);
1424 		data = rb_entry(node, struct work_atoms, node);
1425 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1426 	}
1427 	if (root == &sched->atom_root) {
1428 		root = &sched->merged_atom_root;
1429 		goto again;
1430 	}
1431 }
1432 
1433 static int process_sched_wakeup_event(struct perf_tool *tool,
1434 				      struct perf_evsel *evsel,
1435 				      struct perf_sample *sample,
1436 				      struct machine *machine)
1437 {
1438 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1439 
1440 	if (sched->tp_handler->wakeup_event)
1441 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1442 
1443 	return 0;
1444 }
1445 
1446 union map_priv {
1447 	void	*ptr;
1448 	bool	 color;
1449 };
1450 
1451 static bool thread__has_color(struct thread *thread)
1452 {
1453 	union map_priv priv = {
1454 		.ptr = thread__priv(thread),
1455 	};
1456 
1457 	return priv.color;
1458 }
1459 
1460 static struct thread*
1461 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1462 {
1463 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1464 	union map_priv priv = {
1465 		.color = false,
1466 	};
1467 
1468 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1469 		return thread;
1470 
1471 	if (thread_map__has(sched->map.color_pids, tid))
1472 		priv.color = true;
1473 
1474 	thread__set_priv(thread, priv.ptr);
1475 	return thread;
1476 }
1477 
1478 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1479 			    struct perf_sample *sample, struct machine *machine)
1480 {
1481 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1482 	struct thread *sched_in;
1483 	int new_shortname;
1484 	u64 timestamp0, timestamp = sample->time;
1485 	s64 delta;
1486 	int i, this_cpu = sample->cpu;
1487 	int cpus_nr;
1488 	bool new_cpu = false;
1489 	const char *color = PERF_COLOR_NORMAL;
1490 	char stimestamp[32];
1491 
1492 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1493 
1494 	if (this_cpu > sched->max_cpu)
1495 		sched->max_cpu = this_cpu;
1496 
1497 	if (sched->map.comp) {
1498 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1499 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1500 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1501 			new_cpu = true;
1502 		}
1503 	} else
1504 		cpus_nr = sched->max_cpu;
1505 
1506 	timestamp0 = sched->cpu_last_switched[this_cpu];
1507 	sched->cpu_last_switched[this_cpu] = timestamp;
1508 	if (timestamp0)
1509 		delta = timestamp - timestamp0;
1510 	else
1511 		delta = 0;
1512 
1513 	if (delta < 0) {
1514 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1515 		return -1;
1516 	}
1517 
1518 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1519 	if (sched_in == NULL)
1520 		return -1;
1521 
1522 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1523 
1524 	printf("  ");
1525 
1526 	new_shortname = 0;
1527 	if (!sched_in->shortname[0]) {
1528 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1529 			/*
1530 			 * Don't allocate a letter-number for swapper:0
1531 			 * as a shortname. Instead, we use '.' for it.
1532 			 */
1533 			sched_in->shortname[0] = '.';
1534 			sched_in->shortname[1] = ' ';
1535 		} else {
1536 			sched_in->shortname[0] = sched->next_shortname1;
1537 			sched_in->shortname[1] = sched->next_shortname2;
1538 
1539 			if (sched->next_shortname1 < 'Z') {
1540 				sched->next_shortname1++;
1541 			} else {
1542 				sched->next_shortname1 = 'A';
1543 				if (sched->next_shortname2 < '9')
1544 					sched->next_shortname2++;
1545 				else
1546 					sched->next_shortname2 = '0';
1547 			}
1548 		}
1549 		new_shortname = 1;
1550 	}
1551 
1552 	for (i = 0; i < cpus_nr; i++) {
1553 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1554 		struct thread *curr_thread = sched->curr_thread[cpu];
1555 		const char *pid_color = color;
1556 		const char *cpu_color = color;
1557 
1558 		if (curr_thread && thread__has_color(curr_thread))
1559 			pid_color = COLOR_PIDS;
1560 
1561 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1562 			continue;
1563 
1564 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1565 			cpu_color = COLOR_CPUS;
1566 
1567 		if (cpu != this_cpu)
1568 			color_fprintf(stdout, color, " ");
1569 		else
1570 			color_fprintf(stdout, cpu_color, "*");
1571 
1572 		if (sched->curr_thread[cpu])
1573 			color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1574 		else
1575 			color_fprintf(stdout, color, "   ");
1576 	}
1577 
1578 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1579 		goto out;
1580 
1581 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1582 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1583 	if (new_shortname || (verbose > 0 && sched_in->tid)) {
1584 		const char *pid_color = color;
1585 
1586 		if (thread__has_color(sched_in))
1587 			pid_color = COLOR_PIDS;
1588 
1589 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1590 		       sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1591 	}
1592 
1593 	if (sched->map.comp && new_cpu)
1594 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1595 
1596 out:
1597 	color_fprintf(stdout, color, "\n");
1598 
1599 	thread__put(sched_in);
1600 
1601 	return 0;
1602 }
1603 
1604 static int process_sched_switch_event(struct perf_tool *tool,
1605 				      struct perf_evsel *evsel,
1606 				      struct perf_sample *sample,
1607 				      struct machine *machine)
1608 {
1609 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1610 	int this_cpu = sample->cpu, err = 0;
1611 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1612 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1613 
1614 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1615 		/*
1616 		 * Are we trying to switch away a PID that is
1617 		 * not current?
1618 		 */
1619 		if (sched->curr_pid[this_cpu] != prev_pid)
1620 			sched->nr_context_switch_bugs++;
1621 	}
1622 
1623 	if (sched->tp_handler->switch_event)
1624 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1625 
1626 	sched->curr_pid[this_cpu] = next_pid;
1627 	return err;
1628 }
1629 
1630 static int process_sched_runtime_event(struct perf_tool *tool,
1631 				       struct perf_evsel *evsel,
1632 				       struct perf_sample *sample,
1633 				       struct machine *machine)
1634 {
1635 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1636 
1637 	if (sched->tp_handler->runtime_event)
1638 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1639 
1640 	return 0;
1641 }
1642 
1643 static int perf_sched__process_fork_event(struct perf_tool *tool,
1644 					  union perf_event *event,
1645 					  struct perf_sample *sample,
1646 					  struct machine *machine)
1647 {
1648 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1649 
1650 	/* run the fork event through the perf machineruy */
1651 	perf_event__process_fork(tool, event, sample, machine);
1652 
1653 	/* and then run additional processing needed for this command */
1654 	if (sched->tp_handler->fork_event)
1655 		return sched->tp_handler->fork_event(sched, event, machine);
1656 
1657 	return 0;
1658 }
1659 
1660 static int process_sched_migrate_task_event(struct perf_tool *tool,
1661 					    struct perf_evsel *evsel,
1662 					    struct perf_sample *sample,
1663 					    struct machine *machine)
1664 {
1665 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1666 
1667 	if (sched->tp_handler->migrate_task_event)
1668 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1669 
1670 	return 0;
1671 }
1672 
1673 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1674 				  struct perf_evsel *evsel,
1675 				  struct perf_sample *sample,
1676 				  struct machine *machine);
1677 
1678 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1679 						 union perf_event *event __maybe_unused,
1680 						 struct perf_sample *sample,
1681 						 struct perf_evsel *evsel,
1682 						 struct machine *machine)
1683 {
1684 	int err = 0;
1685 
1686 	if (evsel->handler != NULL) {
1687 		tracepoint_handler f = evsel->handler;
1688 		err = f(tool, evsel, sample, machine);
1689 	}
1690 
1691 	return err;
1692 }
1693 
1694 static int perf_sched__read_events(struct perf_sched *sched)
1695 {
1696 	const struct perf_evsel_str_handler handlers[] = {
1697 		{ "sched:sched_switch",	      process_sched_switch_event, },
1698 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1699 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1700 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1701 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1702 	};
1703 	struct perf_session *session;
1704 	struct perf_data_file file = {
1705 		.path = input_name,
1706 		.mode = PERF_DATA_MODE_READ,
1707 		.force = sched->force,
1708 	};
1709 	int rc = -1;
1710 
1711 	session = perf_session__new(&file, false, &sched->tool);
1712 	if (session == NULL) {
1713 		pr_debug("No Memory for session\n");
1714 		return -1;
1715 	}
1716 
1717 	symbol__init(&session->header.env);
1718 
1719 	if (perf_session__set_tracepoints_handlers(session, handlers))
1720 		goto out_delete;
1721 
1722 	if (perf_session__has_traces(session, "record -R")) {
1723 		int err = perf_session__process_events(session);
1724 		if (err) {
1725 			pr_err("Failed to process events, error %d", err);
1726 			goto out_delete;
1727 		}
1728 
1729 		sched->nr_events      = session->evlist->stats.nr_events[0];
1730 		sched->nr_lost_events = session->evlist->stats.total_lost;
1731 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1732 	}
1733 
1734 	rc = 0;
1735 out_delete:
1736 	perf_session__delete(session);
1737 	return rc;
1738 }
1739 
1740 /*
1741  * scheduling times are printed as msec.usec
1742  */
1743 static inline void print_sched_time(unsigned long long nsecs, int width)
1744 {
1745 	unsigned long msecs;
1746 	unsigned long usecs;
1747 
1748 	msecs  = nsecs / NSEC_PER_MSEC;
1749 	nsecs -= msecs * NSEC_PER_MSEC;
1750 	usecs  = nsecs / NSEC_PER_USEC;
1751 	printf("%*lu.%03lu ", width, msecs, usecs);
1752 }
1753 
1754 /*
1755  * returns runtime data for event, allocating memory for it the
1756  * first time it is used.
