xref: /linux/tools/perf/builtin-sched.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf-sys.h"
4 
5 #include "util/cpumap.h"
6 #include "util/evlist.h"
7 #include "util/evsel.h"
8 #include "util/evsel_fprintf.h"
9 #include "util/mutex.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
22 
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
26 
27 #include "util/debug.h"
28 #include "util/event.h"
29 #include "util/util.h"
30 
31 #include <linux/kernel.h>
32 #include <linux/log2.h>
33 #include <linux/zalloc.h>
34 #include <sys/prctl.h>
35 #include <sys/resource.h>
36 #include <inttypes.h>
37 
38 #include <errno.h>
39 #include <semaphore.h>
40 #include <pthread.h>
41 #include <math.h>
42 #include <api/fs/fs.h>
43 #include <perf/cpumap.h>
44 #include <linux/time64.h>
45 #include <linux/err.h>
46 
47 #include <linux/ctype.h>
48 
49 #define PR_SET_NAME		15               /* Set process name */
50 #define MAX_CPUS		4096
51 #define COMM_LEN		20
52 #define SYM_LEN			129
53 #define MAX_PID			1024000
54 #define MAX_PRIO		140
55 
56 static const char *cpu_list;
57 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
58 
59 struct sched_atom;
60 
61 struct task_desc {
62 	unsigned long		nr;
63 	unsigned long		pid;
64 	char			comm[COMM_LEN];
65 
66 	unsigned long		nr_events;
67 	unsigned long		curr_event;
68 	struct sched_atom	**atoms;
69 
70 	pthread_t		thread;
71 	sem_t			sleep_sem;
72 
73 	sem_t			ready_for_work;
74 	sem_t			work_done_sem;
75 
76 	u64			cpu_usage;
77 };
78 
79 enum sched_event_type {
80 	SCHED_EVENT_RUN,
81 	SCHED_EVENT_SLEEP,
82 	SCHED_EVENT_WAKEUP,
83 	SCHED_EVENT_MIGRATION,
84 };
85 
86 struct sched_atom {
87 	enum sched_event_type	type;
88 	int			specific_wait;
89 	u64			timestamp;
90 	u64			duration;
91 	unsigned long		nr;
92 	sem_t			*wait_sem;
93 	struct task_desc	*wakee;
94 };
95 
96 enum thread_state {
97 	THREAD_SLEEPING = 0,
98 	THREAD_WAIT_CPU,
99 	THREAD_SCHED_IN,
100 	THREAD_IGNORE
101 };
102 
103 struct work_atom {
104 	struct list_head	list;
105 	enum thread_state	state;
106 	u64			sched_out_time;
107 	u64			wake_up_time;
108 	u64			sched_in_time;
109 	u64			runtime;
110 };
111 
112 struct work_atoms {
113 	struct list_head	work_list;
114 	struct thread		*thread;
115 	struct rb_node		node;
116 	u64			max_lat;
117 	u64			max_lat_start;
118 	u64			max_lat_end;
119 	u64			total_lat;
120 	u64			nb_atoms;
121 	u64			total_runtime;
122 	int			num_merged;
123 };
124 
125 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
126 
127 struct perf_sched;
128 
129 struct trace_sched_handler {
130 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
131 			    struct perf_sample *sample, struct machine *machine);
132 
133 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
134 			     struct perf_sample *sample, struct machine *machine);
135 
136 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
137 			    struct perf_sample *sample, struct machine *machine);
138 
139 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
140 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
141 			  struct machine *machine);
142 
143 	int (*migrate_task_event)(struct perf_sched *sched,
144 				  struct evsel *evsel,
145 				  struct perf_sample *sample,
146 				  struct machine *machine);
147 };
148 
149 #define COLOR_PIDS PERF_COLOR_BLUE
150 #define COLOR_CPUS PERF_COLOR_BG_RED
151 
152 struct perf_sched_map {
153 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
154 	struct perf_cpu		*comp_cpus;
155 	bool			 comp;
156 	struct perf_thread_map *color_pids;
157 	const char		*color_pids_str;
158 	struct perf_cpu_map	*color_cpus;
159 	const char		*color_cpus_str;
160 	const char		*task_name;
161 	struct strlist		*task_names;
162 	bool			fuzzy;
163 	struct perf_cpu_map	*cpus;
164 	const char		*cpus_str;
165 };
166 
167 struct perf_sched {
168 	struct perf_tool tool;
169 	const char	 *sort_order;
170 	unsigned long	 nr_tasks;
171 	struct task_desc **pid_to_task;
172 	struct task_desc **tasks;
173 	const struct trace_sched_handler *tp_handler;
174 	struct mutex	 start_work_mutex;
175 	struct mutex	 work_done_wait_mutex;
176 	int		 profile_cpu;
177 /*
178  * Track the current task - that way we can know whether there's any
179  * weird events, such as a task being switched away that is not current.
180  */
181 	struct perf_cpu	 max_cpu;
182 	u32		 *curr_pid;
183 	struct thread	 **curr_thread;
184 	struct thread	 **curr_out_thread;
185 	char		 next_shortname1;
186 	char		 next_shortname2;
187 	unsigned int	 replay_repeat;
188 	unsigned long	 nr_run_events;
189 	unsigned long	 nr_sleep_events;
190 	unsigned long	 nr_wakeup_events;
191 	unsigned long	 nr_sleep_corrections;
192 	unsigned long	 nr_run_events_optimized;
193 	unsigned long	 targetless_wakeups;
194 	unsigned long	 multitarget_wakeups;
195 	unsigned long	 nr_runs;
196 	unsigned long	 nr_timestamps;
197 	unsigned long	 nr_unordered_timestamps;
198 	unsigned long	 nr_context_switch_bugs;
199 	unsigned long	 nr_events;
200 	unsigned long	 nr_lost_chunks;
201 	unsigned long	 nr_lost_events;
202 	u64		 run_measurement_overhead;
203 	u64		 sleep_measurement_overhead;
204 	u64		 start_time;
205 	u64		 cpu_usage;
206 	u64		 runavg_cpu_usage;
207 	u64		 parent_cpu_usage;
208 	u64		 runavg_parent_cpu_usage;
209 	u64		 sum_runtime;
210 	u64		 sum_fluct;
211 	u64		 run_avg;
212 	u64		 all_runtime;
213 	u64		 all_count;
214 	u64		 *cpu_last_switched;
215 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
216 	struct list_head sort_list, cmp_pid;
217 	bool force;
218 	bool skip_merge;
219 	struct perf_sched_map map;
220 
221 	/* options for timehist command */
222 	bool		summary;
223 	bool		summary_only;
224 	bool		idle_hist;
225 	bool		show_callchain;
226 	unsigned int	max_stack;
227 	bool		show_cpu_visual;
228 	bool		show_wakeups;
229 	bool		show_next;
230 	bool		show_migrations;
231 	bool		show_state;
232 	bool		show_prio;
233 	u64		skipped_samples;
234 	const char	*time_str;
235 	struct perf_time_interval ptime;
236 	struct perf_time_interval hist_time;
237 	volatile bool   thread_funcs_exit;
238 	const char	*prio_str;
239 	DECLARE_BITMAP(prio_bitmap, MAX_PRIO);
240 };
241 
242 /* per thread run time data */
243 struct thread_runtime {
244 	u64 last_time;      /* time of previous sched in/out event */
245 	u64 dt_run;         /* run time */
246 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
247 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
248 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
249 	u64 dt_delay;       /* time between wakeup and sched-in */
250 	u64 ready_to_run;   /* time of wakeup */
251 
252 	struct stats run_stats;
253 	u64 total_run_time;
254 	u64 total_sleep_time;
255 	u64 total_iowait_time;
256 	u64 total_preempt_time;
257 	u64 total_delay_time;
258 
259 	char last_state;
260 
261 	char shortname[3];
262 	bool comm_changed;
263 
264 	u64 migrations;
265 
266 	int prio;
267 };
268 
269 /* per event run time data */
270 struct evsel_runtime {
271 	u64 *last_time; /* time this event was last seen per cpu */
272 	u32 ncpu;       /* highest cpu slot allocated */
273 };
274 
275 /* per cpu idle time data */
276 struct idle_thread_runtime {
277 	struct thread_runtime	tr;
278 	struct thread		*last_thread;
279 	struct rb_root_cached	sorted_root;
280 	struct callchain_root	callchain;
281 	struct callchain_cursor	cursor;
282 };
283 
284 /* track idle times per cpu */
285 static struct thread **idle_threads;
286 static int idle_max_cpu;
287 static char idle_comm[] = "<idle>";
288 
289 static u64 get_nsecs(void)
290 {
291 	struct timespec ts;
292 
293 	clock_gettime(CLOCK_MONOTONIC, &ts);
294 
295 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
296 }
297 
298 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
299 {
300 	u64 T0 = get_nsecs(), T1;
301 
302 	do {
303 		T1 = get_nsecs();
304 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
305 }
306 
307 static void sleep_nsecs(u64 nsecs)
308 {
309 	struct timespec ts;
310 
311 	ts.tv_nsec = nsecs % 999999999;
312 	ts.tv_sec = nsecs / 999999999;
313 
314 	nanosleep(&ts, NULL);
315 }
316 
317 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
318 {
319 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
320 	int i;
321 
322 	for (i = 0; i < 10; i++) {
323 		T0 = get_nsecs();
324 		burn_nsecs(sched, 0);
325 		T1 = get_nsecs();
326 		delta = T1-T0;
327 		min_delta = min(min_delta, delta);
328 	}
329 	sched->run_measurement_overhead = min_delta;
330 
331 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
332 }
333 
334 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
335 {
336 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
337 	int i;
338 
339 	for (i = 0; i < 10; i++) {
340 		T0 = get_nsecs();
341 		sleep_nsecs(10000);
342 		T1 = get_nsecs();
343 		delta = T1-T0;
344 		min_delta = min(min_delta, delta);
345 	}
346 	min_delta -= 10000;
347 	sched->sleep_measurement_overhead = min_delta;
348 
349 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
350 }
351 
352 static struct sched_atom *
353 get_new_event(struct task_desc *task, u64 timestamp)
354 {
355 	struct sched_atom *event = zalloc(sizeof(*event));
356 	unsigned long idx = task->nr_events;
357 	size_t size;
358 
359 	event->timestamp = timestamp;
360 	event->nr = idx;
361 
362 	task->nr_events++;
363 	size = sizeof(struct sched_atom *) * task->nr_events;
364 	task->atoms = realloc(task->atoms, size);
365 	BUG_ON(!task->atoms);
366 
367 	task->atoms[idx] = event;
368 
369 	return event;
370 }
371 
372 static struct sched_atom *last_event(struct task_desc *task)
373 {
374 	if (!task->nr_events)
375 		return NULL;
376 
377 	return task->atoms[task->nr_events - 1];
378 }
379 
380 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
381 				u64 timestamp, u64 duration)
382 {
383 	struct sched_atom *event, *curr_event = last_event(task);
384 
385 	/*
386 	 * optimize an existing RUN event by merging this one
387 	 * to it:
388 	 */
389 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
390 		sched->nr_run_events_optimized++;
391 		curr_event->duration += duration;
392 		return;
393 	}
394 
395 	event = get_new_event(task, timestamp);
396 
397 	event->type = SCHED_EVENT_RUN;
398 	event->duration = duration;
399 
400 	sched->nr_run_events++;
401 }
402 
403 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
404 				   u64 timestamp, struct task_desc *wakee)
405 {
406 	struct sched_atom *event, *wakee_event;
407 
408 	event = get_new_event(task, timestamp);
409 	event->type = SCHED_EVENT_WAKEUP;
410 	event->wakee = wakee;
411 
412 	wakee_event = last_event(wakee);
413 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
414 		sched->targetless_wakeups++;
415 		return;
416 	}
417 	if (wakee_event->wait_sem) {
418 		sched->multitarget_wakeups++;
419 		return;
420 	}
421 
422 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
423 	sem_init(wakee_event->wait_sem, 0, 0);
424 	wakee_event->specific_wait = 1;
425 	event->wait_sem = wakee_event->wait_sem;
426 
427 	sched->nr_wakeup_events++;
428 }
429 
430 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
431 				  u64 timestamp, const char task_state __maybe_unused)
432 {
433 	struct sched_atom *event = get_new_event(task, timestamp);
434 
435 	event->type = SCHED_EVENT_SLEEP;
436 
437 	sched->nr_sleep_events++;
438 }
439 
440 static struct task_desc *register_pid(struct perf_sched *sched,
441 				      unsigned long pid, const char *comm)
442 {
443 	struct task_desc *task;
444 	static int pid_max;
445 
446 	if (sched->pid_to_task == NULL) {
447 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
448 			pid_max = MAX_PID;
449 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
450 	}
451 	if (pid >= (unsigned long)pid_max) {
452 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
453 			sizeof(struct task_desc *))) == NULL);
454 		while (pid >= (unsigned long)pid_max)
455 			sched->pid_to_task[pid_max++] = NULL;
456 	}
457 
458 	task = sched->pid_to_task[pid];
459 
460 	if (task)
461 		return task;
462 
463 	task = zalloc(sizeof(*task));
464 	task->pid = pid;
465 	task->nr = sched->nr_tasks;
466 	strcpy(task->comm, comm);
467 	/*
468 	 * every task starts in sleeping state - this gets ignored
469 	 * if there's no wakeup pointing to this sleep state:
470 	 */
471 	add_sched_event_sleep(sched, task, 0, 0);
472 
473 	sched->pid_to_task[pid] = task;
474 	sched->nr_tasks++;
475 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
476 	BUG_ON(!sched->tasks);
477 	sched->tasks[task->nr] = task;
478 
479 	if (verbose > 0)
480 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
481 
482 	return task;
483 }
484 
485 
486 static void print_task_traces(struct perf_sched *sched)
487 {
488 	struct task_desc *task;
489 	unsigned long i;
490 
491 	for (i = 0; i < sched->nr_tasks; i++) {
492 		task = sched->tasks[i];
493 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
494 			task->nr, task->comm, task->pid, task->nr_events);
495 	}
496 }
497 
498 static void add_cross_task_wakeups(struct perf_sched *sched)
499 {
500 	struct task_desc *task1, *task2;
501 	unsigned long i, j;
502 
503 	for (i = 0; i < sched->nr_tasks; i++) {
504 		task1 = sched->tasks[i];
505 		j = i + 1;
506 		if (j == sched->nr_tasks)
507 			j = 0;
508 		task2 = sched->tasks[j];
509 		add_sched_event_wakeup(sched, task1, 0, task2);
510 	}
511 }
512 
513 static void perf_sched__process_event(struct perf_sched *sched,
514 				      struct sched_atom *atom)
515 {
516 	int ret = 0;
517 
518 	switch (atom->type) {
519 		case SCHED_EVENT_RUN:
520 			burn_nsecs(sched, atom->duration);
521 			break;
522 		case SCHED_EVENT_SLEEP:
523 			if (atom->wait_sem)
524 				ret = sem_wait(atom->wait_sem);
525 			BUG_ON(ret);
526 			break;
527 		case SCHED_EVENT_WAKEUP:
528 			if (atom->wait_sem)
529 				ret = sem_post(atom->wait_sem);
530 			BUG_ON(ret);
531 			break;
532 		case SCHED_EVENT_MIGRATION:
533 			break;
534 		default:
535 			BUG_ON(1);
536 	}
537 }
538 
539 static u64 get_cpu_usage_nsec_parent(void)
540 {
541 	struct rusage ru;
542 	u64 sum;
543 	int err;
544 
545 	err = getrusage(RUSAGE_SELF, &ru);
546 	BUG_ON(err);
547 
548 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
549 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
550 
551 	return sum;
552 }
553 
554 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
555 {
556 	struct perf_event_attr attr;
557 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
558 	int fd;
559 	struct rlimit limit;
560 	bool need_privilege = false;
561 
562 	memset(&attr, 0, sizeof(attr));
563 
564 	attr.type = PERF_TYPE_SOFTWARE;
565 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
566 
567 force_again:
568 	fd = sys_perf_event_open(&attr, 0, -1, -1,
569 				 perf_event_open_cloexec_flag());
570 
571 	if (fd < 0) {
572 		if (errno == EMFILE) {
573 			if (sched->force) {
574 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
575 				limit.rlim_cur += sched->nr_tasks - cur_task;
576 				if (limit.rlim_cur > limit.rlim_max) {
577 					limit.rlim_max = limit.rlim_cur;
578 					need_privilege = true;
579 				}
580 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
581 					if (need_privilege && errno == EPERM)
582 						strcpy(info, "Need privilege\n");
583 				} else
584 					goto force_again;
585 			} else
586 				strcpy(info, "Have a try with -f option\n");
587 		}
588 		pr_err("Error: sys_perf_event_open() syscall returned "
589 		       "with %d (%s)\n%s", fd,
590 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
591 		exit(EXIT_FAILURE);
592 	}
593 	return fd;
594 }
595 
596 static u64 get_cpu_usage_nsec_self(int fd)
597 {
598 	u64 runtime;
599 	int ret;
600 
