xref: /linux/tools/perf/builtin-sched.c (revision 4fd18fc38757217c746aa063ba9e4729814dc737)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 #include "perf-sys.h"
5 
6 #include "util/cpumap.h"
7 #include "util/evlist.h"
8 #include "util/evsel.h"
9 #include "util/evsel_fprintf.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
22 
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
26 
27 #include "util/debug.h"
28 #include "util/event.h"
29 
30 #include <linux/kernel.h>
31 #include <linux/log2.h>
32 #include <linux/zalloc.h>
33 #include <sys/prctl.h>
34 #include <sys/resource.h>
35 #include <inttypes.h>
36 
37 #include <errno.h>
38 #include <semaphore.h>
39 #include <pthread.h>
40 #include <math.h>
41 #include <api/fs/fs.h>
42 #include <perf/cpumap.h>
43 #include <linux/time64.h>
44 #include <linux/err.h>
45 
46 #include <linux/ctype.h>
47 
48 #define PR_SET_NAME		15               /* Set process name */
49 #define MAX_CPUS		4096
50 #define COMM_LEN		20
51 #define SYM_LEN			129
52 #define MAX_PID			1024000
53 
54 static const char *cpu_list;
55 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
56 
57 struct sched_atom;
58 
59 struct task_desc {
60 	unsigned long		nr;
61 	unsigned long		pid;
62 	char			comm[COMM_LEN];
63 
64 	unsigned long		nr_events;
65 	unsigned long		curr_event;
66 	struct sched_atom	**atoms;
67 
68 	pthread_t		thread;
69 	sem_t			sleep_sem;
70 
71 	sem_t			ready_for_work;
72 	sem_t			work_done_sem;
73 
74 	u64			cpu_usage;
75 };
76 
77 enum sched_event_type {
78 	SCHED_EVENT_RUN,
79 	SCHED_EVENT_SLEEP,
80 	SCHED_EVENT_WAKEUP,
81 	SCHED_EVENT_MIGRATION,
82 };
83 
84 struct sched_atom {
85 	enum sched_event_type	type;
86 	int			specific_wait;
87 	u64			timestamp;
88 	u64			duration;
89 	unsigned long		nr;
90 	sem_t			*wait_sem;
91 	struct task_desc	*wakee;
92 };
93 
94 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
95 
96 /* task state bitmask, copied from include/linux/sched.h */
97 #define TASK_RUNNING		0
98 #define TASK_INTERRUPTIBLE	1
99 #define TASK_UNINTERRUPTIBLE	2
100 #define __TASK_STOPPED		4
101 #define __TASK_TRACED		8
102 /* in tsk->exit_state */
103 #define EXIT_DEAD		16
104 #define EXIT_ZOMBIE		32
105 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
106 /* in tsk->state again */
107 #define TASK_DEAD		64
108 #define TASK_WAKEKILL		128
109 #define TASK_WAKING		256
110 #define TASK_PARKED		512
111 
112 enum thread_state {
113 	THREAD_SLEEPING = 0,
114 	THREAD_WAIT_CPU,
115 	THREAD_SCHED_IN,
116 	THREAD_IGNORE
117 };
118 
119 struct work_atom {
120 	struct list_head	list;
121 	enum thread_state	state;
122 	u64			sched_out_time;
123 	u64			wake_up_time;
124 	u64			sched_in_time;
125 	u64			runtime;
126 };
127 
128 struct work_atoms {
129 	struct list_head	work_list;
130 	struct thread		*thread;
131 	struct rb_node		node;
132 	u64			max_lat;
133 	u64			max_lat_start;
134 	u64			max_lat_end;
135 	u64			total_lat;
136 	u64			nb_atoms;
137 	u64			total_runtime;
138 	int			num_merged;
139 };
140 
141 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
142 
143 struct perf_sched;
144 
145 struct trace_sched_handler {
146 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
147 			    struct perf_sample *sample, struct machine *machine);
148 
149 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
150 			     struct perf_sample *sample, struct machine *machine);
151 
152 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
153 			    struct perf_sample *sample, struct machine *machine);
154 
155 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
156 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
157 			  struct machine *machine);
158 
159 	int (*migrate_task_event)(struct perf_sched *sched,
160 				  struct evsel *evsel,
161 				  struct perf_sample *sample,
162 				  struct machine *machine);
163 };
164 
165 #define COLOR_PIDS PERF_COLOR_BLUE
166 #define COLOR_CPUS PERF_COLOR_BG_RED
167 
168 struct perf_sched_map {
169 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
170 	int			*comp_cpus;
171 	bool			 comp;
172 	struct perf_thread_map *color_pids;
173 	const char		*color_pids_str;
174 	struct perf_cpu_map	*color_cpus;
175 	const char		*color_cpus_str;
176 	struct perf_cpu_map	*cpus;
177 	const char		*cpus_str;
178 };
179 
180 struct perf_sched {
181 	struct perf_tool tool;
182 	const char	 *sort_order;
183 	unsigned long	 nr_tasks;
184 	struct task_desc **pid_to_task;
185 	struct task_desc **tasks;
186 	const struct trace_sched_handler *tp_handler;
187 	pthread_mutex_t	 start_work_mutex;
188 	pthread_mutex_t	 work_done_wait_mutex;
189 	int		 profile_cpu;
190 /*
191  * Track the current task - that way we can know whether there's any
192  * weird events, such as a task being switched away that is not current.
193  */
194 	int		 max_cpu;
195 	u32		 curr_pid[MAX_CPUS];
196 	struct thread	 *curr_thread[MAX_CPUS];
197 	char		 next_shortname1;
198 	char		 next_shortname2;
199 	unsigned int	 replay_repeat;
200 	unsigned long	 nr_run_events;
201 	unsigned long	 nr_sleep_events;
202 	unsigned long	 nr_wakeup_events;
203 	unsigned long	 nr_sleep_corrections;
204 	unsigned long	 nr_run_events_optimized;
205 	unsigned long	 targetless_wakeups;
206 	unsigned long	 multitarget_wakeups;
207 	unsigned long	 nr_runs;
208 	unsigned long	 nr_timestamps;
209 	unsigned long	 nr_unordered_timestamps;
210 	unsigned long	 nr_context_switch_bugs;
211 	unsigned long	 nr_events;
212 	unsigned long	 nr_lost_chunks;
213 	unsigned long	 nr_lost_events;
214 	u64		 run_measurement_overhead;
215 	u64		 sleep_measurement_overhead;
216 	u64		 start_time;
217 	u64		 cpu_usage;
218 	u64		 runavg_cpu_usage;
219 	u64		 parent_cpu_usage;
220 	u64		 runavg_parent_cpu_usage;
221 	u64		 sum_runtime;
222 	u64		 sum_fluct;
223 	u64		 run_avg;
224 	u64		 all_runtime;
225 	u64		 all_count;
226 	u64		 cpu_last_switched[MAX_CPUS];
227 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
228 	struct list_head sort_list, cmp_pid;
229 	bool force;
230 	bool skip_merge;
231 	struct perf_sched_map map;
232 
233 	/* options for timehist command */
234 	bool		summary;
235 	bool		summary_only;
236 	bool		idle_hist;
237 	bool		show_callchain;
238 	unsigned int	max_stack;
239 	bool		show_cpu_visual;
240 	bool		show_wakeups;
241 	bool		show_next;
242 	bool		show_migrations;
243 	bool		show_state;
244 	u64		skipped_samples;
245 	const char	*time_str;
246 	struct perf_time_interval ptime;
247 	struct perf_time_interval hist_time;
248 };
249 
250 /* per thread run time data */
251 struct thread_runtime {
252 	u64 last_time;      /* time of previous sched in/out event */
253 	u64 dt_run;         /* run time */
254 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
255 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
256 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
257 	u64 dt_delay;       /* time between wakeup and sched-in */
258 	u64 ready_to_run;   /* time of wakeup */
259 
260 	struct stats run_stats;
261 	u64 total_run_time;
262 	u64 total_sleep_time;
263 	u64 total_iowait_time;
264 	u64 total_preempt_time;
265 	u64 total_delay_time;
266 
267 	int last_state;
268 
269 	char shortname[3];
270 	bool comm_changed;
271 
272 	u64 migrations;
273 };
274 
275 /* per event run time data */
276 struct evsel_runtime {
277 	u64 *last_time; /* time this event was last seen per cpu */
278 	u32 ncpu;       /* highest cpu slot allocated */
279 };
280 
281 /* per cpu idle time data */
282 struct idle_thread_runtime {
283 	struct thread_runtime	tr;
284 	struct thread		*last_thread;
285 	struct rb_root_cached	sorted_root;
286 	struct callchain_root	callchain;
287 	struct callchain_cursor	cursor;
288 };
289 
290 /* track idle times per cpu */
291 static struct thread **idle_threads;
292 static int idle_max_cpu;
293 static char idle_comm[] = "<idle>";
294 
295 static u64 get_nsecs(void)
296 {
297 	struct timespec ts;
298 
299 	clock_gettime(CLOCK_MONOTONIC, &ts);
300 
301 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
302 }
303 
304 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
305 {
306 	u64 T0 = get_nsecs(), T1;
307 
308 	do {
309 		T1 = get_nsecs();
310 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
311 }
312 
313 static void sleep_nsecs(u64 nsecs)
314 {
315 	struct timespec ts;
316 
317 	ts.tv_nsec = nsecs % 999999999;
318 	ts.tv_sec = nsecs / 999999999;
319 
320 	nanosleep(&ts, NULL);
321 }
322 
323 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
324 {
325 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
326 	int i;
327 
328 	for (i = 0; i < 10; i++) {
329 		T0 = get_nsecs();
330 		burn_nsecs(sched, 0);
331 		T1 = get_nsecs();
332 		delta = T1-T0;
333 		min_delta = min(min_delta, delta);
334 	}
335 	sched->run_measurement_overhead = min_delta;
336 
337 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
338 }
339 
340 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
341 {
342 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
343 	int i;
344 
345 	for (i = 0; i < 10; i++) {
346 		T0 = get_nsecs();
347 		sleep_nsecs(10000);
348 		T1 = get_nsecs();
349 		delta = T1-T0;
350 		min_delta = min(min_delta, delta);
351 	}
352 	min_delta -= 10000;
353 	sched->sleep_measurement_overhead = min_delta;
354 
355 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
356 }
357 
358 static struct sched_atom *
359 get_new_event(struct task_desc *task, u64 timestamp)
360 {
361 	struct sched_atom *event = zalloc(sizeof(*event));
362 	unsigned long idx = task->nr_events;
363 	size_t size;
364 
365 	event->timestamp = timestamp;
366 	event->nr = idx;
367 
368 	task->nr_events++;
369 	size = sizeof(struct sched_atom *) * task->nr_events;
370 	task->atoms = realloc(task->atoms, size);
371 	BUG_ON(!task->atoms);
372 
373 	task->atoms[idx] = event;
374 
375 	return event;
376 }
377 
378 static struct sched_atom *last_event(struct task_desc *task)
379 {
380 	if (!task->nr_events)
381 		return NULL;
382 
383 	return task->atoms[task->nr_events - 1];
384 }
385 
386 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
387 				u64 timestamp, u64 duration)
388 {
389 	struct sched_atom *event, *curr_event = last_event(task);
390 
391 	/*
392 	 * optimize an existing RUN event by merging this one
393 	 * to it:
394 	 */
395 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
396 		sched->nr_run_events_optimized++;
397 		curr_event->duration += duration;
398 		return;
399 	}
400 
401 	event = get_new_event(task, timestamp);
402 
403 	event->type = SCHED_EVENT_RUN;
404 	event->duration = duration;
405 
406 	sched->nr_run_events++;
407 }
408 
409 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
410 				   u64 timestamp, struct task_desc *wakee)
411 {
412 	struct sched_atom *event, *wakee_event;
413 
414 	event = get_new_event(task, timestamp);
415 	event->type = SCHED_EVENT_WAKEUP;
416 	event->wakee = wakee;
417 
418 	wakee_event = last_event(wakee);
419 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
420 		sched->targetless_wakeups++;
421 		return;
422 	}
423 	if (wakee_event->wait_sem) {
424 		sched->multitarget_wakeups++;
425 		return;
426 	}
427 
428 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
429 	sem_init(wakee_event->wait_sem, 0, 0);
430 	wakee_event->specific_wait = 1;
431 	event->wait_sem = wakee_event->wait_sem;
432 
433 	sched->nr_wakeup_events++;
434 }
435 
436 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
437 				  u64 timestamp, u64 task_state __maybe_unused)
438 {
439 	struct sched_atom *event = get_new_event(task, timestamp);
440 
441 	event->type = SCHED_EVENT_SLEEP;
442 
443 	sched->nr_sleep_events++;
444 }
445 
446 static struct task_desc *register_pid(struct perf_sched *sched,
447 				      unsigned long pid, const char *comm)
448 {