1757  */
1758 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1759 {
1760 	struct evsel_runtime *r = evsel->priv;
1761 
1762 	if (r == NULL) {
1763 		r = zalloc(sizeof(struct evsel_runtime));
1764 		evsel->priv = r;
1765 	}
1766 
1767 	return r;
1768 }
1769 
1770 /*
1771  * save last time event was seen per cpu
1772  */
1773 static void perf_evsel__save_time(struct perf_evsel *evsel,
1774 				  u64 timestamp, u32 cpu)
1775 {
1776 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1777 
1778 	if (r == NULL)
1779 		return;
1780 
1781 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1782 		int i, n = __roundup_pow_of_two(cpu+1);
1783 		void *p = r->last_time;
1784 
1785 		p = realloc(r->last_time, n * sizeof(u64));
1786 		if (!p)
1787 			return;
1788 
1789 		r->last_time = p;
1790 		for (i = r->ncpu; i < n; ++i)
1791 			r->last_time[i] = (u64) 0;
1792 
1793 		r->ncpu = n;
1794 	}
1795 
1796 	r->last_time[cpu] = timestamp;
1797 }
1798 
1799 /* returns last time this event was seen on the given cpu */
1800 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1801 {
1802 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1803 
1804 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1805 		return 0;
1806 
1807 	return r->last_time[cpu];
1808 }
1809 
1810 static int comm_width = 30;
1811 
1812 static char *timehist_get_commstr(struct thread *thread)
1813 {
1814 	static char str[32];
1815 	const char *comm = thread__comm_str(thread);
1816 	pid_t tid = thread->tid;
1817 	pid_t pid = thread->pid_;
1818 	int n;
1819 
1820 	if (pid == 0)
1821 		n = scnprintf(str, sizeof(str), "%s", comm);
1822 
1823 	else if (tid != pid)
1824 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1825 
1826 	else
1827 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1828 
1829 	if (n > comm_width)
1830 		comm_width = n;
1831 
1832 	return str;
1833 }
1834 
1835 static void timehist_header(struct perf_sched *sched)
1836 {
1837 	u32 ncpus = sched->max_cpu + 1;
1838 	u32 i, j;
1839 
1840 	printf("%15s %6s ", "time", "cpu");
1841 
1842 	if (sched->show_cpu_visual) {
1843 		printf(" ");
1844 		for (i = 0, j = 0; i < ncpus; ++i) {
1845 			printf("%x", j++);
1846 			if (j > 15)
1847 				j = 0;
1848 		}
1849 		printf(" ");
1850 	}
1851 
1852 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1853 		"task name", "wait time", "sch delay", "run time");
1854 
1855 	if (sched->show_state)
1856 		printf("  %s", "state");
1857 
1858 	printf("\n");
1859 
1860 	/*
1861 	 * units row
1862 	 */
1863 	printf("%15s %-6s ", "", "");
1864 
1865 	if (sched->show_cpu_visual)
1866 		printf(" %*s ", ncpus, "");
1867 
1868 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1869 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1870 
1871 	if (sched->show_state)
1872 		printf("  %5s", "");
1873 
1874 	printf("\n");
1875 
1876 	/*
1877 	 * separator
1878 	 */
1879 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1880 
1881 	if (sched->show_cpu_visual)
1882 		printf(" %.*s ", ncpus, graph_dotted_line);
1883 
1884 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1885 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1886 		graph_dotted_line);
1887 
1888 	if (sched->show_state)
1889 		printf("  %.5s", graph_dotted_line);
1890 
1891 	printf("\n");
1892 }
1893 
1894 static char task_state_char(struct thread *thread, int state)
1895 {
1896 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1897 	unsigned bit = state ? ffs(state) : 0;
1898 
1899 	/* 'I' for idle */
1900 	if (thread->tid == 0)
1901 		return 'I';
1902 
1903 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1904 }
1905 
1906 static void timehist_print_sample(struct perf_sched *sched,
1907 				  struct perf_evsel *evsel,
1908 				  struct perf_sample *sample,
1909 				  struct addr_location *al,
1910 				  struct thread *thread,
1911 				  u64 t, int state)
1912 {
1913 	struct thread_runtime *tr = thread__priv(thread);
1914 	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
1915 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1916 	u32 max_cpus = sched->max_cpu + 1;
1917 	char tstr[64];
1918 	char nstr[30];
1919 	u64 wait_time;
1920 
1921 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1922 	printf("%15s [%04d] ", tstr, sample->cpu);
1923 
1924 	if (sched->show_cpu_visual) {
1925 		u32 i;
1926 		char c;
1927 
1928 		printf(" ");
1929 		for (i = 0; i < max_cpus; ++i) {
1930 			/* flag idle times with 'i'; others are sched events */
1931 			if (i == sample->cpu)
1932 				c = (thread->tid == 0) ? 'i' : 's';
1933 			else
1934 				c = ' ';
1935 			printf("%c", c);
1936 		}
1937 		printf(" ");
1938 	}
1939 
1940 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1941 
1942 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
1943 	print_sched_time(wait_time, 6);
1944 
1945 	print_sched_time(tr->dt_delay, 6);
1946 	print_sched_time(tr->dt_run, 6);
1947 
1948 	if (sched->show_state)
1949 		printf(" %5c ", task_state_char(thread, state));
1950 
1951 	if (sched->show_next) {
1952 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
1953 		printf(" %-*s", comm_width, nstr);
1954 	}
1955 
1956 	if (sched->show_wakeups && !sched->show_next)
1957 		printf("  %-*s", comm_width, "");
1958 
1959 	if (thread->tid == 0)
1960 		goto out;
1961 
1962 	if (sched->show_callchain)
1963 		printf("  ");
1964 
1965 	sample__fprintf_sym(sample, al, 0,
1966 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1967 			    EVSEL__PRINT_CALLCHAIN_ARROW |
1968 			    EVSEL__PRINT_SKIP_IGNORED,
1969 			    &callchain_cursor, stdout);
1970 
1971 out:
1972 	printf("\n");
1973 }
1974 
1975 /*
1976  * Explanation of delta-time stats:
1977  *
1978  *            t = time of current schedule out event
1979  *        tprev = time of previous sched out event
1980  *                also time of schedule-in event for current task
1981  *    last_time = time of last sched change event for current task
1982  *                (i.e, time process was last scheduled out)
1983  * ready_to_run = time of wakeup for current task
1984  *
1985  * -----|------------|------------|------------|------
1986  *    last         ready        tprev          t
1987  *    time         to run
1988  *
1989  *      |-------- dt_wait --------|
1990  *                   |- dt_delay -|-- dt_run --|
1991  *
1992  *   dt_run = run time of current task
1993  *  dt_wait = time between last schedule out event for task and tprev
1994  *            represents time spent off the cpu
1995  * dt_delay = time between wakeup and schedule-in of task
1996  */
1997 
1998 static void timehist_update_runtime_stats(struct thread_runtime *r,
1999 					 u64 t, u64 tprev)
2000 {
2001 	r->dt_delay   = 0;
2002 	r->dt_sleep   = 0;
2003 	r->dt_iowait  = 0;
2004 	r->dt_preempt = 0;
2005 	r->dt_run     = 0;
2006 
2007 	if (tprev) {
2008 		r->dt_run = t - tprev;
2009 		if (r->ready_to_run) {
2010 			if (r->ready_to_run > tprev)
2011 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2012 			else
2013 				r->dt_delay = tprev - r->ready_to_run;
2014 		}
2015 
2016 		if (r->last_time > tprev)
2017 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2018 		else if (r->last_time) {
2019 			u64 dt_wait = tprev - r->last_time;
