xref: /linux/tools/perf/builtin-sched.c (revision 3f41368fbfe1b3d5922d317fe1a0a0cab6846802)
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 
55 static const char *cpu_list;
56 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57 
58 struct sched_atom;
59 
60 struct task_desc {
61 	unsigned long		nr;
62 	unsigned long		pid;
63 	char			comm[COMM_LEN];
64 
65 	unsigned long		nr_events;
66 	unsigned long		curr_event;
67 	struct sched_atom	**atoms;
68 
69 	pthread_t		thread;
70 	sem_t			sleep_sem;
71 
72 	sem_t			ready_for_work;
73 	sem_t			work_done_sem;
74 
75 	u64			cpu_usage;
76 };
77 
78 enum sched_event_type {
79 	SCHED_EVENT_RUN,
80 	SCHED_EVENT_SLEEP,
81 	SCHED_EVENT_WAKEUP,
82 	SCHED_EVENT_MIGRATION,
83 };
84 
85 struct sched_atom {
86 	enum sched_event_type	type;
87 	int			specific_wait;
88 	u64			timestamp;
89 	u64			duration;
90 	unsigned long		nr;
91 	sem_t			*wait_sem;
92 	struct task_desc	*wakee;
93 };
94 
95 enum thread_state {
96 	THREAD_SLEEPING = 0,
97 	THREAD_WAIT_CPU,
98 	THREAD_SCHED_IN,
99 	THREAD_IGNORE
100 };
101 
102 struct work_atom {
103 	struct list_head	list;
104 	enum thread_state	state;
105 	u64			sched_out_time;
106 	u64			wake_up_time;
107 	u64			sched_in_time;
108 	u64			runtime;
109 };
110 
111 struct work_atoms {
112 	struct list_head	work_list;
113 	struct thread		*thread;
114 	struct rb_node		node;
115 	u64			max_lat;
116 	u64			max_lat_start;
117 	u64			max_lat_end;
118 	u64			total_lat;
119 	u64			nb_atoms;
120 	u64			total_runtime;
121 	int			num_merged;
122 };
123 
124 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
125 
126 struct perf_sched;
127 
128 struct trace_sched_handler {
129 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
130 			    struct perf_sample *sample, struct machine *machine);
131 
132 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
133 			     struct perf_sample *sample, struct machine *machine);
134 
135 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
136 			    struct perf_sample *sample, struct machine *machine);
137 
138 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
139 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
140 			  struct machine *machine);
141 
142 	int (*migrate_task_event)(struct perf_sched *sched,
143 				  struct evsel *evsel,
144 				  struct perf_sample *sample,
145 				  struct machine *machine);
146 };
147 
148 #define COLOR_PIDS PERF_COLOR_BLUE
149 #define COLOR_CPUS PERF_COLOR_BG_RED
150 
151 struct perf_sched_map {
152 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
153 	struct perf_cpu		*comp_cpus;
154 	bool			 comp;
155 	struct perf_thread_map *color_pids;
156 	const char		*color_pids_str;
157 	struct perf_cpu_map	*color_cpus;
158 	const char		*color_cpus_str;
159 	struct perf_cpu_map	*cpus;
160 	const char		*cpus_str;
161 };
162 
163 struct perf_sched {
164 	struct perf_tool tool;
165 	const char	 *sort_order;
166 	unsigned long	 nr_tasks;
167 	struct task_desc **pid_to_task;
168 	struct task_desc **tasks;
169 	const struct trace_sched_handler *tp_handler;
170 	struct mutex	 start_work_mutex;
171 	struct mutex	 work_done_wait_mutex;
172 	int		 profile_cpu;
173 /*
174  * Track the current task - that way we can know whether there's any
175  * weird events, such as a task being switched away that is not current.
176  */
177 	struct perf_cpu	 max_cpu;
178 	u32		 *curr_pid;
179 	struct thread	 **curr_thread;
180 	char		 next_shortname1;
181 	char		 next_shortname2;
182 	unsigned int	 replay_repeat;
183 	unsigned long	 nr_run_events;
184 	unsigned long	 nr_sleep_events;
185 	unsigned long	 nr_wakeup_events;
186 	unsigned long	 nr_sleep_corrections;
187 	unsigned long	 nr_run_events_optimized;
188 	unsigned long	 targetless_wakeups;
189 	unsigned long	 multitarget_wakeups;
190 	unsigned long	 nr_runs;
191 	unsigned long	 nr_timestamps;
192 	unsigned long	 nr_unordered_timestamps;
193 	unsigned long	 nr_context_switch_bugs;
194 	unsigned long	 nr_events;
195 	unsigned long	 nr_lost_chunks;
196 	unsigned long	 nr_lost_events;
197 	u64		 run_measurement_overhead;
198 	u64		 sleep_measurement_overhead;
199 	u64		 start_time;
200 	u64		 cpu_usage;
201 	u64		 runavg_cpu_usage;
202 	u64		 parent_cpu_usage;
203 	u64		 runavg_parent_cpu_usage;
204 	u64		 sum_runtime;
205 	u64		 sum_fluct;
206 	u64		 run_avg;
207 	u64		 all_runtime;
208 	u64		 all_count;
209 	u64		 *cpu_last_switched;
210 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
211 	struct list_head sort_list, cmp_pid;
212 	bool force;
213 	bool skip_merge;
214 	struct perf_sched_map map;
215 
216 	/* options for timehist command */
217 	bool		summary;
218 	bool		summary_only;
219 	bool		idle_hist;
220 	bool		show_callchain;
221 	unsigned int	max_stack;
222 	bool		show_cpu_visual;
223 	bool		show_wakeups;
224 	bool		show_next;
225 	bool		show_migrations;
226 	bool		show_state;
227 	u64		skipped_samples;
228 	const char	*time_str;
229 	struct perf_time_interval ptime;
230 	struct perf_time_interval hist_time;
231 	volatile bool   thread_funcs_exit;
232 };
233 
234 /* per thread run time data */
235 struct thread_runtime {
236 	u64 last_time;      /* time of previous sched in/out event */
237 	u64 dt_run;         /* run time */
238 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
239 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
240 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
241 	u64 dt_delay;       /* time between wakeup and sched-in */
242 	u64 ready_to_run;   /* time of wakeup */
243 
244 	struct stats run_stats;
245 	u64 total_run_time;
246 	u64 total_sleep_time;
247 	u64 total_iowait_time;
248 	u64 total_preempt_time;
249 	u64 total_delay_time;
250 
251 	char last_state;
252 
253 	char shortname[3];
254 	bool comm_changed;
255 
256 	u64 migrations;
257 };
258 
259 /* per event run time data */
260 struct evsel_runtime {
261 	u64 *last_time; /* time this event was last seen per cpu */
262 	u32 ncpu;       /* highest cpu slot allocated */
263 };
264 
265 /* per cpu idle time data */
266 struct idle_thread_runtime {
267 	struct thread_runtime	tr;
268 	struct thread		*last_thread;
269 	struct rb_root_cached	sorted_root;
270 	struct callchain_root	callchain;
271 	struct callchain_cursor	cursor;
272 };
273 
274 /* track idle times per cpu */
275 static struct thread **idle_threads;
276 static int idle_max_cpu;
277 static char idle_comm[] = "<idle>";
278 
279 static u64 get_nsecs(void)
280 {
281 	struct timespec ts;
282 
283 	clock_gettime(CLOCK_MONOTONIC, &ts);
284 
285 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
286 }
287 
288 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
289 {
290 	u64 T0 = get_nsecs(), T1;
291 
292 	do {
293 		T1 = get_nsecs();
294 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
295 }
296 
297 static void sleep_nsecs(u64 nsecs)
298 {
299 	struct timespec ts;
300 
301 	ts.tv_nsec = nsecs % 999999999;
302 	ts.tv_sec = nsecs / 999999999;
303 
304 	nanosleep(&ts, NULL);
305 }
306 
307 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
308 {
309 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
310 	int i;
311 
312 	for (i = 0; i < 10; i++) {
313 		T0 = get_nsecs();
314 		burn_nsecs(sched, 0);
315 		T1 = get_nsecs();
316 		delta = T1-T0;
317 		min_delta = min(min_delta, delta);
318 	}
319 	sched->run_measurement_overhead = min_delta;
320 
321 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
322 }
323 
324 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
325 {
326 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
327 	int i;
328 
329 	for (i = 0; i < 10; i++) {
330 		T0 = get_nsecs();
331 		sleep_nsecs(10000);
332 		T1 = get_nsecs();
333 		delta = T1-T0;
334 		min_delta = min(min_delta, delta);
335 	}
336 	min_delta -= 10000;
337 	sched->sleep_measurement_overhead = min_delta;
338 
339 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340 }
341 
342 static struct sched_atom *
343 get_new_event(struct task_desc *task, u64 timestamp)
344 {
345 	struct sched_atom *event = zalloc(sizeof(*event));
346 	unsigned long idx = task->nr_events;
347 	size_t size;
348 
349 	event->timestamp = timestamp;
350 	event->nr = idx;
351 
352 	task->nr_events++;
353 	size = sizeof(struct sched_atom *) * task->nr_events;
354 	task->atoms = realloc(task->atoms, size);
355 	BUG_ON(!task->atoms);
356 
357 	task->atoms[idx] = event;
358 
359 	return event;
360 }
361 
362 static struct sched_atom *last_event(struct task_desc *task)
363 {
364 	if (!task->nr_events)
365 		return NULL;
366 
367 	return task->atoms[task->nr_events - 1];
368 }
369 
370 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
371 				u64 timestamp, u64 duration)
372 {
373 	struct sched_atom *event, *curr_event = last_event(task);
374 
375 	/*
376 	 * optimize an existing RUN event by merging this one
377 	 * to it:
378 	 */
379 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
380 		sched->nr_run_events_optimized++;
381 		curr_event->duration += duration;
382 		return;
383 	}
384 
385 	event = get_new_event(task, timestamp);
386 
387 	event->type = SCHED_EVENT_RUN;
388 	event->duration = duration;
389 
390 	sched->nr_run_events++;
391 }
392 
393 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
394 				   u64 timestamp, struct task_desc *wakee)
395 {
396 	struct sched_atom *event, *wakee_event;
397 
398 	event = get_new_event(task, timestamp);
399 	event->type = SCHED_EVENT_WAKEUP;
400 	event->wakee = wakee;
401 
402 	wakee_event = last_event(wakee);
403 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
404 		sched->targetless_wakeups++;
405 		return;
406 	}
407 	if (wakee_event->wait_sem) {
408 		sched->multitarget_wakeups++;
409 		return;
410 	}
411 
412 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
413 	sem_init(wakee_event->wait_sem, 0, 0);
414 	wakee_event->specific_wait = 1;
415 	event->wait_sem = wakee_event->wait_sem;
416 
417 	sched->nr_wakeup_events++;
418 }
419 
420 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
421 				  u64 timestamp, const char task_state __maybe_unused)
422 {
423 	struct sched_atom *event = get_new_event(task, timestamp);
424 
425 	event->type = SCHED_EVENT_SLEEP;
426 
427 	sched->nr_sleep_events++;
428 }
429 
430 static struct task_desc *register_pid(struct perf_sched *sched,
431 				      unsigned long pid, const char *comm)
432 {
433 	struct task_desc *task;
434 	static int pid_max;
435 
436 	if (sched->pid_to_task == NULL) {
437 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
438 			pid_max = MAX_PID;
439 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
440 	}
441 	if (pid >= (unsigned long)pid_max) {
442 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
443 			sizeof(struct task_desc *))) == NULL);
444 		while (pid >= (unsigned long)pid_max)
445 			sched->pid_to_task[pid_max++] = NULL;
446 	}
447 
448 	task = sched->pid_to_task[pid];
449 
450 	if (task)
451 		return task;
452 
453 	task = zalloc(sizeof(*task));
454 	task->pid = pid;
455 	task->nr = sched->nr_tasks;
456 	strcpy(task->comm, comm);
457 	/*
458 	 * every task starts in sleeping state - this gets ignored
459 	 * if there's no wakeup pointing to this sleep state:
460 	 */
461 	add_sched_event_sleep(sched, task, 0, 0);
462 
463 	sched->pid_to_task[pid] = task;
464 	sched->nr_tasks++;
465 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
466 	BUG_ON(!sched->tasks);
467 	sched->tasks[task->nr] = task;
468 
469 	if (verbose > 0)
470 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
471 
472 	return task;
473 }
474 
475 
476 static void print_task_traces(struct perf_sched *sched)
477 {
478 	struct task_desc *task;
479 	unsigned long i;
480 
481 	for (i = 0; i < sched->nr_tasks; i++) {
482 		task = sched->tasks[i];
483 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
484 			task->nr, task->comm, task->pid, task->nr_events);
485 	}
486 }
487 
488 static void add_cross_task_wakeups(struct perf_sched *sched)
489 {
490 	struct task_desc *task1, *task2;
491 	unsigned long i, j;
492 
493 	for (i = 0; i < sched->nr_tasks; i++) {
494 		task1 = sched->tasks[i];
495 		j = i + 1;
496 		if (j == sched->nr_tasks)
497 			j = 0;
498 		task2 = sched->tasks[j];
499 		add_sched_event_wakeup(sched, task1, 0, task2);
500 	}
501 }
502 
503 static void perf_sched__process_event(struct perf_sched *sched,
504 				      struct sched_atom *atom)
505 {
506 	int ret = 0;
507 
508 	switch (atom->type) {
509 		case SCHED_EVENT_RUN:
510 			burn_nsecs(sched, atom->duration);
511 			break;
512 		case SCHED_EVENT_SLEEP:
513 			if (atom->wait_sem)
514 				ret = sem_wait(atom->wait_sem);
515 			BUG_ON(ret);
516 			break;
517 		case SCHED_EVENT_WAKEUP:
518 			if (atom->wait_sem)
519 				ret = sem_post(atom->wait_sem);
520 			BUG_ON(ret);
521 			break;
522 		case SCHED_EVENT_MIGRATION:
523 			break;
524 		default:
525 			BUG_ON(1);
526 	}
527 }
528 
529 static u64 get_cpu_usage_nsec_parent(void)
530 {
531 	struct rusage ru;
532 	u64 sum;
533 	int err;
534 
535 	err = getrusage(RUSAGE_SELF, &ru);
536 	BUG_ON(err);
537 
538 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
539 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
540 
541 	return sum;
542 }
543 
544 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
545 {
546 	struct perf_event_attr attr;
547 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
548 	int fd;
549 	struct rlimit limit;
550 	bool need_privilege = false;
551 
552 	memset(&attr, 0, sizeof(attr));
553 
554 	attr.type = PERF_TYPE_SOFTWARE;
555 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
556 
557 force_again:
558 	fd = sys_perf_event_open(&attr, 0, -1, -1,
559 				 perf_event_open_cloexec_flag());
560 
561 	if (fd < 0) {
562 		if (errno == EMFILE) {
563 			if (sched->force) {
564 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
565 				limit.rlim_cur += sched->nr_tasks - cur_task;
566 				if (limit.rlim_cur > limit.rlim_max) {
567 					limit.rlim_max = limit.rlim_cur;
568 					need_privilege = true;
569 				}
570 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
571 					if (need_privilege && errno == EPERM)
572 						strcpy(info, "Need privilege\n");
573 				} else
574 					goto force_again;
575 			} else
576 				strcpy(info, "Have a try with -f option\n");
577 		}
578 		pr_err("Error: sys_perf_event_open() syscall returned "
579 		       "with %d (%s)\n%s", fd,
580 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
581 		exit(EXIT_FAILURE);
582 	}