601 	ret = read(fd, &runtime, sizeof(runtime));
602 	BUG_ON(ret != sizeof(runtime));
603 
604 	return runtime;
605 }
606 
607 struct sched_thread_parms {
608 	struct task_desc  *task;
609 	struct perf_sched *sched;
610 	int fd;
611 };
612 
613 static void *thread_func(void *ctx)
614 {
615 	struct sched_thread_parms *parms = ctx;
616 	struct task_desc *this_task = parms->task;
617 	struct perf_sched *sched = parms->sched;
618 	u64 cpu_usage_0, cpu_usage_1;
619 	unsigned long i, ret;
620 	char comm2[22];
621 	int fd = parms->fd;
622 
623 	zfree(&parms);
624 
625 	sprintf(comm2, ":%s", this_task->comm);
626 	prctl(PR_SET_NAME, comm2);
627 	if (fd < 0)
628 		return NULL;
629 
630 	while (!sched->thread_funcs_exit) {
631 		ret = sem_post(&this_task->ready_for_work);
632 		BUG_ON(ret);
633 		mutex_lock(&sched->start_work_mutex);
634 		mutex_unlock(&sched->start_work_mutex);
635 
636 		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
637 
638 		for (i = 0; i < this_task->nr_events; i++) {
639 			this_task->curr_event = i;
640 			perf_sched__process_event(sched, this_task->atoms[i]);
641 		}
642 
643 		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
644 		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
645 		ret = sem_post(&this_task->work_done_sem);
646 		BUG_ON(ret);
647 
648 		mutex_lock(&sched->work_done_wait_mutex);
649 		mutex_unlock(&sched->work_done_wait_mutex);
650 	}
651 	return NULL;
652 }
653 
654 static void create_tasks(struct perf_sched *sched)
655 	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
656 	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
657 {
658 	struct task_desc *task;
659 	pthread_attr_t attr;
660 	unsigned long i;
661 	int err;
662 
663 	err = pthread_attr_init(&attr);
664 	BUG_ON(err);
665 	err = pthread_attr_setstacksize(&attr,
666 			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
667 	BUG_ON(err);
668 	mutex_lock(&sched->start_work_mutex);
669 	mutex_lock(&sched->work_done_wait_mutex);
670 	for (i = 0; i < sched->nr_tasks; i++) {
671 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
672 		BUG_ON(parms == NULL);
673 		parms->task = task = sched->tasks[i];
674 		parms->sched = sched;
675 		parms->fd = self_open_counters(sched, i);
676 		sem_init(&task->sleep_sem, 0, 0);
677 		sem_init(&task->ready_for_work, 0, 0);
678 		sem_init(&task->work_done_sem, 0, 0);
679 		task->curr_event = 0;
680 		err = pthread_create(&task->thread, &attr, thread_func, parms);
681 		BUG_ON(err);
682 	}
683 }
684 
685 static void destroy_tasks(struct perf_sched *sched)
686 	UNLOCK_FUNCTION(sched->start_work_mutex)
687 	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
688 {
689 	struct task_desc *task;
690 	unsigned long i;
691 	int err;
692 
693 	mutex_unlock(&sched->start_work_mutex);
694 	mutex_unlock(&sched->work_done_wait_mutex);
695 	/* Get rid of threads so they won't be upset by mutex destrunction */
696 	for (i = 0; i < sched->nr_tasks; i++) {
697 		task = sched->tasks[i];
698 		err = pthread_join(task->thread, NULL);
699 		BUG_ON(err);
700 		sem_destroy(&task->sleep_sem);
701 		sem_destroy(&task->ready_for_work);
702 		sem_destroy(&task->work_done_sem);
703 	}
704 }
705 
706 static void wait_for_tasks(struct perf_sched *sched)
707 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
708 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
709 {
710 	u64 cpu_usage_0, cpu_usage_1;
711 	struct task_desc *task;
712 	unsigned long i, ret;
713 
714 	sched->start_time = get_nsecs();
715 	sched->cpu_usage = 0;
716 	mutex_unlock(&sched->work_done_wait_mutex);
717 
718 	for (i = 0; i < sched->nr_tasks; i++) {
719 		task = sched->tasks[i];
720 		ret = sem_wait(&task->ready_for_work);
721 		BUG_ON(ret);
722 		sem_init(&task->ready_for_work, 0, 0);
723 	}
724 	mutex_lock(&sched->work_done_wait_mutex);
725 
726 	cpu_usage_0 = get_cpu_usage_nsec_parent();
727 
728 	mutex_unlock(&sched->start_work_mutex);
729 
730 	for (i = 0; i < sched->nr_tasks; i++) {
731 		task = sched->tasks[i];
732 		ret = sem_wait(&task->work_done_sem);
733 		BUG_ON(ret);
734 		sem_init(&task->work_done_sem, 0, 0);
735 		sched->cpu_usage += task->cpu_usage;
736 		task->cpu_usage = 0;
737 	}
738 
739 	cpu_usage_1 = get_cpu_usage_nsec_parent();
740 	if (!sched->runavg_cpu_usage)
741 		sched->runavg_cpu_usage = sched->cpu_usage;
742 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
743 
744 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
745 	if (!sched->runavg_parent_cpu_usage)
746 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
747 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
748 					 sched->parent_cpu_usage)/sched->replay_repeat;
749 
750 	mutex_lock(&sched->start_work_mutex);
751 
752 	for (i = 0; i < sched->nr_tasks; i++) {
753 		task = sched->tasks[i];
754 		sem_init(&task->sleep_sem, 0, 0);
755 		task->curr_event = 0;
756 	}
757 }
758 
759 static void run_one_test(struct perf_sched *sched)
760 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
761 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
762 {
763 	u64 T0, T1, delta, avg_delta, fluct;
764 
765 	T0 = get_nsecs();
766 	wait_for_tasks(sched);
767 	T1 = get_nsecs();
768 
769 	delta = T1 - T0;
770 	sched->sum_runtime += delta;
771 	sched->nr_runs++;
772 
773 	avg_delta = sched->sum_runtime / sched->nr_runs;
774 	if (delta < avg_delta)
775 		fluct = avg_delta - delta;
776 	else
777 		fluct = delta - avg_delta;
778 	sched->sum_fluct += fluct;
779 	if (!sched->run_avg)
780 		sched->run_avg = delta;
781 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
782 
783 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
784 
785 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
786 
787 	printf("cpu: %0.2f / %0.2f",
788 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
789 
790 #if 0
791 	/*
792 	 * rusage statistics done by the parent, these are less
793 	 * accurate than the sched->sum_exec_runtime based statistics:
794 	 */
795 	printf(" [%0.2f / %0.2f]",
796 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
797 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
798 #endif
799 
800 	printf("\n");
801 
802 	if (sched->nr_sleep_corrections)
803 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
804 	sched->nr_sleep_corrections = 0;
805 }
806 
807 static void test_calibrations(struct perf_sched *sched)
808 {
809 	u64 T0, T1;
810 
811 	T0 = get_nsecs();
812 	burn_nsecs(sched, NSEC_PER_MSEC);
813 	T1 = get_nsecs();
814 
815 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
816 
817 	T0 = get_nsecs();
818 	sleep_nsecs(NSEC_PER_MSEC);
819 	T1 = get_nsecs();
820 
821 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
822 }
823 
824 static int
825 replay_wakeup_event(struct perf_sched *sched,
826 		    struct evsel *evsel, struct perf_sample *sample,
827 		    struct machine *machine __maybe_unused)
828 {
829 	const char *comm = evsel__strval(evsel, sample, "comm");
830 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
831 	struct task_desc *waker, *wakee;
832 
833 	if (verbose > 0) {
834 		printf("sched_wakeup event %p\n", evsel);
835 
836 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
837 	}
838 
839 	waker = register_pid(sched, sample->tid, "<unknown>");
840 	wakee = register_pid(sched, pid, comm);
841 
842 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
843 	return 0;
844 }
845 
846 static int replay_switch_event(struct perf_sched *sched,
847 			       struct evsel *evsel,
848 			       struct perf_sample *sample,
849 			       struct machine *machine __maybe_unused)
850 {
851 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
852 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
853 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
854 		  next_pid = evsel__intval(evsel, sample, "next_pid");
855 	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
856 	struct task_desc *prev, __maybe_unused *next;
857 	u64 timestamp0, timestamp = sample->time;
858 	int cpu = sample->cpu;
859 	s64 delta;
860 
861 	if (verbose > 0)
862 		printf("sched_switch event %p\n", evsel);
863 
864 	if (cpu >= MAX_CPUS || cpu < 0)
865 		return 0;
866 
867 	timestamp0 = sched->cpu_last_switched[cpu];
868 	if (timestamp0)
869 		delta = timestamp - timestamp0;
870 	else
871 		delta = 0;
872 
873 	if (delta < 0) {
874 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
875 		return -1;
876 	}
877 
878 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
879 		 prev_comm, prev_pid, next_comm, next_pid, delta);
880 
881 	prev = register_pid(sched, prev_pid, prev_comm);
882 	next = register_pid(sched, next_pid, next_comm);
883 
884 	sched->cpu_last_switched[cpu] = timestamp;
885 
886 	add_sched_event_run(sched, prev, timestamp, delta);
887 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
888 
889 	return 0;
890 }
891 
892 static int replay_fork_event(struct perf_sched *sched,
893 			     union perf_event *event,
894 			     struct machine *machine)
895 {
896 	struct thread *child, *parent;
897 
898 	child = machine__findnew_thread(machine, event->fork.pid,
899 					event->fork.tid);
900 	parent = machine__findnew_thread(machine, event->fork.ppid,
901 					 event->fork.ptid);
902 
903 	if (child == NULL || parent == NULL) {
904 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
905 				 child, parent);
906 		goto out_put;
907 	}
908 
909 	if (verbose > 0) {
910 		printf("fork event\n");
911 		printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
912 		printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
913 	}
914 
915 	register_pid(sched, thread__tid(parent), thread__comm_str(parent));
916 	register_pid(sched, thread__tid(child), thread__comm_str(child));
917 out_put:
918 	thread__put(child);
919 	thread__put(parent);
920 	return 0;
921 }
922 
923 struct sort_dimension {
924 	const char		*name;
925 	sort_fn_t		cmp;
926 	struct list_head	list;
927 };
928 
929 static inline void init_prio(struct thread_runtime *r)
930 {
931 	r->prio = -1;
932 }
933 
934 /*
935  * handle runtime stats saved per thread
936  */
937 static struct thread_runtime *thread__init_runtime(struct thread *thread)
938 {
939 	struct thread_runtime *r;
940 
941 	r = zalloc(sizeof(struct thread_runtime));
942 	if (!r)
943 		return NULL;
944 
945 	init_stats(&r->run_stats);
946 	init_prio(r);
947 	thread__set_priv(thread, r);
948 
949 	return r;
950 }
951 
952 static struct thread_runtime *thread__get_runtime(struct thread *thread)
953 {
954 	struct thread_runtime *tr;
955 
956 	tr = thread__priv(thread);
957 	if (tr == NULL) {
958 		tr = thread__init_runtime(thread);
959 		if (tr == NULL)
960 			pr_debug("Failed to malloc memory for runtime data.\n");
961 	}
962 
963 	return tr;
964 }
965 
966 static int
967 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
968 {
969 	struct sort_dimension *sort;
970 	int ret = 0;
971 
972 	BUG_ON(list_empty(list));
973 
974 	list_for_each_entry(sort, list, list) {
975 		ret = sort->cmp(l, r);
976 		if (ret)
977 			return ret;
978 	}
979 
980 	return ret;
981 }
982 
983 static struct work_atoms *
984 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
985 			 struct list_head *sort_list)
986 {
987 	struct rb_node *node = root->rb_root.rb_node;
988 	struct work_atoms key = { .thread = thread };
989 
990 	while (node) {
991 		struct work_atoms *atoms;
992 		int cmp;
993 
994 		atoms = container_of(node, struct work_atoms, node);
995 
996 		cmp = thread_lat_cmp(sort_list, &key, atoms);
997 		if (cmp > 0)
998 			node = node->rb_left;
999 		else if (cmp < 0)
1000 			node = node->rb_right;
1001 		else {
1002 			BUG_ON(thread != atoms->thread);
1003 			return atoms;
1004 		}
1005 	}
1006 	return NULL;
1007 }
1008 
1009 static void
1010 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1011 			 struct list_head *sort_list)
1012 {
1013 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1014 	bool leftmost = true;
1015 
1016 	while (*new) {
1017 		struct work_atoms *this;
1018 		int cmp;
1019 
1020 		this = container_of(*new, struct work_atoms, node);
1021 		parent = *new;
1022 
1023 		cmp = thread_lat_cmp(sort_list, data, this);
1024 
1025 		if (cmp > 0)
1026 			new = &((*new)->rb_left);
1027 		else {
1028 			new = &((*new)->rb_right);
1029 			leftmost = false;
1030 		}
1031 	}
1032 
1033 	rb_link_node(&data->node, parent, new);
1034 	rb_insert_color_cached(&data->node, root, leftmost);
1035 }
1036 
1037 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1038 {
1039 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1040 	if (!atoms) {
1041 		pr_err("No memory at %s\n", __func__);
1042 		return -1;
1043 	}
1044 
1045 	atoms->thread = thread__get(thread);
1046 	INIT_LIST_HEAD(&atoms->work_list);
1047 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1048 	return 0;
1049 }
1050 
1051 static int
1052 add_sched_out_event(struct work_atoms *atoms,
1053 		    char run_state,
1054 		    u64 timestamp)
1055 {
1056 	struct work_atom *atom = zalloc(sizeof(*atom));
1057 	if (!atom) {
1058 		pr_err("Non memory at %s", __func__);
1059 		return -1;
1060 	}
1061 
1062 	atom->sched_out_time = timestamp;
1063 
1064 	if (run_state == 'R') {
1065 		atom->state = THREAD_WAIT_CPU;
1066 		atom->wake_up_time = atom->sched_out_time;
1067 	}
1068 
1069 	list_add_tail(&atom->list, &atoms->work_list);
1070 	return 0;
1071 }
1072 
1073 static void
1074 add_runtime_event(struct work_atoms *atoms, u64 delta,
1075 		  u64 timestamp __maybe_unused)
1076 {
1077 	struct work_atom *atom;
1078 
1079 	BUG_ON(list_empty(&atoms->work_list));
1080 
1081 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1082 
1083 	atom->runtime += delta;
1084 	atoms->total_runtime += delta;
1085 }
1086 
1087 static void
1088 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1089 {
1090 	struct work_atom *atom;
1091 	u64 delta;
1092 
1093 	if (list_empty(&atoms->work_list))
1094 		return;
1095 
1096 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1097 
1098 	if (atom->state != THREAD_WAIT_CPU)
1099 		return;
1100 
1101 	if (timestamp < atom->wake_up_time) {
1102 		atom->state = THREAD_IGNORE;
1103 		return;
1104 	}
1105 
1106 	atom->state = THREAD_SCHED_IN;
1107 	atom->sched_in_time = timestamp;
1108 
1109 	delta = atom->sched_in_time - atom->wake_up_time;
1110 	atoms->total_lat += delta;
1111 	if (delta > atoms->max_lat) {
1112 		atoms->max_lat = delta;
1113 		atoms->max_lat_start = atom->wake_up_time;
1114 		atoms->max_lat_end = timestamp;
1115 	}
1116 	atoms->nb_atoms++;
1117 }
1118 
1119 static int latency_switch_event(struct perf_sched *sched,
1120 				struct evsel *evsel,
1121 				struct perf_sample *sample,
1122 				struct machine *machine)
1123 {
1124 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1125 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1126 	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1127 	struct work_atoms *out_events, *in_events;
1128 	struct thread *sched_out, *sched_in;
1129 	u64 timestamp0, timestamp = sample->time;
1130 	int cpu = sample->cpu, err = -1;
1131 	s64 delta;
1132 
1133 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1134 
1135 	timestamp0 = sched->cpu_last_switched[cpu];
1136 	sched->cpu_last_switched[cpu] = timestamp;
1137 	if (timestamp0)
1138 		delta = timestamp - timestamp0;
1139 	else
1140 		delta = 0;
1141 
1142 	if (delta < 0) {
1143 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1144 		return -1;
1145 	}
1146 
1147 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1148 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1149 	if (sched_out == NULL || sched_in == NULL)
1150 		goto out_put;
1151 
1152 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1153 	if (!out_events) {