2020 
2021 			if (r->last_state == TASK_RUNNING)
2022 				r->dt_preempt = dt_wait;
2023 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2024 				r->dt_iowait = dt_wait;
2025 			else
2026 				r->dt_sleep = dt_wait;
2027 		}
2028 	}
2029 
2030 	update_stats(&r->run_stats, r->dt_run);
2031 
2032 	r->total_run_time     += r->dt_run;
2033 	r->total_delay_time   += r->dt_delay;
2034 	r->total_sleep_time   += r->dt_sleep;
2035 	r->total_iowait_time  += r->dt_iowait;
2036 	r->total_preempt_time += r->dt_preempt;
2037 }
2038 
2039 static bool is_idle_sample(struct perf_sample *sample,
2040 			   struct perf_evsel *evsel)
2041 {
2042 	/* pid 0 == swapper == idle task */
2043 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2044 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2045 
2046 	return sample->pid == 0;
2047 }
2048 
2049 static void save_task_callchain(struct perf_sched *sched,
2050 				struct perf_sample *sample,
2051 				struct perf_evsel *evsel,
2052 				struct machine *machine)
2053 {
2054 	struct callchain_cursor *cursor = &callchain_cursor;
2055 	struct thread *thread;
2056 
2057 	/* want main thread for process - has maps */
2058 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2059 	if (thread == NULL) {
2060 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2061 		return;
2062 	}
2063 
2064 	if (!symbol_conf.use_callchain || sample->callchain == NULL)
2065 		return;
2066 
2067 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2068 				      NULL, NULL, sched->max_stack + 2) != 0) {
2069 		if (verbose > 0)
2070 			pr_err("Failed to resolve callchain. Skipping\n");
2071 
2072 		return;
2073 	}
2074 
2075 	callchain_cursor_commit(cursor);
2076 
2077 	while (true) {
2078 		struct callchain_cursor_node *node;
2079 		struct symbol *sym;
2080 
2081 		node = callchain_cursor_current(cursor);
2082 		if (node == NULL)
2083 			break;
2084 
2085 		sym = node->sym;
2086 		if (sym) {
2087 			if (!strcmp(sym->name, "schedule") ||
2088 			    !strcmp(sym->name, "__schedule") ||
2089 			    !strcmp(sym->name, "preempt_schedule"))
2090 				sym->ignore = 1;
2091 		}
2092 
2093 		callchain_cursor_advance(cursor);
2094 	}
2095 }
2096 
2097 static int init_idle_thread(struct thread *thread)
2098 {
2099 	struct idle_thread_runtime *itr;
2100 
2101 	thread__set_comm(thread, idle_comm, 0);
2102 
2103 	itr = zalloc(sizeof(*itr));
2104 	if (itr == NULL)
2105 		return -ENOMEM;
2106 
2107 	init_stats(&itr->tr.run_stats);
2108 	callchain_init(&itr->callchain);
2109 	callchain_cursor_reset(&itr->cursor);
2110 	thread__set_priv(thread, itr);
2111 
2112 	return 0;
2113 }
2114 
2115 /*
2116  * Track idle stats per cpu by maintaining a local thread
2117  * struct for the idle task on each cpu.
2118  */
2119 static int init_idle_threads(int ncpu)
2120 {
2121 	int i, ret;
2122 
2123 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2124 	if (!idle_threads)
2125 		return -ENOMEM;
2126 
2127 	idle_max_cpu = ncpu;
2128 
2129 	/* allocate the actual thread struct if needed */
2130 	for (i = 0; i < ncpu; ++i) {
2131 		idle_threads[i] = thread__new(0, 0);
2132 		if (idle_threads[i] == NULL)
2133 			return -ENOMEM;
2134 
2135 		ret = init_idle_thread(idle_threads[i]);
2136 		if (ret < 0)
2137 			return ret;
2138 	}
2139 
2140 	return 0;
2141 }
2142 
2143 static void free_idle_threads(void)
2144 {
2145 	int i;
2146 
2147 	if (idle_threads == NULL)
2148 		return;
2149 
2150 	for (i = 0; i < idle_max_cpu; ++i) {
2151 		if ((idle_threads[i]))
2152 			thread__delete(idle_threads[i]);
2153 	}
2154 
2155 	free(idle_threads);
2156 }
2157 
2158 static struct thread *get_idle_thread(int cpu)
2159 {
2160 	/*
2161 	 * expand/allocate array of pointers to local thread
2162 	 * structs if needed
2163 	 */
2164 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2165 		int i, j = __roundup_pow_of_two(cpu+1);
2166 		void *p;
2167 
2168 		p = realloc(idle_threads, j * sizeof(struct thread *));
2169 		if (!p)
2170 			return NULL;
2171 
2172 		idle_threads = (struct thread **) p;
2173 		for (i = idle_max_cpu; i < j; ++i)
2174 			idle_threads[i] = NULL;
2175 
2176 		idle_max_cpu = j;
2177 	}
2178 
2179 	/* allocate a new thread struct if needed */
2180 	if (idle_threads[cpu] == NULL) {
2181 		idle_threads[cpu] = thread__new(0, 0);
2182 
2183 		if (idle_threads[cpu]) {
2184 			if (init_idle_thread(idle_threads[cpu]) < 0)
2185 				return NULL;
2186 		}
2187 	}
2188 
2189 	return idle_threads[cpu];
2190 }
2191 
2192 static void save_idle_callchain(struct idle_thread_runtime *itr,
2193 				struct perf_sample *sample)
2194 {
2195 	if (!symbol_conf.use_callchain || sample->callchain == NULL)
2196 		return;
2197 
2198 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2199 }
2200 
2201 /*
2202  * handle runtime stats saved per thread
2203  */
2204 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2205 {
2206 	struct thread_runtime *r;
2207 
2208 	r = zalloc(sizeof(struct thread_runtime));
2209 	if (!r)
2210 		return NULL;
2211 
2212 	init_stats(&r->run_stats);
2213 	thread__set_priv(thread, r);
2214 
2215 	return r;
2216 }
2217 
2218 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2219 {
2220 	struct thread_runtime *tr;
2221 
2222 	tr = thread__priv(thread);
2223 	if (tr == NULL) {
2224 		tr = thread__init_runtime(thread);
2225 		if (tr == NULL)
2226 			pr_debug("Failed to malloc memory for runtime data.\n");
2227 	}
2228 
2229 	return tr;
2230 }
2231 
2232 static struct thread *timehist_get_thread(struct perf_sched *sched,
2233 					  struct perf_sample *sample,
2234 					  struct machine *machine,
2235 					  struct perf_evsel *evsel)
2236 {
2237 	struct thread *thread;
2238 
2239 	if (is_idle_sample(sample, evsel)) {
2240 		thread = get_idle_thread(sample->cpu);
2241 		if (thread == NULL)
2242 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2243 
2244 	} else {
2245 		/* there were samples with tid 0 but non-zero pid */
2246 		thread = machine__findnew_thread(machine, sample->pid,
2247 						 sample->tid ?: sample->pid);
2248 		if (thread == NULL) {
2249 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2250 				 sample->tid);
2251 		}
2252 
2253 		save_task_callchain(sched, sample, evsel, machine);
2254 		if (sched->idle_hist) {
2255 			struct thread *idle;
2256 			struct idle_thread_runtime *itr;
2257 
2258 			idle = get_idle_thread(sample->cpu);
2259 			if (idle == NULL) {
2260 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2261 				return NULL;
2262 			}
2263 
2264 			itr = thread__priv(idle);
2265 			if (itr == NULL)
2266 				return NULL;
2267 
2268 			itr->last_thread = thread;
2269 
2270 			/* copy task callchain when entering to idle */
2271 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2272 				save_idle_callchain(itr, sample);
2273 		}
2274 	}
2275 
2276 	return thread;
2277 }
2278 
2279 static bool timehist_skip_sample(struct perf_sched *sched,
2280 				 struct thread *thread,
2281 				 struct perf_evsel *evsel,
2282 				 struct perf_sample *sample)
2283 {
2284 	bool rc = false;
2285 
2286 	if (thread__is_filtered(thread)) {
2287 		rc = true;
2288 		sched->skipped_samples++;
2289 	}
2290 
2291 	if (sched->idle_hist) {
2292 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2293 			rc = true;
