xref: /openbmc/linux/tools/perf/builtin-sched.c (revision 8730046c)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/util.h"
5 #include "util/evlist.h"
6 #include "util/cache.h"
7 #include "util/evsel.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/tool.h"
13 #include "util/cloexec.h"
14 #include "util/thread_map.h"
15 #include "util/color.h"
16 #include "util/stat.h"
17 #include "util/callchain.h"
18 #include "util/time-utils.h"
19 
20 #include <subcmd/parse-options.h>
21 #include "util/trace-event.h"
22 
23 #include "util/debug.h"
24 
25 #include <linux/log2.h>
26 #include <sys/prctl.h>
27 #include <sys/resource.h>
28 
29 #include <semaphore.h>
30 #include <pthread.h>
31 #include <math.h>
32 #include <api/fs/fs.h>
33 #include <linux/time64.h>
34 
35 #define PR_SET_NAME		15               /* Set process name */
36 #define MAX_CPUS		4096
37 #define COMM_LEN		20
38 #define SYM_LEN			129
39 #define MAX_PID			1024000
40 
41 struct sched_atom;
42 
43 struct task_desc {
44 	unsigned long		nr;
45 	unsigned long		pid;
46 	char			comm[COMM_LEN];
47 
48 	unsigned long		nr_events;
49 	unsigned long		curr_event;
50 	struct sched_atom	**atoms;
51 
52 	pthread_t		thread;
53 	sem_t			sleep_sem;
54 
55 	sem_t			ready_for_work;
56 	sem_t			work_done_sem;
57 
58 	u64			cpu_usage;
59 };
60 
61 enum sched_event_type {
62 	SCHED_EVENT_RUN,
63 	SCHED_EVENT_SLEEP,
64 	SCHED_EVENT_WAKEUP,
65 	SCHED_EVENT_MIGRATION,
66 };
67 
68 struct sched_atom {
69 	enum sched_event_type	type;
70 	int			specific_wait;
71 	u64			timestamp;
72 	u64			duration;
73 	unsigned long		nr;
74 	sem_t			*wait_sem;
75 	struct task_desc	*wakee;
76 };
77 
78 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
79 
80 enum thread_state {
81 	THREAD_SLEEPING = 0,
82 	THREAD_WAIT_CPU,
83 	THREAD_SCHED_IN,
84 	THREAD_IGNORE
85 };
86 
87 struct work_atom {
88 	struct list_head	list;
89 	enum thread_state	state;
90 	u64			sched_out_time;
91 	u64			wake_up_time;
92 	u64			sched_in_time;
93 	u64			runtime;
94 };
95 
96 struct work_atoms {
97 	struct list_head	work_list;
98 	struct thread		*thread;
99 	struct rb_node		node;
100 	u64			max_lat;
101 	u64			max_lat_at;
102 	u64			total_lat;
103 	u64			nb_atoms;
104 	u64			total_runtime;
105 	int			num_merged;
106 };
107 
108 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
109 
110 struct perf_sched;
111 
112 struct trace_sched_handler {
113 	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
114 			    struct perf_sample *sample, struct machine *machine);
115 
116 	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
117 			     struct perf_sample *sample, struct machine *machine);
118 
119 	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
120 			    struct perf_sample *sample, struct machine *machine);
121 
122 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
123 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
124 			  struct machine *machine);
125 
126 	int (*migrate_task_event)(struct perf_sched *sched,
127 				  struct perf_evsel *evsel,
128 				  struct perf_sample *sample,
129 				  struct machine *machine);
130 };
131 
132 #define COLOR_PIDS PERF_COLOR_BLUE
133 #define COLOR_CPUS PERF_COLOR_BG_RED
134 
135 struct perf_sched_map {
136 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
137 	int			*comp_cpus;
138 	bool			 comp;
139 	struct thread_map	*color_pids;
140 	const char		*color_pids_str;
141 	struct cpu_map		*color_cpus;
142 	const char		*color_cpus_str;
143 	struct cpu_map		*cpus;
144 	const char		*cpus_str;
145 };
146 
147 struct perf_sched {
148 	struct perf_tool tool;
149 	const char	 *sort_order;
150 	unsigned long	 nr_tasks;
151 	struct task_desc **pid_to_task;
152 	struct task_desc **tasks;
153 	const struct trace_sched_handler *tp_handler;
154 	pthread_mutex_t	 start_work_mutex;
155 	pthread_mutex_t	 work_done_wait_mutex;
156 	int		 profile_cpu;
157 /*
158  * Track the current task - that way we can know whether there's any
159  * weird events, such as a task being switched away that is not current.
160  */
161 	int		 max_cpu;
162 	u32		 curr_pid[MAX_CPUS];
163 	struct thread	 *curr_thread[MAX_CPUS];
164 	char		 next_shortname1;
165 	char		 next_shortname2;
166 	unsigned int	 replay_repeat;
167 	unsigned long	 nr_run_events;
168 	unsigned long	 nr_sleep_events;
169 	unsigned long	 nr_wakeup_events;
170 	unsigned long	 nr_sleep_corrections;
171 	unsigned long	 nr_run_events_optimized;
172 	unsigned long	 targetless_wakeups;
173 	unsigned long	 multitarget_wakeups;
174 	unsigned long	 nr_runs;
175 	unsigned long	 nr_timestamps;
176 	unsigned long	 nr_unordered_timestamps;
177 	unsigned long	 nr_context_switch_bugs;
178 	unsigned long	 nr_events;
179 	unsigned long	 nr_lost_chunks;
180 	unsigned long	 nr_lost_events;
181 	u64		 run_measurement_overhead;
182 	u64		 sleep_measurement_overhead;
183 	u64		 start_time;
184 	u64		 cpu_usage;
185 	u64		 runavg_cpu_usage;
186 	u64		 parent_cpu_usage;
187 	u64		 runavg_parent_cpu_usage;
188 	u64		 sum_runtime;
189 	u64		 sum_fluct;
190 	u64		 run_avg;
191 	u64		 all_runtime;
192 	u64		 all_count;
193 	u64		 cpu_last_switched[MAX_CPUS];
194 	struct rb_root	 atom_root, sorted_atom_root, merged_atom_root;
195 	struct list_head sort_list, cmp_pid;
196 	bool force;
197 	bool skip_merge;
198 	struct perf_sched_map map;
199 
200 	/* options for timehist command */
201 	bool		summary;
202 	bool		summary_only;
203 	bool		idle_hist;
204 	bool		show_callchain;
205 	unsigned int	max_stack;
206 	bool		show_cpu_visual;
207 	bool		show_wakeups;
208 	bool		show_migrations;
209 	u64		skipped_samples;
210 	const char	*time_str;
211 	struct perf_time_interval ptime;
212 	struct perf_time_interval hist_time;
213 };
214 
215 /* per thread run time data */
216 struct thread_runtime {
217 	u64 last_time;      /* time of previous sched in/out event */
218 	u64 dt_run;         /* run time */
219 	u64 dt_wait;        /* time between CPU access (off cpu) */
220 	u64 dt_delay;       /* time between wakeup and sched-in */
221 	u64 ready_to_run;   /* time of wakeup */
222 
223 	struct stats run_stats;
224 	u64 total_run_time;
225 
226 	u64 migrations;
227 };
228 
229 /* per event run time data */
230 struct evsel_runtime {
231 	u64 *last_time; /* time this event was last seen per cpu */
232 	u32 ncpu;       /* highest cpu slot allocated */
233 };
234 
235 /* per cpu idle time data */
236 struct idle_thread_runtime {
237 	struct thread_runtime	tr;
238 	struct thread		*last_thread;
239 	struct rb_root		sorted_root;
240 	struct callchain_root	callchain;
241 	struct callchain_cursor	cursor;
242 };
243 
244 /* track idle times per cpu */
245 static struct thread **idle_threads;
246 static int idle_max_cpu;
247 static char idle_comm[] = "<idle>";
248 
249 static u64 get_nsecs(void)
250 {
251 	struct timespec ts;
252 
253 	clock_gettime(CLOCK_MONOTONIC, &ts);
254 
255 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
256 }
257 
258 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
259 {
260 	u64 T0 = get_nsecs(), T1;
261 
262 	do {
263 		T1 = get_nsecs();
264 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
265 }
266 
267 static void sleep_nsecs(u64 nsecs)
268 {
269 	struct timespec ts;
270 
271 	ts.tv_nsec = nsecs % 999999999;
272 	ts.tv_sec = nsecs / 999999999;
273 
274 	nanosleep(&ts, NULL);
275 }
276 
277 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
278 {
279 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
280 	int i;
281 
282 	for (i = 0; i < 10; i++) {
283 		T0 = get_nsecs();
284 		burn_nsecs(sched, 0);
285 		T1 = get_nsecs();
286 		delta = T1-T0;
287 		min_delta = min(min_delta, delta);
288 	}
289 	sched->run_measurement_overhead = min_delta;
290 
291 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
292 }
293 
294 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
295 {
296 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
297 	int i;
298 
299 	for (i = 0; i < 10; i++) {
300 		T0 = get_nsecs();
301 		sleep_nsecs(10000);
302 		T1 = get_nsecs();
303 		delta = T1-T0;
304 		min_delta = min(min_delta, delta);
305 	}
306 	min_delta -= 10000;
307 	sched->sleep_measurement_overhead = min_delta;
308 
309 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
310 }
311 
312 static struct sched_atom *
313 get_new_event(struct task_desc *task, u64 timestamp)
314 {
315 	struct sched_atom *event = zalloc(sizeof(*event));
316 	unsigned long idx = task->nr_events;
317 	size_t size;
318 
319 	event->timestamp = timestamp;
320 	event->nr = idx;
321 
322 	task->nr_events++;
323 	size = sizeof(struct sched_atom *) * task->nr_events;
324 	task->atoms = realloc(task->atoms, size);
325 	BUG_ON(!task->atoms);
326 
327 	task->atoms[idx] = event;
328 
329 	return event;
330 }
331 
332 static struct sched_atom *last_event(struct task_desc *task)
333 {
334 	if (!task->nr_events)
335 		return NULL;
336 
337 	return task->atoms[task->nr_events - 1];
338 }
339 
340 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
341 				u64 timestamp, u64 duration)
342 {
343 	struct sched_atom *event, *curr_event = last_event(task);
344 
345 	/*
346 	 * optimize an existing RUN event by merging this one
347 	 * to it:
348 	 */
349 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
350 		sched->nr_run_events_optimized++;
351 		curr_event->duration += duration;
352 		return;
353 	}
354 
355 	event = get_new_event(task, timestamp);
356 
357 	event->type = SCHED_EVENT_RUN;
358 	event->duration = duration;
359 
360 	sched->nr_run_events++;
361 }
362 
363 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
364 				   u64 timestamp, struct task_desc *wakee)
365 {
366 	struct sched_atom *event, *wakee_event;
367 
368 	event = get_new_event(task, timestamp);
369 	event->type = SCHED_EVENT_WAKEUP;
370 	event->wakee = wakee;
371 
372 	wakee_event = last_event(wakee);
373 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
374 		sched->targetless_wakeups++;
375 		return;
376 	}
377 	if (wakee_event->wait_sem) {
378 		sched->multitarget_wakeups++;
379 		return;
380 	}
381 
382 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
383 	sem_init(wakee_event->wait_sem, 0, 0);
384 	wakee_event->specific_wait = 1;
385 	event->wait_sem = wakee_event->wait_sem;
386 
387 	sched->nr_wakeup_events++;
388 }
389 
390 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
391 				  u64 timestamp, u64 task_state __maybe_unused)
392 {
393 	struct sched_atom *event = get_new_event(task, timestamp);
394 
395 	event->type = SCHED_EVENT_SLEEP;
396 
397 	sched->nr_sleep_events++;
398 }
399 
400 static struct task_desc *register_pid(struct perf_sched *sched,
401 				      unsigned long pid, const char *comm)
402 {
403 	struct task_desc *task;
404 	static int pid_max;
405 
406 	if (sched->pid_to_task == NULL) {
407 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
408 			pid_max = MAX_PID;
409 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
410 	}
411 	if (pid >= (unsigned long)pid_max) {
412 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
413 			sizeof(struct task_desc *))) == NULL);
414 		while (pid >= (unsigned long)pid_max)