1154 		if (thread_atoms_insert(sched, sched_out))
1155 			goto out_put;
1156 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1157 		if (!out_events) {
1158 			pr_err("out-event: Internal tree error");
1159 			goto out_put;
1160 		}
1161 	}
1162 	if (add_sched_out_event(out_events, prev_state, timestamp))
1163 		return -1;
1164 
1165 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1166 	if (!in_events) {
1167 		if (thread_atoms_insert(sched, sched_in))
1168 			goto out_put;
1169 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1170 		if (!in_events) {
1171 			pr_err("in-event: Internal tree error");
1172 			goto out_put;
1173 		}
1174 		/*
1175 		 * Take came in we have not heard about yet,
1176 		 * add in an initial atom in runnable state:
1177 		 */
1178 		if (add_sched_out_event(in_events, 'R', timestamp))
1179 			goto out_put;
1180 	}
1181 	add_sched_in_event(in_events, timestamp);
1182 	err = 0;
1183 out_put:
1184 	thread__put(sched_out);
1185 	thread__put(sched_in);
1186 	return err;
1187 }
1188 
1189 static int latency_runtime_event(struct perf_sched *sched,
1190 				 struct evsel *evsel,
1191 				 struct perf_sample *sample,
1192 				 struct machine *machine)
1193 {
1194 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1195 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1196 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1197 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1198 	u64 timestamp = sample->time;
1199 	int cpu = sample->cpu, err = -1;
1200 
1201 	if (thread == NULL)
1202 		return -1;
1203 
1204 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1205 	if (!atoms) {
1206 		if (thread_atoms_insert(sched, thread))
1207 			goto out_put;
1208 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1209 		if (!atoms) {
1210 			pr_err("in-event: Internal tree error");
1211 			goto out_put;
1212 		}
1213 		if (add_sched_out_event(atoms, 'R', timestamp))
1214 			goto out_put;
1215 	}
1216 
1217 	add_runtime_event(atoms, runtime, timestamp);
1218 	err = 0;
1219 out_put:
1220 	thread__put(thread);
1221 	return err;
1222 }
1223 
1224 static int latency_wakeup_event(struct perf_sched *sched,
1225 				struct evsel *evsel,
1226 				struct perf_sample *sample,
1227 				struct machine *machine)
1228 {
1229 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1230 	struct work_atoms *atoms;
1231 	struct work_atom *atom;
1232 	struct thread *wakee;
1233 	u64 timestamp = sample->time;
1234 	int err = -1;
1235 
1236 	wakee = machine__findnew_thread(machine, -1, pid);
1237 	if (wakee == NULL)
1238 		return -1;
1239 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1240 	if (!atoms) {
1241 		if (thread_atoms_insert(sched, wakee))
1242 			goto out_put;
1243 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1244 		if (!atoms) {
1245 			pr_err("wakeup-event: Internal tree error");
1246 			goto out_put;
1247 		}
1248 		if (add_sched_out_event(atoms, 'S', timestamp))
1249 			goto out_put;
1250 	}
1251 
1252 	BUG_ON(list_empty(&atoms->work_list));
1253 
1254 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1255 
1256 	/*
1257 	 * As we do not guarantee the wakeup event happens when
1258 	 * task is out of run queue, also may happen when task is
1259 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1260 	 * then we should not set the ->wake_up_time when wake up a
1261 	 * task which is on run queue.
1262 	 *
1263 	 * You WILL be missing events if you've recorded only
1264 	 * one CPU, or are only looking at only one, so don't
1265 	 * skip in this case.
1266 	 */
1267 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1268 		goto out_ok;
1269 
1270 	sched->nr_timestamps++;
1271 	if (atom->sched_out_time > timestamp) {
1272 		sched->nr_unordered_timestamps++;
1273 		goto out_ok;
1274 	}
1275 
1276 	atom->state = THREAD_WAIT_CPU;
1277 	atom->wake_up_time = timestamp;
1278 out_ok:
1279 	err = 0;
1280 out_put:
1281 	thread__put(wakee);
1282 	return err;
1283 }
1284 
1285 static int latency_migrate_task_event(struct perf_sched *sched,
1286 				      struct evsel *evsel,
1287 				      struct perf_sample *sample,
1288 				      struct machine *machine)
1289 {
1290 	const u32 pid = evsel__intval(evsel, sample, "pid");
1291 	u64 timestamp = sample->time;
1292 	struct work_atoms *atoms;
1293 	struct work_atom *atom;
1294 	struct thread *migrant;
1295 	int err = -1;
1296 
1297 	/*
1298 	 * Only need to worry about migration when profiling one CPU.
1299 	 */
1300 	if (sched->profile_cpu == -1)
1301 		return 0;
1302 
1303 	migrant = machine__findnew_thread(machine, -1, pid);
1304 	if (migrant == NULL)
1305 		return -1;
1306 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1307 	if (!atoms) {
1308 		if (thread_atoms_insert(sched, migrant))
1309 			goto out_put;
1310 		register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1311 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1312 		if (!atoms) {
1313 			pr_err("migration-event: Internal tree error");
1314 			goto out_put;
1315 		}
1316 		if (add_sched_out_event(atoms, 'R', timestamp))
1317 			goto out_put;
1318 	}
1319 
1320 	BUG_ON(list_empty(&atoms->work_list));
1321 
1322 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1323 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1324 
1325 	sched->nr_timestamps++;
1326 
1327 	if (atom->sched_out_time > timestamp)
1328 		sched->nr_unordered_timestamps++;
1329 	err = 0;
1330 out_put:
1331 	thread__put(migrant);
1332 	return err;
1333 }
1334 
1335 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1336 {
1337 	int i;
1338 	int ret;
1339 	u64 avg;
1340 	char max_lat_start[32], max_lat_end[32];
1341 
1342 	if (!work_list->nb_atoms)
1343 		return;
1344 	/*
1345 	 * Ignore idle threads:
1346 	 */
1347 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1348 		return;
1349 
1350 	sched->all_runtime += work_list->total_runtime;
1351 	sched->all_count   += work_list->nb_atoms;
1352 
1353 	if (work_list->num_merged > 1) {
1354 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1355 			     work_list->num_merged);
1356 	} else {
1357 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1358 			     thread__tid(work_list->thread));
1359 	}
1360 
1361 	for (i = 0; i < 24 - ret; i++)
1362 		printf(" ");
1363 
1364 	avg = work_list->total_lat / work_list->nb_atoms;
1365 	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1366 	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1367 
1368 	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1369 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1370 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1371 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1372 		 max_lat_start, max_lat_end);
1373 }
1374 
1375 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1376 {
1377 	pid_t l_tid, r_tid;
1378 
1379 	if (RC_CHK_EQUAL(l->thread, r->thread))
1380 		return 0;
1381 	l_tid = thread__tid(l->thread);
1382 	r_tid = thread__tid(r->thread);
1383 	if (l_tid < r_tid)
1384 		return -1;
1385 	if (l_tid > r_tid)
1386 		return 1;
1387 	return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1388 }
1389 
1390 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1391 {
1392 	u64 avgl, avgr;
1393 
1394 	if (!l->nb_atoms)
1395 		return -1;
1396 
1397 	if (!r->nb_atoms)
1398 		return 1;
1399 
1400 	avgl = l->total_lat / l->nb_atoms;
1401 	avgr = r->total_lat / r->nb_atoms;
1402 
1403 	if (avgl < avgr)
1404 		return -1;
1405 	if (avgl > avgr)
1406 		return 1;
1407 
1408 	return 0;
1409 }
1410 
1411 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1412 {
1413 	if (l->max_lat < r->max_lat)
1414 		return -1;
1415 	if (l->max_lat > r->max_lat)
1416 		return 1;
1417 
1418 	return 0;
1419 }
1420 
1421 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1422 {
1423 	if (l->nb_atoms < r->nb_atoms)
1424 		return -1;
1425 	if (l->nb_atoms > r->nb_atoms)
1426 		return 1;
1427 
1428 	return 0;
1429 }
1430 
1431 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1432 {
1433 	if (l->total_runtime < r->total_runtime)
1434 		return -1;
1435 	if (l->total_runtime > r->total_runtime)
1436 		return 1;
1437 
1438 	return 0;
1439 }
1440 
1441 static int sort_dimension__add(const char *tok, struct list_head *list)
1442 {
1443 	size_t i;
1444 	static struct sort_dimension avg_sort_dimension = {
1445 		.name = "avg",
1446 		.cmp  = avg_cmp,
1447 	};
1448 	static struct sort_dimension max_sort_dimension = {
1449 		.name = "max",
1450 		.cmp  = max_cmp,
1451 	};
1452 	static struct sort_dimension pid_sort_dimension = {
1453 		.name = "pid",
1454 		.cmp  = pid_cmp,
1455 	};
1456 	static struct sort_dimension runtime_sort_dimension = {
1457 		.name = "runtime",
1458 		.cmp  = runtime_cmp,
1459 	};
1460 	static struct sort_dimension switch_sort_dimension = {
1461 		.name = "switch",
1462 		.cmp  = switch_cmp,
1463 	};
1464 	struct sort_dimension *available_sorts[] = {
1465 		&pid_sort_dimension,
1466 		&avg_sort_dimension,
1467 		&max_sort_dimension,
1468 		&switch_sort_dimension,
1469 		&runtime_sort_dimension,
1470 	};
1471 
1472 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1473 		if (!strcmp(available_sorts[i]->name, tok)) {
1474 			list_add_tail(&available_sorts[i]->list, list);
1475 
1476 			return 0;
1477 		}
1478 	}
1479 
1480 	return -1;
1481 }
1482 
1483 static void perf_sched__sort_lat(struct perf_sched *sched)
1484 {
1485 	struct rb_node *node;
1486 	struct rb_root_cached *root = &sched->atom_root;
1487 again:
1488 	for (;;) {
1489 		struct work_atoms *data;
1490 		node = rb_first_cached(root);
1491 		if (!node)
1492 			break;
1493 
1494 		rb_erase_cached(node, root);
1495 		data = rb_entry(node, struct work_atoms, node);
1496 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1497 	}
1498 	if (root == &sched->atom_root) {
1499 		root = &sched->merged_atom_root;
1500 		goto again;
1501 	}
1502 }
1503 
1504 static int process_sched_wakeup_event(const struct perf_tool *tool,
1505 				      struct evsel *evsel,
1506 				      struct perf_sample *sample,
1507 				      struct machine *machine)
1508 {
1509 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1510 
1511 	if (sched->tp_handler->wakeup_event)
1512 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1513 
1514 	return 0;
1515 }
1516 
1517 static int process_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
1518 				      struct evsel *evsel __maybe_unused,
1519 				      struct perf_sample *sample __maybe_unused,
1520 				      struct machine *machine __maybe_unused)
1521 {
1522 	return 0;
1523 }
1524 
1525 union map_priv {
1526 	void	*ptr;
1527 	bool	 color;
1528 };
1529 
1530 static bool thread__has_color(struct thread *thread)
1531 {
1532 	union map_priv priv = {
1533 		.ptr = thread__priv(thread),
1534 	};
1535 
1536 	return priv.color;
1537 }
1538 
1539 static struct thread*
1540 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1541 {
1542 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1543 	union map_priv priv = {
1544 		.color = false,
1545 	};
1546 
1547 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1548 		return thread;
1549 
1550 	if (thread_map__has(sched->map.color_pids, tid))
1551 		priv.color = true;
1552 
1553 	thread__set_priv(thread, priv.ptr);
1554 	return thread;
1555 }
1556 
1557 static bool sched_match_task(struct perf_sched *sched, const char *comm_str)
1558 {
1559 	bool fuzzy_match = sched->map.fuzzy;
1560 	struct strlist *task_names = sched->map.task_names;
1561 	struct str_node *node;
1562 
1563 	strlist__for_each_entry(node, task_names) {
1564 		bool match_found = fuzzy_match ? !!strstr(comm_str, node->s) :
1565 							!strcmp(comm_str, node->s);
1566 		if (match_found)
1567 			return true;
1568 	}
1569 
1570 	return false;
1571 }
1572 
1573 static void print_sched_map(struct perf_sched *sched, struct perf_cpu this_cpu, int cpus_nr,
1574 								const char *color, bool sched_out)
1575 {
1576 	for (int i = 0; i < cpus_nr; i++) {
1577 		struct perf_cpu cpu = {
1578 			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1579 		};
1580 		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1581 		struct thread *curr_out_thread = sched->curr_out_thread[cpu.cpu];
1582 		struct thread_runtime *curr_tr;
1583 		const char *pid_color = color;
1584 		const char *cpu_color = color;
1585 		char symbol = ' ';
1586 		struct thread *thread_to_check = sched_out ? curr_out_thread : curr_thread;
1587 
1588 		if (thread_to_check && thread__has_color(thread_to_check))
1589 			pid_color = COLOR_PIDS;
1590 
1591 		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1592 			cpu_color = COLOR_CPUS;
1593 
1594 		if (cpu.cpu == this_cpu.cpu)
1595 			symbol = '*';
1596 
1597 		color_fprintf(stdout, cpu.cpu != this_cpu.cpu ? color : cpu_color, "%c", symbol);
1598 
1599 		thread_to_check = sched_out ? sched->curr_out_thread[cpu.cpu] :
1600 								sched->curr_thread[cpu.cpu];
1601 
1602 		if (thread_to_check) {
1603 			curr_tr = thread__get_runtime(thread_to_check);
1604 			if (curr_tr == NULL)
1605 				return;
1606 
1607 			if (sched_out) {
1608 				if (cpu.cpu == this_cpu.cpu)
1609 					color_fprintf(stdout, color, "-  ");
1610 				else {
1611 					curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1612 					if (curr_tr != NULL)
1613 						color_fprintf(stdout, pid_color, "%2s ",
1614 										curr_tr->shortname);
1615 				}
1616 			} else
1617 				color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1618 		} else
1619 			color_fprintf(stdout, color, "   ");
1620 	}
1621 }
1622 
1623 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1624 			    struct perf_sample *sample, struct machine *machine)
1625 {
1626 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1627 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid");
1628 	struct thread *sched_in, *sched_out;
1629 	struct thread_runtime *tr;
1630 	int new_shortname;
1631 	u64 timestamp0, timestamp = sample->time;
1632 	s64 delta;
1633 	struct perf_cpu this_cpu = {
1634 		.cpu = sample->cpu,
1635 	};
1636 	int cpus_nr;
1637 	int proceed;
1638 	bool new_cpu = false;
1639 	const char *color = PERF_COLOR_NORMAL;
1640 	char stimestamp[32];
1641 	const char *str;
1642 
1643 	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1644 
1645 	if (this_cpu.cpu > sched->max_cpu.cpu)
1646 		sched->max_cpu = this_cpu;
1647 
1648 	if (sched->map.comp) {
1649 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1650 		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1651 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1652 			new_cpu = true;
1653 		}
1654 	} else
1655 		cpus_nr = sched->max_cpu.cpu;
1656 
1657 	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1658 	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1659 	if (timestamp0)
1660 		delta = timestamp - timestamp0;
1661 	else
1662 		delta = 0;
1663 
1664 	if (delta < 0) {
1665 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1666 		return -1;
1667 	}
1668 
1669 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1670 	sched_out = map__findnew_thread(sched, machine, -1, prev_pid);
1671 	if (sched_in == NULL || sched_out == NULL)
1672 		return -1;
1673 
1674 	tr = thread__get_runtime(sched_in);
1675 	if (tr == NULL) {
1676 		thread__put(sched_in);
1677 		return -1;
1678 	}
1679 
1680 	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1681 	sched->curr_out_thread[this_cpu.cpu] = thread__get(sched_out);
1682 
1683 	str = thread__comm_str(sched_in);
1684 	new_shortname = 0;
1685 	if (!tr->shortname[0]) {
1686 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1687 			/*
1688 			 * Don't allocate a letter-number for swapper:0
1689 			 * as a shortname. Instead, we use '.' for it.