2294 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2295 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2296 			rc = true;
2297 	}
2298 
2299 	return rc;
2300 }
2301 
2302 static void timehist_print_wakeup_event(struct perf_sched *sched,
2303 					struct perf_evsel *evsel,
2304 					struct perf_sample *sample,
2305 					struct machine *machine,
2306 					struct thread *awakened)
2307 {
2308 	struct thread *thread;
2309 	char tstr[64];
2310 
2311 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2312 	if (thread == NULL)
2313 		return;
2314 
2315 	/* show wakeup unless both awakee and awaker are filtered */
2316 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2317 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2318 		return;
2319 	}
2320 
2321 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2322 	printf("%15s [%04d] ", tstr, sample->cpu);
2323 	if (sched->show_cpu_visual)
2324 		printf(" %*s ", sched->max_cpu + 1, "");
2325 
2326 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2327 
2328 	/* dt spacer */
2329 	printf("  %9s  %9s  %9s ", "", "", "");
2330 
2331 	printf("awakened: %s", timehist_get_commstr(awakened));
2332 
2333 	printf("\n");
2334 }
2335 
2336 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2337 				       union perf_event *event __maybe_unused,
2338 				       struct perf_evsel *evsel,
2339 				       struct perf_sample *sample,
2340 				       struct machine *machine)
2341 {
2342 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2343 	struct thread *thread;
2344 	struct thread_runtime *tr = NULL;
2345 	/* want pid of awakened task not pid in sample */
2346 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2347 
2348 	thread = machine__findnew_thread(machine, 0, pid);
2349 	if (thread == NULL)
2350 		return -1;
2351 
2352 	tr = thread__get_runtime(thread);
2353 	if (tr == NULL)
2354 		return -1;
2355 
2356 	if (tr->ready_to_run == 0)
2357 		tr->ready_to_run = sample->time;
2358 
2359 	/* show wakeups if requested */
2360 	if (sched->show_wakeups &&
2361 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2362 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2363 
2364 	return 0;
2365 }
2366 
2367 static void timehist_print_migration_event(struct perf_sched *sched,
2368 					struct perf_evsel *evsel,
2369 					struct perf_sample *sample,
2370 					struct machine *machine,
2371 					struct thread *migrated)
2372 {
2373 	struct thread *thread;
2374 	char tstr[64];
2375 	u32 max_cpus = sched->max_cpu + 1;
2376 	u32 ocpu, dcpu;
2377 
2378 	if (sched->summary_only)
2379 		return;
2380 
2381 	max_cpus = sched->max_cpu + 1;
2382 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2383 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2384 
2385 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2386 	if (thread == NULL)
2387 		return;
2388 
2389 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2390 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2391 		return;
2392 	}
2393 
2394 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2395 	printf("%15s [%04d] ", tstr, sample->cpu);
2396 
2397 	if (sched->show_cpu_visual) {
2398 		u32 i;
2399 		char c;
2400 
2401 		printf("  ");
2402 		for (i = 0; i < max_cpus; ++i) {
2403 			c = (i == sample->cpu) ? 'm' : ' ';
2404 			printf("%c", c);
2405 		}
2406 		printf("  ");
2407 	}
2408 
2409 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2410 
2411 	/* dt spacer */
2412 	printf("  %9s  %9s  %9s ", "", "", "");
2413 
2414 	printf("migrated: %s", timehist_get_commstr(migrated));
2415 	printf(" cpu %d => %d", ocpu, dcpu);
2416 
2417 	printf("\n");
2418 }
2419 
2420 static int timehist_migrate_task_event(struct perf_tool *tool,
2421 				       union perf_event *event __maybe_unused,
2422 				       struct perf_evsel *evsel,
2423 				       struct perf_sample *sample,
2424 				       struct machine *machine)
2425 {
2426 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2427 	struct thread *thread;
2428 	struct thread_runtime *tr = NULL;
2429 	/* want pid of migrated task not pid in sample */
2430 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2431 
2432 	thread = machine__findnew_thread(machine, 0, pid);
2433 	if (thread == NULL)
2434 		return -1;
2435 
2436 	tr = thread__get_runtime(thread);
2437 	if (tr == NULL)
2438 		return -1;
2439 
2440 	tr->migrations++;
2441 
2442 	/* show migrations if requested */
2443 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2444 
2445 	return 0;
2446 }
2447 
2448 static int timehist_sched_change_event(struct perf_tool *tool,
2449 				       union perf_event *event,
2450 				       struct perf_evsel *evsel,
2451 				       struct perf_sample *sample,
2452 				       struct machine *machine)
2453 {
2454 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2455 	struct perf_time_interval *ptime = &sched->ptime;
2456 	struct addr_location al;
2457 	struct thread *thread;
2458 	struct thread_runtime *tr = NULL;
2459 	u64 tprev, t = sample->time;
2460 	int rc = 0;
2461 	int state = perf_evsel__intval(evsel, sample, "prev_state");
2462 
2463 
2464 	if (machine__resolve(machine, &al, sample) < 0) {
2465 		pr_err("problem processing %d event. skipping it\n",
2466 		       event->header.type);
2467 		rc = -1;
2468 		goto out;
2469 	}
2470 
2471 	thread = timehist_get_thread(sched, sample, machine, evsel);
2472 	if (thread == NULL) {
2473 		rc = -1;
2474 		goto out;
2475 	}
2476 
2477 	if (timehist_skip_sample(sched, thread, evsel, sample))
2478 		goto out;
2479 
2480 	tr = thread__get_runtime(thread);
2481 	if (tr == NULL) {
2482 		rc = -1;
2483 		goto out;
2484 	}
2485 
2486 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2487 
2488 	/*
2489 	 * If start time given:
2490 	 * - sample time is under window user cares about - skip sample
2491 	 * - tprev is under window user cares about  - reset to start of window
2492 	 */
2493 	if (ptime->start && ptime->start > t)
2494 		goto out;
2495 
2496 	if (tprev && ptime->start > tprev)
2497 		tprev = ptime->start;
2498 
2499 	/*
2500 	 * If end time given:
2501 	 * - previous sched event is out of window - we are done
2502 	 * - sample time is beyond window user cares about - reset it
2503 	 *   to close out stats for time window interest
2504 	 */
2505 	if (ptime->end) {
2506 		if (tprev > ptime->end)
2507 			goto out;
2508 
2509 		if (t > ptime->end)
2510 			t = ptime->end;
2511 	}
2512 
2513 	if (!sched->idle_hist || thread->tid == 0) {
2514 		timehist_update_runtime_stats(tr, t, tprev);
2515 
2516 		if (sched->idle_hist) {
2517 			struct idle_thread_runtime *itr = (void *)tr;
2518 			struct thread_runtime *last_tr;
2519 
2520 			BUG_ON(thread->tid != 0);
2521 
2522 			if (itr->last_thread == NULL)
2523 				goto out;
2524 
2525 			/* add current idle time as last thread's runtime */
2526 			last_tr = thread__get_runtime(itr->last_thread);
2527 			if (last_tr == NULL)
2528 				goto out;
2529 
2530 			timehist_update_runtime_stats(last_tr, t, tprev);
2531 			/*
2532 			 * remove delta time of last thread as it's not updated
2533 			 * and otherwise it will show an invalid value next
2534 			 * time.  we only care total run time and run stat.
