xref: /openbmc/linux/tools/perf/builtin-sched.c (revision abb84c46)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf-sys.h"
4 
5 #include "util/cpumap.h"
6 #include "util/evlist.h"
7 #include "util/evsel.h"
8 #include "util/evsel_fprintf.h"
9 #include "util/mutex.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/cloexec.h"
16 #include "util/thread_map.h"
17 #include "util/color.h"
18 #include "util/stat.h"
19 #include "util/string2.h"
20 #include "util/callchain.h"
21 #include "util/time-utils.h"
22 
23 #include <subcmd/pager.h>
24 #include <subcmd/parse-options.h>
25 #include "util/trace-event.h"
26 
27 #include "util/debug.h"
28 #include "util/event.h"
29 #include "util/util.h"
30 
31 #include <linux/kernel.h>
32 #include <linux/log2.h>
33 #include <linux/zalloc.h>
34 #include <sys/prctl.h>
35 #include <sys/resource.h>
36 #include <inttypes.h>
37 
38 #include <errno.h>
39 #include <semaphore.h>
40 #include <pthread.h>
41 #include <math.h>
42 #include <api/fs/fs.h>
43 #include <perf/cpumap.h>
44 #include <linux/time64.h>
45 #include <linux/err.h>
46 
47 #include <linux/ctype.h>
48 
49 #define PR_SET_NAME		15               /* Set process name */
50 #define MAX_CPUS		4096
51 #define COMM_LEN		20
52 #define SYM_LEN			129
53 #define MAX_PID			1024000
54 
55 static const char *cpu_list;
56 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57 
58 struct sched_atom;
59 
60 struct task_desc {
61 	unsigned long		nr;
62 	unsigned long		pid;
63 	char			comm[COMM_LEN];
64 
65 	unsigned long		nr_events;
66 	unsigned long		curr_event;
67 	struct sched_atom	**atoms;
68 
69 	pthread_t		thread;
70 	sem_t			sleep_sem;
71 
72 	sem_t			ready_for_work;
73 	sem_t			work_done_sem;
74 
75 	u64			cpu_usage;
76 };
77 
78 enum sched_event_type {
79 	SCHED_EVENT_RUN,
80 	SCHED_EVENT_SLEEP,
81 	SCHED_EVENT_WAKEUP,
82 	SCHED_EVENT_MIGRATION,
83 };
84 
85 struct sched_atom {
86 	enum sched_event_type	type;
87 	int			specific_wait;
88 	u64			timestamp;
89 	u64			duration;
90 	unsigned long		nr;
91 	sem_t			*wait_sem;
92 	struct task_desc	*wakee;
93 };
94 
95 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
96 
97 /* task state bitmask, copied from include/linux/sched.h */
98 #define TASK_RUNNING		0
99 #define TASK_INTERRUPTIBLE	1
100 #define TASK_UNINTERRUPTIBLE	2
101 #define __TASK_STOPPED		4
102 #define __TASK_TRACED		8
103 /* in tsk->exit_state */
104 #define EXIT_DEAD		16
105 #define EXIT_ZOMBIE		32
106 #define EXIT_TRACE		(EXIT_ZOMBIE | EXIT_DEAD)
107 /* in tsk->state again */
108 #define TASK_DEAD		64
109 #define TASK_WAKEKILL		128
110 #define TASK_WAKING		256
111 #define TASK_PARKED		512
112 
113 enum thread_state {
114 	THREAD_SLEEPING = 0,
115 	THREAD_WAIT_CPU,
116 	THREAD_SCHED_IN,
117 	THREAD_IGNORE
118 };
119 
120 struct work_atom {
121 	struct list_head	list;
122 	enum thread_state	state;
123 	u64			sched_out_time;
124 	u64			wake_up_time;
125 	u64			sched_in_time;
126 	u64			runtime;
127 };
128 
129 struct work_atoms {
130 	struct list_head	work_list;
131 	struct thread		*thread;
132 	struct rb_node		node;
133 	u64			max_lat;
134 	u64			max_lat_start;
135 	u64			max_lat_end;
136 	u64			total_lat;
137 	u64			nb_atoms;
138 	u64			total_runtime;
139 	int			num_merged;
140 };
141 
142 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
143 
144 struct perf_sched;
145 
146 struct trace_sched_handler {
147 	int (*switch_event)(struct perf_sched *sched, struct evsel *evsel,
148 			    struct perf_sample *sample, struct machine *machine);
149 
150 	int (*runtime_event)(struct perf_sched *sched, struct evsel *evsel,
151 			     struct perf_sample *sample, struct machine *machine);
152 
153 	int (*wakeup_event)(struct perf_sched *sched, struct evsel *evsel,
154 			    struct perf_sample *sample, struct machine *machine);
155 
156 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
157 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
158 			  struct machine *machine);
159 
160 	int (*migrate_task_event)(struct perf_sched *sched,
161 				  struct evsel *evsel,
162 				  struct perf_sample *sample,
163 				  struct machine *machine);
164 };
165 
166 #define COLOR_PIDS PERF_COLOR_BLUE
167 #define COLOR_CPUS PERF_COLOR_BG_RED
168 
169 struct perf_sched_map {
170 	DECLARE_BITMAP(comp_cpus_mask, MAX_CPUS);
171 	struct perf_cpu		*comp_cpus;
172 	bool			 comp;
173 	struct perf_thread_map *color_pids;
174 	const char		*color_pids_str;
175 	struct perf_cpu_map	*color_cpus;
176 	const char		*color_cpus_str;
177 	struct perf_cpu_map	*cpus;
178 	const char		*cpus_str;
179 };
180 
181 struct perf_sched {
182 	struct perf_tool tool;
183 	const char	 *sort_order;
184 	unsigned long	 nr_tasks;
185 	struct task_desc **pid_to_task;
186 	struct task_desc **tasks;
187 	const struct trace_sched_handler *tp_handler;
188 	struct mutex	 start_work_mutex;
189 	struct mutex	 work_done_wait_mutex;
190 	int		 profile_cpu;
191 /*
192  * Track the current task - that way we can know whether there's any
193  * weird events, such as a task being switched away that is not current.
194  */
195 	struct perf_cpu	 max_cpu;
196 	u32		 *curr_pid;
197 	struct thread	 **curr_thread;
198 	char		 next_shortname1;
199 	char		 next_shortname2;
200 	unsigned int	 replay_repeat;
201 	unsigned long	 nr_run_events;
202 	unsigned long	 nr_sleep_events;
203 	unsigned long	 nr_wakeup_events;
204 	unsigned long	 nr_sleep_corrections;
205 	unsigned long	 nr_run_events_optimized;
206 	unsigned long	 targetless_wakeups;
207 	unsigned long	 multitarget_wakeups;
208 	unsigned long	 nr_runs;
209 	unsigned long	 nr_timestamps;
210 	unsigned long	 nr_unordered_timestamps;
211 	unsigned long	 nr_context_switch_bugs;
212 	unsigned long	 nr_events;
213 	unsigned long	 nr_lost_chunks;
214 	unsigned long	 nr_lost_events;
215 	u64		 run_measurement_overhead;
216 	u64		 sleep_measurement_overhead;
217 	u64		 start_time;
218 	u64		 cpu_usage;
219 	u64		 runavg_cpu_usage;
220 	u64		 parent_cpu_usage;
221 	u64		 runavg_parent_cpu_usage;
222 	u64		 sum_runtime;
223 	u64		 sum_fluct;
224 	u64		 run_avg;
225 	u64		 all_runtime;
226 	u64		 all_count;
227 	u64		 *cpu_last_switched;
228 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
229 	struct list_head sort_list, cmp_pid;
230 	bool force;
231 	bool skip_merge;
232 	struct perf_sched_map map;
233 
234 	/* options for timehist command */
235 	bool		summary;
236 	bool		summary_only;
237 	bool		idle_hist;
238 	bool		show_callchain;
239 	unsigned int	max_stack;
240 	bool		show_cpu_visual;
241 	bool		show_wakeups;
242 	bool		show_next;
243 	bool		show_migrations;
244 	bool		show_state;
245 	u64		skipped_samples;
246 	const char	*time_str;
247 	struct perf_time_interval ptime;
248 	struct perf_time_interval hist_time;
249 	volatile bool   thread_funcs_exit;
250 };
251 
252 /* per thread run time data */
253 struct thread_runtime {
254 	u64 last_time;      /* time of previous sched in/out event */
255 	u64 dt_run;         /* run time */
256 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
257 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
258 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
259 	u64 dt_delay;       /* time between wakeup and sched-in */
260 	u64 ready_to_run;   /* time of wakeup */
261 
262 	struct stats run_stats;
263 	u64 total_run_time;
264 	u64 total_sleep_time;
265 	u64 total_iowait_time;
266 	u64 total_preempt_time;
267 	u64 total_delay_time;
268 
269 	int last_state;
270 
271 	char shortname[3];
272 	bool comm_changed;
273 
274 	u64 migrations;
275 };
276 
277 /* per event run time data */
278 struct evsel_runtime {
279 	u64 *last_time; /* time this event was last seen per cpu */
280 	u32 ncpu;       /* highest cpu slot allocated */
281 };
282 
283 /* per cpu idle time data */
284 struct idle_thread_runtime {
285 	struct thread_runtime	tr;
286 	struct thread		*last_thread;
287 	struct rb_root_cached	sorted_root;
288 	struct callchain_root	callchain;
289 	struct callchain_cursor	cursor;
290 };
291 
292 /* track idle times per cpu */
293 static struct thread **idle_threads;
294 static int idle_max_cpu;
295 static char idle_comm[] = "<idle>";
296 
297 static u64 get_nsecs(void)
298 {
299 	struct timespec ts;
300 
301 	clock_gettime(CLOCK_MONOTONIC, &ts);
302 
303 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
304 }
305 
306 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
307 {
308 	u64 T0 = get_nsecs(), T1;
309 
310 	do {
311 		T1 = get_nsecs();
312 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
313 }
314 
315 static void sleep_nsecs(u64 nsecs)
316 {
317 	struct timespec ts;
318 
319 	ts.tv_nsec = nsecs % 999999999;
320 	ts.tv_sec = nsecs / 999999999;
321 
322 	nanosleep(&ts, NULL);
323 }
324 
325 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
326 {
327 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 	int i;
329 
330 	for (i = 0; i < 10; i++) {
331 		T0 = get_nsecs();
332 		burn_nsecs(sched, 0);
333 		T1 = get_nsecs();
334 		delta = T1-T0;
335 		min_delta = min(min_delta, delta);
336 	}
337 	sched->run_measurement_overhead = min_delta;
338 
339 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340 }
341 
342 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
343 {
344 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
345 	int i;
346 
347 	for (i = 0; i < 10; i++) {
348 		T0 = get_nsecs();
349 		sleep_nsecs(10000);
350 		T1 = get_nsecs();
351 		delta = T1-T0;
352 		min_delta = min(min_delta, delta);
353 	}
354 	min_delta -= 10000;
355 	sched->sleep_measurement_overhead = min_delta;
356 
357 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
358 }
359 
360 static struct sched_atom *
361 get_new_event(struct task_desc *task, u64 timestamp)
362 {
363 	struct sched_atom *event = zalloc(sizeof(*event));
364 	unsigned long idx = task->nr_events;
365 	size_t size;
366 
367 	event->timestamp = timestamp;
368 	event->nr = idx;
369 
370 	task->nr_events++;
371 	size = sizeof(struct sched_atom *) * task->nr_events;
372 	task->atoms = realloc(task->atoms, size);
373 	BUG_ON(!task->atoms);
374 
375 	task->atoms[idx] = event;
376 
377 	return event;
378 }
379 
380 static struct sched_atom *last_event(struct task_desc *task)
381 {
382 	if (!task->nr_events)
383 		return NULL;
384 
385 	return task->atoms[task->nr_events - 1];
386 }
387 
388 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
389 				u64 timestamp, u64 duration)
390 {
391 	struct sched_atom *event, *curr_event = last_event(task);
392 
393 	/*
394 	 * optimize an existing RUN event by merging this one
395 	 * to it:
396 	 */
397 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
398 		sched->nr_run_events_optimized++;
399 		curr_event->duration += duration;
400 		return;
401 	}
402 
403 	event = get_new_event(task, timestamp);
404 
405 	event->type = SCHED_EVENT_RUN;
406 	event->duration = duration;
407 
408 	sched->nr_run_events++;
409 }
410 
411 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
412 				   u64 timestamp, struct task_desc *wakee)
413 {
414 	struct sched_atom *event, *wakee_event;
415 
416 	event = get_new_event(task, timestamp);
417 	event->type = SCHED_EVENT_WAKEUP;
418 	event->wakee = wakee;
419 
420 	wakee_event = last_event(wakee);
421 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
422 		sched->targetless_wakeups++;
423 		return;
424 	}
425 	if (wakee_event->wait_sem) {
426 		sched->multitarget_wakeups++;
427 		return;
428 	}
429 
430 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
431 	sem_init(wakee_event->wait_sem, 0, 0);
432 	wakee_event->specific_wait = 1;
433 	event->wait_sem = wakee_event->wait_sem;
434 
435 	sched->nr_wakeup_events++;
436 }
437 
438 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
439 				  u64 timestamp, u64 task_state __maybe_unused)
440 {
441 	struct sched_atom *event = get_new_event(task, timestamp);
442 
443 	event->type = SCHED_EVENT_SLEEP;
444 
445 	sched->nr_sleep_events++;
446 }
447 
448 static struct task_desc *register_pid(struct perf_sched *sched,
449 				      unsigned long pid, const char *comm)
450 {
451 	struct task_desc *task;
452 	static int pid_max;
453 
454 	if (sched->pid_to_task == NULL) {
