xref: /openbmc/linux/tools/perf/builtin-sched.c (revision be2b81b5)
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/mutex.h"
11 #include "util/symbol.h"
12 #include "util/thread.h"
13 #include "util/header.h"
14 #include "util/session.h"
15 #include "util/tool.h"
16 #include "util/cloexec.h"
17 #include "util/thread_map.h"
18 #include "util/color.h"
19 #include "util/stat.h"
20 #include "util/string2.h"
21 #include "util/callchain.h"
22 #include "util/time-utils.h"
23 
24 #include <subcmd/pager.h>
25 #include <subcmd/parse-options.h>
26 #include "util/trace-event.h"
27 
28 #include "util/debug.h"
29 #include "util/event.h"
30 
31 #include <linux/kernel.h>
32 #include <linux/log2.h>
33 #include <linux/zalloc.h>
34 #include <sys/prctl.h>
35 #include <sys/resource.h>
36 #include <inttypes.h>
37 
38 #include <errno.h>
39 #include <semaphore.h>
40 #include <pthread.h>
41 #include <math.h>
42 #include <api/fs/fs.h>
43 #include <perf/cpumap.h>
44 #include <linux/time64.h>
45 #include <linux/err.h>
46 
47 #include <linux/ctype.h>
48 
49 #define PR_SET_NAME		15               /* Set process name */
50 #define MAX_CPUS		4096
51 #define COMM_LEN		20
52 #define SYM_LEN			129
53 #define MAX_PID			1024000
54 
55 static const char *cpu_list;
56 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57 
58 struct sched_atom;
59 
60 struct task_desc {
61 	unsigned long		nr;
62 	unsigned long		pid;
63 	char			comm[COMM_LEN];
64 
65 	unsigned long		nr_events;
66 	unsigned long		curr_event;
67 	struct sched_atom	**atoms;
68 
69 	pthread_t		thread;
70 	sem_t			sleep_sem;
71 
72 	sem_t			ready_for_work;
73 	sem_t			work_done_sem;
74 
75 	u64			cpu_usage;
76 };
77 
78 enum sched_event_type {
79 	SCHED_EVENT_RUN,
80 	SCHED_EVENT_SLEEP,
81 	SCHED_EVENT_WAKEUP,
82 	SCHED_EVENT_MIGRATION,
83 };
84 
85 struct sched_atom {
86 	enum sched_event_type	type;
87 	int			specific_wait;
88 	u64			timestamp;
89 	u64			duration;
90 	unsigned long		nr;
91 	sem_t			*wait_sem;
92 	struct task_desc	*wakee;
93 };
94 
95 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
96 
97 /* task state bitmask, copied from include/linux/sched.h */
98 #define TASK_RUNNING		0
99 #define TASK_INTERRUPTIBLE	1
100 #define TASK_UNINTERRUPTIBLE	2
101 #define __TASK_STOPPED		4
102 #define __TASK_TRACED		8
103 /* in tsk->exit_state */
104 #define EXIT_DEAD		16
105 #define EXIT_ZOMBIE		32
106 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
107 /* in tsk->state again */
108 #define TASK_DEAD		64
109 #define TASK_WAKEKILL		128
110 #define TASK_WAKING		256
111 #define TASK_PARKED		512
112 
113 enum thread_state {
114 	THREAD_SLEEPING = 0,
115 	THREAD_WAIT_CPU,
116 	THREAD_SCHED_IN,
117 	THREAD_IGNORE
118 };
119 
120 struct work_atom {
121 	struct list_head	list;
122 	enum thread_state	state;
123 	u64			sched_out_time;
124 	u64			wake_up_time;
125 	u64			sched_in_time;
126 	u64			runtime;
127 };
128 
129 struct work_atoms {
130 	struct list_head	work_list;
131 	struct thread		*thread;
132 	struct rb_node		node;
133 	u64			max_lat;
134 	u64			max_lat_start;
135 	u64			max_lat_end;
136 	u64			total_lat;
137 	u64			nb_atoms;
138 	u64			total_runtime;
139 	int			num_merged;
140 };
141 
142 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
143 
144 struct perf_sched;
145 
146 struct trace_sched_handler {
147 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
148 			    struct perf_sample *sample, struct machine *machine);
149 
150 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
151 			     struct perf_sample *sample, struct machine *machine);
152 
153 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
154 			    struct perf_sample *sample, struct machine *machine);
155 
156 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
157 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
158 			  struct machine *machine);
159 
160 	int (*migrate_task_event)(struct perf_sched *sched,
161 				  struct evsel *evsel,
162 				  struct perf_sample *sample,
163 				  struct machine *machine);
164 };
165 
166 #define COLOR_PIDS PERF_COLOR_BLUE
167 #define COLOR_CPUS PERF_COLOR_BG_RED
168 
169 struct perf_sched_map {
170 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
171 	struct perf_cpu		*comp_cpus;
172 	bool			 comp;
173 	struct perf_thread_map *color_pids;
174 	const char		*color_pids_str;
175 	struct perf_cpu_map	*color_cpus;
176 	const char		*color_cpus_str;
177 	struct perf_cpu_map	*cpus;
178 	const char		*cpus_str;
179 };
180 
181 struct perf_sched {
182 	struct perf_tool tool;
183 	const char	 *sort_order;
184 	unsigned long	 nr_tasks;
185 	struct task_desc **pid_to_task;
186 	struct task_desc **tasks;
187 	const struct trace_sched_handler *tp_handler;
188 	struct mutex	 start_work_mutex;
189 	struct mutex	 work_done_wait_mutex;
190 	int		 profile_cpu;
191 /*
192  * Track the current task - that way we can know whether there's any
193  * weird events, such as a task being switched away that is not current.
194  */
195 	struct perf_cpu	 max_cpu;
196 	u32		 curr_pid[MAX_CPUS];
197 	struct thread	 *curr_thread[MAX_CPUS];
198 	char		 next_shortname1;
199 	char		 next_shortname2;
200 	unsigned int	 replay_repeat;
201 	unsigned long	 nr_run_events;
202 	unsigned long	 nr_sleep_events;
203 	unsigned long	 nr_wakeup_events;
204 	unsigned long	 nr_sleep_corrections;
205 	unsigned long	 nr_run_events_optimized;
206 	unsigned long	 targetless_wakeups;
207 	unsigned long	 multitarget_wakeups;
208 	unsigned long	 nr_runs;
209 	unsigned long	 nr_timestamps;
210 	unsigned long	 nr_unordered_timestamps;
211 	unsigned long	 nr_context_switch_bugs;
212 	unsigned long	 nr_events;
213 	unsigned long	 nr_lost_chunks;
214 	unsigned long	 nr_lost_events;
215 	u64		 run_measurement_overhead;
216 	u64		 sleep_measurement_overhead;
217 	u64		 start_time;
218 	u64		 cpu_usage;
219 	u64		 runavg_cpu_usage;
220 	u64		 parent_cpu_usage;
221 	u64		 runavg_parent_cpu_usage;
222 	u64		 sum_runtime;
223 	u64		 sum_fluct;
224 	u64		 run_avg;
225 	u64		 all_runtime;
226 	u64		 all_count;
227 	u64		 cpu_last_switched[MAX_CPUS];
228 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
229 	struct list_head sort_list, cmp_pid;
230 	bool force;
231 	bool skip_merge;
232 	struct perf_sched_map map;
233 
234 	/* options for timehist command */
235 	bool		summary;
236 	bool		summary_only;
237 	bool		idle_hist;
238 	bool		show_callchain;
239 	unsigned int	max_stack;
240 	bool		show_cpu_visual;
241 	bool		show_wakeups;
242 	bool		show_next;
243 	bool		show_migrations;
244 	bool		show_state;
245 	u64		skipped_samples;
246 	const char	*time_str;
247 	struct perf_time_interval ptime;
248 	struct perf_time_interval hist_time;
249 	volatile bool   thread_funcs_exit;
250 };
251 
252 /* per thread run time data */
253 struct thread_runtime {
254 	u64 last_time;      /* time of previous sched in/out event */
255 	u64 dt_run;         /* run time */
256 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
257 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
258 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
259 	u64 dt_delay;       /* time between wakeup and sched-in */
260 	u64 ready_to_run;   /* time of wakeup */
261 
262 	struct stats run_stats;
263 	u64 total_run_time;
264 	u64 total_sleep_time;
265 	u64 total_iowait_time;
266 	u64 total_preempt_time;
267 	u64 total_delay_time;
268 
269 	int last_state;
270 
271 	char shortname[3];
272 	bool comm_changed;
273 
274 	u64 migrations;
275 };
276 
277 /* per event run time data */
278 struct evsel_runtime {
279 	u64 *last_time; /* time this event was last seen per cpu */
280 	u32 ncpu;       /* highest cpu slot allocated */
281 };
282 
283 /* per cpu idle time data */
284 struct idle_thread_runtime {
285 	struct thread_runtime	tr;
286 	struct thread		*last_thread;
287 	struct rb_root_cached	sorted_root;
288 	struct callchain_root	callchain;
289 	struct callchain_cursor	cursor;
290 };
291 
292 /* track idle times per cpu */
293 static struct thread **idle_threads;
294 static int idle_max_cpu;
295 static char idle_comm[] = "<idle>";
296 
297 static u64 get_nsecs(void)
298 {
299 	struct timespec ts;
300 
301 	clock_gettime(CLOCK_MONOTONIC, &ts);
302 
303 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
304 }
305 
306 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
307 {
308 	u64 T0 = get_nsecs(), T1;
309 
310 	do {
311 		T1 = get_nsecs();
312 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
313 }
314 
315 static void sleep_nsecs(u64 nsecs)
316 {
317 	struct timespec ts;
318 
319 	ts.tv_nsec = nsecs % 999999999;
320 	ts.tv_sec = nsecs / 999999999;
321 
322 	nanosleep(&ts, NULL);
323 }
324 
325 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
326 {
327 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 	int i;
329 
330 	for (i = 0; i < 10; i++) {
331 		T0 = get_nsecs();
332 		burn_nsecs(sched, 0);
333 		T1 = get_nsecs();
334 		delta = T1-T0;
335 		min_delta = min(min_delta, delta);
336 	}
337 	sched->run_measurement_overhead = min_delta;
338 
339 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340 }
341 
342 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
343 {
344 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
345 	int i;
346 
347 	for (i = 0; i < 10; i++) {
348 		T0 = get_nsecs();
349 		sleep_nsecs(10000);
350 		T1 = get_nsecs();
351 		delta = T1-T0;
352 		min_delta = min(min_delta, delta);
353 	}
354 	min_delta -= 10000;
355 	sched->sleep_measurement_overhead = min_delta;
356 
357 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
358 }
359 
360 static struct sched_atom *
361 get_new_event(struct task_desc *task, u64 timestamp)
362 {
363 	struct sched_atom *event = zalloc(sizeof(*event));
364 	unsigned long idx = task->nr_events;
365 	size_t size;
366 
367 	event->timestamp = timestamp;
368 	event->nr = idx;
369 
370 	task->nr_events++;
371 	size = sizeof(struct sched_atom *) * task->nr_events;
372 	task->atoms = realloc(task->atoms, size);
373 	BUG_ON(!task->atoms);
374 
375 	task->atoms[idx] = event;
376 
377 	return event;
378 }
379 
380 static struct sched_atom *last_event(struct task_desc *task)
381 {
382 	if (!task->nr_events)
383 		return NULL;
384 
385 	return task->atoms[task->nr_events - 1];
386 }
387 
388 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
389 				u64 timestamp, u64 duration)
390 {
391 	struct sched_atom *event, *curr_event = last_event(task);
392 
393 	/*
394 	 * optimize an existing RUN event by merging this one
395 	 * to it:
396 	 */
397 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
398 		sched->nr_run_events_optimized++;
399 		curr_event->duration += duration;
400 		return;
401 	}
402 
403 	event = get_new_event(task, timestamp);
404 
405 	event->type = SCHED_EVENT_RUN;
406 	event->duration = duration;
407 
408 	sched->nr_run_events++;
409 }
410 
411 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
412 				   u64 timestamp, struct task_desc *wakee)
413 {
414 	struct sched_atom *event, *wakee_event;
415 
416 	event = get_new_event(task, timestamp);
417 	event->type = SCHED_EVENT_WAKEUP;
418 	event->wakee = wakee;
419 
420 	wakee_event = last_event(wakee);
421 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
422 		sched->targetless_wakeups++;
423 		return;
424 	}
425 	if (wakee_event->wait_sem) {
426 		sched->multitarget_wakeups++;
427 		return;
428 	}
429 
430 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
431 	sem_init(wakee_event->wait_sem, 0, 0);
432 	wakee_event->specific_wait = 1;
433 	event->wait_sem = wakee_event->wait_sem;
434 
435 	sched->nr_wakeup_events++;
436 }
437 
438 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
439 				  u64 timestamp, u64 task_state __maybe_unused)
440 {
441 	struct sched_atom *event = get_new_event(task, timestamp);
442 
443 	event->type = SCHED_EVENT_SLEEP;
444 
445 	sched->nr_sleep_events++;
446 }
447 
448 static struct task_desc *register_pid(struct perf_sched *sched,
449 				      unsigned long pid, const char *comm)
450 {
451 	struct task_desc *task;
452 	static int pid_max;
453 
454 	if (sched->pid_to_task == NULL) {
455 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