583 	return fd;
584 }
585 
586 static u64 get_cpu_usage_nsec_self(int fd)
587 {
588 	u64 runtime;
589 	int ret;
590 
591 	ret = read(fd, &runtime, sizeof(runtime));
592 	BUG_ON(ret != sizeof(runtime));
593 
594 	return runtime;
595 }
596 
597 struct sched_thread_parms {
598 	struct task_desc  *task;
599 	struct perf_sched *sched;
600 	int fd;
601 };
602 
603 static void *thread_func(void *ctx)
604 {
605 	struct sched_thread_parms *parms = ctx;
606 	struct task_desc *this_task = parms->task;
607 	struct perf_sched *sched = parms->sched;
608 	u64 cpu_usage_0, cpu_usage_1;
609 	unsigned long i, ret;
610 	char comm2[22];
611 	int fd = parms->fd;
612 
613 	zfree(&parms);
614 
615 	sprintf(comm2, ":%s", this_task->comm);
616 	prctl(PR_SET_NAME, comm2);
617 	if (fd < 0)
618 		return NULL;
619 
620 	while (!sched->thread_funcs_exit) {
621 		ret = sem_post(&this_task->ready_for_work);
622 		BUG_ON(ret);
623 		mutex_lock(&sched->start_work_mutex);
624 		mutex_unlock(&sched->start_work_mutex);
625 
626 		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
627 
628 		for (i = 0; i < this_task->nr_events; i++) {
629 			this_task->curr_event = i;
630 			perf_sched__process_event(sched, this_task->atoms[i]);
631 		}
632 
633 		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
634 		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
635 		ret = sem_post(&this_task->work_done_sem);
636 		BUG_ON(ret);
637 
638 		mutex_lock(&sched->work_done_wait_mutex);
639 		mutex_unlock(&sched->work_done_wait_mutex);
640 	}
641 	return NULL;
642 }
643 
644 static void create_tasks(struct perf_sched *sched)
645 	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
646 	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
647 {
648 	struct task_desc *task;
649 	pthread_attr_t attr;
650 	unsigned long i;
651 	int err;
652 
653 	err = pthread_attr_init(&attr);
654 	BUG_ON(err);
655 	err = pthread_attr_setstacksize(&attr,
656 			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
657 	BUG_ON(err);
658 	mutex_lock(&sched->start_work_mutex);
659 	mutex_lock(&sched->work_done_wait_mutex);
660 	for (i = 0; i < sched->nr_tasks; i++) {
661 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
662 		BUG_ON(parms == NULL);
663 		parms->task = task = sched->tasks[i];
664 		parms->sched = sched;
665 		parms->fd = self_open_counters(sched, i);
666 		sem_init(&task->sleep_sem, 0, 0);
667 		sem_init(&task->ready_for_work, 0, 0);
668 		sem_init(&task->work_done_sem, 0, 0);
669 		task->curr_event = 0;
670 		err = pthread_create(&task->thread, &attr, thread_func, parms);
671 		BUG_ON(err);
672 	}
673 }
674 
675 static void destroy_tasks(struct perf_sched *sched)
676 	UNLOCK_FUNCTION(sched->start_work_mutex)
677 	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
678 {
679 	struct task_desc *task;
680 	unsigned long i;
681 	int err;
682 
683 	mutex_unlock(&sched->start_work_mutex);
684 	mutex_unlock(&sched->work_done_wait_mutex);
685 	/* Get rid of threads so they won't be upset by mutex destrunction */
686 	for (i = 0; i < sched->nr_tasks; i++) {
687 		task = sched->tasks[i];
688 		err = pthread_join(task->thread, NULL);
689 		BUG_ON(err);
690 		sem_destroy(&task->sleep_sem);
691 		sem_destroy(&task->ready_for_work);
692 		sem_destroy(&task->work_done_sem);
693 	}
694 }
695 
696 static void wait_for_tasks(struct perf_sched *sched)
697 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
698 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
699 {
700 	u64 cpu_usage_0, cpu_usage_1;
701 	struct task_desc *task;
702 	unsigned long i, ret;
703 
704 	sched->start_time = get_nsecs();
705 	sched->cpu_usage = 0;
706 	mutex_unlock(&sched->work_done_wait_mutex);
707 
708 	for (i = 0; i < sched->nr_tasks; i++) {
709 		task = sched->tasks[i];
710 		ret = sem_wait(&task->ready_for_work);
711 		BUG_ON(ret);
712 		sem_init(&task->ready_for_work, 0, 0);
713 	}
714 	mutex_lock(&sched->work_done_wait_mutex);
715 
716 	cpu_usage_0 = get_cpu_usage_nsec_parent();
717 
718 	mutex_unlock(&sched->start_work_mutex);
719 
720 	for (i = 0; i < sched->nr_tasks; i++) {
721 		task = sched->tasks[i];
722 		ret = sem_wait(&task->work_done_sem);
723 		BUG_ON(ret);
724 		sem_init(&task->work_done_sem, 0, 0);
725 		sched->cpu_usage += task->cpu_usage;
726 		task->cpu_usage = 0;
727 	}
728 
729 	cpu_usage_1 = get_cpu_usage_nsec_parent();
730 	if (!sched->runavg_cpu_usage)
731 		sched->runavg_cpu_usage = sched->cpu_usage;
732 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
733 
734 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
735 	if (!sched->runavg_parent_cpu_usage)
736 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
737 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
738 					 sched->parent_cpu_usage)/sched->replay_repeat;
739 
740 	mutex_lock(&sched->start_work_mutex);
741 
742 	for (i = 0; i < sched->nr_tasks; i++) {
743 		task = sched->tasks[i];
744 		sem_init(&task->sleep_sem, 0, 0);
745 		task->curr_event = 0;
746 	}
747 }
748 
749 static void run_one_test(struct perf_sched *sched)
750 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
751 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
752 {
753 	u64 T0, T1, delta, avg_delta, fluct;
754 
755 	T0 = get_nsecs();
756 	wait_for_tasks(sched);
757 	T1 = get_nsecs();
758 
759 	delta = T1 - T0;
760 	sched->sum_runtime += delta;
761 	sched->nr_runs++;
762 
763 	avg_delta = sched->sum_runtime / sched->nr_runs;
764 	if (delta < avg_delta)
765 		fluct = avg_delta - delta;
766 	else
767 		fluct = delta - avg_delta;
768 	sched->sum_fluct += fluct;
769 	if (!sched->run_avg)
770 		sched->run_avg = delta;
771 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
772 
773 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
774 
775 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
776 
777 	printf("cpu: %0.2f / %0.2f",
778 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
779 
780 #if 0
781 	/*
782 	 * rusage statistics done by the parent, these are less
783 	 * accurate than the sched->sum_exec_runtime based statistics:
784 	 */
785 	printf(" [%0.2f / %0.2f]",
786 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
787 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
788 #endif
789 
790 	printf("\n");
791 
792 	if (sched->nr_sleep_corrections)
793 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
794 	sched->nr_sleep_corrections = 0;
795 }
796 
797 static void test_calibrations(struct perf_sched *sched)
798 {
799 	u64 T0, T1;
800 
801 	T0 = get_nsecs();
802 	burn_nsecs(sched, NSEC_PER_MSEC);
803 	T1 = get_nsecs();
804 
805 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
806 
807 	T0 = get_nsecs();
808 	sleep_nsecs(NSEC_PER_MSEC);
809 	T1 = get_nsecs();
810 
811 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
812 }
813 
814 static int
815 replay_wakeup_event(struct perf_sched *sched,
816 		    struct evsel *evsel, struct perf_sample *sample,
817 		    struct machine *machine __maybe_unused)
818 {
819 	const char *comm = evsel__strval(evsel, sample, "comm");
820 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
821 	struct task_desc *waker, *wakee;
822 
823 	if (verbose > 0) {
824 		printf("sched_wakeup event %p\n", evsel);
825 
826 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
827 	}
828 
829 	waker = register_pid(sched, sample->tid, "<unknown>");
830 	wakee = register_pid(sched, pid, comm);
831 
832 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
833 	return 0;
834 }
835 
836 static int replay_switch_event(struct perf_sched *sched,
837 			       struct evsel *evsel,
838 			       struct perf_sample *sample,
839 			       struct machine *machine __maybe_unused)
840 {
841 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
842 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
843 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
844 		  next_pid = evsel__intval(evsel, sample, "next_pid");
845 	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
846 	struct task_desc *prev, __maybe_unused *next;
847 	u64 timestamp0, timestamp = sample->time;
848 	int cpu = sample->cpu;
849 	s64 delta;
850 
851 	if (verbose > 0)
852 		printf("sched_switch event %p\n", evsel);
853 
854 	if (cpu >= MAX_CPUS || cpu < 0)
855 		return 0;
856 
857 	timestamp0 = sched->cpu_last_switched[cpu];
858 	if (timestamp0)
859 		delta = timestamp - timestamp0;
860 	else
861 		delta = 0;
862 
863 	if (delta < 0) {
864 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
865 		return -1;
866 	}
867 
868 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
869 		 prev_comm, prev_pid, next_comm, next_pid, delta);
870 
871 	prev = register_pid(sched, prev_pid, prev_comm);
872 	next = register_pid(sched, next_pid, next_comm);
873 
874 	sched->cpu_last_switched[cpu] = timestamp;
875 
876 	add_sched_event_run(sched, prev, timestamp, delta);
877 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
878 
879 	return 0;
880 }
881 
882 static int replay_fork_event(struct perf_sched *sched,
883 			     union perf_event *event,
884 			     struct machine *machine)
885 {
886 	struct thread *child, *parent;
887 
888 	child = machine__findnew_thread(machine, event->fork.pid,
889 					event->fork.tid);
890 	parent = machine__findnew_thread(machine, event->fork.ppid,
891 					 event->fork.ptid);
892 
893 	if (child == NULL || parent == NULL) {
894 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
895 				 child, parent);
896 		goto out_put;
897 	}
898 
899 	if (verbose > 0) {
900 		printf("fork event\n");
901 		printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
902 		printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
903 	}
904 
905 	register_pid(sched, thread__tid(parent), thread__comm_str(parent));
906 	register_pid(sched, thread__tid(child), thread__comm_str(child));
907 out_put:
908 	thread__put(child);
909 	thread__put(parent);
910 	return 0;
911 }
912 
913 struct sort_dimension {
914 	const char		*name;
915 	sort_fn_t		cmp;
916 	struct list_head	list;
917 };
918 
919 /*
920  * handle runtime stats saved per thread
921  */
922 static struct thread_runtime *thread__init_runtime(struct thread *thread)
923 {
924 	struct thread_runtime *r;
925 
926 	r = zalloc(sizeof(struct thread_runtime));
927 	if (!r)
928 		return NULL;
929 
930 	init_stats(&r->run_stats);
931 	thread__set_priv(thread, r);
932 
933 	return r;
934 }
935 
936 static struct thread_runtime *thread__get_runtime(struct thread *thread)
937 {
938 	struct thread_runtime *tr;
939 
940 	tr = thread__priv(thread);
941 	if (tr == NULL) {
942 		tr = thread__init_runtime(thread);
943 		if (tr == NULL)
944 			pr_debug("Failed to malloc memory for runtime data.\n");
945 	}
946 
947 	return tr;
948 }
949 
950 static int
951 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
952 {
953 	struct sort_dimension *sort;
954 	int ret = 0;
955 
956 	BUG_ON(list_empty(list));
957 
958 	list_for_each_entry(sort, list, list) {
959 		ret = sort->cmp(l, r);
960 		if (ret)
961 			return ret;
962 	}
963 
964 	return ret;
965 }
966 
967 static struct work_atoms *
968 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
969 			 struct list_head *sort_list)
970 {
971 	struct rb_node *node = root->rb_root.rb_node;
972 	struct work_atoms key = { .thread = thread };
973 
974 	while (node) {
975 		struct work_atoms *atoms;
976 		int cmp;
977 
978 		atoms = container_of(node, struct work_atoms, node);
979 
980 		cmp = thread_lat_cmp(sort_list, &key, atoms);
981 		if (cmp > 0)
982 			node = node->rb_left;
983 		else if (cmp < 0)
984 			node = node->rb_right;
985 		else {
986 			BUG_ON(thread != atoms->thread);
987 			return atoms;
988 		}
989 	}
990 	return NULL;
991 }
992 
993 static void
994 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
995 			 struct list_head *sort_list)
996 {
997 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
998 	bool leftmost = true;
999 
1000 	while (*new) {
1001 		struct work_atoms *this;
1002 		int cmp;
1003 
1004 		this = container_of(*new, struct work_atoms, node);
1005 		parent = *new;
1006 
1007 		cmp = thread_lat_cmp(sort_list, data, this);
1008 
1009 		if (cmp > 0)
1010 			new = &((*new)->rb_left);
1011 		else {
1012 			new = &((*new)->rb_right);
1013 			leftmost = false;
1014 		}
1015 	}
1016 
1017 	rb_link_node(&data->node, parent, new);
1018 	rb_insert_color_cached(&data->node, root, leftmost);
1019 }
1020 
1021 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1022 {
1023 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1024 	if (!atoms) {
1025 		pr_err("No memory at %s\n", __func__);
1026 		return -1;
1027 	}
1028 
1029 	atoms->thread = thread__get(thread);
1030 	INIT_LIST_HEAD(&atoms->work_list);
1031 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1032 	return 0;
1033 }
1034 
1035 static int
1036 add_sched_out_event(struct work_atoms *atoms,
1037 		    char run_state,
1038 		    u64 timestamp)
1039 {
1040 	struct work_atom *atom = zalloc(sizeof(*atom));
1041 	if (!atom) {
1042 		pr_err("Non memory at %s", __func__);
1043 		return -1;
1044 	}
1045 
1046 	atom->sched_out_time = timestamp;
1047 
1048 	if (run_state == 'R') {
1049 		atom->state = THREAD_WAIT_CPU;
1050 		atom->wake_up_time = atom->sched_out_time;
1051 	}
1052 
1053 	list_add_tail(&atom->list, &atoms->work_list);
1054 	return 0;
1055 }
1056 
1057 static void
1058 add_runtime_event(struct work_atoms *atoms, u64 delta,
1059 		  u64 timestamp __maybe_unused)
1060 {
1061 	struct work_atom *atom;
1062 
1063 	BUG_ON(list_empty(&atoms->work_list));
1064 
1065 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1066 
1067 	atom->runtime += delta;
1068 	atoms->total_runtime += delta;
1069 }
1070 
1071 static void
1072 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1073 {
1074 	struct work_atom *atom;
1075 	u64 delta;
1076 
1077 	if (list_empty(&atoms->work_list))
1078 		return;
1079 
1080 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1081 
1082 	if (atom->state != THREAD_WAIT_CPU)
1083 		return;
1084 
1085 	if (timestamp < atom->wake_up_time) {
1086 		atom->state = THREAD_IGNORE;
1087 		return;
1088 	}
1089 
1090 	atom->state = THREAD_SCHED_IN;
1091 	atom->sched_in_time = timestamp;
1092 
1093 	delta = atom->sched_in_time - atom->wake_up_time;
1094 	atoms->total_lat += delta;
1095 	if (delta > atoms->max_lat) {
1096 		atoms->max_lat = delta;
1097 		atoms->max_lat_start = atom->wake_up_time;
1098 		atoms->max_lat_end = timestamp;