455 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
456 			pid_max = MAX_PID;
457 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
458 	}
459 	if (pid >= (unsigned long)pid_max) {
460 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
461 			sizeof(struct task_desc *))) == NULL);
462 		while (pid >= (unsigned long)pid_max)
463 			sched->pid_to_task[pid_max++] = NULL;
464 	}
465 
466 	task = sched->pid_to_task[pid];
467 
468 	if (task)
469 		return task;
470 
471 	task = zalloc(sizeof(*task));
472 	task->pid = pid;
473 	task->nr = sched->nr_tasks;
474 	strcpy(task->comm, comm);
475 	/*
476 	 * every task starts in sleeping state - this gets ignored
477 	 * if there's no wakeup pointing to this sleep state:
478 	 */
479 	add_sched_event_sleep(sched, task, 0, 0);
480 
481 	sched->pid_to_task[pid] = task;
482 	sched->nr_tasks++;
483 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
484 	BUG_ON(!sched->tasks);
485 	sched->tasks[task->nr] = task;
486 
487 	if (verbose > 0)
488 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
489 
490 	return task;
491 }
492 
493 
494 static void print_task_traces(struct perf_sched *sched)
495 {
496 	struct task_desc *task;
497 	unsigned long i;
498 
499 	for (i = 0; i < sched->nr_tasks; i++) {
500 		task = sched->tasks[i];
501 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
502 			task->nr, task->comm, task->pid, task->nr_events);
503 	}
504 }
505 
506 static void add_cross_task_wakeups(struct perf_sched *sched)
507 {
508 	struct task_desc *task1, *task2;
509 	unsigned long i, j;
510 
511 	for (i = 0; i < sched->nr_tasks; i++) {
512 		task1 = sched->tasks[i];
513 		j = i + 1;
514 		if (j == sched->nr_tasks)
515 			j = 0;
516 		task2 = sched->tasks[j];
517 		add_sched_event_wakeup(sched, task1, 0, task2);
518 	}
519 }
520 
521 static void perf_sched__process_event(struct perf_sched *sched,
522 				      struct sched_atom *atom)
523 {
524 	int ret = 0;
525 
526 	switch (atom->type) {
527 		case SCHED_EVENT_RUN:
528 			burn_nsecs(sched, atom->duration);
529 			break;
530 		case SCHED_EVENT_SLEEP:
531 			if (atom->wait_sem)
532 				ret = sem_wait(atom->wait_sem);
533 			BUG_ON(ret);
534 			break;
535 		case SCHED_EVENT_WAKEUP:
536 			if (atom->wait_sem)
537 				ret = sem_post(atom->wait_sem);
538 			BUG_ON(ret);
539 			break;
540 		case SCHED_EVENT_MIGRATION:
541 			break;
542 		default:
543 			BUG_ON(1);
544 	}
545 }
546 
547 static u64 get_cpu_usage_nsec_parent(void)
548 {
549 	struct rusage ru;
550 	u64 sum;
551 	int err;
552 
553 	err = getrusage(RUSAGE_SELF, &ru);
554 	BUG_ON(err);
555 
556 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
557 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
558 
559 	return sum;
560 }
561 
562 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
563 {
564 	struct perf_event_attr attr;
565 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
566 	int fd;
567 	struct rlimit limit;
568 	bool need_privilege = false;
569 
570 	memset(&attr, 0, sizeof(attr));
571 
572 	attr.type = PERF_TYPE_SOFTWARE;
573 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
574 
575 force_again:
576 	fd = sys_perf_event_open(&attr, 0, -1, -1,
577 				 perf_event_open_cloexec_flag());
578 
579 	if (fd < 0) {
580 		if (errno == EMFILE) {
581 			if (sched->force) {
582 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
583 				limit.rlim_cur += sched->nr_tasks - cur_task;
584 				if (limit.rlim_cur > limit.rlim_max) {
585 					limit.rlim_max = limit.rlim_cur;
586 					need_privilege = true;
587 				}
588 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
589 					if (need_privilege && errno == EPERM)
590 						strcpy(info, "Need privilege\n");
591 				} else
592 					goto force_again;
593 			} else
594 				strcpy(info, "Have a try with -f option\n");
595 		}
596 		pr_err("Error: sys_perf_event_open() syscall returned "
597 		       "with %d (%s)\n%s", fd,
598 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
599 		exit(EXIT_FAILURE);
600 	}
601 	return fd;
602 }
603 
604 static u64 get_cpu_usage_nsec_self(int fd)
605 {
606 	u64 runtime;
607 	int ret;
608 
609 	ret = read(fd, &runtime, sizeof(runtime));
610 	BUG_ON(ret != sizeof(runtime));
611 
612 	return runtime;
613 }
614 
615 struct sched_thread_parms {
616 	struct task_desc  *task;
617 	struct perf_sched *sched;
618 	int fd;
619 };
620 
621 static void *thread_func(void *ctx)
622 {
623 	struct sched_thread_parms *parms = ctx;
624 	struct task_desc *this_task = parms->task;
625 	struct perf_sched *sched = parms->sched;
626 	u64 cpu_usage_0, cpu_usage_1;
627 	unsigned long i, ret;
628 	char comm2[22];
629 	int fd = parms->fd;
630 
631 	zfree(&parms);
632 
633 	sprintf(comm2, ":%s", this_task->comm);
634 	prctl(PR_SET_NAME, comm2);
635 	if (fd < 0)
636 		return NULL;
637 
638 	while (!sched->thread_funcs_exit) {
639 		ret = sem_post(&this_task->ready_for_work);
640 		BUG_ON(ret);
641 		mutex_lock(&sched->start_work_mutex);
642 		mutex_unlock(&sched->start_work_mutex);
643 
644 		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
645 
646 		for (i = 0; i < this_task->nr_events; i++) {
647 			this_task->curr_event = i;
648 			perf_sched__process_event(sched, this_task->atoms[i]);
649 		}
650 
651 		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
652 		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
653 		ret = sem_post(&this_task->work_done_sem);
654 		BUG_ON(ret);
655 
656 		mutex_lock(&sched->work_done_wait_mutex);
657 		mutex_unlock(&sched->work_done_wait_mutex);
658 	}
659 	return NULL;
660 }
661 
662 static void create_tasks(struct perf_sched *sched)
663 	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
664 	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
665 {
666 	struct task_desc *task;
667 	pthread_attr_t attr;
668 	unsigned long i;
669 	int err;
670 
671 	err = pthread_attr_init(&attr);
672 	BUG_ON(err);
673 	err = pthread_attr_setstacksize(&attr,
674 			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
675 	BUG_ON(err);
676 	mutex_lock(&sched->start_work_mutex);
677 	mutex_lock(&sched->work_done_wait_mutex);
678 	for (i = 0; i < sched->nr_tasks; i++) {
679 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
680 		BUG_ON(parms == NULL);
681 		parms->task = task = sched->tasks[i];
682 		parms->sched = sched;
683 		parms->fd = self_open_counters(sched, i);
684 		sem_init(&task->sleep_sem, 0, 0);
685 		sem_init(&task->ready_for_work, 0, 0);
686 		sem_init(&task->work_done_sem, 0, 0);
687 		task->curr_event = 0;
688 		err = pthread_create(&task->thread, &attr, thread_func, parms);
689 		BUG_ON(err);
690 	}
691 }
692 
693 static void destroy_tasks(struct perf_sched *sched)
694 	UNLOCK_FUNCTION(sched->start_work_mutex)
695 	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
696 {
697 	struct task_desc *task;
698 	unsigned long i;
699 	int err;
700 
701 	mutex_unlock(&sched->start_work_mutex);
702 	mutex_unlock(&sched->work_done_wait_mutex);
703 	/* Get rid of threads so they won't be upset by mutex destrunction */
704 	for (i = 0; i < sched->nr_tasks; i++) {
705 		task = sched->tasks[i];
706 		err = pthread_join(task->thread, NULL);
707 		BUG_ON(err);
708 		sem_destroy(&task->sleep_sem);
709 		sem_destroy(&task->ready_for_work);
710 		sem_destroy(&task->work_done_sem);
711 	}
712 }
713 
714 static void wait_for_tasks(struct perf_sched *sched)
715 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
716 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
717 {
718 	u64 cpu_usage_0, cpu_usage_1;
719 	struct task_desc *task;
720 	unsigned long i, ret;
721 
722 	sched->start_time = get_nsecs();
723 	sched->cpu_usage = 0;
724 	mutex_unlock(&sched->work_done_wait_mutex);
725 
726 	for (i = 0; i < sched->nr_tasks; i++) {
727 		task = sched->tasks[i];
728 		ret = sem_wait(&task->ready_for_work);
729 		BUG_ON(ret);
730 		sem_init(&task->ready_for_work, 0, 0);
731 	}
732 	mutex_lock(&sched->work_done_wait_mutex);
733 
734 	cpu_usage_0 = get_cpu_usage_nsec_parent();
735 
736 	mutex_unlock(&sched->start_work_mutex);
737 
738 	for (i = 0; i < sched->nr_tasks; i++) {
739 		task = sched->tasks[i];
740 		ret = sem_wait(&task->work_done_sem);
741 		BUG_ON(ret);
742 		sem_init(&task->work_done_sem, 0, 0);
743 		sched->cpu_usage += task->cpu_usage;
744 		task->cpu_usage = 0;
745 	}
746 
747 	cpu_usage_1 = get_cpu_usage_nsec_parent();
748 	if (!sched->runavg_cpu_usage)
749 		sched->runavg_cpu_usage = sched->cpu_usage;
750 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
751 
752 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
753 	if (!sched->runavg_parent_cpu_usage)
754 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
755 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
756 					 sched->parent_cpu_usage)/sched->replay_repeat;
757 
758 	mutex_lock(&sched->start_work_mutex);
759 
760 	for (i = 0; i < sched->nr_tasks; i++) {
761 		task = sched->tasks[i];
762 		sem_init(&task->sleep_sem, 0, 0);
763 		task->curr_event = 0;
764 	}
765 }
766 
767 static void run_one_test(struct perf_sched *sched)
768 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
769 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
770 {
771 	u64 T0, T1, delta, avg_delta, fluct;
772 
773 	T0 = get_nsecs();
774 	wait_for_tasks(sched);
775 	T1 = get_nsecs();
776 
777 	delta = T1 - T0;
778 	sched->sum_runtime += delta;
779 	sched->nr_runs++;
780 
781 	avg_delta = sched->sum_runtime / sched->nr_runs;
782 	if (delta < avg_delta)
783 		fluct = avg_delta - delta;
784 	else
785 		fluct = delta - avg_delta;
786 	sched->sum_fluct += fluct;
787 	if (!sched->run_avg)
788 		sched->run_avg = delta;
789 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
790 
791 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
792 
793 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
794 
795 	printf("cpu: %0.2f / %0.2f",
796 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
797 
798 #if 0
799 	/*
800 	 * rusage statistics done by the parent, these are less
801 	 * accurate than the sched->sum_exec_runtime based statistics:
802 	 */
803 	printf(" [%0.2f / %0.2f]",
804 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
805 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
806 #endif
807 
808 	printf("\n");
809 
810 	if (sched->nr_sleep_corrections)
811 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
812 	sched->nr_sleep_corrections = 0;
813 }
814 
815 static void test_calibrations(struct perf_sched *sched)
816 {
817 	u64 T0, T1;
818 
819 	T0 = get_nsecs();
820 	burn_nsecs(sched, NSEC_PER_MSEC);
821 	T1 = get_nsecs();
822 
823 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
824 
825 	T0 = get_nsecs();
826 	sleep_nsecs(NSEC_PER_MSEC);
827 	T1 = get_nsecs();
828 
829 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
830 }
831 
832 static int
833 replay_wakeup_event(struct perf_sched *sched,
834 		    struct evsel *evsel, struct perf_sample *sample,
835 		    struct machine *machine __maybe_unused)
836 {
837 	const char *comm = evsel__strval(evsel, sample, "comm");
838 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
839 	struct task_desc *waker, *wakee;
840 
841 	if (verbose > 0) {
842 		printf("sched_wakeup event %p\n", evsel);
843 
844 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
845 	}
846 
847 	waker = register_pid(sched, sample->tid, "<unknown>");
848 	wakee = register_pid(sched, pid, comm);
849 
850 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
851 	return 0;
852 }
853 
854 static int replay_switch_event(struct perf_sched *sched,
855 			       struct evsel *evsel,
856 			       struct perf_sample *sample,
857 			       struct machine *machine __maybe_unused)
858 {
859 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
860 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
861 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
862 		  next_pid = evsel__intval(evsel, sample, "next_pid");
863 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
864 	struct task_desc *prev, __maybe_unused *next;
865 	u64 timestamp0, timestamp = sample->time;
866 	int cpu = sample->cpu;
867 	s64 delta;
868 
869 	if (verbose > 0)
870 		printf("sched_switch event %p\n", evsel);