1690 			 */
1691 			tr->shortname[0] = '.';
1692 			tr->shortname[1] = ' ';
1693 		} else if (!sched->map.task_name || sched_match_task(sched, str)) {
1694 			tr->shortname[0] = sched->next_shortname1;
1695 			tr->shortname[1] = sched->next_shortname2;
1696 
1697 			if (sched->next_shortname1 < 'Z') {
1698 				sched->next_shortname1++;
1699 			} else {
1700 				sched->next_shortname1 = 'A';
1701 				if (sched->next_shortname2 < '9')
1702 					sched->next_shortname2++;
1703 				else
1704 					sched->next_shortname2 = '0';
1705 			}
1706 		} else {
1707 			tr->shortname[0] = '-';
1708 			tr->shortname[1] = ' ';
1709 		}
1710 		new_shortname = 1;
1711 	}
1712 
1713 	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1714 		goto out;
1715 
1716 	proceed = 0;
1717 	str = thread__comm_str(sched_in);
1718 	/*
1719 	 * Check which of sched_in and sched_out matches the passed --task-name
1720 	 * arguments and call the corresponding print_sched_map.
1721 	 */
1722 	if (sched->map.task_name && !sched_match_task(sched, str)) {
1723 		if (!sched_match_task(sched, thread__comm_str(sched_out)))
1724 			goto out;
1725 		else
1726 			goto sched_out;
1727 
1728 	} else {
1729 		str = thread__comm_str(sched_out);
1730 		if (!(sched->map.task_name && !sched_match_task(sched, str)))
1731 			proceed = 1;
1732 	}
1733 
1734 	printf("  ");
1735 
1736 	print_sched_map(sched, this_cpu, cpus_nr, color, false);
1737 
1738 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1739 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1740 	if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1741 		const char *pid_color = color;
1742 
1743 		if (thread__has_color(sched_in))
1744 			pid_color = COLOR_PIDS;
1745 
1746 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1747 			tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1748 		tr->comm_changed = false;
1749 	}
1750 
1751 	if (sched->map.comp && new_cpu)
1752 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1753 
1754 	if (proceed != 1) {
1755 		color_fprintf(stdout, color, "\n");
1756 		goto out;
1757 	}
1758 
1759 sched_out:
1760 	if (sched->map.task_name) {
1761 		tr = thread__get_runtime(sched->curr_out_thread[this_cpu.cpu]);
1762 		if (strcmp(tr->shortname, "") == 0)
1763 			goto out;
1764 
1765 		if (proceed == 1)
1766 			color_fprintf(stdout, color, "\n");
1767 
1768 		printf("  ");
1769 		print_sched_map(sched, this_cpu, cpus_nr, color, true);
1770 		timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1771 		color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1772 	}
1773 
1774 	color_fprintf(stdout, color, "\n");
1775 
1776 out:
1777 	if (sched->map.task_name)
1778 		thread__put(sched_out);
1779 
1780 	thread__put(sched_in);
1781 
1782 	return 0;
1783 }
1784 
1785 static int process_sched_switch_event(const struct perf_tool *tool,
1786 				      struct evsel *evsel,
1787 				      struct perf_sample *sample,
1788 				      struct machine *machine)
1789 {
1790 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1791 	int this_cpu = sample->cpu, err = 0;
1792 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1793 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1794 
1795 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1796 		/*
1797 		 * Are we trying to switch away a PID that is
1798 		 * not current?
1799 		 */
1800 		if (sched->curr_pid[this_cpu] != prev_pid)
1801 			sched->nr_context_switch_bugs++;
1802 	}
1803 
1804 	if (sched->tp_handler->switch_event)
1805 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1806 
1807 	sched->curr_pid[this_cpu] = next_pid;
1808 	return err;
1809 }
1810 
1811 static int process_sched_runtime_event(const struct perf_tool *tool,
1812 				       struct evsel *evsel,
1813 				       struct perf_sample *sample,
1814 				       struct machine *machine)
1815 {
1816 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1817 
1818 	if (sched->tp_handler->runtime_event)
1819 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1820 
1821 	return 0;
1822 }
1823 
1824 static int perf_sched__process_fork_event(const struct perf_tool *tool,
1825 					  union perf_event *event,
1826 					  struct perf_sample *sample,
1827 					  struct machine *machine)
1828 {
1829 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1830 
1831 	/* run the fork event through the perf machinery */
1832 	perf_event__process_fork(tool, event, sample, machine);
1833 
1834 	/* and then run additional processing needed for this command */
1835 	if (sched->tp_handler->fork_event)
1836 		return sched->tp_handler->fork_event(sched, event, machine);
1837 
1838 	return 0;
1839 }
1840 
1841 static int process_sched_migrate_task_event(const struct perf_tool *tool,
1842 					    struct evsel *evsel,
1843 					    struct perf_sample *sample,
1844 					    struct machine *machine)
1845 {
1846 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1847 
1848 	if (sched->tp_handler->migrate_task_event)
1849 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1850 
1851 	return 0;
1852 }
1853 
1854 typedef int (*tracepoint_handler)(const struct perf_tool *tool,
1855 				  struct evsel *evsel,
1856 				  struct perf_sample *sample,
1857 				  struct machine *machine);
1858 
1859 static int perf_sched__process_tracepoint_sample(const struct perf_tool *tool __maybe_unused,
1860 						 union perf_event *event __maybe_unused,
1861 						 struct perf_sample *sample,
1862 						 struct evsel *evsel,
1863 						 struct machine *machine)
1864 {
1865 	int err = 0;
1866 
1867 	if (evsel->handler != NULL) {
1868 		tracepoint_handler f = evsel->handler;
1869 		err = f(tool, evsel, sample, machine);
1870 	}
1871 
1872 	return err;
1873 }
1874 
1875 static int perf_sched__process_comm(const struct perf_tool *tool __maybe_unused,
1876 				    union perf_event *event,
1877 				    struct perf_sample *sample,
1878 				    struct machine *machine)
1879 {
1880 	struct thread *thread;
1881 	struct thread_runtime *tr;
1882 	int err;
1883 
1884 	err = perf_event__process_comm(tool, event, sample, machine);
1885 	if (err)
1886 		return err;
1887 
1888 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1889 	if (!thread) {
1890 		pr_err("Internal error: can't find thread\n");
1891 		return -1;
1892 	}
1893 
1894 	tr = thread__get_runtime(thread);
1895 	if (tr == NULL) {
1896 		thread__put(thread);
1897 		return -1;
1898 	}
1899 
1900 	tr->comm_changed = true;
1901 	thread__put(thread);
1902 
1903 	return 0;
1904 }
1905 
1906 static int perf_sched__read_events(struct perf_sched *sched)
1907 {
1908 	struct evsel_str_handler handlers[] = {
1909 		{ "sched:sched_switch",	      process_sched_switch_event, },
1910 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1911 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1912 		{ "sched:sched_waking",	      process_sched_wakeup_event, },
1913 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1914 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1915 	};
1916 	struct perf_session *session;
1917 	struct perf_data data = {
1918 		.path  = input_name,
1919 		.mode  = PERF_DATA_MODE_READ,
1920 		.force = sched->force,
1921 	};
1922 	int rc = -1;
1923 
1924 	session = perf_session__new(&data, &sched->tool);
1925 	if (IS_ERR(session)) {
1926 		pr_debug("Error creating perf session");
1927 		return PTR_ERR(session);
1928 	}
1929 
1930 	symbol__init(&session->header.env);
1931 
1932 	/* prefer sched_waking if it is captured */
1933 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1934 		handlers[2].handler = process_sched_wakeup_ignore;
1935 
1936 	if (perf_session__set_tracepoints_handlers(session, handlers))
1937 		goto out_delete;
1938 
1939 	if (perf_session__has_traces(session, "record -R")) {
1940 		int err = perf_session__process_events(session);
1941 		if (err) {
1942 			pr_err("Failed to process events, error %d", err);
1943 			goto out_delete;
1944 		}
1945 
1946 		sched->nr_events      = session->evlist->stats.nr_events[0];
1947 		sched->nr_lost_events = session->evlist->stats.total_lost;
1948 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1949 	}
1950 
1951 	rc = 0;
1952 out_delete:
1953 	perf_session__delete(session);
1954 	return rc;
1955 }
1956 
1957 /*
1958  * scheduling times are printed as msec.usec
1959  */
1960 static inline void print_sched_time(unsigned long long nsecs, int width)
1961 {
1962 	unsigned long msecs;
1963 	unsigned long usecs;
1964 
1965 	msecs  = nsecs / NSEC_PER_MSEC;
1966 	nsecs -= msecs * NSEC_PER_MSEC;
1967 	usecs  = nsecs / NSEC_PER_USEC;
1968 	printf("%*lu.%03lu ", width, msecs, usecs);
1969 }
1970 
1971 /*
1972  * returns runtime data for event, allocating memory for it the
1973  * first time it is used.