415 			sched->pid_to_task[pid_max++] = NULL;
416 	}
417 
418 	task = sched->pid_to_task[pid];
419 
420 	if (task)
421 		return task;
422 
423 	task = zalloc(sizeof(*task));
424 	task->pid = pid;
425 	task->nr = sched->nr_tasks;
426 	strcpy(task->comm, comm);
427 	/*
428 	 * every task starts in sleeping state - this gets ignored
429 	 * if there's no wakeup pointing to this sleep state:
430 	 */
431 	add_sched_event_sleep(sched, task, 0, 0);
432 
433 	sched->pid_to_task[pid] = task;
434 	sched->nr_tasks++;
435 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
436 	BUG_ON(!sched->tasks);
437 	sched->tasks[task->nr] = task;
438 
439 	if (verbose)
440 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
441 
442 	return task;
443 }
444 
445 
446 static void print_task_traces(struct perf_sched *sched)
447 {
448 	struct task_desc *task;
449 	unsigned long i;
450 
451 	for (i = 0; i < sched->nr_tasks; i++) {
452 		task = sched->tasks[i];
453 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
454 			task->nr, task->comm, task->pid, task->nr_events);
455 	}
456 }
457 
458 static void add_cross_task_wakeups(struct perf_sched *sched)
459 {
460 	struct task_desc *task1, *task2;
461 	unsigned long i, j;
462 
463 	for (i = 0; i < sched->nr_tasks; i++) {
464 		task1 = sched->tasks[i];
465 		j = i + 1;
466 		if (j == sched->nr_tasks)
467 			j = 0;
468 		task2 = sched->tasks[j];
469 		add_sched_event_wakeup(sched, task1, 0, task2);
470 	}
471 }
472 
473 static void perf_sched__process_event(struct perf_sched *sched,
474 				      struct sched_atom *atom)
475 {
476 	int ret = 0;
477 
478 	switch (atom->type) {
479 		case SCHED_EVENT_RUN:
480 			burn_nsecs(sched, atom->duration);
481 			break;
482 		case SCHED_EVENT_SLEEP:
483 			if (atom->wait_sem)
484 				ret = sem_wait(atom->wait_sem);
485 			BUG_ON(ret);
486 			break;
487 		case SCHED_EVENT_WAKEUP:
488 			if (atom->wait_sem)
489 				ret = sem_post(atom->wait_sem);
490 			BUG_ON(ret);
491 			break;
492 		case SCHED_EVENT_MIGRATION:
493 			break;
494 		default:
495 			BUG_ON(1);
496 	}
497 }
498 
499 static u64 get_cpu_usage_nsec_parent(void)
500 {
501 	struct rusage ru;
502 	u64 sum;
503 	int err;
504 
505 	err = getrusage(RUSAGE_SELF, &ru);
506 	BUG_ON(err);
507 
508 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
509 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
510 
511 	return sum;
512 }
513 
514 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
515 {
516 	struct perf_event_attr attr;
517 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
518 	int fd;
519 	struct rlimit limit;
520 	bool need_privilege = false;
521 
522 	memset(&attr, 0, sizeof(attr));
523 
524 	attr.type = PERF_TYPE_SOFTWARE;
525 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
526 
527 force_again:
528 	fd = sys_perf_event_open(&attr, 0, -1, -1,
529 				 perf_event_open_cloexec_flag());
530 
531 	if (fd < 0) {
532 		if (errno == EMFILE) {
533 			if (sched->force) {
534 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
535 				limit.rlim_cur += sched->nr_tasks - cur_task;
536 				if (limit.rlim_cur > limit.rlim_max) {
537 					limit.rlim_max = limit.rlim_cur;
538 					need_privilege = true;
539 				}
540 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
541 					if (need_privilege && errno == EPERM)
542 						strcpy(info, "Need privilege\n");
543 				} else
544 					goto force_again;
545 			} else
546 				strcpy(info, "Have a try with -f option\n");
547 		}
548 		pr_err("Error: sys_perf_event_open() syscall returned "
549 		       "with %d (%s)\n%s", fd,
550 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
551 		exit(EXIT_FAILURE);
552 	}
553 	return fd;
554 }
555 
556 static u64 get_cpu_usage_nsec_self(int fd)
557 {
558 	u64 runtime;
559 	int ret;
560 
561 	ret = read(fd, &runtime, sizeof(runtime));
562 	BUG_ON(ret != sizeof(runtime));
563 
564 	return runtime;
565 }
566 
567 struct sched_thread_parms {
568 	struct task_desc  *task;
569 	struct perf_sched *sched;
570 	int fd;
571 };
572 
573 static void *thread_func(void *ctx)
574 {
575 	struct sched_thread_parms *parms = ctx;
576 	struct task_desc *this_task = parms->task;
577 	struct perf_sched *sched = parms->sched;
578 	u64 cpu_usage_0, cpu_usage_1;
579 	unsigned long i, ret;
580 	char comm2[22];
581 	int fd = parms->fd;
582 
583 	zfree(&parms);
584 
585 	sprintf(comm2, ":%s", this_task->comm);
586 	prctl(PR_SET_NAME, comm2);
587 	if (fd < 0)
588 		return NULL;
589 again:
590 	ret = sem_post(&this_task->ready_for_work);
591 	BUG_ON(ret);
592 	ret = pthread_mutex_lock(&sched->start_work_mutex);
593 	BUG_ON(ret);
594 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
595 	BUG_ON(ret);
596 
597 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
598 
599 	for (i = 0; i < this_task->nr_events; i++) {
600 		this_task->curr_event = i;
601 		perf_sched__process_event(sched, this_task->atoms[i]);
602 	}
603 
604 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
605 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
606 	ret = sem_post(&this_task->work_done_sem);
607 	BUG_ON(ret);
608 
609 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
610 	BUG_ON(ret);
611 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
612 	BUG_ON(ret);
613 
614 	goto again;
615 }
616 
617 static void create_tasks(struct perf_sched *sched)
618 {
619 	struct task_desc *task;
620 	pthread_attr_t attr;
621 	unsigned long i;
622 	int err;
623 
624 	err = pthread_attr_init(&attr);
625 	BUG_ON(err);
626 	err = pthread_attr_setstacksize(&attr,
627 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
628 	BUG_ON(err);
629 	err = pthread_mutex_lock(&sched->start_work_mutex);
630 	BUG_ON(err);
631 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
632 	BUG_ON(err);
633 	for (i = 0; i < sched->nr_tasks; i++) {
634 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
635 		BUG_ON(parms == NULL);
636 		parms->task = task = sched->tasks[i];
637 		parms->sched = sched;
638 		parms->fd = self_open_counters(sched, i);
639 		sem_init(&task->sleep_sem, 0, 0);
640 		sem_init(&task->ready_for_work, 0, 0);
641 		sem_init(&task->work_done_sem, 0, 0);
642 		task->curr_event = 0;
643 		err = pthread_create(&task->thread, &attr, thread_func, parms);
644 		BUG_ON(err);
645 	}
646 }
647 
648 static void wait_for_tasks(struct perf_sched *sched)
649 {
650 	u64 cpu_usage_0, cpu_usage_1;
651 	struct task_desc *task;
652 	unsigned long i, ret;
653 
654 	sched->start_time = get_nsecs();
655 	sched->cpu_usage = 0;
656 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
657 
658 	for (i = 0; i < sched->nr_tasks; i++) {
659 		task = sched->tasks[i];
660 		ret = sem_wait(&task->ready_for_work);
661 		BUG_ON(ret);
662 		sem_init(&task->ready_for_work, 0, 0);
663 	}
664 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
665 	BUG_ON(ret);
666 
667 	cpu_usage_0 = get_cpu_usage_nsec_parent();
668 
669 	pthread_mutex_unlock(&sched->start_work_mutex);
670 
671 	for (i = 0; i < sched->nr_tasks; i++) {
672 		task = sched->tasks[i];
673 		ret = sem_wait(&task->work_done_sem);
674 		BUG_ON(ret);
675 		sem_init(&task->work_done_sem, 0, 0);
676 		sched->cpu_usage += task->cpu_usage;
677 		task->cpu_usage = 0;
678 	}
679 
680 	cpu_usage_1 = get_cpu_usage_nsec_parent();
681 	if (!sched->runavg_cpu_usage)
682 		sched->runavg_cpu_usage = sched->cpu_usage;
683 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
684 
685 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
686 	if (!sched->runavg_parent_cpu_usage)
687 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
688 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
689 					 sched->parent_cpu_usage)/sched->replay_repeat;
690 
691 	ret = pthread_mutex_lock(&sched->start_work_mutex);
692 	BUG_ON(ret);
693 
694 	for (i = 0; i < sched->nr_tasks; i++) {
695 		task = sched->tasks[i];
696 		sem_init(&task->sleep_sem, 0, 0);
697 		task->curr_event = 0;
698 	}
699 }
700 
701 static void run_one_test(struct perf_sched *sched)
702 {
703 	u64 T0, T1, delta, avg_delta, fluct;
704 
705 	T0 = get_nsecs();
706 	wait_for_tasks(sched);
707 	T1 = get_nsecs();
708 
709 	delta = T1 - T0;
710 	sched->sum_runtime += delta;
711 	sched->nr_runs++;
712 
713 	avg_delta = sched->sum_runtime / sched->nr_runs;
714 	if (delta < avg_delta)
715 		fluct = avg_delta - delta;
716 	else
717 		fluct = delta - avg_delta;
718 	sched->sum_fluct += fluct;
719 	if (!sched->run_avg)
720 		sched->run_avg = delta;
721 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
722 
723 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
724 
725 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
726 
727 	printf("cpu: %0.2f / %0.2f",
728 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
729 
730 #if 0
731 	/*
732 	 * rusage statistics done by the parent, these are less
733 	 * accurate than the sched->sum_exec_runtime based statistics:
734 	 */
735 	printf(" [%0.2f / %0.2f]",
736 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
737 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
738 #endif
739 
740 	printf("\n");
741 
742 	if (sched->nr_sleep_corrections)
743 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
744 	sched->nr_sleep_corrections = 0;
745 }
746 
747 static void test_calibrations(struct perf_sched *sched)
748 {
749 	u64 T0, T1;
750 
751 	T0 = get_nsecs();
752 	burn_nsecs(sched, NSEC_PER_MSEC);
753 	T1 = get_nsecs();
754 
755 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
756 
757 	T0 = get_nsecs();
758 	sleep_nsecs(NSEC_PER_MSEC);
759 	T1 = get_nsecs();
760 
761 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
762 }
763 
764 static int
765 replay_wakeup_event(struct perf_sched *sched,
766 		    struct perf_evsel *evsel, struct perf_sample *sample,
767 		    struct machine *machine __maybe_unused)
768 {
769 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
770 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
771 	struct task_desc *waker, *wakee;
772 
773 	if (verbose) {
774 		printf("sched_wakeup event %p\n", evsel);
775 
776 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
777 	}
778 
779 	waker = register_pid(sched, sample->tid, "<unknown>");
780 	wakee = register_pid(sched, pid, comm);
781 
782 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
783 	return 0;
784 }
785 
786 static int replay_switch_event(struct perf_sched *sched,
787 			       struct perf_evsel *evsel,
788 			       struct perf_sample *sample,
789 			       struct machine *machine __maybe_unused)
790 {
791 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
792 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
793 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
794 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
795 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
796 	struct task_desc *prev, __maybe_unused *next;
797 	u64 timestamp0, timestamp = sample->time;
798 	int cpu = sample->cpu;
799 	s64 delta;
800 
801 	if (verbose)
802 		printf("sched_switch event %p\n", evsel);
803 
804 	if (cpu >= MAX_CPUS || cpu < 0)
805 		return 0;
806 