449 	struct task_desc *task;
450 	static int pid_max;
451 
452 	if (sched->pid_to_task == NULL) {
453 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
454 			pid_max = MAX_PID;
455 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
456 	}
457 	if (pid >= (unsigned long)pid_max) {
458 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
459 			sizeof(struct task_desc *))) == NULL);
460 		while (pid >= (unsigned long)pid_max)
461 			sched->pid_to_task[pid_max++] = NULL;
462 	}
463 
464 	task = sched->pid_to_task[pid];
465 
466 	if (task)
467 		return task;
468 
469 	task = zalloc(sizeof(*task));
470 	task->pid = pid;
471 	task->nr = sched->nr_tasks;
472 	strcpy(task->comm, comm);
473 	/*
474 	 * every task starts in sleeping state - this gets ignored
475 	 * if there's no wakeup pointing to this sleep state:
476 	 */
477 	add_sched_event_sleep(sched, task, 0, 0);
478 
479 	sched->pid_to_task[pid] = task;
480 	sched->nr_tasks++;
481 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
482 	BUG_ON(!sched->tasks);
483 	sched->tasks[task->nr] = task;
484 
485 	if (verbose > 0)
486 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
487 
488 	return task;
489 }
490 
491 
492 static void print_task_traces(struct perf_sched *sched)
493 {
494 	struct task_desc *task;
495 	unsigned long i;
496 
497 	for (i = 0; i < sched->nr_tasks; i++) {
498 		task = sched->tasks[i];
499 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
500 			task->nr, task->comm, task->pid, task->nr_events);
501 	}
502 }
503 
504 static void add_cross_task_wakeups(struct perf_sched *sched)
505 {
506 	struct task_desc *task1, *task2;
507 	unsigned long i, j;
508 
509 	for (i = 0; i < sched->nr_tasks; i++) {
510 		task1 = sched->tasks[i];
511 		j = i + 1;
512 		if (j == sched->nr_tasks)
513 			j = 0;
514 		task2 = sched->tasks[j];
515 		add_sched_event_wakeup(sched, task1, 0, task2);
516 	}
517 }
518 
519 static void perf_sched__process_event(struct perf_sched *sched,
520 				      struct sched_atom *atom)
521 {
522 	int ret = 0;
523 
524 	switch (atom->type) {
525 		case SCHED_EVENT_RUN:
526 			burn_nsecs(sched, atom->duration);
527 			break;
528 		case SCHED_EVENT_SLEEP:
529 			if (atom->wait_sem)
530 				ret = sem_wait(atom->wait_sem);
531 			BUG_ON(ret);
532 			break;
533 		case SCHED_EVENT_WAKEUP:
534 			if (atom->wait_sem)
535 				ret = sem_post(atom->wait_sem);
536 			BUG_ON(ret);
537 			break;
538 		case SCHED_EVENT_MIGRATION:
539 			break;
540 		default:
541 			BUG_ON(1);
542 	}
543 }
544 
545 static u64 get_cpu_usage_nsec_parent(void)
546 {
547 	struct rusage ru;
548 	u64 sum;
549 	int err;
550 
551 	err = getrusage(RUSAGE_SELF, &ru);
552 	BUG_ON(err);
553 
554 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
555 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
556 
557 	return sum;
558 }
559 
560 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
561 {
562 	struct perf_event_attr attr;
563 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
564 	int fd;
565 	struct rlimit limit;
566 	bool need_privilege = false;
567 
568 	memset(&attr, 0, sizeof(attr));
569 
570 	attr.type = PERF_TYPE_SOFTWARE;
571 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
572 
573 force_again:
574 	fd = sys_perf_event_open(&attr, 0, -1, -1,
575 				 perf_event_open_cloexec_flag());
576 
577 	if (fd < 0) {
578 		if (errno == EMFILE) {
579 			if (sched->force) {
580 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
581 				limit.rlim_cur += sched->nr_tasks - cur_task;
582 				if (limit.rlim_cur > limit.rlim_max) {
583 					limit.rlim_max = limit.rlim_cur;
584 					need_privilege = true;
585 				}
586 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
587 					if (need_privilege && errno == EPERM)
588 						strcpy(info, "Need privilege\n");
589 				} else
590 					goto force_again;
591 			} else
592 				strcpy(info, "Have a try with -f option\n");
593 		}
594 		pr_err("Error: sys_perf_event_open() syscall returned "
595 		       "with %d (%s)\n%s", fd,
596 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
597 		exit(EXIT_FAILURE);
598 	}
599 	return fd;
600 }
601 
602 static u64 get_cpu_usage_nsec_self(int fd)
603 {
604 	u64 runtime;
605 	int ret;
606 
607 	ret = read(fd, &runtime, sizeof(runtime));
608 	BUG_ON(ret != sizeof(runtime));
609 
610 	return runtime;
611 }
612 
613 struct sched_thread_parms {
614 	struct task_desc  *task;
615 	struct perf_sched *sched;
616 	int fd;
617 };
618 
619 static void *thread_func(void *ctx)
620 {
621 	struct sched_thread_parms *parms = ctx;
622 	struct task_desc *this_task = parms->task;
623 	struct perf_sched *sched = parms->sched;
624 	u64 cpu_usage_0, cpu_usage_1;
625 	unsigned long i, ret;
626 	char comm2[22];
627 	int fd = parms->fd;
628 
629 	zfree(&parms);
630 
631 	sprintf(comm2, ":%s", this_task->comm);
632 	prctl(PR_SET_NAME, comm2);
633 	if (fd < 0)
634 		return NULL;
635 again:
636 	ret = sem_post(&this_task->ready_for_work);
637 	BUG_ON(ret);
638 	ret = pthread_mutex_lock(&sched->start_work_mutex);
639 	BUG_ON(ret);
640 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
641 	BUG_ON(ret);
642 
643 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
644 
645 	for (i = 0; i < this_task->nr_events; i++) {
646 		this_task->curr_event = i;
647 		perf_sched__process_event(sched, this_task->atoms[i]);
648 	}
649 
650 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
651 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
652 	ret = sem_post(&this_task->work_done_sem);
653 	BUG_ON(ret);
654 
655 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
656 	BUG_ON(ret);
657 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
658 	BUG_ON(ret);
659 
660 	goto again;
661 }
662 
663 static void create_tasks(struct perf_sched *sched)
664 {
665 	struct task_desc *task;
666 	pthread_attr_t attr;
667 	unsigned long i;
668 	int err;
669 
670 	err = pthread_attr_init(&attr);
671 	BUG_ON(err);
672 	err = pthread_attr_setstacksize(&attr,
673 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
674 	BUG_ON(err);
675 	err = pthread_mutex_lock(&sched->start_work_mutex);
676 	BUG_ON(err);
677 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
678 	BUG_ON(err);
679 	for (i = 0; i < sched->nr_tasks; i++) {
680 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
681 		BUG_ON(parms == NULL);
682 		parms->task = task = sched->tasks[i];
683 		parms->sched = sched;
684 		parms->fd = self_open_counters(sched, i);
685 		sem_init(&task->sleep_sem, 0, 0);
686 		sem_init(&task->ready_for_work, 0, 0);
687 		sem_init(&task->work_done_sem, 0, 0);
688 		task->curr_event = 0;
689 		err = pthread_create(&task->thread, &attr, thread_func, parms);
690 		BUG_ON(err);
691 	}
692 }
693 
694 static void wait_for_tasks(struct perf_sched *sched)
695 {
696 	u64 cpu_usage_0, cpu_usage_1;
697 	struct task_desc *task;
698 	unsigned long i, ret;
699 
700 	sched->start_time = get_nsecs();
701 	sched->cpu_usage = 0;
702 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
703 
704 	for (i = 0; i < sched->nr_tasks; i++) {
705 		task = sched->tasks[i];
706 		ret = sem_wait(&task->ready_for_work);
707 		BUG_ON(ret);
708 		sem_init(&task->ready_for_work, 0, 0);
709 	}
710 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
711 	BUG_ON(ret);
712 
713 	cpu_usage_0 = get_cpu_usage_nsec_parent();
714 
715 	pthread_mutex_unlock(&sched->start_work_mutex);
716 
717 	for (i = 0; i < sched->nr_tasks; i++) {
718 		task = sched->tasks[i];
719 		ret = sem_wait(&task->work_done_sem);
720 		BUG_ON(ret);
721 		sem_init(&task->work_done_sem, 0, 0);
722 		sched->cpu_usage += task->cpu_usage;
723 		task->cpu_usage = 0;
724 	}
725 
726 	cpu_usage_1 = get_cpu_usage_nsec_parent();
727 	if (!sched->runavg_cpu_usage)
728 		sched->runavg_cpu_usage = sched->cpu_usage;
729 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
730 
731 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
732 	if (!sched->runavg_parent_cpu_usage)
733 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
734 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
735 					 sched->parent_cpu_usage)/sched->replay_repeat;
736 
737 	ret = pthread_mutex_lock(&sched->start_work_mutex);
738 	BUG_ON(ret);
739 
740 	for (i = 0; i < sched->nr_tasks; i++) {
741 		task = sched->tasks[i];
742 		sem_init(&task->sleep_sem, 0, 0);
743 		task->curr_event = 0;
744 	}
745 }
746 
747 static void run_one_test(struct perf_sched *sched)
748 {
749 	u64 T0, T1, delta, avg_delta, fluct;
750 
751 	T0 = get_nsecs();
752 	wait_for_tasks(sched);
753 	T1 = get_nsecs();
754 
755 	delta = T1 - T0;
756 	sched->sum_runtime += delta;
757 	sched->nr_runs++;
758 
759 	avg_delta = sched->sum_runtime / sched->nr_runs;
760 	if (delta < avg_delta)
761 		fluct = avg_delta - delta;
762 	else
763 		fluct = delta - avg_delta;
764 	sched->sum_fluct += fluct;
765 	if (!sched->run_avg)
766 		sched->run_avg = delta;
767 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
768 
769 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
770 
771 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
772 
773 	printf("cpu: %0.2f / %0.2f",
774 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
775 
776 #if 0
777 	/*
778 	 * rusage statistics done by the parent, these are less
779 	 * accurate than the sched->sum_exec_runtime based statistics:
780 	 */
781 	printf(" [%0.2f / %0.2f]",
782 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
783 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
784 #endif
785 
786 	printf("\n");
787 
788 	if (sched->nr_sleep_corrections)
789 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
790 	sched->nr_sleep_corrections = 0;
791 }
792 
793 static void test_calibrations(struct perf_sched *sched)
794 {
795 	u64 T0, T1;
796 
797 	T0 = get_nsecs();
798 	burn_nsecs(sched, NSEC_PER_MSEC);
799 	T1 = get_nsecs();
800 
801 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
802 
803 	T0 = get_nsecs();
804 	sleep_nsecs(NSEC_PER_MSEC);
805 	T1 = get_nsecs();
806 
807 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
808 }
809 
810 static int
811 replay_wakeup_event(struct perf_sched *sched,
812 		    struct evsel *evsel, struct perf_sample *sample,
813 		    struct machine *machine __maybe_unused)
814 {
815 	const char *comm = evsel__strval(evsel, sample, "comm");
816 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
817 	struct task_desc *waker, *wakee;
818 
819 	if (verbose > 0) {
820 		printf("sched_wakeup event %p\n", evsel);
821 
822 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
823 	}
824 
825 	waker = register_pid(sched, sample->tid, "<unknown>");
826 	wakee = register_pid(sched, pid, comm);
827 
828 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
829 	return 0;
830 }
831 
832 static int replay_switch_event(struct perf_sched *sched,
833 			       struct evsel *evsel,
834 			       struct perf_sample *sample,
835 			       struct machine *machine __maybe_unused)
836 {
837 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
838 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
839 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
840 		  next_pid = evsel__intval(evsel, sample, "next_pid");
841 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
842 	struct task_desc *prev, __maybe_unused *next;
843 	u64 timestamp0, timestamp = sample->time;
844 	int cpu = sample->cpu;
845 	s64 delta;
846 
847 	if (verbose > 0)
848 		printf("sched_switch event %p\n", evsel);
849 
850 	if (cpu >= MAX_CPUS || cpu < 0)
851 		return 0;
852 
853 	timestamp0 = sched->cpu_last_switched[cpu];
854 	if (timestamp0)
855 		delta = timestamp - timestamp0;
856 	else
857 		delta = 0;
858 
859 	if (delta < 0) {
860 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
861 		return -1;