1974  */
1975 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1976 {
1977 	struct evsel_runtime *r = evsel->priv;
1978 
1979 	if (r == NULL) {
1980 		r = zalloc(sizeof(struct evsel_runtime));
1981 		evsel->priv = r;
1982 	}
1983 
1984 	return r;
1985 }
1986 
1987 /*
1988  * save last time event was seen per cpu
1989  */
1990 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1991 {
1992 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1993 
1994 	if (r == NULL)
1995 		return;
1996 
1997 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1998 		int i, n = __roundup_pow_of_two(cpu+1);
1999 		void *p = r->last_time;
2000 
2001 		p = realloc(r->last_time, n * sizeof(u64));
2002 		if (!p)
2003 			return;
2004 
2005 		r->last_time = p;
2006 		for (i = r->ncpu; i < n; ++i)
2007 			r->last_time[i] = (u64) 0;
2008 
2009 		r->ncpu = n;
2010 	}
2011 
2012 	r->last_time[cpu] = timestamp;
2013 }
2014 
2015 /* returns last time this event was seen on the given cpu */
2016 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
2017 {
2018 	struct evsel_runtime *r = evsel__get_runtime(evsel);
2019 
2020 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
2021 		return 0;
2022 
2023 	return r->last_time[cpu];
2024 }
2025 
2026 static int comm_width = 30;
2027 
2028 static char *timehist_get_commstr(struct thread *thread)
2029 {
2030 	static char str[32];
2031 	const char *comm = thread__comm_str(thread);
2032 	pid_t tid = thread__tid(thread);
2033 	pid_t pid = thread__pid(thread);
2034 	int n;
2035 
2036 	if (pid == 0)
2037 		n = scnprintf(str, sizeof(str), "%s", comm);
2038 
2039 	else if (tid != pid)
2040 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
2041 
2042 	else
2043 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
2044 
2045 	if (n > comm_width)
2046 		comm_width = n;
2047 
2048 	return str;
2049 }
2050 
2051 /* prio field format: xxx or xxx->yyy */
2052 #define MAX_PRIO_STR_LEN 8
2053 static char *timehist_get_priostr(struct evsel *evsel,
2054 				  struct thread *thread,
2055 				  struct perf_sample *sample)
2056 {
2057 	static char prio_str[16];
2058 	int prev_prio = (int)evsel__intval(evsel, sample, "prev_prio");
2059 	struct thread_runtime *tr = thread__priv(thread);
2060 
2061 	if (tr->prio != prev_prio && tr->prio != -1)
2062 		scnprintf(prio_str, sizeof(prio_str), "%d->%d", tr->prio, prev_prio);
2063 	else
2064 		scnprintf(prio_str, sizeof(prio_str), "%d", prev_prio);
2065 
2066 	return prio_str;
2067 }
2068 
2069 static void timehist_header(struct perf_sched *sched)
2070 {
2071 	u32 ncpus = sched->max_cpu.cpu + 1;
2072 	u32 i, j;
2073 
2074 	printf("%15s %6s ", "time", "cpu");
2075 
2076 	if (sched->show_cpu_visual) {
2077 		printf(" ");
2078 		for (i = 0, j = 0; i < ncpus; ++i) {
2079 			printf("%x", j++);
2080 			if (j > 15)
2081 				j = 0;
2082 		}
2083 		printf(" ");
2084 	}
2085 
2086 	if (sched->show_prio) {
2087 		printf(" %-*s  %-*s  %9s  %9s  %9s",
2088 		       comm_width, "task name", MAX_PRIO_STR_LEN, "prio",
2089 		       "wait time", "sch delay", "run time");
2090 	} else {
2091 		printf(" %-*s  %9s  %9s  %9s", comm_width,
2092 		       "task name", "wait time", "sch delay", "run time");
2093 	}
2094 
2095 	if (sched->show_state)
2096 		printf("  %s", "state");
2097 
2098 	printf("\n");
2099 
2100 	/*
2101 	 * units row
2102 	 */
2103 	printf("%15s %-6s ", "", "");
2104 
2105 	if (sched->show_cpu_visual)
2106 		printf(" %*s ", ncpus, "");
2107 
2108 	if (sched->show_prio) {
2109 		printf(" %-*s  %-*s  %9s  %9s  %9s",
2110 		       comm_width, "[tid/pid]", MAX_PRIO_STR_LEN, "",
2111 		       "(msec)", "(msec)", "(msec)");
2112 	} else {
2113 		printf(" %-*s  %9s  %9s  %9s", comm_width,
2114 		       "[tid/pid]", "(msec)", "(msec)", "(msec)");
2115 	}
2116 
2117 	if (sched->show_state)
2118 		printf("  %5s", "");
2119 
2120 	printf("\n");
2121 
2122 	/*
2123 	 * separator
2124 	 */
2125 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2126 
2127 	if (sched->show_cpu_visual)
2128 		printf(" %.*s ", ncpus, graph_dotted_line);
2129 
2130 	if (sched->show_prio) {
2131 		printf(" %.*s  %.*s  %.9s  %.9s  %.9s",
2132 		       comm_width, graph_dotted_line, MAX_PRIO_STR_LEN, graph_dotted_line,
2133 		       graph_dotted_line, graph_dotted_line, graph_dotted_line);
2134 	} else {
2135 		printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2136 		       graph_dotted_line, graph_dotted_line, graph_dotted_line,
2137 		       graph_dotted_line);
2138 	}
2139 
2140 	if (sched->show_state)
2141 		printf("  %.5s", graph_dotted_line);
2142 
2143 	printf("\n");
2144 }
2145 
2146 static void timehist_print_sample(struct perf_sched *sched,
2147 				  struct evsel *evsel,
2148 				  struct perf_sample *sample,
2149 				  struct addr_location *al,
2150 				  struct thread *thread,
2151 				  u64 t, const char state)
2152 {
2153 	struct thread_runtime *tr = thread__priv(thread);
2154 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2155 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2156 	u32 max_cpus = sched->max_cpu.cpu + 1;
2157 	char tstr[64];
2158 	char nstr[30];
2159 	u64 wait_time;
2160 
2161 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2162 		return;
2163 
2164 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2165 	printf("%15s [%04d] ", tstr, sample->cpu);
2166 
2167 	if (sched->show_cpu_visual) {
2168 		u32 i;
2169 		char c;
2170 
2171 		printf(" ");
2172 		for (i = 0; i < max_cpus; ++i) {
2173 			/* flag idle times with 'i'; others are sched events */
2174 			if (i == sample->cpu)
2175 				c = (thread__tid(thread) == 0) ? 'i' : 's';
2176 			else
2177 				c = ' ';
2178 			printf("%c", c);
2179 		}
2180 		printf(" ");
2181 	}
2182 
2183 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2184 
2185 	if (sched->show_prio)
2186 		printf(" %-*s ", MAX_PRIO_STR_LEN, timehist_get_priostr(evsel, thread, sample));
2187 
2188 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2189 	print_sched_time(wait_time, 6);
2190 
2191 	print_sched_time(tr->dt_delay, 6);
2192 	print_sched_time(tr->dt_run, 6);
2193 
2194 	if (sched->show_state)
2195 		printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2196 
2197 	if (sched->show_next) {
2198 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2199 		printf(" %-*s", comm_width, nstr);
2200 	}
2201 
2202 	if (sched->show_wakeups && !sched->show_next)
2203 		printf("  %-*s", comm_width, "");
2204 
2205 	if (thread__tid(thread) == 0)
2206 		goto out;
2207 
2208 	if (sched->show_callchain)
2209 		printf("  ");
2210 
2211 	sample__fprintf_sym(sample, al, 0,
2212 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2213 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2214 			    EVSEL__PRINT_SKIP_IGNORED,
2215 			    get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2216 
2217 out:
2218 	printf("\n");
2219 }
2220 
2221 /*
2222  * Explanation of delta-time stats:
2223  *
2224  *            t = time of current schedule out event
2225  *        tprev = time of previous sched out event
2226  *                also time of schedule-in event for current task
2227  *    last_time = time of last sched change event for current task
2228  *                (i.e, time process was last scheduled out)
2229  * ready_to_run = time of wakeup for current task
2230  *
2231  * -----|------------|------------|------------|------
2232  *    last         ready        tprev          t
2233  *    time         to run
2234  *
2235  *      |-------- dt_wait --------|
2236  *                   |- dt_delay -|-- dt_run --|
2237  *
2238  *   dt_run = run time of current task
2239  *  dt_wait = time between last schedule out event for task and tprev
2240  *            represents time spent off the cpu
2241  * dt_delay = time between wakeup and schedule-in of task
2242  */
2243 
2244 static void timehist_update_runtime_stats(struct thread_runtime *r,
2245 					 u64 t, u64 tprev)
2246 {
2247 	r->dt_delay   = 0;
2248 	r->dt_sleep   = 0;
2249 	r->dt_iowait  = 0;
2250 	r->dt_preempt = 0;
2251 	r->dt_run     = 0;
2252 
2253 	if (tprev) {
2254 		r->dt_run = t - tprev;
2255 		if (r->ready_to_run) {
2256 			if (r->ready_to_run > tprev)
2257 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2258 			else
2259 				r->dt_delay = tprev - r->ready_to_run;
2260 		}
2261 
2262 		if (r->last_time > tprev)
2263 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2264 		else if (r->last_time) {
2265 			u64 dt_wait = tprev - r->last_time;
2266 
2267 			if (r->last_state == 'R')
2268 				r->dt_preempt = dt_wait;
2269 			else if (r->last_state == 'D')
2270 				r->dt_iowait = dt_wait;
2271 			else
2272 				r->dt_sleep = dt_wait;
2273 		}
2274 	}
2275 
2276 	update_stats(&r->run_stats, r->dt_run);
2277 
2278 	r->total_run_time     += r->dt_run;
2279 	r->total_delay_time   += r->dt_delay;
2280 	r->total_sleep_time   += r->dt_sleep;
2281 	r->total_iowait_time  += r->dt_iowait;
2282 	r->total_preempt_time += r->dt_preempt;
2283 }
2284 
2285 static bool is_idle_sample(struct perf_sample *sample,
2286 			   struct evsel *evsel)
2287 {
2288 	/* pid 0 == swapper == idle task */
2289 	if (evsel__name_is(evsel, "sched:sched_switch"))
2290 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2291 
2292 	return sample->pid == 0;
2293 }
2294 
2295 static void save_task_callchain(struct perf_sched *sched,
2296 				struct perf_sample *sample,
2297 				struct evsel *evsel,
2298 				struct machine *machine)
2299 {
2300 	struct callchain_cursor *cursor;
2301 	struct thread *thread;
2302 
2303 	/* want main thread for process - has maps */
2304 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2305 	if (thread == NULL) {
2306 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2307 		return;
2308 	}
2309 
2310 	if (!sched->show_callchain || sample->callchain == NULL)
2311 		return;
2312 
2313 	cursor = get_tls_callchain_cursor();
2314 
2315 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2316 				      NULL, NULL, sched->max_stack + 2) != 0) {
2317 		if (verbose > 0)
2318 			pr_err("Failed to resolve callchain. Skipping\n");
2319 
2320 		return;
2321 	}
2322 
2323 	callchain_cursor_commit(cursor);
2324 
2325 	while (true) {
2326 		struct callchain_cursor_node *node;
2327 		struct symbol *sym;
2328 
2329 		node = callchain_cursor_current(cursor);
2330 		if (node == NULL)
2331 			break;
2332 
2333 		sym = node->ms.sym;
2334 		if (sym) {
2335 			if (!strcmp(sym->name, "schedule") ||
2336 			    !strcmp(sym->name, "__schedule") ||
2337 			    !strcmp(sym->name, "preempt_schedule"))
2338 				sym->ignore = 1;
2339 		}
2340 
2341 		callchain_cursor_advance(cursor);
2342 	}
2343 }
2344 
2345 static int init_idle_thread(struct thread *thread)
2346 {
2347 	struct idle_thread_runtime *itr;
2348 
2349 	thread__set_comm(thread, idle_comm, 0);
2350 
2351 	itr = zalloc(sizeof(*itr));
2352 	if (itr == NULL)
2353 		return -ENOMEM;
2354 
2355 	init_prio(&itr->tr);
2356 	init_stats(&itr->tr.run_stats);
2357 	callchain_init(&itr->callchain);
2358 	callchain_cursor_reset(&itr->cursor);
2359 	thread__set_priv(thread, itr);
2360 
2361 	return 0;
2362 }
2363 
2364 /*
2365  * Track idle stats per cpu by maintaining a local thread
2366  * struct for the idle task on each cpu.
2367  */
2368 static int init_idle_threads(int ncpu)
2369 {
2370 	int i, ret;
2371 
2372 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2373 	if (!idle_threads)
2374 		return -ENOMEM;
2375 
2376 	idle_max_cpu = ncpu;
2377 
2378 	/* allocate the actual thread struct if needed */
2379 	for (i = 0; i < ncpu; ++i) {
2380 		idle_threads[i] = thread__new(0, 0);
2381 		if (idle_threads[i] == NULL)
2382 			return -ENOMEM;
2383 
2384 		ret = init_idle_thread(idle_threads[i]);
2385 		if (ret < 0)
2386 			return ret;
2387 	}
2388 
2389 	return 0;
2390 }
2391 
2392 static void free_idle_threads(void)
2393 {
2394 	int i;
2395 
2396 	if (idle_threads == NULL)
2397 		return;
2398 
2399 	for (i = 0; i < idle_max_cpu; ++i) {
2400 		if ((idle_threads[i]))
2401 			thread__delete(idle_threads[i]);
2402 	}
2403 
2404 	free(idle_threads);
2405 }
2406 
2407 static struct thread *get_idle_thread(int cpu)
2408 {
2409 	/*
2410 	 * expand/allocate array of pointers to local thread
2411 	 * structs if needed
2412 	 */
2413 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2414 		int i, j = __roundup_pow_of_two(cpu+1);
2415 		void *p;
2416 
2417 		p = realloc(idle_threads, j * sizeof(struct thread *));
2418 		if (!p)
2419 			return NULL;
2420 
2421 		idle_threads = (struct thread **) p;
2422 		for (i = idle_max_cpu; i < j; ++i)
2423 			idle_threads[i] = NULL;
2424 
2425 		idle_max_cpu = j;
2426 	}
2427 
2428 	/* allocate a new thread struct if needed */
2429 	if (idle_threads[cpu] == NULL) {
2430 		idle_threads[cpu] = thread__new(0, 0);
2431 
2432 		if (idle_threads[cpu]) {
2433 			if (init_idle_thread(idle_threads[cpu]) < 0)
2434 				return NULL;
2435 		}
2436 	}
2437 
2438 	return idle_threads[cpu];
2439 }
2440 
2441 static void save_idle_callchain(struct perf_sched *sched,
2442 				struct idle_thread_runtime *itr,
2443 				struct perf_sample *sample)
2444 {
2445 	struct callchain_cursor *cursor;
2446 
2447 	if (!sched->show_callchain || sample->callchain == NULL)
2448 		return;
2449 
2450 	cursor = get_tls_callchain_cursor();
2451 	if (cursor == NULL)
2452 		return;
2453 
2454 	callchain_cursor__copy(&itr->cursor, cursor);
2455 }
2456 
2457 static struct thread *timehist_get_thread(struct perf_sched *sched,
2458 					  struct perf_sample *sample,
2459 					  struct machine *machine,
2460 					  struct evsel *evsel)
2461 {
2462 	struct thread *thread;
2463 
2464 	if (is_idle_sample(sample, evsel)) {
2465 		thread = get_idle_thread(sample->cpu);
2466 		if (thread == NULL)
2467 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2468 
2469 	} else {
2470 		/* there were samples with tid 0 but non-zero pid */
2471 		thread = machine__findnew_thread(machine, sample->pid,
2472 						 sample->tid ?: sample->pid);
2473 		if (thread == NULL) {
2474 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2475 				 sample->tid);
2476 		}
2477 
2478 		save_task_callchain(sched, sample, evsel, machine);
2479 		if (sched->idle_hist) {
2480 			struct thread *idle;
2481 			struct idle_thread_runtime *itr;
2482 
2483 			idle = get_idle_thread(sample->cpu);
2484 			if (idle == NULL) {
2485 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2486 				return NULL;
2487 			}
2488 
2489 			itr = thread__priv(idle);
2490 			if (itr == NULL)
2491 				return NULL;
2492 
2493 			itr->last_thread = thread;
2494 
2495 			/* copy task callchain when entering to idle */
2496 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2497 				save_idle_callchain(sched, itr, sample);
2498 		}
2499 	}
2500 
2501 	return thread;
2502 }
2503 
2504 static bool timehist_skip_sample(struct perf_sched *sched,
2505 				 struct thread *thread,
2506 				 struct evsel *evsel,
2507 				 struct perf_sample *sample)
2508 {
2509 	bool rc = false;
2510 	int prio = -1;
2511 	struct thread_runtime *tr = NULL;
2512 
2513 	if (thread__is_filtered(thread)) {
2514 		rc = true;
2515 		sched->skipped_samples++;
2516 	}
2517 
2518 	if (sched->prio_str) {
2519 		/*
2520 		 * Because priority may be changed during task execution,
2521 		 * first read priority from prev sched_in event for current task.
2522 		 * If prev sched_in event is not saved, then read priority from
2523 		 * current task sched_out event.