1099 	}
1100 	atoms->nb_atoms++;
1101 }
1102 
1103 static int latency_switch_event(struct perf_sched *sched,
1104 				struct evsel *evsel,
1105 				struct perf_sample *sample,
1106 				struct machine *machine)
1107 {
1108 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1109 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1110 	const char prev_state = evsel__taskstate(evsel, sample, "prev_state");
1111 	struct work_atoms *out_events, *in_events;
1112 	struct thread *sched_out, *sched_in;
1113 	u64 timestamp0, timestamp = sample->time;
1114 	int cpu = sample->cpu, err = -1;
1115 	s64 delta;
1116 
1117 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1118 
1119 	timestamp0 = sched->cpu_last_switched[cpu];
1120 	sched->cpu_last_switched[cpu] = timestamp;
1121 	if (timestamp0)
1122 		delta = timestamp - timestamp0;
1123 	else
1124 		delta = 0;
1125 
1126 	if (delta < 0) {
1127 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1128 		return -1;
1129 	}
1130 
1131 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1132 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1133 	if (sched_out == NULL || sched_in == NULL)
1134 		goto out_put;
1135 
1136 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1137 	if (!out_events) {
1138 		if (thread_atoms_insert(sched, sched_out))
1139 			goto out_put;
1140 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1141 		if (!out_events) {
1142 			pr_err("out-event: Internal tree error");
1143 			goto out_put;
1144 		}
1145 	}
1146 	if (add_sched_out_event(out_events, prev_state, timestamp))
1147 		return -1;
1148 
1149 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1150 	if (!in_events) {
1151 		if (thread_atoms_insert(sched, sched_in))
1152 			goto out_put;
1153 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1154 		if (!in_events) {
1155 			pr_err("in-event: Internal tree error");
1156 			goto out_put;
1157 		}
1158 		/*
1159 		 * Take came in we have not heard about yet,
1160 		 * add in an initial atom in runnable state:
1161 		 */
1162 		if (add_sched_out_event(in_events, 'R', timestamp))
1163 			goto out_put;
1164 	}
1165 	add_sched_in_event(in_events, timestamp);
1166 	err = 0;
1167 out_put:
1168 	thread__put(sched_out);
1169 	thread__put(sched_in);
1170 	return err;
1171 }
1172 
1173 static int latency_runtime_event(struct perf_sched *sched,
1174 				 struct evsel *evsel,
1175 				 struct perf_sample *sample,
1176 				 struct machine *machine)
1177 {
1178 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1179 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1180 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1181 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1182 	u64 timestamp = sample->time;
1183 	int cpu = sample->cpu, err = -1;
1184 
1185 	if (thread == NULL)
1186 		return -1;
1187 
1188 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1189 	if (!atoms) {
1190 		if (thread_atoms_insert(sched, thread))
1191 			goto out_put;
1192 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1193 		if (!atoms) {
1194 			pr_err("in-event: Internal tree error");
1195 			goto out_put;
1196 		}
1197 		if (add_sched_out_event(atoms, 'R', timestamp))
1198 			goto out_put;
1199 	}
1200 
1201 	add_runtime_event(atoms, runtime, timestamp);
1202 	err = 0;
1203 out_put:
1204 	thread__put(thread);
1205 	return err;
1206 }
1207 
1208 static int latency_wakeup_event(struct perf_sched *sched,
1209 				struct evsel *evsel,
1210 				struct perf_sample *sample,
1211 				struct machine *machine)
1212 {
1213 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1214 	struct work_atoms *atoms;
1215 	struct work_atom *atom;
1216 	struct thread *wakee;
1217 	u64 timestamp = sample->time;
1218 	int err = -1;
1219 
1220 	wakee = machine__findnew_thread(machine, -1, pid);
1221 	if (wakee == NULL)
1222 		return -1;
1223 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1224 	if (!atoms) {
1225 		if (thread_atoms_insert(sched, wakee))
1226 			goto out_put;
1227 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1228 		if (!atoms) {
1229 			pr_err("wakeup-event: Internal tree error");
1230 			goto out_put;
1231 		}
1232 		if (add_sched_out_event(atoms, 'S', timestamp))
1233 			goto out_put;
1234 	}
1235 
1236 	BUG_ON(list_empty(&atoms->work_list));
1237 
1238 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1239 
1240 	/*
1241 	 * As we do not guarantee the wakeup event happens when
1242 	 * task is out of run queue, also may happen when task is
1243 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1244 	 * then we should not set the ->wake_up_time when wake up a
1245 	 * task which is on run queue.
1246 	 *
1247 	 * You WILL be missing events if you've recorded only
1248 	 * one CPU, or are only looking at only one, so don't
1249 	 * skip in this case.
1250 	 */
1251 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1252 		goto out_ok;
1253 
1254 	sched->nr_timestamps++;
1255 	if (atom->sched_out_time > timestamp) {
1256 		sched->nr_unordered_timestamps++;
1257 		goto out_ok;
1258 	}
1259 
1260 	atom->state = THREAD_WAIT_CPU;
1261 	atom->wake_up_time = timestamp;
1262 out_ok:
1263 	err = 0;
1264 out_put:
1265 	thread__put(wakee);
1266 	return err;
1267 }
1268 
1269 static int latency_migrate_task_event(struct perf_sched *sched,
1270 				      struct evsel *evsel,
1271 				      struct perf_sample *sample,
1272 				      struct machine *machine)
1273 {
1274 	const u32 pid = evsel__intval(evsel, sample, "pid");
1275 	u64 timestamp = sample->time;
1276 	struct work_atoms *atoms;
1277 	struct work_atom *atom;
1278 	struct thread *migrant;
1279 	int err = -1;
1280 
1281 	/*
1282 	 * Only need to worry about migration when profiling one CPU.
1283 	 */
1284 	if (sched->profile_cpu == -1)
1285 		return 0;
1286 
1287 	migrant = machine__findnew_thread(machine, -1, pid);
1288 	if (migrant == NULL)
1289 		return -1;
1290 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1291 	if (!atoms) {
1292 		if (thread_atoms_insert(sched, migrant))
1293 			goto out_put;
1294 		register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1295 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1296 		if (!atoms) {
1297 			pr_err("migration-event: Internal tree error");
1298 			goto out_put;
1299 		}
1300 		if (add_sched_out_event(atoms, 'R', timestamp))
1301 			goto out_put;
1302 	}
1303 
1304 	BUG_ON(list_empty(&atoms->work_list));
1305 
1306 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1307 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1308 
1309 	sched->nr_timestamps++;
1310 
1311 	if (atom->sched_out_time > timestamp)
1312 		sched->nr_unordered_timestamps++;
1313 	err = 0;
1314 out_put:
1315 	thread__put(migrant);
1316 	return err;
1317 }
1318 
1319 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1320 {
1321 	int i;
1322 	int ret;
1323 	u64 avg;
1324 	char max_lat_start[32], max_lat_end[32];
1325 
1326 	if (!work_list->nb_atoms)
1327 		return;
1328 	/*
1329 	 * Ignore idle threads:
1330 	 */
1331 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1332 		return;
1333 
1334 	sched->all_runtime += work_list->total_runtime;
1335 	sched->all_count   += work_list->nb_atoms;
1336 
1337 	if (work_list->num_merged > 1) {
1338 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1339 			     work_list->num_merged);
1340 	} else {
1341 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1342 			     thread__tid(work_list->thread));
1343 	}
1344 
1345 	for (i = 0; i < 24 - ret; i++)
1346 		printf(" ");
1347 
1348 	avg = work_list->total_lat / work_list->nb_atoms;
1349 	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1350 	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1351 
1352 	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1353 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1354 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1355 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1356 		 max_lat_start, max_lat_end);
1357 }
1358 
1359 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1360 {
1361 	pid_t l_tid, r_tid;
1362 
1363 	if (RC_CHK_EQUAL(l->thread, r->thread))
1364 		return 0;
1365 	l_tid = thread__tid(l->thread);
1366 	r_tid = thread__tid(r->thread);
1367 	if (l_tid < r_tid)
1368 		return -1;
1369 	if (l_tid > r_tid)
1370 		return 1;
1371 	return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1372 }
1373 
1374 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1375 {
1376 	u64 avgl, avgr;
1377 
1378 	if (!l->nb_atoms)
1379 		return -1;
1380 
1381 	if (!r->nb_atoms)
1382 		return 1;
1383 
1384 	avgl = l->total_lat / l->nb_atoms;
1385 	avgr = r->total_lat / r->nb_atoms;
1386 
1387 	if (avgl < avgr)
1388 		return -1;
1389 	if (avgl > avgr)
1390 		return 1;
1391 
1392 	return 0;
1393 }
1394 
1395 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1396 {
1397 	if (l->max_lat < r->max_lat)
1398 		return -1;
1399 	if (l->max_lat > r->max_lat)
1400 		return 1;
1401 
1402 	return 0;
1403 }
1404 
1405 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1406 {
1407 	if (l->nb_atoms < r->nb_atoms)
1408 		return -1;
1409 	if (l->nb_atoms > r->nb_atoms)
1410 		return 1;
1411 
1412 	return 0;
1413 }
1414 
1415 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1416 {
1417 	if (l->total_runtime < r->total_runtime)
1418 		return -1;
1419 	if (l->total_runtime > r->total_runtime)
1420 		return 1;
1421 
1422 	return 0;
1423 }
1424 
1425 static int sort_dimension__add(const char *tok, struct list_head *list)
1426 {
1427 	size_t i;
1428 	static struct sort_dimension avg_sort_dimension = {
1429 		.name = "avg",
1430 		.cmp  = avg_cmp,
1431 	};
1432 	static struct sort_dimension max_sort_dimension = {
1433 		.name = "max",
1434 		.cmp  = max_cmp,
1435 	};
1436 	static struct sort_dimension pid_sort_dimension = {
1437 		.name = "pid",
1438 		.cmp  = pid_cmp,
1439 	};
1440 	static struct sort_dimension runtime_sort_dimension = {
1441 		.name = "runtime",
1442 		.cmp  = runtime_cmp,
1443 	};
1444 	static struct sort_dimension switch_sort_dimension = {
1445 		.name = "switch",
1446 		.cmp  = switch_cmp,
1447 	};
1448 	struct sort_dimension *available_sorts[] = {
1449 		&pid_sort_dimension,
1450 		&avg_sort_dimension,
1451 		&max_sort_dimension,
1452 		&switch_sort_dimension,
1453 		&runtime_sort_dimension,
1454 	};
1455 
1456 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1457 		if (!strcmp(available_sorts[i]->name, tok)) {
1458 			list_add_tail(&available_sorts[i]->list, list);
1459 
1460 			return 0;
1461 		}
1462 	}
1463 
1464 	return -1;
1465 }
1466 
1467 static void perf_sched__sort_lat(struct perf_sched *sched)
1468 {
1469 	struct rb_node *node;
1470 	struct rb_root_cached *root = &sched->atom_root;
1471 again:
1472 	for (;;) {
1473 		struct work_atoms *data;
1474 		node = rb_first_cached(root);
1475 		if (!node)
1476 			break;
1477 
1478 		rb_erase_cached(node, root);
1479 		data = rb_entry(node, struct work_atoms, node);
1480 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1481 	}
1482 	if (root == &sched->atom_root) {
1483 		root = &sched->merged_atom_root;
1484 		goto again;
1485 	}
1486 }
1487 
1488 static int process_sched_wakeup_event(struct perf_tool *tool,
1489 				      struct evsel *evsel,
1490 				      struct perf_sample *sample,
1491 				      struct machine *machine)
1492 {
1493 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1494 
1495 	if (sched->tp_handler->wakeup_event)
1496 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1497 
1498 	return 0;
1499 }
1500 
1501 static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1502 				      struct evsel *evsel __maybe_unused,
1503 				      struct perf_sample *sample __maybe_unused,
1504 				      struct machine *machine __maybe_unused)
1505 {
1506 	return 0;
1507 }
1508 
1509 union map_priv {
1510 	void	*ptr;
1511 	bool	 color;
1512 };
1513 
1514 static bool thread__has_color(struct thread *thread)
1515 {
1516 	union map_priv priv = {
1517 		.ptr = thread__priv(thread),
1518 	};
1519 
1520 	return priv.color;
1521 }
1522 
1523 static struct thread*
1524 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1525 {
1526 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1527 	union map_priv priv = {
1528 		.color = false,
1529 	};
1530 
1531 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1532 		return thread;
1533 
1534 	if (thread_map__has(sched->map.color_pids, tid))
1535 		priv.color = true;
1536 
1537 	thread__set_priv(thread, priv.ptr);
1538 	return thread;
1539 }
1540 
1541 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1542 			    struct perf_sample *sample, struct machine *machine)
1543 {
1544 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1545 	struct thread *sched_in;
1546 	struct thread_runtime *tr;
1547 	int new_shortname;
1548 	u64 timestamp0, timestamp = sample->time;
1549 	s64 delta;
1550 	int i;
1551 	struct perf_cpu this_cpu = {
1552 		.cpu = sample->cpu,
1553 	};
1554 	int cpus_nr;
1555 	bool new_cpu = false;
1556 	const char *color = PERF_COLOR_NORMAL;
1557 	char stimestamp[32];
1558 
1559 	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1560 
1561 	if (this_cpu.cpu > sched->max_cpu.cpu)
1562 		sched->max_cpu = this_cpu;
1563 
1564 	if (sched->map.comp) {
1565 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1566 		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1567 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1568 			new_cpu = true;
1569 		}
1570 	} else
1571 		cpus_nr = sched->max_cpu.cpu;
1572 
1573 	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1574 	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1575 	if (timestamp0)
1576 		delta = timestamp - timestamp0;
1577 	else
1578 		delta = 0;
1579 
1580 	if (delta < 0) {
1581 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1582 		return -1;
1583 	}
1584 
1585 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1586 	if (sched_in == NULL)
1587 		return -1;
1588 
1589 	tr = thread__get_runtime(sched_in);
1590 	if (tr == NULL) {
1591 		thread__put(sched_in);
1592 		return -1;
1593 	}
1594 
1595 	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1596 
1597 	printf("  ");
1598 
1599 	new_shortname = 0;
1600 	if (!tr->shortname[0]) {
1601 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1602 			/*
1603 			 * Don't allocate a letter-number for swapper:0
1604 			 * as a shortname. Instead, we use '.' for it.