862 	}
863 
864 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
865 		 prev_comm, prev_pid, next_comm, next_pid, delta);
866 
867 	prev = register_pid(sched, prev_pid, prev_comm);
868 	next = register_pid(sched, next_pid, next_comm);
869 
870 	sched->cpu_last_switched[cpu] = timestamp;
871 
872 	add_sched_event_run(sched, prev, timestamp, delta);
873 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
874 
875 	return 0;
876 }
877 
878 static int replay_fork_event(struct perf_sched *sched,
879 			     union perf_event *event,
880 			     struct machine *machine)
881 {
882 	struct thread *child, *parent;
883 
884 	child = machine__findnew_thread(machine, event->fork.pid,
885 					event->fork.tid);
886 	parent = machine__findnew_thread(machine, event->fork.ppid,
887 					 event->fork.ptid);
888 
889 	if (child == NULL || parent == NULL) {
890 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
891 				 child, parent);
892 		goto out_put;
893 	}
894 
895 	if (verbose > 0) {
896 		printf("fork event\n");
897 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
898 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
899 	}
900 
901 	register_pid(sched, parent->tid, thread__comm_str(parent));
902 	register_pid(sched, child->tid, thread__comm_str(child));
903 out_put:
904 	thread__put(child);
905 	thread__put(parent);
906 	return 0;
907 }
908 
909 struct sort_dimension {
910 	const char		*name;
911 	sort_fn_t		cmp;
912 	struct list_head	list;
913 };
914 
915 /*
916  * handle runtime stats saved per thread
917  */
918 static struct thread_runtime *thread__init_runtime(struct thread *thread)
919 {
920 	struct thread_runtime *r;
921 
922 	r = zalloc(sizeof(struct thread_runtime));
923 	if (!r)
924 		return NULL;
925 
926 	init_stats(&r->run_stats);
927 	thread__set_priv(thread, r);
928 
929 	return r;
930 }
931 
932 static struct thread_runtime *thread__get_runtime(struct thread *thread)
933 {
934 	struct thread_runtime *tr;
935 
936 	tr = thread__priv(thread);
937 	if (tr == NULL) {
938 		tr = thread__init_runtime(thread);
939 		if (tr == NULL)
940 			pr_debug("Failed to malloc memory for runtime data.\n");
941 	}
942 
943 	return tr;
944 }
945 
946 static int
947 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
948 {
949 	struct sort_dimension *sort;
950 	int ret = 0;
951 
952 	BUG_ON(list_empty(list));
953 
954 	list_for_each_entry(sort, list, list) {
955 		ret = sort->cmp(l, r);
956 		if (ret)
957 			return ret;
958 	}
959 
960 	return ret;
961 }
962 
963 static struct work_atoms *
964 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
965 			 struct list_head *sort_list)
966 {
967 	struct rb_node *node = root->rb_root.rb_node;
968 	struct work_atoms key = { .thread = thread };
969 
970 	while (node) {
971 		struct work_atoms *atoms;
972 		int cmp;
973 
974 		atoms = container_of(node, struct work_atoms, node);
975 
976 		cmp = thread_lat_cmp(sort_list, &key, atoms);
977 		if (cmp > 0)
978 			node = node->rb_left;
979 		else if (cmp < 0)
980 			node = node->rb_right;
981 		else {
982 			BUG_ON(thread != atoms->thread);
983 			return atoms;
984 		}
985 	}
986 	return NULL;
987 }
988 
989 static void
990 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
991 			 struct list_head *sort_list)
992 {
993 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
994 	bool leftmost = true;
995 
996 	while (*new) {
997 		struct work_atoms *this;
998 		int cmp;
999 
1000 		this = container_of(*new, struct work_atoms, node);
1001 		parent = *new;
1002 
1003 		cmp = thread_lat_cmp(sort_list, data, this);
1004 
1005 		if (cmp > 0)
1006 			new = &((*new)->rb_left);
1007 		else {
1008 			new = &((*new)->rb_right);
1009 			leftmost = false;
1010 		}
1011 	}
1012 
1013 	rb_link_node(&data->node, parent, new);
1014 	rb_insert_color_cached(&data->node, root, leftmost);
1015 }
1016 
1017 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1018 {
1019 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1020 	if (!atoms) {
1021 		pr_err("No memory at %s\n", __func__);
1022 		return -1;
1023 	}
1024 
1025 	atoms->thread = thread__get(thread);
1026 	INIT_LIST_HEAD(&atoms->work_list);
1027 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1028 	return 0;
1029 }
1030 
1031 static char sched_out_state(u64 prev_state)
1032 {
1033 	const char *str = TASK_STATE_TO_CHAR_STR;
1034 
1035 	return str[prev_state];
1036 }
1037 
1038 static int
1039 add_sched_out_event(struct work_atoms *atoms,
1040 		    char run_state,
1041 		    u64 timestamp)
1042 {
1043 	struct work_atom *atom = zalloc(sizeof(*atom));
1044 	if (!atom) {
1045 		pr_err("Non memory at %s", __func__);
1046 		return -1;
1047 	}
1048 
1049 	atom->sched_out_time = timestamp;
1050 
1051 	if (run_state == 'R') {
1052 		atom->state = THREAD_WAIT_CPU;
1053 		atom->wake_up_time = atom->sched_out_time;
1054 	}
1055 
1056 	list_add_tail(&atom->list, &atoms->work_list);
1057 	return 0;
1058 }
1059 
1060 static void
1061 add_runtime_event(struct work_atoms *atoms, u64 delta,
1062 		  u64 timestamp __maybe_unused)
1063 {
1064 	struct work_atom *atom;
1065 
1066 	BUG_ON(list_empty(&atoms->work_list));
1067 
1068 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1069 
1070 	atom->runtime += delta;
1071 	atoms->total_runtime += delta;
1072 }
1073 
1074 static void
1075 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1076 {
1077 	struct work_atom *atom;
1078 	u64 delta;
1079 
1080 	if (list_empty(&atoms->work_list))
1081 		return;
1082 
1083 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1084 
1085 	if (atom->state != THREAD_WAIT_CPU)
1086 		return;
1087 
1088 	if (timestamp < atom->wake_up_time) {
1089 		atom->state = THREAD_IGNORE;
1090 		return;
1091 	}
1092 
1093 	atom->state = THREAD_SCHED_IN;
1094 	atom->sched_in_time = timestamp;
1095 
1096 	delta = atom->sched_in_time - atom->wake_up_time;
1097 	atoms->total_lat += delta;
1098 	if (delta > atoms->max_lat) {
1099 		atoms->max_lat = delta;
1100 		atoms->max_lat_start = atom->wake_up_time;
1101 		atoms->max_lat_end = timestamp;
1102 	}
1103 	atoms->nb_atoms++;
1104 }
1105 
1106 static int latency_switch_event(struct perf_sched *sched,
1107 				struct evsel *evsel,
1108 				struct perf_sample *sample,
1109 				struct machine *machine)
1110 {
1111 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1112 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1113 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1114 	struct work_atoms *out_events, *in_events;
1115 	struct thread *sched_out, *sched_in;
1116 	u64 timestamp0, timestamp = sample->time;
1117 	int cpu = sample->cpu, err = -1;
1118 	s64 delta;
1119 
1120 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1121 
1122 	timestamp0 = sched->cpu_last_switched[cpu];
1123 	sched->cpu_last_switched[cpu] = timestamp;
1124 	if (timestamp0)
1125 		delta = timestamp - timestamp0;
1126 	else
1127 		delta = 0;
1128 
1129 	if (delta < 0) {
1130 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1131 		return -1;
1132 	}
1133 
1134 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1135 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1136 	if (sched_out == NULL || sched_in == NULL)
1137 		goto out_put;
1138 
1139 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1140 	if (!out_events) {
1141 		if (thread_atoms_insert(sched, sched_out))
1142 			goto out_put;
1143 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1144 		if (!out_events) {
1145 			pr_err("out-event: Internal tree error");
1146 			goto out_put;
1147 		}
1148 	}
1149 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1150 		return -1;
1151 
1152 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1153 	if (!in_events) {
1154 		if (thread_atoms_insert(sched, sched_in))
1155 			goto out_put;
1156 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1157 		if (!in_events) {
1158 			pr_err("in-event: Internal tree error");
1159 			goto out_put;
1160 		}
1161 		/*
1162 		 * Take came in we have not heard about yet,
1163 		 * add in an initial atom in runnable state:
1164 		 */
1165 		if (add_sched_out_event(in_events, 'R', timestamp))
1166 			goto out_put;
1167 	}
1168 	add_sched_in_event(in_events, timestamp);
1169 	err = 0;
1170 out_put:
1171 	thread__put(sched_out);
1172 	thread__put(sched_in);
1173 	return err;
1174 }
1175 
1176 static int latency_runtime_event(struct perf_sched *sched,
1177 				 struct evsel *evsel,
1178 				 struct perf_sample *sample,
1179 				 struct machine *machine)
1180 {
1181 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1182 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1183 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1184 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1185 	u64 timestamp = sample->time;
1186 	int cpu = sample->cpu, err = -1;
1187 
1188 	if (thread == NULL)
1189 		return -1;
1190 
1191 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1192 	if (!atoms) {
1193 		if (thread_atoms_insert(sched, thread))
1194 			goto out_put;
1195 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1196 		if (!atoms) {
1197 			pr_err("in-event: Internal tree error");
1198 			goto out_put;
1199 		}
1200 		if (add_sched_out_event(atoms, 'R', timestamp))
1201 			goto out_put;
1202 	}
1203 
1204 	add_runtime_event(atoms, runtime, timestamp);
1205 	err = 0;
1206 out_put:
1207 	thread__put(thread);
1208 	return err;
1209 }
1210 
1211 static int latency_wakeup_event(struct perf_sched *sched,
1212 				struct evsel *evsel,
1213 				struct perf_sample *sample,
1214 				struct machine *machine)
1215 {
1216 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1217 	struct work_atoms *atoms;
1218 	struct work_atom *atom;
1219 	struct thread *wakee;
1220 	u64 timestamp = sample->time;
1221 	int err = -1;
1222 
1223 	wakee = machine__findnew_thread(machine, -1, pid);
1224 	if (wakee == NULL)
1225 		return -1;
1226 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1227 	if (!atoms) {
1228 		if (thread_atoms_insert(sched, wakee))
1229 			goto out_put;
1230 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1231 		if (!atoms) {
1232 			pr_err("wakeup-event: Internal tree error");
1233 			goto out_put;
1234 		}
1235 		if (add_sched_out_event(atoms, 'S', timestamp))
1236 			goto out_put;
1237 	}
1238 
1239 	BUG_ON(list_empty(&atoms->work_list));
1240 
1241 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1242 
1243 	/*
1244 	 * As we do not guarantee the wakeup event happens when
1245 	 * task is out of run queue, also may happen when task is
1246 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1247 	 * then we should not set the ->wake_up_time when wake up a
1248 	 * task which is on run queue.
1249 	 *
1250 	 * You WILL be missing events if you've recorded only
1251 	 * one CPU, or are only looking at only one, so don't
1252 	 * skip in this case.
1253 	 */
1254 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1255 		goto out_ok;
1256 
1257 	sched->nr_timestamps++;
1258 	if (atom->sched_out_time > timestamp) {
1259 		sched->nr_unordered_timestamps++;
1260 		goto out_ok;
1261 	}
1262 
1263 	atom->state = THREAD_WAIT_CPU;
1264 	atom->wake_up_time = timestamp;
1265 out_ok:
1266 	err = 0;
1267 out_put:
1268 	thread__put(wakee);
1269 	return err;
1270 }
1271 
1272 static int latency_migrate_task_event(struct perf_sched *sched,
1273 				      struct evsel *evsel,
1274 				      struct perf_sample *sample,
1275 				      struct machine *machine)
1276 {
1277 	const u32 pid = evsel__intval(evsel, sample, "pid");
1278 	u64 timestamp = sample->time;
1279 	struct work_atoms *atoms;
1280 	struct work_atom *atom;
1281 	struct thread *migrant;
1282 	int err = -1;
1283 
1284 	/*
1285 	 * Only need to worry about migration when profiling one CPU.