807 	timestamp0 = sched->cpu_last_switched[cpu];
808 	if (timestamp0)
809 		delta = timestamp - timestamp0;
810 	else
811 		delta = 0;
812 
813 	if (delta < 0) {
814 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
815 		return -1;
816 	}
817 
818 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
819 		 prev_comm, prev_pid, next_comm, next_pid, delta);
820 
821 	prev = register_pid(sched, prev_pid, prev_comm);
822 	next = register_pid(sched, next_pid, next_comm);
823 
824 	sched->cpu_last_switched[cpu] = timestamp;
825 
826 	add_sched_event_run(sched, prev, timestamp, delta);
827 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
828 
829 	return 0;
830 }
831 
832 static int replay_fork_event(struct perf_sched *sched,
833 			     union perf_event *event,
834 			     struct machine *machine)
835 {
836 	struct thread *child, *parent;
837 
838 	child = machine__findnew_thread(machine, event->fork.pid,
839 					event->fork.tid);
840 	parent = machine__findnew_thread(machine, event->fork.ppid,
841 					 event->fork.ptid);
842 
843 	if (child == NULL || parent == NULL) {
844 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
845 				 child, parent);
846 		goto out_put;
847 	}
848 
849 	if (verbose) {
850 		printf("fork event\n");
851 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
852 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
853 	}
854 
855 	register_pid(sched, parent->tid, thread__comm_str(parent));
856 	register_pid(sched, child->tid, thread__comm_str(child));
857 out_put:
858 	thread__put(child);
859 	thread__put(parent);
860 	return 0;
861 }
862 
863 struct sort_dimension {
864 	const char		*name;
865 	sort_fn_t		cmp;
866 	struct list_head	list;
867 };
868 
869 static int
870 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
871 {
872 	struct sort_dimension *sort;
873 	int ret = 0;
874 
875 	BUG_ON(list_empty(list));
876 
877 	list_for_each_entry(sort, list, list) {
878 		ret = sort->cmp(l, r);
879 		if (ret)
880 			return ret;
881 	}
882 
883 	return ret;
884 }
885 
886 static struct work_atoms *
887 thread_atoms_search(struct rb_root *root, struct thread *thread,
888 			 struct list_head *sort_list)
889 {
890 	struct rb_node *node = root->rb_node;
891 	struct work_atoms key = { .thread = thread };
892 
893 	while (node) {
894 		struct work_atoms *atoms;
895 		int cmp;
896 
897 		atoms = container_of(node, struct work_atoms, node);
898 
899 		cmp = thread_lat_cmp(sort_list, &key, atoms);
900 		if (cmp > 0)
901 			node = node->rb_left;
902 		else if (cmp < 0)
903 			node = node->rb_right;
904 		else {
905 			BUG_ON(thread != atoms->thread);
906 			return atoms;
907 		}
908 	}
909 	return NULL;
910 }
911 
912 static void
913 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
914 			 struct list_head *sort_list)
915 {
916 	struct rb_node **new = &(root->rb_node), *parent = NULL;
917 
918 	while (*new) {
919 		struct work_atoms *this;
920 		int cmp;
921 
922 		this = container_of(*new, struct work_atoms, node);
923 		parent = *new;
924 
925 		cmp = thread_lat_cmp(sort_list, data, this);
926 
927 		if (cmp > 0)
928 			new = &((*new)->rb_left);
929 		else
930 			new = &((*new)->rb_right);
931 	}
932 
933 	rb_link_node(&data->node, parent, new);
934 	rb_insert_color(&data->node, root);
935 }
936 
937 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
938 {
939 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
940 	if (!atoms) {
941 		pr_err("No memory at %s\n", __func__);
942 		return -1;
943 	}
944 
945 	atoms->thread = thread__get(thread);
946 	INIT_LIST_HEAD(&atoms->work_list);
947 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
948 	return 0;
949 }
950 
951 static char sched_out_state(u64 prev_state)
952 {
953 	const char *str = TASK_STATE_TO_CHAR_STR;
954 
955 	return str[prev_state];
956 }
957 
958 static int
959 add_sched_out_event(struct work_atoms *atoms,
960 		    char run_state,
961 		    u64 timestamp)
962 {
963 	struct work_atom *atom = zalloc(sizeof(*atom));
964 	if (!atom) {
965 		pr_err("Non memory at %s", __func__);
966 		return -1;
967 	}
968 
969 	atom->sched_out_time = timestamp;
970 
971 	if (run_state == 'R') {
972 		atom->state = THREAD_WAIT_CPU;
973 		atom->wake_up_time = atom->sched_out_time;
974 	}
975 
976 	list_add_tail(&atom->list, &atoms->work_list);
977 	return 0;
978 }
979 
980 static void
981 add_runtime_event(struct work_atoms *atoms, u64 delta,
982 		  u64 timestamp __maybe_unused)
983 {
984 	struct work_atom *atom;
985 
986 	BUG_ON(list_empty(&atoms->work_list));
987 
988 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
989 
990 	atom->runtime += delta;
991 	atoms->total_runtime += delta;
992 }
993 
994 static void
995 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
996 {
997 	struct work_atom *atom;
998 	u64 delta;
999 
1000 	if (list_empty(&atoms->work_list))
1001 		return;
1002 
1003 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1004 
1005 	if (atom->state != THREAD_WAIT_CPU)
1006 		return;
1007 
1008 	if (timestamp < atom->wake_up_time) {
1009 		atom->state = THREAD_IGNORE;
1010 		return;
1011 	}
1012 
1013 	atom->state = THREAD_SCHED_IN;
1014 	atom->sched_in_time = timestamp;
1015 
1016 	delta = atom->sched_in_time - atom->wake_up_time;
1017 	atoms->total_lat += delta;
1018 	if (delta > atoms->max_lat) {
1019 		atoms->max_lat = delta;
1020 		atoms->max_lat_at = timestamp;
1021 	}
1022 	atoms->nb_atoms++;
1023 }
1024 
1025 static int latency_switch_event(struct perf_sched *sched,
1026 				struct perf_evsel *evsel,
1027 				struct perf_sample *sample,
1028 				struct machine *machine)
1029 {
1030 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1031 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1032 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
1033 	struct work_atoms *out_events, *in_events;
1034 	struct thread *sched_out, *sched_in;
1035 	u64 timestamp0, timestamp = sample->time;
1036 	int cpu = sample->cpu, err = -1;
1037 	s64 delta;
1038 
1039 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040 
1041 	timestamp0 = sched->cpu_last_switched[cpu];
1042 	sched->cpu_last_switched[cpu] = timestamp;
1043 	if (timestamp0)
1044 		delta = timestamp - timestamp0;
1045 	else
1046 		delta = 0;
1047 
1048 	if (delta < 0) {
1049 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1050 		return -1;
1051 	}
1052 
1053 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1054 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1055 	if (sched_out == NULL || sched_in == NULL)
1056 		goto out_put;
1057 
1058 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1059 	if (!out_events) {
1060 		if (thread_atoms_insert(sched, sched_out))
1061 			goto out_put;
1062 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1063 		if (!out_events) {
1064 			pr_err("out-event: Internal tree error");
1065 			goto out_put;
1066 		}
1067 	}
1068 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1069 		return -1;
1070 
1071 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1072 	if (!in_events) {
1073 		if (thread_atoms_insert(sched, sched_in))
1074 			goto out_put;
1075 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1076 		if (!in_events) {
1077 			pr_err("in-event: Internal tree error");
1078 			goto out_put;
1079 		}
1080 		/*
1081 		 * Take came in we have not heard about yet,
1082 		 * add in an initial atom in runnable state:
1083 		 */
1084 		if (add_sched_out_event(in_events, 'R', timestamp))
1085 			goto out_put;
1086 	}
1087 	add_sched_in_event(in_events, timestamp);
1088 	err = 0;
1089 out_put:
1090 	thread__put(sched_out);
1091 	thread__put(sched_in);
1092 	return err;
1093 }
1094 
1095 static int latency_runtime_event(struct perf_sched *sched,
1096 				 struct perf_evsel *evsel,
1097 				 struct perf_sample *sample,
1098 				 struct machine *machine)
1099 {
1100 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
1101 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
1102 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1103 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1104 	u64 timestamp = sample->time;
1105 	int cpu = sample->cpu, err = -1;
1106 
1107 	if (thread == NULL)
1108 		return -1;
1109 
1110 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1111 	if (!atoms) {
1112 		if (thread_atoms_insert(sched, thread))
1113 			goto out_put;
1114 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1115 		if (!atoms) {
1116 			pr_err("in-event: Internal tree error");
1117 			goto out_put;
1118 		}
1119 		if (add_sched_out_event(atoms, 'R', timestamp))
1120 			goto out_put;
1121 	}
1122 
1123 	add_runtime_event(atoms, runtime, timestamp);
1124 	err = 0;
1125 out_put:
1126 	thread__put(thread);
1127 	return err;
1128 }
1129 
1130 static int latency_wakeup_event(struct perf_sched *sched,
1131 				struct perf_evsel *evsel,
1132 				struct perf_sample *sample,
1133 				struct machine *machine)
1134 {
1135 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1136 	struct work_atoms *atoms;
1137 	struct work_atom *atom;
1138 	struct thread *wakee;
1139 	u64 timestamp = sample->time;
1140 	int err = -1;
1141 
1142 	wakee = machine__findnew_thread(machine, -1, pid);
1143 	if (wakee == NULL)
1144 		return -1;
1145 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1146 	if (!atoms) {
1147 		if (thread_atoms_insert(sched, wakee))
1148 			goto out_put;
1149 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1150 		if (!atoms) {
1151 			pr_err("wakeup-event: Internal tree error");
1152 			goto out_put;
1153 		}
1154 		if (add_sched_out_event(atoms, 'S', timestamp))
1155 			goto out_put;
1156 	}
1157 
1158 	BUG_ON(list_empty(&atoms->work_list));
1159 
1160 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1161 
1162 	/*
1163 	 * As we do not guarantee the wakeup event happens when
1164 	 * task is out of run queue, also may happen when task is
1165 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1166 	 * then we should not set the ->wake_up_time when wake up a
1167 	 * task which is on run queue.
1168 	 *
1169 	 * You WILL be missing events if you've recorded only
1170 	 * one CPU, or are only looking at only one, so don't
1171 	 * skip in this case.
1172 	 */
1173 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1174 		goto out_ok;
1175 
1176 	sched->nr_timestamps++;
1177 	if (atom->sched_out_time > timestamp) {
1178 		sched->nr_unordered_timestamps++;
1179 		goto out_ok;
1180 	}
1181 
1182 	atom->state = THREAD_WAIT_CPU;
1183 	atom->wake_up_time = timestamp;
1184 out_ok:
1185 	err = 0;
1186 out_put:
1187 	thread__put(wakee);
1188 	return err;
1189 }
1190 
1191 static int latency_migrate_task_event(struct perf_sched *sched,
1192 				      struct perf_evsel *evsel,
1193 				      struct perf_sample *sample,
1194 				      struct machine *machine)
1195 {
1196 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1197 	u64 timestamp = sample->time;
1198 	struct work_atoms *atoms;
1199 	struct work_atom *atom;
1200 	struct thread *migrant;
1201 	int err = -1;
1202 
1203 	/*
1204 	 * Only need to worry about migration when profiling one CPU.