456 			pid_max = MAX_PID;
457 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
458 	}
459 	if (pid >= (unsigned long)pid_max) {
460 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
461 			sizeof(struct task_desc *))) == NULL);
462 		while (pid >= (unsigned long)pid_max)
463 			sched->pid_to_task[pid_max++] = NULL;
464 	}
465 
466 	task = sched->pid_to_task[pid];
467 
468 	if (task)
469 		return task;
470 
471 	task = zalloc(sizeof(*task));
472 	task->pid = pid;
473 	task->nr = sched->nr_tasks;
474 	strcpy(task->comm, comm);
475 	/*
476 	 * every task starts in sleeping state - this gets ignored
477 	 * if there's no wakeup pointing to this sleep state:
478 	 */
479 	add_sched_event_sleep(sched, task, 0, 0);
480 
481 	sched->pid_to_task[pid] = task;
482 	sched->nr_tasks++;
483 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
484 	BUG_ON(!sched->tasks);
485 	sched->tasks[task->nr] = task;
486 
487 	if (verbose > 0)
488 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
489 
490 	return task;
491 }
492 
493 
494 static void print_task_traces(struct perf_sched *sched)
495 {
496 	struct task_desc *task;
497 	unsigned long i;
498 
499 	for (i = 0; i < sched->nr_tasks; i++) {
500 		task = sched->tasks[i];
501 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
502 			task->nr, task->comm, task->pid, task->nr_events);
503 	}
504 }
505 
506 static void add_cross_task_wakeups(struct perf_sched *sched)
507 {
508 	struct task_desc *task1, *task2;
509 	unsigned long i, j;
510 
511 	for (i = 0; i < sched->nr_tasks; i++) {
512 		task1 = sched->tasks[i];
513 		j = i + 1;
514 		if (j == sched->nr_tasks)
515 			j = 0;
516 		task2 = sched->tasks[j];
517 		add_sched_event_wakeup(sched, task1, 0, task2);
518 	}
519 }
520 
521 static void perf_sched__process_event(struct perf_sched *sched,
522 				      struct sched_atom *atom)
523 {
524 	int ret = 0;
525 
526 	switch (atom->type) {
527 		case SCHED_EVENT_RUN:
528 			burn_nsecs(sched, atom->duration);
529 			break;
530 		case SCHED_EVENT_SLEEP:
531 			if (atom->wait_sem)
532 				ret = sem_wait(atom->wait_sem);
533 			BUG_ON(ret);
534 			break;
535 		case SCHED_EVENT_WAKEUP:
536 			if (atom->wait_sem)
537 				ret = sem_post(atom->wait_sem);
538 			BUG_ON(ret);
539 			break;
540 		case SCHED_EVENT_MIGRATION:
541 			break;
542 		default:
543 			BUG_ON(1);
544 	}
545 }
546 
547 static u64 get_cpu_usage_nsec_parent(void)
548 {
549 	struct rusage ru;
550 	u64 sum;
551 	int err;
552 
553 	err = getrusage(RUSAGE_SELF, &ru);
554 	BUG_ON(err);
555 
556 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
557 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
558 
559 	return sum;
560 }
561 
562 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
563 {
564 	struct perf_event_attr attr;
565 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
566 	int fd;
567 	struct rlimit limit;
568 	bool need_privilege = false;
569 
570 	memset(&attr, 0, sizeof(attr));
571 
572 	attr.type = PERF_TYPE_SOFTWARE;
573 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
574 
575 force_again:
576 	fd = sys_perf_event_open(&attr, 0, -1, -1,
577 				 perf_event_open_cloexec_flag());
578 
579 	if (fd < 0) {
580 		if (errno == EMFILE) {
581 			if (sched->force) {
582 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
583 				limit.rlim_cur += sched->nr_tasks - cur_task;
584 				if (limit.rlim_cur > limit.rlim_max) {
585 					limit.rlim_max = limit.rlim_cur;
586 					need_privilege = true;
587 				}
588 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
589 					if (need_privilege && errno == EPERM)
590 						strcpy(info, "Need privilege\n");
591 				} else
592 					goto force_again;
593 			} else
594 				strcpy(info, "Have a try with -f option\n");
595 		}
596 		pr_err("Error: sys_perf_event_open() syscall returned "
597 		       "with %d (%s)\n%s", fd,
598 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
599 		exit(EXIT_FAILURE);
600 	}
601 	return fd;
602 }
603 
604 static u64 get_cpu_usage_nsec_self(int fd)
605 {
606 	u64 runtime;
607 	int ret;
608 
609 	ret = read(fd, &runtime, sizeof(runtime));
610 	BUG_ON(ret != sizeof(runtime));
611 
612 	return runtime;
613 }
614 
615 struct sched_thread_parms {
616 	struct task_desc  *task;
617 	struct perf_sched *sched;
618 	int fd;
619 };
620 
621 static void *thread_func(void *ctx)
622 {
623 	struct sched_thread_parms *parms = ctx;
624 	struct task_desc *this_task = parms->task;
625 	struct perf_sched *sched = parms->sched;
626 	u64 cpu_usage_0, cpu_usage_1;
627 	unsigned long i, ret;
628 	char comm2[22];
629 	int fd = parms->fd;
630 
631 	zfree(&parms);
632 
633 	sprintf(comm2, ":%s", this_task->comm);
634 	prctl(PR_SET_NAME, comm2);
635 	if (fd < 0)
636 		return NULL;
637 
638 	while (!sched->thread_funcs_exit) {
639 		ret = sem_post(&this_task->ready_for_work);
640 		BUG_ON(ret);
641 		mutex_lock(&sched->start_work_mutex);
642 		mutex_unlock(&sched->start_work_mutex);
643 
644 		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
645 
646 		for (i = 0; i < this_task->nr_events; i++) {
647 			this_task->curr_event = i;
648 			perf_sched__process_event(sched, this_task->atoms[i]);
649 		}
650 
651 		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
652 		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
653 		ret = sem_post(&this_task->work_done_sem);
654 		BUG_ON(ret);
655 
656 		mutex_lock(&sched->work_done_wait_mutex);
657 		mutex_unlock(&sched->work_done_wait_mutex);
658 	}
659 	return NULL;
660 }
661 
662 static void create_tasks(struct perf_sched *sched)
663 	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
664 	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
665 {
666 	struct task_desc *task;
667 	pthread_attr_t attr;
668 	unsigned long i;
669 	int err;
670 
671 	err = pthread_attr_init(&attr);
672 	BUG_ON(err);
673 	err = pthread_attr_setstacksize(&attr,
674 			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
675 	BUG_ON(err);
676 	mutex_lock(&sched->start_work_mutex);
677 	mutex_lock(&sched->work_done_wait_mutex);
678 	for (i = 0; i < sched->nr_tasks; i++) {
679 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
680 		BUG_ON(parms == NULL);
681 		parms->task = task = sched->tasks[i];
682 		parms->sched = sched;
683 		parms->fd = self_open_counters(sched, i);
684 		sem_init(&task->sleep_sem, 0, 0);
685 		sem_init(&task->ready_for_work, 0, 0);
686 		sem_init(&task->work_done_sem, 0, 0);
687 		task->curr_event = 0;
688 		err = pthread_create(&task->thread, &attr, thread_func, parms);
689 		BUG_ON(err);
690 	}
691 }
692 
693 static void destroy_tasks(struct perf_sched *sched)
694 	UNLOCK_FUNCTION(sched->start_work_mutex)
695 	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
696 {
697 	struct task_desc *task;
698 	unsigned long i;
699 	int err;
700 
701 	mutex_unlock(&sched->start_work_mutex);
702 	mutex_unlock(&sched->work_done_wait_mutex);
703 	/* Get rid of threads so they won't be upset by mutex destrunction */
704 	for (i = 0; i < sched->nr_tasks; i++) {
705 		task = sched->tasks[i];
706 		err = pthread_join(task->thread, NULL);
707 		BUG_ON(err);
708 		sem_destroy(&task->sleep_sem);
709 		sem_destroy(&task->ready_for_work);
710 		sem_destroy(&task->work_done_sem);
711 	}
712 }
713 
714 static void wait_for_tasks(struct perf_sched *sched)
715 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
716 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
717 {
718 	u64 cpu_usage_0, cpu_usage_1;
719 	struct task_desc *task;
720 	unsigned long i, ret;
721 
722 	sched->start_time = get_nsecs();
723 	sched->cpu_usage = 0;
724 	mutex_unlock(&sched->work_done_wait_mutex);
725 
726 	for (i = 0; i < sched->nr_tasks; i++) {
727 		task = sched->tasks[i];
728 		ret = sem_wait(&task->ready_for_work);
729 		BUG_ON(ret);
730 		sem_init(&task->ready_for_work, 0, 0);
731 	}
732 	mutex_lock(&sched->work_done_wait_mutex);
733 
734 	cpu_usage_0 = get_cpu_usage_nsec_parent();
735 
736 	mutex_unlock(&sched->start_work_mutex);
737 
738 	for (i = 0; i < sched->nr_tasks; i++) {
739 		task = sched->tasks[i];
740 		ret = sem_wait(&task->work_done_sem);
741 		BUG_ON(ret);
742 		sem_init(&task->work_done_sem, 0, 0);
743 		sched->cpu_usage += task->cpu_usage;
744 		task->cpu_usage = 0;
745 	}
746 
747 	cpu_usage_1 = get_cpu_usage_nsec_parent();
748 	if (!sched->runavg_cpu_usage)
749 		sched->runavg_cpu_usage = sched->cpu_usage;
750 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
751 
752 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
753 	if (!sched->runavg_parent_cpu_usage)
754 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
755 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
756 					 sched->parent_cpu_usage)/sched->replay_repeat;
757 
758 	mutex_lock(&sched->start_work_mutex);
759 
760 	for (i = 0; i < sched->nr_tasks; i++) {
761 		task = sched->tasks[i];
762 		sem_init(&task->sleep_sem, 0, 0);
763 		task->curr_event = 0;
764 	}
765 }
766 
767 static void run_one_test(struct perf_sched *sched)
768 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
769 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
770 {
771 	u64 T0, T1, delta, avg_delta, fluct;
772 
773 	T0 = get_nsecs();
774 	wait_for_tasks(sched);
775 	T1 = get_nsecs();
776 
777 	delta = T1 - T0;
778 	sched->sum_runtime += delta;
779 	sched->nr_runs++;
780 
781 	avg_delta = sched->sum_runtime / sched->nr_runs;
782 	if (delta < avg_delta)
783 		fluct = avg_delta - delta;
784 	else
785 		fluct = delta - avg_delta;
786 	sched->sum_fluct += fluct;
787 	if (!sched->run_avg)
788 		sched->run_avg = delta;
789 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
790 
791 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
792 
793 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
794 
795 	printf("cpu: %0.2f / %0.2f",
796 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
797 
798 #if 0
799 	/*
800 	 * rusage statistics done by the parent, these are less
801 	 * accurate than the sched->sum_exec_runtime based statistics:
802 	 */
803 	printf(" [%0.2f / %0.2f]",
804 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
805 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
806 #endif
807 
808 	printf("\n");
809 
810 	if (sched->nr_sleep_corrections)
811 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
812 	sched->nr_sleep_corrections = 0;
813 }
814 
815 static void test_calibrations(struct perf_sched *sched)
816 {
817 	u64 T0, T1;
818 
819 	T0 = get_nsecs();
820 	burn_nsecs(sched, NSEC_PER_MSEC);
821 	T1 = get_nsecs();
822 
823 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
824 
825 	T0 = get_nsecs();
826 	sleep_nsecs(NSEC_PER_MSEC);
827 	T1 = get_nsecs();
828 
829 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
830 }
831 
832 static int
833 replay_wakeup_event(struct perf_sched *sched,
834 		    struct evsel *evsel, struct perf_sample *sample,
835 		    struct machine *machine __maybe_unused)
836 {
837 	const char *comm = evsel__strval(evsel, sample, "comm");
838 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
839 	struct task_desc *waker, *wakee;
840 
841 	if (verbose > 0) {
842 		printf("sched_wakeup event %p\n", evsel);
843 
844 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
845 	}
846 
847 	waker = register_pid(sched, sample->tid, "<unknown>");
848 	wakee = register_pid(sched, pid, comm);
849 
850 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
851 	return 0;
852 }
853 
854 static int replay_switch_event(struct perf_sched *sched,
855 			       struct evsel *evsel,
856 			       struct perf_sample *sample,
857 			       struct machine *machine __maybe_unused)
858 {
859 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
860 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
861 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
862 		  next_pid = evsel__intval(evsel, sample, "next_pid");
863 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
864 	struct task_desc *prev, __maybe_unused *next;
865 	u64 timestamp0, timestamp = sample->time;
866 	int cpu = sample->cpu;
867 	s64 delta;
868 
869 	if (verbose > 0)