871 
872 	if (cpu >= MAX_CPUS || cpu < 0)
873 		return 0;
874 
875 	timestamp0 = sched->cpu_last_switched[cpu];
876 	if (timestamp0)
877 		delta = timestamp - timestamp0;
878 	else
879 		delta = 0;
880 
881 	if (delta < 0) {
882 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
883 		return -1;
884 	}
885 
886 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
887 		 prev_comm, prev_pid, next_comm, next_pid, delta);
888 
889 	prev = register_pid(sched, prev_pid, prev_comm);
890 	next = register_pid(sched, next_pid, next_comm);
891 
892 	sched->cpu_last_switched[cpu] = timestamp;
893 
894 	add_sched_event_run(sched, prev, timestamp, delta);
895 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
896 
897 	return 0;
898 }
899 
900 static int replay_fork_event(struct perf_sched *sched,
901 			     union perf_event *event,
902 			     struct machine *machine)
903 {
904 	struct thread *child, *parent;
905 
906 	child = machine__findnew_thread(machine, event->fork.pid,
907 					event->fork.tid);
908 	parent = machine__findnew_thread(machine, event->fork.ppid,
909 					 event->fork.ptid);
910 
911 	if (child == NULL || parent == NULL) {
912 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
913 				 child, parent);
914 		goto out_put;
915 	}
916 
917 	if (verbose > 0) {
918 		printf("fork event\n");
919 		printf("... parent: %s/%d\n", thread__comm_str(parent), thread__tid(parent));
920 		printf("...  child: %s/%d\n", thread__comm_str(child), thread__tid(child));
921 	}
922 
923 	register_pid(sched, thread__tid(parent), thread__comm_str(parent));
924 	register_pid(sched, thread__tid(child), thread__comm_str(child));
925 out_put:
926 	thread__put(child);
927 	thread__put(parent);
928 	return 0;
929 }
930 
931 struct sort_dimension {
932 	const char		*name;
933 	sort_fn_t		cmp;
934 	struct list_head	list;
935 };
936 
937 /*
938  * handle runtime stats saved per thread
939  */
940 static struct thread_runtime *thread__init_runtime(struct thread *thread)
941 {
942 	struct thread_runtime *r;
943 
944 	r = zalloc(sizeof(struct thread_runtime));
945 	if (!r)
946 		return NULL;
947 
948 	init_stats(&r->run_stats);
949 	thread__set_priv(thread, r);
950 
951 	return r;
952 }
953 
954 static struct thread_runtime *thread__get_runtime(struct thread *thread)
955 {
956 	struct thread_runtime *tr;
957 
958 	tr = thread__priv(thread);
959 	if (tr == NULL) {
960 		tr = thread__init_runtime(thread);
961 		if (tr == NULL)
962 			pr_debug("Failed to malloc memory for runtime data.\n");
963 	}
964 
965 	return tr;
966 }
967 
968 static int
969 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
970 {
971 	struct sort_dimension *sort;
972 	int ret = 0;
973 
974 	BUG_ON(list_empty(list));
975 
976 	list_for_each_entry(sort, list, list) {
977 		ret = sort->cmp(l, r);
978 		if (ret)
979 			return ret;
980 	}
981 
982 	return ret;
983 }
984 
985 static struct work_atoms *
986 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
987 			 struct list_head *sort_list)
988 {
989 	struct rb_node *node = root->rb_root.rb_node;
990 	struct work_atoms key = { .thread = thread };
991 
992 	while (node) {
993 		struct work_atoms *atoms;
994 		int cmp;
995 
996 		atoms = container_of(node, struct work_atoms, node);
997 
998 		cmp = thread_lat_cmp(sort_list, &key, atoms);
999 		if (cmp > 0)
1000 			node = node->rb_left;
1001 		else if (cmp < 0)
1002 			node = node->rb_right;
1003 		else {
1004 			BUG_ON(thread != atoms->thread);
1005 			return atoms;
1006 		}
1007 	}
1008 	return NULL;
1009 }
1010 
1011 static void
1012 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1013 			 struct list_head *sort_list)
1014 {
1015 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1016 	bool leftmost = true;
1017 
1018 	while (*new) {
1019 		struct work_atoms *this;
1020 		int cmp;
1021 
1022 		this = container_of(*new, struct work_atoms, node);
1023 		parent = *new;
1024 
1025 		cmp = thread_lat_cmp(sort_list, data, this);
1026 
1027 		if (cmp > 0)
1028 			new = &((*new)->rb_left);
1029 		else {
1030 			new = &((*new)->rb_right);
1031 			leftmost = false;
1032 		}
1033 	}
1034 
1035 	rb_link_node(&data->node, parent, new);
1036 	rb_insert_color_cached(&data->node, root, leftmost);
1037 }
1038 
1039 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1040 {
1041 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1042 	if (!atoms) {
1043 		pr_err("No memory at %s\n", __func__);
1044 		return -1;
1045 	}
1046 
1047 	atoms->thread = thread__get(thread);
1048 	INIT_LIST_HEAD(&atoms->work_list);
1049 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1050 	return 0;
1051 }
1052 
1053 static char sched_out_state(u64 prev_state)
1054 {
1055 	const char *str = TASK_STATE_TO_CHAR_STR;
1056 
1057 	return str[prev_state];
1058 }
1059 
1060 static int
1061 add_sched_out_event(struct work_atoms *atoms,
1062 		    char run_state,
1063 		    u64 timestamp)
1064 {
1065 	struct work_atom *atom = zalloc(sizeof(*atom));
1066 	if (!atom) {
1067 		pr_err("Non memory at %s", __func__);
1068 		return -1;
1069 	}
1070 
1071 	atom->sched_out_time = timestamp;
1072 
1073 	if (run_state == 'R') {
1074 		atom->state = THREAD_WAIT_CPU;
1075 		atom->wake_up_time = atom->sched_out_time;
1076 	}
1077 
1078 	list_add_tail(&atom->list, &atoms->work_list);
1079 	return 0;
1080 }
1081 
1082 static void
1083 add_runtime_event(struct work_atoms *atoms, u64 delta,
1084 		  u64 timestamp __maybe_unused)
1085 {
1086 	struct work_atom *atom;
1087 
1088 	BUG_ON(list_empty(&atoms->work_list));
1089 
1090 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091 
1092 	atom->runtime += delta;
1093 	atoms->total_runtime += delta;
1094 }
1095 
1096 static void
1097 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1098 {
1099 	struct work_atom *atom;
1100 	u64 delta;
1101 
1102 	if (list_empty(&atoms->work_list))
1103 		return;
1104 
1105 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1106 
1107 	if (atom->state != THREAD_WAIT_CPU)
1108 		return;
1109 
1110 	if (timestamp < atom->wake_up_time) {
1111 		atom->state = THREAD_IGNORE;
1112 		return;
1113 	}
1114 
1115 	atom->state = THREAD_SCHED_IN;
1116 	atom->sched_in_time = timestamp;
1117 
1118 	delta = atom->sched_in_time - atom->wake_up_time;
1119 	atoms->total_lat += delta;
1120 	if (delta > atoms->max_lat) {
1121 		atoms->max_lat = delta;
1122 		atoms->max_lat_start = atom->wake_up_time;
1123 		atoms->max_lat_end = timestamp;
1124 	}
1125 	atoms->nb_atoms++;
1126 }
1127 
1128 static int latency_switch_event(struct perf_sched *sched,
1129 				struct evsel *evsel,
1130 				struct perf_sample *sample,
1131 				struct machine *machine)
1132 {
1133 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1134 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1135 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1136 	struct work_atoms *out_events, *in_events;
1137 	struct thread *sched_out, *sched_in;
1138 	u64 timestamp0, timestamp = sample->time;
1139 	int cpu = sample->cpu, err = -1;
1140 	s64 delta;
1141 
1142 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1143 
1144 	timestamp0 = sched->cpu_last_switched[cpu];
1145 	sched->cpu_last_switched[cpu] = timestamp;
1146 	if (timestamp0)
1147 		delta = timestamp - timestamp0;
1148 	else
1149 		delta = 0;
1150 
1151 	if (delta < 0) {
1152 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1153 		return -1;
1154 	}
1155 
1156 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1157 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1158 	if (sched_out == NULL || sched_in == NULL)
1159 		goto out_put;
1160 
1161 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1162 	if (!out_events) {
1163 		if (thread_atoms_insert(sched, sched_out))
1164 			goto out_put;
1165 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1166 		if (!out_events) {
1167 			pr_err("out-event: Internal tree error");
1168 			goto out_put;
1169 		}
1170 	}
1171 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1172 		return -1;
1173 
1174 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1175 	if (!in_events) {
1176 		if (thread_atoms_insert(sched, sched_in))
1177 			goto out_put;
1178 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1179 		if (!in_events) {
1180 			pr_err("in-event: Internal tree error");
1181 			goto out_put;
1182 		}
1183 		/*
1184 		 * Take came in we have not heard about yet,
1185 		 * add in an initial atom in runnable state:
1186 		 */
1187 		if (add_sched_out_event(in_events, 'R', timestamp))
1188 			goto out_put;
1189 	}
1190 	add_sched_in_event(in_events, timestamp);
1191 	err = 0;
1192 out_put:
1193 	thread__put(sched_out);
1194 	thread__put(sched_in);
1195 	return err;
1196 }
1197 
1198 static int latency_runtime_event(struct perf_sched *sched,
1199 				 struct evsel *evsel,
1200 				 struct perf_sample *sample,
1201 				 struct machine *machine)
1202 {
1203 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1204 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1205 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1206 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1207 	u64 timestamp = sample->time;
1208 	int cpu = sample->cpu, err = -1;
1209 
1210 	if (thread == NULL)
1211 		return -1;
1212 
1213 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1214 	if (!atoms) {
1215 		if (thread_atoms_insert(sched, thread))
1216 			goto out_put;
1217 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1218 		if (!atoms) {
1219 			pr_err("in-event: Internal tree error");
1220 			goto out_put;
1221 		}
1222 		if (add_sched_out_event(atoms, 'R', timestamp))
1223 			goto out_put;
1224 	}
1225 
1226 	add_runtime_event(atoms, runtime, timestamp);
1227 	err = 0;
1228 out_put:
1229 	thread__put(thread);
1230 	return err;
1231 }
1232 
1233 static int latency_wakeup_event(struct perf_sched *sched,
1234 				struct evsel *evsel,
1235 				struct perf_sample *sample,
1236 				struct machine *machine)
1237 {
1238 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1239 	struct work_atoms *atoms;
1240 	struct work_atom *atom;
1241 	struct thread *wakee;
1242 	u64 timestamp = sample->time;
1243 	int err = -1;
1244 
1245 	wakee = machine__findnew_thread(machine, -1, pid);
1246 	if (wakee == NULL)
1247 		return -1;
1248 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1249 	if (!atoms) {
1250 		if (thread_atoms_insert(sched, wakee))
1251 			goto out_put;
1252 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1253 		if (!atoms) {
1254 			pr_err("wakeup-event: Internal tree error");
1255 			goto out_put;
1256 		}
1257 		if (add_sched_out_event(atoms, 'S', timestamp))
1258 			goto out_put;
1259 	}
1260 
1261 	BUG_ON(list_empty(&atoms->work_list));
1262 
1263 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1264 
1265 	/*
1266 	 * As we do not guarantee the wakeup event happens when
1267 	 * task is out of run queue, also may happen when task is
1268 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1269 	 * then we should not set the ->wake_up_time when wake up a
1270 	 * task which is on run queue.
1271 	 *
1272 	 * You WILL be missing events if you've recorded only
1273 	 * one CPU, or are only looking at only one, so don't
1274 	 * skip in this case.
1275 	 */
1276 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1277 		goto out_ok;
1278 
1279 	sched->nr_timestamps++;
1280 	if (atom->sched_out_time > timestamp) {
1281 		sched->nr_unordered_timestamps++;
1282 		goto out_ok;
1283 	}
1284 
1285 	atom->state = THREAD_WAIT_CPU;
1286 	atom->wake_up_time = timestamp;
1287 out_ok:
1288 	err = 0;
1289 out_put:
1290 	thread__put(wakee);
1291 	return err;
1292 }
1293 
1294 static int latency_migrate_task_event(struct perf_sched *sched,
1295 				      struct evsel *evsel,
1296 				      struct perf_sample *sample,
1297 				      struct machine *machine)
1298 {
1299 	const u32 pid = evsel__intval(evsel, sample, "pid");
1300 	u64 timestamp = sample->time;
1301 	struct work_atoms *atoms;
1302 	struct work_atom *atom;
1303 	struct thread *migrant;
1304 	int err = -1;
1305 
1306 	/*
1307 	 * Only need to worry about migration when profiling one CPU.