2535 			 */
2536 			last_tr->dt_run = 0;
2537 			last_tr->dt_delay = 0;
2538 			last_tr->dt_sleep = 0;
2539 			last_tr->dt_iowait = 0;
2540 			last_tr->dt_preempt = 0;
2541 
2542 			if (itr->cursor.nr)
2543 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2544 
2545 			itr->last_thread = NULL;
2546 		}
2547 	}
2548 
2549 	if (!sched->summary_only)
2550 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2551 
2552 out:
2553 	if (sched->hist_time.start == 0 && t >= ptime->start)
2554 		sched->hist_time.start = t;
2555 	if (ptime->end == 0 || t <= ptime->end)
2556 		sched->hist_time.end = t;
2557 
2558 	if (tr) {
2559 		/* time of this sched_switch event becomes last time task seen */
2560 		tr->last_time = sample->time;
2561 
2562 		/* last state is used to determine where to account wait time */
2563 		tr->last_state = state;
2564 
2565 		/* sched out event for task so reset ready to run time */
2566 		tr->ready_to_run = 0;
2567 	}
2568 
2569 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2570 
2571 	return rc;
2572 }
2573 
2574 static int timehist_sched_switch_event(struct perf_tool *tool,
2575 			     union perf_event *event,
2576 			     struct perf_evsel *evsel,
2577 			     struct perf_sample *sample,
2578 			     struct machine *machine __maybe_unused)
2579 {
2580 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2581 }
2582 
2583 static int process_lost(struct perf_tool *tool __maybe_unused,
2584 			union perf_event *event,
2585 			struct perf_sample *sample,
2586 			struct machine *machine __maybe_unused)
2587 {
2588 	char tstr[64];
2589 
2590 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2591 	printf("%15s ", tstr);
2592 	printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2593 
2594 	return 0;
2595 }
2596 
2597 
2598 static void print_thread_runtime(struct thread *t,
2599 				 struct thread_runtime *r)
2600 {
2601 	double mean = avg_stats(&r->run_stats);
2602 	float stddev;
2603 
2604 	printf("%*s   %5d  %9" PRIu64 " ",
2605 	       comm_width, timehist_get_commstr(t), t->ppid,
2606 	       (u64) r->run_stats.n);
2607 
2608 	print_sched_time(r->total_run_time, 8);
2609 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2610 	print_sched_time(r->run_stats.min, 6);
2611 	printf(" ");
2612 	print_sched_time((u64) mean, 6);
2613 	printf(" ");
2614 	print_sched_time(r->run_stats.max, 6);
2615 	printf("  ");
2616 	printf("%5.2f", stddev);
2617 	printf("   %5" PRIu64, r->migrations);
2618 	printf("\n");
2619 }
2620 
2621 static void print_thread_waittime(struct thread *t,
2622 				  struct thread_runtime *r)
2623 {
2624 	printf("%*s   %5d  %9" PRIu64 " ",
2625 	       comm_width, timehist_get_commstr(t), t->ppid,
2626 	       (u64) r->run_stats.n);
2627 
2628 	print_sched_time(r->total_run_time, 8);
2629 	print_sched_time(r->total_sleep_time, 6);
2630 	printf(" ");
2631 	print_sched_time(r->total_iowait_time, 6);
2632 	printf(" ");
2633 	print_sched_time(r->total_preempt_time, 6);
2634 	printf(" ");
2635 	print_sched_time(r->total_delay_time, 6);
2636 	printf("\n");
2637 }
2638 
2639 struct total_run_stats {
2640 	struct perf_sched *sched;
2641 	u64  sched_count;
2642 	u64  task_count;
2643 	u64  total_run_time;
2644 };
2645 
2646 static int __show_thread_runtime(struct thread *t, void *priv)
2647 {
2648 	struct total_run_stats *stats = priv;
2649 	struct thread_runtime *r;
2650 
2651 	if (thread__is_filtered(t))
2652 		return 0;
2653 
2654 	r = thread__priv(t);
2655 	if (r && r->run_stats.n) {
2656 		stats->task_count++;
2657 		stats->sched_count += r->run_stats.n;
2658 		stats->total_run_time += r->total_run_time;
2659 
2660 		if (stats->sched->show_state)
2661 			print_thread_waittime(t, r);
2662 		else
2663 			print_thread_runtime(t, r);
2664 	}
2665 
2666 	return 0;
2667 }
2668 
2669 static int show_thread_runtime(struct thread *t, void *priv)
2670 {
2671 	if (t->dead)
2672 		return 0;
2673 
2674 	return __show_thread_runtime(t, priv);
2675 }
2676 
2677 static int show_deadthread_runtime(struct thread *t, void *priv)
2678 {
2679 	if (!t->dead)
2680 		return 0;
2681 
2682 	return __show_thread_runtime(t, priv);
2683 }
2684 
2685 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2686 {
2687 	const char *sep = " <- ";
2688 	struct callchain_list *chain;
2689 	size_t ret = 0;
2690 	char bf[1024];
2691 	bool first;
2692 
2693 	if (node == NULL)
2694 		return 0;
2695 
2696 	ret = callchain__fprintf_folded(fp, node->parent);
2697 	first = (ret == 0);
2698 
2699 	list_for_each_entry(chain, &node->val, list) {
2700 		if (chain->ip >= PERF_CONTEXT_MAX)
2701 			continue;
2702 		if (chain->ms.sym && chain->ms.sym->ignore)
2703 			continue;
2704 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2705 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2706 							false));
2707 		first = false;
2708 	}
2709 
2710 	return ret;
2711 }
2712 
2713 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2714 {
2715 	size_t ret = 0;
2716 	FILE *fp = stdout;
2717 	struct callchain_node *chain;
2718 	struct rb_node *rb_node = rb_first(root);
2719 
2720 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2721 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2722 	       graph_dotted_line);
2723 
2724 	while (rb_node) {
2725 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2726 		rb_node = rb_next(rb_node);
2727 
2728 		ret += fprintf(fp, "  ");
2729 		print_sched_time(chain->hit, 12);
2730 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2731 		ret += fprintf(fp, " %8d  ", chain->count);
2732 		ret += callchain__fprintf_folded(fp, chain);
2733 		ret += fprintf(fp, "\n");
2734 	}
2735 
2736 	return ret;
2737 }
2738 
2739 static void timehist_print_summary(struct perf_sched *sched,
2740 				   struct perf_session *session)
2741 {
2742 	struct machine *m = &session->machines.host;
2743 	struct total_run_stats totals;
2744 	u64 task_count;
2745 	struct thread *t;
2746 	struct thread_runtime *r;
2747 	int i;
2748 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2749 
2750 	memset(&totals, 0, sizeof(totals));
2751 	totals.sched = sched;
2752 
2753 	if (sched->idle_hist) {
2754 		printf("\nIdle-time summary\n");
2755 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2756 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2757 	} else if (sched->show_state) {
2758 		printf("\nWait-time summary\n");
2759 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2760 		printf("   run-time      sleep      iowait     preempt       delay\n");
2761 	} else {
2762 		printf("\nRuntime summary\n");
2763 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2764 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2765 	}
2766 	printf("%*s            (count)  ", comm_width, "");
2767 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2768 	       sched->show_state ? "(msec)" : "%");
2769 	printf("%.117s\n", graph_dotted_line);
2770 
2771 	machine__for_each_thread(m, show_thread_runtime, &totals);
2772 	task_count = totals.task_count;
2773 	if (!task_count)
2774 		printf("<no still running tasks>\n");
2775 
2776 	printf("\nTerminated tasks:\n");
2777 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2778 	if (task_count == totals.task_count)
2779 		printf("<no terminated tasks>\n");
2780 
2781 	/* CPU idle stats not tracked when samples were skipped */
2782 	if (sched->skipped_samples && !sched->idle_hist)