1205 	 */
1206 	if (sched->profile_cpu == -1)
1207 		return 0;
1208 
1209 	migrant = machine__findnew_thread(machine, -1, pid);
1210 	if (migrant == NULL)
1211 		return -1;
1212 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1213 	if (!atoms) {
1214 		if (thread_atoms_insert(sched, migrant))
1215 			goto out_put;
1216 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1217 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1218 		if (!atoms) {
1219 			pr_err("migration-event: Internal tree error");
1220 			goto out_put;
1221 		}
1222 		if (add_sched_out_event(atoms, 'R', timestamp))
1223 			goto out_put;
1224 	}
1225 
1226 	BUG_ON(list_empty(&atoms->work_list));
1227 
1228 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1229 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1230 
1231 	sched->nr_timestamps++;
1232 
1233 	if (atom->sched_out_time > timestamp)
1234 		sched->nr_unordered_timestamps++;
1235 	err = 0;
1236 out_put:
1237 	thread__put(migrant);
1238 	return err;
1239 }
1240 
1241 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1242 {
1243 	int i;
1244 	int ret;
1245 	u64 avg;
1246 	char max_lat_at[32];
1247 
1248 	if (!work_list->nb_atoms)
1249 		return;
1250 	/*
1251 	 * Ignore idle threads:
1252 	 */
1253 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1254 		return;
1255 
1256 	sched->all_runtime += work_list->total_runtime;
1257 	sched->all_count   += work_list->nb_atoms;
1258 
1259 	if (work_list->num_merged > 1)
1260 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1261 	else
1262 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1263 
1264 	for (i = 0; i < 24 - ret; i++)
1265 		printf(" ");
1266 
1267 	avg = work_list->total_lat / work_list->nb_atoms;
1268 	timestamp__scnprintf_usec(work_list->max_lat_at, max_lat_at, sizeof(max_lat_at));
1269 
1270 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13s s\n",
1271 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1272 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1273 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1274 		 max_lat_at);
1275 }
1276 
1277 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1278 {
1279 	if (l->thread == r->thread)
1280 		return 0;
1281 	if (l->thread->tid < r->thread->tid)
1282 		return -1;
1283 	if (l->thread->tid > r->thread->tid)
1284 		return 1;
1285 	return (int)(l->thread - r->thread);
1286 }
1287 
1288 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1289 {
1290 	u64 avgl, avgr;
1291 
1292 	if (!l->nb_atoms)
1293 		return -1;
1294 
1295 	if (!r->nb_atoms)
1296 		return 1;
1297 
1298 	avgl = l->total_lat / l->nb_atoms;
1299 	avgr = r->total_lat / r->nb_atoms;
1300 
1301 	if (avgl < avgr)
1302 		return -1;
1303 	if (avgl > avgr)
1304 		return 1;
1305 
1306 	return 0;
1307 }
1308 
1309 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1310 {
1311 	if (l->max_lat < r->max_lat)
1312 		return -1;
1313 	if (l->max_lat > r->max_lat)
1314 		return 1;
1315 
1316 	return 0;
1317 }
1318 
1319 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1320 {
1321 	if (l->nb_atoms < r->nb_atoms)
1322 		return -1;
1323 	if (l->nb_atoms > r->nb_atoms)
1324 		return 1;
1325 
1326 	return 0;
1327 }
1328 
1329 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1330 {
1331 	if (l->total_runtime < r->total_runtime)
1332 		return -1;
1333 	if (l->total_runtime > r->total_runtime)
1334 		return 1;
1335 
1336 	return 0;
1337 }
1338 
1339 static int sort_dimension__add(const char *tok, struct list_head *list)
1340 {
1341 	size_t i;
1342 	static struct sort_dimension avg_sort_dimension = {
1343 		.name = "avg",
1344 		.cmp  = avg_cmp,
1345 	};
1346 	static struct sort_dimension max_sort_dimension = {
1347 		.name = "max",
1348 		.cmp  = max_cmp,
1349 	};
1350 	static struct sort_dimension pid_sort_dimension = {
1351 		.name = "pid",
1352 		.cmp  = pid_cmp,
1353 	};
1354 	static struct sort_dimension runtime_sort_dimension = {
1355 		.name = "runtime",
1356 		.cmp  = runtime_cmp,
1357 	};
1358 	static struct sort_dimension switch_sort_dimension = {
1359 		.name = "switch",
1360 		.cmp  = switch_cmp,
1361 	};
1362 	struct sort_dimension *available_sorts[] = {
1363 		&pid_sort_dimension,
1364 		&avg_sort_dimension,
1365 		&max_sort_dimension,
1366 		&switch_sort_dimension,
1367 		&runtime_sort_dimension,
1368 	};
1369 
1370 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1371 		if (!strcmp(available_sorts[i]->name, tok)) {
1372 			list_add_tail(&available_sorts[i]->list, list);
1373 
1374 			return 0;
1375 		}
1376 	}
1377 
1378 	return -1;
1379 }
1380 
1381 static void perf_sched__sort_lat(struct perf_sched *sched)
1382 {
1383 	struct rb_node *node;
1384 	struct rb_root *root = &sched->atom_root;
1385 again:
1386 	for (;;) {
1387 		struct work_atoms *data;
1388 		node = rb_first(root);
1389 		if (!node)
1390 			break;
1391 
1392 		rb_erase(node, root);
1393 		data = rb_entry(node, struct work_atoms, node);
1394 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1395 	}
1396 	if (root == &sched->atom_root) {
1397 		root = &sched->merged_atom_root;
1398 		goto again;
1399 	}
1400 }
1401 
1402 static int process_sched_wakeup_event(struct perf_tool *tool,
1403 				      struct perf_evsel *evsel,
1404 				      struct perf_sample *sample,
1405 				      struct machine *machine)
1406 {
1407 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1408 
1409 	if (sched->tp_handler->wakeup_event)
1410 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1411 
1412 	return 0;
1413 }
1414 
1415 union map_priv {
1416 	void	*ptr;
1417 	bool	 color;
1418 };
1419 
1420 static bool thread__has_color(struct thread *thread)
1421 {
1422 	union map_priv priv = {
1423 		.ptr = thread__priv(thread),
1424 	};
1425 
1426 	return priv.color;
1427 }
1428 
1429 static struct thread*
1430 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1431 {
1432 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1433 	union map_priv priv = {
1434 		.color = false,
1435 	};
1436 
1437 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1438 		return thread;
1439 
1440 	if (thread_map__has(sched->map.color_pids, tid))
1441 		priv.color = true;
1442 
1443 	thread__set_priv(thread, priv.ptr);
1444 	return thread;
1445 }
1446 
1447 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1448 			    struct perf_sample *sample, struct machine *machine)
1449 {
1450 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1451 	struct thread *sched_in;
1452 	int new_shortname;
1453 	u64 timestamp0, timestamp = sample->time;
1454 	s64 delta;
1455 	int i, this_cpu = sample->cpu;
1456 	int cpus_nr;
1457 	bool new_cpu = false;
1458 	const char *color = PERF_COLOR_NORMAL;
1459 	char stimestamp[32];
1460 
1461 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1462 
1463 	if (this_cpu > sched->max_cpu)
1464 		sched->max_cpu = this_cpu;
1465 
1466 	if (sched->map.comp) {
1467 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1468 		if (!test_and_set_bit(this_cpu, sched->map.comp_cpus_mask)) {
1469 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1470 			new_cpu = true;
1471 		}
1472 	} else
1473 		cpus_nr = sched->max_cpu;
1474 
1475 	timestamp0 = sched->cpu_last_switched[this_cpu];
1476 	sched->cpu_last_switched[this_cpu] = timestamp;
1477 	if (timestamp0)
1478 		delta = timestamp - timestamp0;
1479 	else
1480 		delta = 0;
1481 
1482 	if (delta < 0) {
1483 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1484 		return -1;
1485 	}
1486 
1487 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1488 	if (sched_in == NULL)
1489 		return -1;
1490 
1491 	sched->curr_thread[this_cpu] = thread__get(sched_in);
1492 
1493 	printf("  ");
1494 
1495 	new_shortname = 0;
1496 	if (!sched_in->shortname[0]) {
1497 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1498 			/*
1499 			 * Don't allocate a letter-number for swapper:0
1500 			 * as a shortname. Instead, we use '.' for it.
1501 			 */
1502 			sched_in->shortname[0] = '.';
1503 			sched_in->shortname[1] = ' ';
1504 		} else {
1505 			sched_in->shortname[0] = sched->next_shortname1;
1506 			sched_in->shortname[1] = sched->next_shortname2;
1507 
1508 			if (sched->next_shortname1 < 'Z') {
1509 				sched->next_shortname1++;
1510 			} else {
1511 				sched->next_shortname1 = 'A';
1512 				if (sched->next_shortname2 < '9')
1513 					sched->next_shortname2++;
1514 				else
1515 					sched->next_shortname2 = '0';
1516 			}
1517 		}
1518 		new_shortname = 1;
1519 	}
1520 
1521 	for (i = 0; i < cpus_nr; i++) {
1522 		int cpu = sched->map.comp ? sched->map.comp_cpus[i] : i;
1523 		struct thread *curr_thread = sched->curr_thread[cpu];
1524 		const char *pid_color = color;
1525 		const char *cpu_color = color;
1526 
1527 		if (curr_thread && thread__has_color(curr_thread))
1528 			pid_color = COLOR_PIDS;
1529 
1530 		if (sched->map.cpus && !cpu_map__has(sched->map.cpus, cpu))
1531 			continue;
1532 
1533 		if (sched->map.color_cpus && cpu_map__has(sched->map.color_cpus, cpu))
1534 			cpu_color = COLOR_CPUS;
1535 
1536 		if (cpu != this_cpu)
1537 			color_fprintf(stdout, color, " ");
1538 		else
1539 			color_fprintf(stdout, cpu_color, "*");
1540 
1541 		if (sched->curr_thread[cpu])
1542 			color_fprintf(stdout, pid_color, "%2s ", sched->curr_thread[cpu]->shortname);
1543 		else
1544 			color_fprintf(stdout, color, "   ");
1545 	}
1546 
1547 	if (sched->map.cpus && !cpu_map__has(sched->map.cpus, this_cpu))
1548 		goto out;
1549 
1550 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1551 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1552 	if (new_shortname || (verbose && sched_in->tid)) {
1553 		const char *pid_color = color;
1554 
1555 		if (thread__has_color(sched_in))
1556 			pid_color = COLOR_PIDS;
1557 
1558 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1559 		       sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1560 	}
1561 
1562 	if (sched->map.comp && new_cpu)
1563 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1564 
1565 out:
1566 	color_fprintf(stdout, color, "\n");
1567 
1568 	thread__put(sched_in);
1569 
1570 	return 0;
1571 }
1572 
1573 static int process_sched_switch_event(struct perf_tool *tool,
1574 				      struct perf_evsel *evsel,
1575 				      struct perf_sample *sample,
1576 				      struct machine *machine)
1577 {
1578 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1579 	int this_cpu = sample->cpu, err = 0;
1580 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1581 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1582 
1583 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1584 		/*
1585 		 * Are we trying to switch away a PID that is
1586 		 * not current?
1587 		 */
1588 		if (sched->curr_pid[this_cpu] != prev_pid)
1589 			sched->nr_context_switch_bugs++;
1590 	}
1591 
1592 	if (sched->tp_handler->switch_event)
1593 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1594 
1595 	sched->curr_pid[this_cpu] = next_pid;
1596 	return err;
1597 }
1598 
1599 static int process_sched_runtime_event(struct perf_tool *tool,
1600 				       struct perf_evsel *evsel,
1601 				       struct perf_sample *sample,
1602 				       struct machine *machine)
1603 {
1604 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1605 
1606 	if (sched->tp_handler->runtime_event)
1607 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1608 
1609 	return 0;
1610 }
1611 
1612 static int perf_sched__process_fork_event(struct perf_tool *tool,
1613 					  union perf_event *event,
1614 					  struct perf_sample *sample,
1615 					  struct machine *machine)
1616 {
1617 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1618 
1619 	/* run the fork event through the perf machineruy */
1620 	perf_event__process_fork(tool, event, sample, machine);
1621 
1622 	/* and then run additional processing needed for this command */
1623 	if (sched->tp_handler->fork_event)
1624 		return sched->tp_handler->fork_event(sched, event, machine);
1625 
1626 	return 0;
1627 }
1628 
1629 static int process_sched_migrate_task_event(struct perf_tool *tool,
1630 					    struct perf_evsel *evsel,
1631 					    struct perf_sample *sample,
1632 					    struct machine *machine)
1633 {
1634 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1635 
1636 	if (sched->tp_handler->migrate_task_event)
1637 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1638 
1639 	return 0;
1640 }
1641 
1642 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1643 				  struct perf_evsel *evsel,
1644 				  struct perf_sample *sample,
1645 				  struct machine *machine);
1646 
1647 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1648 						 union perf_event *event __maybe_unused,
1649 						 struct perf_sample *sample,
1650 						 struct perf_evsel *evsel,
1651 						 struct machine *machine)
1652 {
1653 	int err = 0;
1654 
1655 	if (evsel->handler != NULL) {
1656 		tracepoint_handler f = evsel->handler;
1657 		err = f(tool, evsel, sample, machine);
1658 	}
1659 
1660 	return err;
1661 }
1662 
1663 static int perf_sched__read_events(struct perf_sched *sched)
1664 {
1665 	const struct perf_evsel_str_handler handlers[] = {
1666 		{ "sched:sched_switch",	      process_sched_switch_event, },
1667 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1668 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1669 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1670 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1671 	};
1672 	struct perf_session *session;
1673 	struct perf_data_file file = {
1674 		.path = input_name,
1675 		.mode = PERF_DATA_MODE_READ,
1676 		.force = sched->force,
1677 	};
1678 	int rc = -1;
1679 
1680 	session = perf_session__new(&file, false, &sched->tool);
1681 	if (session == NULL) {
1682 		pr_debug("No Memory for session\n");
1683 		return -1;
1684 	}
1685 
1686 	symbol__init(&session->header.env);
1687 
1688 	if (perf_session__set_tracepoints_handlers(session, handlers))
1689 		goto out_delete;
1690 
1691 	if (perf_session__has_traces(session, "record -R")) {
1692 		int err = perf_session__process_events(session);
1693 		if (err) {
1694 			pr_err("Failed to process events, error %d", err);
1695 			goto out_delete;
1696 		}
1697 
1698 		sched->nr_events      = session->evlist->stats.nr_events[0];
1699 		sched->nr_lost_events = session->evlist->stats.total_lost;
1700 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1701 	}
1702 
1703 	rc = 0;
1704 out_delete:
1705 	perf_session__delete(session);
1706 	return rc;
1707 }
1708 
1709 /*
1710  * scheduling times are printed as msec.usec
1711  */
1712 static inline void print_sched_time(unsigned long long nsecs, int width)
1713 {
1714 	unsigned long msecs;
1715 	unsigned long usecs;
1716 
1717 	msecs  = nsecs / NSEC_PER_MSEC;
1718 	nsecs -= msecs * NSEC_PER_MSEC;
1719 	usecs  = nsecs / NSEC_PER_USEC;
1720 	printf("%*lu.%03lu ", width, msecs, usecs);
1721 }
1722 
1723 /*
1724  * returns runtime data for event, allocating memory for it the
1725  * first time it is used.