1308 	 */
1309 	if (sched->profile_cpu == -1)
1310 		return 0;
1311 
1312 	migrant = machine__findnew_thread(machine, -1, pid);
1313 	if (migrant == NULL)
1314 		return -1;
1315 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1316 	if (!atoms) {
1317 		if (thread_atoms_insert(sched, migrant))
1318 			goto out_put;
1319 		register_pid(sched, thread__tid(migrant), thread__comm_str(migrant));
1320 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1321 		if (!atoms) {
1322 			pr_err("migration-event: Internal tree error");
1323 			goto out_put;
1324 		}
1325 		if (add_sched_out_event(atoms, 'R', timestamp))
1326 			goto out_put;
1327 	}
1328 
1329 	BUG_ON(list_empty(&atoms->work_list));
1330 
1331 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1332 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1333 
1334 	sched->nr_timestamps++;
1335 
1336 	if (atom->sched_out_time > timestamp)
1337 		sched->nr_unordered_timestamps++;
1338 	err = 0;
1339 out_put:
1340 	thread__put(migrant);
1341 	return err;
1342 }
1343 
1344 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1345 {
1346 	int i;
1347 	int ret;
1348 	u64 avg;
1349 	char max_lat_start[32], max_lat_end[32];
1350 
1351 	if (!work_list->nb_atoms)
1352 		return;
1353 	/*
1354 	 * Ignore idle threads:
1355 	 */
1356 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1357 		return;
1358 
1359 	sched->all_runtime += work_list->total_runtime;
1360 	sched->all_count   += work_list->nb_atoms;
1361 
1362 	if (work_list->num_merged > 1) {
1363 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread),
1364 			     work_list->num_merged);
1365 	} else {
1366 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread),
1367 			     thread__tid(work_list->thread));
1368 	}
1369 
1370 	for (i = 0; i < 24 - ret; i++)
1371 		printf(" ");
1372 
1373 	avg = work_list->total_lat / work_list->nb_atoms;
1374 	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1375 	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1376 
1377 	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1378 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1379 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1380 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1381 		 max_lat_start, max_lat_end);
1382 }
1383 
1384 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1385 {
1386 	pid_t l_tid, r_tid;
1387 
1388 	if (RC_CHK_ACCESS(l->thread) == RC_CHK_ACCESS(r->thread))
1389 		return 0;
1390 	l_tid = thread__tid(l->thread);
1391 	r_tid = thread__tid(r->thread);
1392 	if (l_tid < r_tid)
1393 		return -1;
1394 	if (l_tid > r_tid)
1395 		return 1;
1396 	return (int)(RC_CHK_ACCESS(l->thread) - RC_CHK_ACCESS(r->thread));
1397 }
1398 
1399 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1400 {
1401 	u64 avgl, avgr;
1402 
1403 	if (!l->nb_atoms)
1404 		return -1;
1405 
1406 	if (!r->nb_atoms)
1407 		return 1;
1408 
1409 	avgl = l->total_lat / l->nb_atoms;
1410 	avgr = r->total_lat / r->nb_atoms;
1411 
1412 	if (avgl < avgr)
1413 		return -1;
1414 	if (avgl > avgr)
1415 		return 1;
1416 
1417 	return 0;
1418 }
1419 
1420 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1421 {
1422 	if (l->max_lat < r->max_lat)
1423 		return -1;
1424 	if (l->max_lat > r->max_lat)
1425 		return 1;
1426 
1427 	return 0;
1428 }
1429 
1430 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1431 {
1432 	if (l->nb_atoms < r->nb_atoms)
1433 		return -1;
1434 	if (l->nb_atoms > r->nb_atoms)
1435 		return 1;
1436 
1437 	return 0;
1438 }
1439 
1440 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1441 {
1442 	if (l->total_runtime < r->total_runtime)
1443 		return -1;
1444 	if (l->total_runtime > r->total_runtime)
1445 		return 1;
1446 
1447 	return 0;
1448 }
1449 
1450 static int sort_dimension__add(const char *tok, struct list_head *list)
1451 {
1452 	size_t i;
1453 	static struct sort_dimension avg_sort_dimension = {
1454 		.name = "avg",
1455 		.cmp  = avg_cmp,
1456 	};
1457 	static struct sort_dimension max_sort_dimension = {
1458 		.name = "max",
1459 		.cmp  = max_cmp,
1460 	};
1461 	static struct sort_dimension pid_sort_dimension = {
1462 		.name = "pid",
1463 		.cmp  = pid_cmp,
1464 	};
1465 	static struct sort_dimension runtime_sort_dimension = {
1466 		.name = "runtime",
1467 		.cmp  = runtime_cmp,
1468 	};
1469 	static struct sort_dimension switch_sort_dimension = {
1470 		.name = "switch",
1471 		.cmp  = switch_cmp,
1472 	};
1473 	struct sort_dimension *available_sorts[] = {
1474 		&pid_sort_dimension,
1475 		&avg_sort_dimension,
1476 		&max_sort_dimension,
1477 		&switch_sort_dimension,
1478 		&runtime_sort_dimension,
1479 	};
1480 
1481 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1482 		if (!strcmp(available_sorts[i]->name, tok)) {
1483 			list_add_tail(&available_sorts[i]->list, list);
1484 
1485 			return 0;
1486 		}
1487 	}
1488 
1489 	return -1;
1490 }
1491 
1492 static void perf_sched__sort_lat(struct perf_sched *sched)
1493 {
1494 	struct rb_node *node;
1495 	struct rb_root_cached *root = &sched->atom_root;
1496 again:
1497 	for (;;) {
1498 		struct work_atoms *data;
1499 		node = rb_first_cached(root);
1500 		if (!node)
1501 			break;
1502 
1503 		rb_erase_cached(node, root);
1504 		data = rb_entry(node, struct work_atoms, node);
1505 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1506 	}
1507 	if (root == &sched->atom_root) {
1508 		root = &sched->merged_atom_root;
1509 		goto again;
1510 	}
1511 }
1512 
1513 static int process_sched_wakeup_event(struct perf_tool *tool,
1514 				      struct evsel *evsel,
1515 				      struct perf_sample *sample,
1516 				      struct machine *machine)
1517 {
1518 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1519 
1520 	if (sched->tp_handler->wakeup_event)
1521 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1522 
1523 	return 0;
1524 }
1525 
1526 static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1527 				      struct evsel *evsel __maybe_unused,
1528 				      struct perf_sample *sample __maybe_unused,
1529 				      struct machine *machine __maybe_unused)
1530 {
1531 	return 0;
1532 }
1533 
1534 union map_priv {
1535 	void	*ptr;
1536 	bool	 color;
1537 };
1538 
1539 static bool thread__has_color(struct thread *thread)
1540 {
1541 	union map_priv priv = {
1542 		.ptr = thread__priv(thread),
1543 	};
1544 
1545 	return priv.color;
1546 }
1547 
1548 static struct thread*
1549 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1550 {
1551 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1552 	union map_priv priv = {
1553 		.color = false,
1554 	};
1555 
1556 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1557 		return thread;
1558 
1559 	if (thread_map__has(sched->map.color_pids, tid))
1560 		priv.color = true;
1561 
1562 	thread__set_priv(thread, priv.ptr);
1563 	return thread;
1564 }
1565 
1566 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1567 			    struct perf_sample *sample, struct machine *machine)
1568 {
1569 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1570 	struct thread *sched_in;
1571 	struct thread_runtime *tr;
1572 	int new_shortname;
1573 	u64 timestamp0, timestamp = sample->time;
1574 	s64 delta;
1575 	int i;
1576 	struct perf_cpu this_cpu = {
1577 		.cpu = sample->cpu,
1578 	};
1579 	int cpus_nr;
1580 	bool new_cpu = false;
1581 	const char *color = PERF_COLOR_NORMAL;
1582 	char stimestamp[32];
1583 
1584 	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1585 
1586 	if (this_cpu.cpu > sched->max_cpu.cpu)
1587 		sched->max_cpu = this_cpu;
1588 
1589 	if (sched->map.comp) {
1590 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1591 		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1592 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1593 			new_cpu = true;
1594 		}
1595 	} else
1596 		cpus_nr = sched->max_cpu.cpu;
1597 
1598 	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1599 	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1600 	if (timestamp0)
1601 		delta = timestamp - timestamp0;
1602 	else
1603 		delta = 0;
1604 
1605 	if (delta < 0) {
1606 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1607 		return -1;
1608 	}
1609 
1610 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1611 	if (sched_in == NULL)
1612 		return -1;
1613 
1614 	tr = thread__get_runtime(sched_in);
1615 	if (tr == NULL) {
1616 		thread__put(sched_in);
1617 		return -1;
1618 	}
1619 
1620 	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1621 
1622 	printf("  ");
1623 
1624 	new_shortname = 0;
1625 	if (!tr->shortname[0]) {
1626 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1627 			/*
1628 			 * Don't allocate a letter-number for swapper:0
1629 			 * as a shortname. Instead, we use '.' for it.
1630 			 */
1631 			tr->shortname[0] = '.';
1632 			tr->shortname[1] = ' ';
1633 		} else {
1634 			tr->shortname[0] = sched->next_shortname1;
1635 			tr->shortname[1] = sched->next_shortname2;
1636 
1637 			if (sched->next_shortname1 < 'Z') {
1638 				sched->next_shortname1++;
1639 			} else {
1640 				sched->next_shortname1 = 'A';
1641 				if (sched->next_shortname2 < '9')
1642 					sched->next_shortname2++;
1643 				else
1644 					sched->next_shortname2 = '0';
1645 			}
1646 		}
1647 		new_shortname = 1;
1648 	}
1649 
1650 	for (i = 0; i < cpus_nr; i++) {
1651 		struct perf_cpu cpu = {
1652 			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1653 		};
1654 		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1655 		struct thread_runtime *curr_tr;
1656 		const char *pid_color = color;
1657 		const char *cpu_color = color;
1658 
1659 		if (curr_thread && thread__has_color(curr_thread))
1660 			pid_color = COLOR_PIDS;
1661 
1662 		if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1663 			continue;
1664 
1665 		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1666 			cpu_color = COLOR_CPUS;
1667 
1668 		if (cpu.cpu != this_cpu.cpu)
1669 			color_fprintf(stdout, color, " ");
1670 		else
1671 			color_fprintf(stdout, cpu_color, "*");
1672 
1673 		if (sched->curr_thread[cpu.cpu]) {
1674 			curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1675 			if (curr_tr == NULL) {
1676 				thread__put(sched_in);
1677 				return -1;
1678 			}
1679 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1680 		} else
1681 			color_fprintf(stdout, color, "   ");
1682 	}
1683 
1684 	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1685 		goto out;
1686 
1687 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1688 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1689 	if (new_shortname || tr->comm_changed || (verbose > 0 && thread__tid(sched_in))) {
1690 		const char *pid_color = color;
1691 
1692 		if (thread__has_color(sched_in))
1693 			pid_color = COLOR_PIDS;
1694 
1695 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1696 			tr->shortname, thread__comm_str(sched_in), thread__tid(sched_in));
1697 		tr->comm_changed = false;
1698 	}
1699 
1700 	if (sched->map.comp && new_cpu)
1701 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1702 
1703 out:
1704 	color_fprintf(stdout, color, "\n");
1705 
1706 	thread__put(sched_in);
1707 
1708 	return 0;
1709 }
1710 
1711 static int process_sched_switch_event(struct perf_tool *tool,
1712 				      struct evsel *evsel,
1713 				      struct perf_sample *sample,
1714 				      struct machine *machine)
1715 {
1716 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1717 	int this_cpu = sample->cpu, err = 0;
1718 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1719 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1720 
1721 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1722 		/*
1723 		 * Are we trying to switch away a PID that is
1724 		 * not current?
