xref: /openbmc/linux/tools/perf/builtin-sched.c (revision 3e30a927)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 #include "perf-sys.h"
5 
6 #include "util/cpumap.h"
7 #include "util/evlist.h"
8 #include "util/evsel.h"
9 #include "util/evsel_fprintf.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
22 
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
26 
27 #include "util/debug.h"
28 #include "util/event.h"
29 
30 #include <linux/kernel.h>
31 #include <linux/log2.h>
32 #include <linux/zalloc.h>
33 #include <sys/prctl.h>
34 #include <sys/resource.h>
35 #include <inttypes.h>
36 
37 #include <errno.h>
38 #include <semaphore.h>
39 #include <pthread.h>
40 #include <math.h>
41 #include <api/fs/fs.h>
42 #include <perf/cpumap.h>
43 #include <linux/time64.h>
44 #include <linux/err.h>
45 
46 #include <linux/ctype.h>
47 
48 #define PR_SET_NAME		15               /* Set process name */
49 #define MAX_CPUS		4096
50 #define COMM_LEN		20
51 #define SYM_LEN			129
52 #define MAX_PID			1024000
53 
54 static const char *cpu_list;
55 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
56 
57 struct sched_atom;
58 
59 struct task_desc {
60 	unsigned long		nr;
61 	unsigned long		pid;
62 	char			comm[COMM_LEN];
63 
64 	unsigned long		nr_events;
65 	unsigned long		curr_event;
66 	struct sched_atom	**atoms;
67 
68 	pthread_t		thread;
69 	sem_t			sleep_sem;
70 
71 	sem_t			ready_for_work;
72 	sem_t			work_done_sem;
73 
74 	u64			cpu_usage;
75 };
76 
77 enum sched_event_type {
78 	SCHED_EVENT_RUN,
79 	SCHED_EVENT_SLEEP,
80 	SCHED_EVENT_WAKEUP,
81 	SCHED_EVENT_MIGRATION,
82 };
83 
84 struct sched_atom {
85 	enum sched_event_type	type;
86 	int			specific_wait;
87 	u64			timestamp;
88 	u64			duration;
89 	unsigned long		nr;
90 	sem_t			*wait_sem;
91 	struct task_desc	*wakee;
92 };
93 
94 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
95 
96 /* task state bitmask, copied from include/linux/sched.h */
97 #define TASK_RUNNING		0
98 #define TASK_INTERRUPTIBLE	1
99 #define TASK_UNINTERRUPTIBLE	2
100 #define __TASK_STOPPED		4
101 #define __TASK_TRACED		8
102 /* in tsk->exit_state */
103 #define EXIT_DEAD		16
104 #define EXIT_ZOMBIE		32
105 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
106 /* in tsk->state again */
107 #define TASK_DEAD		64
108 #define TASK_WAKEKILL		128
109 #define TASK_WAKING		256
110 #define TASK_PARKED		512
111 
112 enum thread_state {
113 	THREAD_SLEEPING = 0,
114 	THREAD_WAIT_CPU,
115 	THREAD_SCHED_IN,
116 	THREAD_IGNORE
117 };
118 
119 struct work_atom {
120 	struct list_head	list;
121 	enum thread_state	state;
122 	u64			sched_out_time;
123 	u64			wake_up_time;
124 	u64			sched_in_time;
125 	u64			runtime;
126 };
127 
128 struct work_atoms {
129 	struct list_head	work_list;
130 	struct thread		*thread;
131 	struct rb_node		node;
132 	u64			max_lat;
133 	u64			max_lat_at;
134 	u64			total_lat;
135 	u64			nb_atoms;
136 	u64			total_runtime;
137 	int			num_merged;
138 };
139 
140 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
141 
142 struct perf_sched;
143 
144 struct trace_sched_handler {
145 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
146 			    struct perf_sample *sample, struct machine *machine);
147 
148 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
149 			     struct perf_sample *sample, struct machine *machine);
150 
151 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
152 			    struct perf_sample *sample, struct machine *machine);
153 
154 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
155 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
156 			  struct machine *machine);
157 
158 	int (*migrate_task_event)(struct perf_sched *sched,
159 				  struct evsel *evsel,
160 				  struct perf_sample *sample,
161 				  struct machine *machine);
162 };
163 
164 #define COLOR_PIDS PERF_COLOR_BLUE
165 #define COLOR_CPUS PERF_COLOR_BG_RED
166 
167 struct perf_sched_map {
168 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
169 	int			*comp_cpus;
170 	bool			 comp;
171 	struct perf_thread_map *color_pids;
172 	const char		*color_pids_str;
173 	struct perf_cpu_map	*color_cpus;
174 	const char		*color_cpus_str;
175 	struct perf_cpu_map	*cpus;
176 	const char		*cpus_str;
177 };
178 
179 struct perf_sched {
180 	struct perf_tool tool;
181 	const char	 *sort_order;
182 	unsigned long	 nr_tasks;
183 	struct task_desc **pid_to_task;
184 	struct task_desc **tasks;
185 	const struct trace_sched_handler *tp_handler;
186 	pthread_mutex_t	 start_work_mutex;
187 	pthread_mutex_t	 work_done_wait_mutex;
188 	int		 profile_cpu;
189 /*
190  * Track the current task - that way we can know whether there's any
191  * weird events, such as a task being switched away that is not current.
192  */
193 	int		 max_cpu;
194 	u32		 curr_pid[MAX_CPUS];
195 	struct thread	 *curr_thread[MAX_CPUS];
196 	char		 next_shortname1;
197 	char		 next_shortname2;
198 	unsigned int	 replay_repeat;
199 	unsigned long	 nr_run_events;
200 	unsigned long	 nr_sleep_events;
201 	unsigned long	 nr_wakeup_events;
202 	unsigned long	 nr_sleep_corrections;
203 	unsigned long	 nr_run_events_optimized;
204 	unsigned long	 targetless_wakeups;
205 	unsigned long	 multitarget_wakeups;
206 	unsigned long	 nr_runs;
207 	unsigned long	 nr_timestamps;
208 	unsigned long	 nr_unordered_timestamps;
209 	unsigned long	 nr_context_switch_bugs;
210 	unsigned long	 nr_events;
211 	unsigned long	 nr_lost_chunks;
212 	unsigned long	 nr_lost_events;
213 	u64		 run_measurement_overhead;
214 	u64		 sleep_measurement_overhead;
215 	u64		 start_time;
216 	u64		 cpu_usage;
217 	u64		 runavg_cpu_usage;
218 	u64		 parent_cpu_usage;
219 	u64		 runavg_parent_cpu_usage;
220 	u64		 sum_runtime;
221 	u64		 sum_fluct;
222 	u64		 run_avg;
223 	u64		 all_runtime;
224 	u64		 all_count;
225 	u64		 cpu_last_switched[MAX_CPUS];
226 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
227 	struct list_head sort_list, cmp_pid;
228 	bool force;
229 	bool skip_merge;
230 	struct perf_sched_map map;
231 
232 	/* options for timehist command */
233 	bool		summary;
234 	bool		summary_only;
235 	bool		idle_hist;
236 	bool		show_callchain;
237 	unsigned int	max_stack;
238 	bool		show_cpu_visual;
239 	bool		show_wakeups;
240 	bool		show_next;
241 	bool		show_migrations;
242 	bool		show_state;
243 	u64		skipped_samples;
244 	const char	*time_str;
245 	struct perf_time_interval ptime;
246 	struct perf_time_interval hist_time;
247 };
248 
249 /* per thread run time data */
250 struct thread_runtime {
251 	u64 last_time;      /* time of previous sched in/out event */
252 	u64 dt_run;         /* run time */
253 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
254 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
255 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
256 	u64 dt_delay;       /* time between wakeup and sched-in */
257 	u64 ready_to_run;   /* time of wakeup */
258 
259 	struct stats run_stats;
260 	u64 total_run_time;
261 	u64 total_sleep_time;
262 	u64 total_iowait_time;
263 	u64 total_preempt_time;
264 	u64 total_delay_time;
265 
266 	int last_state;
267 
268 	char shortname[3];
269 	bool comm_changed;
270 
271 	u64 migrations;
272 };
273 
274 /* per event run time data */
275 struct evsel_runtime {
276 	u64 *last_time; /* time this event was last seen per cpu */
277 	u32 ncpu;       /* highest cpu slot allocated */
278 };
279 
280 /* per cpu idle time data */
281 struct idle_thread_runtime {
282 	struct thread_runtime	tr;
283 	struct thread		*last_thread;
284 	struct rb_root_cached	sorted_root;
285 	struct callchain_root	callchain;
286 	struct callchain_cursor	cursor;
287 };
288 
289 /* track idle times per cpu */
290 static struct thread **idle_threads;
291 static int idle_max_cpu;
292 static char idle_comm[] = "<idle>";
293 
294 static u64 get_nsecs(void)
295 {
296 	struct timespec ts;
297 
298 	clock_gettime(CLOCK_MONOTONIC, &ts);
299 
300 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
301 }
302 
303 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
304 {
305 	u64 T0 = get_nsecs(), T1;
306 
307 	do {
308 		T1 = get_nsecs();
309 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
310 }
311 
312 static void sleep_nsecs(u64 nsecs)
313 {
314 	struct timespec ts;
315 
316 	ts.tv_nsec = nsecs % 999999999;
317 	ts.tv_sec = nsecs / 999999999;
318 
319 	nanosleep(&ts, NULL);
320 }
321 
322 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
323 {
324 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
325 	int i;
326 
327 	for (i = 0; i < 10; i++) {
328 		T0 = get_nsecs();
329 		burn_nsecs(sched, 0);
330 		T1 = get_nsecs();
331 		delta = T1-T0;
332 		min_delta = min(min_delta, delta);
333 	}
334 	sched->run_measurement_overhead = min_delta;
335 
336 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
337 }
338 
339 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
340 {
341 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
342 	int i;
343 
344 	for (i = 0; i < 10; i++) {
345 		T0 = get_nsecs();
346 		sleep_nsecs(10000);
347 		T1 = get_nsecs();
348 		delta = T1-T0;
349 		min_delta = min(min_delta, delta);
350 	}
351 	min_delta -= 10000;
352 	sched->sleep_measurement_overhead = min_delta;
353 
354 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
355 }
356 
357 static struct sched_atom *
358 get_new_event(struct task_desc *task, u64 timestamp)
359 {
360 	struct sched_atom *event = zalloc(sizeof(*event));
361 	unsigned long idx = task->nr_events;
362 	size_t size;
363 
364 	event->timestamp = timestamp;
365 	event->nr = idx;
366 
367 	task->nr_events++;
368 	size = sizeof(struct sched_atom *) * task->nr_events;
369 	task->atoms = realloc(task->atoms, size);
370 	BUG_ON(!task->atoms);
371 
372 	task->atoms[idx] = event;
373 
374 	return event;
375 }
376 
377 static struct sched_atom *last_event(struct task_desc *task)
378 {
379 	if (!task->nr_events)
380 		return NULL;
381 
382 	return task->atoms[task->nr_events - 1];
383 }
384 
385 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
386 				u64 timestamp, u64 duration)
387 {
388 	struct sched_atom *event, *curr_event = last_event(task);
389 
390 	/*
391 	 * optimize an existing RUN event by merging this one
392 	 * to it:
393 	 */
394 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
395 		sched->nr_run_events_optimized++;
396 		curr_event->duration += duration;
397 		return;
398 	}
399 
400 	event = get_new_event(task, timestamp);
401 
402 	event->type = SCHED_EVENT_RUN;
403 	event->duration = duration;
404 
405 	sched->nr_run_events++;
406 }
407 
408 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
409 				   u64 timestamp, struct task_desc *wakee)
410 {
411 	struct sched_atom *event, *wakee_event;
412 
413 	event = get_new_event(task, timestamp);
414 	event->type = SCHED_EVENT_WAKEUP;
415 	event->wakee = wakee;
416 
417 	wakee_event = last_event(wakee);
418 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
419 		sched->targetless_wakeups++;
420 		return;
421 	}
422 	if (wakee_event->wait_sem) {
423 		sched->multitarget_wakeups++;
424 		return;
425 	}
426 
427 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
428 	sem_init(wakee_event->wait_sem, 0, 0);
429 	wakee_event->specific_wait = 1;
430 	event->wait_sem = wakee_event->wait_sem;
431 
432 	sched->nr_wakeup_events++;
433 }
434 
435 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
436 				  u64 timestamp, u64 task_state __maybe_unused)
437 {
438 	struct sched_atom *event = get_new_event(task, timestamp);
439 
440 	event->type = SCHED_EVENT_SLEEP;
441 
442 	sched->nr_sleep_events++;