2783 		return;
2784 
2785 	printf("\nIdle stats:\n");
2786 	for (i = 0; i < idle_max_cpu; ++i) {
2787 		t = idle_threads[i];
2788 		if (!t)
2789 			continue;
2790 
2791 		r = thread__priv(t);
2792 		if (r && r->run_stats.n) {
2793 			totals.sched_count += r->run_stats.n;
2794 			printf("    CPU %2d idle for ", i);
2795 			print_sched_time(r->total_run_time, 6);
2796 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2797 		} else
2798 			printf("    CPU %2d idle entire time window\n", i);
2799 	}
2800 
2801 	if (sched->idle_hist && symbol_conf.use_callchain) {
2802 		callchain_param.mode  = CHAIN_FOLDED;
2803 		callchain_param.value = CCVAL_PERIOD;
2804 
2805 		callchain_register_param(&callchain_param);
2806 
2807 		printf("\nIdle stats by callchain:\n");
2808 		for (i = 0; i < idle_max_cpu; ++i) {
2809 			struct idle_thread_runtime *itr;
2810 
2811 			t = idle_threads[i];
2812 			if (!t)
2813 				continue;
2814 
2815 			itr = thread__priv(t);
2816 			if (itr == NULL)
2817 				continue;
2818 
2819 			callchain_param.sort(&itr->sorted_root, &itr->callchain,
2820 					     0, &callchain_param);
2821 
2822 			printf("  CPU %2d:", i);
2823 			print_sched_time(itr->tr.total_run_time, 6);
2824 			printf(" msec\n");
2825 			timehist_print_idlehist_callchain(&itr->sorted_root);
2826 			printf("\n");
2827 		}
2828 	}
2829 
2830 	printf("\n"
2831 	       "    Total number of unique tasks: %" PRIu64 "\n"
2832 	       "Total number of context switches: %" PRIu64 "\n",
2833 	       totals.task_count, totals.sched_count);
2834 
2835 	printf("           Total run time (msec): ");
2836 	print_sched_time(totals.total_run_time, 2);
2837 	printf("\n");
2838 
2839 	printf("    Total scheduling time (msec): ");
2840 	print_sched_time(hist_time, 2);
2841 	printf(" (x %d)\n", sched->max_cpu);
2842 }
2843 
2844 typedef int (*sched_handler)(struct perf_tool *tool,
2845 			  union perf_event *event,
2846 			  struct perf_evsel *evsel,
2847 			  struct perf_sample *sample,
2848 			  struct machine *machine);
2849 
2850 static int perf_timehist__process_sample(struct perf_tool *tool,
2851 					 union perf_event *event,
2852 					 struct perf_sample *sample,
2853 					 struct perf_evsel *evsel,
2854 					 struct machine *machine)
2855 {
2856 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2857 	int err = 0;
2858 	int this_cpu = sample->cpu;
2859 
2860 	if (this_cpu > sched->max_cpu)
2861 		sched->max_cpu = this_cpu;
2862 
2863 	if (evsel->handler != NULL) {
2864 		sched_handler f = evsel->handler;
2865 
2866 		err = f(tool, event, evsel, sample, machine);
2867 	}
2868 
2869 	return err;
2870 }
2871 
2872 static int timehist_check_attr(struct perf_sched *sched,
2873 			       struct perf_evlist *evlist)
2874 {
2875 	struct perf_evsel *evsel;
2876 	struct evsel_runtime *er;
2877 
2878 	list_for_each_entry(evsel, &evlist->entries, node) {
2879 		er = perf_evsel__get_runtime(evsel);
2880 		if (er == NULL) {
2881 			pr_err("Failed to allocate memory for evsel runtime data\n");
2882 			return -1;
2883 		}
2884 
2885 		if (sched->show_callchain &&
2886 		    !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2887 			pr_info("Samples do not have callchains.\n");
2888 			sched->show_callchain = 0;
2889 			symbol_conf.use_callchain = 0;
2890 		}
2891 	}
2892 
2893 	return 0;
2894 }
2895 
2896 static int perf_sched__timehist(struct perf_sched *sched)
2897 {
2898 	const struct perf_evsel_str_handler handlers[] = {
2899 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2900 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2901 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2902 	};
2903 	const struct perf_evsel_str_handler migrate_handlers[] = {
2904 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2905 	};
2906 	struct perf_data_file file = {
2907 		.path = input_name,
2908 		.mode = PERF_DATA_MODE_READ,
2909 		.force = sched->force,
2910 	};
2911 
2912 	struct perf_session *session;
2913 	struct perf_evlist *evlist;
2914 	int err = -1;
2915 
2916 	/*
2917 	 * event handlers for timehist option
2918 	 */
2919 	sched->tool.sample	 = perf_timehist__process_sample;
2920 	sched->tool.mmap	 = perf_event__process_mmap;
2921 	sched->tool.comm	 = perf_event__process_comm;
2922 	sched->tool.exit	 = perf_event__process_exit;
2923 	sched->tool.fork	 = perf_event__process_fork;
2924 	sched->tool.lost	 = process_lost;
2925 	sched->tool.attr	 = perf_event__process_attr;
2926 	sched->tool.tracing_data = perf_event__process_tracing_data;
2927 	sched->tool.build_id	 = perf_event__process_build_id;
2928 
2929 	sched->tool.ordered_events = true;
2930 	sched->tool.ordering_requires_timestamps = true;
2931 
2932 	symbol_conf.use_callchain = sched->show_callchain;
2933 
2934 	session = perf_session__new(&file, false, &sched->tool);
2935 	if (session == NULL)
2936 		return -ENOMEM;
2937 
2938 	evlist = session->evlist;
2939 
2940 	symbol__init(&session->header.env);
2941 
2942 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2943 		pr_err("Invalid time string\n");
2944 		return -EINVAL;
2945 	}
2946 
2947 	if (timehist_check_attr(sched, evlist) != 0)
2948 		goto out;
2949 
2950 	setup_pager();
2951 
2952 	/* setup per-evsel handlers */
2953 	if (perf_session__set_tracepoints_handlers(session, handlers))
2954 		goto out;
2955 
2956 	/* sched_switch event at a minimum needs to exist */
2957 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2958 						  "sched:sched_switch")) {
2959 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2960 		goto out;
2961 	}
2962 
2963 	if (sched->show_migrations &&
2964 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
2965 		goto out;
2966 
2967 	/* pre-allocate struct for per-CPU idle stats */
2968 	sched->max_cpu = session->header.env.nr_cpus_online;
2969 	if (sched->max_cpu == 0)
2970 		sched->max_cpu = 4;
2971 	if (init_idle_threads(sched->max_cpu))
2972 		goto out;
2973 
2974 	/* summary_only implies summary option, but don't overwrite summary if set */
2975 	if (sched->summary_only)
2976 		sched->summary = sched->summary_only;
2977 
2978 	if (!sched->summary_only)
2979 		timehist_header(sched);
2980 
2981 	err = perf_session__process_events(session);
2982 	if (err) {
2983 		pr_err("Failed to process events, error %d", err);
2984 		goto out;
2985 	}
2986 
2987 	sched->nr_events      = evlist->stats.nr_events[0];
2988 	sched->nr_lost_events = evlist->stats.total_lost;
2989 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2990 
2991 	if (sched->summary)
2992 		timehist_print_summary(sched, session);
2993 
2994 out:
2995 	free_idle_threads();
2996 	perf_session__delete(session);
2997 
2998 	return err;
2999 }
3000 
3001 
3002 static void print_bad_events(struct perf_sched *sched)
3003 {
3004 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3005 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3006 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3007 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3008 	}
3009 	if (sched->nr_lost_events && sched->nr_events) {
3010 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3011 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3012 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3013 	}
3014 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3015 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3016 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3017 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3018 		if (sched->nr_lost_events)