1286 	 */
1287 	if (sched->profile_cpu == -1)
1288 		return 0;
1289 
1290 	migrant = machine__findnew_thread(machine, -1, pid);
1291 	if (migrant == NULL)
1292 		return -1;
1293 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1294 	if (!atoms) {
1295 		if (thread_atoms_insert(sched, migrant))
1296 			goto out_put;
1297 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1298 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1299 		if (!atoms) {
1300 			pr_err("migration-event: Internal tree error");
1301 			goto out_put;
1302 		}
1303 		if (add_sched_out_event(atoms, 'R', timestamp))
1304 			goto out_put;
1305 	}
1306 
1307 	BUG_ON(list_empty(&atoms->work_list));
1308 
1309 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1310 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1311 
1312 	sched->nr_timestamps++;
1313 
1314 	if (atom->sched_out_time > timestamp)
1315 		sched->nr_unordered_timestamps++;
1316 	err = 0;
1317 out_put:
1318 	thread__put(migrant);
1319 	return err;
1320 }
1321 
1322 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1323 {
1324 	int i;
1325 	int ret;
1326 	u64 avg;
1327 	char max_lat_start[32], max_lat_end[32];
1328 
1329 	if (!work_list->nb_atoms)
1330 		return;
1331 	/*
1332 	 * Ignore idle threads:
1333 	 */
1334 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1335 		return;
1336 
1337 	sched->all_runtime += work_list->total_runtime;
1338 	sched->all_count   += work_list->nb_atoms;
1339 
1340 	if (work_list->num_merged > 1)
1341 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1342 	else
1343 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
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 	if (l->thread == r->thread)
1362 		return 0;
1363 	if (l->thread->tid < r->thread->tid)
1364 		return -1;
1365 	if (l->thread->tid > r->thread->tid)
1366 		return 1;
1367 	return (int)(l->thread - r->thread);
1368 }
1369 
1370 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1371 {
1372 	u64 avgl, avgr;
1373 
1374 	if (!l->nb_atoms)
1375 		return -1;
1376 
1377 	if (!r->nb_atoms)
1378 		return 1;
1379 
1380 	avgl = l->total_lat / l->nb_atoms;
1381 	avgr = r->total_lat / r->nb_atoms;
1382 
1383 	if (avgl < avgr)
1384 		return -1;
1385 	if (avgl > avgr)
1386 		return 1;
1387 
1388 	return 0;
1389 }
1390 
1391 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1392 {
1393 	if (l->max_lat < r->max_lat)
1394 		return -1;
1395 	if (l->max_lat > r->max_lat)
1396 		return 1;
1397 
1398 	return 0;
1399 }
1400 
1401 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1402 {
1403 	if (l->nb_atoms < r->nb_atoms)
1404 		return -1;
1405 	if (l->nb_atoms > r->nb_atoms)
1406 		return 1;
1407 
1408 	return 0;
1409 }
1410 
1411 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1412 {
1413 	if (l->total_runtime < r->total_runtime)
1414 		return -1;
1415 	if (l->total_runtime > r->total_runtime)
1416 		return 1;
1417 
1418 	return 0;
1419 }
1420 
1421 static int sort_dimension__add(const char *tok, struct list_head *list)
1422 {
1423 	size_t i;
1424 	static struct sort_dimension avg_sort_dimension = {
1425 		.name = "avg",
1426 		.cmp  = avg_cmp,
1427 	};
1428 	static struct sort_dimension max_sort_dimension = {
1429 		.name = "max",
1430 		.cmp  = max_cmp,
1431 	};
1432 	static struct sort_dimension pid_sort_dimension = {
1433 		.name = "pid",
1434 		.cmp  = pid_cmp,
1435 	};
1436 	static struct sort_dimension runtime_sort_dimension = {
1437 		.name = "runtime",
1438 		.cmp  = runtime_cmp,
1439 	};
1440 	static struct sort_dimension switch_sort_dimension = {
1441 		.name = "switch",
1442 		.cmp  = switch_cmp,
1443 	};
1444 	struct sort_dimension *available_sorts[] = {
1445 		&pid_sort_dimension,
1446 		&avg_sort_dimension,
1447 		&max_sort_dimension,
1448 		&switch_sort_dimension,
1449 		&runtime_sort_dimension,
1450 	};
1451 
1452 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1453 		if (!strcmp(available_sorts[i]->name, tok)) {
1454 			list_add_tail(&available_sorts[i]->list, list);
1455 
1456 			return 0;
1457 		}
1458 	}
1459 
1460 	return -1;
1461 }
1462 
1463 static void perf_sched__sort_lat(struct perf_sched *sched)
1464 {
1465 	struct rb_node *node;
1466 	struct rb_root_cached *root = &sched->atom_root;
1467 again:
1468 	for (;;) {
1469 		struct work_atoms *data;
1470 		node = rb_first_cached(root);
1471 		if (!node)
1472 			break;
1473 
1474 		rb_erase_cached(node, root);
1475 		data = rb_entry(node, struct work_atoms, node);
1476 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1477 	}
1478 	if (root == &sched->atom_root) {
1479 		root = &sched->merged_atom_root;
1480 		goto again;
1481 	}
1482 }
1483 
1484 static int process_sched_wakeup_event(struct perf_tool *tool,
1485 				      struct evsel *evsel,
1486 				      struct perf_sample *sample,
1487 				      struct machine *machine)
1488 {
1489 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1490 
1491 	if (sched->tp_handler->wakeup_event)
1492 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1493 
1494 	return 0;
1495 }
1496 
1497 union map_priv {
1498 	void	*ptr;
1499 	bool	 color;
1500 };
1501 
1502 static bool thread__has_color(struct thread *thread)
1503 {
1504 	union map_priv priv = {
1505 		.ptr = thread__priv(thread),
1506 	};
1507 
1508 	return priv.color;
1509 }
1510 
1511 static struct thread*
1512 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1513 {
1514 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1515 	union map_priv priv = {
1516 		.color = false,
1517 	};
1518 
1519 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1520 		return thread;
1521 
1522 	if (thread_map__has(sched->map.color_pids, tid))
1523 		priv.color = true;
1524 
1525 	thread__set_priv(thread, priv.ptr);
1526 	return thread;
1527 }
1528 
1529 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1530 			    struct perf_sample *sample, struct machine *machine)
1531 {
1532 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1533 	struct thread *sched_in;
1534 	struct thread_runtime *tr;
1535 	int new_shortname;
1536 	u64 timestamp0, timestamp = sample->time;
1537 	s64 delta;
1538 	int i, this_cpu = sample->cpu;
1539 	int cpus_nr;
1540 	bool new_cpu = false;
1541 	const char *color = PERF_COLOR_NORMAL;
1542 	char stimestamp[32];
1543 
1544 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1545 
1546 	if (this_cpu > sched->max_cpu)
1547 		sched->max_cpu = this_cpu;
1548 
1549 	if (sched->map.comp) {
1550 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1551 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1552 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1553 			new_cpu = true;
1554 		}
1555 	} else
1556 		cpus_nr = sched->max_cpu;
1557 
1558 	timestamp0 = sched->cpu_last_switched[this_cpu];
1559 	sched->cpu_last_switched[this_cpu] = timestamp;
1560 	if (timestamp0)
1561 		delta = timestamp - timestamp0;
1562 	else
1563 		delta = 0;
1564 
1565 	if (delta < 0) {
1566 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1567 		return -1;
1568 	}
1569 
1570 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1571 	if (sched_in == NULL)
1572 		return -1;
1573 
1574 	tr = thread__get_runtime(sched_in);
1575 	if (tr == NULL) {
1576 		thread__put(sched_in);
1577 		return -1;
1578 	}
1579 
1580 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1581 
1582 	printf("  ");
1583 
1584 	new_shortname = 0;
1585 	if (!tr->shortname[0]) {
1586 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1587 			/*
1588 			 * Don't allocate a letter-number for swapper:0
1589 			 * as a shortname. Instead, we use '.' for it.
1590 			 */
1591 			tr->shortname[0] = '.';
1592 			tr->shortname[1] = ' ';
1593 		} else {
1594 			tr->shortname[0] = sched->next_shortname1;
1595 			tr->shortname[1] = sched->next_shortname2;
1596 
1597 			if (sched->next_shortname1 < 'Z') {
1598 				sched->next_shortname1++;
1599 			} else {
1600 				sched->next_shortname1 = 'A';
1601 				if (sched->next_shortname2 < '9')
1602 					sched->next_shortname2++;
1603 				else
1604 					sched->next_shortname2 = '0';
1605 			}
1606 		}
1607 		new_shortname = 1;
1608 	}
1609 
1610 	for (i = 0; i < cpus_nr; i++) {
1611 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1612 		struct thread *curr_thread = sched->curr_thread[cpu];
1613 		struct thread_runtime *curr_tr;
1614 		const char *pid_color = color;
1615 		const char *cpu_color = color;
1616 
1617 		if (curr_thread && thread__has_color(curr_thread))
1618 			pid_color = COLOR_PIDS;
1619 
1620 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1621 			continue;
1622 
1623 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1624 			cpu_color = COLOR_CPUS;
1625 
1626 		if (cpu != this_cpu)
1627 			color_fprintf(stdout, color, " ");
1628 		else
1629 			color_fprintf(stdout, cpu_color, "*");
1630 
1631 		if (sched->curr_thread[cpu]) {
1632 			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1633 			if (curr_tr == NULL) {
1634 				thread__put(sched_in);
1635 				return -1;
1636 			}
1637 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1638 		} else
1639 			color_fprintf(stdout, color, "   ");
1640 	}
1641 
1642 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1643 		goto out;
1644 
1645 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1646 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1647 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1648 		const char *pid_color = color;
1649 
1650 		if (thread__has_color(sched_in))
1651 			pid_color = COLOR_PIDS;
1652 
1653 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1654 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1655 		tr->comm_changed = false;
1656 	}
1657 
1658 	if (sched->map.comp && new_cpu)
1659 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1660 
1661 out:
1662 	color_fprintf(stdout, color, "\n");
1663 
1664 	thread__put(sched_in);
1665 
1666 	return 0;
1667 }
1668 
1669 static int process_sched_switch_event(struct perf_tool *tool,
1670 				      struct evsel *evsel,
1671 				      struct perf_sample *sample,
1672 				      struct machine *machine)
1673 {
1674 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1675 	int this_cpu = sample->cpu, err = 0;
1676 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1677 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1678 
1679 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1680 		/*
1681 		 * Are we trying to switch away a PID that is
1682 		 * not current?
1683 		 */
1684 		if (sched->curr_pid[this_cpu] != prev_pid)
1685 			sched->nr_context_switch_bugs++;
1686 	}
1687 
1688 	if (sched->tp_handler->switch_event)
1689 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1690 
1691 	sched->curr_pid[this_cpu] = next_pid;
1692 	return err;
1693 }
1694 
1695 static int process_sched_runtime_event(struct perf_tool *tool,
1696 				       struct evsel *evsel,
1697 				       struct perf_sample *sample,
1698 				       struct machine *machine)
1699 {
1700 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1701 
1702 	if (sched->tp_handler->runtime_event)
1703 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1704 
1705 	return 0;
1706 }
1707 
1708 static int perf_sched__process_fork_event(struct perf_tool *tool,
1709 					  union perf_event *event,
1710 					  struct perf_sample *sample,
1711 					  struct machine *machine)
1712 {
1713 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1714 
1715 	/* run the fork event through the perf machineruy */
1716 	perf_event__process_fork(tool, event, sample, machine);
1717 
1718 	/* and then run additional processing needed for this command */
1719 	if (sched->tp_handler->fork_event)
1720 		return sched->tp_handler->fork_event(sched, event, machine);
1721 
1722 	return 0;
1723 }
1724 
1725 static int process_sched_migrate_task_event(struct perf_tool *tool,
1726 					    struct evsel *evsel,
1727 					    struct perf_sample *sample,
1728 					    struct machine *machine)
1729 {
1730 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1731 
1732 	if (sched->tp_handler->migrate_task_event)
1733 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1734 
1735 	return 0;
1736 }
1737 
1738 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1739 				  struct evsel *evsel,
1740 				  struct perf_sample *sample,
1741 				  struct machine *machine);
1742 
1743 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1744 						 union perf_event *event __maybe_unused,
1745 						 struct perf_sample *sample,
1746 						 struct evsel *evsel,
1747 						 struct machine *machine)
1748 {
1749 	int err = 0;
1750 
1751 	if (evsel->handler != NULL) {
1752 		tracepoint_handler f = evsel->handler;
1753 		err = f(tool, evsel, sample, machine);
1754 	}
1755 
1756 	return err;
1757 }
1758 
1759 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1760 				    union perf_event *event,
1761 				    struct perf_sample *sample,
1762 				    struct machine *machine)
1763 {
1764 	struct thread *thread;
1765 	struct thread_runtime *tr;
1766 	int err;
1767 
1768 	err = perf_event__process_comm(tool, event, sample, machine);
1769 	if (err)
1770 		return err;
1771 
1772 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1773 	if (!thread) {
1774 		pr_err("Internal error: can't find thread\n");
1775 		return -1;
1776 	}
1777 
1778 	tr = thread__get_runtime(thread);
1779 	if (tr == NULL) {
1780 		thread__put(thread);
1781 		return -1;
1782 	}
1783 
1784 	tr->comm_changed = true;
1785 	thread__put(thread);
1786 
1787 	return 0;
1788 }
1789 
1790 static int perf_sched__read_events(struct perf_sched *sched)
1791 {
1792 	const struct evsel_str_handler handlers[] = {
1793 		{ "sched:sched_switch",	      process_sched_switch_event, },
1794 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1795 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1796 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1797 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1798 	};
1799 	struct perf_session *session;
1800 	struct perf_data data = {
1801 		.path  = input_name,
1802 		.mode  = PERF_DATA_MODE_READ,
1803 		.force = sched->force,
1804 	};
1805 	int rc = -1;
1806 
1807 	session = perf_session__new(&data, false, &sched->tool);
1808 	if (IS_ERR(session)) {
1809 		pr_debug("Error creating perf session");
1810 		return PTR_ERR(session);
1811 	}
1812 
1813 	symbol__init(&session->header.env);
1814 
1815 	if (perf_session__set_tracepoints_handlers(session, handlers))
1816 		goto out_delete;
1817 
1818 	if (perf_session__has_traces(session, "record -R")) {
1819 		int err = perf_session__process_events(session);
1820 		if (err) {
1821 			pr_err("Failed to process events, error %d", err);
1822 			goto out_delete;
1823 		}
1824 
1825 		sched->nr_events      = session->evlist->stats.nr_events[0];
1826 		sched->nr_lost_events = session->evlist->stats.total_lost;
1827 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1828 	}
1829 
1830 	rc = 0;
1831 out_delete:
1832 	perf_session__delete(session);
1833 	return rc;
1834 }
1835 
1836 /*
1837  * scheduling times are printed as msec.usec
1838  */
1839 static inline void print_sched_time(unsigned long long nsecs, int width)
1840 {
1841 	unsigned long msecs;
1842 	unsigned long usecs;
1843 
1844 	msecs  = nsecs / NSEC_PER_MSEC;
1845 	nsecs -= msecs * NSEC_PER_MSEC;
1846 	usecs  = nsecs / NSEC_PER_USEC;
1847 	printf("%*lu.%03lu ", width, msecs, usecs);
1848 }
1849 
1850 /*
1851  * returns runtime data for event, allocating memory for it the
1852  * first time it is used.