1605 			 */
1606 			tr->shortname[0] = '.';
1607 			tr->shortname[1] = ' ';
1608 		} else {
1609 			tr->shortname[0] = sched->next_shortname1;
1610 			tr->shortname[1] = sched->next_shortname2;
1611 
1612 			if (sched->next_shortname1 < 'Z') {
1613 				sched->next_shortname1++;
1614 			} else {
1615 				sched->next_shortname1 = 'A';
1616 				if (sched->next_shortname2 < '9')
1617 					sched->next_shortname2++;
1618 				else
1619 					sched->next_shortname2 = '0';
1620 			}
1621 		}
1622 		new_shortname = 1;
1623 	}
1624 
1625 	for (i = 0; i < cpus_nr; i++) {
1626 		struct perf_cpu cpu = {
1627 			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1628 		};
1629 		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1630 		struct thread_runtime *curr_tr;
1631 		const char *pid_color = color;
1632 		const char *cpu_color = color;
1633 
1634 		if (curr_thread && thread__has_color(curr_thread))
1635 			pid_color = COLOR_PIDS;
1636 
1637 		if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1638 			continue;
1639 
1640 		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1641 			cpu_color = COLOR_CPUS;
1642 
1643 		if (cpu.cpu != this_cpu.cpu)
1644 			color_fprintf(stdout, color, " ");
1645 		else
1646 			color_fprintf(stdout, cpu_color, "*");
1647 
1648 		if (sched->curr_thread[cpu.cpu]) {
1649 			curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1650 			if (curr_tr == NULL) {
1651 				thread__put(sched_in);
1652 				return -1;
1653 			}
1654 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1655 		} else
1656 			color_fprintf(stdout, color, "   ");
1657 	}
1658 
1659 	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1660 		goto out;
1661 
1662 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1663 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1664 	if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1665 		const char *pid_color = color;
1666 
1667 		if (thread__has_color(sched_in))
1668 			pid_color = COLOR_PIDS;
1669 
1670 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1671 			tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1672 		tr->comm_changed = false;
1673 	}
1674 
1675 	if (sched->map.comp && new_cpu)
1676 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1677 
1678 out:
1679 	color_fprintf(stdout, color, "\n");
1680 
1681 	thread__put(sched_in);
1682 
1683 	return 0;
1684 }
1685 
1686 static int process_sched_switch_event(struct perf_tool *tool,
1687 				      struct evsel *evsel,
1688 				      struct perf_sample *sample,
1689 				      struct machine *machine)
1690 {
1691 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1692 	int this_cpu = sample->cpu, err = 0;
1693 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1694 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1695 
1696 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1697 		/*
1698 		 * Are we trying to switch away a PID that is
1699 		 * not current?
1700 		 */
1701 		if (sched->curr_pid[this_cpu] != prev_pid)
1702 			sched->nr_context_switch_bugs++;
1703 	}
1704 
1705 	if (sched->tp_handler->switch_event)
1706 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1707 
1708 	sched->curr_pid[this_cpu] = next_pid;
1709 	return err;
1710 }
1711 
1712 static int process_sched_runtime_event(struct perf_tool *tool,
1713 				       struct evsel *evsel,
1714 				       struct perf_sample *sample,
1715 				       struct machine *machine)
1716 {
1717 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1718 
1719 	if (sched->tp_handler->runtime_event)
1720 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1721 
1722 	return 0;
1723 }
1724 
1725 static int perf_sched__process_fork_event(struct perf_tool *tool,
1726 					  union perf_event *event,
1727 					  struct perf_sample *sample,
1728 					  struct machine *machine)
1729 {
1730 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731 
1732 	/* run the fork event through the perf machinery */
1733 	perf_event__process_fork(tool, event, sample, machine);
1734 
1735 	/* and then run additional processing needed for this command */
1736 	if (sched->tp_handler->fork_event)
1737 		return sched->tp_handler->fork_event(sched, event, machine);
1738 
1739 	return 0;
1740 }
1741 
1742 static int process_sched_migrate_task_event(struct perf_tool *tool,
1743 					    struct evsel *evsel,
1744 					    struct perf_sample *sample,
1745 					    struct machine *machine)
1746 {
1747 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1748 
1749 	if (sched->tp_handler->migrate_task_event)
1750 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1751 
1752 	return 0;
1753 }
1754 
1755 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1756 				  struct evsel *evsel,
1757 				  struct perf_sample *sample,
1758 				  struct machine *machine);
1759 
1760 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1761 						 union perf_event *event __maybe_unused,
1762 						 struct perf_sample *sample,
1763 						 struct evsel *evsel,
1764 						 struct machine *machine)
1765 {
1766 	int err = 0;
1767 
1768 	if (evsel->handler != NULL) {
1769 		tracepoint_handler f = evsel->handler;
1770 		err = f(tool, evsel, sample, machine);
1771 	}
1772 
1773 	return err;
1774 }
1775 
1776 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1777 				    union perf_event *event,
1778 				    struct perf_sample *sample,
1779 				    struct machine *machine)
1780 {
1781 	struct thread *thread;
1782 	struct thread_runtime *tr;
1783 	int err;
1784 
1785 	err = perf_event__process_comm(tool, event, sample, machine);
1786 	if (err)
1787 		return err;
1788 
1789 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1790 	if (!thread) {
1791 		pr_err("Internal error: can't find thread\n");
1792 		return -1;
1793 	}
1794 
1795 	tr = thread__get_runtime(thread);
1796 	if (tr == NULL) {
1797 		thread__put(thread);
1798 		return -1;
1799 	}
1800 
1801 	tr->comm_changed = true;
1802 	thread__put(thread);
1803 
1804 	return 0;
1805 }
1806 
1807 static int perf_sched__read_events(struct perf_sched *sched)
1808 {
1809 	struct evsel_str_handler handlers[] = {
1810 		{ "sched:sched_switch",	      process_sched_switch_event, },
1811 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1812 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1813 		{ "sched:sched_waking",	      process_sched_wakeup_event, },
1814 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1815 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1816 	};
1817 	struct perf_session *session;
1818 	struct perf_data data = {
1819 		.path  = input_name,
1820 		.mode  = PERF_DATA_MODE_READ,
1821 		.force = sched->force,
1822 	};
1823 	int rc = -1;
1824 
1825 	session = perf_session__new(&data, &sched->tool);
1826 	if (IS_ERR(session)) {
1827 		pr_debug("Error creating perf session");
1828 		return PTR_ERR(session);
1829 	}
1830 
1831 	symbol__init(&session->header.env);
1832 
1833 	/* prefer sched_waking if it is captured */
1834 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1835 		handlers[2].handler = process_sched_wakeup_ignore;
1836 
1837 	if (perf_session__set_tracepoints_handlers(session, handlers))
1838 		goto out_delete;
1839 
1840 	if (perf_session__has_traces(session, "record -R")) {
1841 		int err = perf_session__process_events(session);
1842 		if (err) {
1843 			pr_err("Failed to process events, error %d", err);
1844 			goto out_delete;
1845 		}
1846 
1847 		sched->nr_events      = session->evlist->stats.nr_events[0];
1848 		sched->nr_lost_events = session->evlist->stats.total_lost;
1849 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1850 	}
1851 
1852 	rc = 0;
1853 out_delete:
1854 	perf_session__delete(session);
1855 	return rc;
1856 }
1857 
1858 /*
1859  * scheduling times are printed as msec.usec
1860  */
1861 static inline void print_sched_time(unsigned long long nsecs, int width)
1862 {
1863 	unsigned long msecs;
1864 	unsigned long usecs;
1865 
1866 	msecs  = nsecs / NSEC_PER_MSEC;
1867 	nsecs -= msecs * NSEC_PER_MSEC;
1868 	usecs  = nsecs / NSEC_PER_USEC;
1869 	printf("%*lu.%03lu ", width, msecs, usecs);
1870 }
1871 
1872 /*
1873  * returns runtime data for event, allocating memory for it the
1874  * first time it is used.