1725 		 */
1726 		if (sched->curr_pid[this_cpu] != prev_pid)
1727 			sched->nr_context_switch_bugs++;
1728 	}
1729 
1730 	if (sched->tp_handler->switch_event)
1731 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1732 
1733 	sched->curr_pid[this_cpu] = next_pid;
1734 	return err;
1735 }
1736 
1737 static int process_sched_runtime_event(struct perf_tool *tool,
1738 				       struct evsel *evsel,
1739 				       struct perf_sample *sample,
1740 				       struct machine *machine)
1741 {
1742 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1743 
1744 	if (sched->tp_handler->runtime_event)
1745 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1746 
1747 	return 0;
1748 }
1749 
1750 static int perf_sched__process_fork_event(struct perf_tool *tool,
1751 					  union perf_event *event,
1752 					  struct perf_sample *sample,
1753 					  struct machine *machine)
1754 {
1755 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1756 
1757 	/* run the fork event through the perf machinery */
1758 	perf_event__process_fork(tool, event, sample, machine);
1759 
1760 	/* and then run additional processing needed for this command */
1761 	if (sched->tp_handler->fork_event)
1762 		return sched->tp_handler->fork_event(sched, event, machine);
1763 
1764 	return 0;
1765 }
1766 
1767 static int process_sched_migrate_task_event(struct perf_tool *tool,
1768 					    struct evsel *evsel,
1769 					    struct perf_sample *sample,
1770 					    struct machine *machine)
1771 {
1772 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1773 
1774 	if (sched->tp_handler->migrate_task_event)
1775 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1776 
1777 	return 0;
1778 }
1779 
1780 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1781 				  struct evsel *evsel,
1782 				  struct perf_sample *sample,
1783 				  struct machine *machine);
1784 
1785 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1786 						 union perf_event *event __maybe_unused,
1787 						 struct perf_sample *sample,
1788 						 struct evsel *evsel,
1789 						 struct machine *machine)
1790 {
1791 	int err = 0;
1792 
1793 	if (evsel->handler != NULL) {
1794 		tracepoint_handler f = evsel->handler;
1795 		err = f(tool, evsel, sample, machine);
1796 	}
1797 
1798 	return err;
1799 }
1800 
1801 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1802 				    union perf_event *event,
1803 				    struct perf_sample *sample,
1804 				    struct machine *machine)
1805 {
1806 	struct thread *thread;
1807 	struct thread_runtime *tr;
1808 	int err;
1809 
1810 	err = perf_event__process_comm(tool, event, sample, machine);
1811 	if (err)
1812 		return err;
1813 
1814 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1815 	if (!thread) {
1816 		pr_err("Internal error: can't find thread\n");
1817 		return -1;
1818 	}
1819 
1820 	tr = thread__get_runtime(thread);
1821 	if (tr == NULL) {
1822 		thread__put(thread);
1823 		return -1;
1824 	}
1825 
1826 	tr->comm_changed = true;
1827 	thread__put(thread);
1828 
1829 	return 0;
1830 }
1831 
1832 static int perf_sched__read_events(struct perf_sched *sched)
1833 {
1834 	struct evsel_str_handler handlers[] = {
1835 		{ "sched:sched_switch",	      process_sched_switch_event, },
1836 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1837 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1838 		{ "sched:sched_waking",	      process_sched_wakeup_event, },
1839 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1840 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1841 	};
1842 	struct perf_session *session;
1843 	struct perf_data data = {
1844 		.path  = input_name,
1845 		.mode  = PERF_DATA_MODE_READ,
1846 		.force = sched->force,
1847 	};
1848 	int rc = -1;
1849 
1850 	session = perf_session__new(&data, &sched->tool);
1851 	if (IS_ERR(session)) {
1852 		pr_debug("Error creating perf session");
1853 		return PTR_ERR(session);
1854 	}
1855 
1856 	symbol__init(&session->header.env);
1857 
1858 	/* prefer sched_waking if it is captured */
1859 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1860 		handlers[2].handler = process_sched_wakeup_ignore;
1861 
1862 	if (perf_session__set_tracepoints_handlers(session, handlers))
1863 		goto out_delete;
1864 
1865 	if (perf_session__has_traces(session, "record -R")) {
1866 		int err = perf_session__process_events(session);
1867 		if (err) {
1868 			pr_err("Failed to process events, error %d", err);
1869 			goto out_delete;
1870 		}
1871 
1872 		sched->nr_events      = session->evlist->stats.nr_events[0];
1873 		sched->nr_lost_events = session->evlist->stats.total_lost;
1874 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1875 	}
1876 
1877 	rc = 0;
1878 out_delete:
1879 	perf_session__delete(session);
1880 	return rc;
1881 }
1882 
1883 /*
1884  * scheduling times are printed as msec.usec
1885  */
1886 static inline void print_sched_time(unsigned long long nsecs, int width)
1887 {
1888 	unsigned long msecs;
1889 	unsigned long usecs;
1890 
1891 	msecs  = nsecs / NSEC_PER_MSEC;
1892 	nsecs -= msecs * NSEC_PER_MSEC;
1893 	usecs  = nsecs / NSEC_PER_USEC;
1894 	printf("%*lu.%03lu ", width, msecs, usecs);
1895 }
1896 
1897 /*
1898  * returns runtime data for event, allocating memory for it the
1899  * first time it is used.
1900  */
1901 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1902 {
1903 	struct evsel_runtime *r = evsel->priv;
1904 
1905 	if (r == NULL) {
1906 		r = zalloc(sizeof(struct evsel_runtime));
1907 		evsel->priv = r;
1908 	}
1909 
1910 	return r;
1911 }
1912 
1913 /*
1914  * save last time event was seen per cpu
1915  */
1916 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1917 {
1918 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1919 
1920 	if (r == NULL)
1921 		return;
1922 
1923 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1924 		int i, n = __roundup_pow_of_two(cpu+1);
1925 		void *p = r->last_time;
1926 
1927 		p = realloc(r->last_time, n * sizeof(u64));
1928 		if (!p)
1929 			return;
1930 
1931 		r->last_time = p;
1932 		for (i = r->ncpu; i < n; ++i)
1933 			r->last_time[i] = (u64) 0;
1934 
1935 		r->ncpu = n;
1936 	}
1937 
1938 	r->last_time[cpu] = timestamp;
1939 }
1940 
1941 /* returns last time this event was seen on the given cpu */
1942 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1943 {
1944 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1945 
1946 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1947 		return 0;
1948 
1949 	return r->last_time[cpu];
1950 }
1951 
1952 static int comm_width = 30;
1953 
1954 static char *timehist_get_commstr(struct thread *thread)
1955 {
1956 	static char str[32];
1957 	const char *comm = thread__comm_str(thread);
1958 	pid_t tid = thread__tid(thread);
1959 	pid_t pid = thread__pid(thread);
1960 	int n;
1961 
1962 	if (pid == 0)
1963 		n = scnprintf(str, sizeof(str), "%s", comm);
1964 
1965 	else if (tid != pid)
1966 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1967 
1968 	else
1969 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1970 
1971 	if (n > comm_width)
1972 		comm_width = n;
1973 
1974 	return str;
1975 }
1976 
1977 static void timehist_header(struct perf_sched *sched)
1978 {
1979 	u32 ncpus = sched->max_cpu.cpu + 1;
1980 	u32 i, j;
1981 
1982 	printf("%15s %6s ", "time", "cpu");
1983 
1984 	if (sched->show_cpu_visual) {
1985 		printf(" ");
1986 		for (i = 0, j = 0; i < ncpus; ++i) {
1987 			printf("%x", j++);
1988 			if (j > 15)
1989 				j = 0;
1990 		}
1991 		printf(" ");
1992 	}
1993 
1994 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1995 		"task name", "wait time", "sch delay", "run time");
1996 
1997 	if (sched->show_state)
1998 		printf("  %s", "state");
1999 
2000 	printf("\n");
2001 
2002 	/*
2003 	 * units row
2004 	 */
2005 	printf("%15s %-6s ", "", "");
2006 
2007 	if (sched->show_cpu_visual)
2008 		printf(" %*s ", ncpus, "");
2009 
2010 	printf(" %-*s  %9s  %9s  %9s", comm_width,
2011 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
2012 
2013 	if (sched->show_state)
2014 		printf("  %5s", "");
2015 
2016 	printf("\n");
2017 
2018 	/*
2019 	 * separator
2020 	 */
2021 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2022 
2023 	if (sched->show_cpu_visual)
2024 		printf(" %.*s ", ncpus, graph_dotted_line);
2025 
2026 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2027 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
2028 		graph_dotted_line);
2029 
2030 	if (sched->show_state)
2031 		printf("  %.5s", graph_dotted_line);
2032 
2033 	printf("\n");
2034 }
2035 
2036 static char task_state_char(struct thread *thread, int state)
2037 {
2038 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
2039 	unsigned bit = state ? ffs(state) : 0;
2040 
2041 	/* 'I' for idle */
2042 	if (thread__tid(thread) == 0)
2043 		return 'I';
2044 
2045 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
2046 }
2047 
2048 static void timehist_print_sample(struct perf_sched *sched,
2049 				  struct evsel *evsel,
2050 				  struct perf_sample *sample,
2051 				  struct addr_location *al,
2052 				  struct thread *thread,
2053 				  u64 t, int state)
2054 {
2055 	struct thread_runtime *tr = thread__priv(thread);
2056 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2057 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2058 	u32 max_cpus = sched->max_cpu.cpu + 1;
2059 	char tstr[64];
2060 	char nstr[30];
2061 	u64 wait_time;
2062 
2063 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2064 		return;
2065 
2066 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2067 	printf("%15s [%04d] ", tstr, sample->cpu);
2068 
2069 	if (sched->show_cpu_visual) {
2070 		u32 i;
2071 		char c;
2072 
2073 		printf(" ");
2074 		for (i = 0; i < max_cpus; ++i) {
2075 			/* flag idle times with 'i'; others are sched events */
2076 			if (i == sample->cpu)
2077 				c = (thread__tid(thread) == 0) ? 'i' : 's';
2078 			else
2079 				c = ' ';
2080 			printf("%c", c);
2081 		}
2082 		printf(" ");
2083 	}
2084 
2085 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2086 
2087 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2088 	print_sched_time(wait_time, 6);
2089 
2090 	print_sched_time(tr->dt_delay, 6);
2091 	print_sched_time(tr->dt_run, 6);
2092 
2093 	if (sched->show_state)
2094 		printf(" %5c ", task_state_char(thread, state));
2095 
2096 	if (sched->show_next) {
2097 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2098 		printf(" %-*s", comm_width, nstr);
2099 	}
2100 
2101 	if (sched->show_wakeups && !sched->show_next)
2102 		printf("  %-*s", comm_width, "");
2103 
2104 	if (thread__tid(thread) == 0)
2105 		goto out;
2106 
2107 	if (sched->show_callchain)
2108 		printf("  ");
2109 
2110 	sample__fprintf_sym(sample, al, 0,
2111 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2112 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2113 			    EVSEL__PRINT_SKIP_IGNORED,
2114 			    get_tls_callchain_cursor(), symbol_conf.bt_stop_list,  stdout);
2115 
2116 out:
2117 	printf("\n");
2118 }
2119 
2120 /*
2121  * Explanation of delta-time stats:
2122  *
2123  *            t = time of current schedule out event
2124  *        tprev = time of previous sched out event
2125  *                also time of schedule-in event for current task
2126  *    last_time = time of last sched change event for current task
2127  *                (i.e, time process was last scheduled out)
2128  * ready_to_run = time of wakeup for current task
2129  *
2130  * -----|------------|------------|------------|------
2131  *    last         ready        tprev          t
2132  *    time         to run
2133  *
2134  *      |-------- dt_wait --------|
2135  *                   |- dt_delay -|-- dt_run --|
2136  *
2137  *   dt_run = run time of current task
2138  *  dt_wait = time between last schedule out event for task and tprev
2139  *            represents time spent off the cpu
2140  * dt_delay = time between wakeup and schedule-in of task
2141  */
2142 
2143 static void timehist_update_runtime_stats(struct thread_runtime *r,
2144 					 u64 t, u64 tprev)
2145 {
2146 	r->dt_delay   = 0;
2147 	r->dt_sleep   = 0;
2148 	r->dt_iowait  = 0;
2149 	r->dt_preempt = 0;
2150 	r->dt_run     = 0;
2151 
2152 	if (tprev) {
2153 		r->dt_run = t - tprev;
2154 		if (r->ready_to_run) {
2155 			if (r->ready_to_run > tprev)
2156 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2157 			else
2158 				r->dt_delay = tprev - r->ready_to_run;
2159 		}
2160 
2161 		if (r->last_time > tprev)
2162 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2163 		else if (r->last_time) {
2164 			u64 dt_wait = tprev - r->last_time;
2165 
2166 			if (r->last_state == TASK_RUNNING)
2167 				r->dt_preempt = dt_wait;
2168 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2169 				r->dt_iowait = dt_wait;
2170 			else
2171 				r->dt_sleep = dt_wait;
2172 		}
2173 	}
2174 
2175 	update_stats(&r->run_stats, r->dt_run);
2176 
2177 	r->total_run_time     += r->dt_run;
2178 	r->total_delay_time   += r->dt_delay;
2179 	r->total_sleep_time   += r->dt_sleep;
2180 	r->total_iowait_time  += r->dt_iowait;
2181 	r->total_preempt_time += r->dt_preempt;
2182 }
2183 
2184 static bool is_idle_sample(struct perf_sample *sample,
2185 			   struct evsel *evsel)
2186 {
2187 	/* pid 0 == swapper == idle task */
2188 	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2189 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2190 
2191 	return sample->pid == 0;
2192 }
2193 
2194 static void save_task_callchain(struct perf_sched *sched,
2195 				struct perf_sample *sample,
2196 				struct evsel *evsel,
2197 				struct machine *machine)
2198 {
2199 	struct callchain_cursor *cursor;
2200 	struct thread *thread;
2201 
2202 	/* want main thread for process - has maps */
2203 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2204 	if (thread == NULL) {
2205 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2206 		return;
2207 	}
2208 
2209 	if (!sched->show_callchain || sample->callchain == NULL)
2210 		return;
2211 
2212 	cursor = get_tls_callchain_cursor();
2213 
2214 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2215 				      NULL, NULL, sched->max_stack + 2) != 0) {
2216 		if (verbose > 0)
2217 			pr_err("Failed to resolve callchain. Skipping\n");
2218 
2219 		return;
2220 	}
2221 
2222 	callchain_cursor_commit(cursor);
2223 
2224 	while (true) {
2225 		struct callchain_cursor_node *node;
2226 		struct symbol *sym;
2227 
2228 		node = callchain_cursor_current(cursor);
2229 		if (node == NULL)
2230 			break;
2231 
2232 		sym = node->ms.sym;
2233 		if (sym) {
2234 			if (!strcmp(sym->name, "schedule") ||
2235 			    !strcmp(sym->name, "__schedule") ||
2236 			    !strcmp(sym->name, "preempt_schedule"))
2237 				sym->ignore = 1;
2238 		}
2239 
2240 		callchain_cursor_advance(cursor);
2241 	}
2242 }
2243 
2244 static int init_idle_thread(struct thread *thread)
2245 {
2246 	struct idle_thread_runtime *itr;
2247 
2248 	thread__set_comm(thread, idle_comm, 0);
2249 
2250 	itr = zalloc(sizeof(*itr));
2251 	if (itr == NULL)
2252 		return -ENOMEM;
2253 
2254 	init_stats(&itr->tr.run_stats);
2255 	callchain_init(&itr->callchain);
2256 	callchain_cursor_reset(&itr->cursor);
2257 	thread__set_priv(thread, itr);
2258 
2259 	return 0;
2260 }
2261 
2262 /*
2263  * Track idle stats per cpu by maintaining a local thread
2264  * struct for the idle task on each cpu.