2524 		 */
2525 		tr = thread__get_runtime(thread);
2526 		if (tr && tr->prio != -1)
2527 			prio = tr->prio;
2528 		else if (evsel__name_is(evsel, "sched:sched_switch"))
2529 			prio = evsel__intval(evsel, sample, "prev_prio");
2530 
2531 		if (prio != -1 && !test_bit(prio, sched->prio_bitmap)) {
2532 			rc = true;
2533 			sched->skipped_samples++;
2534 		}
2535 	}
2536 
2537 	if (sched->idle_hist) {
2538 		if (!evsel__name_is(evsel, "sched:sched_switch"))
2539 			rc = true;
2540 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2541 			 evsel__intval(evsel, sample, "next_pid") != 0)
2542 			rc = true;
2543 	}
2544 
2545 	return rc;
2546 }
2547 
2548 static void timehist_print_wakeup_event(struct perf_sched *sched,
2549 					struct evsel *evsel,
2550 					struct perf_sample *sample,
2551 					struct machine *machine,
2552 					struct thread *awakened)
2553 {
2554 	struct thread *thread;
2555 	char tstr[64];
2556 
2557 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2558 	if (thread == NULL)
2559 		return;
2560 
2561 	/* show wakeup unless both awakee and awaker are filtered */
2562 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2563 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2564 		return;
2565 	}
2566 
2567 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2568 	printf("%15s [%04d] ", tstr, sample->cpu);
2569 	if (sched->show_cpu_visual)
2570 		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2571 
2572 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2573 
2574 	/* dt spacer */
2575 	printf("  %9s  %9s  %9s ", "", "", "");
2576 
2577 	printf("awakened: %s", timehist_get_commstr(awakened));
2578 
2579 	printf("\n");
2580 }
2581 
2582 static int timehist_sched_wakeup_ignore(const struct perf_tool *tool __maybe_unused,
2583 					union perf_event *event __maybe_unused,
2584 					struct evsel *evsel __maybe_unused,
2585 					struct perf_sample *sample __maybe_unused,
2586 					struct machine *machine __maybe_unused)
2587 {
2588 	return 0;
2589 }
2590 
2591 static int timehist_sched_wakeup_event(const struct perf_tool *tool,
2592 				       union perf_event *event __maybe_unused,
2593 				       struct evsel *evsel,
2594 				       struct perf_sample *sample,
2595 				       struct machine *machine)
2596 {
2597 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2598 	struct thread *thread;
2599 	struct thread_runtime *tr = NULL;
2600 	/* want pid of awakened task not pid in sample */
2601 	const u32 pid = evsel__intval(evsel, sample, "pid");
2602 
2603 	thread = machine__findnew_thread(machine, 0, pid);
2604 	if (thread == NULL)
2605 		return -1;
2606 
2607 	tr = thread__get_runtime(thread);
2608 	if (tr == NULL)
2609 		return -1;
2610 
2611 	if (tr->ready_to_run == 0)
2612 		tr->ready_to_run = sample->time;
2613 
2614 	/* show wakeups if requested */
2615 	if (sched->show_wakeups &&
2616 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2617 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2618 
2619 	return 0;
2620 }
2621 
2622 static void timehist_print_migration_event(struct perf_sched *sched,
2623 					struct evsel *evsel,
2624 					struct perf_sample *sample,
2625 					struct machine *machine,
2626 					struct thread *migrated)
2627 {
2628 	struct thread *thread;
2629 	char tstr[64];
2630 	u32 max_cpus;
2631 	u32 ocpu, dcpu;
2632 
2633 	if (sched->summary_only)
2634 		return;
2635 
2636 	max_cpus = sched->max_cpu.cpu + 1;
2637 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2638 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2639 
2640 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2641 	if (thread == NULL)
2642 		return;
2643 
2644 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2645 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2646 		return;
2647 	}
2648 
2649 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2650 	printf("%15s [%04d] ", tstr, sample->cpu);
2651 
2652 	if (sched->show_cpu_visual) {
2653 		u32 i;
2654 		char c;
2655 
2656 		printf("  ");
2657 		for (i = 0; i < max_cpus; ++i) {
2658 			c = (i == sample->cpu) ? 'm' : ' ';
2659 			printf("%c", c);
2660 		}
2661 		printf("  ");
2662 	}
2663 
2664 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2665 
2666 	/* dt spacer */
2667 	printf("  %9s  %9s  %9s ", "", "", "");
2668 
2669 	printf("migrated: %s", timehist_get_commstr(migrated));
2670 	printf(" cpu %d => %d", ocpu, dcpu);
2671 
2672 	printf("\n");
2673 }
2674 
2675 static int timehist_migrate_task_event(const struct perf_tool *tool,
2676 				       union perf_event *event __maybe_unused,
2677 				       struct evsel *evsel,
2678 				       struct perf_sample *sample,
2679 				       struct machine *machine)
2680 {
2681 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2682 	struct thread *thread;
2683 	struct thread_runtime *tr = NULL;
2684 	/* want pid of migrated task not pid in sample */
2685 	const u32 pid = evsel__intval(evsel, sample, "pid");
2686 
2687 	thread = machine__findnew_thread(machine, 0, pid);
2688 	if (thread == NULL)
2689 		return -1;
2690 
2691 	tr = thread__get_runtime(thread);
2692 	if (tr == NULL)
2693 		return -1;
2694 
2695 	tr->migrations++;
2696 
2697 	/* show migrations if requested */
2698 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2699 
2700 	return 0;
2701 }
2702 
2703 static void timehist_update_task_prio(struct evsel *evsel,
2704 				      struct perf_sample *sample,
2705 				      struct machine *machine)
2706 {
2707 	struct thread *thread;
2708 	struct thread_runtime *tr = NULL;
2709 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2710 	const u32 next_prio = evsel__intval(evsel, sample, "next_prio");
2711 
2712 	if (next_pid == 0)
2713 		thread = get_idle_thread(sample->cpu);
2714 	else
2715 		thread = machine__findnew_thread(machine, -1, next_pid);
2716 
2717 	if (thread == NULL)
2718 		return;
2719 
2720 	tr = thread__get_runtime(thread);
2721 	if (tr == NULL)
2722 		return;
2723 
2724 	tr->prio = next_prio;
2725 }
2726 
2727 static int timehist_sched_change_event(const struct perf_tool *tool,
2728 				       union perf_event *event,
2729 				       struct evsel *evsel,
2730 				       struct perf_sample *sample,
2731 				       struct machine *machine)
2732 {
2733 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2734 	struct perf_time_interval *ptime = &sched->ptime;
2735 	struct addr_location al;
2736 	struct thread *thread;
2737 	struct thread_runtime *tr = NULL;
2738 	u64 tprev, t = sample->time;
2739 	int rc = 0;
2740 	const char state = evsel__taskstate(evsel, sample, "prev_state");
2741 
2742 	addr_location__init(&al);
2743 	if (machine__resolve(machine, &al, sample) < 0) {
2744 		pr_err("problem processing %d event. skipping it\n",
2745 		       event->header.type);
2746 		rc = -1;
2747 		goto out;
2748 	}
2749 
2750 	if (sched->show_prio || sched->prio_str)
2751 		timehist_update_task_prio(evsel, sample, machine);
2752 
2753 	thread = timehist_get_thread(sched, sample, machine, evsel);
2754 	if (thread == NULL) {
2755 		rc = -1;
2756 		goto out;
2757 	}
2758 
2759 	if (timehist_skip_sample(sched, thread, evsel, sample))
2760 		goto out;
2761 
2762 	tr = thread__get_runtime(thread);
2763 	if (tr == NULL) {
2764 		rc = -1;
2765 		goto out;
2766 	}
2767 
2768 	tprev = evsel__get_time(evsel, sample->cpu);
2769 
2770 	/*
2771 	 * If start time given:
2772 	 * - sample time is under window user cares about - skip sample
2773 	 * - tprev is under window user cares about  - reset to start of window
2774 	 */
2775 	if (ptime->start && ptime->start > t)
2776 		goto out;
2777 
2778 	if (tprev && ptime->start > tprev)
2779 		tprev = ptime->start;
2780 
2781 	/*
2782 	 * If end time given:
2783 	 * - previous sched event is out of window - we are done
2784 	 * - sample time is beyond window user cares about - reset it
2785 	 *   to close out stats for time window interest
2786 	 * - If tprev is 0, that is, sched_in event for current task is
2787 	 *   not recorded, cannot determine whether sched_in event is
2788 	 *   within time window interest - ignore it
2789 	 */
2790 	if (ptime->end) {
2791 		if (!tprev || tprev > ptime->end)
2792 			goto out;
2793 
2794 		if (t > ptime->end)
2795 			t = ptime->end;
2796 	}
2797 
2798 	if (!sched->idle_hist || thread__tid(thread) == 0) {
2799 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2800 			timehist_update_runtime_stats(tr, t, tprev);
2801 
2802 		if (sched->idle_hist) {
2803 			struct idle_thread_runtime *itr = (void *)tr;
2804 			struct thread_runtime *last_tr;
2805 
2806 			if (itr->last_thread == NULL)
2807 				goto out;
2808 
2809 			/* add current idle time as last thread's runtime */
2810 			last_tr = thread__get_runtime(itr->last_thread);
2811 			if (last_tr == NULL)
2812 				goto out;
2813 
2814 			timehist_update_runtime_stats(last_tr, t, tprev);
2815 			/*
2816 			 * remove delta time of last thread as it's not updated
2817 			 * and otherwise it will show an invalid value next
2818 			 * time.  we only care total run time and run stat.
2819 			 */
2820 			last_tr->dt_run = 0;
2821 			last_tr->dt_delay = 0;
2822 			last_tr->dt_sleep = 0;
2823 			last_tr->dt_iowait = 0;
2824 			last_tr->dt_preempt = 0;
2825 
2826 			if (itr->cursor.nr)
2827 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2828 
2829 			itr->last_thread = NULL;
2830 		}
2831 
2832 		if (!sched->summary_only)
2833 			timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2834 	}
2835 
2836 out:
2837 	if (sched->hist_time.start == 0 && t >= ptime->start)
2838 		sched->hist_time.start = t;
2839 	if (ptime->end == 0 || t <= ptime->end)
2840 		sched->hist_time.end = t;
2841 
2842 	if (tr) {
2843 		/* time of this sched_switch event becomes last time task seen */
2844 		tr->last_time = sample->time;
2845 
2846 		/* last state is used to determine where to account wait time */
2847 		tr->last_state = state;
2848 
2849 		/* sched out event for task so reset ready to run time */
2850 		if (state == 'R')
2851 			tr->ready_to_run = t;
2852 		else
2853 			tr->ready_to_run = 0;
2854 	}
2855 
2856 	evsel__save_time(evsel, sample->time, sample->cpu);
2857 
2858 	addr_location__exit(&al);
2859 	return rc;
2860 }
2861 
2862 static int timehist_sched_switch_event(const struct perf_tool *tool,
2863 			     union perf_event *event,
2864 			     struct evsel *evsel,
2865 			     struct perf_sample *sample,
2866 			     struct machine *machine __maybe_unused)
2867 {
2868 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2869 }
2870 
2871 static int process_lost(const struct perf_tool *tool __maybe_unused,
2872 			union perf_event *event,
2873 			struct perf_sample *sample,
2874 			struct machine *machine __maybe_unused)
2875 {
2876 	char tstr[64];
2877 
2878 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2879 	printf("%15s ", tstr);
2880 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2881 
2882 	return 0;
2883 }
2884 
2885 
2886 static void print_thread_runtime(struct thread *t,
2887 				 struct thread_runtime *r)
2888 {
2889 	double mean = avg_stats(&r->run_stats);
2890 	float stddev;
2891 
2892 	printf("%*s   %5d  %9" PRIu64 " ",
2893 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2894 	       (u64) r->run_stats.n);
2895 
2896 	print_sched_time(r->total_run_time, 8);
2897 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2898 	print_sched_time(r->run_stats.min, 6);
2899 	printf(" ");
2900 	print_sched_time((u64) mean, 6);
2901 	printf(" ");
2902 	print_sched_time(r->run_stats.max, 6);
2903 	printf("  ");
2904 	printf("%5.2f", stddev);
2905 	printf("   %5" PRIu64, r->migrations);
2906 	printf("\n");
2907 }
2908 
2909 static void print_thread_waittime(struct thread *t,
2910 				  struct thread_runtime *r)
2911 {
2912 	printf("%*s   %5d  %9" PRIu64 " ",
2913 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2914 	       (u64) r->run_stats.n);
2915 
2916 	print_sched_time(r->total_run_time, 8);
2917 	print_sched_time(r->total_sleep_time, 6);
2918 	printf(" ");
2919 	print_sched_time(r->total_iowait_time, 6);
2920 	printf(" ");
2921 	print_sched_time(r->total_preempt_time, 6);
2922 	printf(" ");
2923 	print_sched_time(r->total_delay_time, 6);
2924 	printf("\n");
2925 }
2926 
2927 struct total_run_stats {
2928 	struct perf_sched *sched;
2929 	u64  sched_count;
2930 	u64  task_count;
2931 	u64  total_run_time;
2932 };
2933 
2934 static int show_thread_runtime(struct thread *t, void *priv)
2935 {
2936 	struct total_run_stats *stats = priv;
2937 	struct thread_runtime *r;
2938 
2939 	if (thread__is_filtered(t))
2940 		return 0;
2941 
2942 	r = thread__priv(t);
2943 	if (r && r->run_stats.n) {
2944 		stats->task_count++;
2945 		stats->sched_count += r->run_stats.n;
2946 		stats->total_run_time += r->total_run_time;
2947 
2948 		if (stats->sched->show_state)
2949 			print_thread_waittime(t, r);
2950 		else
2951 			print_thread_runtime(t, r);
2952 	}
2953 
2954 	return 0;
2955 }
2956 
2957 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2958 {
2959 	const char *sep = " <- ";
2960 	struct callchain_list *chain;
2961 	size_t ret = 0;
2962 	char bf[1024];
2963 	bool first;
2964 
2965 	if (node == NULL)
2966 		return 0;
2967 
2968 	ret = callchain__fprintf_folded(fp, node->parent);
2969 	first = (ret == 0);
2970 
2971 	list_for_each_entry(chain, &node->val, list) {
2972 		if (chain->ip >= PERF_CONTEXT_MAX)
2973 			continue;
2974 		if (chain->ms.sym && chain->ms.sym->ignore)
2975 			continue;
2976 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2977 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2978 							false));
2979 		first = false;
2980 	}
2981 
2982 	return ret;
2983 }
2984 
2985 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2986 {
2987 	size_t ret = 0;
2988 	FILE *fp = stdout;
2989 	struct callchain_node *chain;
2990 	struct rb_node *rb_node = rb_first_cached(root);
2991 
2992 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2993 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2994 	       graph_dotted_line);
2995 
2996 	while (rb_node) {
2997 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2998 		rb_node = rb_next(rb_node);
2999 
3000 		ret += fprintf(fp, "  ");
3001 		print_sched_time(chain->hit, 12);
3002 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
3003 		ret += fprintf(fp, " %8d  ", chain->count);
3004 		ret += callchain__fprintf_folded(fp, chain);
3005 		ret += fprintf(fp, "\n");
3006 	}
3007 
3008 	return ret;
3009 }
3010 
3011 static void timehist_print_summary(struct perf_sched *sched,
3012 				   struct perf_session *session)
3013 {
3014 	struct machine *m = &session->machines.host;
3015 	struct total_run_stats totals;
3016 	u64 task_count;
3017 	struct thread *t;
3018 	struct thread_runtime *r;
3019 	int i;
3020 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
3021 
3022 	memset(&totals, 0, sizeof(totals));
3023 	totals.sched = sched;
3024 
3025 	if (sched->idle_hist) {
3026 		printf("\nIdle-time summary\n");