443 }
444 
445 static struct task_desc *register_pid(struct perf_sched *sched,
446 				      unsigned long pid, const char *comm)
447 {
448 	struct task_desc *task;
449 	static int pid_max;
450 
451 	if (sched->pid_to_task == NULL) {
452 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
453 			pid_max = MAX_PID;
454 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
455 	}
456 	if (pid >= (unsigned long)pid_max) {
457 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
458 			sizeof(struct task_desc *))) == NULL);
459 		while (pid >= (unsigned long)pid_max)
460 			sched->pid_to_task[pid_max++] = NULL;
461 	}
462 
463 	task = sched->pid_to_task[pid];
464 
465 	if (task)
466 		return task;
467 
468 	task = zalloc(sizeof(*task));
469 	task->pid = pid;
470 	task->nr = sched->nr_tasks;
471 	strcpy(task->comm, comm);
472 	/*
473 	 * every task starts in sleeping state - this gets ignored
474 	 * if there's no wakeup pointing to this sleep state:
475 	 */
476 	add_sched_event_sleep(sched, task, 0, 0);
477 
478 	sched->pid_to_task[pid] = task;
479 	sched->nr_tasks++;
480 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
481 	BUG_ON(!sched->tasks);
482 	sched->tasks[task->nr] = task;
483 
484 	if (verbose > 0)
485 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
486 
487 	return task;
488 }
489 
490 
491 static void print_task_traces(struct perf_sched *sched)
492 {
493 	struct task_desc *task;
494 	unsigned long i;
495 
496 	for (i = 0; i < sched->nr_tasks; i++) {
497 		task = sched->tasks[i];
498 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
499 			task->nr, task->comm, task->pid, task->nr_events);
500 	}
501 }
502 
503 static void add_cross_task_wakeups(struct perf_sched *sched)
504 {
505 	struct task_desc *task1, *task2;
506 	unsigned long i, j;
507 
508 	for (i = 0; i < sched->nr_tasks; i++) {
509 		task1 = sched->tasks[i];
510 		j = i + 1;
511 		if (j == sched->nr_tasks)
512 			j = 0;
513 		task2 = sched->tasks[j];
514 		add_sched_event_wakeup(sched, task1, 0, task2);
515 	}
516 }
517 
518 static void perf_sched__process_event(struct perf_sched *sched,
519 				      struct sched_atom *atom)
520 {
521 	int ret = 0;
522 
523 	switch (atom->type) {
524 		case SCHED_EVENT_RUN:
525 			burn_nsecs(sched, atom->duration);
526 			break;
527 		case SCHED_EVENT_SLEEP:
528 			if (atom->wait_sem)
529 				ret = sem_wait(atom->wait_sem);
530 			BUG_ON(ret);
531 			break;
532 		case SCHED_EVENT_WAKEUP:
533 			if (atom->wait_sem)
534 				ret = sem_post(atom->wait_sem);
535 			BUG_ON(ret);
536 			break;
537 		case SCHED_EVENT_MIGRATION:
538 			break;
539 		default:
540 			BUG_ON(1);
541 	}
542 }
543 
544 static u64 get_cpu_usage_nsec_parent(void)
545 {
546 	struct rusage ru;
547 	u64 sum;
548 	int err;
549 
550 	err = getrusage(RUSAGE_SELF, &ru);
551 	BUG_ON(err);
552 
553 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
554 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
555 
556 	return sum;
557 }
558 
559 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
560 {
561 	struct perf_event_attr attr;
562 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
563 	int fd;
564 	struct rlimit limit;
565 	bool need_privilege = false;
566 
567 	memset(&attr, 0, sizeof(attr));
568 
569 	attr.type = PERF_TYPE_SOFTWARE;
570 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
571 
572 force_again:
573 	fd = sys_perf_event_open(&attr, 0, -1, -1,
574 				 perf_event_open_cloexec_flag());
575 
576 	if (fd < 0) {
577 		if (errno == EMFILE) {
578 			if (sched->force) {
579 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
580 				limit.rlim_cur += sched->nr_tasks - cur_task;
581 				if (limit.rlim_cur > limit.rlim_max) {
582 					limit.rlim_max = limit.rlim_cur;
583 					need_privilege = true;
584 				}
585 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
586 					if (need_privilege && errno == EPERM)
587 						strcpy(info, "Need privilege\n");
588 				} else
589 					goto force_again;
590 			} else
591 				strcpy(info, "Have a try with -f option\n");
592 		}
593 		pr_err("Error: sys_perf_event_open() syscall returned "
594 		       "with %d (%s)\n%s", fd,
595 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
596 		exit(EXIT_FAILURE);
597 	}
598 	return fd;
599 }
600 
601 static u64 get_cpu_usage_nsec_self(int fd)
602 {
603 	u64 runtime;
604 	int ret;
605 
606 	ret = read(fd, &runtime, sizeof(runtime));
607 	BUG_ON(ret != sizeof(runtime));
608 
609 	return runtime;
610 }
611 
612 struct sched_thread_parms {
613 	struct task_desc  *task;
614 	struct perf_sched *sched;
615 	int fd;
616 };
617 
618 static void *thread_func(void *ctx)
619 {
620 	struct sched_thread_parms *parms = ctx;
621 	struct task_desc *this_task = parms->task;
622 	struct perf_sched *sched = parms->sched;
623 	u64 cpu_usage_0, cpu_usage_1;
624 	unsigned long i, ret;
625 	char comm2[22];
626 	int fd = parms->fd;
627 
628 	zfree(&parms);
629 
630 	sprintf(comm2, ":%s", this_task->comm);
631 	prctl(PR_SET_NAME, comm2);
632 	if (fd < 0)
633 		return NULL;
634 again:
635 	ret = sem_post(&this_task->ready_for_work);
636 	BUG_ON(ret);
637 	ret = pthread_mutex_lock(&sched->start_work_mutex);
638 	BUG_ON(ret);
639 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
640 	BUG_ON(ret);
641 
642 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
643 
644 	for (i = 0; i < this_task->nr_events; i++) {
645 		this_task->curr_event = i;
646 		perf_sched__process_event(sched, this_task->atoms[i]);
647 	}
648 
649 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
650 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
651 	ret = sem_post(&this_task->work_done_sem);
652 	BUG_ON(ret);
653 
654 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
655 	BUG_ON(ret);
656 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
657 	BUG_ON(ret);
658 
659 	goto again;
660 }
661 
662 static void create_tasks(struct perf_sched *sched)
663 {
664 	struct task_desc *task;
665 	pthread_attr_t attr;
666 	unsigned long i;
667 	int err;
668 
669 	err = pthread_attr_init(&attr);
670 	BUG_ON(err);
671 	err = pthread_attr_setstacksize(&attr,
672 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
673 	BUG_ON(err);
674 	err = pthread_mutex_lock(&sched->start_work_mutex);
675 	BUG_ON(err);
676 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
677 	BUG_ON(err);
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 wait_for_tasks(struct perf_sched *sched)
694 {
695 	u64 cpu_usage_0, cpu_usage_1;
696 	struct task_desc *task;
697 	unsigned long i, ret;
698 
699 	sched->start_time = get_nsecs();
700 	sched->cpu_usage = 0;
701 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
702 
703 	for (i = 0; i < sched->nr_tasks; i++) {
704 		task = sched->tasks[i];
705 		ret = sem_wait(&task->ready_for_work);
706 		BUG_ON(ret);
707 		sem_init(&task->ready_for_work, 0, 0);
708 	}
709 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
710 	BUG_ON(ret);
711 
712 	cpu_usage_0 = get_cpu_usage_nsec_parent();
713 
714 	pthread_mutex_unlock(&sched->start_work_mutex);
715 
716 	for (i = 0; i < sched->nr_tasks; i++) {
717 		task = sched->tasks[i];
718 		ret = sem_wait(&task->work_done_sem);
719 		BUG_ON(ret);
720 		sem_init(&task->work_done_sem, 0, 0);
721 		sched->cpu_usage += task->cpu_usage;
722 		task->cpu_usage = 0;
723 	}
724 
725 	cpu_usage_1 = get_cpu_usage_nsec_parent();
726 	if (!sched->runavg_cpu_usage)
727 		sched->runavg_cpu_usage = sched->cpu_usage;
728 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
729 
730 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
731 	if (!sched->runavg_parent_cpu_usage)
732 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
733 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
734 					 sched->parent_cpu_usage)/sched->replay_repeat;
735 
736 	ret = pthread_mutex_lock(&sched->start_work_mutex);
737 	BUG_ON(ret);
738 
739 	for (i = 0; i < sched->nr_tasks; i++) {
740 		task = sched->tasks[i];
741 		sem_init(&task->sleep_sem, 0, 0);
742 		task->curr_event = 0;
743 	}
744 }
745 
746 static void run_one_test(struct perf_sched *sched)
747 {
748 	u64 T0, T1, delta, avg_delta, fluct;
749 
750 	T0 = get_nsecs();
751 	wait_for_tasks(sched);
752 	T1 = get_nsecs();
753 
754 	delta = T1 - T0;
755 	sched->sum_runtime += delta;
756 	sched->nr_runs++;
757 
758 	avg_delta = sched->sum_runtime / sched->nr_runs;
759 	if (delta < avg_delta)
760 		fluct = avg_delta - delta;
761 	else
762 		fluct = delta - avg_delta;
763 	sched->sum_fluct += fluct;
764 	if (!sched->run_avg)
765 		sched->run_avg = delta;
766 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
767 
768 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
769 
770 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
771 
772 	printf("cpu: %0.2f / %0.2f",
773 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
774 
775 #if 0
776 	/*
777 	 * rusage statistics done by the parent, these are less
778 	 * accurate than the sched->sum_exec_runtime based statistics:
779 	 */
780 	printf(" [%0.2f / %0.2f]",
781 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
782 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
783 #endif
784 
785 	printf("\n");
786 
787 	if (sched->nr_sleep_corrections)
788 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
789 	sched->nr_sleep_corrections = 0;
790 }
791 
792 static void test_calibrations(struct perf_sched *sched)
793 {
794 	u64 T0, T1;
795 
796 	T0 = get_nsecs();
797 	burn_nsecs(sched, NSEC_PER_MSEC);
798 	T1 = get_nsecs();
799 
800 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
801 
802 	T0 = get_nsecs();
803 	sleep_nsecs(NSEC_PER_MSEC);
804 	T1 = get_nsecs();
805 
806 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
807 }
808 
809 static int
810 replay_wakeup_event(struct perf_sched *sched,
811 		    struct evsel *evsel, struct perf_sample *sample,
812 		    struct machine *machine __maybe_unused)
813 {
814 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
815 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
816 	struct task_desc *waker, *wakee;
817 
818 	if (verbose > 0) {
819 		printf("sched_wakeup event %p\n", evsel);
820 
821 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
822 	}
823 
824 	waker = register_pid(sched, sample->tid, "<unknown>");
825 	wakee = register_pid(sched, pid, comm);
826 
827 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
828 	return 0;
829 }
830 
831 static int replay_switch_event(struct perf_sched *sched,
832 			       struct evsel *evsel,
833 			       struct perf_sample *sample,
834 			       struct machine *machine __maybe_unused)
835 {
836 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
837 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
838 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
839 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
840 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
841 	struct task_desc *prev, __maybe_unused *next;
842 	u64 timestamp0, timestamp = sample->time;
843 	int cpu = sample->cpu;
844 	s64 delta;
845 
846 	if (verbose > 0)
847 		printf("sched_switch event %p\n", evsel);
848 
849 	if (cpu >= MAX_CPUS || cpu < 0)
850 		return 0;
851 
852 	timestamp0 = sched->cpu_last_switched[cpu];
853 	if (timestamp0)
854 		delta = timestamp - timestamp0;
855 	else
856 		delta = 0;
857 
858 	if (delta < 0) {
859 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
860 		return -1;
861 	}
862 
863 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
864 		 prev_comm, prev_pid, next_comm, next_pid, delta);
865 