2265  */
2266 static int init_idle_threads(int ncpu)
2267 {
2268 	int i, ret;
2269 
2270 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2271 	if (!idle_threads)
2272 		return -ENOMEM;
2273 
2274 	idle_max_cpu = ncpu;
2275 
2276 	/* allocate the actual thread struct if needed */
2277 	for (i = 0; i < ncpu; ++i) {
2278 		idle_threads[i] = thread__new(0, 0);
2279 		if (idle_threads[i] == NULL)
2280 			return -ENOMEM;
2281 
2282 		ret = init_idle_thread(idle_threads[i]);
2283 		if (ret < 0)
2284 			return ret;
2285 	}
2286 
2287 	return 0;
2288 }
2289 
2290 static void free_idle_threads(void)
2291 {
2292 	int i;
2293 
2294 	if (idle_threads == NULL)
2295 		return;
2296 
2297 	for (i = 0; i < idle_max_cpu; ++i) {
2298 		if ((idle_threads[i]))
2299 			thread__delete(idle_threads[i]);
2300 	}
2301 
2302 	free(idle_threads);
2303 }
2304 
2305 static struct thread *get_idle_thread(int cpu)
2306 {
2307 	/*
2308 	 * expand/allocate array of pointers to local thread
2309 	 * structs if needed
2310 	 */
2311 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2312 		int i, j = __roundup_pow_of_two(cpu+1);
2313 		void *p;
2314 
2315 		p = realloc(idle_threads, j * sizeof(struct thread *));
2316 		if (!p)
2317 			return NULL;
2318 
2319 		idle_threads = (struct thread **) p;
2320 		for (i = idle_max_cpu; i < j; ++i)
2321 			idle_threads[i] = NULL;
2322 
2323 		idle_max_cpu = j;
2324 	}
2325 
2326 	/* allocate a new thread struct if needed */
2327 	if (idle_threads[cpu] == NULL) {
2328 		idle_threads[cpu] = thread__new(0, 0);
2329 
2330 		if (idle_threads[cpu]) {
2331 			if (init_idle_thread(idle_threads[cpu]) < 0)
2332 				return NULL;
2333 		}
2334 	}
2335 
2336 	return idle_threads[cpu];
2337 }
2338 
2339 static void save_idle_callchain(struct perf_sched *sched,
2340 				struct idle_thread_runtime *itr,
2341 				struct perf_sample *sample)
2342 {
2343 	struct callchain_cursor *cursor;
2344 
2345 	if (!sched->show_callchain || sample->callchain == NULL)
2346 		return;
2347 
2348 	cursor = get_tls_callchain_cursor();
2349 	if (cursor == NULL)
2350 		return;
2351 
2352 	callchain_cursor__copy(&itr->cursor, cursor);
2353 }
2354 
2355 static struct thread *timehist_get_thread(struct perf_sched *sched,
2356 					  struct perf_sample *sample,
2357 					  struct machine *machine,
2358 					  struct evsel *evsel)
2359 {
2360 	struct thread *thread;
2361 
2362 	if (is_idle_sample(sample, evsel)) {
2363 		thread = get_idle_thread(sample->cpu);
2364 		if (thread == NULL)
2365 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2366 
2367 	} else {
2368 		/* there were samples with tid 0 but non-zero pid */
2369 		thread = machine__findnew_thread(machine, sample->pid,
2370 						 sample->tid ?: sample->pid);
2371 		if (thread == NULL) {
2372 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2373 				 sample->tid);
2374 		}
2375 
2376 		save_task_callchain(sched, sample, evsel, machine);
2377 		if (sched->idle_hist) {
2378 			struct thread *idle;
2379 			struct idle_thread_runtime *itr;
2380 
2381 			idle = get_idle_thread(sample->cpu);
2382 			if (idle == NULL) {
2383 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2384 				return NULL;
2385 			}
2386 
2387 			itr = thread__priv(idle);
2388 			if (itr == NULL)
2389 				return NULL;
2390 
2391 			itr->last_thread = thread;
2392 
2393 			/* copy task callchain when entering to idle */
2394 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2395 				save_idle_callchain(sched, itr, sample);
2396 		}
2397 	}
2398 
2399 	return thread;
2400 }
2401 
2402 static bool timehist_skip_sample(struct perf_sched *sched,
2403 				 struct thread *thread,
2404 				 struct evsel *evsel,
2405 				 struct perf_sample *sample)
2406 {
2407 	bool rc = false;
2408 
2409 	if (thread__is_filtered(thread)) {
2410 		rc = true;
2411 		sched->skipped_samples++;
2412 	}
2413 
2414 	if (sched->idle_hist) {
2415 		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2416 			rc = true;
2417 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2418 			 evsel__intval(evsel, sample, "next_pid") != 0)
2419 			rc = true;
2420 	}
2421 
2422 	return rc;
2423 }
2424 
2425 static void timehist_print_wakeup_event(struct perf_sched *sched,
2426 					struct evsel *evsel,
2427 					struct perf_sample *sample,
2428 					struct machine *machine,
2429 					struct thread *awakened)
2430 {
2431 	struct thread *thread;
2432 	char tstr[64];
2433 
2434 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2435 	if (thread == NULL)
2436 		return;
2437 
2438 	/* show wakeup unless both awakee and awaker are filtered */
2439 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2440 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2441 		return;
2442 	}
2443 
2444 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2445 	printf("%15s [%04d] ", tstr, sample->cpu);
2446 	if (sched->show_cpu_visual)
2447 		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2448 
2449 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2450 
2451 	/* dt spacer */
2452 	printf("  %9s  %9s  %9s ", "", "", "");
2453 
2454 	printf("awakened: %s", timehist_get_commstr(awakened));
2455 
2456 	printf("\n");
2457 }
2458 
2459 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2460 					union perf_event *event __maybe_unused,
2461 					struct evsel *evsel __maybe_unused,
2462 					struct perf_sample *sample __maybe_unused,
2463 					struct machine *machine __maybe_unused)
2464 {
2465 	return 0;
2466 }
2467 
2468 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2469 				       union perf_event *event __maybe_unused,
2470 				       struct evsel *evsel,
2471 				       struct perf_sample *sample,
2472 				       struct machine *machine)
2473 {
2474 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2475 	struct thread *thread;
2476 	struct thread_runtime *tr = NULL;
2477 	/* want pid of awakened task not pid in sample */
2478 	const u32 pid = evsel__intval(evsel, sample, "pid");
2479 
2480 	thread = machine__findnew_thread(machine, 0, pid);
2481 	if (thread == NULL)
2482 		return -1;
2483 
2484 	tr = thread__get_runtime(thread);
2485 	if (tr == NULL)
2486 		return -1;
2487 
2488 	if (tr->ready_to_run == 0)
2489 		tr->ready_to_run = sample->time;
2490 
2491 	/* show wakeups if requested */
2492 	if (sched->show_wakeups &&
2493 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2494 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2495 
2496 	return 0;
2497 }
2498 
2499 static void timehist_print_migration_event(struct perf_sched *sched,
2500 					struct evsel *evsel,
2501 					struct perf_sample *sample,
2502 					struct machine *machine,
2503 					struct thread *migrated)
2504 {
2505 	struct thread *thread;
2506 	char tstr[64];
2507 	u32 max_cpus;
2508 	u32 ocpu, dcpu;
2509 
2510 	if (sched->summary_only)
2511 		return;
2512 
2513 	max_cpus = sched->max_cpu.cpu + 1;
2514 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2515 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2516 
2517 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2518 	if (thread == NULL)
2519 		return;
2520 
2521 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2522 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2523 		return;
2524 	}
2525 
2526 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2527 	printf("%15s [%04d] ", tstr, sample->cpu);
2528 
2529 	if (sched->show_cpu_visual) {
2530 		u32 i;
2531 		char c;
2532 
2533 		printf("  ");
2534 		for (i = 0; i < max_cpus; ++i) {
2535 			c = (i == sample->cpu) ? 'm' : ' ';
2536 			printf("%c", c);
2537 		}
2538 		printf("  ");
2539 	}
2540 
2541 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2542 
2543 	/* dt spacer */
2544 	printf("  %9s  %9s  %9s ", "", "", "");
2545 
2546 	printf("migrated: %s", timehist_get_commstr(migrated));
2547 	printf(" cpu %d => %d", ocpu, dcpu);
2548 
2549 	printf("\n");
2550 }
2551 
2552 static int timehist_migrate_task_event(struct perf_tool *tool,
2553 				       union perf_event *event __maybe_unused,
2554 				       struct evsel *evsel,
2555 				       struct perf_sample *sample,
2556 				       struct machine *machine)
2557 {
2558 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2559 	struct thread *thread;
2560 	struct thread_runtime *tr = NULL;
2561 	/* want pid of migrated task not pid in sample */
2562 	const u32 pid = evsel__intval(evsel, sample, "pid");
2563 
2564 	thread = machine__findnew_thread(machine, 0, pid);
2565 	if (thread == NULL)
2566 		return -1;
2567 
2568 	tr = thread__get_runtime(thread);
2569 	if (tr == NULL)
2570 		return -1;
2571 
2572 	tr->migrations++;
2573 
2574 	/* show migrations if requested */
2575 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2576 
2577 	return 0;
2578 }
2579 
2580 static int timehist_sched_change_event(struct perf_tool *tool,
2581 				       union perf_event *event,
2582 				       struct evsel *evsel,
2583 				       struct perf_sample *sample,
2584 				       struct machine *machine)
2585 {
2586 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2587 	struct perf_time_interval *ptime = &sched->ptime;
2588 	struct addr_location al;
2589 	struct thread *thread;
2590 	struct thread_runtime *tr = NULL;
2591 	u64 tprev, t = sample->time;
2592 	int rc = 0;
2593 	int state = evsel__intval(evsel, sample, "prev_state");
2594 
2595 	addr_location__init(&al);
2596 	if (machine__resolve(machine, &al, sample) < 0) {
2597 		pr_err("problem processing %d event. skipping it\n",
2598 		       event->header.type);
2599 		rc = -1;
2600 		goto out;
2601 	}
2602 
2603 	thread = timehist_get_thread(sched, sample, machine, evsel);
2604 	if (thread == NULL) {
2605 		rc = -1;
2606 		goto out;
2607 	}
2608 
2609 	if (timehist_skip_sample(sched, thread, evsel, sample))
2610 		goto out;
2611 
2612 	tr = thread__get_runtime(thread);
2613 	if (tr == NULL) {
2614 		rc = -1;
2615 		goto out;
2616 	}
2617 
2618 	tprev = evsel__get_time(evsel, sample->cpu);
2619 
2620 	/*
2621 	 * If start time given:
2622 	 * - sample time is under window user cares about - skip sample
2623 	 * - tprev is under window user cares about  - reset to start of window
2624 	 */
2625 	if (ptime->start && ptime->start > t)
2626 		goto out;
2627 
2628 	if (tprev && ptime->start > tprev)
2629 		tprev = ptime->start;
2630 
2631 	/*
2632 	 * If end time given:
2633 	 * - previous sched event is out of window - we are done
2634 	 * - sample time is beyond window user cares about - reset it
2635 	 *   to close out stats for time window interest
2636 	 */
2637 	if (ptime->end) {
2638 		if (tprev > ptime->end)
2639 			goto out;
2640 
2641 		if (t > ptime->end)
2642 			t = ptime->end;
2643 	}
2644 
2645 	if (!sched->idle_hist || thread__tid(thread) == 0) {
2646 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2647 			timehist_update_runtime_stats(tr, t, tprev);
2648 
2649 		if (sched->idle_hist) {
2650 			struct idle_thread_runtime *itr = (void *)tr;
2651 			struct thread_runtime *last_tr;
2652 
2653 			BUG_ON(thread__tid(thread) != 0);
2654 
2655 			if (itr->last_thread == NULL)
2656 				goto out;
2657 
2658 			/* add current idle time as last thread's runtime */
2659 			last_tr = thread__get_runtime(itr->last_thread);
2660 			if (last_tr == NULL)
2661 				goto out;
2662 
2663 			timehist_update_runtime_stats(last_tr, t, tprev);
2664 			/*
2665 			 * remove delta time of last thread as it's not updated
2666 			 * and otherwise it will show an invalid value next
2667 			 * time.  we only care total run time and run stat.