870 		printf("sched_switch event %p\n", evsel);
871 
872 	if (cpu >= MAX_CPUS || cpu < 0)
873 		return 0;
874 
875 	timestamp0 = sched->cpu_last_switched[cpu];
876 	if (timestamp0)
877 		delta = timestamp - timestamp0;
878 	else
879 		delta = 0;
880 
881 	if (delta < 0) {
882 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
883 		return -1;
884 	}
885 
886 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
887 		 prev_comm, prev_pid, next_comm, next_pid, delta);
888 
889 	prev = register_pid(sched, prev_pid, prev_comm);
890 	next = register_pid(sched, next_pid, next_comm);
891 
892 	sched->cpu_last_switched[cpu] = timestamp;
893 
894 	add_sched_event_run(sched, prev, timestamp, delta);
895 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
896 
897 	return 0;
898 }
899 
900 static int replay_fork_event(struct perf_sched *sched,
901 			     union perf_event *event,
902 			     struct machine *machine)
903 {
904 	struct thread *child, *parent;
905 
906 	child = machine__findnew_thread(machine, event->fork.pid,
907 					event->fork.tid);
908 	parent = machine__findnew_thread(machine, event->fork.ppid,
909 					 event->fork.ptid);
910 
911 	if (child == NULL || parent == NULL) {
912 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
913 				 child, parent);
914 		goto out_put;
915 	}
916 
917 	if (verbose > 0) {
918 		printf("fork event\n");
919 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
920 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
921 	}
922 
923 	register_pid(sched, parent->tid, thread__comm_str(parent));
924 	register_pid(sched, child->tid, thread__comm_str(child));
925 out_put:
926 	thread__put(child);
927 	thread__put(parent);
928 	return 0;
929 }
930 
931 struct sort_dimension {
932 	const char		*name;
933 	sort_fn_t		cmp;
934 	struct list_head	list;
935 };
936 
937 /*
938  * handle runtime stats saved per thread
939  */
940 static struct thread_runtime *thread__init_runtime(struct thread *thread)
941 {
942 	struct thread_runtime *r;
943 
944 	r = zalloc(sizeof(struct thread_runtime));
945 	if (!r)
946 		return NULL;
947 
948 	init_stats(&r->run_stats);
949 	thread__set_priv(thread, r);
950 
951 	return r;
952 }
953 
954 static struct thread_runtime *thread__get_runtime(struct thread *thread)
955 {
956 	struct thread_runtime *tr;
957 
958 	tr = thread__priv(thread);
959 	if (tr == NULL) {
960 		tr = thread__init_runtime(thread);
961 		if (tr == NULL)
962 			pr_debug("Failed to malloc memory for runtime data.\n");
963 	}
964 
965 	return tr;
966 }
967 
968 static int
969 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
970 {
971 	struct sort_dimension *sort;
972 	int ret = 0;
973 
974 	BUG_ON(list_empty(list));
975 
976 	list_for_each_entry(sort, list, list) {
977 		ret = sort->cmp(l, r);
978 		if (ret)
979 			return ret;
980 	}
981 
982 	return ret;
983 }
984 
985 static struct work_atoms *
986 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
987 			 struct list_head *sort_list)
988 {
989 	struct rb_node *node = root->rb_root.rb_node;
990 	struct work_atoms key = { .thread = thread };
991 
992 	while (node) {
993 		struct work_atoms *atoms;
994 		int cmp;
995 
996 		atoms = container_of(node, struct work_atoms, node);
997 
998 		cmp = thread_lat_cmp(sort_list, &key, atoms);
999 		if (cmp > 0)
1000 			node = node->rb_left;
1001 		else if (cmp < 0)
1002 			node = node->rb_right;
1003 		else {
1004 			BUG_ON(thread != atoms->thread);
1005 			return atoms;
1006 		}
1007 	}
1008 	return NULL;
1009 }
1010 
1011 static void
1012 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1013 			 struct list_head *sort_list)
1014 {
1015 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1016 	bool leftmost = true;
1017 
1018 	while (*new) {
1019 		struct work_atoms *this;
1020 		int cmp;
1021 
1022 		this = container_of(*new, struct work_atoms, node);
1023 		parent = *new;
1024 
1025 		cmp = thread_lat_cmp(sort_list, data, this);
1026 
1027 		if (cmp > 0)
1028 			new = &((*new)->rb_left);
1029 		else {
1030 			new = &((*new)->rb_right);
1031 			leftmost = false;
1032 		}
1033 	}
1034 
1035 	rb_link_node(&data->node, parent, new);
1036 	rb_insert_color_cached(&data->node, root, leftmost);
1037 }
1038 
1039 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1040 {
1041 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1042 	if (!atoms) {
1043 		pr_err("No memory at %s\n", __func__);
1044 		return -1;
1045 	}
1046 
1047 	atoms->thread = thread__get(thread);
1048 	INIT_LIST_HEAD(&atoms->work_list);
1049 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1050 	return 0;
1051 }
1052 
1053 static char sched_out_state(u64 prev_state)
1054 {
1055 	const char *str = TASK_STATE_TO_CHAR_STR;
1056 
1057 	return str[prev_state];
1058 }
1059 
1060 static int
1061 add_sched_out_event(struct work_atoms *atoms,
1062 		    char run_state,
1063 		    u64 timestamp)
1064 {
1065 	struct work_atom *atom = zalloc(sizeof(*atom));
1066 	if (!atom) {
1067 		pr_err("Non memory at %s", __func__);
1068 		return -1;
1069 	}
1070 
1071 	atom->sched_out_time = timestamp;
1072 
1073 	if (run_state == 'R') {
1074 		atom->state = THREAD_WAIT_CPU;
1075 		atom->wake_up_time = atom->sched_out_time;
1076 	}
1077 
1078 	list_add_tail(&atom->list, &atoms->work_list);
1079 	return 0;
1080 }
1081 
1082 static void
1083 add_runtime_event(struct work_atoms *atoms, u64 delta,
1084 		  u64 timestamp __maybe_unused)
1085 {
1086 	struct work_atom *atom;
1087 
1088 	BUG_ON(list_empty(&atoms->work_list));
1089 
1090 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091 
1092 	atom->runtime += delta;
1093 	atoms->total_runtime += delta;
1094 }
1095 
1096 static void
1097 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1098 {
1099 	struct work_atom *atom;
1100 	u64 delta;
1101 
1102 	if (list_empty(&atoms->work_list))
1103 		return;
1104 
1105 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1106 
1107 	if (atom->state != THREAD_WAIT_CPU)
1108 		return;
1109 
1110 	if (timestamp < atom->wake_up_time) {
1111 		atom->state = THREAD_IGNORE;
1112 		return;
1113 	}
1114 
1115 	atom->state = THREAD_SCHED_IN;
1116 	atom->sched_in_time = timestamp;
1117 
1118 	delta = atom->sched_in_time - atom->wake_up_time;
1119 	atoms->total_lat += delta;
1120 	if (delta > atoms->max_lat) {
1121 		atoms->max_lat = delta;
1122 		atoms->max_lat_start = atom->wake_up_time;
1123 		atoms->max_lat_end = timestamp;
1124 	}
1125 	atoms->nb_atoms++;
1126 }
1127 
1128 static int latency_switch_event(struct perf_sched *sched,
1129 				struct evsel *evsel,
1130 				struct perf_sample *sample,
1131 				struct machine *machine)
1132 {
1133 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1134 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1135 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1136 	struct work_atoms *out_events, *in_events;
1137 	struct thread *sched_out, *sched_in;
1138 	u64 timestamp0, timestamp = sample->time;
1139 	int cpu = sample->cpu, err = -1;
1140 	s64 delta;
1141 
1142 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1143 
1144 	timestamp0 = sched->cpu_last_switched[cpu];
1145 	sched->cpu_last_switched[cpu] = timestamp;
1146 	if (timestamp0)
1147 		delta = timestamp - timestamp0;
1148 	else
1149 		delta = 0;
1150 
1151 	if (delta < 0) {
1152 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1153 		return -1;
1154 	}
1155 
1156 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1157 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1158 	if (sched_out == NULL || sched_in == NULL)
1159 		goto out_put;
1160 
1161 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1162 	if (!out_events) {
1163 		if (thread_atoms_insert(sched, sched_out))
1164 			goto out_put;
1165 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1166 		if (!out_events) {
1167 			pr_err("out-event: Internal tree error");
1168 			goto out_put;
1169 		}
1170 	}
1171 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1172 		return -1;
1173 
1174 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1175 	if (!in_events) {
1176 		if (thread_atoms_insert(sched, sched_in))
1177 			goto out_put;
1178 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1179 		if (!in_events) {
1180 			pr_err("in-event: Internal tree error");
1181 			goto out_put;
1182 		}
1183 		/*
1184 		 * Take came in we have not heard about yet,
1185 		 * add in an initial atom in runnable state:
1186 		 */
1187 		if (add_sched_out_event(in_events, 'R', timestamp))
1188 			goto out_put;
1189 	}
1190 	add_sched_in_event(in_events, timestamp);
1191 	err = 0;
1192 out_put:
1193 	thread__put(sched_out);
1194 	thread__put(sched_in);
1195 	return err;
1196 }
1197 
1198 static int latency_runtime_event(struct perf_sched *sched,
1199 				 struct evsel *evsel,
1200 				 struct perf_sample *sample,
1201 				 struct machine *machine)
1202 {
1203 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1204 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1205 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1206 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1207 	u64 timestamp = sample->time;
1208 	int cpu = sample->cpu, err = -1;
1209 
1210 	if (thread == NULL)
1211 		return -1;
1212 
1213 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1214 	if (!atoms) {
1215 		if (thread_atoms_insert(sched, thread))
1216 			goto out_put;
1217 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1218 		if (!atoms) {
1219 			pr_err("in-event: Internal tree error");
1220 			goto out_put;
1221 		}
1222 		if (add_sched_out_event(atoms, 'R', timestamp))
1223 			goto out_put;
1224 	}
1225 
1226 	add_runtime_event(atoms, runtime, timestamp);
1227 	err = 0;
1228 out_put:
1229 	thread__put(thread);
1230 	return err;
1231 }
1232 
1233 static int latency_wakeup_event(struct perf_sched *sched,
1234 				struct evsel *evsel,
1235 				struct perf_sample *sample,
1236 				struct machine *machine)
1237 {
1238 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1239 	struct work_atoms *atoms;
1240 	struct work_atom *atom;
1241 	struct thread *wakee;
1242 	u64 timestamp = sample->time;
1243 	int err = -1;
1244 
1245 	wakee = machine__findnew_thread(machine, -1, pid);
1246 	if (wakee == NULL)
1247 		return -1;
1248 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1249 	if (!atoms) {
1250 		if (thread_atoms_insert(sched, wakee))
1251 			goto out_put;
1252 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1253 		if (!atoms) {
1254 			pr_err("wakeup-event: Internal tree error");
1255 			goto out_put;
1256 		}
1257 		if (add_sched_out_event(atoms, 'S', timestamp))
1258 			goto out_put;
1259 	}
1260 
1261 	BUG_ON(list_empty(&atoms->work_list));
1262 
1263 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1264 
1265 	/*
1266 	 * As we do not guarantee the wakeup event happens when
1267 	 * task is out of run queue, also may happen when task is
1268 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1269 	 * then we should not set the ->wake_up_time when wake up a
1270 	 * task which is on run queue.
1271 	 *
1272 	 * You WILL be missing events if you've recorded only
1273 	 * one CPU, or are only looking at only one, so don't
1274 	 * skip in this case.
1275 	 */
1276 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1277 		goto out_ok;
1278 
1279 	sched->nr_timestamps++;
1280 	if (atom->sched_out_time > timestamp) {
1281 		sched->nr_unordered_timestamps++;
1282 		goto out_ok;
1283 	}
1284 
1285 	atom->state = THREAD_WAIT_CPU;
1286 	atom->wake_up_time = timestamp;
1287 out_ok:
1288 	err = 0;
1289 out_put:
1290 	thread__put(wakee);
1291 	return err;
1292 }
1293 
1294 static int latency_migrate_task_event(struct perf_sched *sched,
1295 				      struct evsel *evsel,
1296 				      struct perf_sample *sample,
1297 				      struct machine *machine)
1298 {
1299 	const u32 pid = evsel__intval(evsel, sample, "pid");
1300 	u64 timestamp = sample->time;
1301 	struct work_atoms *atoms;
1302 	struct work_atom *atom;
1303 	struct thread *migrant;
1304 	int err = -1;
1305 
1306 	/*
1307 	 * Only need to worry about migration when profiling one CPU.