866 	prev = register_pid(sched, prev_pid, prev_comm);
867 	next = register_pid(sched, next_pid, next_comm);
868 
869 	sched->cpu_last_switched[cpu] = timestamp;
870 
871 	add_sched_event_run(sched, prev, timestamp, delta);
872 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
873 
874 	return 0;
875 }
876 
877 static int replay_fork_event(struct perf_sched *sched,
878 			     union perf_event *event,
879 			     struct machine *machine)
880 {
881 	struct thread *child, *parent;
882 
883 	child = machine__findnew_thread(machine, event->fork.pid,
884 					event->fork.tid);
885 	parent = machine__findnew_thread(machine, event->fork.ppid,
886 					 event->fork.ptid);
887 
888 	if (child == NULL || parent == NULL) {
889 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
890 				 child, parent);
891 		goto out_put;
892 	}
893 
894 	if (verbose > 0) {
895 		printf("fork event\n");
896 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
897 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
898 	}
899 
900 	register_pid(sched, parent->tid, thread__comm_str(parent));
901 	register_pid(sched, child->tid, thread__comm_str(child));
902 out_put:
903 	thread__put(child);
904 	thread__put(parent);
905 	return 0;
906 }
907 
908 struct sort_dimension {
909 	const char		*name;
910 	sort_fn_t		cmp;
911 	struct list_head	list;
912 };
913 
914 /*
915  * handle runtime stats saved per thread
916  */
917 static struct thread_runtime *thread__init_runtime(struct thread *thread)
918 {
919 	struct thread_runtime *r;
920 
921 	r = zalloc(sizeof(struct thread_runtime));
922 	if (!r)
923 		return NULL;
924 
925 	init_stats(&r->run_stats);
926 	thread__set_priv(thread, r);
927 
928 	return r;
929 }
930 
931 static struct thread_runtime *thread__get_runtime(struct thread *thread)
932 {
933 	struct thread_runtime *tr;
934 
935 	tr = thread__priv(thread);
936 	if (tr == NULL) {
937 		tr = thread__init_runtime(thread);
938 		if (tr == NULL)
939 			pr_debug("Failed to malloc memory for runtime data.\n");
940 	}
941 
942 	return tr;
943 }
944 
945 static int
946 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
947 {
948 	struct sort_dimension *sort;
949 	int ret = 0;
950 
951 	BUG_ON(list_empty(list));
952 
953 	list_for_each_entry(sort, list, list) {
954 		ret = sort->cmp(l, r);
955 		if (ret)
956 			return ret;
957 	}
958 
959 	return ret;
960 }
961 
962 static struct work_atoms *
963 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
964 			 struct list_head *sort_list)
965 {
966 	struct rb_node *node = root->rb_root.rb_node;
967 	struct work_atoms key = { .thread = thread };
968 
969 	while (node) {
970 		struct work_atoms *atoms;
971 		int cmp;
972 
973 		atoms = container_of(node, struct work_atoms, node);
974 
975 		cmp = thread_lat_cmp(sort_list, &key, atoms);
976 		if (cmp > 0)
977 			node = node->rb_left;
978 		else if (cmp < 0)
979 			node = node->rb_right;
980 		else {
981 			BUG_ON(thread != atoms->thread);
982 			return atoms;
983 		}
984 	}
985 	return NULL;
986 }
987 
988 static void
989 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
990 			 struct list_head *sort_list)
991 {
992 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
993 	bool leftmost = true;
994 
995 	while (*new) {
996 		struct work_atoms *this;
997 		int cmp;
998 
999 		this = container_of(*new, struct work_atoms, node);
1000 		parent = *new;
1001 
1002 		cmp = thread_lat_cmp(sort_list, data, this);
1003 
1004 		if (cmp > 0)
1005 			new = &((*new)->rb_left);
1006 		else {
1007 			new = &((*new)->rb_right);
1008 			leftmost = false;
1009 		}
1010 	}
1011 
1012 	rb_link_node(&data->node, parent, new);
1013 	rb_insert_color_cached(&data->node, root, leftmost);
1014 }
1015 
1016 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1017 {
1018 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1019 	if (!atoms) {
1020 		pr_err("No memory at %s\n", __func__);
1021 		return -1;
1022 	}
1023 
1024 	atoms->thread = thread__get(thread);
1025 	INIT_LIST_HEAD(&atoms->work_list);
1026 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1027 	return 0;
1028 }
1029 
1030 static char sched_out_state(u64 prev_state)
1031 {
1032 	const char *str = TASK_STATE_TO_CHAR_STR;
1033 
1034 	return str[prev_state];
1035 }
1036 
1037 static int
1038 add_sched_out_event(struct work_atoms *atoms,
1039 		    char run_state,
1040 		    u64 timestamp)
1041 {
1042 	struct work_atom *atom = zalloc(sizeof(*atom));
1043 	if (!atom) {
1044 		pr_err("Non memory at %s", __func__);
1045 		return -1;
1046 	}
1047 
1048 	atom->sched_out_time = timestamp;
1049 
1050 	if (run_state == 'R') {
1051 		atom->state = THREAD_WAIT_CPU;
1052 		atom->wake_up_time = atom->sched_out_time;
1053 	}
1054 
1055 	list_add_tail(&atom->list, &atoms->work_list);
1056 	return 0;
1057 }
1058 
1059 static void
1060 add_runtime_event(struct work_atoms *atoms, u64 delta,
1061 		  u64 timestamp __maybe_unused)
1062 {
1063 	struct work_atom *atom;
1064 
1065 	BUG_ON(list_empty(&atoms->work_list));
1066 
1067 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1068 
1069 	atom->runtime += delta;
1070 	atoms->total_runtime += delta;
1071 }
1072 
1073 static void
1074 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1075 {
1076 	struct work_atom *atom;
1077 	u64 delta;
1078 
1079 	if (list_empty(&atoms->work_list))
1080 		return;
1081 
1082 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1083 
1084 	if (atom->state != THREAD_WAIT_CPU)
1085 		return;
1086 
1087 	if (timestamp < atom->wake_up_time) {
1088 		atom->state = THREAD_IGNORE;
1089 		return;
1090 	}
1091 
1092 	atom->state = THREAD_SCHED_IN;
1093 	atom->sched_in_time = timestamp;
1094 
1095 	delta = atom->sched_in_time - atom->wake_up_time;
1096 	atoms->total_lat += delta;
1097 	if (delta > atoms->max_lat) {
1098 		atoms->max_lat = delta;
1099 		atoms->max_lat_at = timestamp;
1100 	}
1101 	atoms->nb_atoms++;
1102 }
1103 
1104 static int latency_switch_event(struct perf_sched *sched,
1105 				struct evsel *evsel,
1106 				struct perf_sample *sample,
1107 				struct machine *machine)
1108 {
1109 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1110 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1111 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1112 	struct work_atoms *out_events, *in_events;
1113 	struct thread *sched_out, *sched_in;
1114 	u64 timestamp0, timestamp = sample->time;
1115 	int cpu = sample->cpu, err = -1;
1116 	s64 delta;
1117 
1118 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1119 
1120 	timestamp0 = sched->cpu_last_switched[cpu];
1121 	sched->cpu_last_switched[cpu] = timestamp;
1122 	if (timestamp0)
1123 		delta = timestamp - timestamp0;
1124 	else
1125 		delta = 0;
1126 
1127 	if (delta < 0) {
1128 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1129 		return -1;
1130 	}
1131 
1132 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1133 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1134 	if (sched_out == NULL || sched_in == NULL)
1135 		goto out_put;
1136 
1137 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1138 	if (!out_events) {
1139 		if (thread_atoms_insert(sched, sched_out))
1140 			goto out_put;
1141 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1142 		if (!out_events) {
1143 			pr_err("out-event: Internal tree error");
1144 			goto out_put;
1145 		}
1146 	}
1147 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1148 		return -1;
1149 
1150 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1151 	if (!in_events) {
1152 		if (thread_atoms_insert(sched, sched_in))
1153 			goto out_put;
1154 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1155 		if (!in_events) {
1156 			pr_err("in-event: Internal tree error");
1157 			goto out_put;
1158 		}
1159 		/*
1160 		 * Take came in we have not heard about yet,
1161 		 * add in an initial atom in runnable state:
1162 		 */
1163 		if (add_sched_out_event(in_events, 'R', timestamp))
1164 			goto out_put;
1165 	}
1166 	add_sched_in_event(in_events, timestamp);
1167 	err = 0;
1168 out_put:
1169 	thread__put(sched_out);
1170 	thread__put(sched_in);
1171 	return err;
1172 }
1173 
1174 static int latency_runtime_event(struct perf_sched *sched,
1175 				 struct evsel *evsel,
1176 				 struct perf_sample *sample,
1177 				 struct machine *machine)
1178 {
1179 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1180 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1181 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1182 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1183 	u64 timestamp = sample->time;
1184 	int cpu = sample->cpu, err = -1;
1185 
1186 	if (thread == NULL)
1187 		return -1;
1188 
1189 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1190 	if (!atoms) {
1191 		if (thread_atoms_insert(sched, thread))
1192 			goto out_put;
1193 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1194 		if (!atoms) {
1195 			pr_err("in-event: Internal tree error");
1196 			goto out_put;
1197 		}
1198 		if (add_sched_out_event(atoms, 'R', timestamp))
1199 			goto out_put;
1200 	}
1201 
1202 	add_runtime_event(atoms, runtime, timestamp);
1203 	err = 0;
1204 out_put:
1205 	thread__put(thread);
1206 	return err;
1207 }
1208 
1209 static int latency_wakeup_event(struct perf_sched *sched,
1210 				struct evsel *evsel,
1211 				struct perf_sample *sample,
1212 				struct machine *machine)
1213 {
1214 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1215 	struct work_atoms *atoms;
1216 	struct work_atom *atom;
1217 	struct thread *wakee;
1218 	u64 timestamp = sample->time;
1219 	int err = -1;
1220 
1221 	wakee = machine__findnew_thread(machine, -1, pid);
1222 	if (wakee == NULL)
1223 		return -1;
1224 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1225 	if (!atoms) {
1226 		if (thread_atoms_insert(sched, wakee))
1227 			goto out_put;
1228 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1229 		if (!atoms) {
1230 			pr_err("wakeup-event: Internal tree error");
1231 			goto out_put;
1232 		}
1233 		if (add_sched_out_event(atoms, 'S', timestamp))
1234 			goto out_put;
1235 	}
1236 
1237 	BUG_ON(list_empty(&atoms->work_list));
1238 
1239 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1240 
1241 	/*
1242 	 * As we do not guarantee the wakeup event happens when
1243 	 * task is out of run queue, also may happen when task is
1244 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1245 	 * then we should not set the ->wake_up_time when wake up a
1246 	 * task which is on run queue.
1247 	 *
1248 	 * You WILL be missing events if you've recorded only
1249 	 * one CPU, or are only looking at only one, so don't
1250 	 * skip in this case.
1251 	 */
1252 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1253 		goto out_ok;
1254 
1255 	sched->nr_timestamps++;
1256 	if (atom->sched_out_time > timestamp) {
1257 		sched->nr_unordered_timestamps++;
1258 		goto out_ok;
1259 	}
1260 
1261 	atom->state = THREAD_WAIT_CPU;
1262 	atom->wake_up_time = timestamp;
1263 out_ok:
1264 	err = 0;
1265 out_put:
1266 	thread__put(wakee);
1267 	return err;
1268 }
1269 
1270 static int latency_migrate_task_event(struct perf_sched *sched,
1271 				      struct evsel *evsel,
1272 				      struct perf_sample *sample,
1273 				      struct machine *machine)
1274 {
1275 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1276 	u64 timestamp = sample->time;
1277 	struct work_atoms *atoms;
1278 	struct work_atom *atom;
1279 	struct thread *migrant;
1280 	int err = -1;
1281 
1282 	/*
1283 	 * Only need to worry about migration when profiling one CPU.