1726  */
1727 static struct evsel_runtime *perf_evsel__get_runtime(struct perf_evsel *evsel)
1728 {
1729 	struct evsel_runtime *r = evsel->priv;
1730 
1731 	if (r == NULL) {
1732 		r = zalloc(sizeof(struct evsel_runtime));
1733 		evsel->priv = r;
1734 	}
1735 
1736 	return r;
1737 }
1738 
1739 /*
1740  * save last time event was seen per cpu
1741  */
1742 static void perf_evsel__save_time(struct perf_evsel *evsel,
1743 				  u64 timestamp, u32 cpu)
1744 {
1745 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1746 
1747 	if (r == NULL)
1748 		return;
1749 
1750 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1751 		int i, n = __roundup_pow_of_two(cpu+1);
1752 		void *p = r->last_time;
1753 
1754 		p = realloc(r->last_time, n * sizeof(u64));
1755 		if (!p)
1756 			return;
1757 
1758 		r->last_time = p;
1759 		for (i = r->ncpu; i < n; ++i)
1760 			r->last_time[i] = (u64) 0;
1761 
1762 		r->ncpu = n;
1763 	}
1764 
1765 	r->last_time[cpu] = timestamp;
1766 }
1767 
1768 /* returns last time this event was seen on the given cpu */
1769 static u64 perf_evsel__get_time(struct perf_evsel *evsel, u32 cpu)
1770 {
1771 	struct evsel_runtime *r = perf_evsel__get_runtime(evsel);
1772 
1773 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1774 		return 0;
1775 
1776 	return r->last_time[cpu];
1777 }
1778 
1779 static int comm_width = 30;
1780 
1781 static char *timehist_get_commstr(struct thread *thread)
1782 {
1783 	static char str[32];
1784 	const char *comm = thread__comm_str(thread);
1785 	pid_t tid = thread->tid;
1786 	pid_t pid = thread->pid_;
1787 	int n;
1788 
1789 	if (pid == 0)
1790 		n = scnprintf(str, sizeof(str), "%s", comm);
1791 
1792 	else if (tid != pid)
1793 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1794 
1795 	else
1796 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1797 
1798 	if (n > comm_width)
1799 		comm_width = n;
1800 
1801 	return str;
1802 }
1803 
1804 static void timehist_header(struct perf_sched *sched)
1805 {
1806 	u32 ncpus = sched->max_cpu + 1;
1807 	u32 i, j;
1808 
1809 	printf("%15s %6s ", "time", "cpu");
1810 
1811 	if (sched->show_cpu_visual) {
1812 		printf(" ");
1813 		for (i = 0, j = 0; i < ncpus; ++i) {
1814 			printf("%x", j++);
1815 			if (j > 15)
1816 				j = 0;
1817 		}
1818 		printf(" ");
1819 	}
1820 
1821 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1822 		"task name", "wait time", "sch delay", "run time");
1823 
1824 	printf("\n");
1825 
1826 	/*
1827 	 * units row
1828 	 */
1829 	printf("%15s %-6s ", "", "");
1830 
1831 	if (sched->show_cpu_visual)
1832 		printf(" %*s ", ncpus, "");
1833 
1834 	printf(" %-*s  %9s  %9s  %9s\n", comm_width,
1835 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
1836 
1837 	/*
1838 	 * separator
1839 	 */
1840 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
1841 
1842 	if (sched->show_cpu_visual)
1843 		printf(" %.*s ", ncpus, graph_dotted_line);
1844 
1845 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
1846 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
1847 		graph_dotted_line);
1848 
1849 	printf("\n");
1850 }
1851 
1852 static void timehist_print_sample(struct perf_sched *sched,
1853 				  struct perf_sample *sample,
1854 				  struct addr_location *al,
1855 				  struct thread *thread,
1856 				  u64 t)
1857 {
1858 	struct thread_runtime *tr = thread__priv(thread);
1859 	u32 max_cpus = sched->max_cpu + 1;
1860 	char tstr[64];
1861 
1862 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
1863 	printf("%15s [%04d] ", tstr, sample->cpu);
1864 
1865 	if (sched->show_cpu_visual) {
1866 		u32 i;
1867 		char c;
1868 
1869 		printf(" ");
1870 		for (i = 0; i < max_cpus; ++i) {
1871 			/* flag idle times with 'i'; others are sched events */
1872 			if (i == sample->cpu)
1873 				c = (thread->tid == 0) ? 'i' : 's';
1874 			else
1875 				c = ' ';
1876 			printf("%c", c);
1877 		}
1878 		printf(" ");
1879 	}
1880 
1881 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
1882 
1883 	print_sched_time(tr->dt_wait, 6);
1884 	print_sched_time(tr->dt_delay, 6);
1885 	print_sched_time(tr->dt_run, 6);
1886 
1887 	if (sched->show_wakeups)
1888 		printf("  %-*s", comm_width, "");
1889 
1890 	if (thread->tid == 0)
1891 		goto out;
1892 
1893 	if (sched->show_callchain)
1894 		printf("  ");
1895 
1896 	sample__fprintf_sym(sample, al, 0,
1897 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
1898 			    EVSEL__PRINT_CALLCHAIN_ARROW |
1899 			    EVSEL__PRINT_SKIP_IGNORED,
1900 			    &callchain_cursor, stdout);
1901 
1902 out:
1903 	printf("\n");
1904 }
1905 
1906 /*
1907  * Explanation of delta-time stats:
1908  *
1909  *            t = time of current schedule out event
1910  *        tprev = time of previous sched out event
1911  *                also time of schedule-in event for current task
1912  *    last_time = time of last sched change event for current task
1913  *                (i.e, time process was last scheduled out)
1914  * ready_to_run = time of wakeup for current task
1915  *
1916  * -----|------------|------------|------------|------
1917  *    last         ready        tprev          t
1918  *    time         to run
1919  *
1920  *      |-------- dt_wait --------|
1921  *                   |- dt_delay -|-- dt_run --|
1922  *
1923  *   dt_run = run time of current task
1924  *  dt_wait = time between last schedule out event for task and tprev
1925  *            represents time spent off the cpu
1926  * dt_delay = time between wakeup and schedule-in of task
1927  */
1928 
1929 static void timehist_update_runtime_stats(struct thread_runtime *r,
1930 					 u64 t, u64 tprev)
1931 {
1932 	r->dt_delay   = 0;
1933 	r->dt_wait    = 0;
1934 	r->dt_run     = 0;
1935 	if (tprev) {
1936 		r->dt_run = t - tprev;
1937 		if (r->ready_to_run) {
1938 			if (r->ready_to_run > tprev)
1939 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
1940 			else
1941 				r->dt_delay = tprev - r->ready_to_run;
1942 		}
1943 
1944 		if (r->last_time > tprev)
1945 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
1946 		else if (r->last_time)
1947 			r->dt_wait = tprev - r->last_time;
1948 	}
1949 
1950 	update_stats(&r->run_stats, r->dt_run);
1951 	r->total_run_time += r->dt_run;
1952 }
1953 
1954 static bool is_idle_sample(struct perf_sample *sample,
1955 			   struct perf_evsel *evsel)
1956 {
1957 	/* pid 0 == swapper == idle task */
1958 	if (strcmp(perf_evsel__name(evsel), "sched:sched_switch") == 0)
1959 		return perf_evsel__intval(evsel, sample, "prev_pid") == 0;
1960 
1961 	return sample->pid == 0;
1962 }
1963 
1964 static void save_task_callchain(struct perf_sched *sched,
1965 				struct perf_sample *sample,
1966 				struct perf_evsel *evsel,
1967 				struct machine *machine)
1968 {
1969 	struct callchain_cursor *cursor = &callchain_cursor;
1970 	struct thread *thread;
1971 
1972 	/* want main thread for process - has maps */
1973 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
1974 	if (thread == NULL) {
1975 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
1976 		return;
1977 	}
1978 
1979 	if (!symbol_conf.use_callchain || sample->callchain == NULL)
1980 		return;
1981 
1982 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
1983 				      NULL, NULL, sched->max_stack + 2) != 0) {
1984 		if (verbose)
1985 			error("Failed to resolve callchain. Skipping\n");
1986 
1987 		return;
1988 	}
1989 
1990 	callchain_cursor_commit(cursor);
1991 
1992 	while (true) {
1993 		struct callchain_cursor_node *node;
1994 		struct symbol *sym;
1995 
1996 		node = callchain_cursor_current(cursor);
1997 		if (node == NULL)
1998 			break;
1999 
2000 		sym = node->sym;
2001 		if (sym && sym->name) {
2002 			if (!strcmp(sym->name, "schedule") ||
2003 			    !strcmp(sym->name, "__schedule") ||
2004 			    !strcmp(sym->name, "preempt_schedule"))
2005 				sym->ignore = 1;
2006 		}
2007 
2008 		callchain_cursor_advance(cursor);
2009 	}
2010 }
2011 
2012 static int init_idle_thread(struct thread *thread)
2013 {
2014 	struct idle_thread_runtime *itr;
2015 
2016 	thread__set_comm(thread, idle_comm, 0);
2017 
2018 	itr = zalloc(sizeof(*itr));
2019 	if (itr == NULL)
2020 		return -ENOMEM;
2021 
2022 	init_stats(&itr->tr.run_stats);
2023 	callchain_init(&itr->callchain);
2024 	callchain_cursor_reset(&itr->cursor);
2025 	thread__set_priv(thread, itr);
2026 
2027 	return 0;
2028 }
2029 
2030 /*
2031  * Track idle stats per cpu by maintaining a local thread
2032  * struct for the idle task on each cpu.