1308 	 */
1309 	if (sched->profile_cpu == -1)
1310 		return 0;
1311 
1312 	migrant = machine__findnew_thread(machine, -1, pid);
1313 	if (migrant == NULL)
1314 		return -1;
1315 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1316 	if (!atoms) {
1317 		if (thread_atoms_insert(sched, migrant))
1318 			goto out_put;
1319 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1320 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1321 		if (!atoms) {
1322 			pr_err("migration-event: Internal tree error");
1323 			goto out_put;
1324 		}
1325 		if (add_sched_out_event(atoms, 'R', timestamp))
1326 			goto out_put;
1327 	}
1328 
1329 	BUG_ON(list_empty(&atoms->work_list));
1330 
1331 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1332 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1333 
1334 	sched->nr_timestamps++;
1335 
1336 	if (atom->sched_out_time > timestamp)
1337 		sched->nr_unordered_timestamps++;
1338 	err = 0;
1339 out_put:
1340 	thread__put(migrant);
1341 	return err;
1342 }
1343 
1344 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1345 {
1346 	int i;
1347 	int ret;
1348 	u64 avg;
1349 	char max_lat_start[32], max_lat_end[32];
1350 
1351 	if (!work_list->nb_atoms)
1352 		return;
1353 	/*
1354 	 * Ignore idle threads:
1355 	 */
1356 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1357 		return;
1358 
1359 	sched->all_runtime += work_list->total_runtime;
1360 	sched->all_count   += work_list->nb_atoms;
1361 
1362 	if (work_list->num_merged > 1)
1363 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1364 	else
1365 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1366 
1367 	for (i = 0; i < 24 - ret; i++)
1368 		printf(" ");
1369 
1370 	avg = work_list->total_lat / work_list->nb_atoms;
1371 	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1372 	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1373 
1374 	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1375 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1376 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1377 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1378 		 max_lat_start, max_lat_end);
1379 }
1380 
1381 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1382 {
1383 	if (l->thread == r->thread)
1384 		return 0;
1385 	if (l->thread->tid < r->thread->tid)
1386 		return -1;
1387 	if (l->thread->tid > r->thread->tid)
1388 		return 1;
1389 	return (int)(l->thread - r->thread);
1390 }
1391 
1392 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1393 {
1394 	u64 avgl, avgr;
1395 
1396 	if (!l->nb_atoms)
1397 		return -1;
1398 
1399 	if (!r->nb_atoms)
1400 		return 1;
1401 
1402 	avgl = l->total_lat / l->nb_atoms;
1403 	avgr = r->total_lat / r->nb_atoms;
1404 
1405 	if (avgl < avgr)
1406 		return -1;
1407 	if (avgl > avgr)
1408 		return 1;
1409 
1410 	return 0;
1411 }
1412 
1413 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1414 {
1415 	if (l->max_lat < r->max_lat)
1416 		return -1;
1417 	if (l->max_lat > r->max_lat)
1418 		return 1;
1419 
1420 	return 0;
1421 }
1422 
1423 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1424 {
1425 	if (l->nb_atoms < r->nb_atoms)
1426 		return -1;
1427 	if (l->nb_atoms > r->nb_atoms)
1428 		return 1;
1429 
1430 	return 0;
1431 }
1432 
1433 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1434 {
1435 	if (l->total_runtime < r->total_runtime)
1436 		return -1;
1437 	if (l->total_runtime > r->total_runtime)
1438 		return 1;
1439 
1440 	return 0;
1441 }
1442 
1443 static int sort_dimension__add(const char *tok, struct list_head *list)
1444 {
1445 	size_t i;
1446 	static struct sort_dimension avg_sort_dimension = {
1447 		.name = "avg",
1448 		.cmp  = avg_cmp,
1449 	};
1450 	static struct sort_dimension max_sort_dimension = {
1451 		.name = "max",
1452 		.cmp  = max_cmp,
1453 	};
1454 	static struct sort_dimension pid_sort_dimension = {
1455 		.name = "pid",
1456 		.cmp  = pid_cmp,
1457 	};
1458 	static struct sort_dimension runtime_sort_dimension = {
1459 		.name = "runtime",
1460 		.cmp  = runtime_cmp,
1461 	};
1462 	static struct sort_dimension switch_sort_dimension = {
1463 		.name = "switch",
1464 		.cmp  = switch_cmp,
1465 	};
1466 	struct sort_dimension *available_sorts[] = {
1467 		&pid_sort_dimension,
1468 		&avg_sort_dimension,
1469 		&max_sort_dimension,
1470 		&switch_sort_dimension,
1471 		&runtime_sort_dimension,
1472 	};
1473 
1474 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1475 		if (!strcmp(available_sorts[i]->name, tok)) {
1476 			list_add_tail(&available_sorts[i]->list, list);
1477 
1478 			return 0;
1479 		}
1480 	}
1481 
1482 	return -1;
1483 }
1484 
1485 static void perf_sched__sort_lat(struct perf_sched *sched)
1486 {
1487 	struct rb_node *node;
1488 	struct rb_root_cached *root = &sched->atom_root;
1489 again:
1490 	for (;;) {
1491 		struct work_atoms *data;
1492 		node = rb_first_cached(root);
1493 		if (!node)
1494 			break;
1495 
1496 		rb_erase_cached(node, root);
1497 		data = rb_entry(node, struct work_atoms, node);
1498 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1499 	}
1500 	if (root == &sched->atom_root) {
1501 		root = &sched->merged_atom_root;
1502 		goto again;
1503 	}
1504 }
1505 
1506 static int process_sched_wakeup_event(struct perf_tool *tool,
1507 				      struct evsel *evsel,
1508 				      struct perf_sample *sample,
1509 				      struct machine *machine)
1510 {
1511 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1512 
1513 	if (sched->tp_handler->wakeup_event)
1514 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1515 
1516 	return 0;
1517 }
1518 
1519 union map_priv {
1520 	void	*ptr;
1521 	bool	 color;
1522 };
1523 
1524 static bool thread__has_color(struct thread *thread)
1525 {
1526 	union map_priv priv = {
1527 		.ptr = thread__priv(thread),
1528 	};
1529 
1530 	return priv.color;
1531 }
1532 
1533 static struct thread*
1534 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1535 {
1536 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1537 	union map_priv priv = {
1538 		.color = false,
1539 	};
1540 
1541 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1542 		return thread;
1543 
1544 	if (thread_map__has(sched->map.color_pids, tid))
1545 		priv.color = true;
1546 
1547 	thread__set_priv(thread, priv.ptr);
1548 	return thread;
1549 }
1550 
1551 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1552 			    struct perf_sample *sample, struct machine *machine)
1553 {
1554 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1555 	struct thread *sched_in;
1556 	struct thread_runtime *tr;
1557 	int new_shortname;
1558 	u64 timestamp0, timestamp = sample->time;
1559 	s64 delta;
1560 	int i;
1561 	struct perf_cpu this_cpu = {
1562 		.cpu = sample->cpu,
1563 	};
1564 	int cpus_nr;
1565 	bool new_cpu = false;
1566 	const char *color = PERF_COLOR_NORMAL;
1567 	char stimestamp[32];
1568 
1569 	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1570 
1571 	if (this_cpu.cpu > sched->max_cpu.cpu)
1572 		sched->max_cpu = this_cpu;
1573 
1574 	if (sched->map.comp) {
1575 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1576 		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1577 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1578 			new_cpu = true;
1579 		}
1580 	} else
1581 		cpus_nr = sched->max_cpu.cpu;
1582 
1583 	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1584 	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1585 	if (timestamp0)
1586 		delta = timestamp - timestamp0;
1587 	else
1588 		delta = 0;
1589 
1590 	if (delta < 0) {
1591 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1592 		return -1;
1593 	}
1594 
1595 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1596 	if (sched_in == NULL)
1597 		return -1;
1598 
1599 	tr = thread__get_runtime(sched_in);
1600 	if (tr == NULL) {
1601 		thread__put(sched_in);
1602 		return -1;
1603 	}
1604 
1605 	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1606 
1607 	printf("  ");
1608 
1609 	new_shortname = 0;
1610 	if (!tr->shortname[0]) {
1611 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1612 			/*
1613 			 * Don't allocate a letter-number for swapper:0
1614 			 * as a shortname. Instead, we use '.' for it.
1615 			 */
1616 			tr->shortname[0] = '.';
1617 			tr->shortname[1] = ' ';
1618 		} else {
1619 			tr->shortname[0] = sched->next_shortname1;
1620 			tr->shortname[1] = sched->next_shortname2;
1621 
1622 			if (sched->next_shortname1 < 'Z') {
1623 				sched->next_shortname1++;
1624 			} else {
1625 				sched->next_shortname1 = 'A';
1626 				if (sched->next_shortname2 < '9')
1627 					sched->next_shortname2++;
1628 				else
1629 					sched->next_shortname2 = '0';
1630 			}
1631 		}
1632 		new_shortname = 1;
1633 	}
1634 
1635 	for (i = 0; i < cpus_nr; i++) {
1636 		struct perf_cpu cpu = {
1637 			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1638 		};
1639 		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1640 		struct thread_runtime *curr_tr;
1641 		const char *pid_color = color;
1642 		const char *cpu_color = color;
1643 
1644 		if (curr_thread && thread__has_color(curr_thread))
1645 			pid_color = COLOR_PIDS;
1646 
1647 		if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1648 			continue;
1649 
1650 		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1651 			cpu_color = COLOR_CPUS;
1652 
1653 		if (cpu.cpu != this_cpu.cpu)
1654 			color_fprintf(stdout, color, " ");
1655 		else
1656 			color_fprintf(stdout, cpu_color, "*");
1657 
1658 		if (sched->curr_thread[cpu.cpu]) {
1659 			curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1660 			if (curr_tr == NULL) {
1661 				thread__put(sched_in);
1662 				return -1;
1663 			}
1664 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1665 		} else
1666 			color_fprintf(stdout, color, "   ");
1667 	}
1668 
1669 	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1670 		goto out;
1671 
1672 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1673 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1674 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1675 		const char *pid_color = color;
1676 
1677 		if (thread__has_color(sched_in))
1678 			pid_color = COLOR_PIDS;
1679 
1680 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1681 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1682 		tr->comm_changed = false;
1683 	}
1684 
1685 	if (sched->map.comp && new_cpu)
1686 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1687 
1688 out:
1689 	color_fprintf(stdout, color, "\n");
1690 
1691 	thread__put(sched_in);
1692 
1693 	return 0;
1694 }
1695 
1696 static int process_sched_switch_event(struct perf_tool *tool,
1697 				      struct evsel *evsel,
1698 				      struct perf_sample *sample,
1699 				      struct machine *machine)
1700 {
1701 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1702 	int this_cpu = sample->cpu, err = 0;
1703 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1704 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1705 
1706 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1707 		/*
1708 		 * Are we trying to switch away a PID that is
1709 		 * not current?
1710 		 */
1711 		if (sched->curr_pid[this_cpu] != prev_pid)
1712 			sched->nr_context_switch_bugs++;
1713 	}
1714 
1715 	if (sched->tp_handler->switch_event)
1716 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1717 
1718 	sched->curr_pid[this_cpu] = next_pid;
1719 	return err;
1720 }
1721 
1722 static int process_sched_runtime_event(struct perf_tool *tool,
1723 				       struct evsel *evsel,
1724 				       struct perf_sample *sample,
1725 				       struct machine *machine)
1726 {
1727 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1728 
1729 	if (sched->tp_handler->runtime_event)
1730 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1731 
1732 	return 0;
1733 }
1734 
1735 static int perf_sched__process_fork_event(struct perf_tool *tool,
1736 					  union perf_event *event,
1737 					  struct perf_sample *sample,
1738 					  struct machine *machine)
1739 {
1740 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1741 
1742 	/* run the fork event through the perf machinery */
1743 	perf_event__process_fork(tool, event, sample, machine);
1744 
1745 	/* and then run additional processing needed for this command */
1746 	if (sched->tp_handler->fork_event)
1747 		return sched->tp_handler->fork_event(sched, event, machine);
1748 
1749 	return 0;
1750 }
1751 
1752 static int process_sched_migrate_task_event(struct perf_tool *tool,
1753 					    struct evsel *evsel,
1754 					    struct perf_sample *sample,
1755 					    struct machine *machine)
1756 {
1757 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1758 
1759 	if (sched->tp_handler->migrate_task_event)
1760 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1761 
1762 	return 0;
1763 }
1764 
1765 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1766 				  struct evsel *evsel,
1767 				  struct perf_sample *sample,
1768 				  struct machine *machine);
1769 
1770 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1771 						 union perf_event *event __maybe_unused,
1772 						 struct perf_sample *sample,
1773 						 struct evsel *evsel,
1774 						 struct machine *machine)
1775 {
1776 	int err = 0;
1777 
1778 	if (evsel->handler != NULL) {
1779 		tracepoint_handler f = evsel->handler;
1780 		err = f(tool, evsel, sample, machine);
1781 	}
1782 
1783 	return err;
1784 }
1785 
1786 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1787 				    union perf_event *event,
1788 				    struct perf_sample *sample,
1789 				    struct machine *machine)
1790 {
1791 	struct thread *thread;
1792 	struct thread_runtime *tr;
1793 	int err;
1794 
1795 	err = perf_event__process_comm(tool, event, sample, machine);
1796 	if (err)
1797 		return err;
1798 
1799 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1800 	if (!thread) {
1801 		pr_err("Internal error: can't find thread\n");
1802 		return -1;
1803 	}
1804 
1805 	tr = thread__get_runtime(thread);
1806 	if (tr == NULL) {
1807 		thread__put(thread);
1808 		return -1;
1809 	}
1810 
1811 	tr->comm_changed = true;
1812 	thread__put(thread);
1813 
1814 	return 0;
1815 }
1816 
1817 static int perf_sched__read_events(struct perf_sched *sched)
1818 {
1819 	const struct evsel_str_handler handlers[] = {
1820 		{ "sched:sched_switch",	      process_sched_switch_event, },
1821 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1822 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1823 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1824 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1825 	};
1826 	struct perf_session *session;
1827 	struct perf_data data = {
1828 		.path  = input_name,
1829 		.mode  = PERF_DATA_MODE_READ,
1830 		.force = sched->force,
1831 	};
1832 	int rc = -1;
1833 
1834 	session = perf_session__new(&data, &sched->tool);
1835 	if (IS_ERR(session)) {
1836 		pr_debug("Error creating perf session");
1837 		return PTR_ERR(session);
1838 	}
1839 
1840 	symbol__init(&session->header.env);
1841 
1842 	if (perf_session__set_tracepoints_handlers(session, handlers))
1843 		goto out_delete;
1844 
1845 	if (perf_session__has_traces(session, "record -R")) {
1846 		int err = perf_session__process_events(session);
1847 		if (err) {
1848 			pr_err("Failed to process events, error %d", err);
1849 			goto out_delete;
1850 		}
1851 
1852 		sched->nr_events      = session->evlist->stats.nr_events[0];
1853 		sched->nr_lost_events = session->evlist->stats.total_lost;
1854 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1855 	}
1856 
1857 	rc = 0;
1858 out_delete:
1859 	perf_session__delete(session);
1860 	return rc;
1861 }
1862 
1863 /*
1864  * scheduling times are printed as msec.usec
1865  */
1866 static inline void print_sched_time(unsigned long long nsecs, int width)
1867 {
1868 	unsigned long msecs;
1869 	unsigned long usecs;
1870 
1871 	msecs  = nsecs / NSEC_PER_MSEC;
1872 	nsecs -= msecs * NSEC_PER_MSEC;
1873 	usecs  = nsecs / NSEC_PER_USEC;
1874 	printf("%*lu.%03lu ", width, msecs, usecs);
1875 }
1876 
1877 /*
1878  * returns runtime data for event, allocating memory for it the
1879  * first time it is used.