1284 	 */
1285 	if (sched->profile_cpu == -1)
1286 		return 0;
1287 
1288 	migrant = machine__findnew_thread(machine, -1, pid);
1289 	if (migrant == NULL)
1290 		return -1;
1291 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1292 	if (!atoms) {
1293 		if (thread_atoms_insert(sched, migrant))
1294 			goto out_put;
1295 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1296 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1297 		if (!atoms) {
1298 			pr_err("migration-event: Internal tree error");
1299 			goto out_put;
1300 		}
1301 		if (add_sched_out_event(atoms, 'R', timestamp))
1302 			goto out_put;
1303 	}
1304 
1305 	BUG_ON(list_empty(&atoms->work_list));
1306 
1307 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1308 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1309 
1310 	sched->nr_timestamps++;
1311 
1312 	if (atom->sched_out_time > timestamp)
1313 		sched->nr_unordered_timestamps++;
1314 	err = 0;
1315 out_put:
1316 	thread__put(migrant);
1317 	return err;
1318 }
1319 
1320 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1321 {
1322 	int i;
1323 	int ret;
1324 	u64 avg;
1325 	char max_lat_at[32];
1326 
1327 	if (!work_list->nb_atoms)
1328 		return;
1329 	/*
1330 	 * Ignore idle threads:
1331 	 */
1332 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1333 		return;
1334 
1335 	sched->all_runtime += work_list->total_runtime;
1336 	sched->all_count   += work_list->nb_atoms;
1337 
1338 	if (work_list->num_merged > 1)
1339 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1340 	else
1341 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1342 
1343 	for (i = 0; i < 24 - ret; i++)
1344 		printf(" ");
1345 
1346 	avg = work_list->total_lat / work_list->nb_atoms;
1347 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1348 
1349 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1350 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1351 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1352 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1353 		 max_lat_at);
1354 }
1355 
1356 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1357 {
1358 	if (l->thread == r->thread)
1359 		return 0;
1360 	if (l->thread->tid < r->thread->tid)
1361 		return -1;
1362 	if (l->thread->tid > r->thread->tid)
1363 		return 1;
1364 	return (int)(l->thread - r->thread);
1365 }
1366 
1367 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1368 {
1369 	u64 avgl, avgr;
1370 
1371 	if (!l->nb_atoms)
1372 		return -1;
1373 
1374 	if (!r->nb_atoms)
1375 		return 1;
1376 
1377 	avgl = l->total_lat / l->nb_atoms;
1378 	avgr = r->total_lat / r->nb_atoms;
1379 
1380 	if (avgl < avgr)
1381 		return -1;
1382 	if (avgl > avgr)
1383 		return 1;
1384 
1385 	return 0;
1386 }
1387 
1388 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1389 {
1390 	if (l->max_lat < r->max_lat)
1391 		return -1;
1392 	if (l->max_lat > r->max_lat)
1393 		return 1;
1394 
1395 	return 0;
1396 }
1397 
1398 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1399 {
1400 	if (l->nb_atoms < r->nb_atoms)
1401 		return -1;
1402 	if (l->nb_atoms > r->nb_atoms)
1403 		return 1;
1404 
1405 	return 0;
1406 }
1407 
1408 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1409 {
1410 	if (l->total_runtime < r->total_runtime)
1411 		return -1;
1412 	if (l->total_runtime > r->total_runtime)
1413 		return 1;
1414 
1415 	return 0;
1416 }
1417 
1418 static int sort_dimension__add(const char *tok, struct list_head *list)
1419 {
1420 	size_t i;
1421 	static struct sort_dimension avg_sort_dimension = {
1422 		.name = "avg",
1423 		.cmp  = avg_cmp,
1424 	};
1425 	static struct sort_dimension max_sort_dimension = {
1426 		.name = "max",
1427 		.cmp  = max_cmp,
1428 	};
1429 	static struct sort_dimension pid_sort_dimension = {
1430 		.name = "pid",
1431 		.cmp  = pid_cmp,
1432 	};
1433 	static struct sort_dimension runtime_sort_dimension = {
1434 		.name = "runtime",
1435 		.cmp  = runtime_cmp,
1436 	};
1437 	static struct sort_dimension switch_sort_dimension = {
1438 		.name = "switch",
1439 		.cmp  = switch_cmp,
1440 	};
1441 	struct sort_dimension *available_sorts[] = {
1442 		&pid_sort_dimension,
1443 		&avg_sort_dimension,
1444 		&max_sort_dimension,
1445 		&switch_sort_dimension,
1446 		&runtime_sort_dimension,
1447 	};
1448 
1449 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1450 		if (!strcmp(available_sorts[i]->name, tok)) {
1451 			list_add_tail(&available_sorts[i]->list, list);
1452 
1453 			return 0;
1454 		}
1455 	}
1456 
1457 	return -1;
1458 }
1459 
1460 static void perf_sched__sort_lat(struct perf_sched *sched)
1461 {
1462 	struct rb_node *node;
1463 	struct rb_root_cached *root = &sched->atom_root;
1464 again:
1465 	for (;;) {
1466 		struct work_atoms *data;
1467 		node = rb_first_cached(root);
1468 		if (!node)
1469 			break;
1470 
1471 		rb_erase_cached(node, root);
1472 		data = rb_entry(node, struct work_atoms, node);
1473 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1474 	}
1475 	if (root == &sched->atom_root) {
1476 		root = &sched->merged_atom_root;
1477 		goto again;
1478 	}
1479 }
1480 
1481 static int process_sched_wakeup_event(struct perf_tool *tool,
1482 				      struct evsel *evsel,
1483 				      struct perf_sample *sample,
1484 				      struct machine *machine)
1485 {
1486 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1487 
1488 	if (sched->tp_handler->wakeup_event)
1489 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1490 
1491 	return 0;
1492 }
1493 
1494 union map_priv {
1495 	void	*ptr;
1496 	bool	 color;
1497 };
1498 
1499 static bool thread__has_color(struct thread *thread)
1500 {
1501 	union map_priv priv = {
1502 		.ptr = thread__priv(thread),
1503 	};
1504 
1505 	return priv.color;
1506 }
1507 
1508 static struct thread*
1509 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1510 {
1511 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1512 	union map_priv priv = {
1513 		.color = false,
1514 	};
1515 
1516 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1517 		return thread;
1518 
1519 	if (thread_map__has(sched->map.color_pids, tid))
1520 		priv.color = true;
1521 
1522 	thread__set_priv(thread, priv.ptr);
1523 	return thread;
1524 }
1525 
1526 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1527 			    struct perf_sample *sample, struct machine *machine)
1528 {
1529 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1530 	struct thread *sched_in;
1531 	struct thread_runtime *tr;
1532 	int new_shortname;
1533 	u64 timestamp0, timestamp = sample->time;
1534 	s64 delta;
1535 	int i, this_cpu = sample->cpu;
1536 	int cpus_nr;
1537 	bool new_cpu = false;
1538 	const char *color = PERF_COLOR_NORMAL;
1539 	char stimestamp[32];
1540 
1541 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1542 
1543 	if (this_cpu > sched->max_cpu)
1544 		sched->max_cpu = this_cpu;
1545 
1546 	if (sched->map.comp) {
1547 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1548 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1549 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1550 			new_cpu = true;
1551 		}
1552 	} else
1553 		cpus_nr = sched->max_cpu;
1554 
1555 	timestamp0 = sched->cpu_last_switched[this_cpu];
1556 	sched->cpu_last_switched[this_cpu] = timestamp;
1557 	if (timestamp0)
1558 		delta = timestamp - timestamp0;
1559 	else
1560 		delta = 0;
1561 
1562 	if (delta < 0) {
1563 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1564 		return -1;
1565 	}
1566 
1567 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1568 	if (sched_in == NULL)
1569 		return -1;
1570 
1571 	tr = thread__get_runtime(sched_in);
1572 	if (tr == NULL) {
1573 		thread__put(sched_in);
1574 		return -1;
1575 	}
1576 
1577 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1578 
1579 	printf("  ");
1580 
1581 	new_shortname = 0;
1582 	if (!tr->shortname[0]) {
1583 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1584 			/*
1585 			 * Don't allocate a letter-number for swapper:0
1586 			 * as a shortname. Instead, we use '.' for it.
1587 			 */
1588 			tr->shortname[0] = '.';
1589 			tr->shortname[1] = ' ';
1590 		} else {
1591 			tr->shortname[0] = sched->next_shortname1;
1592 			tr->shortname[1] = sched->next_shortname2;
1593 
1594 			if (sched->next_shortname1 < 'Z') {
1595 				sched->next_shortname1++;
1596 			} else {
1597 				sched->next_shortname1 = 'A';
1598 				if (sched->next_shortname2 < '9')
1599 					sched->next_shortname2++;
1600 				else
1601 					sched->next_shortname2 = '0';
1602 			}
1603 		}
1604 		new_shortname = 1;
1605 	}
1606 
1607 	for (i = 0; i < cpus_nr; i++) {
1608 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1609 		struct thread *curr_thread = sched->curr_thread[cpu];
1610 		struct thread_runtime *curr_tr;
1611 		const char *pid_color = color;
1612 		const char *cpu_color = color;
1613 
1614 		if (curr_thread && thread__has_color(curr_thread))
1615 			pid_color = COLOR_PIDS;
1616 
1617 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1618 			continue;
1619 
1620 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1621 			cpu_color = COLOR_CPUS;
1622 
1623 		if (cpu != this_cpu)
1624 			color_fprintf(stdout, color, " ");
1625 		else
1626 			color_fprintf(stdout, cpu_color, "*");
1627 
1628 		if (sched->curr_thread[cpu]) {
1629 			curr_tr = thread__get_runtime(sched->curr_thread[cpu]);
1630 			if (curr_tr == NULL) {
1631 				thread__put(sched_in);
1632 				return -1;
1633 			}
1634 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1635 		} else
1636 			color_fprintf(stdout, color, "   ");
1637 	}
1638 
1639 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1640 		goto out;
1641 
1642 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1643 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1644 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1645 		const char *pid_color = color;
1646 
1647 		if (thread__has_color(sched_in))
1648 			pid_color = COLOR_PIDS;
1649 
1650 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1651 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1652 		tr->comm_changed = false;
1653 	}
1654 
1655 	if (sched->map.comp && new_cpu)
1656 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1657 
1658 out:
1659 	color_fprintf(stdout, color, "\n");
1660 
1661 	thread__put(sched_in);
1662 
1663 	return 0;
1664 }
1665 
1666 static int process_sched_switch_event(struct perf_tool *tool,
1667 				      struct evsel *evsel,
1668 				      struct perf_sample *sample,
1669 				      struct machine *machine)
1670 {
1671 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1672 	int this_cpu = sample->cpu, err = 0;
1673 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1674 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1675 
1676 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1677 		/*
1678 		 * Are we trying to switch away a PID that is
1679 		 * not current?
1680 		 */
1681 		if (sched->curr_pid[this_cpu] != prev_pid)
1682 			sched->nr_context_switch_bugs++;
1683 	}
1684 
1685 	if (sched->tp_handler->switch_event)
1686 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1687 
1688 	sched->curr_pid[this_cpu] = next_pid;
1689 	return err;
1690 }
1691 
1692 static int process_sched_runtime_event(struct perf_tool *tool,
1693 				       struct evsel *evsel,
1694 				       struct perf_sample *sample,
1695 				       struct machine *machine)
1696 {
1697 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1698 
1699 	if (sched->tp_handler->runtime_event)
1700 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1701 
1702 	return 0;
1703 }
1704 
1705 static int perf_sched__process_fork_event(struct perf_tool *tool,
1706 					  union perf_event *event,
1707 					  struct perf_sample *sample,
1708 					  struct machine *machine)
1709 {
1710 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1711 
1712 	/* run the fork event through the perf machineruy */
1713 	perf_event__process_fork(tool, event, sample, machine);
1714 
1715 	/* and then run additional processing needed for this command */
1716 	if (sched->tp_handler->fork_event)
1717 		return sched->tp_handler->fork_event(sched, event, machine);
1718 
1719 	return 0;
1720 }
1721 
1722 static int process_sched_migrate_task_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->migrate_task_event)
1730 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1731 
1732 	return 0;
1733 }
1734 
1735 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1736 				  struct evsel *evsel,
1737 				  struct perf_sample *sample,
1738 				  struct machine *machine);
1739 
1740 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1741 						 union perf_event *event __maybe_unused,
1742 						 struct perf_sample *sample,
1743 						 struct evsel *evsel,
1744 						 struct machine *machine)
1745 {
1746 	int err = 0;
1747 
1748 	if (evsel->handler != NULL) {
1749 		tracepoint_handler f = evsel->handler;
1750 		err = f(tool, evsel, sample, machine);
1751 	}
1752 
1753 	return err;
1754 }
1755 
1756 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1757 				    union perf_event *event,
1758 				    struct perf_sample *sample,
1759 				    struct machine *machine)
1760 {
1761 	struct thread *thread;
1762 	struct thread_runtime *tr;
1763 	int err;
1764 
1765 	err = perf_event__process_comm(tool, event, sample, machine);
1766 	if (err)
1767 		return err;
1768 
1769 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1770 	if (!thread) {
1771 		pr_err("Internal error: can't find thread\n");
1772 		return -1;
1773 	}
1774 
1775 	tr = thread__get_runtime(thread);
1776 	if (tr == NULL) {
1777 		thread__put(thread);
1778 		return -1;
1779 	}
1780 
1781 	tr->comm_changed = true;
1782 	thread__put(thread);
1783 
1784 	return 0;
1785 }
1786 
1787 static int perf_sched__read_events(struct perf_sched *sched)
1788 {
1789 	const struct evsel_str_handler handlers[] = {
1790 		{ "sched:sched_switch",	      process_sched_switch_event, },
1791 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1792 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1793 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1794 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1795 	};
1796 	struct perf_session *session;
1797 	struct perf_data data = {
1798 		.path  = input_name,
1799 		.mode  = PERF_DATA_MODE_READ,
1800 		.force = sched->force,
1801 	};
1802 	int rc = -1;
1803 
1804 	session = perf_session__new(&data, false, &sched->tool);
1805 	if (IS_ERR(session)) {
1806 		pr_debug("Error creating perf session");
1807 		return PTR_ERR(session);
1808 	}
1809 
1810 	symbol__init(&session->header.env);
1811 
1812 	if (perf_session__set_tracepoints_handlers(session, handlers))
1813 		goto out_delete;
1814 
1815 	if (perf_session__has_traces(session, "record -R")) {
1816 		int err = perf_session__process_events(session);
1817 		if (err) {
1818 			pr_err("Failed to process events, error %d", err);
1819 			goto out_delete;
1820 		}
1821 
1822 		sched->nr_events      = session->evlist->stats.nr_events[0];
1823 		sched->nr_lost_events = session->evlist->stats.total_lost;
1824 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1825 	}
1826 
1827 	rc = 0;
1828 out_delete:
1829 	perf_session__delete(session);
1830 	return rc;
1831 }
1832 
1833 /*
1834  * scheduling times are printed as msec.usec
1835  */
1836 static inline void print_sched_time(unsigned long long nsecs, int width)
1837 {
1838 	unsigned long msecs;
1839 	unsigned long usecs;
1840 
1841 	msecs  = nsecs / NSEC_PER_MSEC;
1842 	nsecs -= msecs * NSEC_PER_MSEC;
1843 	usecs  = nsecs / NSEC_PER_USEC;
1844 	printf("%*lu.%03lu ", width, msecs, usecs);
1845 }
1846 
1847 /*
1848  * returns runtime data for event, allocating memory for it the
1849  * first time it is used.