3019 			printf(" (due to lost events?)");
3020 		printf("\n");
3021 	}
3022 }
3023 
3024 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
3025 {
3026 	struct rb_node **new = &(root->rb_node), *parent = NULL;
3027 	struct work_atoms *this;
3028 	const char *comm = thread__comm_str(data->thread), *this_comm;
3029 
3030 	while (*new) {
3031 		int cmp;
3032 
3033 		this = container_of(*new, struct work_atoms, node);
3034 		parent = *new;
3035 
3036 		this_comm = thread__comm_str(this->thread);
3037 		cmp = strcmp(comm, this_comm);
3038 		if (cmp > 0) {
3039 			new = &((*new)->rb_left);
3040 		} else if (cmp < 0) {
3041 			new = &((*new)->rb_right);
3042 		} else {
3043 			this->num_merged++;
3044 			this->total_runtime += data->total_runtime;
3045 			this->nb_atoms += data->nb_atoms;
3046 			this->total_lat += data->total_lat;
3047 			list_splice(&data->work_list, &this->work_list);
3048 			if (this->max_lat < data->max_lat) {
3049 				this->max_lat = data->max_lat;
3050 				this->max_lat_at = data->max_lat_at;
3051 			}
3052 			zfree(&data);
3053 			return;
3054 		}
3055 	}
3056 
3057 	data->num_merged++;
3058 	rb_link_node(&data->node, parent, new);
3059 	rb_insert_color(&data->node, root);
3060 }
3061 
3062 static void perf_sched__merge_lat(struct perf_sched *sched)
3063 {
3064 	struct work_atoms *data;
3065 	struct rb_node *node;
3066 
3067 	if (sched->skip_merge)
3068 		return;
3069 
3070 	while ((node = rb_first(&sched->atom_root))) {
3071 		rb_erase(node, &sched->atom_root);
3072 		data = rb_entry(node, struct work_atoms, node);
3073 		__merge_work_atoms(&sched->merged_atom_root, data);
3074 	}
3075 }
3076 
3077 static int perf_sched__lat(struct perf_sched *sched)
3078 {
3079 	struct rb_node *next;
3080 
3081 	setup_pager();
3082 
3083 	if (perf_sched__read_events(sched))
3084 		return -1;
3085 
3086 	perf_sched__merge_lat(sched);
3087 	perf_sched__sort_lat(sched);
3088 
3089 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3090 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3091 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3092 
3093 	next = rb_first(&sched->sorted_atom_root);
3094 
3095 	while (next) {
3096 		struct work_atoms *work_list;
3097 
3098 		work_list = rb_entry(next, struct work_atoms, node);
3099 		output_lat_thread(sched, work_list);
3100 		next = rb_next(next);
3101 		thread__zput(work_list->thread);
3102 	}
3103 
3104 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3105 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3106 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3107 
3108 	printf(" ---------------------------------------------------\n");
3109 
3110 	print_bad_events(sched);
3111 	printf("\n");
3112 
3113 	return 0;
3114 }
3115 
3116 static int setup_map_cpus(struct perf_sched *sched)
3117 {
3118 	struct cpu_map *map;
3119 
3120 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3121 
3122 	if (sched->map.comp) {
3123 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3124 		if (!sched->map.comp_cpus)
3125 			return -1;
3126 	}
3127 
3128 	if (!sched->map.cpus_str)
3129 		return 0;
3130 
3131 	map = cpu_map__new(sched->map.cpus_str);
3132 	if (!map) {
3133 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3134 		return -1;
3135 	}
3136 
3137 	sched->map.cpus = map;
3138 	return 0;
3139 }
3140 
3141 static int setup_color_pids(struct perf_sched *sched)
3142 {
3143 	struct thread_map *map;
3144 
3145 	if (!sched->map.color_pids_str)
3146 		return 0;
3147 
3148 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3149 	if (!map) {
3150 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3151 		return -1;
3152 	}
3153 
3154 	sched->map.color_pids = map;
3155 	return 0;
3156 }
3157 
3158 static int setup_color_cpus(struct perf_sched *sched)
3159 {
3160 	struct cpu_map *map;
3161 
3162 	if (!sched->map.color_cpus_str)
3163 		return 0;
3164 
3165 	map = cpu_map__new(sched->map.color_cpus_str);
3166 	if (!map) {
3167 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3168 		return -1;
3169 	}
3170 
3171 	sched->map.color_cpus = map;
3172 	return 0;
3173 }
3174 
3175 static int perf_sched__map(struct perf_sched *sched)
3176 {
3177 	if (setup_map_cpus(sched))
3178 		return -1;
3179 
3180 	if (setup_color_pids(sched))
3181 		return -1;
3182 
3183 	if (setup_color_cpus(sched))
3184 		return -1;
3185 
3186 	setup_pager();
3187 	if (perf_sched__read_events(sched))
3188 		return -1;
3189 	print_bad_events(sched);
3190 	return 0;
3191 }
3192 
3193 static int perf_sched__replay(struct perf_sched *sched)
3194 {
3195 	unsigned long i;
3196 
3197 	calibrate_run_measurement_overhead(sched);
3198 	calibrate_sleep_measurement_overhead(sched);
3199 
3200 	test_calibrations(sched);
3201 
3202 	if (perf_sched__read_events(sched))
3203 		return -1;
3204 
3205 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3206 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3207 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3208 
3209 	if (sched->targetless_wakeups)
3210 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3211 	if (sched->multitarget_wakeups)
3212 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3213 	if (sched->nr_run_events_optimized)
3214 		printf("run atoms optimized: %ld\n",
3215 			sched->nr_run_events_optimized);
3216 
3217 	print_task_traces(sched);
3218 	add_cross_task_wakeups(sched);
3219 
3220 	create_tasks(sched);
3221 	printf("------------------------------------------------------------\n");
3222 	for (i = 0; i < sched->replay_repeat; i++)
3223 		run_one_test(sched);
3224 
3225 	return 0;
3226 }
3227 
3228 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3229 			  const char * const usage_msg[])
3230 {
3231 	char *tmp, *tok, *str = strdup(sched->sort_order);
3232 
3233 	for (tok = strtok_r(str, ", ", &tmp);
3234 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3235 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3236 			usage_with_options_msg(usage_msg, options,
3237 					"Unknown --sort key: `%s'", tok);
3238 		}
3239 	}
3240 
3241 	free(str);
3242 
3243 	sort_dimension__add("pid", &sched->cmp_pid);
3244 }
3245 
3246 static int __cmd_record(int argc, const char **argv)
3247 {
3248 	unsigned int rec_argc, i, j;
3249 	const char **rec_argv;
3250 	const char * const record_args[] = {
3251 		"record",
3252 		"-a",
3253 		"-R",
3254 		"-m", "1024",
3255 		"-c", "1",
3256 		"-e", "sched:sched_switch",
3257 		"-e", "sched:sched_stat_wait",
3258 		"-e", "sched:sched_stat_sleep",
3259 		"-e", "sched:sched_stat_iowait",
3260 		"-e", "sched:sched_stat_runtime",
3261 		"-e", "sched:sched_process_fork",
3262 		"-e", "sched:sched_wakeup",
3263 		"-e", "sched:sched_wakeup_new",
3264 		"-e", "sched:sched_migrate_task",
3265 	};
3266 
3267 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3268 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3269 
3270 	if (rec_argv == NULL)
3271 		return -ENOMEM;
3272 
3273 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3274 		rec_argv[i] = strdup(record_args[i]);
3275 
3276 	for (j = 1; j < (unsigned int)argc; j++, i++)
3277 		rec_argv[i] = argv[j];
3278 
3279 	BUG_ON(i != rec_argc);
3280 
3281 	return cmd_record(i, rec_argv);
3282 }
3283 
3284 int cmd_sched(int argc, const char **argv)
3285 {