1875  */
1876 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1877 {
1878 	struct evsel_runtime *r = evsel->priv;
1879 
1880 	if (r == NULL) {
1881 		r = zalloc(sizeof(struct evsel_runtime));
1882 		evsel->priv = r;
1883 	}
1884 
1885 	return r;
1886 }
1887 
1888 /*
1889  * save last time event was seen per cpu
1890  */
1891 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1892 {
1893 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1894 
1895 	if (r == NULL)
1896 		return;
1897 
1898 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1899 		int i, n = __roundup_pow_of_two(cpu+1);
1900 		void *p = r->last_time;
1901 
1902 		p = realloc(r->last_time, n * sizeof(u64));
1903 		if (!p)
1904 			return;
1905 
1906 		r->last_time = p;
1907 		for (i = r->ncpu; i < n; ++i)
1908 			r->last_time[i] = (u64) 0;
1909 
1910 		r->ncpu = n;
1911 	}
1912 
1913 	r->last_time[cpu] = timestamp;
1914 }
1915 
1916 /* returns last time this event was seen on the given cpu */
1917 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1918 {
1919 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1920 
1921 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1922 		return 0;
1923 
1924 	return r->last_time[cpu];
1925 }
1926 
1927 static int comm_width = 30;
1928 
1929 static char *timehist_get_commstr(struct thread *thread)
1930 {
1931 	static char str[32];
1932 	const char *comm = thread__comm_str(thread);
1933 	pid_t tid = thread__tid(thread);
1934 	pid_t pid = thread__pid(thread);
1935 	int n;
1936 
1937 	if (pid == 0)
1938 		n = scnprintf(str, sizeof(str), "%s", comm);
1939 
1940 	else if (tid != pid)
1941 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1942 
1943 	else
1944 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1945 
1946 	if (n > comm_width)
1947 		comm_width = n;
1948 
1949 	return str;
1950 }
1951 
1952 static void timehist_header(struct perf_sched *sched)
1953 {
1954 	u32 ncpus = sched->max_cpu.cpu + 1;
1955 	u32 i, j;
1956 
1957 	printf("%15s %6s ", "time", "cpu");
1958 
1959 	if (sched->show_cpu_visual) {
1960 		printf(" ");
1961 		for (i = 0, j = 0; i < ncpus; ++i) {
1962 			printf("%x", j++);
1963 			if (j > 15)
1964 				j = 0;
1965 		}
1966 		printf(" ");
1967 	}
1968 
1969 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1970 		"task name", "wait time", "sch delay", "run time");
1971 
1972 	if (sched->show_state)
1973 		printf("  %s", "state");
1974 
1975 	printf("\n");
1976 
1977 	/*
1978 	 * units row
1979 	 */
1980 	printf("%15s %-6s ", "", "");
1981 
1982 	if (sched->show_cpu_visual)
1983 		printf(" %*s ", ncpus, "");
1984 
1985 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1986 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1987 
1988 	if (sched->show_state)
1989 		printf("  %5s", "");
1990 
1991 	printf("\n");
1992 
1993 	/*
1994 	 * separator
1995 	 */
1996 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1997 
1998 	if (sched->show_cpu_visual)
1999 		printf(" %.*s ", ncpus, graph_dotted_line);
2000 
2001 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2002 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
2003 		graph_dotted_line);
2004 
2005 	if (sched->show_state)
2006 		printf("  %.5s", graph_dotted_line);
2007 
2008 	printf("\n");
2009 }
2010 
2011 static void timehist_print_sample(struct perf_sched *sched,
2012 				  struct evsel *evsel,
2013 				  struct perf_sample *sample,
2014 				  struct addr_location *al,
2015 				  struct thread *thread,
2016 				  u64 t, const char state)
2017 {
2018 	struct thread_runtime *tr = thread__priv(thread);
2019 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2020 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2021 	u32 max_cpus = sched->max_cpu.cpu + 1;
2022 	char tstr[64];
2023 	char nstr[30];
2024 	u64 wait_time;
2025 
2026 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2027 		return;
2028 
2029 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2030 	printf("%15s [%04d] ", tstr, sample->cpu);
2031 
2032 	if (sched->show_cpu_visual) {
2033 		u32 i;
2034 		char c;
2035 
2036 		printf(" ");
2037 		for (i = 0; i < max_cpus; ++i) {
2038 			/* flag idle times with 'i'; others are sched events */
2039 			if (i == sample->cpu)
2040 				c = (thread__tid(thread) == 0) ? 'i' : 's';
2041 			else
2042 				c = ' ';
2043 			printf("%c", c);
2044 		}
2045 		printf(" ");
2046 	}
2047 
2048 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2049 
2050 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2051 	print_sched_time(wait_time, 6);
2052 
2053 	print_sched_time(tr->dt_delay, 6);
2054 	print_sched_time(tr->dt_run, 6);
2055 
2056 	if (sched->show_state)
2057 		printf(" %5c ", thread__tid(thread) == 0 ? 'I' : state);
2058 
2059 	if (sched->show_next) {
2060 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2061 		printf(" %-*s", comm_width, nstr);
2062 	}
2063 
2064 	if (sched->show_wakeups && !sched->show_next)
2065 		printf("  %-*s", comm_width, "");
2066 
2067 	if (thread__tid(thread) == 0)
2068 		goto out;
2069 
2070 	if (sched->show_callchain)
2071 		printf("  ");
2072 
2073 	sample__fprintf_sym(sample, al, 0,
2074 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2075 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2076 			    EVSEL__PRINT_SKIP_IGNORED,
2077 			    get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2078 
2079 out:
2080 	printf("\n");
2081 }
2082 
2083 /*
2084  * Explanation of delta-time stats:
2085  *
2086  *            t = time of current schedule out event
2087  *        tprev = time of previous sched out event
2088  *                also time of schedule-in event for current task
2089  *    last_time = time of last sched change event for current task
2090  *                (i.e, time process was last scheduled out)
2091  * ready_to_run = time of wakeup for current task
2092  *
2093  * -----|------------|------------|------------|------
2094  *    last         ready        tprev          t
2095  *    time         to run
2096  *
2097  *      |-------- dt_wait --------|
2098  *                   |- dt_delay -|-- dt_run --|
2099  *
2100  *   dt_run = run time of current task
2101  *  dt_wait = time between last schedule out event for task and tprev
2102  *            represents time spent off the cpu
2103  * dt_delay = time between wakeup and schedule-in of task
2104  */
2105 
2106 static void timehist_update_runtime_stats(struct thread_runtime *r,
2107 					 u64 t, u64 tprev)
2108 {
2109 	r->dt_delay   = 0;
2110 	r->dt_sleep   = 0;
2111 	r->dt_iowait  = 0;
2112 	r->dt_preempt = 0;
2113 	r->dt_run     = 0;
2114 
2115 	if (tprev) {
2116 		r->dt_run = t - tprev;
2117 		if (r->ready_to_run) {
2118 			if (r->ready_to_run > tprev)
2119 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2120 			else
2121 				r->dt_delay = tprev - r->ready_to_run;
2122 		}
2123 
2124 		if (r->last_time > tprev)
2125 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2126 		else if (r->last_time) {
2127 			u64 dt_wait = tprev - r->last_time;
2128 
2129 			if (r->last_state == 'R')
2130 				r->dt_preempt = dt_wait;
2131 			else if (r->last_state == 'D')
2132 				r->dt_iowait = dt_wait;
2133 			else
2134 				r->dt_sleep = dt_wait;
2135 		}
2136 	}
2137 
2138 	update_stats(&r->run_stats, r->dt_run);
2139 
2140 	r->total_run_time     += r->dt_run;
2141 	r->total_delay_time   += r->dt_delay;
2142 	r->total_sleep_time   += r->dt_sleep;
2143 	r->total_iowait_time  += r->dt_iowait;
2144 	r->total_preempt_time += r->dt_preempt;
2145 }
2146 
2147 static bool is_idle_sample(struct perf_sample *sample,
2148 			   struct evsel *evsel)
2149 {
2150 	/* pid 0 == swapper == idle task */
2151 	if (evsel__name_is(evsel, "sched:sched_switch"))
2152 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2153 
2154 	return sample->pid == 0;
2155 }
2156 
2157 static void save_task_callchain(struct perf_sched *sched,
2158 				struct perf_sample *sample,
2159 				struct evsel *evsel,
2160 				struct machine *machine)
2161 {
2162 	struct callchain_cursor *cursor;
2163 	struct thread *thread;
2164 
2165 	/* want main thread for process - has maps */
2166 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2167 	if (thread == NULL) {
2168 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2169 		return;
2170 	}
2171 
2172 	if (!sched->show_callchain || sample->callchain == NULL)
2173 		return;
2174 
2175 	cursor = get_tls_callchain_cursor();
2176 
2177 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2178 				      NULL, NULL, sched->max_stack + 2) != 0) {
2179 		if (verbose > 0)
2180 			pr_err("Failed to resolve callchain. Skipping\n");
2181 
2182 		return;
2183 	}
2184 
2185 	callchain_cursor_commit(cursor);
2186 
2187 	while (true) {
2188 		struct callchain_cursor_node *node;
2189 		struct symbol *sym;
2190 
2191 		node = callchain_cursor_current(cursor);
2192 		if (node == NULL)
2193 			break;
2194 
2195 		sym = node->ms.sym;
2196 		if (sym) {
2197 			if (!strcmp(sym->name, "schedule") ||
2198 			    !strcmp(sym->name, "__schedule") ||
2199 			    !strcmp(sym->name, "preempt_schedule"))
2200 				sym->ignore = 1;
2201 		}
2202 
2203 		callchain_cursor_advance(cursor);
2204 	}
2205 }
2206 
2207 static int init_idle_thread(struct thread *thread)
2208 {
2209 	struct idle_thread_runtime *itr;
2210 
2211 	thread__set_comm(thread, idle_comm, 0);
2212 
2213 	itr = zalloc(sizeof(*itr));
2214 	if (itr == NULL)
2215 		return -ENOMEM;
2216 
2217 	init_stats(&itr->tr.run_stats);
2218 	callchain_init(&itr->callchain);
2219 	callchain_cursor_reset(&itr->cursor);
2220 	thread__set_priv(thread, itr);
2221 
2222 	return 0;
2223 }
2224 
2225 /*
2226  * Track idle stats per cpu by maintaining a local thread
2227  * struct for the idle task on each cpu.
2228  */
2229 static int init_idle_threads(int ncpu)
2230 {
2231 	int i, ret;
2232 
2233 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2234 	if (!idle_threads)
2235 		return -ENOMEM;
2236 
2237 	idle_max_cpu = ncpu;
2238 
2239 	/* allocate the actual thread struct if needed */
2240 	for (i = 0; i < ncpu; ++i) {
2241 		idle_threads[i] = thread__new(0, 0);
2242 		if (idle_threads[i] == NULL)
2243 			return -ENOMEM;
2244 
2245 		ret = init_idle_thread(idle_threads[i]);
2246 		if (ret < 0)
2247 			return ret;
2248 	}
2249 
2250 	return 0;
2251 }
2252 
2253 static void free_idle_threads(void)
2254 {
2255 	int i;
2256 
2257 	if (idle_threads == NULL)
2258 		return;
2259 
2260 	for (i = 0; i < idle_max_cpu; ++i) {
2261 		if ((idle_threads[i]))
2262 			thread__delete(idle_threads[i]);
2263 	}
2264 
2265 	free(idle_threads);
2266 }
2267 
2268 static struct thread *get_idle_thread(int cpu)
2269 {
2270 	/*
2271 	 * expand/allocate array of pointers to local thread
2272 	 * structs if needed
2273 	 */
2274 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2275 		int i, j = __roundup_pow_of_two(cpu+1);
2276 		void *p;
2277 
2278 		p = realloc(idle_threads, j * sizeof(struct thread *));
2279 		if (!p)
2280 			return NULL;
2281 
2282 		idle_threads = (struct thread **) p;
2283 		for (i = idle_max_cpu; i < j; ++i)
2284 			idle_threads[i] = NULL;
2285 
2286 		idle_max_cpu = j;
2287 	}
2288 
2289 	/* allocate a new thread struct if needed */
2290 	if (idle_threads[cpu] == NULL) {
2291 		idle_threads[cpu] = thread__new(0, 0);
2292 
2293 		if (idle_threads[cpu]) {
2294 			if (init_idle_thread(idle_threads[cpu]) < 0)
2295 				return NULL;
2296 		}
2297 	}
2298 
2299 	return idle_threads[cpu];
2300 }
2301 
2302 static void save_idle_callchain(struct perf_sched *sched,
2303 				struct idle_thread_runtime *itr,
2304 				struct perf_sample *sample)
2305 {
2306 	struct callchain_cursor *cursor;
2307 
2308 	if (!sched->show_callchain || sample->callchain == NULL)
2309 		return;
2310 
2311 	cursor = get_tls_callchain_cursor();
2312 	if (cursor == NULL)
2313 		return;
2314 
2315 	callchain_cursor__copy(&itr->cursor, cursor);
2316 }
2317 
2318 static struct thread *timehist_get_thread(struct perf_sched *sched,
2319 					  struct perf_sample *sample,
2320 					  struct machine *machine,
2321 					  struct evsel *evsel)
2322 {
2323 	struct thread *thread;
2324 
2325 	if (is_idle_sample(sample, evsel)) {
2326 		thread = get_idle_thread(sample->cpu);
2327 		if (thread == NULL)
2328 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329 
2330 	} else {
2331 		/* there were samples with tid 0 but non-zero pid */
2332 		thread = machine__findnew_thread(machine, sample->pid,
2333 						 sample->tid ?: sample->pid);
2334 		if (thread == NULL) {
2335 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2336 				 sample->tid);
2337 		}
2338 
2339 		save_task_callchain(sched, sample, evsel, machine);
2340 		if (sched->idle_hist) {
2341 			struct thread *idle;
2342 			struct idle_thread_runtime *itr;
2343 
2344 			idle = get_idle_thread(sample->cpu);
2345 			if (idle == NULL) {
2346 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2347 				return NULL;
2348 			}
2349 
2350 			itr = thread__priv(idle);
2351 			if (itr == NULL)
2352 				return NULL;
2353 
2354 			itr->last_thread = thread;
2355 
2356 			/* copy task callchain when entering to idle */
2357 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2358 				save_idle_callchain(sched, itr, sample);
2359 		}
2360 	}
2361 
2362 	return thread;
2363 }
2364 
2365 static bool timehist_skip_sample(struct perf_sched *sched,
2366 				 struct thread *thread,
2367 				 struct evsel *evsel,
2368 				 struct perf_sample *sample)
2369 {
2370 	bool rc = false;
2371 
2372 	if (thread__is_filtered(thread)) {
2373 		rc = true;
2374 		sched->skipped_samples++;
2375 	}
2376 
2377 	if (sched->idle_hist) {
2378 		if (!evsel__name_is(evsel, "sched:sched_switch"))
2379 			rc = true;