1850  */
1851 static struct evsel_runtime *perf_evsel__get_runtime(struct evsel *evsel)
1852 {
1853 	struct evsel_runtime *r = evsel->priv;
1854 
1855 	if (r == NULL) {
1856 		r = zalloc(sizeof(struct evsel_runtime));
1857 		evsel->priv = r;
1858 	}
1859 
1860 	return r;
1861 }
1862 
1863 /*
1864  * save last time event was seen per cpu
1865  */
1866 static void perf_evsel__save_time(struct evsel *evsel,
1867 				  u64 timestamp, u32 cpu)
1868 {
1869 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1870 
1871 	if (r == NULL)
1872 		return;
1873 
1874 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1875 		int i, n = __roundup_pow_of_two(cpu+1);
1876 		void *p = r->last_time;
1877 
1878 		p = realloc(r->last_time, n * sizeof(u64));
1879 		if (!p)
1880 			return;
1881 
1882 		r->last_time = p;
1883 		for (i = r->ncpu; i < n; ++i)
1884 			r->last_time[i] = (u64) 0;
1885 
1886 		r->ncpu = n;
1887 	}
1888 
1889 	r->last_time[cpu] = timestamp;
1890 }
1891 
1892 /* returns last time this event was seen on the given cpu */
1893 static u64 perf_evsel__get_time(struct evsel *evsel, u32 cpu)
1894 {
1895 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1896 
1897 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1898 		return 0;
1899 
1900 	return r->last_time[cpu];
1901 }
1902 
1903 static int comm_width = 30;
1904 
1905 static char *timehist_get_commstr(struct thread *thread)
1906 {
1907 	static char str[32];
1908 	const char *comm = thread__comm_str(thread);
1909 	pid_t tid = thread->tid;
1910 	pid_t pid = thread->pid_;
1911 	int n;
1912 
1913 	if (pid == 0)
1914 		n = scnprintf(str, sizeof(str), "%s", comm);
1915 
1916 	else if (tid != pid)
1917 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1918 
1919 	else
1920 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1921 
1922 	if (n > comm_width)
1923 		comm_width = n;
1924 
1925 	return str;
1926 }
1927 
1928 static void timehist_header(struct perf_sched *sched)
1929 {
1930 	u32 ncpus = sched->max_cpu + 1;
1931 	u32 i, j;
1932 
1933 	printf("%15s %6s ", "time", "cpu");
1934 
1935 	if (sched->show_cpu_visual) {
1936 		printf(" ");
1937 		for (i = 0, j = 0; i < ncpus; ++i) {
1938 			printf("%x", j++);
1939 			if (j > 15)
1940 				j = 0;
1941 		}
1942 		printf(" ");
1943 	}
1944 
1945 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1946 		"task name", "wait time", "sch delay", "run time");
1947 
1948 	if (sched->show_state)
1949 		printf("  %s", "state");
1950 
1951 	printf("\n");
1952 
1953 	/*
1954 	 * units row
1955 	 */
1956 	printf("%15s %-6s ", "", "");
1957 
1958 	if (sched->show_cpu_visual)
1959 		printf(" %*s ", ncpus, "");
1960 
1961 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1962 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1963 
1964 	if (sched->show_state)
1965 		printf("  %5s", "");
1966 
1967 	printf("\n");
1968 
1969 	/*
1970 	 * separator
1971 	 */
1972 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1973 
1974 	if (sched->show_cpu_visual)
1975 		printf(" %.*s ", ncpus, graph_dotted_line);
1976 
1977 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1978 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1979 		graph_dotted_line);
1980 
1981 	if (sched->show_state)
1982 		printf("  %.5s", graph_dotted_line);
1983 
1984 	printf("\n");
1985 }
1986 
1987 static char task_state_char(struct thread *thread, int state)
1988 {
1989 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
1990 	unsigned bit = state ? ffs(state) : 0;
1991 
1992 	/* 'I' for idle */
1993 	if (thread->tid == 0)
1994 		return 'I';
1995 
1996 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
1997 }
1998 
1999 static void timehist_print_sample(struct perf_sched *sched,
2000 				  struct evsel *evsel,
2001 				  struct perf_sample *sample,
2002 				  struct addr_location *al,
2003 				  struct thread *thread,
2004 				  u64 t, int state)
2005 {
2006 	struct thread_runtime *tr = thread__priv(thread);
2007 	const char *next_comm = perf_evsel__strval(evsel, sample, "next_comm");
2008 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
2009 	u32 max_cpus = sched->max_cpu + 1;
2010 	char tstr[64];
2011 	char nstr[30];
2012 	u64 wait_time;
2013 
2014 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2015 		return;
2016 
2017 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2018 	printf("%15s [%04d] ", tstr, sample->cpu);
2019 
2020 	if (sched->show_cpu_visual) {
2021 		u32 i;
2022 		char c;
2023 
2024 		printf(" ");
2025 		for (i = 0; i < max_cpus; ++i) {
2026 			/* flag idle times with 'i'; others are sched events */
2027 			if (i == sample->cpu)
2028 				c = (thread->tid == 0) ? 'i' : 's';
2029 			else
2030 				c = ' ';
2031 			printf("%c", c);
2032 		}
2033 		printf(" ");
2034 	}
2035 
2036 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2037 
2038 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2039 	print_sched_time(wait_time, 6);
2040 
2041 	print_sched_time(tr->dt_delay, 6);
2042 	print_sched_time(tr->dt_run, 6);
2043 
2044 	if (sched->show_state)
2045 		printf(" %5c ", task_state_char(thread, state));
2046 
2047 	if (sched->show_next) {
2048 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2049 		printf(" %-*s", comm_width, nstr);
2050 	}
2051 
2052 	if (sched->show_wakeups && !sched->show_next)
2053 		printf("  %-*s", comm_width, "");
2054 
2055 	if (thread->tid == 0)
2056 		goto out;
2057 
2058 	if (sched->show_callchain)
2059 		printf("  ");
2060 
2061 	sample__fprintf_sym(sample, al, 0,
2062 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2063 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2064 			    EVSEL__PRINT_SKIP_IGNORED,
2065 			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2066 
2067 out:
2068 	printf("\n");
2069 }
2070 
2071 /*
2072  * Explanation of delta-time stats:
2073  *
2074  *            t = time of current schedule out event
2075  *        tprev = time of previous sched out event
2076  *                also time of schedule-in event for current task
2077  *    last_time = time of last sched change event for current task
2078  *                (i.e, time process was last scheduled out)
2079  * ready_to_run = time of wakeup for current task
2080  *
2081  * -----|------------|------------|------------|------
2082  *    last         ready        tprev          t
2083  *    time         to run
2084  *
2085  *      |-------- dt_wait --------|
2086  *                   |- dt_delay -|-- dt_run --|
2087  *
2088  *   dt_run = run time of current task
2089  *  dt_wait = time between last schedule out event for task and tprev
2090  *            represents time spent off the cpu
2091  * dt_delay = time between wakeup and schedule-in of task
2092  */
2093 
2094 static void timehist_update_runtime_stats(struct thread_runtime *r,
2095 					 u64 t, u64 tprev)
2096 {
2097 	r->dt_delay   = 0;
2098 	r->dt_sleep   = 0;
2099 	r->dt_iowait  = 0;
2100 	r->dt_preempt = 0;
2101 	r->dt_run     = 0;
2102 
2103 	if (tprev) {
2104 		r->dt_run = t - tprev;
2105 		if (r->ready_to_run) {
2106 			if (r->ready_to_run > tprev)
2107 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2108 			else
2109 				r->dt_delay = tprev - r->ready_to_run;
2110 		}
2111 
2112 		if (r->last_time > tprev)
2113 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2114 		else if (r->last_time) {
2115 			u64 dt_wait = tprev - r->last_time;
2116 
2117 			if (r->last_state == TASK_RUNNING)
2118 				r->dt_preempt = dt_wait;
2119 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2120 				r->dt_iowait = dt_wait;
2121 			else
2122 				r->dt_sleep = dt_wait;
2123 		}
2124 	}
2125 
2126 	update_stats(&r->run_stats, r->dt_run);
2127 
2128 	r->total_run_time     += r->dt_run;
2129 	r->total_delay_time   += r->dt_delay;
2130 	r->total_sleep_time   += r->dt_sleep;
2131 	r->total_iowait_time  += r->dt_iowait;
2132 	r->total_preempt_time += r->dt_preempt;
2133 }
2134 
2135 static bool is_idle_sample(struct perf_sample *sample,
2136 			   struct evsel *evsel)
2137 {
2138 	/* pid 0 == swapper == idle task */
2139 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
2140 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
2141 
2142 	return sample->pid == 0;
2143 }
2144 
2145 static void save_task_callchain(struct perf_sched *sched,
2146 				struct perf_sample *sample,
2147 				struct evsel *evsel,
2148 				struct machine *machine)
2149 {
2150 	struct callchain_cursor *cursor = &callchain_cursor;
2151 	struct thread *thread;
2152 
2153 	/* want main thread for process - has maps */
2154 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2155 	if (thread == NULL) {
2156 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2157 		return;
2158 	}
2159 
2160 	if (!sched->show_callchain || sample->callchain == NULL)
2161 		return;
2162 
2163 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2164 				      NULL, NULL, sched->max_stack + 2) != 0) {
2165 		if (verbose > 0)
2166 			pr_err("Failed to resolve callchain. Skipping\n");
2167 
2168 		return;
2169 	}
2170 
2171 	callchain_cursor_commit(cursor);
2172 
2173 	while (true) {
2174 		struct callchain_cursor_node *node;
2175 		struct symbol *sym;
2176 
2177 		node = callchain_cursor_current(cursor);
2178 		if (node == NULL)
2179 			break;
2180 
2181 		sym = node->ms.sym;
2182 		if (sym) {
2183 			if (!strcmp(sym->name, "schedule") ||
2184 			    !strcmp(sym->name, "__schedule") ||
2185 			    !strcmp(sym->name, "preempt_schedule"))
2186 				sym->ignore = 1;
2187 		}
2188 
2189 		callchain_cursor_advance(cursor);
2190 	}
2191 }
2192 
2193 static int init_idle_thread(struct thread *thread)
2194 {
2195 	struct idle_thread_runtime *itr;
2196 
2197 	thread__set_comm(thread, idle_comm, 0);
2198 
2199 	itr = zalloc(sizeof(*itr));
2200 	if (itr == NULL)
2201 		return -ENOMEM;
2202 
2203 	init_stats(&itr->tr.run_stats);
2204 	callchain_init(&itr->callchain);
2205 	callchain_cursor_reset(&itr->cursor);
2206 	thread__set_priv(thread, itr);
2207 
2208 	return 0;
2209 }
2210 
2211 /*
2212  * Track idle stats per cpu by maintaining a local thread
2213  * struct for the idle task on each cpu.