1880  */
1881 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1882 {
1883 	struct evsel_runtime *r = evsel->priv;
1884 
1885 	if (r == NULL) {
1886 		r = zalloc(sizeof(struct evsel_runtime));
1887 		evsel->priv = r;
1888 	}
1889 
1890 	return r;
1891 }
1892 
1893 /*
1894  * save last time event was seen per cpu
1895  */
1896 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1897 {
1898 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1899 
1900 	if (r == NULL)
1901 		return;
1902 
1903 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1904 		int i, n = __roundup_pow_of_two(cpu+1);
1905 		void *p = r->last_time;
1906 
1907 		p = realloc(r->last_time, n * sizeof(u64));
1908 		if (!p)
1909 			return;
1910 
1911 		r->last_time = p;
1912 		for (i = r->ncpu; i < n; ++i)
1913 			r->last_time[i] = (u64) 0;
1914 
1915 		r->ncpu = n;
1916 	}
1917 
1918 	r->last_time[cpu] = timestamp;
1919 }
1920 
1921 /* returns last time this event was seen on the given cpu */
1922 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1923 {
1924 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1925 
1926 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1927 		return 0;
1928 
1929 	return r->last_time[cpu];
1930 }
1931 
1932 static int comm_width = 30;
1933 
1934 static char *timehist_get_commstr(struct thread *thread)
1935 {
1936 	static char str[32];
1937 	const char *comm = thread__comm_str(thread);
1938 	pid_t tid = thread->tid;
1939 	pid_t pid = thread->pid_;
1940 	int n;
1941 
1942 	if (pid == 0)
1943 		n = scnprintf(str, sizeof(str), "%s", comm);
1944 
1945 	else if (tid != pid)
1946 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1947 
1948 	else
1949 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1950 
1951 	if (n > comm_width)
1952 		comm_width = n;
1953 
1954 	return str;
1955 }
1956 
1957 static void timehist_header(struct perf_sched *sched)
1958 {
1959 	u32 ncpus = sched->max_cpu.cpu + 1;
1960 	u32 i, j;
1961 
1962 	printf("%15s %6s ", "time", "cpu");
1963 
1964 	if (sched->show_cpu_visual) {
1965 		printf(" ");
1966 		for (i = 0, j = 0; i < ncpus; ++i) {
1967 			printf("%x", j++);
1968 			if (j > 15)
1969 				j = 0;
1970 		}
1971 		printf(" ");
1972 	}
1973 
1974 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1975 		"task name", "wait time", "sch delay", "run time");
1976 
1977 	if (sched->show_state)
1978 		printf("  %s", "state");
1979 
1980 	printf("\n");
1981 
1982 	/*
1983 	 * units row
1984 	 */
1985 	printf("%15s %-6s ", "", "");
1986 
1987 	if (sched->show_cpu_visual)
1988 		printf(" %*s ", ncpus, "");
1989 
1990 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1991 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1992 
1993 	if (sched->show_state)
1994 		printf("  %5s", "");
1995 
1996 	printf("\n");
1997 
1998 	/*
1999 	 * separator
2000 	 */
2001 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2002 
2003 	if (sched->show_cpu_visual)
2004 		printf(" %.*s ", ncpus, graph_dotted_line);
2005 
2006 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2007 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
2008 		graph_dotted_line);
2009 
2010 	if (sched->show_state)
2011 		printf("  %.5s", graph_dotted_line);
2012 
2013 	printf("\n");
2014 }
2015 
2016 static char task_state_char(struct thread *thread, int state)
2017 {
2018 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
2019 	unsigned bit = state ? ffs(state) : 0;
2020 
2021 	/* 'I' for idle */
2022 	if (thread->tid == 0)
2023 		return 'I';
2024 
2025 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
2026 }
2027 
2028 static void timehist_print_sample(struct perf_sched *sched,
2029 				  struct evsel *evsel,
2030 				  struct perf_sample *sample,
2031 				  struct addr_location *al,
2032 				  struct thread *thread,
2033 				  u64 t, int state)
2034 {
2035 	struct thread_runtime *tr = thread__priv(thread);
2036 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2037 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2038 	u32 max_cpus = sched->max_cpu.cpu + 1;
2039 	char tstr[64];
2040 	char nstr[30];
2041 	u64 wait_time;
2042 
2043 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2044 		return;
2045 
2046 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2047 	printf("%15s [%04d] ", tstr, sample->cpu);
2048 
2049 	if (sched->show_cpu_visual) {
2050 		u32 i;
2051 		char c;
2052 
2053 		printf(" ");
2054 		for (i = 0; i < max_cpus; ++i) {
2055 			/* flag idle times with 'i'; others are sched events */
2056 			if (i == sample->cpu)
2057 				c = (thread->tid == 0) ? 'i' : 's';
2058 			else
2059 				c = ' ';
2060 			printf("%c", c);
2061 		}
2062 		printf(" ");
2063 	}
2064 
2065 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2066 
2067 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2068 	print_sched_time(wait_time, 6);
2069 
2070 	print_sched_time(tr->dt_delay, 6);
2071 	print_sched_time(tr->dt_run, 6);
2072 
2073 	if (sched->show_state)
2074 		printf(" %5c ", task_state_char(thread, state));
2075 
2076 	if (sched->show_next) {
2077 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2078 		printf(" %-*s", comm_width, nstr);
2079 	}
2080 
2081 	if (sched->show_wakeups && !sched->show_next)
2082 		printf("  %-*s", comm_width, "");
2083 
2084 	if (thread->tid == 0)
2085 		goto out;
2086 
2087 	if (sched->show_callchain)
2088 		printf("  ");
2089 
2090 	sample__fprintf_sym(sample, al, 0,
2091 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2092 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2093 			    EVSEL__PRINT_SKIP_IGNORED,
2094 			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2095 
2096 out:
2097 	printf("\n");
2098 }
2099 
2100 /*
2101  * Explanation of delta-time stats:
2102  *
2103  *            t = time of current schedule out event
2104  *        tprev = time of previous sched out event
2105  *                also time of schedule-in event for current task
2106  *    last_time = time of last sched change event for current task
2107  *                (i.e, time process was last scheduled out)
2108  * ready_to_run = time of wakeup for current task
2109  *
2110  * -----|------------|------------|------------|------
2111  *    last         ready        tprev          t
2112  *    time         to run
2113  *
2114  *      |-------- dt_wait --------|
2115  *                   |- dt_delay -|-- dt_run --|
2116  *
2117  *   dt_run = run time of current task
2118  *  dt_wait = time between last schedule out event for task and tprev
2119  *            represents time spent off the cpu
2120  * dt_delay = time between wakeup and schedule-in of task
2121  */
2122 
2123 static void timehist_update_runtime_stats(struct thread_runtime *r,
2124 					 u64 t, u64 tprev)
2125 {
2126 	r->dt_delay   = 0;
2127 	r->dt_sleep   = 0;
2128 	r->dt_iowait  = 0;
2129 	r->dt_preempt = 0;
2130 	r->dt_run     = 0;
2131 
2132 	if (tprev) {
2133 		r->dt_run = t - tprev;
2134 		if (r->ready_to_run) {
2135 			if (r->ready_to_run > tprev)
2136 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2137 			else
2138 				r->dt_delay = tprev - r->ready_to_run;
2139 		}
2140 
2141 		if (r->last_time > tprev)
2142 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2143 		else if (r->last_time) {
2144 			u64 dt_wait = tprev - r->last_time;
2145 
2146 			if (r->last_state == TASK_RUNNING)
2147 				r->dt_preempt = dt_wait;
2148 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2149 				r->dt_iowait = dt_wait;
2150 			else
2151 				r->dt_sleep = dt_wait;
2152 		}
2153 	}
2154 
2155 	update_stats(&r->run_stats, r->dt_run);
2156 
2157 	r->total_run_time     += r->dt_run;
2158 	r->total_delay_time   += r->dt_delay;
2159 	r->total_sleep_time   += r->dt_sleep;
2160 	r->total_iowait_time  += r->dt_iowait;
2161 	r->total_preempt_time += r->dt_preempt;
2162 }
2163 
2164 static bool is_idle_sample(struct perf_sample *sample,
2165 			   struct evsel *evsel)
2166 {
2167 	/* pid 0 == swapper == idle task */
2168 	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2169 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2170 
2171 	return sample->pid == 0;
2172 }
2173 
2174 static void save_task_callchain(struct perf_sched *sched,
2175 				struct perf_sample *sample,
2176 				struct evsel *evsel,
2177 				struct machine *machine)
2178 {
2179 	struct callchain_cursor *cursor = &callchain_cursor;
2180 	struct thread *thread;
2181 
2182 	/* want main thread for process - has maps */
2183 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2184 	if (thread == NULL) {
2185 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2186 		return;
2187 	}
2188 
2189 	if (!sched->show_callchain || sample->callchain == NULL)
2190 		return;
2191 
2192 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2193 				      NULL, NULL, sched->max_stack + 2) != 0) {
2194 		if (verbose > 0)
2195 			pr_err("Failed to resolve callchain. Skipping\n");
2196 
2197 		return;
2198 	}
2199 
2200 	callchain_cursor_commit(cursor);
2201 
2202 	while (true) {
2203 		struct callchain_cursor_node *node;
2204 		struct symbol *sym;
2205 
2206 		node = callchain_cursor_current(cursor);
2207 		if (node == NULL)
2208 			break;
2209 
2210 		sym = node->ms.sym;
2211 		if (sym) {
2212 			if (!strcmp(sym->name, "schedule") ||
2213 			    !strcmp(sym->name, "__schedule") ||
2214 			    !strcmp(sym->name, "preempt_schedule"))
2215 				sym->ignore = 1;
2216 		}
2217 
2218 		callchain_cursor_advance(cursor);
2219 	}
2220 }
2221 
2222 static int init_idle_thread(struct thread *thread)
2223 {
2224 	struct idle_thread_runtime *itr;
2225 
2226 	thread__set_comm(thread, idle_comm, 0);
2227 
2228 	itr = zalloc(sizeof(*itr));
2229 	if (itr == NULL)
2230 		return -ENOMEM;
2231 
2232 	init_stats(&itr->tr.run_stats);
2233 	callchain_init(&itr->callchain);
2234 	callchain_cursor_reset(&itr->cursor);
2235 	thread__set_priv(thread, itr);
2236 
2237 	return 0;
2238 }
2239 
2240 /*
2241  * Track idle stats per cpu by maintaining a local thread
2242  * struct for the idle task on each cpu.