2380 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2381 			 evsel__intval(evsel, sample, "next_pid") != 0)
2382 			rc = true;
2383 	}
2384 
2385 	return rc;
2386 }
2387 
2388 static void timehist_print_wakeup_event(struct perf_sched *sched,
2389 					struct evsel *evsel,
2390 					struct perf_sample *sample,
2391 					struct machine *machine,
2392 					struct thread *awakened)
2393 {
2394 	struct thread *thread;
2395 	char tstr[64];
2396 
2397 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2398 	if (thread == NULL)
2399 		return;
2400 
2401 	/* show wakeup unless both awakee and awaker are filtered */
2402 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2403 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2404 		return;
2405 	}
2406 
2407 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2408 	printf("%15s [%04d] ", tstr, sample->cpu);
2409 	if (sched->show_cpu_visual)
2410 		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2411 
2412 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2413 
2414 	/* dt spacer */
2415 	printf("  %9s  %9s  %9s ", "", "", "");
2416 
2417 	printf("awakened: %s", timehist_get_commstr(awakened));
2418 
2419 	printf("\n");
2420 }
2421 
2422 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2423 					union perf_event *event __maybe_unused,
2424 					struct evsel *evsel __maybe_unused,
2425 					struct perf_sample *sample __maybe_unused,
2426 					struct machine *machine __maybe_unused)
2427 {
2428 	return 0;
2429 }
2430 
2431 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2432 				       union perf_event *event __maybe_unused,
2433 				       struct evsel *evsel,
2434 				       struct perf_sample *sample,
2435 				       struct machine *machine)
2436 {
2437 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2438 	struct thread *thread;
2439 	struct thread_runtime *tr = NULL;
2440 	/* want pid of awakened task not pid in sample */
2441 	const u32 pid = evsel__intval(evsel, sample, "pid");
2442 
2443 	thread = machine__findnew_thread(machine, 0, pid);
2444 	if (thread == NULL)
2445 		return -1;
2446 
2447 	tr = thread__get_runtime(thread);
2448 	if (tr == NULL)
2449 		return -1;
2450 
2451 	if (tr->ready_to_run == 0)
2452 		tr->ready_to_run = sample->time;
2453 
2454 	/* show wakeups if requested */
2455 	if (sched->show_wakeups &&
2456 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2457 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2458 
2459 	return 0;
2460 }
2461 
2462 static void timehist_print_migration_event(struct perf_sched *sched,
2463 					struct evsel *evsel,
2464 					struct perf_sample *sample,
2465 					struct machine *machine,
2466 					struct thread *migrated)
2467 {
2468 	struct thread *thread;
2469 	char tstr[64];
2470 	u32 max_cpus;
2471 	u32 ocpu, dcpu;
2472 
2473 	if (sched->summary_only)
2474 		return;
2475 
2476 	max_cpus = sched->max_cpu.cpu + 1;
2477 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2478 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2479 
2480 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2481 	if (thread == NULL)
2482 		return;
2483 
2484 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2485 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2486 		return;
2487 	}
2488 
2489 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2490 	printf("%15s [%04d] ", tstr, sample->cpu);
2491 
2492 	if (sched->show_cpu_visual) {
2493 		u32 i;
2494 		char c;
2495 
2496 		printf("  ");
2497 		for (i = 0; i < max_cpus; ++i) {
2498 			c = (i == sample->cpu) ? 'm' : ' ';
2499 			printf("%c", c);
2500 		}
2501 		printf("  ");
2502 	}
2503 
2504 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2505 
2506 	/* dt spacer */
2507 	printf("  %9s  %9s  %9s ", "", "", "");
2508 
2509 	printf("migrated: %s", timehist_get_commstr(migrated));
2510 	printf(" cpu %d => %d", ocpu, dcpu);
2511 
2512 	printf("\n");
2513 }
2514 
2515 static int timehist_migrate_task_event(struct perf_tool *tool,
2516 				       union perf_event *event __maybe_unused,
2517 				       struct evsel *evsel,
2518 				       struct perf_sample *sample,
2519 				       struct machine *machine)
2520 {
2521 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2522 	struct thread *thread;
2523 	struct thread_runtime *tr = NULL;
2524 	/* want pid of migrated task not pid in sample */
2525 	const u32 pid = evsel__intval(evsel, sample, "pid");
2526 
2527 	thread = machine__findnew_thread(machine, 0, pid);
2528 	if (thread == NULL)
2529 		return -1;
2530 
2531 	tr = thread__get_runtime(thread);
2532 	if (tr == NULL)
2533 		return -1;
2534 
2535 	tr->migrations++;
2536 
2537 	/* show migrations if requested */
2538 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2539 
2540 	return 0;
2541 }
2542 
2543 static int timehist_sched_change_event(struct perf_tool *tool,
2544 				       union perf_event *event,
2545 				       struct evsel *evsel,
2546 				       struct perf_sample *sample,
2547 				       struct machine *machine)
2548 {
2549 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2550 	struct perf_time_interval *ptime = &sched->ptime;
2551 	struct addr_location al;
2552 	struct thread *thread;
2553 	struct thread_runtime *tr = NULL;
2554 	u64 tprev, t = sample->time;
2555 	int rc = 0;
2556 	const char state = evsel__taskstate(evsel, sample, "prev_state");
2557 
2558 	addr_location__init(&al);
2559 	if (machine__resolve(machine, &al, sample) < 0) {
2560 		pr_err("problem processing %d event. skipping it\n",
2561 		       event->header.type);
2562 		rc = -1;
2563 		goto out;
2564 	}
2565 
2566 	thread = timehist_get_thread(sched, sample, machine, evsel);
2567 	if (thread == NULL) {
2568 		rc = -1;
2569 		goto out;
2570 	}
2571 
2572 	if (timehist_skip_sample(sched, thread, evsel, sample))
2573 		goto out;
2574 
2575 	tr = thread__get_runtime(thread);
2576 	if (tr == NULL) {
2577 		rc = -1;
2578 		goto out;
2579 	}
2580 
2581 	tprev = evsel__get_time(evsel, sample->cpu);
2582 
2583 	/*
2584 	 * If start time given:
2585 	 * - sample time is under window user cares about - skip sample
2586 	 * - tprev is under window user cares about  - reset to start of window
2587 	 */
2588 	if (ptime->start && ptime->start > t)
2589 		goto out;
2590 
2591 	if (tprev && ptime->start > tprev)
2592 		tprev = ptime->start;
2593 
2594 	/*
2595 	 * If end time given:
2596 	 * - previous sched event is out of window - we are done
2597 	 * - sample time is beyond window user cares about - reset it
2598 	 *   to close out stats for time window interest
2599 	 */
2600 	if (ptime->end) {
2601 		if (tprev > ptime->end)
2602 			goto out;
2603 
2604 		if (t > ptime->end)
2605 			t = ptime->end;
2606 	}
2607 
2608 	if (!sched->idle_hist || thread__tid(thread) == 0) {
2609 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2610 			timehist_update_runtime_stats(tr, t, tprev);
2611 
2612 		if (sched->idle_hist) {
2613 			struct idle_thread_runtime *itr = (void *)tr;
2614 			struct thread_runtime *last_tr;
2615 
2616 			BUG_ON(thread__tid(thread) != 0);
2617 
2618 			if (itr->last_thread == NULL)
2619 				goto out;
2620 
2621 			/* add current idle time as last thread's runtime */
2622 			last_tr = thread__get_runtime(itr->last_thread);
2623 			if (last_tr == NULL)
2624 				goto out;
2625 
2626 			timehist_update_runtime_stats(last_tr, t, tprev);
2627 			/*
2628 			 * remove delta time of last thread as it's not updated
2629 			 * and otherwise it will show an invalid value next
2630 			 * time.  we only care total run time and run stat.
2631 			 */
2632 			last_tr->dt_run = 0;
2633 			last_tr->dt_delay = 0;
2634 			last_tr->dt_sleep = 0;
2635 			last_tr->dt_iowait = 0;
2636 			last_tr->dt_preempt = 0;
2637 
2638 			if (itr->cursor.nr)
2639 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2640 
2641 			itr->last_thread = NULL;
2642 		}
2643 	}
2644 
2645 	if (!sched->summary_only)
2646 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2647 
2648 out:
2649 	if (sched->hist_time.start == 0 && t >= ptime->start)
2650 		sched->hist_time.start = t;
2651 	if (ptime->end == 0 || t <= ptime->end)
2652 		sched->hist_time.end = t;
2653 
2654 	if (tr) {
2655 		/* time of this sched_switch event becomes last time task seen */
2656 		tr->last_time = sample->time;
2657 
2658 		/* last state is used to determine where to account wait time */
2659 		tr->last_state = state;
2660 
2661 		/* sched out event for task so reset ready to run time */
2662 		tr->ready_to_run = 0;
2663 	}
2664 
2665 	evsel__save_time(evsel, sample->time, sample->cpu);
2666 
2667 	addr_location__exit(&al);
2668 	return rc;
2669 }
2670 
2671 static int timehist_sched_switch_event(struct perf_tool *tool,
2672 			     union perf_event *event,
2673 			     struct evsel *evsel,
2674 			     struct perf_sample *sample,
2675 			     struct machine *machine __maybe_unused)
2676 {
2677 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2678 }
2679 
2680 static int process_lost(struct perf_tool *tool __maybe_unused,
2681 			union perf_event *event,
2682 			struct perf_sample *sample,
2683 			struct machine *machine __maybe_unused)
2684 {
2685 	char tstr[64];
2686 
2687 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2688 	printf("%15s ", tstr);
2689 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2690 
2691 	return 0;
2692 }
2693 
2694 
2695 static void print_thread_runtime(struct thread *t,
2696 				 struct thread_runtime *r)
2697 {
2698 	double mean = avg_stats(&r->run_stats);
2699 	float stddev;
2700 
2701 	printf("%*s   %5d  %9" PRIu64 " ",
2702 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2703 	       (u64) r->run_stats.n);
2704 
2705 	print_sched_time(r->total_run_time, 8);
2706 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2707 	print_sched_time(r->run_stats.min, 6);
2708 	printf(" ");
2709 	print_sched_time((u64) mean, 6);
2710 	printf(" ");
2711 	print_sched_time(r->run_stats.max, 6);
2712 	printf("  ");
2713 	printf("%5.2f", stddev);
2714 	printf("   %5" PRIu64, r->migrations);
2715 	printf("\n");
2716 }
2717 
2718 static void print_thread_waittime(struct thread *t,
2719 				  struct thread_runtime *r)
2720 {
2721 	printf("%*s   %5d  %9" PRIu64 " ",
2722 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2723 	       (u64) r->run_stats.n);
2724 
2725 	print_sched_time(r->total_run_time, 8);
2726 	print_sched_time(r->total_sleep_time, 6);
2727 	printf(" ");
2728 	print_sched_time(r->total_iowait_time, 6);
2729 	printf(" ");
2730 	print_sched_time(r->total_preempt_time, 6);
2731 	printf(" ");
2732 	print_sched_time(r->total_delay_time, 6);
2733 	printf("\n");
2734 }
2735 
2736 struct total_run_stats {
2737 	struct perf_sched *sched;
2738 	u64  sched_count;
2739 	u64  task_count;
2740 	u64  total_run_time;
2741 };
2742 
2743 static int show_thread_runtime(struct thread *t, void *priv)
2744 {
2745 	struct total_run_stats *stats = priv;
2746 	struct thread_runtime *r;
2747 
2748 	if (thread__is_filtered(t))
2749 		return 0;
2750 
2751 	r = thread__priv(t);
2752 	if (r && r->run_stats.n) {
2753 		stats->task_count++;
2754 		stats->sched_count += r->run_stats.n;
2755 		stats->total_run_time += r->total_run_time;
2756 
2757 		if (stats->sched->show_state)
2758 			print_thread_waittime(t, r);
2759 		else
2760 			print_thread_runtime(t, r);
2761 	}
2762 
2763 	return 0;
2764 }
2765 
2766 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2767 {
2768 	const char *sep = " <- ";
2769 	struct callchain_list *chain;
2770 	size_t ret = 0;
2771 	char bf[1024];
2772 	bool first;
2773 
2774 	if (node == NULL)
2775 		return 0;
2776 
2777 	ret = callchain__fprintf_folded(fp, node->parent);
2778 	first = (ret == 0);
2779 
2780 	list_for_each_entry(chain, &node->val, list) {
2781 		if (chain->ip >= PERF_CONTEXT_MAX)
2782 			continue;
2783 		if (chain->ms.sym && chain->ms.sym->ignore)
2784 			continue;
2785 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2786 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2787 							false));
2788 		first = false;
2789 	}
2790 
2791 	return ret;
2792 }
2793 
2794 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2795 {
2796 	size_t ret = 0;
2797 	FILE *fp = stdout;
2798 	struct callchain_node *chain;
2799 	struct rb_node *rb_node = rb_first_cached(root);
2800 
2801 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2802 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2803 	       graph_dotted_line);
2804 
2805 	while (rb_node) {
2806 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2807 		rb_node = rb_next(rb_node);
2808 
2809 		ret += fprintf(fp, "  ");
2810 		print_sched_time(chain->hit, 12);
2811 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2812 		ret += fprintf(fp, " %8d  ", chain->count);
2813 		ret += callchain__fprintf_folded(fp, chain);
2814 		ret += fprintf(fp, "\n");
2815 	}
2816 
2817 	return ret;
2818 }
2819 
2820 static void timehist_print_summary(struct perf_sched *sched,
2821 				   struct perf_session *session)
2822 {
2823 	struct machine *m = &session->machines.host;
2824 	struct total_run_stats totals;
2825 	u64 task_count;
2826 	struct thread *t;
2827 	struct thread_runtime *r;
2828 	int i;
2829 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2830 
2831 	memset(&totals, 0, sizeof(totals));
2832 	totals.sched = sched;
2833 
2834 	if (sched->idle_hist) {
2835 		printf("\nIdle-time summary\n");
2836 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2837 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2838 	} else if (sched->show_state) {
2839 		printf("\nWait-time summary\n");
2840 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2841 		printf("   run-time      sleep      iowait     preempt       delay\n");
2842 	} else {
2843 		printf("\nRuntime summary\n");