1853  */
1854 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1855 {
1856 	struct evsel_runtime *r = evsel->priv;
1857 
1858 	if (r == NULL) {
1859 		r = zalloc(sizeof(struct evsel_runtime));
1860 		evsel->priv = r;
1861 	}
1862 
1863 	return r;
1864 }
1865 
1866 /*
1867  * save last time event was seen per cpu
1868  */
1869 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1870 {
1871 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1872 
1873 	if (r == NULL)
1874 		return;
1875 
1876 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1877 		int i, n = __roundup_pow_of_two(cpu+1);
1878 		void *p = r->last_time;
1879 
1880 		p = realloc(r->last_time, n * sizeof(u64));
1881 		if (!p)
1882 			return;
1883 
1884 		r->last_time = p;
1885 		for (i = r->ncpu; i < n; ++i)
1886 			r->last_time[i] = (u64) 0;
1887 
1888 		r->ncpu = n;
1889 	}
1890 
1891 	r->last_time[cpu] = timestamp;
1892 }
1893 
1894 /* returns last time this event was seen on the given cpu */
1895 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1896 {
1897 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1898 
1899 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1900 		return 0;
1901 
1902 	return r->last_time[cpu];
1903 }
1904 
1905 static int comm_width = 30;
1906 
1907 static char *timehist_get_commstr(struct thread *thread)
1908 {
1909 	static char str[32];
1910 	const char *comm = thread__comm_str(thread);
1911 	pid_t tid = thread->tid;
1912 	pid_t pid = thread->pid_;
1913 	int n;
1914 
1915 	if (pid == 0)
1916 		n = scnprintf(str, sizeof(str), "%s", comm);
1917 
1918 	else if (tid != pid)
1919 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1920 
1921 	else
1922 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1923 
1924 	if (n > comm_width)
1925 		comm_width = n;
1926 
1927 	return str;
1928 }
1929 
1930 static void timehist_header(struct perf_sched *sched)
1931 {
1932 	u32 ncpus = sched->max_cpu + 1;
1933 	u32 i, j;
1934 
1935 	printf("%15s %6s ", "time", "cpu");
1936 
1937 	if (sched->show_cpu_visual) {
1938 		printf(" ");
1939 		for (i = 0, j = 0; i < ncpus; ++i) {
1940 			printf("%x", j++);
1941 			if (j > 15)
1942 				j = 0;
1943 		}
1944 		printf(" ");
1945 	}
1946 
1947 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1948 		"task name", "wait time", "sch delay", "run time");
1949 
1950 	if (sched->show_state)
1951 		printf("  %s", "state");
1952 
1953 	printf("\n");
1954 
1955 	/*
1956 	 * units row
1957 	 */
1958 	printf("%15s %-6s ", "", "");
1959 
1960 	if (sched->show_cpu_visual)
1961 		printf(" %*s ", ncpus, "");
1962 
1963 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1964 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1965 
1966 	if (sched->show_state)
1967 		printf("  %5s", "");
1968 
1969 	printf("\n");
1970 
1971 	/*
1972 	 * separator
1973 	 */
1974 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1975 
1976 	if (sched->show_cpu_visual)
1977 		printf(" %.*s ", ncpus, graph_dotted_line);
1978 
1979 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1980 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1981 		graph_dotted_line);
1982 
1983 	if (sched->show_state)
1984 		printf("  %.5s", graph_dotted_line);
1985 
1986 	printf("\n");
1987 }
1988 
1989 static char task_state_char(struct thread *thread, int state)
1990 {
1991 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1992 	unsigned bit = state ? ffs(state) : 0;
1993 
1994 	/* 'I' for idle */
1995 	if (thread->tid == 0)
1996 		return 'I';
1997 
1998 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1999 }
2000 
2001 static void timehist_print_sample(struct perf_sched *sched,
2002 				  struct evsel *evsel,
2003 				  struct perf_sample *sample,
2004 				  struct addr_location *al,
2005 				  struct thread *thread,
2006 				  u64 t, int state)
2007 {
2008 	struct thread_runtime *tr = thread__priv(thread);
2009 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2010 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2011 	u32 max_cpus = sched->max_cpu + 1;
2012 	char tstr[64];
2013 	char nstr[30];
2014 	u64 wait_time;
2015 
2016 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2017 		return;
2018 
2019 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2020 	printf("%15s [%04d] ", tstr, sample->cpu);
2021 
2022 	if (sched->show_cpu_visual) {
2023 		u32 i;
2024 		char c;
2025 
2026 		printf(" ");
2027 		for (i = 0; i < max_cpus; ++i) {
2028 			/* flag idle times with 'i'; others are sched events */
2029 			if (i == sample->cpu)
2030 				c = (thread->tid == 0) ? 'i' : 's';
2031 			else
2032 				c = ' ';
2033 			printf("%c", c);
2034 		}
2035 		printf(" ");
2036 	}
2037 
2038 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2039 
2040 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2041 	print_sched_time(wait_time, 6);
2042 
2043 	print_sched_time(tr->dt_delay, 6);
2044 	print_sched_time(tr->dt_run, 6);
2045 
2046 	if (sched->show_state)
2047 		printf(" %5c ", task_state_char(thread, state));
2048 
2049 	if (sched->show_next) {
2050 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2051 		printf(" %-*s", comm_width, nstr);
2052 	}
2053 
2054 	if (sched->show_wakeups && !sched->show_next)
2055 		printf("  %-*s", comm_width, "");
2056 
2057 	if (thread->tid == 0)
2058 		goto out;
2059 
2060 	if (sched->show_callchain)
2061 		printf("  ");
2062 
2063 	sample__fprintf_sym(sample, al, 0,
2064 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2065 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2066 			    EVSEL__PRINT_SKIP_IGNORED,
2067 			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2068 
2069 out:
2070 	printf("\n");
2071 }
2072 
2073 /*
2074  * Explanation of delta-time stats:
2075  *
2076  *            t = time of current schedule out event
2077  *        tprev = time of previous sched out event
2078  *                also time of schedule-in event for current task
2079  *    last_time = time of last sched change event for current task
2080  *                (i.e, time process was last scheduled out)
2081  * ready_to_run = time of wakeup for current task
2082  *
2083  * -----|------------|------------|------------|------
2084  *    last         ready        tprev          t
2085  *    time         to run
2086  *
2087  *      |-------- dt_wait --------|
2088  *                   |- dt_delay -|-- dt_run --|
2089  *
2090  *   dt_run = run time of current task
2091  *  dt_wait = time between last schedule out event for task and tprev
2092  *            represents time spent off the cpu
2093  * dt_delay = time between wakeup and schedule-in of task
2094  */
2095 
2096 static void timehist_update_runtime_stats(struct thread_runtime *r,
2097 					 u64 t, u64 tprev)
2098 {
2099 	r->dt_delay   = 0;
2100 	r->dt_sleep   = 0;
2101 	r->dt_iowait  = 0;
2102 	r->dt_preempt = 0;
2103 	r->dt_run     = 0;
2104 
2105 	if (tprev) {
2106 		r->dt_run = t - tprev;
2107 		if (r->ready_to_run) {
2108 			if (r->ready_to_run > tprev)
2109 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2110 			else
2111 				r->dt_delay = tprev - r->ready_to_run;
2112 		}
2113 
2114 		if (r->last_time > tprev)
2115 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2116 		else if (r->last_time) {
2117 			u64 dt_wait = tprev - r->last_time;
2118 
2119 			if (r->last_state == TASK_RUNNING)
2120 				r->dt_preempt = dt_wait;
2121 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2122 				r->dt_iowait = dt_wait;
2123 			else
2124 				r->dt_sleep = dt_wait;
2125 		}
2126 	}
2127 
2128 	update_stats(&r->run_stats, r->dt_run);
2129 
2130 	r->total_run_time     += r->dt_run;
2131 	r->total_delay_time   += r->dt_delay;
2132 	r->total_sleep_time   += r->dt_sleep;
2133 	r->total_iowait_time  += r->dt_iowait;
2134 	r->total_preempt_time += r->dt_preempt;
2135 }
2136 
2137 static bool is_idle_sample(struct perf_sample *sample,
2138 			   struct evsel *evsel)
2139 {
2140 	/* pid 0 == swapper == idle task */
2141 	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2142 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2143 
2144 	return sample->pid == 0;
2145 }
2146 
2147 static void save_task_callchain(struct perf_sched *sched,
2148 				struct perf_sample *sample,
2149 				struct evsel *evsel,
2150 				struct machine *machine)
2151 {
2152 	struct callchain_cursor *cursor = &callchain_cursor;
2153 	struct thread *thread;
2154 
2155 	/* want main thread for process - has maps */
2156 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2157 	if (thread == NULL) {
2158 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2159 		return;
2160 	}
2161 
2162 	if (!sched->show_callchain || sample->callchain == NULL)
2163 		return;
2164 
2165 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2166 				      NULL, NULL, sched->max_stack + 2) != 0) {
2167 		if (verbose > 0)
2168 			pr_err("Failed to resolve callchain. Skipping\n");
2169 
2170 		return;
2171 	}
2172 
2173 	callchain_cursor_commit(cursor);
2174 
2175 	while (true) {
2176 		struct callchain_cursor_node *node;
2177 		struct symbol *sym;
2178 
2179 		node = callchain_cursor_current(cursor);
2180 		if (node == NULL)
2181 			break;
2182 
2183 		sym = node->ms.sym;
2184 		if (sym) {
2185 			if (!strcmp(sym->name, "schedule") ||
2186 			    !strcmp(sym->name, "__schedule") ||
2187 			    !strcmp(sym->name, "preempt_schedule"))
2188 				sym->ignore = 1;
2189 		}
2190 
2191 		callchain_cursor_advance(cursor);
2192 	}
2193 }
2194 
2195 static int init_idle_thread(struct thread *thread)
2196 {
2197 	struct idle_thread_runtime *itr;
2198 
2199 	thread__set_comm(thread, idle_comm, 0);
2200 
2201 	itr = zalloc(sizeof(*itr));
2202 	if (itr == NULL)
2203 		return -ENOMEM;
2204 
2205 	init_stats(&itr->tr.run_stats);
2206 	callchain_init(&itr->callchain);
2207 	callchain_cursor_reset(&itr->cursor);
2208 	thread__set_priv(thread, itr);
2209 
2210 	return 0;
2211 }
2212 
2213 /*
2214  * Track idle stats per cpu by maintaining a local thread
2215  * struct for the idle task on each cpu.