3027 		printf("%*s  parent  sched-out  ", comm_width, "comm");
3028 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
3029 	} else if (sched->show_state) {
3030 		printf("\nWait-time summary\n");
3031 		printf("%*s  parent   sched-in  ", comm_width, "comm");
3032 		printf("   run-time      sleep      iowait     preempt       delay\n");
3033 	} else {
3034 		printf("\nRuntime summary\n");
3035 		printf("%*s  parent   sched-in  ", comm_width, "comm");
3036 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
3037 	}
3038 	printf("%*s            (count)  ", comm_width, "");
3039 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
3040 	       sched->show_state ? "(msec)" : "%");
3041 	printf("%.117s\n", graph_dotted_line);
3042 
3043 	machine__for_each_thread(m, show_thread_runtime, &totals);
3044 	task_count = totals.task_count;
3045 	if (!task_count)
3046 		printf("<no still running tasks>\n");
3047 
3048 	/* CPU idle stats not tracked when samples were skipped */
3049 	if (sched->skipped_samples && !sched->idle_hist)
3050 		return;
3051 
3052 	printf("\nIdle stats:\n");
3053 	for (i = 0; i < idle_max_cpu; ++i) {
3054 		if (cpu_list && !test_bit(i, cpu_bitmap))
3055 			continue;
3056 
3057 		t = idle_threads[i];
3058 		if (!t)
3059 			continue;
3060 
3061 		r = thread__priv(t);
3062 		if (r && r->run_stats.n) {
3063 			totals.sched_count += r->run_stats.n;
3064 			printf("    CPU %2d idle for ", i);
3065 			print_sched_time(r->total_run_time, 6);
3066 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
3067 		} else
3068 			printf("    CPU %2d idle entire time window\n", i);
3069 	}
3070 
3071 	if (sched->idle_hist && sched->show_callchain) {
3072 		callchain_param.mode  = CHAIN_FOLDED;
3073 		callchain_param.value = CCVAL_PERIOD;
3074 
3075 		callchain_register_param(&callchain_param);
3076 
3077 		printf("\nIdle stats by callchain:\n");
3078 		for (i = 0; i < idle_max_cpu; ++i) {
3079 			struct idle_thread_runtime *itr;
3080 
3081 			t = idle_threads[i];
3082 			if (!t)
3083 				continue;
3084 
3085 			itr = thread__priv(t);
3086 			if (itr == NULL)
3087 				continue;
3088 
3089 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
3090 					     0, &callchain_param);
3091 
3092 			printf("  CPU %2d:", i);
3093 			print_sched_time(itr->tr.total_run_time, 6);
3094 			printf(" msec\n");
3095 			timehist_print_idlehist_callchain(&itr->sorted_root);
3096 			printf("\n");
3097 		}
3098 	}
3099 
3100 	printf("\n"
3101 	       "    Total number of unique tasks: %" PRIu64 "\n"
3102 	       "Total number of context switches: %" PRIu64 "\n",
3103 	       totals.task_count, totals.sched_count);
3104 
3105 	printf("           Total run time (msec): ");
3106 	print_sched_time(totals.total_run_time, 2);
3107 	printf("\n");
3108 
3109 	printf("    Total scheduling time (msec): ");
3110 	print_sched_time(hist_time, 2);
3111 	printf(" (x %d)\n", sched->max_cpu.cpu);
3112 }
3113 
3114 typedef int (*sched_handler)(const struct perf_tool *tool,
3115 			  union perf_event *event,
3116 			  struct evsel *evsel,
3117 			  struct perf_sample *sample,
3118 			  struct machine *machine);
3119 
3120 static int perf_timehist__process_sample(const struct perf_tool *tool,
3121 					 union perf_event *event,
3122 					 struct perf_sample *sample,
3123 					 struct evsel *evsel,
3124 					 struct machine *machine)
3125 {
3126 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
3127 	int err = 0;
3128 	struct perf_cpu this_cpu = {
3129 		.cpu = sample->cpu,
3130 	};
3131 
3132 	if (this_cpu.cpu > sched->max_cpu.cpu)
3133 		sched->max_cpu = this_cpu;
3134 
3135 	if (evsel->handler != NULL) {
3136 		sched_handler f = evsel->handler;
3137 
3138 		err = f(tool, event, evsel, sample, machine);
3139 	}
3140 
3141 	return err;
3142 }
3143 
3144 static int timehist_check_attr(struct perf_sched *sched,
3145 			       struct evlist *evlist)
3146 {
3147 	struct evsel *evsel;
3148 	struct evsel_runtime *er;
3149 
3150 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
3151 		er = evsel__get_runtime(evsel);
3152 		if (er == NULL) {
3153 			pr_err("Failed to allocate memory for evsel runtime data\n");
3154 			return -1;
3155 		}
3156 
3157 		/* only need to save callchain related to sched_switch event */
3158 		if (sched->show_callchain &&
3159 		    evsel__name_is(evsel, "sched:sched_switch") &&
3160 		    !evsel__has_callchain(evsel)) {
3161 			pr_info("Samples of sched_switch event do not have callchains.\n");
3162 			sched->show_callchain = 0;
3163 			symbol_conf.use_callchain = 0;
3164 		}
3165 	}
3166 
3167 	return 0;
3168 }
3169 
3170 static int timehist_parse_prio_str(struct perf_sched *sched)
3171 {
3172 	char *p;
3173 	unsigned long start_prio, end_prio;
3174 	const char *str = sched->prio_str;
3175 
3176 	if (!str)
3177 		return 0;
3178 
3179 	while (isdigit(*str)) {
3180 		p = NULL;
3181 		start_prio = strtoul(str, &p, 0);
3182 		if (start_prio >= MAX_PRIO || (*p != '\0' && *p != ',' && *p != '-'))
3183 			return -1;
3184 
3185 		if (*p == '-') {
3186 			str = ++p;
3187 			p = NULL;
3188 			end_prio = strtoul(str, &p, 0);
3189 
3190 			if (end_prio >= MAX_PRIO || (*p != '\0' && *p != ','))
3191 				return -1;
3192 
3193 			if (end_prio < start_prio)
3194 				return -1;
3195 		} else {
3196 			end_prio = start_prio;
3197 		}
3198 
3199 		for (; start_prio <= end_prio; start_prio++)
3200 			__set_bit(start_prio, sched->prio_bitmap);
3201 
3202 		if (*p)
3203 			++p;
3204 
3205 		str = p;
3206 	}
3207 
3208 	return 0;
3209 }
3210 
3211 static int perf_sched__timehist(struct perf_sched *sched)
3212 {
3213 	struct evsel_str_handler handlers[] = {
3214 		{ "sched:sched_switch",       timehist_sched_switch_event, },
3215 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
3216 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
3217 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
3218 	};
3219 	const struct evsel_str_handler migrate_handlers[] = {
3220 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
3221 	};
3222 	struct perf_data data = {
3223 		.path  = input_name,
3224 		.mode  = PERF_DATA_MODE_READ,
3225 		.force = sched->force,
3226 	};
3227 
3228 	struct perf_session *session;
3229 	struct evlist *evlist;
3230 	int err = -1;
3231 
3232 	/*
3233 	 * event handlers for timehist option
3234 	 */
3235 	sched->tool.sample	 = perf_timehist__process_sample;
3236 	sched->tool.mmap	 = perf_event__process_mmap;
3237 	sched->tool.comm	 = perf_event__process_comm;
3238 	sched->tool.exit	 = perf_event__process_exit;
3239 	sched->tool.fork	 = perf_event__process_fork;
3240 	sched->tool.lost	 = process_lost;
3241 	sched->tool.attr	 = perf_event__process_attr;
3242 	sched->tool.tracing_data = perf_event__process_tracing_data;
3243 	sched->tool.build_id	 = perf_event__process_build_id;
3244 
3245 	sched->tool.ordering_requires_timestamps = true;
3246 
3247 	symbol_conf.use_callchain = sched->show_callchain;
3248 
3249 	session = perf_session__new(&data, &sched->tool);
3250 	if (IS_ERR(session))
3251 		return PTR_ERR(session);
3252 
3253 	if (cpu_list) {
3254 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3255 		if (err < 0)
3256 			goto out;
3257 	}
3258 
3259 	evlist = session->evlist;
3260 
3261 	symbol__init(&session->header.env);
3262 
3263 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3264 		pr_err("Invalid time string\n");
3265 		err = -EINVAL;
3266 		goto out;
3267 	}
3268 
3269 	if (timehist_check_attr(sched, evlist) != 0)
3270 		goto out;
3271 
3272 	if (timehist_parse_prio_str(sched) != 0) {
3273 		pr_err("Invalid prio string\n");
3274 		goto out;
3275 	}
3276 
3277 	setup_pager();
3278 
3279 	/* prefer sched_waking if it is captured */
3280 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3281 		handlers[1].handler = timehist_sched_wakeup_ignore;
3282 
3283 	/* setup per-evsel handlers */
3284 	if (perf_session__set_tracepoints_handlers(session, handlers))
3285 		goto out;
3286 
3287 	/* sched_switch event at a minimum needs to exist */
3288 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3289 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3290 		goto out;
3291 	}
3292 
3293 	if (sched->show_migrations &&
3294 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3295 		goto out;
3296 
3297 	/* pre-allocate struct for per-CPU idle stats */
3298 	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3299 	if (sched->max_cpu.cpu == 0)
3300 		sched->max_cpu.cpu = 4;
3301 	if (init_idle_threads(sched->max_cpu.cpu))
3302 		goto out;
3303 
3304 	/* summary_only implies summary option, but don't overwrite summary if set */
3305 	if (sched->summary_only)
3306 		sched->summary = sched->summary_only;
3307 
3308 	if (!sched->summary_only)
3309 		timehist_header(sched);
3310 
3311 	err = perf_session__process_events(session);
3312 	if (err) {
3313 		pr_err("Failed to process events, error %d", err);
3314 		goto out;
3315 	}
3316 
3317 	sched->nr_events      = evlist->stats.nr_events[0];
3318 	sched->nr_lost_events = evlist->stats.total_lost;
3319 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3320 
3321 	if (sched->summary)
3322 		timehist_print_summary(sched, session);
3323 
3324 out:
3325 	free_idle_threads();
3326 	perf_session__delete(session);
3327 
3328 	return err;
3329 }
3330 
3331 
3332 static void print_bad_events(struct perf_sched *sched)
3333 {
3334 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3335 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3336 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3337 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3338 	}
3339 	if (sched->nr_lost_events && sched->nr_events) {
3340 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3341 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3342 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3343 	}
3344 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3345 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3346 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3347 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3348 		if (sched->nr_lost_events)
3349 			printf(" (due to lost events?)");
3350 		printf("\n");
3351 	}
3352 }
3353 
3354 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3355 {
3356 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3357 	struct work_atoms *this;
3358 	const char *comm = thread__comm_str(data->thread), *this_comm;
3359 	bool leftmost = true;
3360 
3361 	while (*new) {
3362 		int cmp;
3363 
3364 		this = container_of(*new, struct work_atoms, node);
3365 		parent = *new;
3366 
3367 		this_comm = thread__comm_str(this->thread);
3368 		cmp = strcmp(comm, this_comm);
3369 		if (cmp > 0) {
3370 			new = &((*new)->rb_left);
3371 		} else if (cmp < 0) {
3372 			new = &((*new)->rb_right);
3373 			leftmost = false;
3374 		} else {
3375 			this->num_merged++;
3376 			this->total_runtime += data->total_runtime;
3377 			this->nb_atoms += data->nb_atoms;
3378 			this->total_lat += data->total_lat;
3379 			list_splice(&data->work_list, &this->work_list);
3380 			if (this->max_lat < data->max_lat) {
3381 				this->max_lat = data->max_lat;
3382 				this->max_lat_start = data->max_lat_start;
3383 				this->max_lat_end = data->max_lat_end;
3384 			}
3385 			zfree(&data);
3386 			return;
3387 		}
3388 	}
3389 
3390 	data->num_merged++;
3391 	rb_link_node(&data->node, parent, new);
3392 	rb_insert_color_cached(&data->node, root, leftmost);
3393 }
3394 
3395 static void perf_sched__merge_lat(struct perf_sched *sched)
3396 {
3397 	struct work_atoms *data;
3398 	struct rb_node *node;
3399 
3400 	if (sched->skip_merge)
3401 		return;
3402 
3403 	while ((node = rb_first_cached(&sched->atom_root))) {
3404 		rb_erase_cached(node, &sched->atom_root);
3405 		data = rb_entry(node, struct work_atoms, node);
3406 		__merge_work_atoms(&sched->merged_atom_root, data);
3407 	}
3408 }
3409 
3410 static int setup_cpus_switch_event(struct perf_sched *sched)
3411 {
3412 	unsigned int i;
3413 
3414 	sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3415 	if (!sched->cpu_last_switched)
3416 		return -1;
3417 
3418 	sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3419 	if (!sched->curr_pid) {
3420 		zfree(&sched->cpu_last_switched);
3421 		return -1;
3422 	}
3423 
3424 	for (i = 0; i < MAX_CPUS; i++)
3425 		sched->curr_pid[i] = -1;
3426 
3427 	return 0;
3428 }
3429 
3430 static void free_cpus_switch_event(struct perf_sched *sched)
3431 {
3432 	zfree(&sched->curr_pid);
3433 	zfree(&sched->cpu_last_switched);
3434 }
3435 
3436 static int perf_sched__lat(struct perf_sched *sched)
3437 {
3438 	int rc = -1;
3439 	struct rb_node *next;
3440 
3441 	setup_pager();
3442 
3443 	if (setup_cpus_switch_event(sched))
3444 		return rc;
3445 
3446 	if (perf_sched__read_events(sched))
3447 		goto out_free_cpus_switch_event;
3448 
3449 	perf_sched__merge_lat(sched);
3450 	perf_sched__sort_lat(sched);
3451 
3452 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3453 	printf("  Task                  |   Runtime ms  |  Count   | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3454 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3455 
3456 	next = rb_first_cached(&sched->sorted_atom_root);
3457 
3458 	while (next) {
3459 		struct work_atoms *work_list;
3460 
3461 		work_list = rb_entry(next, struct work_atoms, node);
3462 		output_lat_thread(sched, work_list);
3463 		next = rb_next(next);
3464 		thread__zput(work_list->thread);
3465 	}
3466 
3467 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3468 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3469 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3470 
3471 	printf(" ---------------------------------------------------\n");
3472 
3473 	print_bad_events(sched);
3474 	printf("\n");
3475 
3476 	rc = 0;
3477 
3478 out_free_cpus_switch_event:
3479 	free_cpus_switch_event(sched);
3480 	return rc;
3481 }
3482 
3483 static int setup_map_cpus(struct perf_sched *sched)
3484 {
3485 	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3486 
3487 	if (sched->map.comp) {
3488 		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3489 		if (!sched->map.comp_cpus)
3490 			return -1;
3491 	}
3492 
3493 	if (sched->map.cpus_str) {
3494 		sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3495 		if (!sched->map.cpus) {
3496 			pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3497 			zfree(&sched->map.comp_cpus);
3498 			return -1;
3499 		}
3500 	}
3501 
3502 	return 0;
3503 }
3504 
3505 static int setup_color_pids(struct perf_sched *sched)
3506 {
3507 	struct perf_thread_map *map;
3508 
3509 	if (!sched->map.color_pids_str)
3510 		return 0;
3511 
3512 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3513 	if (!map) {
3514 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3515 		return -1;
3516 	}
3517 
3518 	sched->map.color_pids = map;
3519 	return 0;
3520 }
3521 
3522 static int setup_color_cpus(struct perf_sched *sched)
3523 {
3524 	struct perf_cpu_map *map;
3525 
3526 	if (!sched->map.color_cpus_str)