2844 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2845 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2846 	}
2847 	printf("%*s            (count)  ", comm_width, "");
2848 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2849 	       sched->show_state ? "(msec)" : "%");
2850 	printf("%.117s\n", graph_dotted_line);
2851 
2852 	machine__for_each_thread(m, show_thread_runtime, &totals);
2853 	task_count = totals.task_count;
2854 	if (!task_count)
2855 		printf("<no still running tasks>\n");
2856 
2857 	/* CPU idle stats not tracked when samples were skipped */
2858 	if (sched->skipped_samples && !sched->idle_hist)
2859 		return;
2860 
2861 	printf("\nIdle stats:\n");
2862 	for (i = 0; i < idle_max_cpu; ++i) {
2863 		if (cpu_list && !test_bit(i, cpu_bitmap))
2864 			continue;
2865 
2866 		t = idle_threads[i];
2867 		if (!t)
2868 			continue;
2869 
2870 		r = thread__priv(t);
2871 		if (r && r->run_stats.n) {
2872 			totals.sched_count += r->run_stats.n;
2873 			printf("    CPU %2d idle for ", i);
2874 			print_sched_time(r->total_run_time, 6);
2875 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2876 		} else
2877 			printf("    CPU %2d idle entire time window\n", i);
2878 	}
2879 
2880 	if (sched->idle_hist && sched->show_callchain) {
2881 		callchain_param.mode  = CHAIN_FOLDED;
2882 		callchain_param.value = CCVAL_PERIOD;
2883 
2884 		callchain_register_param(&callchain_param);
2885 
2886 		printf("\nIdle stats by callchain:\n");
2887 		for (i = 0; i < idle_max_cpu; ++i) {
2888 			struct idle_thread_runtime *itr;
2889 
2890 			t = idle_threads[i];
2891 			if (!t)
2892 				continue;
2893 
2894 			itr = thread__priv(t);
2895 			if (itr == NULL)
2896 				continue;
2897 
2898 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2899 					     0, &callchain_param);
2900 
2901 			printf("  CPU %2d:", i);
2902 			print_sched_time(itr->tr.total_run_time, 6);
2903 			printf(" msec\n");
2904 			timehist_print_idlehist_callchain(&itr->sorted_root);
2905 			printf("\n");
2906 		}
2907 	}
2908 
2909 	printf("\n"
2910 	       "    Total number of unique tasks: %" PRIu64 "\n"
2911 	       "Total number of context switches: %" PRIu64 "\n",
2912 	       totals.task_count, totals.sched_count);
2913 
2914 	printf("           Total run time (msec): ");
2915 	print_sched_time(totals.total_run_time, 2);
2916 	printf("\n");
2917 
2918 	printf("    Total scheduling time (msec): ");
2919 	print_sched_time(hist_time, 2);
2920 	printf(" (x %d)\n", sched->max_cpu.cpu);
2921 }
2922 
2923 typedef int (*sched_handler)(struct perf_tool *tool,
2924 			  union perf_event *event,
2925 			  struct evsel *evsel,
2926 			  struct perf_sample *sample,
2927 			  struct machine *machine);
2928 
2929 static int perf_timehist__process_sample(struct perf_tool *tool,
2930 					 union perf_event *event,
2931 					 struct perf_sample *sample,
2932 					 struct evsel *evsel,
2933 					 struct machine *machine)
2934 {
2935 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2936 	int err = 0;
2937 	struct perf_cpu this_cpu = {
2938 		.cpu = sample->cpu,
2939 	};
2940 
2941 	if (this_cpu.cpu > sched->max_cpu.cpu)
2942 		sched->max_cpu = this_cpu;
2943 
2944 	if (evsel->handler != NULL) {
2945 		sched_handler f = evsel->handler;
2946 
2947 		err = f(tool, event, evsel, sample, machine);
2948 	}
2949 
2950 	return err;
2951 }
2952 
2953 static int timehist_check_attr(struct perf_sched *sched,
2954 			       struct evlist *evlist)
2955 {
2956 	struct evsel *evsel;
2957 	struct evsel_runtime *er;
2958 
2959 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2960 		er = evsel__get_runtime(evsel);
2961 		if (er == NULL) {
2962 			pr_err("Failed to allocate memory for evsel runtime data\n");
2963 			return -1;
2964 		}
2965 
2966 		/* only need to save callchain related to sched_switch event */
2967 		if (sched->show_callchain &&
2968 		    evsel__name_is(evsel, "sched:sched_switch") &&
2969 		    !evsel__has_callchain(evsel)) {
2970 			pr_info("Samples of sched_switch event do not have callchains.\n");
2971 			sched->show_callchain = 0;
2972 			symbol_conf.use_callchain = 0;
2973 		}
2974 	}
2975 
2976 	return 0;
2977 }
2978 
2979 static int perf_sched__timehist(struct perf_sched *sched)
2980 {
2981 	struct evsel_str_handler handlers[] = {
2982 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2983 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2984 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
2985 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2986 	};
2987 	const struct evsel_str_handler migrate_handlers[] = {
2988 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2989 	};
2990 	struct perf_data data = {
2991 		.path  = input_name,
2992 		.mode  = PERF_DATA_MODE_READ,
2993 		.force = sched->force,
2994 	};
2995 
2996 	struct perf_session *session;
2997 	struct evlist *evlist;
2998 	int err = -1;
2999 
3000 	/*
3001 	 * event handlers for timehist option
3002 	 */
3003 	sched->tool.sample	 = perf_timehist__process_sample;
3004 	sched->tool.mmap	 = perf_event__process_mmap;
3005 	sched->tool.comm	 = perf_event__process_comm;
3006 	sched->tool.exit	 = perf_event__process_exit;
3007 	sched->tool.fork	 = perf_event__process_fork;
3008 	sched->tool.lost	 = process_lost;
3009 	sched->tool.attr	 = perf_event__process_attr;
3010 	sched->tool.tracing_data = perf_event__process_tracing_data;
3011 	sched->tool.build_id	 = perf_event__process_build_id;
3012 
3013 	sched->tool.ordered_events = true;
3014 	sched->tool.ordering_requires_timestamps = true;
3015 
3016 	symbol_conf.use_callchain = sched->show_callchain;
3017 
3018 	session = perf_session__new(&data, &sched->tool);
3019 	if (IS_ERR(session))
3020 		return PTR_ERR(session);
3021 
3022 	if (cpu_list) {
3023 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3024 		if (err < 0)
3025 			goto out;
3026 	}
3027 
3028 	evlist = session->evlist;
3029 
3030 	symbol__init(&session->header.env);
3031 
3032 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3033 		pr_err("Invalid time string\n");
3034 		return -EINVAL;
3035 	}
3036 
3037 	if (timehist_check_attr(sched, evlist) != 0)
3038 		goto out;
3039 
3040 	setup_pager();
3041 
3042 	/* prefer sched_waking if it is captured */
3043 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3044 		handlers[1].handler = timehist_sched_wakeup_ignore;
3045 
3046 	/* setup per-evsel handlers */
3047 	if (perf_session__set_tracepoints_handlers(session, handlers))
3048 		goto out;
3049 
3050 	/* sched_switch event at a minimum needs to exist */
3051 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3052 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3053 		goto out;
3054 	}
3055 
3056 	if (sched->show_migrations &&
3057 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3058 		goto out;
3059 
3060 	/* pre-allocate struct for per-CPU idle stats */
3061 	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3062 	if (sched->max_cpu.cpu == 0)
3063 		sched->max_cpu.cpu = 4;
3064 	if (init_idle_threads(sched->max_cpu.cpu))
3065 		goto out;
3066 
3067 	/* summary_only implies summary option, but don't overwrite summary if set */
3068 	if (sched->summary_only)
3069 		sched->summary = sched->summary_only;
3070 
3071 	if (!sched->summary_only)
3072 		timehist_header(sched);
3073 
3074 	err = perf_session__process_events(session);
3075 	if (err) {
3076 		pr_err("Failed to process events, error %d", err);
3077 		goto out;
3078 	}
3079 
3080 	sched->nr_events      = evlist->stats.nr_events[0];
3081 	sched->nr_lost_events = evlist->stats.total_lost;
3082 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3083 
3084 	if (sched->summary)
3085 		timehist_print_summary(sched, session);
3086 
3087 out:
3088 	free_idle_threads();
3089 	perf_session__delete(session);
3090 
3091 	return err;
3092 }
3093 
3094 
3095 static void print_bad_events(struct perf_sched *sched)
3096 {
3097 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3098 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3099 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3100 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3101 	}
3102 	if (sched->nr_lost_events && sched->nr_events) {
3103 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3104 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3105 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3106 	}
3107 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3108 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3109 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3110 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3111 		if (sched->nr_lost_events)
3112 			printf(" (due to lost events?)");
3113 		printf("\n");
3114 	}
3115 }
3116 
3117 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3118 {
3119 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3120 	struct work_atoms *this;
3121 	const char *comm = thread__comm_str(data->thread), *this_comm;
3122 	bool leftmost = true;
3123 
3124 	while (*new) {
3125 		int cmp;
3126 
3127 		this = container_of(*new, struct work_atoms, node);
3128 		parent = *new;
3129 
3130 		this_comm = thread__comm_str(this->thread);
3131 		cmp = strcmp(comm, this_comm);
3132 		if (cmp > 0) {
3133 			new = &((*new)->rb_left);
3134 		} else if (cmp < 0) {
3135 			new = &((*new)->rb_right);
3136 			leftmost = false;
3137 		} else {
3138 			this->num_merged++;
3139 			this->total_runtime += data->total_runtime;
3140 			this->nb_atoms += data->nb_atoms;
3141 			this->total_lat += data->total_lat;
3142 			list_splice(&data->work_list, &this->work_list);
3143 			if (this->max_lat < data->max_lat) {
3144 				this->max_lat = data->max_lat;
3145 				this->max_lat_start = data->max_lat_start;
3146 				this->max_lat_end = data->max_lat_end;
3147 			}
3148 			zfree(&data);
3149 			return;
3150 		}
3151 	}
3152 
3153 	data->num_merged++;
3154 	rb_link_node(&data->node, parent, new);
3155 	rb_insert_color_cached(&data->node, root, leftmost);
3156 }
3157 
3158 static void perf_sched__merge_lat(struct perf_sched *sched)
3159 {
3160 	struct work_atoms *data;
3161 	struct rb_node *node;
3162 
3163 	if (sched->skip_merge)
3164 		return;
3165 
3166 	while ((node = rb_first_cached(&sched->atom_root))) {
3167 		rb_erase_cached(node, &sched->atom_root);
3168 		data = rb_entry(node, struct work_atoms, node);
3169 		__merge_work_atoms(&sched->merged_atom_root, data);
3170 	}
3171 }
3172 
3173 static int setup_cpus_switch_event(struct perf_sched *sched)
3174 {
3175 	unsigned int i;
3176 
3177 	sched->cpu_last_switched = calloc(MAX_CPUS, sizeof(*(sched->cpu_last_switched)));
3178 	if (!sched->cpu_last_switched)
3179 		return -1;
3180 
3181 	sched->curr_pid = malloc(MAX_CPUS * sizeof(*(sched->curr_pid)));
3182 	if (!sched->curr_pid) {
3183 		zfree(&sched->cpu_last_switched);
3184 		return -1;
3185 	}
3186 
3187 	for (i = 0; i < MAX_CPUS; i++)
3188 		sched->curr_pid[i] = -1;
3189 
3190 	return 0;
3191 }
3192 
3193 static void free_cpus_switch_event(struct perf_sched *sched)
3194 {
3195 	zfree(&sched->curr_pid);
3196 	zfree(&sched->cpu_last_switched);
3197 }
3198 
3199 static int perf_sched__lat(struct perf_sched *sched)
3200 {
3201 	int rc = -1;
3202 	struct rb_node *next;
3203 
3204 	setup_pager();
3205 
3206 	if (setup_cpus_switch_event(sched))
3207 		return rc;
3208 
3209 	if (perf_sched__read_events(sched))
3210 		goto out_free_cpus_switch_event;
3211 
3212 	perf_sched__merge_lat(sched);
3213 	perf_sched__sort_lat(sched);
3214 
3215 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3216 	printf("  Task                  |   Runtime ms  |  Count   | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3217 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3218 
3219 	next = rb_first_cached(&sched->sorted_atom_root);
3220 
3221 	while (next) {
3222 		struct work_atoms *work_list;
3223 
3224 		work_list = rb_entry(next, struct work_atoms, node);
3225 		output_lat_thread(sched, work_list);
3226 		next = rb_next(next);
3227 		thread__zput(work_list->thread);
3228 	}
3229 
3230 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3231 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3232 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3233 
3234 	printf(" ---------------------------------------------------\n");
3235 
3236 	print_bad_events(sched);
3237 	printf("\n");
3238 
3239 	rc = 0;
3240 
3241 out_free_cpus_switch_event:
3242 	free_cpus_switch_event(sched);
3243 	return rc;
3244 }
3245 
3246 static int setup_map_cpus(struct perf_sched *sched)
3247 {
3248 	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3249 
3250 	if (sched->map.comp) {
3251 		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3252 		if (!sched->map.comp_cpus)
3253 			return -1;
3254 	}
3255 
3256 	if (sched->map.cpus_str) {
3257 		sched->map.cpus = perf_cpu_map__new(sched->map.cpus_str);
3258 		if (!sched->map.cpus) {
3259 			pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3260 			zfree(&sched->map.comp_cpus);
3261 			return -1;
3262 		}
3263 	}
3264 
3265 	return 0;
3266 }
3267 
3268 static int setup_color_pids(struct perf_sched *sched)
3269 {
3270 	struct perf_thread_map *map;
3271 
3272 	if (!sched->map.color_pids_str)
3273 		return 0;
3274 
3275 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3276 	if (!map) {
3277 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3278 		return -1;
3279 	}
3280 
3281 	sched->map.color_pids = map;
3282 	return 0;
3283 }
3284 
3285 static int setup_color_cpus(struct perf_sched *sched)
3286 {
3287 	struct perf_cpu_map *map;
3288 
3289 	if (!sched->map.color_cpus_str)
3290 		return 0;
3291 
3292 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3293 	if (!map) {
3294 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3295 		return -1;