2214  */
2215 static int init_idle_threads(int ncpu)
2216 {
2217 	int i, ret;
2218 
2219 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2220 	if (!idle_threads)
2221 		return -ENOMEM;
2222 
2223 	idle_max_cpu = ncpu;
2224 
2225 	/* allocate the actual thread struct if needed */
2226 	for (i = 0; i < ncpu; ++i) {
2227 		idle_threads[i] = thread__new(0, 0);
2228 		if (idle_threads[i] == NULL)
2229 			return -ENOMEM;
2230 
2231 		ret = init_idle_thread(idle_threads[i]);
2232 		if (ret < 0)
2233 			return ret;
2234 	}
2235 
2236 	return 0;
2237 }
2238 
2239 static void free_idle_threads(void)
2240 {
2241 	int i;
2242 
2243 	if (idle_threads == NULL)
2244 		return;
2245 
2246 	for (i = 0; i < idle_max_cpu; ++i) {
2247 		if ((idle_threads[i]))
2248 			thread__delete(idle_threads[i]);
2249 	}
2250 
2251 	free(idle_threads);
2252 }
2253 
2254 static struct thread *get_idle_thread(int cpu)
2255 {
2256 	/*
2257 	 * expand/allocate array of pointers to local thread
2258 	 * structs if needed
2259 	 */
2260 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2261 		int i, j = __roundup_pow_of_two(cpu+1);
2262 		void *p;
2263 
2264 		p = realloc(idle_threads, j * sizeof(struct thread *));
2265 		if (!p)
2266 			return NULL;
2267 
2268 		idle_threads = (struct thread **) p;
2269 		for (i = idle_max_cpu; i < j; ++i)
2270 			idle_threads[i] = NULL;
2271 
2272 		idle_max_cpu = j;
2273 	}
2274 
2275 	/* allocate a new thread struct if needed */
2276 	if (idle_threads[cpu] == NULL) {
2277 		idle_threads[cpu] = thread__new(0, 0);
2278 
2279 		if (idle_threads[cpu]) {
2280 			if (init_idle_thread(idle_threads[cpu]) < 0)
2281 				return NULL;
2282 		}
2283 	}
2284 
2285 	return idle_threads[cpu];
2286 }
2287 
2288 static void save_idle_callchain(struct perf_sched *sched,
2289 				struct idle_thread_runtime *itr,
2290 				struct perf_sample *sample)
2291 {
2292 	if (!sched->show_callchain || sample->callchain == NULL)
2293 		return;
2294 
2295 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2296 }
2297 
2298 static struct thread *timehist_get_thread(struct perf_sched *sched,
2299 					  struct perf_sample *sample,
2300 					  struct machine *machine,
2301 					  struct evsel *evsel)
2302 {
2303 	struct thread *thread;
2304 
2305 	if (is_idle_sample(sample, evsel)) {
2306 		thread = get_idle_thread(sample->cpu);
2307 		if (thread == NULL)
2308 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2309 
2310 	} else {
2311 		/* there were samples with tid 0 but non-zero pid */
2312 		thread = machine__findnew_thread(machine, sample->pid,
2313 						 sample->tid ?: sample->pid);
2314 		if (thread == NULL) {
2315 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2316 				 sample->tid);
2317 		}
2318 
2319 		save_task_callchain(sched, sample, evsel, machine);
2320 		if (sched->idle_hist) {
2321 			struct thread *idle;
2322 			struct idle_thread_runtime *itr;
2323 
2324 			idle = get_idle_thread(sample->cpu);
2325 			if (idle == NULL) {
2326 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2327 				return NULL;
2328 			}
2329 
2330 			itr = thread__priv(idle);
2331 			if (itr == NULL)
2332 				return NULL;
2333 
2334 			itr->last_thread = thread;
2335 
2336 			/* copy task callchain when entering to idle */
2337 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2338 				save_idle_callchain(sched, itr, sample);
2339 		}
2340 	}
2341 
2342 	return thread;
2343 }
2344 
2345 static bool timehist_skip_sample(struct perf_sched *sched,
2346 				 struct thread *thread,
2347 				 struct evsel *evsel,
2348 				 struct perf_sample *sample)
2349 {
2350 	bool rc = false;
2351 
2352 	if (thread__is_filtered(thread)) {
2353 		rc = true;
2354 		sched->skipped_samples++;
2355 	}
2356 
2357 	if (sched->idle_hist) {
2358 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2359 			rc = true;
2360 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2361 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2362 			rc = true;
2363 	}
2364 
2365 	return rc;
2366 }
2367 
2368 static void timehist_print_wakeup_event(struct perf_sched *sched,
2369 					struct evsel *evsel,
2370 					struct perf_sample *sample,
2371 					struct machine *machine,
2372 					struct thread *awakened)
2373 {
2374 	struct thread *thread;
2375 	char tstr[64];
2376 
2377 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2378 	if (thread == NULL)
2379 		return;
2380 
2381 	/* show wakeup unless both awakee and awaker are filtered */
2382 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2383 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2384 		return;
2385 	}
2386 
2387 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2388 	printf("%15s [%04d] ", tstr, sample->cpu);
2389 	if (sched->show_cpu_visual)
2390 		printf(" %*s ", sched->max_cpu + 1, "");
2391 
2392 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2393 
2394 	/* dt spacer */
2395 	printf("  %9s  %9s  %9s ", "", "", "");
2396 
2397 	printf("awakened: %s", timehist_get_commstr(awakened));
2398 
2399 	printf("\n");
2400 }
2401 
2402 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2403 				       union perf_event *event __maybe_unused,
2404 				       struct evsel *evsel,
2405 				       struct perf_sample *sample,
2406 				       struct machine *machine)
2407 {
2408 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2409 	struct thread *thread;
2410 	struct thread_runtime *tr = NULL;
2411 	/* want pid of awakened task not pid in sample */
2412 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2413 
2414 	thread = machine__findnew_thread(machine, 0, pid);
2415 	if (thread == NULL)
2416 		return -1;
2417 
2418 	tr = thread__get_runtime(thread);
2419 	if (tr == NULL)
2420 		return -1;
2421 
2422 	if (tr->ready_to_run == 0)
2423 		tr->ready_to_run = sample->time;
2424 
2425 	/* show wakeups if requested */
2426 	if (sched->show_wakeups &&
2427 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2428 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2429 
2430 	return 0;
2431 }
2432 
2433 static void timehist_print_migration_event(struct perf_sched *sched,
2434 					struct evsel *evsel,
2435 					struct perf_sample *sample,
2436 					struct machine *machine,
2437 					struct thread *migrated)
2438 {
2439 	struct thread *thread;
2440 	char tstr[64];
2441 	u32 max_cpus = sched->max_cpu + 1;
2442 	u32 ocpu, dcpu;
2443 
2444 	if (sched->summary_only)
2445 		return;
2446 
2447 	max_cpus = sched->max_cpu + 1;
2448 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2449 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2450 
2451 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2452 	if (thread == NULL)
2453 		return;
2454 
2455 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2456 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2457 		return;
2458 	}
2459 
2460 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2461 	printf("%15s [%04d] ", tstr, sample->cpu);
2462 
2463 	if (sched->show_cpu_visual) {
2464 		u32 i;
2465 		char c;
2466 
2467 		printf("  ");
2468 		for (i = 0; i < max_cpus; ++i) {
2469 			c = (i == sample->cpu) ? 'm' : ' ';
2470 			printf("%c", c);
2471 		}
2472 		printf("  ");
2473 	}
2474 
2475 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2476 
2477 	/* dt spacer */
2478 	printf("  %9s  %9s  %9s ", "", "", "");
2479 
2480 	printf("migrated: %s", timehist_get_commstr(migrated));
2481 	printf(" cpu %d => %d", ocpu, dcpu);
2482 
2483 	printf("\n");
2484 }
2485 
2486 static int timehist_migrate_task_event(struct perf_tool *tool,
2487 				       union perf_event *event __maybe_unused,
2488 				       struct evsel *evsel,
2489 				       struct perf_sample *sample,
2490 				       struct machine *machine)
2491 {
2492 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2493 	struct thread *thread;
2494 	struct thread_runtime *tr = NULL;
2495 	/* want pid of migrated task not pid in sample */
2496 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2497 
2498 	thread = machine__findnew_thread(machine, 0, pid);
2499 	if (thread == NULL)
2500 		return -1;
2501 
2502 	tr = thread__get_runtime(thread);
2503 	if (tr == NULL)
2504 		return -1;
2505 
2506 	tr->migrations++;
2507 
2508 	/* show migrations if requested */
2509 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2510 
2511 	return 0;
2512 }
2513 
2514 static int timehist_sched_change_event(struct perf_tool *tool,
2515 				       union perf_event *event,
2516 				       struct evsel *evsel,
2517 				       struct perf_sample *sample,
2518 				       struct machine *machine)
2519 {
2520 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2521 	struct perf_time_interval *ptime = &sched->ptime;
2522 	struct addr_location al;
2523 	struct thread *thread;
2524 	struct thread_runtime *tr = NULL;
2525 	u64 tprev, t = sample->time;
2526 	int rc = 0;
2527 	int state = perf_evsel__intval(evsel, sample, "prev_state");
2528 
2529 
2530 	if (machine__resolve(machine, &al, sample) < 0) {
2531 		pr_err("problem processing %d event. skipping it\n",
2532 		       event->header.type);
2533 		rc = -1;
2534 		goto out;
2535 	}
2536 
2537 	thread = timehist_get_thread(sched, sample, machine, evsel);
2538 	if (thread == NULL) {
2539 		rc = -1;
2540 		goto out;
2541 	}
2542 
2543 	if (timehist_skip_sample(sched, thread, evsel, sample))
2544 		goto out;
2545 
2546 	tr = thread__get_runtime(thread);
2547 	if (tr == NULL) {
2548 		rc = -1;
2549 		goto out;
2550 	}
2551 
2552 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2553 
2554 	/*
2555 	 * If start time given:
2556 	 * - sample time is under window user cares about - skip sample
2557 	 * - tprev is under window user cares about  - reset to start of window
2558 	 */
2559 	if (ptime->start && ptime->start > t)
2560 		goto out;
2561 
2562 	if (tprev && ptime->start > tprev)
2563 		tprev = ptime->start;
2564 
2565 	/*
2566 	 * If end time given:
2567 	 * - previous sched event is out of window - we are done
2568 	 * - sample time is beyond window user cares about - reset it
2569 	 *   to close out stats for time window interest
2570 	 */
2571 	if (ptime->end) {
2572 		if (tprev > ptime->end)
2573 			goto out;
2574 
2575 		if (t > ptime->end)
2576 			t = ptime->end;
2577 	}
2578 
2579 	if (!sched->idle_hist || thread->tid == 0) {
2580 		timehist_update_runtime_stats(tr, t, tprev);
2581 
2582 		if (sched->idle_hist) {
2583 			struct idle_thread_runtime *itr = (void *)tr;
2584 			struct thread_runtime *last_tr;
2585 
2586 			BUG_ON(thread->tid != 0);
2587 
2588 			if (itr->last_thread == NULL)
2589 				goto out;
2590 
2591 			/* add current idle time as last thread's runtime */
2592 			last_tr = thread__get_runtime(itr->last_thread);
2593 			if (last_tr == NULL)
2594 				goto out;
2595 
2596 			timehist_update_runtime_stats(last_tr, t, tprev);
2597 			/*
2598 			 * remove delta time of last thread as it's not updated
2599 			 * and otherwise it will show an invalid value next
2600 			 * time.  we only care total run time and run stat.