3527 		return 0;
3528 
3529 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3530 	if (!map) {
3531 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3532 		return -1;
3533 	}
3534 
3535 	sched->map.color_cpus = map;
3536 	return 0;
3537 }
3538 
3539 static int perf_sched__map(struct perf_sched *sched)
3540 {
3541 	int rc = -1;
3542 
3543 	sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3544 	if (!sched->curr_thread)
3545 		return rc;
3546 
3547 	sched->curr_out_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_out_thread)));
3548 	if (!sched->curr_out_thread)
3549 		return rc;
3550 
3551 	if (setup_cpus_switch_event(sched))
3552 		goto out_free_curr_thread;
3553 
3554 	if (setup_map_cpus(sched))
3555 		goto out_free_cpus_switch_event;
3556 
3557 	if (setup_color_pids(sched))
3558 		goto out_put_map_cpus;
3559 
3560 	if (setup_color_cpus(sched))
3561 		goto out_put_color_pids;
3562 
3563 	setup_pager();
3564 	if (perf_sched__read_events(sched))
3565 		goto out_put_color_cpus;
3566 
3567 	rc = 0;
3568 	print_bad_events(sched);
3569 
3570 out_put_color_cpus:
3571 	perf_cpu_map__put(sched->map.color_cpus);
3572 
3573 out_put_color_pids:
3574 	perf_thread_map__put(sched->map.color_pids);
3575 
3576 out_put_map_cpus:
3577 	zfree(&sched->map.comp_cpus);
3578 	perf_cpu_map__put(sched->map.cpus);
3579 
3580 out_free_cpus_switch_event:
3581 	free_cpus_switch_event(sched);
3582 
3583 out_free_curr_thread:
3584 	zfree(&sched->curr_thread);
3585 	return rc;
3586 }
3587 
3588 static int perf_sched__replay(struct perf_sched *sched)
3589 {
3590 	int ret;
3591 	unsigned long i;
3592 
3593 	mutex_init(&sched->start_work_mutex);
3594 	mutex_init(&sched->work_done_wait_mutex);
3595 
3596 	ret = setup_cpus_switch_event(sched);
3597 	if (ret)
3598 		goto out_mutex_destroy;
3599 
3600 	calibrate_run_measurement_overhead(sched);
3601 	calibrate_sleep_measurement_overhead(sched);
3602 
3603 	test_calibrations(sched);
3604 
3605 	ret = perf_sched__read_events(sched);
3606 	if (ret)
3607 		goto out_free_cpus_switch_event;
3608 
3609 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3610 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3611 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3612 
3613 	if (sched->targetless_wakeups)
3614 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3615 	if (sched->multitarget_wakeups)
3616 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3617 	if (sched->nr_run_events_optimized)
3618 		printf("run atoms optimized: %ld\n",
3619 			sched->nr_run_events_optimized);
3620 
3621 	print_task_traces(sched);
3622 	add_cross_task_wakeups(sched);
3623 
3624 	sched->thread_funcs_exit = false;
3625 	create_tasks(sched);
3626 	printf("------------------------------------------------------------\n");
3627 	if (sched->replay_repeat == 0)
3628 		sched->replay_repeat = UINT_MAX;
3629 
3630 	for (i = 0; i < sched->replay_repeat; i++)
3631 		run_one_test(sched);
3632 
3633 	sched->thread_funcs_exit = true;
3634 	destroy_tasks(sched);
3635 
3636 out_free_cpus_switch_event:
3637 	free_cpus_switch_event(sched);
3638 
3639 out_mutex_destroy:
3640 	mutex_destroy(&sched->start_work_mutex);
3641 	mutex_destroy(&sched->work_done_wait_mutex);
3642 	return ret;
3643 }
3644 
3645 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3646 			  const char * const usage_msg[])
3647 {
3648 	char *tmp, *tok, *str = strdup(sched->sort_order);
3649 
3650 	for (tok = strtok_r(str, ", ", &tmp);
3651 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3652 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3653 			usage_with_options_msg(usage_msg, options,
3654 					"Unknown --sort key: `%s'", tok);
3655 		}
3656 	}
3657 
3658 	free(str);
3659 
3660 	sort_dimension__add("pid", &sched->cmp_pid);
3661 }
3662 
3663 static bool schedstat_events_exposed(void)
3664 {
3665 	/*
3666 	 * Select "sched:sched_stat_wait" event to check
3667 	 * whether schedstat tracepoints are exposed.
3668 	 */
3669 	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3670 		false : true;
3671 }
3672 
3673 static int __cmd_record(int argc, const char **argv)
3674 {
3675 	unsigned int rec_argc, i, j;
3676 	char **rec_argv;
3677 	const char **rec_argv_copy;
3678 	const char * const record_args[] = {
3679 		"record",
3680 		"-a",
3681 		"-R",
3682 		"-m", "1024",
3683 		"-c", "1",
3684 		"-e", "sched:sched_switch",
3685 		"-e", "sched:sched_stat_runtime",
3686 		"-e", "sched:sched_process_fork",
3687 		"-e", "sched:sched_wakeup_new",
3688 		"-e", "sched:sched_migrate_task",
3689 	};
3690 
3691 	/*
3692 	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3693 	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3694 	 * to prevent "perf sched record" execution failure, determine
3695 	 * whether to record schedstat events according to actual situation.
3696 	 */
3697 	const char * const schedstat_args[] = {
3698 		"-e", "sched:sched_stat_wait",
3699 		"-e", "sched:sched_stat_sleep",
3700 		"-e", "sched:sched_stat_iowait",
3701 	};
3702 	unsigned int schedstat_argc = schedstat_events_exposed() ?
3703 		ARRAY_SIZE(schedstat_args) : 0;
3704 
3705 	struct tep_event *waking_event;
3706 	int ret;
3707 
3708 	/*
3709 	 * +2 for either "-e", "sched:sched_wakeup" or
3710 	 * "-e", "sched:sched_waking"
3711 	 */
3712 	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3713 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3714 	if (rec_argv == NULL)
3715 		return -ENOMEM;
3716 	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3717 	if (rec_argv_copy == NULL) {
3718 		free(rec_argv);
3719 		return -ENOMEM;
3720 	}
3721 
3722 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3723 		rec_argv[i] = strdup(record_args[i]);
3724 
3725 	rec_argv[i++] = strdup("-e");
3726 	waking_event = trace_event__tp_format("sched", "sched_waking");
3727 	if (!IS_ERR(waking_event))
3728 		rec_argv[i++] = strdup("sched:sched_waking");
3729 	else
3730 		rec_argv[i++] = strdup("sched:sched_wakeup");
3731 
3732 	for (j = 0; j < schedstat_argc; j++)
3733 		rec_argv[i++] = strdup(schedstat_args[j]);
3734 
3735 	for (j = 1; j < (unsigned int)argc; j++, i++)
3736 		rec_argv[i] = strdup(argv[j]);
3737 
3738 	BUG_ON(i != rec_argc);
3739 
3740 	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3741 	ret = cmd_record(rec_argc, rec_argv_copy);
3742 
3743 	for (i = 0; i < rec_argc; i++)
3744 		free(rec_argv[i]);
3745 	free(rec_argv);
3746 	free(rec_argv_copy);
3747 
3748 	return ret;
3749 }
3750 
3751 int cmd_sched(int argc, const char **argv)
3752 {
3753 	static const char default_sort_order[] = "avg, max, switch, runtime";
3754 	struct perf_sched sched = {
3755 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3756 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3757 		.sort_order	      = default_sort_order,
3758 		.replay_repeat	      = 10,
3759 		.profile_cpu	      = -1,
3760 		.next_shortname1      = 'A',
3761 		.next_shortname2      = '0',
3762 		.skip_merge           = 0,
3763 		.show_callchain	      = 1,
3764 		.max_stack            = 5,
3765 	};
3766 	const struct option sched_options[] = {
3767 	OPT_STRING('i', "input", &input_name, "file",
3768 		    "input file name"),
3769 	OPT_INCR('v', "verbose", &verbose,
3770 		    "be more verbose (show symbol address, etc)"),
3771 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3772 		    "dump raw trace in ASCII"),
3773 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3774 	OPT_END()
3775 	};
3776 	const struct option latency_options[] = {
3777 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3778 		   "sort by key(s): runtime, switch, avg, max"),
3779 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3780 		    "CPU to profile on"),
3781 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3782 		    "latency stats per pid instead of per comm"),
3783 	OPT_PARENT(sched_options)
3784 	};
3785 	const struct option replay_options[] = {
3786 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3787 		     "repeat the workload replay N times (0: infinite)"),
3788 	OPT_PARENT(sched_options)
3789 	};
3790 	const struct option map_options[] = {
3791 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3792 		    "map output in compact mode"),
3793 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3794 		   "highlight given pids in map"),
3795 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3796                     "highlight given CPUs in map"),
3797 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3798                     "display given CPUs in map"),
3799 	OPT_STRING(0, "task-name", &sched.map.task_name, "task",
3800 		"map output only for the given task name(s)."),
3801 	OPT_BOOLEAN(0, "fuzzy-name", &sched.map.fuzzy,
3802 		"given command name can be partially matched (fuzzy matching)"),
3803 	OPT_PARENT(sched_options)
3804 	};
3805 	const struct option timehist_options[] = {
3806 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3807 		   "file", "vmlinux pathname"),
3808 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3809 		   "file", "kallsyms pathname"),
3810 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3811 		    "Display call chains if present (default on)"),
3812 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3813 		   "Maximum number of functions to display backtrace."),
3814 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3815 		    "Look for files with symbols relative to this directory"),
3816 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3817 		    "Show only syscall summary with statistics"),
3818 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3819 		    "Show all syscalls and summary with statistics"),
3820 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3821 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3822 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3823 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3824 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3825 	OPT_STRING(0, "time", &sched.time_str, "str",
3826 		   "Time span for analysis (start,stop)"),
3827 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3828 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3829 		   "analyze events only for given process id(s)"),
3830 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3831 		   "analyze events only for given thread id(s)"),
3832 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3833 	OPT_BOOLEAN(0, "show-prio", &sched.show_prio, "Show task priority"),
3834 	OPT_STRING(0, "prio", &sched.prio_str, "prio",
3835 		   "analyze events only for given task priority(ies)"),
3836 	OPT_PARENT(sched_options)
3837 	};
3838 
3839 	const char * const latency_usage[] = {
3840 		"perf sched latency [<options>]",
3841 		NULL
3842 	};
3843 	const char * const replay_usage[] = {
3844 		"perf sched replay [<options>]",
3845 		NULL
3846 	};
3847 	const char * const map_usage[] = {
3848 		"perf sched map [<options>]",
3849 		NULL
3850 	};
3851 	const char * const timehist_usage[] = {
3852 		"perf sched timehist [<options>]",
3853 		NULL
3854 	};
3855 	const char *const sched_subcommands[] = { "record", "latency", "map",
3856 						  "replay", "script",
3857 						  "timehist", NULL };
3858 	const char *sched_usage[] = {
3859 		NULL,
3860 		NULL
3861 	};
3862 	struct trace_sched_handler lat_ops  = {
3863 		.wakeup_event	    = latency_wakeup_event,
3864 		.switch_event	    = latency_switch_event,
3865 		.runtime_event	    = latency_runtime_event,
3866 		.migrate_task_event = latency_migrate_task_event,
3867 	};
3868 	struct trace_sched_handler map_ops  = {
3869 		.switch_event	    = map_switch_event,
3870 	};
3871 	struct trace_sched_handler replay_ops  = {
3872 		.wakeup_event	    = replay_wakeup_event,
3873 		.switch_event	    = replay_switch_event,
3874 		.fork_event	    = replay_fork_event,
3875 	};
3876 	int ret;
3877 
3878 	perf_tool__init(&sched.tool, /*ordered_events=*/true);
3879 	sched.tool.sample	 = perf_sched__process_tracepoint_sample;
3880 	sched.tool.comm		 = perf_sched__process_comm;
3881 	sched.tool.namespaces	 = perf_event__process_namespaces;
3882 	sched.tool.lost		 = perf_event__process_lost;
3883 	sched.tool.fork		 = perf_sched__process_fork_event;
3884 
3885 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3886 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3887 	if (!argc)
3888 		usage_with_options(sched_usage, sched_options);
3889 
3890 	/*
3891 	 * Aliased to 'perf script' for now:
3892 	 */
3893 	if (!strcmp(argv[0], "script")) {
3894 		return cmd_script(argc, argv);
3895 	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3896 		return __cmd_record(argc, argv);
3897 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3898 		sched.tp_handler = &lat_ops;
3899 		if (argc > 1) {
3900 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3901 			if (argc)
3902 				usage_with_options(latency_usage, latency_options);
3903 		}
3904 		setup_sorting(&sched, latency_options, latency_usage);
3905 		return perf_sched__lat(&sched);
3906 	} else if (!strcmp(argv[0], "map")) {
3907 		if (argc) {
3908 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3909 			if (argc)
3910 				usage_with_options(map_usage, map_options);
3911 
3912 			if (sched.map.task_name) {
3913 				sched.map.task_names = strlist__new(sched.map.task_name, NULL);
3914 				if (sched.map.task_names == NULL) {
3915 					fprintf(stderr, "Failed to parse task names\n");
3916 					return -1;
3917 				}
3918 			}
3919 		}
3920 		sched.tp_handler = &map_ops;
3921 		setup_sorting(&sched, latency_options, latency_usage);
3922 		return perf_sched__map(&sched);
3923 	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3924 		sched.tp_handler = &replay_ops;
3925 		if (argc) {
3926 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3927 			if (argc)
3928 				usage_with_options(replay_usage, replay_options);
3929 		}
3930 		return perf_sched__replay(&sched);
3931 	} else if (!strcmp(argv[0], "timehist")) {
3932 		if (argc) {
3933 			argc = parse_options(argc, argv, timehist_options,
3934 					     timehist_usage, 0);
3935 			if (argc)
3936 				usage_with_options(timehist_usage, timehist_options);
3937 		}
3938 		if ((sched.show_wakeups || sched.show_next) &&
3939 		    sched.summary_only) {
3940 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3941 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3942 			if (sched.show_wakeups)
3943 				parse_options_usage(NULL, timehist_options, "w", true);
3944 			if (sched.show_next)
3945 				parse_options_usage(NULL, timehist_options, "n", true);
3946 			return -EINVAL;
3947 		}
3948 		ret = symbol__validate_sym_arguments();
3949 		if (ret)
3950 			return ret;
3951 
3952 		return perf_sched__timehist(&sched);
3953 	} else {
3954 		usage_with_options(sched_usage, sched_options);
3955 	}
3956 
3957 	/* free usage string allocated by parse_options_subcommand */
3958 	free((void *)sched_usage[0]);
3959 
3960 	return 0;
3961 }
3962