3296 	}
3297 
3298 	sched->map.color_cpus = map;
3299 	return 0;
3300 }
3301 
3302 static int perf_sched__map(struct perf_sched *sched)
3303 {
3304 	int rc = -1;
3305 
3306 	sched->curr_thread = calloc(MAX_CPUS, sizeof(*(sched->curr_thread)));
3307 	if (!sched->curr_thread)
3308 		return rc;
3309 
3310 	if (setup_cpus_switch_event(sched))
3311 		goto out_free_curr_thread;
3312 
3313 	if (setup_map_cpus(sched))
3314 		goto out_free_cpus_switch_event;
3315 
3316 	if (setup_color_pids(sched))
3317 		goto out_put_map_cpus;
3318 
3319 	if (setup_color_cpus(sched))
3320 		goto out_put_color_pids;
3321 
3322 	setup_pager();
3323 	if (perf_sched__read_events(sched))
3324 		goto out_put_color_cpus;
3325 
3326 	rc = 0;
3327 	print_bad_events(sched);
3328 
3329 out_put_color_cpus:
3330 	perf_cpu_map__put(sched->map.color_cpus);
3331 
3332 out_put_color_pids:
3333 	perf_thread_map__put(sched->map.color_pids);
3334 
3335 out_put_map_cpus:
3336 	zfree(&sched->map.comp_cpus);
3337 	perf_cpu_map__put(sched->map.cpus);
3338 
3339 out_free_cpus_switch_event:
3340 	free_cpus_switch_event(sched);
3341 
3342 out_free_curr_thread:
3343 	zfree(&sched->curr_thread);
3344 	return rc;
3345 }
3346 
3347 static int perf_sched__replay(struct perf_sched *sched)
3348 {
3349 	int ret;
3350 	unsigned long i;
3351 
3352 	mutex_init(&sched->start_work_mutex);
3353 	mutex_init(&sched->work_done_wait_mutex);
3354 
3355 	ret = setup_cpus_switch_event(sched);
3356 	if (ret)
3357 		goto out_mutex_destroy;
3358 
3359 	calibrate_run_measurement_overhead(sched);
3360 	calibrate_sleep_measurement_overhead(sched);
3361 
3362 	test_calibrations(sched);
3363 
3364 	ret = perf_sched__read_events(sched);
3365 	if (ret)
3366 		goto out_free_cpus_switch_event;
3367 
3368 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3369 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3370 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3371 
3372 	if (sched->targetless_wakeups)
3373 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3374 	if (sched->multitarget_wakeups)
3375 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3376 	if (sched->nr_run_events_optimized)
3377 		printf("run atoms optimized: %ld\n",
3378 			sched->nr_run_events_optimized);
3379 
3380 	print_task_traces(sched);
3381 	add_cross_task_wakeups(sched);
3382 
3383 	sched->thread_funcs_exit = false;
3384 	create_tasks(sched);
3385 	printf("------------------------------------------------------------\n");
3386 	for (i = 0; i < sched->replay_repeat; i++)
3387 		run_one_test(sched);
3388 
3389 	sched->thread_funcs_exit = true;
3390 	destroy_tasks(sched);
3391 
3392 out_free_cpus_switch_event:
3393 	free_cpus_switch_event(sched);
3394 
3395 out_mutex_destroy:
3396 	mutex_destroy(&sched->start_work_mutex);
3397 	mutex_destroy(&sched->work_done_wait_mutex);
3398 	return ret;
3399 }
3400 
3401 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3402 			  const char * const usage_msg[])
3403 {
3404 	char *tmp, *tok, *str = strdup(sched->sort_order);
3405 
3406 	for (tok = strtok_r(str, ", ", &tmp);
3407 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3408 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3409 			usage_with_options_msg(usage_msg, options,
3410 					"Unknown --sort key: `%s'", tok);
3411 		}
3412 	}
3413 
3414 	free(str);
3415 
3416 	sort_dimension__add("pid", &sched->cmp_pid);
3417 }
3418 
3419 static bool schedstat_events_exposed(void)
3420 {
3421 	/*
3422 	 * Select "sched:sched_stat_wait" event to check
3423 	 * whether schedstat tracepoints are exposed.
3424 	 */
3425 	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3426 		false : true;
3427 }
3428 
3429 static int __cmd_record(int argc, const char **argv)
3430 {
3431 	unsigned int rec_argc, i, j;
3432 	char **rec_argv;
3433 	const char **rec_argv_copy;
3434 	const char * const record_args[] = {
3435 		"record",
3436 		"-a",
3437 		"-R",
3438 		"-m", "1024",
3439 		"-c", "1",
3440 		"-e", "sched:sched_switch",
3441 		"-e", "sched:sched_stat_runtime",
3442 		"-e", "sched:sched_process_fork",
3443 		"-e", "sched:sched_wakeup_new",
3444 		"-e", "sched:sched_migrate_task",
3445 	};
3446 
3447 	/*
3448 	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3449 	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3450 	 * to prevent "perf sched record" execution failure, determine
3451 	 * whether to record schedstat events according to actual situation.
3452 	 */
3453 	const char * const schedstat_args[] = {
3454 		"-e", "sched:sched_stat_wait",
3455 		"-e", "sched:sched_stat_sleep",
3456 		"-e", "sched:sched_stat_iowait",
3457 	};
3458 	unsigned int schedstat_argc = schedstat_events_exposed() ?
3459 		ARRAY_SIZE(schedstat_args) : 0;
3460 
3461 	struct tep_event *waking_event;
3462 	int ret;
3463 
3464 	/*
3465 	 * +2 for either "-e", "sched:sched_wakeup" or
3466 	 * "-e", "sched:sched_waking"
3467 	 */
3468 	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3469 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3470 	if (rec_argv == NULL)
3471 		return -ENOMEM;
3472 	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3473 	if (rec_argv_copy == NULL) {
3474 		free(rec_argv);
3475 		return -ENOMEM;
3476 	}
3477 
3478 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3479 		rec_argv[i] = strdup(record_args[i]);
3480 
3481 	rec_argv[i++] = strdup("-e");
3482 	waking_event = trace_event__tp_format("sched", "sched_waking");
3483 	if (!IS_ERR(waking_event))
3484 		rec_argv[i++] = strdup("sched:sched_waking");
3485 	else
3486 		rec_argv[i++] = strdup("sched:sched_wakeup");
3487 
3488 	for (j = 0; j < schedstat_argc; j++)
3489 		rec_argv[i++] = strdup(schedstat_args[j]);
3490 
3491 	for (j = 1; j < (unsigned int)argc; j++, i++)
3492 		rec_argv[i] = strdup(argv[j]);
3493 
3494 	BUG_ON(i != rec_argc);
3495 
3496 	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3497 	ret = cmd_record(rec_argc, rec_argv_copy);
3498 
3499 	for (i = 0; i < rec_argc; i++)
3500 		free(rec_argv[i]);
3501 	free(rec_argv);
3502 	free(rec_argv_copy);
3503 
3504 	return ret;
3505 }
3506 
3507 int cmd_sched(int argc, const char **argv)
3508 {
3509 	static const char default_sort_order[] = "avg, max, switch, runtime";
3510 	struct perf_sched sched = {
3511 		.tool = {
3512 			.sample		 = perf_sched__process_tracepoint_sample,
3513 			.comm		 = perf_sched__process_comm,
3514 			.namespaces	 = perf_event__process_namespaces,
3515 			.lost		 = perf_event__process_lost,
3516 			.fork		 = perf_sched__process_fork_event,
3517 			.ordered_events = true,
3518 		},
3519 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3520 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3521 		.sort_order	      = default_sort_order,
3522 		.replay_repeat	      = 10,
3523 		.profile_cpu	      = -1,
3524 		.next_shortname1      = 'A',
3525 		.next_shortname2      = '0',
3526 		.skip_merge           = 0,
3527 		.show_callchain	      = 1,
3528 		.max_stack            = 5,
3529 	};
3530 	const struct option sched_options[] = {
3531 	OPT_STRING('i', "input", &input_name, "file",
3532 		    "input file name"),
3533 	OPT_INCR('v', "verbose", &verbose,
3534 		    "be more verbose (show symbol address, etc)"),
3535 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3536 		    "dump raw trace in ASCII"),
3537 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3538 	OPT_END()
3539 	};
3540 	const struct option latency_options[] = {
3541 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3542 		   "sort by key(s): runtime, switch, avg, max"),
3543 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3544 		    "CPU to profile on"),
3545 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3546 		    "latency stats per pid instead of per comm"),
3547 	OPT_PARENT(sched_options)
3548 	};
3549 	const struct option replay_options[] = {
3550 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3551 		     "repeat the workload replay N times (-1: infinite)"),
3552 	OPT_PARENT(sched_options)
3553 	};
3554 	const struct option map_options[] = {
3555 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3556 		    "map output in compact mode"),
3557 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3558 		   "highlight given pids in map"),
3559 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3560                     "highlight given CPUs in map"),
3561 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3562                     "display given CPUs in map"),
3563 	OPT_PARENT(sched_options)
3564 	};
3565 	const struct option timehist_options[] = {
3566 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3567 		   "file", "vmlinux pathname"),
3568 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3569 		   "file", "kallsyms pathname"),
3570 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3571 		    "Display call chains if present (default on)"),
3572 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3573 		   "Maximum number of functions to display backtrace."),
3574 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3575 		    "Look for files with symbols relative to this directory"),
3576 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3577 		    "Show only syscall summary with statistics"),
3578 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3579 		    "Show all syscalls and summary with statistics"),
3580 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3581 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3582 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3583 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3584 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3585 	OPT_STRING(0, "time", &sched.time_str, "str",
3586 		   "Time span for analysis (start,stop)"),
3587 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3588 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3589 		   "analyze events only for given process id(s)"),
3590 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3591 		   "analyze events only for given thread id(s)"),
3592 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3593 	OPT_PARENT(sched_options)
3594 	};
3595 
3596 	const char * const latency_usage[] = {
3597 		"perf sched latency [<options>]",
3598 		NULL
3599 	};
3600 	const char * const replay_usage[] = {
3601 		"perf sched replay [<options>]",
3602 		NULL
3603 	};
3604 	const char * const map_usage[] = {
3605 		"perf sched map [<options>]",
3606 		NULL
3607 	};
3608 	const char * const timehist_usage[] = {
3609 		"perf sched timehist [<options>]",
3610 		NULL
3611 	};
3612 	const char *const sched_subcommands[] = { "record", "latency", "map",
3613 						  "replay", "script",
3614 						  "timehist", NULL };
3615 	const char *sched_usage[] = {
3616 		NULL,
3617 		NULL
3618 	};
3619 	struct trace_sched_handler lat_ops  = {
3620 		.wakeup_event	    = latency_wakeup_event,
3621 		.switch_event	    = latency_switch_event,
3622 		.runtime_event	    = latency_runtime_event,
3623 		.migrate_task_event = latency_migrate_task_event,
3624 	};
3625 	struct trace_sched_handler map_ops  = {
3626 		.switch_event	    = map_switch_event,
3627 	};
3628 	struct trace_sched_handler replay_ops  = {
3629 		.wakeup_event	    = replay_wakeup_event,
3630 		.switch_event	    = replay_switch_event,
3631 		.fork_event	    = replay_fork_event,
3632 	};
3633 	int ret;
3634 
3635 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3636 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3637 	if (!argc)
3638 		usage_with_options(sched_usage, sched_options);
3639 
3640 	/*
3641 	 * Aliased to 'perf script' for now:
3642 	 */
3643 	if (!strcmp(argv[0], "script")) {
3644 		return cmd_script(argc, argv);
3645 	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3646 		return __cmd_record(argc, argv);
3647 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3648 		sched.tp_handler = &lat_ops;
3649 		if (argc > 1) {
3650 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3651 			if (argc)
3652 				usage_with_options(latency_usage, latency_options);
3653 		}
3654 		setup_sorting(&sched, latency_options, latency_usage);
3655 		return perf_sched__lat(&sched);
3656 	} else if (!strcmp(argv[0], "map")) {
3657 		if (argc) {
3658 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3659 			if (argc)
3660 				usage_with_options(map_usage, map_options);
3661 		}
3662 		sched.tp_handler = &map_ops;
3663 		setup_sorting(&sched, latency_options, latency_usage);
3664 		return perf_sched__map(&sched);
3665 	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3666 		sched.tp_handler = &replay_ops;
3667 		if (argc) {
3668 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3669 			if (argc)
3670 				usage_with_options(replay_usage, replay_options);
3671 		}
3672 		return perf_sched__replay(&sched);
3673 	} else if (!strcmp(argv[0], "timehist")) {
3674 		if (argc) {
3675 			argc = parse_options(argc, argv, timehist_options,
3676 					     timehist_usage, 0);
3677 			if (argc)
3678 				usage_with_options(timehist_usage, timehist_options);
3679 		}
3680 		if ((sched.show_wakeups || sched.show_next) &&
3681 		    sched.summary_only) {
3682 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3683 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3684 			if (sched.show_wakeups)
3685 				parse_options_usage(NULL, timehist_options, "w", true);
3686 			if (sched.show_next)
3687 				parse_options_usage(NULL, timehist_options, "n", true);
3688 			return -EINVAL;
3689 		}
3690 		ret = symbol__validate_sym_arguments();
3691 		if (ret)
3692 			return ret;
3693 
3694 		return perf_sched__timehist(&sched);
3695 	} else {
3696 		usage_with_options(sched_usage, sched_options);
3697 	}
3698 
3699 	return 0;
3700 }
3701