2243  */
2244 static int init_idle_threads(int ncpu)
2245 {
2246 	int i, ret;
2247 
2248 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2249 	if (!idle_threads)
2250 		return -ENOMEM;
2251 
2252 	idle_max_cpu = ncpu;
2253 
2254 	/* allocate the actual thread struct if needed */
2255 	for (i = 0; i < ncpu; ++i) {
2256 		idle_threads[i] = thread__new(0, 0);
2257 		if (idle_threads[i] == NULL)
2258 			return -ENOMEM;
2259 
2260 		ret = init_idle_thread(idle_threads[i]);
2261 		if (ret < 0)
2262 			return ret;
2263 	}
2264 
2265 	return 0;
2266 }
2267 
2268 static void free_idle_threads(void)
2269 {
2270 	int i;
2271 
2272 	if (idle_threads == NULL)
2273 		return;
2274 
2275 	for (i = 0; i < idle_max_cpu; ++i) {
2276 		if ((idle_threads[i]))
2277 			thread__delete(idle_threads[i]);
2278 	}
2279 
2280 	free(idle_threads);
2281 }
2282 
2283 static struct thread *get_idle_thread(int cpu)
2284 {
2285 	/*
2286 	 * expand/allocate array of pointers to local thread
2287 	 * structs if needed
2288 	 */
2289 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2290 		int i, j = __roundup_pow_of_two(cpu+1);
2291 		void *p;
2292 
2293 		p = realloc(idle_threads, j * sizeof(struct thread *));
2294 		if (!p)
2295 			return NULL;
2296 
2297 		idle_threads = (struct thread **) p;
2298 		for (i = idle_max_cpu; i < j; ++i)
2299 			idle_threads[i] = NULL;
2300 
2301 		idle_max_cpu = j;
2302 	}
2303 
2304 	/* allocate a new thread struct if needed */
2305 	if (idle_threads[cpu] == NULL) {
2306 		idle_threads[cpu] = thread__new(0, 0);
2307 
2308 		if (idle_threads[cpu]) {
2309 			if (init_idle_thread(idle_threads[cpu]) < 0)
2310 				return NULL;
2311 		}
2312 	}
2313 
2314 	return idle_threads[cpu];
2315 }
2316 
2317 static void save_idle_callchain(struct perf_sched *sched,
2318 				struct idle_thread_runtime *itr,
2319 				struct perf_sample *sample)
2320 {
2321 	if (!sched->show_callchain || sample->callchain == NULL)
2322 		return;
2323 
2324 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2325 }
2326 
2327 static struct thread *timehist_get_thread(struct perf_sched *sched,
2328 					  struct perf_sample *sample,
2329 					  struct machine *machine,
2330 					  struct evsel *evsel)
2331 {
2332 	struct thread *thread;
2333 
2334 	if (is_idle_sample(sample, evsel)) {
2335 		thread = get_idle_thread(sample->cpu);
2336 		if (thread == NULL)
2337 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2338 
2339 	} else {
2340 		/* there were samples with tid 0 but non-zero pid */
2341 		thread = machine__findnew_thread(machine, sample->pid,
2342 						 sample->tid ?: sample->pid);
2343 		if (thread == NULL) {
2344 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2345 				 sample->tid);
2346 		}
2347 
2348 		save_task_callchain(sched, sample, evsel, machine);
2349 		if (sched->idle_hist) {
2350 			struct thread *idle;
2351 			struct idle_thread_runtime *itr;
2352 
2353 			idle = get_idle_thread(sample->cpu);
2354 			if (idle == NULL) {
2355 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2356 				return NULL;
2357 			}
2358 
2359 			itr = thread__priv(idle);
2360 			if (itr == NULL)
2361 				return NULL;
2362 
2363 			itr->last_thread = thread;
2364 
2365 			/* copy task callchain when entering to idle */
2366 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2367 				save_idle_callchain(sched, itr, sample);
2368 		}
2369 	}
2370 
2371 	return thread;
2372 }
2373 
2374 static bool timehist_skip_sample(struct perf_sched *sched,
2375 				 struct thread *thread,
2376 				 struct evsel *evsel,
2377 				 struct perf_sample *sample)
2378 {
2379 	bool rc = false;
2380 
2381 	if (thread__is_filtered(thread)) {
2382 		rc = true;
2383 		sched->skipped_samples++;
2384 	}
2385 
2386 	if (sched->idle_hist) {
2387 		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2388 			rc = true;
2389 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2390 			 evsel__intval(evsel, sample, "next_pid") != 0)
2391 			rc = true;
2392 	}
2393 
2394 	return rc;
2395 }
2396 
2397 static void timehist_print_wakeup_event(struct perf_sched *sched,
2398 					struct evsel *evsel,
2399 					struct perf_sample *sample,
2400 					struct machine *machine,
2401 					struct thread *awakened)
2402 {
2403 	struct thread *thread;
2404 	char tstr[64];
2405 
2406 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2407 	if (thread == NULL)
2408 		return;
2409 
2410 	/* show wakeup unless both awakee and awaker are filtered */
2411 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2412 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2413 		return;
2414 	}
2415 
2416 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2417 	printf("%15s [%04d] ", tstr, sample->cpu);
2418 	if (sched->show_cpu_visual)
2419 		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2420 
2421 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2422 
2423 	/* dt spacer */
2424 	printf("  %9s  %9s  %9s ", "", "", "");
2425 
2426 	printf("awakened: %s", timehist_get_commstr(awakened));
2427 
2428 	printf("\n");
2429 }
2430 
2431 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2432 					union perf_event *event __maybe_unused,
2433 					struct evsel *evsel __maybe_unused,
2434 					struct perf_sample *sample __maybe_unused,
2435 					struct machine *machine __maybe_unused)
2436 {
2437 	return 0;
2438 }
2439 
2440 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2441 				       union perf_event *event __maybe_unused,
2442 				       struct evsel *evsel,
2443 				       struct perf_sample *sample,
2444 				       struct machine *machine)
2445 {
2446 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2447 	struct thread *thread;
2448 	struct thread_runtime *tr = NULL;
2449 	/* want pid of awakened task not pid in sample */
2450 	const u32 pid = evsel__intval(evsel, sample, "pid");
2451 
2452 	thread = machine__findnew_thread(machine, 0, pid);
2453 	if (thread == NULL)
2454 		return -1;
2455 
2456 	tr = thread__get_runtime(thread);
2457 	if (tr == NULL)
2458 		return -1;
2459 
2460 	if (tr->ready_to_run == 0)
2461 		tr->ready_to_run = sample->time;
2462 
2463 	/* show wakeups if requested */
2464 	if (sched->show_wakeups &&
2465 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2466 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2467 
2468 	return 0;
2469 }
2470 
2471 static void timehist_print_migration_event(struct perf_sched *sched,
2472 					struct evsel *evsel,
2473 					struct perf_sample *sample,
2474 					struct machine *machine,
2475 					struct thread *migrated)
2476 {
2477 	struct thread *thread;
2478 	char tstr[64];
2479 	u32 max_cpus;
2480 	u32 ocpu, dcpu;
2481 
2482 	if (sched->summary_only)
2483 		return;
2484 
2485 	max_cpus = sched->max_cpu.cpu + 1;
2486 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2487 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2488 
2489 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2490 	if (thread == NULL)
2491 		return;
2492 
2493 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2494 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2495 		return;
2496 	}
2497 
2498 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2499 	printf("%15s [%04d] ", tstr, sample->cpu);
2500 
2501 	if (sched->show_cpu_visual) {
2502 		u32 i;
2503 		char c;
2504 
2505 		printf("  ");
2506 		for (i = 0; i < max_cpus; ++i) {
2507 			c = (i == sample->cpu) ? 'm' : ' ';
2508 			printf("%c", c);
2509 		}
2510 		printf("  ");
2511 	}
2512 
2513 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2514 
2515 	/* dt spacer */
2516 	printf("  %9s  %9s  %9s ", "", "", "");
2517 
2518 	printf("migrated: %s", timehist_get_commstr(migrated));
2519 	printf(" cpu %d => %d", ocpu, dcpu);
2520 
2521 	printf("\n");
2522 }
2523 
2524 static int timehist_migrate_task_event(struct perf_tool *tool,
2525 				       union perf_event *event __maybe_unused,
2526 				       struct evsel *evsel,
2527 				       struct perf_sample *sample,
2528 				       struct machine *machine)
2529 {
2530 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2531 	struct thread *thread;
2532 	struct thread_runtime *tr = NULL;
2533 	/* want pid of migrated task not pid in sample */
2534 	const u32 pid = evsel__intval(evsel, sample, "pid");
2535 
2536 	thread = machine__findnew_thread(machine, 0, pid);
2537 	if (thread == NULL)
2538 		return -1;
2539 
2540 	tr = thread__get_runtime(thread);
2541 	if (tr == NULL)
2542 		return -1;
2543 
2544 	tr->migrations++;
2545 
2546 	/* show migrations if requested */
2547 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2548 
2549 	return 0;
2550 }
2551 
2552 static int timehist_sched_change_event(struct perf_tool *tool,
2553 				       union perf_event *event,
2554 				       struct evsel *evsel,
2555 				       struct perf_sample *sample,
2556 				       struct machine *machine)
2557 {
2558 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2559 	struct perf_time_interval *ptime = &sched->ptime;
2560 	struct addr_location al;
2561 	struct thread *thread;
2562 	struct thread_runtime *tr = NULL;
2563 	u64 tprev, t = sample->time;
2564 	int rc = 0;
2565 	int state = evsel__intval(evsel, sample, "prev_state");
2566 
2567 	if (machine__resolve(machine, &al, sample) < 0) {
2568 		pr_err("problem processing %d event. skipping it\n",
2569 		       event->header.type);
2570 		rc = -1;
2571 		goto out;
2572 	}
2573 
2574 	thread = timehist_get_thread(sched, sample, machine, evsel);
2575 	if (thread == NULL) {
2576 		rc = -1;
2577 		goto out;
2578 	}
2579 
2580 	if (timehist_skip_sample(sched, thread, evsel, sample))
2581 		goto out;
2582 
2583 	tr = thread__get_runtime(thread);
2584 	if (tr == NULL) {
2585 		rc = -1;
2586 		goto out;
2587 	}
2588 
2589 	tprev = evsel__get_time(evsel, sample->cpu);
2590 
2591 	/*
2592 	 * If start time given:
2593 	 * - sample time is under window user cares about - skip sample
2594 	 * - tprev is under window user cares about  - reset to start of window
2595 	 */
2596 	if (ptime->start && ptime->start > t)
2597 		goto out;
2598 
2599 	if (tprev && ptime->start > tprev)
2600 		tprev = ptime->start;
2601 
2602 	/*
2603 	 * If end time given:
2604 	 * - previous sched event is out of window - we are done
2605 	 * - sample time is beyond window user cares about - reset it
2606 	 *   to close out stats for time window interest
2607 	 */
2608 	if (ptime->end) {
2609 		if (tprev > ptime->end)
2610 			goto out;
2611 
2612 		if (t > ptime->end)
2613 			t = ptime->end;
2614 	}
2615 
2616 	if (!sched->idle_hist || thread->tid == 0) {
2617 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2618 			timehist_update_runtime_stats(tr, t, tprev);
2619 
2620 		if (sched->idle_hist) {
2621 			struct idle_thread_runtime *itr = (void *)tr;
2622 			struct thread_runtime *last_tr;
2623 
2624 			BUG_ON(thread->tid != 0);
2625 
2626 			if (itr->last_thread == NULL)
2627 				goto out;
2628 
2629 			/* add current idle time as last thread's runtime */
2630 			last_tr = thread__get_runtime(itr->last_thread);
2631 			if (last_tr == NULL)
2632 				goto out;
2633 
2634 			timehist_update_runtime_stats(last_tr, t, tprev);
2635 			/*
2636 			 * remove delta time of last thread as it's not updated
2637 			 * and otherwise it will show an invalid value next
2638 			 * time.  we only care total run time and run stat.
2639 			 */
2640 			last_tr->dt_run = 0;
2641 			last_tr->dt_delay = 0;
2642 			last_tr->dt_sleep = 0;
2643 			last_tr->dt_iowait = 0;
2644 			last_tr->dt_preempt = 0;
2645 
2646 			if (itr->cursor.nr)
2647 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2648 
2649 			itr->last_thread = NULL;
2650 		}
2651 	}
2652 
2653 	if (!sched->summary_only)
2654 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2655 
2656 out:
2657 	if (sched->hist_time.start == 0 && t >= ptime->start)
2658 		sched->hist_time.start = t;
2659 	if (ptime->end == 0 || t <= ptime->end)
2660 		sched->hist_time.end = t;
2661 
2662 	if (tr) {
2663 		/* time of this sched_switch event becomes last time task seen */
2664 		tr->last_time = sample->time;
2665 
2666 		/* last state is used to determine where to account wait time */
2667 		tr->last_state = state;
2668 
2669 		/* sched out event for task so reset ready to run time */
2670 		tr->ready_to_run = 0;
2671 	}
2672 
2673 	evsel__save_time(evsel, sample->time, sample->cpu);
2674 
2675 	return rc;
2676 }
2677 
2678 static int timehist_sched_switch_event(struct perf_tool *tool,
2679 			     union perf_event *event,
2680 			     struct evsel *evsel,
2681 			     struct perf_sample *sample,
2682 			     struct machine *machine __maybe_unused)
2683 {
2684 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2685 }
2686 
2687 static int process_lost(struct perf_tool *tool __maybe_unused,
2688 			union perf_event *event,
2689 			struct perf_sample *sample,
2690 			struct machine *machine __maybe_unused)
2691 {
2692 	char tstr[64];
2693 
2694 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2695 	printf("%15s ", tstr);
2696 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2697 
2698 	return 0;
2699 }
2700 
2701 
2702 static void print_thread_runtime(struct thread *t,
2703 				 struct thread_runtime *r)
2704 {
2705 	double mean = avg_stats(&r->run_stats);
2706 	float stddev;
2707 
2708 	printf("%*s   %5d  %9" PRIu64 " ",
2709 	       comm_width, timehist_get_commstr(t), t->ppid,
2710 	       (u64) r->run_stats.n);
2711 
2712 	print_sched_time(r->total_run_time, 8);
2713 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2714 	print_sched_time(r->run_stats.min, 6);
2715 	printf(" ");
2716 	print_sched_time((u64) mean, 6);
2717 	printf(" ");
2718 	print_sched_time(r->run_stats.max, 6);
2719 	printf("  ");
2720 	printf("%5.2f", stddev);
2721 	printf("   %5" PRIu64, r->migrations);
2722 	printf("\n");
2723 }
2724 
2725 static void print_thread_waittime(struct thread *t,
2726 				  struct thread_runtime *r)
2727 {
2728 	printf("%*s   %5d  %9" PRIu64 " ",
2729 	       comm_width, timehist_get_commstr(t), t->ppid,
2730 	       (u64) r->run_stats.n);
2731 
2732 	print_sched_time(r->total_run_time, 8);
2733 	print_sched_time(r->total_sleep_time, 6);
2734 	printf(" ");
2735 	print_sched_time(r->total_iowait_time, 6);
2736 	printf(" ");
2737 	print_sched_time(r->total_preempt_time, 6);
2738 	printf(" ");
2739 	print_sched_time(r->total_delay_time, 6);
2740 	printf("\n");
2741 }
2742 
2743 struct total_run_stats {
2744 	struct perf_sched *sched;
2745 	u64  sched_count;
2746 	u64  task_count;
2747 	u64  total_run_time;
2748 };
2749 
2750 static int __show_thread_runtime(struct thread *t, void *priv)
2751 {
2752 	struct total_run_stats *stats = priv;
2753 	struct thread_runtime *r;
2754 
2755 	if (thread__is_filtered(t))
2756 		return 0;
2757 
2758 	r = thread__priv(t);
2759 	if (r && r->run_stats.n) {
2760 		stats->task_count++;
2761 		stats->sched_count += r->run_stats.n;
2762 		stats->total_run_time += r->total_run_time;
2763 
2764 		if (stats->sched->show_state)
2765 			print_thread_waittime(t, r);
2766 		else
2767 			print_thread_runtime(t, r);
2768 	}
2769 
2770 	return 0;
2771 }
2772 
2773 static int show_thread_runtime(struct thread *t, void *priv)
2774 {
2775 	if (t->dead)
2776 		return 0;
2777 
2778 	return __show_thread_runtime(t, priv);
2779 }
2780 
2781 static int show_deadthread_runtime(struct thread *t, void *priv)
2782 {
2783 	if (!t->dead)
2784 		return 0;
2785 
2786 	return __show_thread_runtime(t, priv);
2787 }
2788 
2789 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2790 {
2791 	const char *sep = " <- ";
2792 	struct callchain_list *chain;
2793 	size_t ret = 0;
2794 	char bf[1024];
2795 	bool first;
2796 
2797 	if (node == NULL)
2798 		return 0;
2799 
2800 	ret = callchain__fprintf_folded(fp, node->parent);
2801 	first = (ret == 0);
2802 
2803 	list_for_each_entry(chain, &node->val, list) {
2804 		if (chain->ip >= PERF_CONTEXT_MAX)
2805 			continue;
2806 		if (chain->ms.sym && chain->ms.sym->ignore)
2807 			continue;
2808 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2809 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2810 							false));
2811 		first = false;
2812 	}
2813 
2814 	return ret;
2815 }
2816 
2817 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2818 {
2819 	size_t ret = 0;
2820 	FILE *fp = stdout;
2821 	struct callchain_node *chain;
2822 	struct rb_node *rb_node = rb_first_cached(root);
2823 
2824 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2825 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2826 	       graph_dotted_line);
2827 
2828 	while (rb_node) {