2601 			 */
2602 			last_tr->dt_run = 0;
2603 			last_tr->dt_delay = 0;
2604 			last_tr->dt_sleep = 0;
2605 			last_tr->dt_iowait = 0;
2606 			last_tr->dt_preempt = 0;
2607 
2608 			if (itr->cursor.nr)
2609 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2610 
2611 			itr->last_thread = NULL;
2612 		}
2613 	}
2614 
2615 	if (!sched->summary_only)
2616 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2617 
2618 out:
2619 	if (sched->hist_time.start == 0 && t >= ptime->start)
2620 		sched->hist_time.start = t;
2621 	if (ptime->end == 0 || t <= ptime->end)
2622 		sched->hist_time.end = t;
2623 
2624 	if (tr) {
2625 		/* time of this sched_switch event becomes last time task seen */
2626 		tr->last_time = sample->time;
2627 
2628 		/* last state is used to determine where to account wait time */
2629 		tr->last_state = state;
2630 
2631 		/* sched out event for task so reset ready to run time */
2632 		tr->ready_to_run = 0;
2633 	}
2634 
2635 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2636 
2637 	return rc;
2638 }
2639 
2640 static int timehist_sched_switch_event(struct perf_tool *tool,
2641 			     union perf_event *event,
2642 			     struct evsel *evsel,
2643 			     struct perf_sample *sample,
2644 			     struct machine *machine __maybe_unused)
2645 {
2646 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2647 }
2648 
2649 static int process_lost(struct perf_tool *tool __maybe_unused,
2650 			union perf_event *event,
2651 			struct perf_sample *sample,
2652 			struct machine *machine __maybe_unused)
2653 {
2654 	char tstr[64];
2655 
2656 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2657 	printf("%15s ", tstr);
2658 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2659 
2660 	return 0;
2661 }
2662 
2663 
2664 static void print_thread_runtime(struct thread *t,
2665 				 struct thread_runtime *r)
2666 {
2667 	double mean = avg_stats(&r->run_stats);
2668 	float stddev;
2669 
2670 	printf("%*s   %5d  %9" PRIu64 " ",
2671 	       comm_width, timehist_get_commstr(t), t->ppid,
2672 	       (u64) r->run_stats.n);
2673 
2674 	print_sched_time(r->total_run_time, 8);
2675 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2676 	print_sched_time(r->run_stats.min, 6);
2677 	printf(" ");
2678 	print_sched_time((u64) mean, 6);
2679 	printf(" ");
2680 	print_sched_time(r->run_stats.max, 6);
2681 	printf("  ");
2682 	printf("%5.2f", stddev);
2683 	printf("   %5" PRIu64, r->migrations);
2684 	printf("\n");
2685 }
2686 
2687 static void print_thread_waittime(struct thread *t,
2688 				  struct thread_runtime *r)
2689 {
2690 	printf("%*s   %5d  %9" PRIu64 " ",
2691 	       comm_width, timehist_get_commstr(t), t->ppid,
2692 	       (u64) r->run_stats.n);
2693 
2694 	print_sched_time(r->total_run_time, 8);
2695 	print_sched_time(r->total_sleep_time, 6);
2696 	printf(" ");
2697 	print_sched_time(r->total_iowait_time, 6);
2698 	printf(" ");
2699 	print_sched_time(r->total_preempt_time, 6);
2700 	printf(" ");
2701 	print_sched_time(r->total_delay_time, 6);
2702 	printf("\n");
2703 }
2704 
2705 struct total_run_stats {
2706 	struct perf_sched *sched;
2707 	u64  sched_count;
2708 	u64  task_count;
2709 	u64  total_run_time;
2710 };
2711 
2712 static int __show_thread_runtime(struct thread *t, void *priv)
2713 {
2714 	struct total_run_stats *stats = priv;
2715 	struct thread_runtime *r;
2716 
2717 	if (thread__is_filtered(t))
2718 		return 0;
2719 
2720 	r = thread__priv(t);
2721 	if (r && r->run_stats.n) {
2722 		stats->task_count++;
2723 		stats->sched_count += r->run_stats.n;
2724 		stats->total_run_time += r->total_run_time;
2725 
2726 		if (stats->sched->show_state)
2727 			print_thread_waittime(t, r);
2728 		else
2729 			print_thread_runtime(t, r);
2730 	}
2731 
2732 	return 0;
2733 }
2734 
2735 static int show_thread_runtime(struct thread *t, void *priv)
2736 {
2737 	if (t->dead)
2738 		return 0;
2739 
2740 	return __show_thread_runtime(t, priv);
2741 }
2742 
2743 static int show_deadthread_runtime(struct thread *t, void *priv)
2744 {
2745 	if (!t->dead)
2746 		return 0;
2747 
2748 	return __show_thread_runtime(t, priv);
2749 }
2750 
2751 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2752 {
2753 	const char *sep = " <- ";
2754 	struct callchain_list *chain;
2755 	size_t ret = 0;
2756 	char bf[1024];
2757 	bool first;
2758 
2759 	if (node == NULL)
2760 		return 0;
2761 
2762 	ret = callchain__fprintf_folded(fp, node->parent);
2763 	first = (ret == 0);
2764 
2765 	list_for_each_entry(chain, &node->val, list) {
2766 		if (chain->ip >= PERF_CONTEXT_MAX)
2767 			continue;
2768 		if (chain->ms.sym && chain->ms.sym->ignore)
2769 			continue;
2770 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2771 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2772 							false));
2773 		first = false;
2774 	}
2775 
2776 	return ret;
2777 }
2778 
2779 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2780 {
2781 	size_t ret = 0;
2782 	FILE *fp = stdout;
2783 	struct callchain_node *chain;
2784 	struct rb_node *rb_node = rb_first_cached(root);
2785 
2786 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2787 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2788 	       graph_dotted_line);
2789 
2790 	while (rb_node) {
2791 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2792 		rb_node = rb_next(rb_node);
2793 
2794 		ret += fprintf(fp, "  ");
2795 		print_sched_time(chain->hit, 12);
2796 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2797 		ret += fprintf(fp, " %8d  ", chain->count);
2798 		ret += callchain__fprintf_folded(fp, chain);
2799 		ret += fprintf(fp, "\n");
2800 	}
2801 
2802 	return ret;
2803 }
2804 
2805 static void timehist_print_summary(struct perf_sched *sched,
2806 				   struct perf_session *session)
2807 {
2808 	struct machine *m = &session->machines.host;
2809 	struct total_run_stats totals;
2810 	u64 task_count;
2811 	struct thread *t;
2812 	struct thread_runtime *r;
2813 	int i;
2814 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2815 
2816 	memset(&totals, 0, sizeof(totals));
2817 	totals.sched = sched;
2818 
2819 	if (sched->idle_hist) {
2820 		printf("\nIdle-time summary\n");
2821 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2822 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2823 	} else if (sched->show_state) {
2824 		printf("\nWait-time summary\n");
2825 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2826 		printf("   run-time      sleep      iowait     preempt       delay\n");
2827 	} else {
2828 		printf("\nRuntime summary\n");
2829 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2830 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2831 	}
2832 	printf("%*s            (count)  ", comm_width, "");
2833 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2834 	       sched->show_state ? "(msec)" : "%");
2835 	printf("%.117s\n", graph_dotted_line);
2836 
2837 	machine__for_each_thread(m, show_thread_runtime, &totals);
2838 	task_count = totals.task_count;
2839 	if (!task_count)
2840 		printf("<no still running tasks>\n");
2841 
2842 	printf("\nTerminated tasks:\n");
2843 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2844 	if (task_count == totals.task_count)
2845 		printf("<no terminated tasks>\n");
2846 
2847 	/* CPU idle stats not tracked when samples were skipped */
2848 	if (sched->skipped_samples && !sched->idle_hist)
2849 		return;
2850 
2851 	printf("\nIdle stats:\n");
2852 	for (i = 0; i < idle_max_cpu; ++i) {
2853 		t = idle_threads[i];
2854 		if (!t)
2855 			continue;
2856 
2857 		r = thread__priv(t);
2858 		if (r && r->run_stats.n) {
2859 			totals.sched_count += r->run_stats.n;
2860 			printf("    CPU %2d idle for ", i);
2861 			print_sched_time(r->total_run_time, 6);
2862 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2863 		} else
2864 			printf("    CPU %2d idle entire time window\n", i);
2865 	}
2866 
2867 	if (sched->idle_hist && sched->show_callchain) {
2868 		callchain_param.mode  = CHAIN_FOLDED;
2869 		callchain_param.value = CCVAL_PERIOD;
2870 
2871 		callchain_register_param(&callchain_param);
2872 
2873 		printf("\nIdle stats by callchain:\n");
2874 		for (i = 0; i < idle_max_cpu; ++i) {
2875 			struct idle_thread_runtime *itr;
2876 
2877 			t = idle_threads[i];
2878 			if (!t)
2879 				continue;
2880 
2881 			itr = thread__priv(t);
2882 			if (itr == NULL)
2883 				continue;
2884 
2885 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2886 					     0, &callchain_param);
2887 
2888 			printf("  CPU %2d:", i);
2889 			print_sched_time(itr->tr.total_run_time, 6);
2890 			printf(" msec\n");
2891 			timehist_print_idlehist_callchain(&itr->sorted_root);
2892 			printf("\n");
2893 		}
2894 	}
2895 
2896 	printf("\n"
2897 	       "    Total number of unique tasks: %" PRIu64 "\n"
2898 	       "Total number of context switches: %" PRIu64 "\n",
2899 	       totals.task_count, totals.sched_count);
2900 
2901 	printf("           Total run time (msec): ");
2902 	print_sched_time(totals.total_run_time, 2);
2903 	printf("\n");
2904 
2905 	printf("    Total scheduling time (msec): ");
2906 	print_sched_time(hist_time, 2);
2907 	printf(" (x %d)\n", sched->max_cpu);
2908 }
2909 
2910 typedef int (*sched_handler)(struct perf_tool *tool,
2911 			  union perf_event *event,
2912 			  struct evsel *evsel,
2913 			  struct perf_sample *sample,
2914 			  struct machine *machine);
2915 
2916 static int perf_timehist__process_sample(struct perf_tool *tool,
2917 					 union perf_event *event,
2918 					 struct perf_sample *sample,
2919 					 struct evsel *evsel,
2920 					 struct machine *machine)
2921 {
2922 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2923 	int err = 0;
2924 	int this_cpu = sample->cpu;
2925 
2926 	if (this_cpu > sched->max_cpu)
2927 		sched->max_cpu = this_cpu;
2928 
2929 	if (evsel->handler != NULL) {
2930 		sched_handler f = evsel->handler;
2931 
2932 		err = f(tool, event, evsel, sample, machine);
2933 	}
2934 
2935 	return err;
2936 }
2937 
2938 static int timehist_check_attr(struct perf_sched *sched,
2939 			       struct evlist *evlist)
2940 {
2941 	struct evsel *evsel;
2942 	struct evsel_runtime *er;
2943 
2944 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2945 		er = perf_evsel__get_runtime(evsel);
2946 		if (er == NULL) {
2947 			pr_err("Failed to allocate memory for evsel runtime data\n");
2948 			return -1;
2949 		}
2950 
2951 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
2952 			pr_info("Samples do not have callchains.\n");
2953 			sched->show_callchain = 0;
2954 			symbol_conf.use_callchain = 0;
2955 		}
2956 	}
2957 
2958 	return 0;
2959 }
2960 
2961 static int perf_sched__timehist(struct perf_sched *sched)
2962 {
2963 	const struct evsel_str_handler handlers[] = {
2964 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2965 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2966 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2967 	};
2968 	const struct evsel_str_handler migrate_handlers[] = {
2969 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2970 	};
2971 	struct perf_data data = {
2972 		.path  = input_name,
2973 		.mode  = PERF_DATA_MODE_READ,
2974 		.force = sched->force,
2975 	};
2976 
2977 	struct perf_session *session;
2978 	struct evlist *evlist;
2979 	int err = -1;
2980 
2981 	/*
2982 	 * event handlers for timehist option
2983 	 */
2984 	sched->tool.sample	 = perf_timehist__process_sample;
2985 	sched->tool.mmap	 = perf_event__process_mmap;
2986 	sched->tool.comm	 = perf_event__process_comm;
2987 	sched->tool.exit	 = perf_event__process_exit;
2988 	sched->tool.fork	 = perf_event__process_fork;
2989 	sched->tool.lost	 = process_lost;
2990 	sched->tool.attr	 = perf_event__process_attr;
2991 	sched->tool.tracing_data = perf_event__process_tracing_data;
2992 	sched->tool.build_id	 = perf_event__process_build_id;
2993 
2994 	sched->tool.ordered_events = true;
2995 	sched->tool.ordering_requires_timestamps = true;
2996 
2997 	symbol_conf.use_callchain = sched->show_callchain;
2998 
2999 	session = perf_session__new(&data, false, &sched->tool);
3000 	if (IS_ERR(session))
3001 		return PTR_ERR(session);
3002 
3003 	if (cpu_list) {
3004 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3005 		if (err < 0)
3006 			goto out;
3007 	}
3008 
3009 	evlist = session->evlist;
3010 
3011 	symbol__init(&session->header.env);
3012 
3013 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3014 		pr_err("Invalid time string\n");
3015 		return -EINVAL;
3016 	}
3017 
3018 	if (timehist_check_attr(sched, evlist) != 0)
3019 		goto out;
3020 
3021 	setup_pager();
3022 
3023 	/* setup per-evsel handlers */
3024 	if (perf_session__set_tracepoints_handlers(session, handlers))
3025 		goto out;
3026 
3027 	/* sched_switch event at a minimum needs to exist */
3028 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
3029 						  "sched:sched_switch")) {
3030 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3031 		goto out;
3032 	}
3033 
3034 	if (sched->show_migrations &&
3035 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3036 		goto out;
3037 
3038 	/* pre-allocate struct for per-CPU idle stats */
3039 	sched->max_cpu = session->header.env.nr_cpus_online;
3040 	if (sched->max_cpu == 0)
3041 		sched->max_cpu = 4;
3042 	if (init_idle_threads(sched->max_cpu))
3043 		goto out;
3044 
3045 	/* summary_only implies summary option, but don't overwrite summary if set */
3046 	if (sched->summary_only)
3047 		sched->summary = sched->summary_only;
3048 
3049 	if (!sched->summary_only)
3050 		timehist_header(sched);
3051 
3052 	err = perf_session__process_events(session);
3053 	if (err) {
3054 		pr_err("Failed to process events, error %d", err);
3055 		goto out;
3056 	}
3057 
3058 	sched->nr_events      = evlist->stats.nr_events[0];
3059 	sched->nr_lost_events = evlist->stats.total_lost;
3060 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3061 
3062 	if (sched->summary)
3063 		timehist_print_summary(sched, session);
3064 
3065 out:
3066 	free_idle_threads();
3067 	perf_session__delete(session);
3068 
3069 	return err;
3070 }
3071 
3072 
3073 static void print_bad_events(struct perf_sched *sched)
3074 {
3075 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3076 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3077 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3078 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3079 	}
3080 	if (sched->nr_lost_events && sched->nr_events) {
3081 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3082 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3083 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3084 	}
3085 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3086 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3087 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3088 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3089 		if (sched->nr_lost_events)
3090 			printf(" (due to lost events?)");
3091 		printf("\n");
3092 	}
3093 }
3094 