3286 	const char default_sort_order[] = "avg, max, switch, runtime";
3287 	struct perf_sched sched = {
3288 		.tool = {
3289 			.sample		 = perf_sched__process_tracepoint_sample,
3290 			.comm		 = perf_event__process_comm,
3291 			.namespaces	 = perf_event__process_namespaces,
3292 			.lost		 = perf_event__process_lost,
3293 			.fork		 = perf_sched__process_fork_event,
3294 			.ordered_events = true,
3295 		},
3296 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3297 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3298 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3299 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3300 		.sort_order	      = default_sort_order,
3301 		.replay_repeat	      = 10,
3302 		.profile_cpu	      = -1,
3303 		.next_shortname1      = 'A',
3304 		.next_shortname2      = '0',
3305 		.skip_merge           = 0,
3306 		.show_callchain	      = 1,
3307 		.max_stack            = 5,
3308 	};
3309 	const struct option sched_options[] = {
3310 	OPT_STRING('i', "input", &input_name, "file",
3311 		    "input file name"),
3312 	OPT_INCR('v', "verbose", &verbose,
3313 		    "be more verbose (show symbol address, etc)"),
3314 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3315 		    "dump raw trace in ASCII"),
3316 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3317 	OPT_END()
3318 	};
3319 	const struct option latency_options[] = {
3320 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3321 		   "sort by key(s): runtime, switch, avg, max"),
3322 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3323 		    "CPU to profile on"),
3324 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3325 		    "latency stats per pid instead of per comm"),
3326 	OPT_PARENT(sched_options)
3327 	};
3328 	const struct option replay_options[] = {
3329 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3330 		     "repeat the workload replay N times (-1: infinite)"),
3331 	OPT_PARENT(sched_options)
3332 	};
3333 	const struct option map_options[] = {
3334 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3335 		    "map output in compact mode"),
3336 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3337 		   "highlight given pids in map"),
3338 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3339                     "highlight given CPUs in map"),
3340 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3341                     "display given CPUs in map"),
3342 	OPT_PARENT(sched_options)
3343 	};
3344 	const struct option timehist_options[] = {
3345 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3346 		   "file", "vmlinux pathname"),
3347 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3348 		   "file", "kallsyms pathname"),
3349 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3350 		    "Display call chains if present (default on)"),
3351 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3352 		   "Maximum number of functions to display backtrace."),
3353 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3354 		    "Look for files with symbols relative to this directory"),
3355 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3356 		    "Show only syscall summary with statistics"),
3357 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3358 		    "Show all syscalls and summary with statistics"),
3359 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3360 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3361 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3362 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3363 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3364 	OPT_STRING(0, "time", &sched.time_str, "str",
3365 		   "Time span for analysis (start,stop)"),
3366 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3367 	OPT_PARENT(sched_options)
3368 	};
3369 
3370 	const char * const latency_usage[] = {
3371 		"perf sched latency [<options>]",
3372 		NULL
3373 	};
3374 	const char * const replay_usage[] = {
3375 		"perf sched replay [<options>]",
3376 		NULL
3377 	};
3378 	const char * const map_usage[] = {
3379 		"perf sched map [<options>]",
3380 		NULL
3381 	};
3382 	const char * const timehist_usage[] = {
3383 		"perf sched timehist [<options>]",
3384 		NULL
3385 	};
3386 	const char *const sched_subcommands[] = { "record", "latency", "map",
3387 						  "replay", "script",
3388 						  "timehist", NULL };
3389 	const char *sched_usage[] = {
3390 		NULL,
3391 		NULL
3392 	};
3393 	struct trace_sched_handler lat_ops  = {
3394 		.wakeup_event	    = latency_wakeup_event,
3395 		.switch_event	    = latency_switch_event,
3396 		.runtime_event	    = latency_runtime_event,
3397 		.migrate_task_event = latency_migrate_task_event,
3398 	};
3399 	struct trace_sched_handler map_ops  = {
3400 		.switch_event	    = map_switch_event,
3401 	};
3402 	struct trace_sched_handler replay_ops  = {
3403 		.wakeup_event	    = replay_wakeup_event,
3404 		.switch_event	    = replay_switch_event,
3405 		.fork_event	    = replay_fork_event,
3406 	};
3407 	unsigned int i;
3408 
3409 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3410 		sched.curr_pid[i] = -1;
3411 
3412 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3413 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3414 	if (!argc)
3415 		usage_with_options(sched_usage, sched_options);
3416 
3417 	/*
3418 	 * Aliased to 'perf script' for now:
3419 	 */
3420 	if (!strcmp(argv[0], "script"))
3421 		return cmd_script(argc, argv);
3422 
3423 	if (!strncmp(argv[0], "rec", 3)) {
3424 		return __cmd_record(argc, argv);
3425 	} else if (!strncmp(argv[0], "lat", 3)) {
3426 		sched.tp_handler = &lat_ops;
3427 		if (argc > 1) {
3428 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3429 			if (argc)
3430 				usage_with_options(latency_usage, latency_options);
3431 		}
3432 		setup_sorting(&sched, latency_options, latency_usage);
3433 		return perf_sched__lat(&sched);
3434 	} else if (!strcmp(argv[0], "map")) {
3435 		if (argc) {
3436 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3437 			if (argc)
3438 				usage_with_options(map_usage, map_options);
3439 		}
3440 		sched.tp_handler = &map_ops;
3441 		setup_sorting(&sched, latency_options, latency_usage);
3442 		return perf_sched__map(&sched);
3443 	} else if (!strncmp(argv[0], "rep", 3)) {
3444 		sched.tp_handler = &replay_ops;
3445 		if (argc) {
3446 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3447 			if (argc)
3448 				usage_with_options(replay_usage, replay_options);
3449 		}
3450 		return perf_sched__replay(&sched);
3451 	} else if (!strcmp(argv[0], "timehist")) {
3452 		if (argc) {
3453 			argc = parse_options(argc, argv, timehist_options,
3454 					     timehist_usage, 0);
3455 			if (argc)
3456 				usage_with_options(timehist_usage, timehist_options);
3457 		}
3458 		if ((sched.show_wakeups || sched.show_next) &&
3459 		    sched.summary_only) {
3460 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3461 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3462 			if (sched.show_wakeups)
3463 				parse_options_usage(NULL, timehist_options, "w", true);
3464 			if (sched.show_next)
3465 				parse_options_usage(NULL, timehist_options, "n", true);
3466 			return -EINVAL;
3467 		}
3468 
3469 		return perf_sched__timehist(&sched);
3470 	} else {
3471 		usage_with_options(sched_usage, sched_options);
3472 	}
3473 
3474 	return 0;
3475 }
3476