2216  */
2217 static int init_idle_threads(int ncpu)
2218 {
2219 	int i, ret;
2220 
2221 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2222 	if (!idle_threads)
2223 		return -ENOMEM;
2224 
2225 	idle_max_cpu = ncpu;
2226 
2227 	/* allocate the actual thread struct if needed */
2228 	for (i = 0; i < ncpu; ++i) {
2229 		idle_threads[i] = thread__new(0, 0);
2230 		if (idle_threads[i] == NULL)
2231 			return -ENOMEM;
2232 
2233 		ret = init_idle_thread(idle_threads[i]);
2234 		if (ret < 0)
2235 			return ret;
2236 	}
2237 
2238 	return 0;
2239 }
2240 
2241 static void free_idle_threads(void)
2242 {
2243 	int i;
2244 
2245 	if (idle_threads == NULL)
2246 		return;
2247 
2248 	for (i = 0; i < idle_max_cpu; ++i) {
2249 		if ((idle_threads[i]))
2250 			thread__delete(idle_threads[i]);
2251 	}
2252 
2253 	free(idle_threads);
2254 }
2255 
2256 static struct thread *get_idle_thread(int cpu)
2257 {
2258 	/*
2259 	 * expand/allocate array of pointers to local thread
2260 	 * structs if needed
2261 	 */
2262 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2263 		int i, j = __roundup_pow_of_two(cpu+1);
2264 		void *p;
2265 
2266 		p = realloc(idle_threads, j * sizeof(struct thread *));
2267 		if (!p)
2268 			return NULL;
2269 
2270 		idle_threads = (struct thread **) p;
2271 		for (i = idle_max_cpu; i < j; ++i)
2272 			idle_threads[i] = NULL;
2273 
2274 		idle_max_cpu = j;
2275 	}
2276 
2277 	/* allocate a new thread struct if needed */
2278 	if (idle_threads[cpu] == NULL) {
2279 		idle_threads[cpu] = thread__new(0, 0);
2280 
2281 		if (idle_threads[cpu]) {
2282 			if (init_idle_thread(idle_threads[cpu]) < 0)
2283 				return NULL;
2284 		}
2285 	}
2286 
2287 	return idle_threads[cpu];
2288 }
2289 
2290 static void save_idle_callchain(struct perf_sched *sched,
2291 				struct idle_thread_runtime *itr,
2292 				struct perf_sample *sample)
2293 {
2294 	if (!sched->show_callchain || sample->callchain == NULL)
2295 		return;
2296 
2297 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2298 }
2299 
2300 static struct thread *timehist_get_thread(struct perf_sched *sched,
2301 					  struct perf_sample *sample,
2302 					  struct machine *machine,
2303 					  struct evsel *evsel)
2304 {
2305 	struct thread *thread;
2306 
2307 	if (is_idle_sample(sample, evsel)) {
2308 		thread = get_idle_thread(sample->cpu);
2309 		if (thread == NULL)
2310 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2311 
2312 	} else {
2313 		/* there were samples with tid 0 but non-zero pid */
2314 		thread = machine__findnew_thread(machine, sample->pid,
2315 						 sample->tid ?: sample->pid);
2316 		if (thread == NULL) {
2317 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2318 				 sample->tid);
2319 		}
2320 
2321 		save_task_callchain(sched, sample, evsel, machine);
2322 		if (sched->idle_hist) {
2323 			struct thread *idle;
2324 			struct idle_thread_runtime *itr;
2325 
2326 			idle = get_idle_thread(sample->cpu);
2327 			if (idle == NULL) {
2328 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2329 				return NULL;
2330 			}
2331 
2332 			itr = thread__priv(idle);
2333 			if (itr == NULL)
2334 				return NULL;
2335 
2336 			itr->last_thread = thread;
2337 
2338 			/* copy task callchain when entering to idle */
2339 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2340 				save_idle_callchain(sched, itr, sample);
2341 		}
2342 	}
2343 
2344 	return thread;
2345 }
2346 
2347 static bool timehist_skip_sample(struct perf_sched *sched,
2348 				 struct thread *thread,
2349 				 struct evsel *evsel,
2350 				 struct perf_sample *sample)
2351 {
2352 	bool rc = false;
2353 
2354 	if (thread__is_filtered(thread)) {
2355 		rc = true;
2356 		sched->skipped_samples++;
2357 	}
2358 
2359 	if (sched->idle_hist) {
2360 		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2361 			rc = true;
2362 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2363 			 evsel__intval(evsel, sample, "next_pid") != 0)
2364 			rc = true;
2365 	}
2366 
2367 	return rc;
2368 }
2369 
2370 static void timehist_print_wakeup_event(struct perf_sched *sched,
2371 					struct evsel *evsel,
2372 					struct perf_sample *sample,
2373 					struct machine *machine,
2374 					struct thread *awakened)
2375 {
2376 	struct thread *thread;
2377 	char tstr[64];
2378 
2379 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2380 	if (thread == NULL)
2381 		return;
2382 
2383 	/* show wakeup unless both awakee and awaker are filtered */
2384 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2385 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2386 		return;
2387 	}
2388 
2389 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2390 	printf("%15s [%04d] ", tstr, sample->cpu);
2391 	if (sched->show_cpu_visual)
2392 		printf(" %*s ", sched->max_cpu + 1, "");
2393 
2394 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2395 
2396 	/* dt spacer */
2397 	printf("  %9s  %9s  %9s ", "", "", "");
2398 
2399 	printf("awakened: %s", timehist_get_commstr(awakened));
2400 
2401 	printf("\n");
2402 }
2403 
2404 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2405 					union perf_event *event __maybe_unused,
2406 					struct evsel *evsel __maybe_unused,
2407 					struct perf_sample *sample __maybe_unused,
2408 					struct machine *machine __maybe_unused)
2409 {
2410 	return 0;
2411 }
2412 
2413 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2414 				       union perf_event *event __maybe_unused,
2415 				       struct evsel *evsel,
2416 				       struct perf_sample *sample,
2417 				       struct machine *machine)
2418 {
2419 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2420 	struct thread *thread;
2421 	struct thread_runtime *tr = NULL;
2422 	/* want pid of awakened task not pid in sample */
2423 	const u32 pid = evsel__intval(evsel, sample, "pid");
2424 
2425 	thread = machine__findnew_thread(machine, 0, pid);
2426 	if (thread == NULL)
2427 		return -1;
2428 
2429 	tr = thread__get_runtime(thread);
2430 	if (tr == NULL)
2431 		return -1;
2432 
2433 	if (tr->ready_to_run == 0)
2434 		tr->ready_to_run = sample->time;
2435 
2436 	/* show wakeups if requested */
2437 	if (sched->show_wakeups &&
2438 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2439 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2440 
2441 	return 0;
2442 }
2443 
2444 static void timehist_print_migration_event(struct perf_sched *sched,
2445 					struct evsel *evsel,
2446 					struct perf_sample *sample,
2447 					struct machine *machine,
2448 					struct thread *migrated)
2449 {
2450 	struct thread *thread;
2451 	char tstr[64];
2452 	u32 max_cpus = sched->max_cpu + 1;
2453 	u32 ocpu, dcpu;
2454 
2455 	if (sched->summary_only)
2456 		return;
2457 
2458 	max_cpus = sched->max_cpu + 1;
2459 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2460 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2461 
2462 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2463 	if (thread == NULL)
2464 		return;
2465 
2466 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2467 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2468 		return;
2469 	}
2470 
2471 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2472 	printf("%15s [%04d] ", tstr, sample->cpu);
2473 
2474 	if (sched->show_cpu_visual) {
2475 		u32 i;
2476 		char c;
2477 
2478 		printf("  ");
2479 		for (i = 0; i < max_cpus; ++i) {
2480 			c = (i == sample->cpu) ? 'm' : ' ';
2481 			printf("%c", c);
2482 		}
2483 		printf("  ");
2484 	}
2485 
2486 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2487 
2488 	/* dt spacer */
2489 	printf("  %9s  %9s  %9s ", "", "", "");
2490 
2491 	printf("migrated: %s", timehist_get_commstr(migrated));
2492 	printf(" cpu %d => %d", ocpu, dcpu);
2493 
2494 	printf("\n");
2495 }
2496 
2497 static int timehist_migrate_task_event(struct perf_tool *tool,
2498 				       union perf_event *event __maybe_unused,
2499 				       struct evsel *evsel,
2500 				       struct perf_sample *sample,
2501 				       struct machine *machine)
2502 {
2503 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2504 	struct thread *thread;
2505 	struct thread_runtime *tr = NULL;
2506 	/* want pid of migrated task not pid in sample */
2507 	const u32 pid = evsel__intval(evsel, sample, "pid");
2508 
2509 	thread = machine__findnew_thread(machine, 0, pid);
2510 	if (thread == NULL)
2511 		return -1;
2512 
2513 	tr = thread__get_runtime(thread);
2514 	if (tr == NULL)
2515 		return -1;
2516 
2517 	tr->migrations++;
2518 
2519 	/* show migrations if requested */
2520 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2521 
2522 	return 0;
2523 }
2524 
2525 static int timehist_sched_change_event(struct perf_tool *tool,
2526 				       union perf_event *event,
2527 				       struct evsel *evsel,
2528 				       struct perf_sample *sample,
2529 				       struct machine *machine)
2530 {
2531 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2532 	struct perf_time_interval *ptime = &sched->ptime;
2533 	struct addr_location al;
2534 	struct thread *thread;
2535 	struct thread_runtime *tr = NULL;
2536 	u64 tprev, t = sample->time;
2537 	int rc = 0;
2538 	int state = evsel__intval(evsel, sample, "prev_state");
2539 
2540 	if (machine__resolve(machine, &al, sample) < 0) {
2541 		pr_err("problem processing %d event. skipping it\n",
2542 		       event->header.type);
2543 		rc = -1;
2544 		goto out;
2545 	}
2546 
2547 	thread = timehist_get_thread(sched, sample, machine, evsel);
2548 	if (thread == NULL) {
2549 		rc = -1;
2550 		goto out;
2551 	}
2552 
2553 	if (timehist_skip_sample(sched, thread, evsel, sample))
2554 		goto out;
2555 
2556 	tr = thread__get_runtime(thread);
2557 	if (tr == NULL) {
2558 		rc = -1;
2559 		goto out;
2560 	}
2561 
2562 	tprev = evsel__get_time(evsel, sample->cpu);
2563 
2564 	/*
2565 	 * If start time given:
2566 	 * - sample time is under window user cares about - skip sample
2567 	 * - tprev is under window user cares about  - reset to start of window
2568 	 */
2569 	if (ptime->start && ptime->start > t)
2570 		goto out;
2571 
2572 	if (tprev && ptime->start > tprev)
2573 		tprev = ptime->start;
2574 
2575 	/*
2576 	 * If end time given:
2577 	 * - previous sched event is out of window - we are done
2578 	 * - sample time is beyond window user cares about - reset it
2579 	 *   to close out stats for time window interest
2580 	 */
2581 	if (ptime->end) {
2582 		if (tprev > ptime->end)
2583 			goto out;
2584 
2585 		if (t > ptime->end)
2586 			t = ptime->end;
2587 	}
2588 
2589 	if (!sched->idle_hist || thread->tid == 0) {
2590 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2591 			timehist_update_runtime_stats(tr, t, tprev);
2592 
2593 		if (sched->idle_hist) {
2594 			struct idle_thread_runtime *itr = (void *)tr;
2595 			struct thread_runtime *last_tr;
2596 
2597 			BUG_ON(thread->tid != 0);
2598 
2599 			if (itr->last_thread == NULL)
2600 				goto out;
2601 
2602 			/* add current idle time as last thread's runtime */
2603 			last_tr = thread__get_runtime(itr->last_thread);
2604 			if (last_tr == NULL)
2605 				goto out;
2606 
2607 			timehist_update_runtime_stats(last_tr, t, tprev);
2608 			/*
2609 			 * remove delta time of last thread as it's not updated
2610 			 * and otherwise it will show an invalid value next
2611 			 * time.  we only care total run time and run stat.
2612 			 */
2613 			last_tr->dt_run = 0;
2614 			last_tr->dt_delay = 0;
2615 			last_tr->dt_sleep = 0;
2616 			last_tr->dt_iowait = 0;
2617 			last_tr->dt_preempt = 0;
2618 
2619 			if (itr->cursor.nr)
2620 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2621 
2622 			itr->last_thread = NULL;
2623 		}
2624 	}
2625 
2626 	if (!sched->summary_only)
2627 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2628 
2629 out:
2630 	if (sched->hist_time.start == 0 && t >= ptime->start)
2631 		sched->hist_time.start = t;
2632 	if (ptime->end == 0 || t <= ptime->end)
2633 		sched->hist_time.end = t;
2634 
2635 	if (tr) {
2636 		/* time of this sched_switch event becomes last time task seen */
2637 		tr->last_time = sample->time;
2638 
2639 		/* last state is used to determine where to account wait time */
2640 		tr->last_state = state;
2641 
2642 		/* sched out event for task so reset ready to run time */
2643 		tr->ready_to_run = 0;
2644 	}
2645 
2646 	evsel__save_time(evsel, sample->time, sample->cpu);
2647 
2648 	return rc;
2649 }
2650 
2651 static int timehist_sched_switch_event(struct perf_tool *tool,
2652 			     union perf_event *event,
2653 			     struct evsel *evsel,
2654 			     struct perf_sample *sample,
2655 			     struct machine *machine __maybe_unused)
2656 {
2657 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2658 }
2659 
2660 static int process_lost(struct perf_tool *tool __maybe_unused,
2661 			union perf_event *event,
2662 			struct perf_sample *sample,
2663 			struct machine *machine __maybe_unused)
2664 {
2665 	char tstr[64];
2666 
2667 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2668 	printf("%15s ", tstr);
2669 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2670 
2671 	return 0;
2672 }
2673 
2674 
2675 static void print_thread_runtime(struct thread *t,
2676 				 struct thread_runtime *r)
2677 {
2678 	double mean = avg_stats(&r->run_stats);
2679 	float stddev;
2680 
2681 	printf("%*s   %5d  %9" PRIu64 " ",
2682 	       comm_width, timehist_get_commstr(t), t->ppid,
2683 	       (u64) r->run_stats.n);
2684 
2685 	print_sched_time(r->total_run_time, 8);
2686 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2687 	print_sched_time(r->run_stats.min, 6);
2688 	printf(" ");
2689 	print_sched_time((u64) mean, 6);
2690 	printf(" ");
2691 	print_sched_time(r->run_stats.max, 6);
2692 	printf("  ");
2693 	printf("%5.2f", stddev);
2694 	printf("   %5" PRIu64, r->migrations);
2695 	printf("\n");
2696 }
2697 
2698 static void print_thread_waittime(struct thread *t,
2699 				  struct thread_runtime *r)
2700 {
2701 	printf("%*s   %5d  %9" PRIu64 " ",
2702 	       comm_width, timehist_get_commstr(t), t->ppid,