2668 			 */
2669 			last_tr->dt_run = 0;
2670 			last_tr->dt_delay = 0;
2671 			last_tr->dt_sleep = 0;
2672 			last_tr->dt_iowait = 0;
2673 			last_tr->dt_preempt = 0;
2674 
2675 			if (itr->cursor.nr)
2676 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2677 
2678 			itr->last_thread = NULL;
2679 		}
2680 	}
2681 
2682 	if (!sched->summary_only)
2683 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2684 
2685 out:
2686 	if (sched->hist_time.start == 0 && t >= ptime->start)
2687 		sched->hist_time.start = t;
2688 	if (ptime->end == 0 || t <= ptime->end)
2689 		sched->hist_time.end = t;
2690 
2691 	if (tr) {
2692 		/* time of this sched_switch event becomes last time task seen */
2693 		tr->last_time = sample->time;
2694 
2695 		/* last state is used to determine where to account wait time */
2696 		tr->last_state = state;
2697 
2698 		/* sched out event for task so reset ready to run time */
2699 		tr->ready_to_run = 0;
2700 	}
2701 
2702 	evsel__save_time(evsel, sample->time, sample->cpu);
2703 
2704 	addr_location__exit(&al);
2705 	return rc;
2706 }
2707 
2708 static int timehist_sched_switch_event(struct perf_tool *tool,
2709 			     union perf_event *event,
2710 			     struct evsel *evsel,
2711 			     struct perf_sample *sample,
2712 			     struct machine *machine __maybe_unused)
2713 {
2714 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2715 }
2716 
2717 static int process_lost(struct perf_tool *tool __maybe_unused,
2718 			union perf_event *event,
2719 			struct perf_sample *sample,
2720 			struct machine *machine __maybe_unused)
2721 {
2722 	char tstr[64];
2723 
2724 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2725 	printf("%15s ", tstr);
2726 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2727 
2728 	return 0;
2729 }
2730 
2731 
2732 static void print_thread_runtime(struct thread *t,
2733 				 struct thread_runtime *r)
2734 {
2735 	double mean = avg_stats(&r->run_stats);
2736 	float stddev;
2737 
2738 	printf("%*s   %5d  %9" PRIu64 " ",
2739 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2740 	       (u64) r->run_stats.n);
2741 
2742 	print_sched_time(r->total_run_time, 8);
2743 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2744 	print_sched_time(r->run_stats.min, 6);
2745 	printf(" ");
2746 	print_sched_time((u64) mean, 6);
2747 	printf(" ");
2748 	print_sched_time(r->run_stats.max, 6);
2749 	printf("  ");
2750 	printf("%5.2f", stddev);
2751 	printf("   %5" PRIu64, r->migrations);
2752 	printf("\n");
2753 }
2754 
2755 static void print_thread_waittime(struct thread *t,
2756 				  struct thread_runtime *r)
2757 {
2758 	printf("%*s   %5d  %9" PRIu64 " ",
2759 	       comm_width, timehist_get_commstr(t), thread__ppid(t),
2760 	       (u64) r->run_stats.n);
2761 
2762 	print_sched_time(r->total_run_time, 8);
2763 	print_sched_time(r->total_sleep_time, 6);
2764 	printf(" ");
2765 	print_sched_time(r->total_iowait_time, 6);
2766 	printf(" ");
2767 	print_sched_time(r->total_preempt_time, 6);
2768 	printf(" ");
2769 	print_sched_time(r->total_delay_time, 6);
2770 	printf("\n");
2771 }
2772 
2773 struct total_run_stats {
2774 	struct perf_sched *sched;
2775 	u64  sched_count;
2776 	u64  task_count;
2777 	u64  total_run_time;
2778 };
2779 
2780 static int show_thread_runtime(struct thread *t, void *priv)
2781 {
2782 	struct total_run_stats *stats = priv;
2783 	struct thread_runtime *r;
2784 
2785 	if (thread__is_filtered(t))
2786 		return 0;
2787 
2788 	r = thread__priv(t);
2789 	if (r && r->run_stats.n) {
2790 		stats->task_count++;
2791 		stats->sched_count += r->run_stats.n;
2792 		stats->total_run_time += r->total_run_time;
2793 
2794 		if (stats->sched->show_state)
2795 			print_thread_waittime(t, r);
2796 		else
2797 			print_thread_runtime(t, r);
2798 	}
2799 
2800 	return 0;
2801 }
2802 
2803 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2804 {
2805 	const char *sep = " <- ";
2806 	struct callchain_list *chain;
2807 	size_t ret = 0;
2808 	char bf[1024];
2809 	bool first;
2810 
2811 	if (node == NULL)
2812 		return 0;
2813 
2814 	ret = callchain__fprintf_folded(fp, node->parent);
2815 	first = (ret == 0);
2816 
2817 	list_for_each_entry(chain, &node->val, list) {
2818 		if (chain->ip >= PERF_CONTEXT_MAX)
2819 			continue;
2820 		if (chain->ms.sym && chain->ms.sym->ignore)
2821 			continue;
2822 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2823 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2824 							false));
2825 		first = false;
2826 	}
2827 
2828 	return ret;
2829 }
2830 
2831 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2832 {
2833 	size_t ret = 0;
2834 	FILE *fp = stdout;
2835 	struct callchain_node *chain;
2836 	struct rb_node *rb_node = rb_first_cached(root);
2837 
2838 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2839 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2840 	       graph_dotted_line);
2841 
2842 	while (rb_node) {
2843 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2844 		rb_node = rb_next(rb_node);
2845 
2846 		ret += fprintf(fp, "  ");
2847 		print_sched_time(chain->hit, 12);
2848 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2849 		ret += fprintf(fp, " %8d  ", chain->count);
2850 		ret += callchain__fprintf_folded(fp, chain);
2851 		ret += fprintf(fp, "\n");
2852 	}
2853 
2854 	return ret;
2855 }
2856 
2857 static void timehist_print_summary(struct perf_sched *sched,
2858 				   struct perf_session *session)
2859 {
2860 	struct machine *m = &session->machines.host;
2861 	struct total_run_stats totals;
2862 	u64 task_count;
2863 	struct thread *t;
2864 	struct thread_runtime *r;
2865 	int i;
2866 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2867 
2868 	memset(&totals, 0, sizeof(totals));
2869 	totals.sched = sched;
2870 
2871 	if (sched->idle_hist) {
2872 		printf("\nIdle-time summary\n");
2873 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2874 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2875 	} else if (sched->show_state) {
2876 		printf("\nWait-time summary\n");
2877 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2878 		printf("   run-time      sleep      iowait     preempt       delay\n");
2879 	} else {
2880 		printf("\nRuntime summary\n");
2881 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2882 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2883 	}
2884 	printf("%*s            (count)  ", comm_width, "");
2885 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2886 	       sched->show_state ? "(msec)" : "%");
2887 	printf("%.117s\n", graph_dotted_line);
2888 
2889 	machine__for_each_thread(m, show_thread_runtime, &totals);
2890 	task_count = totals.task_count;
2891 	if (!task_count)
2892 		printf("<no still running tasks>\n");
2893 
2894 	/* CPU idle stats not tracked when samples were skipped */
2895 	if (sched->skipped_samples && !sched->idle_hist)
2896 		return;
2897 
2898 	printf("\nIdle stats:\n");
2899 	for (i = 0; i < idle_max_cpu; ++i) {
2900 		if (cpu_list && !test_bit(i, cpu_bitmap))
2901 			continue;
2902 
2903 		t = idle_threads[i];
2904 		if (!t)
2905 			continue;
2906 
2907 		r = thread__priv(t);
2908 		if (r && r->run_stats.n) {
2909 			totals.sched_count += r->run_stats.n;
2910 			printf("    CPU %2d idle for ", i);
2911 			print_sched_time(r->total_run_time, 6);
2912 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2913 		} else
2914 			printf("    CPU %2d idle entire time window\n", i);
2915 	}
2916 
2917 	if (sched->idle_hist && sched->show_callchain) {
2918 		callchain_param.mode  = CHAIN_FOLDED;
2919 		callchain_param.value = CCVAL_PERIOD;
2920 
2921 		callchain_register_param(&callchain_param);
2922 
2923 		printf("\nIdle stats by callchain:\n");
2924 		for (i = 0; i < idle_max_cpu; ++i) {
2925 			struct idle_thread_runtime *itr;
2926 
2927 			t = idle_threads[i];
2928 			if (!t)
2929 				continue;
2930 
2931 			itr = thread__priv(t);
2932 			if (itr == NULL)
2933 				continue;
2934 
2935 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2936 					     0, &callchain_param);
2937 
2938 			printf("  CPU %2d:", i);
2939 			print_sched_time(itr->tr.total_run_time, 6);
2940 			printf(" msec\n");
2941 			timehist_print_idlehist_callchain(&itr->sorted_root);
2942 			printf("\n");
2943 		}
2944 	}
2945 
2946 	printf("\n"
2947 	       "    Total number of unique tasks: %" PRIu64 "\n"
2948 	       "Total number of context switches: %" PRIu64 "\n",
2949 	       totals.task_count, totals.sched_count);
2950 
2951 	printf("           Total run time (msec): ");
2952 	print_sched_time(totals.total_run_time, 2);
2953 	printf("\n");
2954 
2955 	printf("    Total scheduling time (msec): ");
2956 	print_sched_time(hist_time, 2);
2957 	printf(" (x %d)\n", sched->max_cpu.cpu);
2958 }
2959 
2960 typedef int (*sched_handler)(struct perf_tool *tool,
2961 			  union perf_event *event,
2962 			  struct evsel *evsel,
2963 			  struct perf_sample *sample,
2964 			  struct machine *machine);
2965 
2966 static int perf_timehist__process_sample(struct perf_tool *tool,
2967 					 union perf_event *event,
2968 					 struct perf_sample *sample,
2969 					 struct evsel *evsel,
2970 					 struct machine *machine)
2971 {
2972 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2973 	int err = 0;
2974 	struct perf_cpu this_cpu = {
2975 		.cpu = sample->cpu,
2976 	};
2977 
2978 	if (this_cpu.cpu > sched->max_cpu.cpu)
2979 		sched->max_cpu = this_cpu;
2980 
2981 	if (evsel->handler != NULL) {
2982 		sched_handler f = evsel->handler;
2983 
2984 		err = f(tool, event, evsel, sample, machine);
2985 	}
2986 
2987 	return err;
2988 }
2989 
2990 static int timehist_check_attr(struct perf_sched *sched,
2991 			       struct evlist *evlist)
2992 {
2993 	struct evsel *evsel;
2994 	struct evsel_runtime *er;
2995 
2996 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
2997 		er = evsel__get_runtime(evsel);
2998 		if (er == NULL) {
2999 			pr_err("Failed to allocate memory for evsel runtime data\n");
3000 			return -1;
3001 		}
3002 
3003 		/* only need to save callchain related to sched_switch event */
3004 		if (sched->show_callchain &&
3005 		    evsel__name_is(evsel, "sched:sched_switch") &&
3006 		    !evsel__has_callchain(evsel)) {
3007 			pr_info("Samples of sched_switch event do not have callchains.\n");
3008 			sched->show_callchain = 0;
3009 			symbol_conf.use_callchain = 0;
3010 		}
3011 	}
3012 
3013 	return 0;
3014 }
3015 
3016 static int perf_sched__timehist(struct perf_sched *sched)
3017 {
3018 	struct evsel_str_handler handlers[] = {
3019 		{ "sched:sched_switch",       timehist_sched_switch_event, },
3020 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
3021 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
3022 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
3023 	};
3024 	const struct evsel_str_handler migrate_handlers[] = {
3025 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
3026 	};
3027 	struct perf_data data = {
3028 		.path  = input_name,
3029 		.mode  = PERF_DATA_MODE_READ,
3030 		.force = sched->force,
3031 	};
3032 
3033 	struct perf_session *session;
3034 	struct evlist *evlist;
3035 	int err = -1;
3036 
3037 	/*
3038 	 * event handlers for timehist option
3039 	 */
3040 	sched->tool.sample	 = perf_timehist__process_sample;
3041 	sched->tool.mmap	 = perf_event__process_mmap;
3042 	sched->tool.comm	 = perf_event__process_comm;
3043 	sched->tool.exit	 = perf_event__process_exit;
3044 	sched->tool.fork	 = perf_event__process_fork;
3045 	sched->tool.lost	 = process_lost;
3046 	sched->tool.attr	 = perf_event__process_attr;
3047 	sched->tool.tracing_data = perf_event__process_tracing_data;
3048 	sched->tool.build_id	 = perf_event__process_build_id;
3049 
3050 	sched->tool.ordered_events = true;
3051 	sched->tool.ordering_requires_timestamps = true;
3052 
3053 	symbol_conf.use_callchain = sched->show_callchain;
3054 
3055 	session = perf_session__new(&data, &sched->tool);
3056 	if (IS_ERR(session))
3057 		return PTR_ERR(session);
3058 
3059 	if (cpu_list) {
3060 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3061 		if (err < 0)
3062 			goto out;
3063 	}
3064 
3065 	evlist = session->evlist;
3066 
3067 	symbol__init(&session->header.env);
3068 
3069 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3070 		pr_err("Invalid time string\n");
3071 		return -EINVAL;
3072 	}
3073 
3074 	if (timehist_check_attr(sched, evlist) != 0)
3075 		goto out;
3076 
3077 	setup_pager();
3078 
3079 	/* prefer sched_waking if it is captured */
3080 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3081 		handlers[1].handler = timehist_sched_wakeup_ignore;
3082 
3083 	/* setup per-evsel handlers */
3084 	if (perf_session__set_tracepoints_handlers(session, handlers))
3085 		goto out;
3086 
3087 	/* sched_switch event at a minimum needs to exist */
3088 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3089 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3090 		goto out;
3091 	}
3092 
3093 	if (sched->show_migrations &&
3094 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3095 		goto out;
3096 
3097 	/* pre-allocate struct for per-CPU idle stats */
3098 	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3099 	if (sched->max_cpu.cpu == 0)