2033  */
2034 static int init_idle_threads(int ncpu)
2035 {
2036 	int i, ret;
2037 
2038 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2039 	if (!idle_threads)
2040 		return -ENOMEM;
2041 
2042 	idle_max_cpu = ncpu;
2043 
2044 	/* allocate the actual thread struct if needed */
2045 	for (i = 0; i < ncpu; ++i) {
2046 		idle_threads[i] = thread__new(0, 0);
2047 		if (idle_threads[i] == NULL)
2048 			return -ENOMEM;
2049 
2050 		ret = init_idle_thread(idle_threads[i]);
2051 		if (ret < 0)
2052 			return ret;
2053 	}
2054 
2055 	return 0;
2056 }
2057 
2058 static void free_idle_threads(void)
2059 {
2060 	int i;
2061 
2062 	if (idle_threads == NULL)
2063 		return;
2064 
2065 	for (i = 0; i < idle_max_cpu; ++i) {
2066 		if ((idle_threads[i]))
2067 			thread__delete(idle_threads[i]);
2068 	}
2069 
2070 	free(idle_threads);
2071 }
2072 
2073 static struct thread *get_idle_thread(int cpu)
2074 {
2075 	/*
2076 	 * expand/allocate array of pointers to local thread
2077 	 * structs if needed
2078 	 */
2079 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2080 		int i, j = __roundup_pow_of_two(cpu+1);
2081 		void *p;
2082 
2083 		p = realloc(idle_threads, j * sizeof(struct thread *));
2084 		if (!p)
2085 			return NULL;
2086 
2087 		idle_threads = (struct thread **) p;
2088 		for (i = idle_max_cpu; i < j; ++i)
2089 			idle_threads[i] = NULL;
2090 
2091 		idle_max_cpu = j;
2092 	}
2093 
2094 	/* allocate a new thread struct if needed */
2095 	if (idle_threads[cpu] == NULL) {
2096 		idle_threads[cpu] = thread__new(0, 0);
2097 
2098 		if (idle_threads[cpu]) {
2099 			if (init_idle_thread(idle_threads[cpu]) < 0)
2100 				return NULL;
2101 		}
2102 	}
2103 
2104 	return idle_threads[cpu];
2105 }
2106 
2107 static void save_idle_callchain(struct idle_thread_runtime *itr,
2108 				struct perf_sample *sample)
2109 {
2110 	if (!symbol_conf.use_callchain || sample->callchain == NULL)
2111 		return;
2112 
2113 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2114 }
2115 
2116 /*
2117  * handle runtime stats saved per thread
2118  */
2119 static struct thread_runtime *thread__init_runtime(struct thread *thread)
2120 {
2121 	struct thread_runtime *r;
2122 
2123 	r = zalloc(sizeof(struct thread_runtime));
2124 	if (!r)
2125 		return NULL;
2126 
2127 	init_stats(&r->run_stats);
2128 	thread__set_priv(thread, r);
2129 
2130 	return r;
2131 }
2132 
2133 static struct thread_runtime *thread__get_runtime(struct thread *thread)
2134 {
2135 	struct thread_runtime *tr;
2136 
2137 	tr = thread__priv(thread);
2138 	if (tr == NULL) {
2139 		tr = thread__init_runtime(thread);
2140 		if (tr == NULL)
2141 			pr_debug("Failed to malloc memory for runtime data.\n");
2142 	}
2143 
2144 	return tr;
2145 }
2146 
2147 static struct thread *timehist_get_thread(struct perf_sched *sched,
2148 					  struct perf_sample *sample,
2149 					  struct machine *machine,
2150 					  struct perf_evsel *evsel)
2151 {
2152 	struct thread *thread;
2153 
2154 	if (is_idle_sample(sample, evsel)) {
2155 		thread = get_idle_thread(sample->cpu);
2156 		if (thread == NULL)
2157 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2158 
2159 	} else {
2160 		/* there were samples with tid 0 but non-zero pid */
2161 		thread = machine__findnew_thread(machine, sample->pid,
2162 						 sample->tid ?: sample->pid);
2163 		if (thread == NULL) {
2164 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2165 				 sample->tid);
2166 		}
2167 
2168 		save_task_callchain(sched, sample, evsel, machine);
2169 		if (sched->idle_hist) {
2170 			struct thread *idle;
2171 			struct idle_thread_runtime *itr;
2172 
2173 			idle = get_idle_thread(sample->cpu);
2174 			if (idle == NULL) {
2175 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2176 				return NULL;
2177 			}
2178 
2179 			itr = thread__priv(idle);
2180 			if (itr == NULL)
2181 				return NULL;
2182 
2183 			itr->last_thread = thread;
2184 
2185 			/* copy task callchain when entering to idle */
2186 			if (perf_evsel__intval(evsel, sample, "next_pid") == 0)
2187 				save_idle_callchain(itr, sample);
2188 		}
2189 	}
2190 
2191 	return thread;
2192 }
2193 
2194 static bool timehist_skip_sample(struct perf_sched *sched,
2195 				 struct thread *thread,
2196 				 struct perf_evsel *evsel,
2197 				 struct perf_sample *sample)
2198 {
2199 	bool rc = false;
2200 
2201 	if (thread__is_filtered(thread)) {
2202 		rc = true;
2203 		sched->skipped_samples++;
2204 	}
2205 
2206 	if (sched->idle_hist) {
2207 		if (strcmp(perf_evsel__name(evsel), "sched:sched_switch"))
2208 			rc = true;
2209 		else if (perf_evsel__intval(evsel, sample, "prev_pid") != 0 &&
2210 			 perf_evsel__intval(evsel, sample, "next_pid") != 0)
2211 			rc = true;
2212 	}
2213 
2214 	return rc;
2215 }
2216 
2217 static void timehist_print_wakeup_event(struct perf_sched *sched,
2218 					struct perf_evsel *evsel,
2219 					struct perf_sample *sample,
2220 					struct machine *machine,
2221 					struct thread *awakened)
2222 {
2223 	struct thread *thread;
2224 	char tstr[64];
2225 
2226 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2227 	if (thread == NULL)
2228 		return;
2229 
2230 	/* show wakeup unless both awakee and awaker are filtered */
2231 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2232 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2233 		return;
2234 	}
2235 
2236 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2237 	printf("%15s [%04d] ", tstr, sample->cpu);
2238 	if (sched->show_cpu_visual)
2239 		printf(" %*s ", sched->max_cpu + 1, "");
2240 
2241 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2242 
2243 	/* dt spacer */
2244 	printf("  %9s  %9s  %9s ", "", "", "");
2245 
2246 	printf("awakened: %s", timehist_get_commstr(awakened));
2247 
2248 	printf("\n");
2249 }
2250 
2251 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2252 				       union perf_event *event __maybe_unused,
2253 				       struct perf_evsel *evsel,
2254 				       struct perf_sample *sample,
2255 				       struct machine *machine)
2256 {
2257 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2258 	struct thread *thread;
2259 	struct thread_runtime *tr = NULL;
2260 	/* want pid of awakened task not pid in sample */
2261 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2262 
2263 	thread = machine__findnew_thread(machine, 0, pid);
2264 	if (thread == NULL)
2265 		return -1;
2266 
2267 	tr = thread__get_runtime(thread);
2268 	if (tr == NULL)
2269 		return -1;
2270 
2271 	if (tr->ready_to_run == 0)
2272 		tr->ready_to_run = sample->time;
2273 
2274 	/* show wakeups if requested */
2275 	if (sched->show_wakeups &&
2276 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2277 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2278 
2279 	return 0;
2280 }
2281 
2282 static void timehist_print_migration_event(struct perf_sched *sched,
2283 					struct perf_evsel *evsel,
2284 					struct perf_sample *sample,
2285 					struct machine *machine,
2286 					struct thread *migrated)
2287 {
2288 	struct thread *thread;
2289 	char tstr[64];
2290 	u32 max_cpus = sched->max_cpu + 1;
2291 	u32 ocpu, dcpu;
2292 
2293 	if (sched->summary_only)
2294 		return;
2295 
2296 	max_cpus = sched->max_cpu + 1;
2297 	ocpu = perf_evsel__intval(evsel, sample, "orig_cpu");
2298 	dcpu = perf_evsel__intval(evsel, sample, "dest_cpu");
2299 
2300 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2301 	if (thread == NULL)
2302 		return;
2303 
2304 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2305 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2306 		return;
2307 	}
2308 
2309 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2310 	printf("%15s [%04d] ", tstr, sample->cpu);
2311 
2312 	if (sched->show_cpu_visual) {
2313 		u32 i;
2314 		char c;
2315 
2316 		printf("  ");
2317 		for (i = 0; i < max_cpus; ++i) {
2318 			c = (i == sample->cpu) ? 'm' : ' ';
2319 			printf("%c", c);
2320 		}
2321 		printf("  ");
2322 	}
2323 
2324 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2325 
2326 	/* dt spacer */
2327 	printf("  %9s  %9s  %9s ", "", "", "");
2328 
2329 	printf("migrated: %s", timehist_get_commstr(migrated));
2330 	printf(" cpu %d => %d", ocpu, dcpu);
2331 
2332 	printf("\n");
2333 }
2334 
2335 static int timehist_migrate_task_event(struct perf_tool *tool,
2336 				       union perf_event *event __maybe_unused,
2337 				       struct perf_evsel *evsel,
2338 				       struct perf_sample *sample,
2339 				       struct machine *machine)
2340 {
2341 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2342 	struct thread *thread;
2343 	struct thread_runtime *tr = NULL;
2344 	/* want pid of migrated task not pid in sample */
2345 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
2346 
2347 	thread = machine__findnew_thread(machine, 0, pid);
2348 	if (thread == NULL)
2349 		return -1;
2350 
2351 	tr = thread__get_runtime(thread);
2352 	if (tr == NULL)
2353 		return -1;
2354 
2355 	tr->migrations++;
2356 
2357 	/* show migrations if requested */
2358 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2359 
2360 	return 0;
2361 }
2362 
2363 static int timehist_sched_change_event(struct perf_tool *tool,
2364 				       union perf_event *event,
2365 				       struct perf_evsel *evsel,
2366 				       struct perf_sample *sample,
2367 				       struct machine *machine)
2368 {
2369 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2370 	struct perf_time_interval *ptime = &sched->ptime;
2371 	struct addr_location al;
2372 	struct thread *thread;
2373 	struct thread_runtime *tr = NULL;
2374 	u64 tprev, t = sample->time;
2375 	int rc = 0;
2376 
2377 	if (machine__resolve(machine, &al, sample) < 0) {
2378 		pr_err("problem processing %d event. skipping it\n",
2379 		       event->header.type);
2380 		rc = -1;
2381 		goto out;
2382 	}
2383 
2384 	thread = timehist_get_thread(sched, sample, machine, evsel);
2385 	if (thread == NULL) {
2386 		rc = -1;
2387 		goto out;
2388 	}
2389 
2390 	if (timehist_skip_sample(sched, thread, evsel, sample))
2391 		goto out;
2392 
2393 	tr = thread__get_runtime(thread);
2394 	if (tr == NULL) {
2395 		rc = -1;
2396 		goto out;
2397 	}
2398 
2399 	tprev = perf_evsel__get_time(evsel, sample->cpu);
2400 
2401 	/*
2402 	 * If start time given:
2403 	 * - sample time is under window user cares about - skip sample
2404 	 * - tprev is under window user cares about  - reset to start of window
2405 	 */
2406 	if (ptime->start && ptime->start > t)
2407 		goto out;
2408 
2409 	if (tprev && ptime->start > tprev)
2410 		tprev = ptime->start;
2411 
2412 	/*
2413 	 * If end time given:
2414 	 * - previous sched event is out of window - we are done
2415 	 * - sample time is beyond window user cares about - reset it
2416 	 *   to close out stats for time window interest
2417 	 */
2418 	if (ptime->end) {
2419 		if (tprev > ptime->end)
2420 			goto out;
2421 
2422 		if (t > ptime->end)
2423 			t = ptime->end;
2424 	}
2425 
2426 	if (!sched->idle_hist || thread->tid == 0) {
2427 		timehist_update_runtime_stats(tr, t, tprev);
2428 
2429 		if (sched->idle_hist) {
2430 			struct idle_thread_runtime *itr = (void *)tr;
2431 			struct thread_runtime *last_tr;
2432 
2433 			BUG_ON(thread->tid != 0);
2434 
2435 			if (itr->last_thread == NULL)
2436 				goto out;
2437 
2438 			/* add current idle time as last thread's runtime */
2439 			last_tr = thread__get_runtime(itr->last_thread);
2440 			if (last_tr == NULL)
2441 				goto out;
2442 
2443 			timehist_update_runtime_stats(last_tr, t, tprev);
2444 			/*
2445 			 * remove delta time of last thread as it's not updated
2446 			 * and otherwise it will show an invalid value next
2447 			 * time.  we only care total run time and run stat.