2829 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2830 		rb_node = rb_next(rb_node);
2831 
2832 		ret += fprintf(fp, "  ");
2833 		print_sched_time(chain->hit, 12);
2834 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2835 		ret += fprintf(fp, " %8d  ", chain->count);
2836 		ret += callchain__fprintf_folded(fp, chain);
2837 		ret += fprintf(fp, "\n");
2838 	}
2839 
2840 	return ret;
2841 }
2842 
2843 static void timehist_print_summary(struct perf_sched *sched,
2844 				   struct perf_session *session)
2845 {
2846 	struct machine *m = &session->machines.host;
2847 	struct total_run_stats totals;
2848 	u64 task_count;
2849 	struct thread *t;
2850 	struct thread_runtime *r;
2851 	int i;
2852 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2853 
2854 	memset(&totals, 0, sizeof(totals));
2855 	totals.sched = sched;
2856 
2857 	if (sched->idle_hist) {
2858 		printf("\nIdle-time summary\n");
2859 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2860 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2861 	} else if (sched->show_state) {
2862 		printf("\nWait-time summary\n");
2863 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2864 		printf("   run-time      sleep      iowait     preempt       delay\n");
2865 	} else {
2866 		printf("\nRuntime summary\n");
2867 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2868 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2869 	}
2870 	printf("%*s            (count)  ", comm_width, "");
2871 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2872 	       sched->show_state ? "(msec)" : "%");
2873 	printf("%.117s\n", graph_dotted_line);
2874 
2875 	machine__for_each_thread(m, show_thread_runtime, &totals);
2876 	task_count = totals.task_count;
2877 	if (!task_count)
2878 		printf("<no still running tasks>\n");
2879 
2880 	printf("\nTerminated tasks:\n");
2881 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2882 	if (task_count == totals.task_count)
2883 		printf("<no terminated tasks>\n");
2884 
2885 	/* CPU idle stats not tracked when samples were skipped */
2886 	if (sched->skipped_samples && !sched->idle_hist)
2887 		return;
2888 
2889 	printf("\nIdle stats:\n");
2890 	for (i = 0; i < idle_max_cpu; ++i) {
2891 		if (cpu_list && !test_bit(i, cpu_bitmap))
2892 			continue;
2893 
2894 		t = idle_threads[i];
2895 		if (!t)
2896 			continue;
2897 
2898 		r = thread__priv(t);
2899 		if (r && r->run_stats.n) {
2900 			totals.sched_count += r->run_stats.n;
2901 			printf("    CPU %2d idle for ", i);
2902 			print_sched_time(r->total_run_time, 6);
2903 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2904 		} else
2905 			printf("    CPU %2d idle entire time window\n", i);
2906 	}
2907 
2908 	if (sched->idle_hist && sched->show_callchain) {
2909 		callchain_param.mode  = CHAIN_FOLDED;
2910 		callchain_param.value = CCVAL_PERIOD;
2911 
2912 		callchain_register_param(&callchain_param);
2913 
2914 		printf("\nIdle stats by callchain:\n");
2915 		for (i = 0; i < idle_max_cpu; ++i) {
2916 			struct idle_thread_runtime *itr;
2917 
2918 			t = idle_threads[i];
2919 			if (!t)
2920 				continue;
2921 
2922 			itr = thread__priv(t);
2923 			if (itr == NULL)
2924 				continue;
2925 
2926 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2927 					     0, &callchain_param);
2928 
2929 			printf("  CPU %2d:", i);
2930 			print_sched_time(itr->tr.total_run_time, 6);
2931 			printf(" msec\n");
2932 			timehist_print_idlehist_callchain(&itr->sorted_root);
2933 			printf("\n");
2934 		}
2935 	}
2936 
2937 	printf("\n"
2938 	       "    Total number of unique tasks: %" PRIu64 "\n"
2939 	       "Total number of context switches: %" PRIu64 "\n",
2940 	       totals.task_count, totals.sched_count);
2941 
2942 	printf("           Total run time (msec): ");
2943 	print_sched_time(totals.total_run_time, 2);
2944 	printf("\n");
2945 
2946 	printf("    Total scheduling time (msec): ");
2947 	print_sched_time(hist_time, 2);
2948 	printf(" (x %d)\n", sched->max_cpu.cpu);
2949 }
2950 
2951 typedef int (*sched_handler)(struct perf_tool *tool,
2952 			  union perf_event *event,
2953 			  struct evsel *evsel,
2954 			  struct perf_sample *sample,
2955 			  struct machine *machine);
2956 
2957 static int perf_timehist__process_sample(struct perf_tool *tool,
2958 					 union perf_event *event,
2959 					 struct perf_sample *sample,
2960 					 struct evsel *evsel,
2961 					 struct machine *machine)
2962 {
2963 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2964 	int err = 0;
2965 	struct perf_cpu this_cpu = {
2966 		.cpu = sample->cpu,
2967 	};
2968 
2969 	if (this_cpu.cpu > sched->max_cpu.cpu)
2970 		sched->max_cpu = this_cpu;
2971 
2972 	if (evsel->handler != NULL) {
2973 		sched_handler f = evsel->handler;
2974 
2975 		err = f(tool, event, evsel, sample, machine);
2976 	}
2977 
2978 	return err;
2979 }
2980 
2981 static int timehist_check_attr(struct perf_sched *sched,
2982 			       struct evlist *evlist)
2983 {
2984 	struct evsel *evsel;
2985 	struct evsel_runtime *er;
2986 
2987 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2988 		er = evsel__get_runtime(evsel);
2989 		if (er == NULL) {
2990 			pr_err("Failed to allocate memory for evsel runtime data\n");
2991 			return -1;
2992 		}
2993 
2994 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2995 			pr_info("Samples do not have callchains.\n");
2996 			sched->show_callchain = 0;
2997 			symbol_conf.use_callchain = 0;
2998 		}
2999 	}
3000 
3001 	return 0;
3002 }
3003 
3004 static int perf_sched__timehist(struct perf_sched *sched)
3005 {
3006 	struct evsel_str_handler handlers[] = {
3007 		{ "sched:sched_switch",       timehist_sched_switch_event, },
3008 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
3009 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
3010 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
3011 	};
3012 	const struct evsel_str_handler migrate_handlers[] = {
3013 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
3014 	};
3015 	struct perf_data data = {
3016 		.path  = input_name,
3017 		.mode  = PERF_DATA_MODE_READ,
3018 		.force = sched->force,
3019 	};
3020 
3021 	struct perf_session *session;
3022 	struct evlist *evlist;
3023 	int err = -1;
3024 
3025 	/*
3026 	 * event handlers for timehist option
3027 	 */
3028 	sched->tool.sample	 = perf_timehist__process_sample;
3029 	sched->tool.mmap	 = perf_event__process_mmap;
3030 	sched->tool.comm	 = perf_event__process_comm;
3031 	sched->tool.exit	 = perf_event__process_exit;
3032 	sched->tool.fork	 = perf_event__process_fork;
3033 	sched->tool.lost	 = process_lost;
3034 	sched->tool.attr	 = perf_event__process_attr;
3035 	sched->tool.tracing_data = perf_event__process_tracing_data;
3036 	sched->tool.build_id	 = perf_event__process_build_id;
3037 
3038 	sched->tool.ordered_events = true;
3039 	sched->tool.ordering_requires_timestamps = true;
3040 
3041 	symbol_conf.use_callchain = sched->show_callchain;
3042 
3043 	session = perf_session__new(&data, &sched->tool);
3044 	if (IS_ERR(session))
3045 		return PTR_ERR(session);
3046 
3047 	if (cpu_list) {
3048 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3049 		if (err < 0)
3050 			goto out;
3051 	}
3052 
3053 	evlist = session->evlist;
3054 
3055 	symbol__init(&session->header.env);
3056 
3057 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3058 		pr_err("Invalid time string\n");
3059 		return -EINVAL;
3060 	}
3061 
3062 	if (timehist_check_attr(sched, evlist) != 0)
3063 		goto out;
3064 
3065 	setup_pager();
3066 
3067 	/* prefer sched_waking if it is captured */
3068 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3069 		handlers[1].handler = timehist_sched_wakeup_ignore;
3070 
3071 	/* setup per-evsel handlers */
3072 	if (perf_session__set_tracepoints_handlers(session, handlers))
3073 		goto out;
3074 
3075 	/* sched_switch event at a minimum needs to exist */
3076 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3077 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3078 		goto out;
3079 	}
3080 
3081 	if (sched->show_migrations &&
3082 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3083 		goto out;
3084 
3085 	/* pre-allocate struct for per-CPU idle stats */
3086 	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3087 	if (sched->max_cpu.cpu == 0)
3088 		sched->max_cpu.cpu = 4;
3089 	if (init_idle_threads(sched->max_cpu.cpu))
3090 		goto out;
3091 
3092 	/* summary_only implies summary option, but don't overwrite summary if set */
3093 	if (sched->summary_only)
3094 		sched->summary = sched->summary_only;
3095 
3096 	if (!sched->summary_only)
3097 		timehist_header(sched);
3098 
3099 	err = perf_session__process_events(session);
3100 	if (err) {
3101 		pr_err("Failed to process events, error %d", err);
3102 		goto out;
3103 	}
3104 
3105 	sched->nr_events      = evlist->stats.nr_events[0];
3106 	sched->nr_lost_events = evlist->stats.total_lost;
3107 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3108 
3109 	if (sched->summary)
3110 		timehist_print_summary(sched, session);
3111 
3112 out:
3113 	free_idle_threads();
3114 	perf_session__delete(session);
3115 
3116 	return err;
3117 }
3118 
3119 
3120 static void print_bad_events(struct perf_sched *sched)
3121 {
3122 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3123 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3124 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3125 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3126 	}
3127 	if (sched->nr_lost_events && sched->nr_events) {
3128 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3129 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3130 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3131 	}
3132 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3133 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3134 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3135 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3136 		if (sched->nr_lost_events)
3137 			printf(" (due to lost events?)");
3138 		printf("\n");
3139 	}
3140 }
3141 
3142 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3143 {
3144 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3145 	struct work_atoms *this;
3146 	const char *comm = thread__comm_str(data->thread), *this_comm;
3147 	bool leftmost = true;
3148 
3149 	while (*new) {
3150 		int cmp;
3151 
3152 		this = container_of(*new, struct work_atoms, node);
3153 		parent = *new;
3154 
3155 		this_comm = thread__comm_str(this->thread);
3156 		cmp = strcmp(comm, this_comm);
3157 		if (cmp > 0) {
3158 			new = &((*new)->rb_left);
3159 		} else if (cmp < 0) {
3160 			new = &((*new)->rb_right);
3161 			leftmost = false;
3162 		} else {
3163 			this->num_merged++;
3164 			this->total_runtime += data->total_runtime;
3165 			this->nb_atoms += data->nb_atoms;
3166 			this->total_lat += data->total_lat;
3167 			list_splice(&data->work_list, &this->work_list);
3168 			if (this->max_lat < data->max_lat) {
3169 				this->max_lat = data->max_lat;
3170 				this->max_lat_start = data->max_lat_start;
3171 				this->max_lat_end = data->max_lat_end;
3172 			}
3173 			zfree(&data);
3174 			return;
3175 		}
3176 	}
3177 
3178 	data->num_merged++;
3179 	rb_link_node(&data->node, parent, new);
3180 	rb_insert_color_cached(&data->node, root, leftmost);
3181 }
3182 
3183 static void perf_sched__merge_lat(struct perf_sched *sched)
3184 {
3185 	struct work_atoms *data;
3186 	struct rb_node *node;
3187 
3188 	if (sched->skip_merge)
3189 		return;
3190 
3191 	while ((node = rb_first_cached(&sched->atom_root))) {
3192 		rb_erase_cached(node, &sched->atom_root);
3193 		data = rb_entry(node, struct work_atoms, node);
3194 		__merge_work_atoms(&sched->merged_atom_root, data);
3195 	}
3196 }
3197 
3198 static int perf_sched__lat(struct perf_sched *sched)
3199 {
3200 	struct rb_node *next;
3201 
3202 	setup_pager();
3203 
3204 	if (perf_sched__read_events(sched))
3205 		return -1;
3206 
3207 	perf_sched__merge_lat(sched);
3208 	perf_sched__sort_lat(sched);
3209 
3210 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3211 	printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3212 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3213 
3214 	next = rb_first_cached(&sched->sorted_atom_root);
3215 
3216 	while (next) {
3217 		struct work_atoms *work_list;
3218 
3219 		work_list = rb_entry(next, struct work_atoms, node);
3220 		output_lat_thread(sched, work_list);
3221 		next = rb_next(next);
3222 		thread__zput(work_list->thread);
3223 	}
3224 
3225 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3226 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3227 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3228 
3229 	printf(" ---------------------------------------------------\n");
3230 
3231 	print_bad_events(sched);
3232 	printf("\n");
3233 
3234 	return 0;
3235 }
3236 
3237 static int setup_map_cpus(struct perf_sched *sched)
3238 {
3239 	struct perf_cpu_map *map;
3240 
3241 	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3242 
3243 	if (sched->map.comp) {
3244 		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3245 		if (!sched->map.comp_cpus)
3246 			return -1;
3247 	}
3248 
3249 	if (!sched->map.cpus_str)
3250 		return 0;
3251 
3252 	map = perf_cpu_map__new(sched->map.cpus_str);
3253 	if (!map) {
3254 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3255 		return -1;
3256 	}
3257 
3258 	sched->map.cpus = map;
3259 	return 0;
3260 }
3261 
3262 static int setup_color_pids(struct perf_sched *sched)
3263 {
3264 	struct perf_thread_map *map;
3265 
3266 	if (!sched->map.color_pids_str)
3267 		return 0;
3268 
3269 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3270 	if (!map) {
3271 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3272 		return -1;
3273 	}
3274 
3275 	sched->map.color_pids = map;
3276 	return 0;
3277 }
3278 
3279 static int setup_color_cpus(struct perf_sched *sched)
3280 {
3281 	struct perf_cpu_map *map;
3282 
3283 	if (!sched->map.color_cpus_str)
3284 		return 0;
3285 
3286 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3287 	if (!map) {
3288 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3289 		return -1;
3290 	}
3291 
3292 	sched->map.color_cpus = map;
3293 	return 0;
3294 }
3295 
3296 static int perf_sched__map(struct perf_sched *sched)
3297 {
3298 	if (setup_map_cpus(sched))
3299 		return -1;
3300 
3301 	if (setup_color_pids(sched))
3302 		return -1;
3303 
3304 	if (setup_color_cpus(sched))
3305 		return -1;
3306 
3307 	setup_pager();
3308 	if (perf_sched__read_events(sched))
3309 		return -1;
3310 	print_bad_events(sched);
3311 	return 0;
3312 }
3313 
3314 static int perf_sched__replay(struct perf_sched *sched)
3315 {
3316 	unsigned long i;
3317 
3318 	calibrate_run_measurement_overhead(sched);
3319 	calibrate_sleep_measurement_overhead(sched);
3320 
3321 	test_calibrations(sched);
3322 
3323 	if (perf_sched__read_events(sched))
3324 		return -1;
3325 
3326 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3327 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3328 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3329 
3330 	if (sched->targetless_wakeups)
3331 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3332 	if (sched->multitarget_wakeups)
3333 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3334 	if (sched->nr_run_events_optimized)
3335 		printf("run atoms optimized: %ld\n",
3336 			sched->nr_run_events_optimized);
3337 
3338 	print_task_traces(sched);
3339 	add_cross_task_wakeups(sched);
3340 
3341 	sched->thread_funcs_exit = false;
3342 	create_tasks(sched);
3343 	printf("------------------------------------------------------------\n");
3344 	for (i = 0; i < sched->replay_repeat; i++)
3345 		run_one_test(sched);
3346 
3347 	sched->thread_funcs_exit = true;
3348 	destroy_tasks(sched);
3349 	return 0;
3350 }
3351 
3352 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3353 			  const char * const usage_msg[])
3354 {
3355 	char *tmp, *tok, *str = strdup(sched->sort_order);
3356 
3357 	for (tok = strtok_r(str, ", ", &tmp);
3358 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3359 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3360 			usage_with_options_msg(usage_msg, options,
3361 					"Unknown --sort key: `%s'", tok);
3362 		}
3363 	}
3364 
3365 	free(str);
3366 
3367 	sort_dimension__add("pid", &sched->cmp_pid);
3368 }
3369 
3370 static bool schedstat_events_exposed(void)
3371 {
3372 	/*
3373 	 * Select "sched:sched_stat_wait" event to check
3374 	 * whether schedstat tracepoints are exposed.