2703 	       (u64) r->run_stats.n);
2704 
2705 	print_sched_time(r->total_run_time, 8);
2706 	print_sched_time(r->total_sleep_time, 6);
2707 	printf(" ");
2708 	print_sched_time(r->total_iowait_time, 6);
2709 	printf(" ");
2710 	print_sched_time(r->total_preempt_time, 6);
2711 	printf(" ");
2712 	print_sched_time(r->total_delay_time, 6);
2713 	printf("\n");
2714 }
2715 
2716 struct total_run_stats {
2717 	struct perf_sched *sched;
2718 	u64  sched_count;
2719 	u64  task_count;
2720 	u64  total_run_time;
2721 };
2722 
2723 static int __show_thread_runtime(struct thread *t, void *priv)
2724 {
2725 	struct total_run_stats *stats = priv;
2726 	struct thread_runtime *r;
2727 
2728 	if (thread__is_filtered(t))
2729 		return 0;
2730 
2731 	r = thread__priv(t);
2732 	if (r && r->run_stats.n) {
2733 		stats->task_count++;
2734 		stats->sched_count += r->run_stats.n;
2735 		stats->total_run_time += r->total_run_time;
2736 
2737 		if (stats->sched->show_state)
2738 			print_thread_waittime(t, r);
2739 		else
2740 			print_thread_runtime(t, r);
2741 	}
2742 
2743 	return 0;
2744 }
2745 
2746 static int show_thread_runtime(struct thread *t, void *priv)
2747 {
2748 	if (t->dead)
2749 		return 0;
2750 
2751 	return __show_thread_runtime(t, priv);
2752 }
2753 
2754 static int show_deadthread_runtime(struct thread *t, void *priv)
2755 {
2756 	if (!t->dead)
2757 		return 0;
2758 
2759 	return __show_thread_runtime(t, priv);
2760 }
2761 
2762 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2763 {
2764 	const char *sep = " <- ";
2765 	struct callchain_list *chain;
2766 	size_t ret = 0;
2767 	char bf[1024];
2768 	bool first;
2769 
2770 	if (node == NULL)
2771 		return 0;
2772 
2773 	ret = callchain__fprintf_folded(fp, node->parent);
2774 	first = (ret == 0);
2775 
2776 	list_for_each_entry(chain, &node->val, list) {
2777 		if (chain->ip >= PERF_CONTEXT_MAX)
2778 			continue;
2779 		if (chain->ms.sym && chain->ms.sym->ignore)
2780 			continue;
2781 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2782 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2783 							false));
2784 		first = false;
2785 	}
2786 
2787 	return ret;
2788 }
2789 
2790 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2791 {
2792 	size_t ret = 0;
2793 	FILE *fp = stdout;
2794 	struct callchain_node *chain;
2795 	struct rb_node *rb_node = rb_first_cached(root);
2796 
2797 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2798 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2799 	       graph_dotted_line);
2800 
2801 	while (rb_node) {
2802 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2803 		rb_node = rb_next(rb_node);
2804 
2805 		ret += fprintf(fp, "  ");
2806 		print_sched_time(chain->hit, 12);
2807 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2808 		ret += fprintf(fp, " %8d  ", chain->count);
2809 		ret += callchain__fprintf_folded(fp, chain);
2810 		ret += fprintf(fp, "\n");
2811 	}
2812 
2813 	return ret;
2814 }
2815 
2816 static void timehist_print_summary(struct perf_sched *sched,
2817 				   struct perf_session *session)
2818 {
2819 	struct machine *m = &session->machines.host;
2820 	struct total_run_stats totals;
2821 	u64 task_count;
2822 	struct thread *t;
2823 	struct thread_runtime *r;
2824 	int i;
2825 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2826 
2827 	memset(&totals, 0, sizeof(totals));
2828 	totals.sched = sched;
2829 
2830 	if (sched->idle_hist) {
2831 		printf("\nIdle-time summary\n");
2832 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2833 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2834 	} else if (sched->show_state) {
2835 		printf("\nWait-time summary\n");
2836 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2837 		printf("   run-time      sleep      iowait     preempt       delay\n");
2838 	} else {
2839 		printf("\nRuntime summary\n");
2840 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2841 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2842 	}
2843 	printf("%*s            (count)  ", comm_width, "");
2844 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2845 	       sched->show_state ? "(msec)" : "%");
2846 	printf("%.117s\n", graph_dotted_line);
2847 
2848 	machine__for_each_thread(m, show_thread_runtime, &totals);
2849 	task_count = totals.task_count;
2850 	if (!task_count)
2851 		printf("<no still running tasks>\n");
2852 
2853 	printf("\nTerminated tasks:\n");
2854 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2855 	if (task_count == totals.task_count)
2856 		printf("<no terminated tasks>\n");
2857 
2858 	/* CPU idle stats not tracked when samples were skipped */
2859 	if (sched->skipped_samples && !sched->idle_hist)
2860 		return;
2861 
2862 	printf("\nIdle stats:\n");
2863 	for (i = 0; i < idle_max_cpu; ++i) {
2864 		if (cpu_list && !test_bit(i, cpu_bitmap))
2865 			continue;
2866 
2867 		t = idle_threads[i];
2868 		if (!t)
2869 			continue;
2870 
2871 		r = thread__priv(t);
2872 		if (r && r->run_stats.n) {
2873 			totals.sched_count += r->run_stats.n;
2874 			printf("    CPU %2d idle for ", i);
2875 			print_sched_time(r->total_run_time, 6);
2876 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2877 		} else
2878 			printf("    CPU %2d idle entire time window\n", i);
2879 	}
2880 
2881 	if (sched->idle_hist && sched->show_callchain) {
2882 		callchain_param.mode  = CHAIN_FOLDED;
2883 		callchain_param.value = CCVAL_PERIOD;
2884 
2885 		callchain_register_param(&callchain_param);
2886 
2887 		printf("\nIdle stats by callchain:\n");
2888 		for (i = 0; i < idle_max_cpu; ++i) {
2889 			struct idle_thread_runtime *itr;
2890 
2891 			t = idle_threads[i];
2892 			if (!t)
2893 				continue;
2894 
2895 			itr = thread__priv(t);
2896 			if (itr == NULL)
2897 				continue;
2898 
2899 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2900 					     0, &callchain_param);
2901 
2902 			printf("  CPU %2d:", i);
2903 			print_sched_time(itr->tr.total_run_time, 6);
2904 			printf(" msec\n");
2905 			timehist_print_idlehist_callchain(&itr->sorted_root);
2906 			printf("\n");
2907 		}
2908 	}
2909 
2910 	printf("\n"
2911 	       "    Total number of unique tasks: %" PRIu64 "\n"
2912 	       "Total number of context switches: %" PRIu64 "\n",
2913 	       totals.task_count, totals.sched_count);
2914 
2915 	printf("           Total run time (msec): ");
2916 	print_sched_time(totals.total_run_time, 2);
2917 	printf("\n");
2918 
2919 	printf("    Total scheduling time (msec): ");
2920 	print_sched_time(hist_time, 2);
2921 	printf(" (x %d)\n", sched->max_cpu);
2922 }
2923 
2924 typedef int (*sched_handler)(struct perf_tool *tool,
2925 			  union perf_event *event,
2926 			  struct evsel *evsel,
2927 			  struct perf_sample *sample,
2928 			  struct machine *machine);
2929 
2930 static int perf_timehist__process_sample(struct perf_tool *tool,
2931 					 union perf_event *event,
2932 					 struct perf_sample *sample,
2933 					 struct evsel *evsel,
2934 					 struct machine *machine)
2935 {
2936 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2937 	int err = 0;
2938 	int this_cpu = sample->cpu;
2939 
2940 	if (this_cpu > sched->max_cpu)
2941 		sched->max_cpu = this_cpu;
2942 
2943 	if (evsel->handler != NULL) {
2944 		sched_handler f = evsel->handler;
2945 
2946 		err = f(tool, event, evsel, sample, machine);
2947 	}
2948 
2949 	return err;
2950 }
2951 
2952 static int timehist_check_attr(struct perf_sched *sched,
2953 			       struct evlist *evlist)
2954 {
2955 	struct evsel *evsel;
2956 	struct evsel_runtime *er;
2957 
2958 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2959 		er = evsel__get_runtime(evsel);
2960 		if (er == NULL) {
2961 			pr_err("Failed to allocate memory for evsel runtime data\n");
2962 			return -1;
2963 		}
2964 
2965 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2966 			pr_info("Samples do not have callchains.\n");
2967 			sched->show_callchain = 0;
2968 			symbol_conf.use_callchain = 0;
2969 		}
2970 	}
2971 
2972 	return 0;
2973 }
2974 
2975 static int perf_sched__timehist(struct perf_sched *sched)
2976 {
2977 	struct evsel_str_handler handlers[] = {
2978 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2979 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2980 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
2981 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2982 	};
2983 	const struct evsel_str_handler migrate_handlers[] = {
2984 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2985 	};
2986 	struct perf_data data = {
2987 		.path  = input_name,
2988 		.mode  = PERF_DATA_MODE_READ,
2989 		.force = sched->force,
2990 	};
2991 
2992 	struct perf_session *session;
2993 	struct evlist *evlist;
2994 	int err = -1;
2995 
2996 	/*
2997 	 * event handlers for timehist option
2998 	 */
2999 	sched->tool.sample	 = perf_timehist__process_sample;
3000 	sched->tool.mmap	 = perf_event__process_mmap;
3001 	sched->tool.comm	 = perf_event__process_comm;
3002 	sched->tool.exit	 = perf_event__process_exit;
3003 	sched->tool.fork	 = perf_event__process_fork;
3004 	sched->tool.lost	 = process_lost;
3005 	sched->tool.attr	 = perf_event__process_attr;
3006 	sched->tool.tracing_data = perf_event__process_tracing_data;
3007 	sched->tool.build_id	 = perf_event__process_build_id;
3008 
3009 	sched->tool.ordered_events = true;
3010 	sched->tool.ordering_requires_timestamps = true;
3011 
3012 	symbol_conf.use_callchain = sched->show_callchain;
3013 
3014 	session = perf_session__new(&data, false, &sched->tool);
3015 	if (IS_ERR(session))
3016 		return PTR_ERR(session);
3017 
3018 	if (cpu_list) {
3019 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3020 		if (err < 0)
3021 			goto out;
3022 	}
3023 
3024 	evlist = session->evlist;
3025 
3026 	symbol__init(&session->header.env);
3027 
3028 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3029 		pr_err("Invalid time string\n");
3030 		return -EINVAL;
3031 	}
3032 
3033 	if (timehist_check_attr(sched, evlist) != 0)
3034 		goto out;
3035 
3036 	setup_pager();
3037 
3038 	/* prefer sched_waking if it is captured */
3039 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3040 		handlers[1].handler = timehist_sched_wakeup_ignore;
3041 
3042 	/* setup per-evsel handlers */
3043 	if (perf_session__set_tracepoints_handlers(session, handlers))
3044 		goto out;
3045 
3046 	/* sched_switch event at a minimum needs to exist */
3047 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3048 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3049 		goto out;
3050 	}
3051 
3052 	if (sched->show_migrations &&
3053 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3054 		goto out;
3055 
3056 	/* pre-allocate struct for per-CPU idle stats */
3057 	sched->max_cpu = session->header.env.nr_cpus_online;
3058 	if (sched->max_cpu == 0)
3059 		sched->max_cpu = 4;
3060 	if (init_idle_threads(sched->max_cpu))
3061 		goto out;
3062 
3063 	/* summary_only implies summary option, but don't overwrite summary if set */
3064 	if (sched->summary_only)
3065 		sched->summary = sched->summary_only;
3066 
3067 	if (!sched->summary_only)
3068 		timehist_header(sched);
3069 
3070 	err = perf_session__process_events(session);
3071 	if (err) {
3072 		pr_err("Failed to process events, error %d", err);
3073 		goto out;
3074 	}
3075 
3076 	sched->nr_events      = evlist->stats.nr_events[0];
3077 	sched->nr_lost_events = evlist->stats.total_lost;
3078 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3079 
3080 	if (sched->summary)
3081 		timehist_print_summary(sched, session);
3082 
3083 out:
3084 	free_idle_threads();
3085 	perf_session__delete(session);
3086 
3087 	return err;
3088 }
3089 
3090 
3091 static void print_bad_events(struct perf_sched *sched)
3092 {
3093 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3094 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3095 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3096 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3097 	}
3098 	if (sched->nr_lost_events && sched->nr_events) {
3099 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3100 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3101 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3102 	}
3103 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3104 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3105 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3106 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3107 		if (sched->nr_lost_events)
3108 			printf(" (due to lost events?)");
3109 		printf("\n");
3110 	}
3111 }
3112 
3113 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3114 {
3115 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3116 	struct work_atoms *this;
3117 	const char *comm = thread__comm_str(data->thread), *this_comm;
3118 	bool leftmost = true;
3119 
3120 	while (*new) {
3121 		int cmp;
3122 
3123 		this = container_of(*new, struct work_atoms, node);
3124 		parent = *new;
3125 
3126 		this_comm = thread__comm_str(this->thread);
3127 		cmp = strcmp(comm, this_comm);
3128 		if (cmp > 0) {
3129 			new = &((*new)->rb_left);
3130 		} else if (cmp < 0) {
3131 			new = &((*new)->rb_right);
3132 			leftmost = false;
3133 		} else {
3134 			this->num_merged++;
3135 			this->total_runtime += data->total_runtime;
3136 			this->nb_atoms += data->nb_atoms;
3137 			this->total_lat += data->total_lat;
3138 			list_splice(&data->work_list, &this->work_list);
3139 			if (this->max_lat < data->max_lat) {
3140 				this->max_lat = data->max_lat;
3141 				this->max_lat_start = data->max_lat_start;
3142 				this->max_lat_end = data->max_lat_end;
3143 			}
3144 			zfree(&data);
3145 			return;
3146 		}
3147 	}
3148 
3149 	data->num_merged++;
3150 	rb_link_node(&data->node, parent, new);
3151 	rb_insert_color_cached(&data->node, root, leftmost);
3152 }
3153 
3154 static void perf_sched__merge_lat(struct perf_sched *sched)
3155 {
3156 	struct work_atoms *data;
3157 	struct rb_node *node;
3158 