2448 			 */
2449 			last_tr->dt_run = 0;
2450 			last_tr->dt_wait = 0;
2451 			last_tr->dt_delay = 0;
2452 
2453 			if (itr->cursor.nr)
2454 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2455 
2456 			itr->last_thread = NULL;
2457 		}
2458 	}
2459 
2460 	if (!sched->summary_only)
2461 		timehist_print_sample(sched, sample, &al, thread, t);
2462 
2463 out:
2464 	if (sched->hist_time.start == 0 && t >= ptime->start)
2465 		sched->hist_time.start = t;
2466 	if (ptime->end == 0 || t <= ptime->end)
2467 		sched->hist_time.end = t;
2468 
2469 	if (tr) {
2470 		/* time of this sched_switch event becomes last time task seen */
2471 		tr->last_time = sample->time;
2472 
2473 		/* sched out event for task so reset ready to run time */
2474 		tr->ready_to_run = 0;
2475 	}
2476 
2477 	perf_evsel__save_time(evsel, sample->time, sample->cpu);
2478 
2479 	return rc;
2480 }
2481 
2482 static int timehist_sched_switch_event(struct perf_tool *tool,
2483 			     union perf_event *event,
2484 			     struct perf_evsel *evsel,
2485 			     struct perf_sample *sample,
2486 			     struct machine *machine __maybe_unused)
2487 {
2488 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2489 }
2490 
2491 static int process_lost(struct perf_tool *tool __maybe_unused,
2492 			union perf_event *event,
2493 			struct perf_sample *sample,
2494 			struct machine *machine __maybe_unused)
2495 {
2496 	char tstr[64];
2497 
2498 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2499 	printf("%15s ", tstr);
2500 	printf("lost %" PRIu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2501 
2502 	return 0;
2503 }
2504 
2505 
2506 static void print_thread_runtime(struct thread *t,
2507 				 struct thread_runtime *r)
2508 {
2509 	double mean = avg_stats(&r->run_stats);
2510 	float stddev;
2511 
2512 	printf("%*s   %5d  %9" PRIu64 " ",
2513 	       comm_width, timehist_get_commstr(t), t->ppid,
2514 	       (u64) r->run_stats.n);
2515 
2516 	print_sched_time(r->total_run_time, 8);
2517 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2518 	print_sched_time(r->run_stats.min, 6);
2519 	printf(" ");
2520 	print_sched_time((u64) mean, 6);
2521 	printf(" ");
2522 	print_sched_time(r->run_stats.max, 6);
2523 	printf("  ");
2524 	printf("%5.2f", stddev);
2525 	printf("   %5" PRIu64, r->migrations);
2526 	printf("\n");
2527 }
2528 
2529 struct total_run_stats {
2530 	u64  sched_count;
2531 	u64  task_count;
2532 	u64  total_run_time;
2533 };
2534 
2535 static int __show_thread_runtime(struct thread *t, void *priv)
2536 {
2537 	struct total_run_stats *stats = priv;
2538 	struct thread_runtime *r;
2539 
2540 	if (thread__is_filtered(t))
2541 		return 0;
2542 
2543 	r = thread__priv(t);
2544 	if (r && r->run_stats.n) {
2545 		stats->task_count++;
2546 		stats->sched_count += r->run_stats.n;
2547 		stats->total_run_time += r->total_run_time;
2548 		print_thread_runtime(t, r);
2549 	}
2550 
2551 	return 0;
2552 }
2553 
2554 static int show_thread_runtime(struct thread *t, void *priv)
2555 {
2556 	if (t->dead)
2557 		return 0;
2558 
2559 	return __show_thread_runtime(t, priv);
2560 }
2561 
2562 static int show_deadthread_runtime(struct thread *t, void *priv)
2563 {
2564 	if (!t->dead)
2565 		return 0;
2566 
2567 	return __show_thread_runtime(t, priv);
2568 }
2569 
2570 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2571 {
2572 	const char *sep = " <- ";
2573 	struct callchain_list *chain;
2574 	size_t ret = 0;
2575 	char bf[1024];
2576 	bool first;
2577 
2578 	if (node == NULL)
2579 		return 0;
2580 
2581 	ret = callchain__fprintf_folded(fp, node->parent);
2582 	first = (ret == 0);
2583 
2584 	list_for_each_entry(chain, &node->val, list) {
2585 		if (chain->ip >= PERF_CONTEXT_MAX)
2586 			continue;
2587 		if (chain->ms.sym && chain->ms.sym->ignore)
2588 			continue;
2589 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2590 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2591 							false));
2592 		first = false;
2593 	}
2594 
2595 	return ret;
2596 }
2597 
2598 static size_t timehist_print_idlehist_callchain(struct rb_root *root)
2599 {
2600 	size_t ret = 0;
2601 	FILE *fp = stdout;
2602 	struct callchain_node *chain;
2603 	struct rb_node *rb_node = rb_first(root);
2604 
2605 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2606 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2607 	       graph_dotted_line);
2608 
2609 	while (rb_node) {
2610 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2611 		rb_node = rb_next(rb_node);
2612 
2613 		ret += fprintf(fp, "  ");
2614 		print_sched_time(chain->hit, 12);
2615 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2616 		ret += fprintf(fp, " %8d  ", chain->count);
2617 		ret += callchain__fprintf_folded(fp, chain);
2618 		ret += fprintf(fp, "\n");
2619 	}
2620 
2621 	return ret;
2622 }
2623 
2624 static void timehist_print_summary(struct perf_sched *sched,
2625 				   struct perf_session *session)
2626 {
2627 	struct machine *m = &session->machines.host;
2628 	struct total_run_stats totals;
2629 	u64 task_count;
2630 	struct thread *t;
2631 	struct thread_runtime *r;
2632 	int i;
2633 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2634 
2635 	memset(&totals, 0, sizeof(totals));
2636 
2637 	if (sched->idle_hist) {
2638 		printf("\nIdle-time summary\n");
2639 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2640 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2641 	} else {
2642 		printf("\nRuntime summary\n");
2643 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2644 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2645 	}
2646 	printf("%*s            (count)  ", comm_width, "");
2647 	printf("     (msec)     (msec)      (msec)      (msec)       %%\n");
2648 	printf("%.117s\n", graph_dotted_line);
2649 
2650 	machine__for_each_thread(m, show_thread_runtime, &totals);
2651 	task_count = totals.task_count;
2652 	if (!task_count)
2653 		printf("<no still running tasks>\n");
2654 
2655 	printf("\nTerminated tasks:\n");
2656 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2657 	if (task_count == totals.task_count)
2658 		printf("<no terminated tasks>\n");
2659 
2660 	/* CPU idle stats not tracked when samples were skipped */
2661 	if (sched->skipped_samples && !sched->idle_hist)
2662 		return;
2663 
2664 	printf("\nIdle stats:\n");
2665 	for (i = 0; i < idle_max_cpu; ++i) {
2666 		t = idle_threads[i];
2667 		if (!t)
2668 			continue;
2669 
2670 		r = thread__priv(t);
2671 		if (r && r->run_stats.n) {
2672 			totals.sched_count += r->run_stats.n;
2673 			printf("    CPU %2d idle for ", i);
2674 			print_sched_time(r->total_run_time, 6);
2675 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2676 		} else
2677 			printf("    CPU %2d idle entire time window\n", i);
2678 	}
2679 
2680 	if (sched->idle_hist && symbol_conf.use_callchain) {
2681 		callchain_param.mode  = CHAIN_FOLDED;
2682 		callchain_param.value = CCVAL_PERIOD;
2683 
2684 		callchain_register_param(&callchain_param);
2685 
2686 		printf("\nIdle stats by callchain:\n");
2687 		for (i = 0; i < idle_max_cpu; ++i) {
2688 			struct idle_thread_runtime *itr;
2689 
2690 			t = idle_threads[i];
2691 			if (!t)
2692 				continue;
2693 
2694 			itr = thread__priv(t);
2695 			if (itr == NULL)
2696 				continue;
2697 
2698 			callchain_param.sort(&itr->sorted_root, &itr->callchain,
2699 					     0, &callchain_param);
2700 
2701 			printf("  CPU %2d:", i);
2702 			print_sched_time(itr->tr.total_run_time, 6);
2703 			printf(" msec\n");
2704 			timehist_print_idlehist_callchain(&itr->sorted_root);
2705 			printf("\n");
2706 		}
2707 	}
2708 
2709 	printf("\n"
2710 	       "    Total number of unique tasks: %" PRIu64 "\n"
2711 	       "Total number of context switches: %" PRIu64 "\n",
2712 	       totals.task_count, totals.sched_count);
2713 
2714 	printf("           Total run time (msec): ");
2715 	print_sched_time(totals.total_run_time, 2);
2716 	printf("\n");
2717 
2718 	printf("    Total scheduling time (msec): ");
2719 	print_sched_time(hist_time, 2);
2720 	printf(" (x %d)\n", sched->max_cpu);
2721 }
2722 
2723 typedef int (*sched_handler)(struct perf_tool *tool,
2724 			  union perf_event *event,
2725 			  struct perf_evsel *evsel,
2726 			  struct perf_sample *sample,
2727 			  struct machine *machine);
2728 
2729 static int perf_timehist__process_sample(struct perf_tool *tool,
2730 					 union perf_event *event,
2731 					 struct perf_sample *sample,
2732 					 struct perf_evsel *evsel,
2733 					 struct machine *machine)
2734 {
2735 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2736 	int err = 0;
2737 	int this_cpu = sample->cpu;
2738 
2739 	if (this_cpu > sched->max_cpu)
2740 		sched->max_cpu = this_cpu;
2741 
2742 	if (evsel->handler != NULL) {
2743 		sched_handler f = evsel->handler;
2744 
2745 		err = f(tool, event, evsel, sample, machine);
2746 	}
2747 
2748 	return err;
2749 }
2750 
2751 static int timehist_check_attr(struct perf_sched *sched,
2752 			       struct perf_evlist *evlist)
2753 {
2754 	struct perf_evsel *evsel;
2755 	struct evsel_runtime *er;
2756 
2757 	list_for_each_entry(evsel, &evlist->entries, node) {
2758 		er = perf_evsel__get_runtime(evsel);
2759 		if (er == NULL) {
2760 			pr_err("Failed to allocate memory for evsel runtime data\n");
2761 			return -1;
2762 		}
2763 
2764 		if (sched->show_callchain &&
2765 		    !(evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN)) {
2766 			pr_info("Samples do not have callchains.\n");
2767 			sched->show_callchain = 0;
2768 			symbol_conf.use_callchain = 0;
2769 		}
2770 	}
2771 
2772 	return 0;
2773 }
2774 
2775 static int perf_sched__timehist(struct perf_sched *sched)
2776 {
2777 	const struct perf_evsel_str_handler handlers[] = {
2778 		{ "sched:sched_switch",       timehist_sched_switch_event, },
2779 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
2780 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
2781 	};
2782 	const struct perf_evsel_str_handler migrate_handlers[] = {
2783 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
2784 	};
2785 	struct perf_data_file file = {
2786 		.path = input_name,
2787 		.mode = PERF_DATA_MODE_READ,
2788 		.force = sched->force,
2789 	};
2790 
2791 	struct perf_session *session;
2792 	struct perf_evlist *evlist;
2793 	int err = -1;
2794 
2795 	/*
2796 	 * event handlers for timehist option
2797 	 */
2798 	sched->tool.sample	 = perf_timehist__process_sample;
2799 	sched->tool.mmap	 = perf_event__process_mmap;
2800 	sched->tool.comm	 = perf_event__process_comm;
2801 	sched->tool.exit	 = perf_event__process_exit;
2802 	sched->tool.fork	 = perf_event__process_fork;
2803 	sched->tool.lost	 = process_lost;
2804 	sched->tool.attr	 = perf_event__process_attr;
2805 	sched->tool.tracing_data = perf_event__process_tracing_data;
2806 	sched->tool.build_id	 = perf_event__process_build_id;
2807 
2808 	sched->tool.ordered_events = true;
2809 	sched->tool.ordering_requires_timestamps = true;
2810 
2811 	symbol_conf.use_callchain = sched->show_callchain;
2812 
2813 	session = perf_session__new(&file, false, &sched->tool);
2814 	if (session == NULL)
2815 		return -ENOMEM;
2816 
2817 	evlist = session->evlist;
2818 
2819 	symbol__init(&session->header.env);
2820 
2821 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
2822 		pr_err("Invalid time string\n");
2823 		return -EINVAL;
2824 	}
2825 
2826 	if (timehist_check_attr(sched, evlist) != 0)
2827 		goto out;
2828 
2829 	setup_pager();
2830 
2831 	/* setup per-evsel handlers */
2832 	if (perf_session__set_tracepoints_handlers(session, handlers))
2833 		goto out;
2834 
2835 	/* sched_switch event at a minimum needs to exist */
2836 	if (!perf_evlist__find_tracepoint_by_name(session->evlist,
2837 						  "sched:sched_switch")) {
2838 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
2839 		goto out;
2840 	}
2841 
2842 	if (sched->show_migrations &&
2843 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
2844 		goto out;
2845 
2846 	/* pre-allocate struct for per-CPU idle stats */
2847 	sched->max_cpu = session->header.env.nr_cpus_online;
2848 	if (sched->max_cpu == 0)
2849 		sched->max_cpu = 4;
2850 	if (init_idle_threads(sched->max_cpu))
2851 		goto out;
2852 
2853 	/* summary_only implies summary option, but don't overwrite summary if set */
2854 	if (sched->summary_only)
2855 		sched->summary = sched->summary_only;
2856 
2857 	if (!sched->summary_only)
2858 		timehist_header(sched);
2859 
2860 	err = perf_session__process_events(session);
2861 	if (err) {
2862 		pr_err("Failed to process events, error %d", err);
2863 		goto out;
2864 	}
2865 
2866 	sched->nr_events      = evlist->stats.nr_events[0];
2867 	sched->nr_lost_events = evlist->stats.total_lost;
2868 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
2869 
2870 	if (sched->summary)
2871 		timehist_print_summary(sched, session);
2872 
2873 out:
2874 	free_idle_threads();
2875 	perf_session__delete(session);
2876 
2877 	return err;
2878 }
2879 
2880 
2881 static void print_bad_events(struct perf_sched *sched)
2882 {
2883 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
2884 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
2885 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
2886 			sched->nr_unordered_timestamps, sched->nr_timestamps);
2887 	}
2888 	if (sched->nr_lost_events && sched->nr_events) {
2889 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
2890 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
2891 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
2892 	}
2893 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
2894 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
2895 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
2896 			sched->nr_context_switch_bugs, sched->nr_timestamps);
2897 		if (sched->nr_lost_events)
2898 			printf(" (due to lost events?)");
2899 		printf("\n");
2900 	}
2901 }
2902 
2903 static void __merge_work_atoms(struct rb_root *root, struct work_atoms *data)
2904 {
2905 	struct rb_node **new = &(root->rb_node), *parent = NULL;
2906 	struct work_atoms *this;