3375 	 */
3376 	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3377 		false : true;
3378 }
3379 
3380 static int __cmd_record(int argc, const char **argv)
3381 {
3382 	unsigned int rec_argc, i, j;
3383 	char **rec_argv;
3384 	const char **rec_argv_copy;
3385 	const char * const record_args[] = {
3386 		"record",
3387 		"-a",
3388 		"-R",
3389 		"-m", "1024",
3390 		"-c", "1",
3391 		"-e", "sched:sched_switch",
3392 		"-e", "sched:sched_stat_runtime",
3393 		"-e", "sched:sched_process_fork",
3394 		"-e", "sched:sched_wakeup_new",
3395 		"-e", "sched:sched_migrate_task",
3396 	};
3397 
3398 	/*
3399 	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3400 	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3401 	 * to prevent "perf sched record" execution failure, determine
3402 	 * whether to record schedstat events according to actual situation.
3403 	 */
3404 	const char * const schedstat_args[] = {
3405 		"-e", "sched:sched_stat_wait",
3406 		"-e", "sched:sched_stat_sleep",
3407 		"-e", "sched:sched_stat_iowait",
3408 	};
3409 	unsigned int schedstat_argc = schedstat_events_exposed() ?
3410 		ARRAY_SIZE(schedstat_args) : 0;
3411 
3412 	struct tep_event *waking_event;
3413 	int ret;
3414 
3415 	/*
3416 	 * +2 for either "-e", "sched:sched_wakeup" or
3417 	 * "-e", "sched:sched_waking"
3418 	 */
3419 	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3420 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3421 	if (rec_argv == NULL)
3422 		return -ENOMEM;
3423 	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3424 	if (rec_argv_copy == NULL) {
3425 		free(rec_argv);
3426 		return -ENOMEM;
3427 	}
3428 
3429 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3430 		rec_argv[i] = strdup(record_args[i]);
3431 
3432 	rec_argv[i++] = strdup("-e");
3433 	waking_event = trace_event__tp_format("sched", "sched_waking");
3434 	if (!IS_ERR(waking_event))
3435 		rec_argv[i++] = strdup("sched:sched_waking");
3436 	else
3437 		rec_argv[i++] = strdup("sched:sched_wakeup");
3438 
3439 	for (j = 0; j < schedstat_argc; j++)
3440 		rec_argv[i++] = strdup(schedstat_args[j]);
3441 
3442 	for (j = 1; j < (unsigned int)argc; j++, i++)
3443 		rec_argv[i] = strdup(argv[j]);
3444 
3445 	BUG_ON(i != rec_argc);
3446 
3447 	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3448 	ret = cmd_record(rec_argc, rec_argv_copy);
3449 
3450 	for (i = 0; i < rec_argc; i++)
3451 		free(rec_argv[i]);
3452 	free(rec_argv);
3453 	free(rec_argv_copy);
3454 
3455 	return ret;
3456 }
3457 
3458 int cmd_sched(int argc, const char **argv)
3459 {
3460 	static const char default_sort_order[] = "avg, max, switch, runtime";
3461 	struct perf_sched sched = {
3462 		.tool = {
3463 			.sample		 = perf_sched__process_tracepoint_sample,
3464 			.comm		 = perf_sched__process_comm,
3465 			.namespaces	 = perf_event__process_namespaces,
3466 			.lost		 = perf_event__process_lost,
3467 			.fork		 = perf_sched__process_fork_event,
3468 			.ordered_events = true,
3469 		},
3470 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3471 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3472 		.sort_order	      = default_sort_order,
3473 		.replay_repeat	      = 10,
3474 		.profile_cpu	      = -1,
3475 		.next_shortname1      = 'A',
3476 		.next_shortname2      = '0',
3477 		.skip_merge           = 0,
3478 		.show_callchain	      = 1,
3479 		.max_stack            = 5,
3480 	};
3481 	const struct option sched_options[] = {
3482 	OPT_STRING('i', "input", &input_name, "file",
3483 		    "input file name"),
3484 	OPT_INCR('v', "verbose", &verbose,
3485 		    "be more verbose (show symbol address, etc)"),
3486 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3487 		    "dump raw trace in ASCII"),
3488 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3489 	OPT_END()
3490 	};
3491 	const struct option latency_options[] = {
3492 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3493 		   "sort by key(s): runtime, switch, avg, max"),
3494 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3495 		    "CPU to profile on"),
3496 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3497 		    "latency stats per pid instead of per comm"),
3498 	OPT_PARENT(sched_options)
3499 	};
3500 	const struct option replay_options[] = {
3501 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3502 		     "repeat the workload replay N times (-1: infinite)"),
3503 	OPT_PARENT(sched_options)
3504 	};
3505 	const struct option map_options[] = {
3506 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3507 		    "map output in compact mode"),
3508 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3509 		   "highlight given pids in map"),
3510 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3511                     "highlight given CPUs in map"),
3512 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3513                     "display given CPUs in map"),
3514 	OPT_PARENT(sched_options)
3515 	};
3516 	const struct option timehist_options[] = {
3517 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3518 		   "file", "vmlinux pathname"),
3519 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3520 		   "file", "kallsyms pathname"),
3521 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3522 		    "Display call chains if present (default on)"),
3523 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3524 		   "Maximum number of functions to display backtrace."),
3525 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3526 		    "Look for files with symbols relative to this directory"),
3527 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3528 		    "Show only syscall summary with statistics"),
3529 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3530 		    "Show all syscalls and summary with statistics"),
3531 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3532 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3533 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3534 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3535 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3536 	OPT_STRING(0, "time", &sched.time_str, "str",
3537 		   "Time span for analysis (start,stop)"),
3538 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3539 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3540 		   "analyze events only for given process id(s)"),
3541 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3542 		   "analyze events only for given thread id(s)"),
3543 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3544 	OPT_PARENT(sched_options)
3545 	};
3546 
3547 	const char * const latency_usage[] = {
3548 		"perf sched latency [<options>]",
3549 		NULL
3550 	};
3551 	const char * const replay_usage[] = {
3552 		"perf sched replay [<options>]",
3553 		NULL
3554 	};
3555 	const char * const map_usage[] = {
3556 		"perf sched map [<options>]",
3557 		NULL
3558 	};
3559 	const char * const timehist_usage[] = {
3560 		"perf sched timehist [<options>]",
3561 		NULL
3562 	};
3563 	const char *const sched_subcommands[] = { "record", "latency", "map",
3564 						  "replay", "script",
3565 						  "timehist", NULL };
3566 	const char *sched_usage[] = {
3567 		NULL,
3568 		NULL
3569 	};
3570 	struct trace_sched_handler lat_ops  = {
3571 		.wakeup_event	    = latency_wakeup_event,
3572 		.switch_event	    = latency_switch_event,
3573 		.runtime_event	    = latency_runtime_event,
3574 		.migrate_task_event = latency_migrate_task_event,
3575 	};
3576 	struct trace_sched_handler map_ops  = {
3577 		.switch_event	    = map_switch_event,
3578 	};
3579 	struct trace_sched_handler replay_ops  = {
3580 		.wakeup_event	    = replay_wakeup_event,
3581 		.switch_event	    = replay_switch_event,
3582 		.fork_event	    = replay_fork_event,
3583 	};
3584 	unsigned int i;
3585 	int ret = 0;
3586 
3587 	mutex_init(&sched.start_work_mutex);
3588 	mutex_init(&sched.work_done_wait_mutex);
3589 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3590 		sched.curr_pid[i] = -1;
3591 
3592 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3593 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3594 	if (!argc)
3595 		usage_with_options(sched_usage, sched_options);
3596 
3597 	/*
3598 	 * Aliased to 'perf script' for now:
3599 	 */
3600 	if (!strcmp(argv[0], "script")) {
3601 		ret = cmd_script(argc, argv);
3602 	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3603 		ret = __cmd_record(argc, argv);
3604 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3605 		sched.tp_handler = &lat_ops;
3606 		if (argc > 1) {
3607 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3608 			if (argc)
3609 				usage_with_options(latency_usage, latency_options);
3610 		}
3611 		setup_sorting(&sched, latency_options, latency_usage);
3612 		ret = perf_sched__lat(&sched);
3613 	} else if (!strcmp(argv[0], "map")) {
3614 		if (argc) {
3615 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3616 			if (argc)
3617 				usage_with_options(map_usage, map_options);
3618 		}
3619 		sched.tp_handler = &map_ops;
3620 		setup_sorting(&sched, latency_options, latency_usage);
3621 		ret = perf_sched__map(&sched);
3622 	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3623 		sched.tp_handler = &replay_ops;
3624 		if (argc) {
3625 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3626 			if (argc)
3627 				usage_with_options(replay_usage, replay_options);
3628 		}
3629 		ret = perf_sched__replay(&sched);
3630 	} else if (!strcmp(argv[0], "timehist")) {
3631 		if (argc) {
3632 			argc = parse_options(argc, argv, timehist_options,
3633 					     timehist_usage, 0);
3634 			if (argc)
3635 				usage_with_options(timehist_usage, timehist_options);
3636 		}
3637 		if ((sched.show_wakeups || sched.show_next) &&
3638 		    sched.summary_only) {
3639 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3640 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3641 			if (sched.show_wakeups)
3642 				parse_options_usage(NULL, timehist_options, "w", true);
3643 			if (sched.show_next)
3644 				parse_options_usage(NULL, timehist_options, "n", true);
3645 			ret = -EINVAL;
3646 			goto out;
3647 		}
3648 		ret = symbol__validate_sym_arguments();
3649 		if (ret)
3650 			goto out;
3651 
3652 		ret = perf_sched__timehist(&sched);
3653 	} else {
3654 		usage_with_options(sched_usage, sched_options);
3655 	}
3656 
3657 out:
3658 	mutex_destroy(&sched.start_work_mutex);
3659 	mutex_destroy(&sched.work_done_wait_mutex);
3660 
3661 	return ret;
3662 }
3663