3159 	if (sched->skip_merge)
3160 		return;
3161 
3162 	while ((node = rb_first_cached(&sched->atom_root))) {
3163 		rb_erase_cached(node, &sched->atom_root);
3164 		data = rb_entry(node, struct work_atoms, node);
3165 		__merge_work_atoms(&sched->merged_atom_root, data);
3166 	}
3167 }
3168 
3169 static int perf_sched__lat(struct perf_sched *sched)
3170 {
3171 	struct rb_node *next;
3172 
3173 	setup_pager();
3174 
3175 	if (perf_sched__read_events(sched))
3176 		return -1;
3177 
3178 	perf_sched__merge_lat(sched);
3179 	perf_sched__sort_lat(sched);
3180 
3181 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3182 	printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3183 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3184 
3185 	next = rb_first_cached(&sched->sorted_atom_root);
3186 
3187 	while (next) {
3188 		struct work_atoms *work_list;
3189 
3190 		work_list = rb_entry(next, struct work_atoms, node);
3191 		output_lat_thread(sched, work_list);
3192 		next = rb_next(next);
3193 		thread__zput(work_list->thread);
3194 	}
3195 
3196 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3197 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3198 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3199 
3200 	printf(" ---------------------------------------------------\n");
3201 
3202 	print_bad_events(sched);
3203 	printf("\n");
3204 
3205 	return 0;
3206 }
3207 
3208 static int setup_map_cpus(struct perf_sched *sched)
3209 {
3210 	struct perf_cpu_map *map;
3211 
3212 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3213 
3214 	if (sched->map.comp) {
3215 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3216 		if (!sched->map.comp_cpus)
3217 			return -1;
3218 	}
3219 
3220 	if (!sched->map.cpus_str)
3221 		return 0;
3222 
3223 	map = perf_cpu_map__new(sched->map.cpus_str);
3224 	if (!map) {
3225 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3226 		return -1;
3227 	}
3228 
3229 	sched->map.cpus = map;
3230 	return 0;
3231 }
3232 
3233 static int setup_color_pids(struct perf_sched *sched)
3234 {
3235 	struct perf_thread_map *map;
3236 
3237 	if (!sched->map.color_pids_str)
3238 		return 0;
3239 
3240 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3241 	if (!map) {
3242 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3243 		return -1;
3244 	}
3245 
3246 	sched->map.color_pids = map;
3247 	return 0;
3248 }
3249 
3250 static int setup_color_cpus(struct perf_sched *sched)
3251 {
3252 	struct perf_cpu_map *map;
3253 
3254 	if (!sched->map.color_cpus_str)
3255 		return 0;
3256 
3257 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3258 	if (!map) {
3259 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3260 		return -1;
3261 	}
3262 
3263 	sched->map.color_cpus = map;
3264 	return 0;
3265 }
3266 
3267 static int perf_sched__map(struct perf_sched *sched)
3268 {
3269 	if (setup_map_cpus(sched))
3270 		return -1;
3271 
3272 	if (setup_color_pids(sched))
3273 		return -1;
3274 
3275 	if (setup_color_cpus(sched))
3276 		return -1;
3277 
3278 	setup_pager();
3279 	if (perf_sched__read_events(sched))
3280 		return -1;
3281 	print_bad_events(sched);
3282 	return 0;
3283 }
3284 
3285 static int perf_sched__replay(struct perf_sched *sched)
3286 {
3287 	unsigned long i;
3288 
3289 	calibrate_run_measurement_overhead(sched);
3290 	calibrate_sleep_measurement_overhead(sched);
3291 
3292 	test_calibrations(sched);
3293 
3294 	if (perf_sched__read_events(sched))
3295 		return -1;
3296 
3297 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3298 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3299 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3300 
3301 	if (sched->targetless_wakeups)
3302 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3303 	if (sched->multitarget_wakeups)
3304 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3305 	if (sched->nr_run_events_optimized)
3306 		printf("run atoms optimized: %ld\n",
3307 			sched->nr_run_events_optimized);
3308 
3309 	print_task_traces(sched);
3310 	add_cross_task_wakeups(sched);
3311 
3312 	create_tasks(sched);
3313 	printf("------------------------------------------------------------\n");
3314 	for (i = 0; i < sched->replay_repeat; i++)
3315 		run_one_test(sched);
3316 
3317 	return 0;
3318 }
3319 
3320 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3321 			  const char * const usage_msg[])
3322 {
3323 	char *tmp, *tok, *str = strdup(sched->sort_order);
3324 
3325 	for (tok = strtok_r(str, ", ", &tmp);
3326 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3327 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3328 			usage_with_options_msg(usage_msg, options,
3329 					"Unknown --sort key: `%s'", tok);
3330 		}
3331 	}
3332 
3333 	free(str);
3334 
3335 	sort_dimension__add("pid", &sched->cmp_pid);
3336 }
3337 
3338 static int __cmd_record(int argc, const char **argv)
3339 {
3340 	unsigned int rec_argc, i, j;
3341 	const char **rec_argv;
3342 	const char * const record_args[] = {
3343 		"record",
3344 		"-a",
3345 		"-R",
3346 		"-m", "1024",
3347 		"-c", "1",
3348 		"-e", "sched:sched_switch",
3349 		"-e", "sched:sched_stat_wait",
3350 		"-e", "sched:sched_stat_sleep",
3351 		"-e", "sched:sched_stat_iowait",
3352 		"-e", "sched:sched_stat_runtime",
3353 		"-e", "sched:sched_process_fork",
3354 		"-e", "sched:sched_wakeup_new",
3355 		"-e", "sched:sched_migrate_task",
3356 	};
3357 	struct tep_event *waking_event;
3358 
3359 	/*
3360 	 * +2 for either "-e", "sched:sched_wakeup" or
3361 	 * "-e", "sched:sched_waking"
3362 	 */
3363 	rec_argc = ARRAY_SIZE(record_args) + 2 + argc - 1;
3364 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3365 
3366 	if (rec_argv == NULL)
3367 		return -ENOMEM;
3368 
3369 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3370 		rec_argv[i] = strdup(record_args[i]);
3371 
3372 	rec_argv[i++] = "-e";
3373 	waking_event = trace_event__tp_format("sched", "sched_waking");
3374 	if (!IS_ERR(waking_event))
3375 		rec_argv[i++] = strdup("sched:sched_waking");
3376 	else
3377 		rec_argv[i++] = strdup("sched:sched_wakeup");
3378 
3379 	for (j = 1; j < (unsigned int)argc; j++, i++)
3380 		rec_argv[i] = argv[j];
3381 
3382 	BUG_ON(i != rec_argc);
3383 
3384 	return cmd_record(i, rec_argv);
3385 }
3386 
3387 int cmd_sched(int argc, const char **argv)
3388 {
3389 	static const char default_sort_order[] = "avg, max, switch, runtime";
3390 	struct perf_sched sched = {
3391 		.tool = {
3392 			.sample		 = perf_sched__process_tracepoint_sample,
3393 			.comm		 = perf_sched__process_comm,
3394 			.namespaces	 = perf_event__process_namespaces,
3395 			.lost		 = perf_event__process_lost,
3396 			.fork		 = perf_sched__process_fork_event,
3397 			.ordered_events = true,
3398 		},
3399 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3400 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3401 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3402 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3403 		.sort_order	      = default_sort_order,
3404 		.replay_repeat	      = 10,
3405 		.profile_cpu	      = -1,
3406 		.next_shortname1      = 'A',
3407 		.next_shortname2      = '0',
3408 		.skip_merge           = 0,
3409 		.show_callchain	      = 1,
3410 		.max_stack            = 5,
3411 	};
3412 	const struct option sched_options[] = {
3413 	OPT_STRING('i', "input", &input_name, "file",
3414 		    "input file name"),
3415 	OPT_INCR('v', "verbose", &verbose,
3416 		    "be more verbose (show symbol address, etc)"),
3417 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3418 		    "dump raw trace in ASCII"),
3419 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3420 	OPT_END()
3421 	};
3422 	const struct option latency_options[] = {
3423 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3424 		   "sort by key(s): runtime, switch, avg, max"),
3425 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3426 		    "CPU to profile on"),
3427 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3428 		    "latency stats per pid instead of per comm"),
3429 	OPT_PARENT(sched_options)
3430 	};
3431 	const struct option replay_options[] = {
3432 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3433 		     "repeat the workload replay N times (-1: infinite)"),
3434 	OPT_PARENT(sched_options)
3435 	};
3436 	const struct option map_options[] = {
3437 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3438 		    "map output in compact mode"),
3439 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3440 		   "highlight given pids in map"),
3441 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3442                     "highlight given CPUs in map"),
3443 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3444                     "display given CPUs in map"),
3445 	OPT_PARENT(sched_options)
3446 	};
3447 	const struct option timehist_options[] = {
3448 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3449 		   "file", "vmlinux pathname"),
3450 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3451 		   "file", "kallsyms pathname"),
3452 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3453 		    "Display call chains if present (default on)"),
3454 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3455 		   "Maximum number of functions to display backtrace."),
3456 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3457 		    "Look for files with symbols relative to this directory"),
3458 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3459 		    "Show only syscall summary with statistics"),
3460 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3461 		    "Show all syscalls and summary with statistics"),
3462 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3463 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3464 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3465 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3466 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3467 	OPT_STRING(0, "time", &sched.time_str, "str",
3468 		   "Time span for analysis (start,stop)"),
3469 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3470 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3471 		   "analyze events only for given process id(s)"),
3472 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3473 		   "analyze events only for given thread id(s)"),
3474 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3475 	OPT_PARENT(sched_options)
3476 	};
3477 
3478 	const char * const latency_usage[] = {
3479 		"perf sched latency [<options>]",
3480 		NULL
3481 	};
3482 	const char * const replay_usage[] = {
3483 		"perf sched replay [<options>]",
3484 		NULL
3485 	};
3486 	const char * const map_usage[] = {
3487 		"perf sched map [<options>]",
3488 		NULL
3489 	};
3490 	const char * const timehist_usage[] = {
3491 		"perf sched timehist [<options>]",
3492 		NULL
3493 	};
3494 	const char *const sched_subcommands[] = { "record", "latency", "map",
3495 						  "replay", "script",
3496 						  "timehist", NULL };
3497 	const char *sched_usage[] = {
3498 		NULL,
3499 		NULL
3500 	};
3501 	struct trace_sched_handler lat_ops  = {
3502 		.wakeup_event	    = latency_wakeup_event,
3503 		.switch_event	    = latency_switch_event,
3504 		.runtime_event	    = latency_runtime_event,
3505 		.migrate_task_event = latency_migrate_task_event,
3506 	};
3507 	struct trace_sched_handler map_ops  = {
3508 		.switch_event	    = map_switch_event,
3509 	};
3510 	struct trace_sched_handler replay_ops  = {
3511 		.wakeup_event	    = replay_wakeup_event,
3512 		.switch_event	    = replay_switch_event,
3513 		.fork_event	    = replay_fork_event,
3514 	};
3515 	unsigned int i;
3516 
3517 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3518 		sched.curr_pid[i] = -1;
3519 
3520 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3521 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3522 	if (!argc)
3523 		usage_with_options(sched_usage, sched_options);
3524 
3525 	/*
3526 	 * Aliased to 'perf script' for now:
3527 	 */
3528 	if (!strcmp(argv[0], "script"))
3529 		return cmd_script(argc, argv);
3530 
3531 	if (!strncmp(argv[0], "rec", 3)) {
3532 		return __cmd_record(argc, argv);
3533 	} else if (!strncmp(argv[0], "lat", 3)) {
3534 		sched.tp_handler = &lat_ops;
3535 		if (argc > 1) {
3536 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3537 			if (argc)
3538 				usage_with_options(latency_usage, latency_options);
3539 		}
3540 		setup_sorting(&sched, latency_options, latency_usage);
3541 		return perf_sched__lat(&sched);
3542 	} else if (!strcmp(argv[0], "map")) {
3543 		if (argc) {
3544 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3545 			if (argc)
3546 				usage_with_options(map_usage, map_options);
3547 		}
3548 		sched.tp_handler = &map_ops;
3549 		setup_sorting(&sched, latency_options, latency_usage);
3550 		return perf_sched__map(&sched);
3551 	} else if (!strncmp(argv[0], "rep", 3)) {
3552 		sched.tp_handler = &replay_ops;
3553 		if (argc) {
3554 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3555 			if (argc)
3556 				usage_with_options(replay_usage, replay_options);
3557 		}
3558 		return perf_sched__replay(&sched);
3559 	} else if (!strcmp(argv[0], "timehist")) {
3560 		if (argc) {
3561 			argc = parse_options(argc, argv, timehist_options,
3562 					     timehist_usage, 0);
3563 			if (argc)
3564 				usage_with_options(timehist_usage, timehist_options);
3565 		}
3566 		if ((sched.show_wakeups || sched.show_next) &&
3567 		    sched.summary_only) {
3568 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3569 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3570 			if (sched.show_wakeups)
3571 				parse_options_usage(NULL, timehist_options, "w", true);
3572 			if (sched.show_next)
3573 				parse_options_usage(NULL, timehist_options, "n", true);
3574 			return -EINVAL;
3575 		}
3576 
3577 		return perf_sched__timehist(&sched);
3578 	} else {
3579 		usage_with_options(sched_usage, sched_options);
3580 	}
3581 
3582 	return 0;
3583 }
3584