3095 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3096 {
3097 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3098 	struct work_atoms *this;
3099 	const char *comm = thread__comm_str(data->thread), *this_comm;
3100 	bool leftmost = true;
3101 
3102 	while (*new) {
3103 		int cmp;
3104 
3105 		this = container_of(*new, struct work_atoms, node);
3106 		parent = *new;
3107 
3108 		this_comm = thread__comm_str(this->thread);
3109 		cmp = strcmp(comm, this_comm);
3110 		if (cmp > 0) {
3111 			new = &((*new)->rb_left);
3112 		} else if (cmp < 0) {
3113 			new = &((*new)->rb_right);
3114 			leftmost = false;
3115 		} else {
3116 			this->num_merged++;
3117 			this->total_runtime += data->total_runtime;
3118 			this->nb_atoms += data->nb_atoms;
3119 			this->total_lat += data->total_lat;
3120 			list_splice(&data->work_list, &this->work_list);
3121 			if (this->max_lat < data->max_lat) {
3122 				this->max_lat = data->max_lat;
3123 				this->max_lat_at = data->max_lat_at;
3124 			}
3125 			zfree(&data);
3126 			return;
3127 		}
3128 	}
3129 
3130 	data->num_merged++;
3131 	rb_link_node(&data->node, parent, new);
3132 	rb_insert_color_cached(&data->node, root, leftmost);
3133 }
3134 
3135 static void perf_sched__merge_lat(struct perf_sched *sched)
3136 {
3137 	struct work_atoms *data;
3138 	struct rb_node *node;
3139 
3140 	if (sched->skip_merge)
3141 		return;
3142 
3143 	while ((node = rb_first_cached(&sched->atom_root))) {
3144 		rb_erase_cached(node, &sched->atom_root);
3145 		data = rb_entry(node, struct work_atoms, node);
3146 		__merge_work_atoms(&sched->merged_atom_root, data);
3147 	}
3148 }
3149 
3150 static int perf_sched__lat(struct perf_sched *sched)
3151 {
3152 	struct rb_node *next;
3153 
3154 	setup_pager();
3155 
3156 	if (perf_sched__read_events(sched))
3157 		return -1;
3158 
3159 	perf_sched__merge_lat(sched);
3160 	perf_sched__sort_lat(sched);
3161 
3162 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
3163 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
3164 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3165 
3166 	next = rb_first_cached(&sched->sorted_atom_root);
3167 
3168 	while (next) {
3169 		struct work_atoms *work_list;
3170 
3171 		work_list = rb_entry(next, struct work_atoms, node);
3172 		output_lat_thread(sched, work_list);
3173 		next = rb_next(next);
3174 		thread__zput(work_list->thread);
3175 	}
3176 
3177 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3178 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3179 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3180 
3181 	printf(" ---------------------------------------------------\n");
3182 
3183 	print_bad_events(sched);
3184 	printf("\n");
3185 
3186 	return 0;
3187 }
3188 
3189 static int setup_map_cpus(struct perf_sched *sched)
3190 {
3191 	struct perf_cpu_map *map;
3192 
3193 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3194 
3195 	if (sched->map.comp) {
3196 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3197 		if (!sched->map.comp_cpus)
3198 			return -1;
3199 	}
3200 
3201 	if (!sched->map.cpus_str)
3202 		return 0;
3203 
3204 	map = perf_cpu_map__new(sched->map.cpus_str);
3205 	if (!map) {
3206 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3207 		return -1;
3208 	}
3209 
3210 	sched->map.cpus = map;
3211 	return 0;
3212 }
3213 
3214 static int setup_color_pids(struct perf_sched *sched)
3215 {
3216 	struct perf_thread_map *map;
3217 
3218 	if (!sched->map.color_pids_str)
3219 		return 0;
3220 
3221 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3222 	if (!map) {
3223 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3224 		return -1;
3225 	}
3226 
3227 	sched->map.color_pids = map;
3228 	return 0;
3229 }
3230 
3231 static int setup_color_cpus(struct perf_sched *sched)
3232 {
3233 	struct perf_cpu_map *map;
3234 
3235 	if (!sched->map.color_cpus_str)
3236 		return 0;
3237 
3238 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3239 	if (!map) {
3240 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3241 		return -1;
3242 	}
3243 
3244 	sched->map.color_cpus = map;
3245 	return 0;
3246 }
3247 
3248 static int perf_sched__map(struct perf_sched *sched)
3249 {
3250 	if (setup_map_cpus(sched))
3251 		return -1;
3252 
3253 	if (setup_color_pids(sched))
3254 		return -1;
3255 
3256 	if (setup_color_cpus(sched))
3257 		return -1;
3258 
3259 	setup_pager();
3260 	if (perf_sched__read_events(sched))
3261 		return -1;
3262 	print_bad_events(sched);
3263 	return 0;
3264 }
3265 
3266 static int perf_sched__replay(struct perf_sched *sched)
3267 {
3268 	unsigned long i;
3269 
3270 	calibrate_run_measurement_overhead(sched);
3271 	calibrate_sleep_measurement_overhead(sched);
3272 
3273 	test_calibrations(sched);
3274 
3275 	if (perf_sched__read_events(sched))
3276 		return -1;
3277 
3278 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3279 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3280 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3281 
3282 	if (sched->targetless_wakeups)
3283 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3284 	if (sched->multitarget_wakeups)
3285 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3286 	if (sched->nr_run_events_optimized)
3287 		printf("run atoms optimized: %ld\n",
3288 			sched->nr_run_events_optimized);
3289 
3290 	print_task_traces(sched);
3291 	add_cross_task_wakeups(sched);
3292 
3293 	create_tasks(sched);
3294 	printf("------------------------------------------------------------\n");
3295 	for (i = 0; i < sched->replay_repeat; i++)
3296 		run_one_test(sched);
3297 
3298 	return 0;
3299 }
3300 
3301 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3302 			  const char * const usage_msg[])
3303 {
3304 	char *tmp, *tok, *str = strdup(sched->sort_order);
3305 
3306 	for (tok = strtok_r(str, ", ", &tmp);
3307 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3308 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3309 			usage_with_options_msg(usage_msg, options,
3310 					"Unknown --sort key: `%s'", tok);
3311 		}
3312 	}
3313 
3314 	free(str);
3315 
3316 	sort_dimension__add("pid", &sched->cmp_pid);
3317 }
3318 
3319 static int __cmd_record(int argc, const char **argv)
3320 {
3321 	unsigned int rec_argc, i, j;
3322 	const char **rec_argv;
3323 	const char * const record_args[] = {
3324 		"record",
3325 		"-a",
3326 		"-R",
3327 		"-m", "1024",
3328 		"-c", "1",
3329 		"-e", "sched:sched_switch",
3330 		"-e", "sched:sched_stat_wait",
3331 		"-e", "sched:sched_stat_sleep",
3332 		"-e", "sched:sched_stat_iowait",
3333 		"-e", "sched:sched_stat_runtime",
3334 		"-e", "sched:sched_process_fork",
3335 		"-e", "sched:sched_wakeup",
3336 		"-e", "sched:sched_wakeup_new",
3337 		"-e", "sched:sched_migrate_task",
3338 	};
3339 
3340 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3341 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3342 
3343 	if (rec_argv == NULL)
3344 		return -ENOMEM;
3345 
3346 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3347 		rec_argv[i] = strdup(record_args[i]);
3348 
3349 	for (j = 1; j < (unsigned int)argc; j++, i++)
3350 		rec_argv[i] = argv[j];
3351 
3352 	BUG_ON(i != rec_argc);
3353 
3354 	return cmd_record(i, rec_argv);
3355 }
3356 
3357 int cmd_sched(int argc, const char **argv)
3358 {
3359 	static const char default_sort_order[] = "avg, max, switch, runtime";
3360 	struct perf_sched sched = {
3361 		.tool = {
3362 			.sample		 = perf_sched__process_tracepoint_sample,
3363 			.comm		 = perf_sched__process_comm,
3364 			.namespaces	 = perf_event__process_namespaces,
3365 			.lost		 = perf_event__process_lost,
3366 			.fork		 = perf_sched__process_fork_event,
3367 			.ordered_events = true,
3368 		},
3369 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3370 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3371 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3372 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3373 		.sort_order	      = default_sort_order,
3374 		.replay_repeat	      = 10,
3375 		.profile_cpu	      = -1,
3376 		.next_shortname1      = 'A',
3377 		.next_shortname2      = '0',
3378 		.skip_merge           = 0,
3379 		.show_callchain	      = 1,
3380 		.max_stack            = 5,
3381 	};
3382 	const struct option sched_options[] = {
3383 	OPT_STRING('i', "input", &input_name, "file",
3384 		    "input file name"),
3385 	OPT_INCR('v', "verbose", &verbose,
3386 		    "be more verbose (show symbol address, etc)"),
3387 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3388 		    "dump raw trace in ASCII"),
3389 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3390 	OPT_END()
3391 	};
3392 	const struct option latency_options[] = {
3393 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3394 		   "sort by key(s): runtime, switch, avg, max"),
3395 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3396 		    "CPU to profile on"),
3397 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3398 		    "latency stats per pid instead of per comm"),
3399 	OPT_PARENT(sched_options)
3400 	};
3401 	const struct option replay_options[] = {
3402 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3403 		     "repeat the workload replay N times (-1: infinite)"),
3404 	OPT_PARENT(sched_options)
3405 	};
3406 	const struct option map_options[] = {
3407 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3408 		    "map output in compact mode"),
3409 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3410 		   "highlight given pids in map"),
3411 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3412                     "highlight given CPUs in map"),
3413 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3414                     "display given CPUs in map"),
3415 	OPT_PARENT(sched_options)
3416 	};
3417 	const struct option timehist_options[] = {
3418 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3419 		   "file", "vmlinux pathname"),
3420 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3421 		   "file", "kallsyms pathname"),
3422 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3423 		    "Display call chains if present (default on)"),
3424 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3425 		   "Maximum number of functions to display backtrace."),
3426 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3427 		    "Look for files with symbols relative to this directory"),
3428 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3429 		    "Show only syscall summary with statistics"),
3430 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3431 		    "Show all syscalls and summary with statistics"),
3432 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3433 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3434 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3435 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3436 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3437 	OPT_STRING(0, "time", &sched.time_str, "str",
3438 		   "Time span for analysis (start,stop)"),
3439 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3440 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3441 		   "analyze events only for given process id(s)"),
3442 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3443 		   "analyze events only for given thread id(s)"),
3444 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3445 	OPT_PARENT(sched_options)
3446 	};
3447 
3448 	const char * const latency_usage[] = {
3449 		"perf sched latency [<options>]",
3450 		NULL
3451 	};
3452 	const char * const replay_usage[] = {
3453 		"perf sched replay [<options>]",
3454 		NULL
3455 	};
3456 	const char * const map_usage[] = {
3457 		"perf sched map [<options>]",
3458 		NULL
3459 	};
3460 	const char * const timehist_usage[] = {
3461 		"perf sched timehist [<options>]",
3462 		NULL
3463 	};
3464 	const char *const sched_subcommands[] = { "record", "latency", "map",
3465 						  "replay", "script",
3466 						  "timehist", NULL };
3467 	const char *sched_usage[] = {
3468 		NULL,
3469 		NULL
3470 	};
3471 	struct trace_sched_handler lat_ops  = {
3472 		.wakeup_event	    = latency_wakeup_event,
3473 		.switch_event	    = latency_switch_event,
3474 		.runtime_event	    = latency_runtime_event,
3475 		.migrate_task_event = latency_migrate_task_event,
3476 	};
3477 	struct trace_sched_handler map_ops  = {
3478 		.switch_event	    = map_switch_event,
3479 	};
3480 	struct trace_sched_handler replay_ops  = {
3481 		.wakeup_event	    = replay_wakeup_event,
3482 		.switch_event	    = replay_switch_event,
3483 		.fork_event	    = replay_fork_event,
3484 	};
3485 	unsigned int i;
3486 
3487 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3488 		sched.curr_pid[i] = -1;
3489 
3490 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3491 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3492 	if (!argc)
3493 		usage_with_options(sched_usage, sched_options);
3494 
3495 	/*
3496 	 * Aliased to 'perf script' for now:
3497 	 */
3498 	if (!strcmp(argv[0], "script"))
3499 		return cmd_script(argc, argv);
3500 
3501 	if (!strncmp(argv[0], "rec", 3)) {
3502 		return __cmd_record(argc, argv);
3503 	} else if (!strncmp(argv[0], "lat", 3)) {
3504 		sched.tp_handler = &lat_ops;
3505 		if (argc > 1) {
3506 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3507 			if (argc)
3508 				usage_with_options(latency_usage, latency_options);
3509 		}
3510 		setup_sorting(&sched, latency_options, latency_usage);
3511 		return perf_sched__lat(&sched);
3512 	} else if (!strcmp(argv[0], "map")) {
3513 		if (argc) {
3514 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3515 			if (argc)
3516 				usage_with_options(map_usage, map_options);
3517 		}
3518 		sched.tp_handler = &map_ops;
3519 		setup_sorting(&sched, latency_options, latency_usage);
3520 		return perf_sched__map(&sched);
3521 	} else if (!strncmp(argv[0], "rep", 3)) {
3522 		sched.tp_handler = &replay_ops;
3523 		if (argc) {
3524 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3525 			if (argc)
3526 				usage_with_options(replay_usage, replay_options);
3527 		}
3528 		return perf_sched__replay(&sched);
3529 	} else if (!strcmp(argv[0], "timehist")) {
3530 		if (argc) {
3531 			argc = parse_options(argc, argv, timehist_options,
3532 					     timehist_usage, 0);
3533 			if (argc)
3534 				usage_with_options(timehist_usage, timehist_options);
3535 		}
3536 		if ((sched.show_wakeups || sched.show_next) &&
3537 		    sched.summary_only) {
3538 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3539 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3540 			if (sched.show_wakeups)
3541 				parse_options_usage(NULL, timehist_options, "w", true);
3542 			if (sched.show_next)
3543 				parse_options_usage(NULL, timehist_options, "n", true);
3544 			return -EINVAL;
3545 		}
3546 
3547 		return perf_sched__timehist(&sched);
3548 	} else {
3549 		usage_with_options(sched_usage, sched_options);
3550 	}
3551 
3552 	return 0;
3553 }
3554