2907 	const char *comm = thread__comm_str(data->thread), *this_comm;
2908 
2909 	while (*new) {
2910 		int cmp;
2911 
2912 		this = container_of(*new, struct work_atoms, node);
2913 		parent = *new;
2914 
2915 		this_comm = thread__comm_str(this->thread);
2916 		cmp = strcmp(comm, this_comm);
2917 		if (cmp > 0) {
2918 			new = &((*new)->rb_left);
2919 		} else if (cmp < 0) {
2920 			new = &((*new)->rb_right);
2921 		} else {
2922 			this->num_merged++;
2923 			this->total_runtime += data->total_runtime;
2924 			this->nb_atoms += data->nb_atoms;
2925 			this->total_lat += data->total_lat;
2926 			list_splice(&data->work_list, &this->work_list);
2927 			if (this->max_lat < data->max_lat) {
2928 				this->max_lat = data->max_lat;
2929 				this->max_lat_at = data->max_lat_at;
2930 			}
2931 			zfree(&data);
2932 			return;
2933 		}
2934 	}
2935 
2936 	data->num_merged++;
2937 	rb_link_node(&data->node, parent, new);
2938 	rb_insert_color(&data->node, root);
2939 }
2940 
2941 static void perf_sched__merge_lat(struct perf_sched *sched)
2942 {
2943 	struct work_atoms *data;
2944 	struct rb_node *node;
2945 
2946 	if (sched->skip_merge)
2947 		return;
2948 
2949 	while ((node = rb_first(&sched->atom_root))) {
2950 		rb_erase(node, &sched->atom_root);
2951 		data = rb_entry(node, struct work_atoms, node);
2952 		__merge_work_atoms(&sched->merged_atom_root, data);
2953 	}
2954 }
2955 
2956 static int perf_sched__lat(struct perf_sched *sched)
2957 {
2958 	struct rb_node *next;
2959 
2960 	setup_pager();
2961 
2962 	if (perf_sched__read_events(sched))
2963 		return -1;
2964 
2965 	perf_sched__merge_lat(sched);
2966 	perf_sched__sort_lat(sched);
2967 
2968 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
2969 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
2970 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
2971 
2972 	next = rb_first(&sched->sorted_atom_root);
2973 
2974 	while (next) {
2975 		struct work_atoms *work_list;
2976 
2977 		work_list = rb_entry(next, struct work_atoms, node);
2978 		output_lat_thread(sched, work_list);
2979 		next = rb_next(next);
2980 		thread__zput(work_list->thread);
2981 	}
2982 
2983 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
2984 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
2985 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
2986 
2987 	printf(" ---------------------------------------------------\n");
2988 
2989 	print_bad_events(sched);
2990 	printf("\n");
2991 
2992 	return 0;
2993 }
2994 
2995 static int setup_map_cpus(struct perf_sched *sched)
2996 {
2997 	struct cpu_map *map;
2998 
2999 	sched->max_cpu  = sysconf(_SC_NPROCESSORS_CONF);
3000 
3001 	if (sched->map.comp) {
3002 		sched->map.comp_cpus = zalloc(sched->max_cpu * sizeof(int));
3003 		if (!sched->map.comp_cpus)
3004 			return -1;
3005 	}
3006 
3007 	if (!sched->map.cpus_str)
3008 		return 0;
3009 
3010 	map = cpu_map__new(sched->map.cpus_str);
3011 	if (!map) {
3012 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3013 		return -1;
3014 	}
3015 
3016 	sched->map.cpus = map;
3017 	return 0;
3018 }
3019 
3020 static int setup_color_pids(struct perf_sched *sched)
3021 {
3022 	struct thread_map *map;
3023 
3024 	if (!sched->map.color_pids_str)
3025 		return 0;
3026 
3027 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3028 	if (!map) {
3029 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3030 		return -1;
3031 	}
3032 
3033 	sched->map.color_pids = map;
3034 	return 0;
3035 }
3036 
3037 static int setup_color_cpus(struct perf_sched *sched)
3038 {
3039 	struct cpu_map *map;
3040 
3041 	if (!sched->map.color_cpus_str)
3042 		return 0;
3043 
3044 	map = cpu_map__new(sched->map.color_cpus_str);
3045 	if (!map) {
3046 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3047 		return -1;
3048 	}
3049 
3050 	sched->map.color_cpus = map;
3051 	return 0;
3052 }
3053 
3054 static int perf_sched__map(struct perf_sched *sched)
3055 {
3056 	if (setup_map_cpus(sched))
3057 		return -1;
3058 
3059 	if (setup_color_pids(sched))
3060 		return -1;
3061 
3062 	if (setup_color_cpus(sched))
3063 		return -1;
3064 
3065 	setup_pager();
3066 	if (perf_sched__read_events(sched))
3067 		return -1;
3068 	print_bad_events(sched);
3069 	return 0;
3070 }
3071 
3072 static int perf_sched__replay(struct perf_sched *sched)
3073 {
3074 	unsigned long i;
3075 
3076 	calibrate_run_measurement_overhead(sched);
3077 	calibrate_sleep_measurement_overhead(sched);
3078 
3079 	test_calibrations(sched);
3080 
3081 	if (perf_sched__read_events(sched))
3082 		return -1;
3083 
3084 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3085 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3086 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3087 
3088 	if (sched->targetless_wakeups)
3089 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3090 	if (sched->multitarget_wakeups)
3091 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3092 	if (sched->nr_run_events_optimized)
3093 		printf("run atoms optimized: %ld\n",
3094 			sched->nr_run_events_optimized);
3095 
3096 	print_task_traces(sched);
3097 	add_cross_task_wakeups(sched);
3098 
3099 	create_tasks(sched);
3100 	printf("------------------------------------------------------------\n");
3101 	for (i = 0; i < sched->replay_repeat; i++)
3102 		run_one_test(sched);
3103 
3104 	return 0;
3105 }
3106 
3107 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3108 			  const char * const usage_msg[])
3109 {
3110 	char *tmp, *tok, *str = strdup(sched->sort_order);
3111 
3112 	for (tok = strtok_r(str, ", ", &tmp);
3113 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3114 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3115 			usage_with_options_msg(usage_msg, options,
3116 					"Unknown --sort key: `%s'", tok);
3117 		}
3118 	}
3119 
3120 	free(str);
3121 
3122 	sort_dimension__add("pid", &sched->cmp_pid);
3123 }
3124 
3125 static int __cmd_record(int argc, const char **argv)
3126 {
3127 	unsigned int rec_argc, i, j;
3128 	const char **rec_argv;
3129 	const char * const record_args[] = {
3130 		"record",
3131 		"-a",
3132 		"-R",
3133 		"-m", "1024",
3134 		"-c", "1",
3135 		"-e", "sched:sched_switch",
3136 		"-e", "sched:sched_stat_wait",
3137 		"-e", "sched:sched_stat_sleep",
3138 		"-e", "sched:sched_stat_iowait",
3139 		"-e", "sched:sched_stat_runtime",
3140 		"-e", "sched:sched_process_fork",
3141 		"-e", "sched:sched_wakeup",
3142 		"-e", "sched:sched_wakeup_new",
3143 		"-e", "sched:sched_migrate_task",
3144 	};
3145 
3146 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
3147 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3148 
3149 	if (rec_argv == NULL)
3150 		return -ENOMEM;
3151 
3152 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3153 		rec_argv[i] = strdup(record_args[i]);
3154 
3155 	for (j = 1; j < (unsigned int)argc; j++, i++)
3156 		rec_argv[i] = argv[j];
3157 
3158 	BUG_ON(i != rec_argc);
3159 
3160 	return cmd_record(i, rec_argv, NULL);
3161 }
3162 
3163 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
3164 {
3165 	const char default_sort_order[] = "avg, max, switch, runtime";
3166 	struct perf_sched sched = {
3167 		.tool = {
3168 			.sample		 = perf_sched__process_tracepoint_sample,
3169 			.comm		 = perf_event__process_comm,
3170 			.lost		 = perf_event__process_lost,
3171 			.fork		 = perf_sched__process_fork_event,
3172 			.ordered_events = true,
3173 		},
3174 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3175 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3176 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
3177 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
3178 		.sort_order	      = default_sort_order,
3179 		.replay_repeat	      = 10,
3180 		.profile_cpu	      = -1,
3181 		.next_shortname1      = 'A',
3182 		.next_shortname2      = '0',
3183 		.skip_merge           = 0,
3184 		.show_callchain	      = 1,
3185 		.max_stack            = 5,
3186 	};
3187 	const struct option sched_options[] = {
3188 	OPT_STRING('i', "input", &input_name, "file",
3189 		    "input file name"),
3190 	OPT_INCR('v', "verbose", &verbose,
3191 		    "be more verbose (show symbol address, etc)"),
3192 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3193 		    "dump raw trace in ASCII"),
3194 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3195 	OPT_END()
3196 	};
3197 	const struct option latency_options[] = {
3198 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3199 		   "sort by key(s): runtime, switch, avg, max"),
3200 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3201 		    "CPU to profile on"),
3202 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3203 		    "latency stats per pid instead of per comm"),
3204 	OPT_PARENT(sched_options)
3205 	};
3206 	const struct option replay_options[] = {
3207 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3208 		     "repeat the workload replay N times (-1: infinite)"),
3209 	OPT_PARENT(sched_options)
3210 	};
3211 	const struct option map_options[] = {
3212 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3213 		    "map output in compact mode"),
3214 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3215 		   "highlight given pids in map"),
3216 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3217                     "highlight given CPUs in map"),
3218 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3219                     "display given CPUs in map"),
3220 	OPT_PARENT(sched_options)
3221 	};
3222 	const struct option timehist_options[] = {
3223 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3224 		   "file", "vmlinux pathname"),
3225 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3226 		   "file", "kallsyms pathname"),
3227 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3228 		    "Display call chains if present (default on)"),
3229 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3230 		   "Maximum number of functions to display backtrace."),
3231 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3232 		    "Look for files with symbols relative to this directory"),
3233 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3234 		    "Show only syscall summary with statistics"),
3235 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3236 		    "Show all syscalls and summary with statistics"),
3237 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3238 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3239 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3240 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3241 	OPT_STRING(0, "time", &sched.time_str, "str",
3242 		   "Time span for analysis (start,stop)"),
3243 	OPT_PARENT(sched_options)
3244 	};
3245 
3246 	const char * const latency_usage[] = {
3247 		"perf sched latency [<options>]",
3248 		NULL
3249 	};
3250 	const char * const replay_usage[] = {
3251 		"perf sched replay [<options>]",
3252 		NULL
3253 	};
3254 	const char * const map_usage[] = {
3255 		"perf sched map [<options>]",
3256 		NULL
3257 	};
3258 	const char * const timehist_usage[] = {
3259 		"perf sched timehist [<options>]",
3260 		NULL
3261 	};
3262 	const char *const sched_subcommands[] = { "record", "latency", "map",
3263 						  "replay", "script",
3264 						  "timehist", NULL };
3265 	const char *sched_usage[] = {
3266 		NULL,
3267 		NULL
3268 	};
3269 	struct trace_sched_handler lat_ops  = {
3270 		.wakeup_event	    = latency_wakeup_event,
3271 		.switch_event	    = latency_switch_event,
3272 		.runtime_event	    = latency_runtime_event,
3273 		.migrate_task_event = latency_migrate_task_event,
3274 	};
3275 	struct trace_sched_handler map_ops  = {
3276 		.switch_event	    = map_switch_event,
3277 	};
3278 	struct trace_sched_handler replay_ops  = {
3279 		.wakeup_event	    = replay_wakeup_event,
3280 		.switch_event	    = replay_switch_event,
3281 		.fork_event	    = replay_fork_event,
3282 	};
3283 	unsigned int i;
3284 
3285 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3286 		sched.curr_pid[i] = -1;
3287 
3288 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3289 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3290 	if (!argc)
3291 		usage_with_options(sched_usage, sched_options);
3292 
3293 	/*
3294 	 * Aliased to 'perf script' for now:
3295 	 */
3296 	if (!strcmp(argv[0], "script"))
3297 		return cmd_script(argc, argv, prefix);
3298 
3299 	if (!strncmp(argv[0], "rec", 3)) {
3300 		return __cmd_record(argc, argv);
3301 	} else if (!strncmp(argv[0], "lat", 3)) {
3302 		sched.tp_handler = &lat_ops;
3303 		if (argc > 1) {
3304 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3305 			if (argc)
3306 				usage_with_options(latency_usage, latency_options);
3307 		}
3308 		setup_sorting(&sched, latency_options, latency_usage);
3309 		return perf_sched__lat(&sched);
3310 	} else if (!strcmp(argv[0], "map")) {
3311 		if (argc) {
3312 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3313 			if (argc)
3314 				usage_with_options(map_usage, map_options);
3315 		}
3316 		sched.tp_handler = &map_ops;
3317 		setup_sorting(&sched, latency_options, latency_usage);
3318 		return perf_sched__map(&sched);
3319 	} else if (!strncmp(argv[0], "rep", 3)) {
3320 		sched.tp_handler = &replay_ops;
3321 		if (argc) {
3322 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3323 			if (argc)
3324 				usage_with_options(replay_usage, replay_options);
3325 		}
3326 		return perf_sched__replay(&sched);
3327 	} else if (!strcmp(argv[0], "timehist")) {
3328 		if (argc) {
3329 			argc = parse_options(argc, argv, timehist_options,
3330 					     timehist_usage, 0);
3331 			if (argc)
3332 				usage_with_options(timehist_usage, timehist_options);
3333 		}
3334 		if (sched.show_wakeups && sched.summary_only) {
3335 			pr_err(" Error: -s and -w are mutually exclusive.\n");
3336 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3337 			parse_options_usage(NULL, timehist_options, "w", true);
3338 			return -EINVAL;
3339 		}
3340 
3341 		return perf_sched__timehist(&sched);
3342 	} else {
3343 		usage_with_options(sched_usage, sched_options);
3344 	}
3345 
3346 	return 0;
3347 }
3348