3100 		sched->max_cpu.cpu = 4;
3101 	if (init_idle_threads(sched->max_cpu.cpu))
3102 		goto out;
3103 
3104 	/* summary_only implies summary option, but don't overwrite summary if set */
3105 	if (sched->summary_only)
3106 		sched->summary = sched->summary_only;
3107 
3108 	if (!sched->summary_only)
3109 		timehist_header(sched);
3110 
3111 	err = perf_session__process_events(session);
3112 	if (err) {
3113 		pr_err("Failed to process events, error %d", err);
3114 		goto out;
3115 	}
3116 
3117 	sched->nr_events      = evlist->stats.nr_events[0];
3118 	sched->nr_lost_events = evlist->stats.total_lost;
3119 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3120 
3121 	if (sched->summary)
3122 		timehist_print_summary(sched, session);
3123 
3124 out:
3125 	free_idle_threads();
3126 	perf_session__delete(session);
3127 
3128 	return err;
3129 }
3130 
3131 
3132 static void print_bad_events(struct perf_sched *sched)
3133 {
3134 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3135 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3136 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3137 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3138 	}
3139 	if (sched->nr_lost_events && sched->nr_events) {
3140 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3141 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3142 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3143 	}
3144 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3145 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3146 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3147 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3148 		if (sched->nr_lost_events)
3149 			printf(" (due to lost events?)");
3150 		printf("\n");
3151 	}
3152 }
3153 
3154 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3155 {
3156 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3157 	struct work_atoms *this;
3158 	const char *comm = thread__comm_str(data->thread), *this_comm;
3159 	bool leftmost = true;
3160 
3161 	while (*new) {
3162 		int cmp;
3163 
3164 		this = container_of(*new, struct work_atoms, node);
3165 		parent = *new;
3166 
3167 		this_comm = thread__comm_str(this->thread);
3168 		cmp = strcmp(comm, this_comm);
3169 		if (cmp > 0) {
3170 			new = &((*new)->rb_left);
3171 		} else if (cmp < 0) {
3172 			new = &((*new)->rb_right);
3173 			leftmost = false;
3174 		} else {
3175 			this->num_merged++;
3176 			this->total_runtime += data->total_runtime;
3177 			this->nb_atoms += data->nb_atoms;
3178 			this->total_lat += data->total_lat;
3179 			list_splice(&data->work_list, &this->work_list);
3180 			if (this->max_lat < data->max_lat) {
3181 				this->max_lat = data->max_lat;
3182 				this->max_lat_start = data->max_lat_start;
3183 				this->max_lat_end = data->max_lat_end;
3184 			}
3185 			zfree(&data);
3186 			return;
3187 		}
3188 	}
3189 
3190 	data->num_merged++;
3191 	rb_link_node(&data->node, parent, new);
3192 	rb_insert_color_cached(&data->node, root, leftmost);
3193 }
3194 
3195 static void perf_sched__merge_lat(struct perf_sched *sched)
3196 {
3197 	struct work_atoms *data;
3198 	struct rb_node *node;
3199 
3200 	if (sched->skip_merge)
3201 		return;
3202 
3203 	while ((node = rb_first_cached(&sched->atom_root))) {
3204 		rb_erase_cached(node, &sched->atom_root);
3205 		data = rb_entry(node, struct work_atoms, node);
3206 		__merge_work_atoms(&sched->merged_atom_root, data);
3207 	}
3208 }
3209 
3210 static int perf_sched__lat(struct perf_sched *sched)
3211 {
3212 	struct rb_node *next;
3213 
3214 	setup_pager();
3215 
3216 	if (perf_sched__read_events(sched))
3217 		return -1;
3218 
3219 	perf_sched__merge_lat(sched);
3220 	perf_sched__sort_lat(sched);
3221 
3222 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3223 	printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3224 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3225 
3226 	next = rb_first_cached(&sched->sorted_atom_root);
3227 
3228 	while (next) {
3229 		struct work_atoms *work_list;
3230 
3231 		work_list = rb_entry(next, struct work_atoms, node);
3232 		output_lat_thread(sched, work_list);
3233 		next = rb_next(next);
3234 		thread__zput(work_list->thread);
3235 	}
3236 
3237 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3238 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3239 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3240 
3241 	printf(" ---------------------------------------------------\n");
3242 
3243 	print_bad_events(sched);
3244 	printf("\n");
3245 
3246 	return 0;
3247 }
3248 
3249 static int setup_map_cpus(struct perf_sched *sched)
3250 {
3251 	struct perf_cpu_map *map;
3252 
3253 	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3254 
3255 	if (sched->map.comp) {
3256 		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3257 		if (!sched->map.comp_cpus)
3258 			return -1;
3259 	}
3260 
3261 	if (!sched->map.cpus_str)
3262 		return 0;
3263 
3264 	map = perf_cpu_map__new(sched->map.cpus_str);
3265 	if (!map) {
3266 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3267 		return -1;
3268 	}
3269 
3270 	sched->map.cpus = map;
3271 	return 0;
3272 }
3273 
3274 static int setup_color_pids(struct perf_sched *sched)
3275 {
3276 	struct perf_thread_map *map;
3277 
3278 	if (!sched->map.color_pids_str)
3279 		return 0;
3280 
3281 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3282 	if (!map) {
3283 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3284 		return -1;
3285 	}
3286 
3287 	sched->map.color_pids = map;
3288 	return 0;
3289 }
3290 
3291 static int setup_color_cpus(struct perf_sched *sched)
3292 {
3293 	struct perf_cpu_map *map;
3294 
3295 	if (!sched->map.color_cpus_str)
3296 		return 0;
3297 
3298 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3299 	if (!map) {
3300 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3301 		return -1;
3302 	}
3303 
3304 	sched->map.color_cpus = map;
3305 	return 0;
3306 }
3307 
3308 static int perf_sched__map(struct perf_sched *sched)
3309 {
3310 	if (setup_map_cpus(sched))
3311 		return -1;
3312 
3313 	if (setup_color_pids(sched))
3314 		return -1;
3315 
3316 	if (setup_color_cpus(sched))
3317 		return -1;
3318 
3319 	setup_pager();
3320 	if (perf_sched__read_events(sched))
3321 		return -1;
3322 	print_bad_events(sched);
3323 	return 0;
3324 }
3325 
3326 static int perf_sched__replay(struct perf_sched *sched)
3327 {
3328 	unsigned long i;
3329 
3330 	calibrate_run_measurement_overhead(sched);
3331 	calibrate_sleep_measurement_overhead(sched);
3332 
3333 	test_calibrations(sched);
3334 
3335 	if (perf_sched__read_events(sched))
3336 		return -1;
3337 
3338 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3339 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3340 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3341 
3342 	if (sched->targetless_wakeups)
3343 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3344 	if (sched->multitarget_wakeups)
3345 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3346 	if (sched->nr_run_events_optimized)
3347 		printf("run atoms optimized: %ld\n",
3348 			sched->nr_run_events_optimized);
3349 
3350 	print_task_traces(sched);
3351 	add_cross_task_wakeups(sched);
3352 
3353 	sched->thread_funcs_exit = false;
3354 	create_tasks(sched);
3355 	printf("------------------------------------------------------------\n");
3356 	for (i = 0; i < sched->replay_repeat; i++)
3357 		run_one_test(sched);
3358 
3359 	sched->thread_funcs_exit = true;
3360 	destroy_tasks(sched);
3361 	return 0;
3362 }
3363 
3364 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3365 			  const char * const usage_msg[])
3366 {
3367 	char *tmp, *tok, *str = strdup(sched->sort_order);
3368 
3369 	for (tok = strtok_r(str, ", ", &tmp);
3370 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3371 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3372 			usage_with_options_msg(usage_msg, options,
3373 					"Unknown --sort key: `%s'", tok);
3374 		}
3375 	}
3376 
3377 	free(str);
3378 
3379 	sort_dimension__add("pid", &sched->cmp_pid);
3380 }
3381 
3382 static bool schedstat_events_exposed(void)
3383 {
3384 	/*
3385 	 * Select "sched:sched_stat_wait" event to check
3386 	 * whether schedstat tracepoints are exposed.
3387 	 */
3388 	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3389 		false : true;
3390 }
3391 
3392 static int __cmd_record(int argc, const char **argv)
3393 {
3394 	unsigned int rec_argc, i, j;
3395 	char **rec_argv;
3396 	const char **rec_argv_copy;
3397 	const char * const record_args[] = {
3398 		"record",
3399 		"-a",
3400 		"-R",
3401 		"-m", "1024",
3402 		"-c", "1",
3403 		"-e", "sched:sched_switch",
3404 		"-e", "sched:sched_stat_runtime",
3405 		"-e", "sched:sched_process_fork",
3406 		"-e", "sched:sched_wakeup_new",
3407 		"-e", "sched:sched_migrate_task",
3408 	};
3409 
3410 	/*
3411 	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3412 	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3413 	 * to prevent "perf sched record" execution failure, determine
3414 	 * whether to record schedstat events according to actual situation.
3415 	 */
3416 	const char * const schedstat_args[] = {
3417 		"-e", "sched:sched_stat_wait",
3418 		"-e", "sched:sched_stat_sleep",
3419 		"-e", "sched:sched_stat_iowait",
3420 	};
3421 	unsigned int schedstat_argc = schedstat_events_exposed() ?
3422 		ARRAY_SIZE(schedstat_args) : 0;
3423 
3424 	struct tep_event *waking_event;
3425 	int ret;
3426 
3427 	/*
3428 	 * +2 for either "-e", "sched:sched_wakeup" or
3429 	 * "-e", "sched:sched_waking"
3430 	 */
3431 	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3432 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3433 	if (rec_argv == NULL)
3434 		return -ENOMEM;
3435 	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3436 	if (rec_argv_copy == NULL) {
3437 		free(rec_argv);
3438 		return -ENOMEM;
3439 	}
3440 
3441 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3442 		rec_argv[i] = strdup(record_args[i]);
3443 
3444 	rec_argv[i++] = strdup("-e");
3445 	waking_event = trace_event__tp_format("sched", "sched_waking");
3446 	if (!IS_ERR(waking_event))
3447 		rec_argv[i++] = strdup("sched:sched_waking");
3448 	else
3449 		rec_argv[i++] = strdup("sched:sched_wakeup");
3450 
3451 	for (j = 0; j < schedstat_argc; j++)
3452 		rec_argv[i++] = strdup(schedstat_args[j]);
3453 
3454 	for (j = 1; j < (unsigned int)argc; j++, i++)
3455 		rec_argv[i] = strdup(argv[j]);
3456 
3457 	BUG_ON(i != rec_argc);
3458 
3459 	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3460 	ret = cmd_record(rec_argc, rec_argv_copy);
3461 
3462 	for (i = 0; i < rec_argc; i++)
3463 		free(rec_argv[i]);
3464 	free(rec_argv);
3465 	free(rec_argv_copy);
3466 
3467 	return ret;
3468 }
3469 
3470 int cmd_sched(int argc, const char **argv)
3471 {
3472 	static const char default_sort_order[] = "avg, max, switch, runtime";
3473 	struct perf_sched sched = {
3474 		.tool = {
3475 			.sample		 = perf_sched__process_tracepoint_sample,
3476 			.comm		 = perf_sched__process_comm,
3477 			.namespaces	 = perf_event__process_namespaces,
3478 			.lost		 = perf_event__process_lost,
3479 			.fork		 = perf_sched__process_fork_event,
3480 			.ordered_events = true,
3481 		},
3482 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3483 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3484 		.sort_order	      = default_sort_order,
3485 		.replay_repeat	      = 10,
3486 		.profile_cpu	      = -1,
3487 		.next_shortname1      = 'A',
3488 		.next_shortname2      = '0',
3489 		.skip_merge           = 0,
3490 		.show_callchain	      = 1,
3491 		.max_stack            = 5,
3492 	};
3493 	const struct option sched_options[] = {
3494 	OPT_STRING('i', "input", &input_name, "file",
3495 		    "input file name"),
3496 	OPT_INCR('v', "verbose", &verbose,
3497 		    "be more verbose (show symbol address, etc)"),
3498 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3499 		    "dump raw trace in ASCII"),
3500 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3501 	OPT_END()
3502 	};
3503 	const struct option latency_options[] = {
3504 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3505 		   "sort by key(s): runtime, switch, avg, max"),
3506 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3507 		    "CPU to profile on"),
3508 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3509 		    "latency stats per pid instead of per comm"),
3510 	OPT_PARENT(sched_options)
3511 	};
3512 	const struct option replay_options[] = {
3513 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3514 		     "repeat the workload replay N times (-1: infinite)"),
3515 	OPT_PARENT(sched_options)
3516 	};
3517 	const struct option map_options[] = {
3518 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3519 		    "map output in compact mode"),
3520 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3521 		   "highlight given pids in map"),
3522 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3523                     "highlight given CPUs in map"),
3524 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3525                     "display given CPUs in map"),
3526 	OPT_PARENT(sched_options)
3527 	};
3528 	const struct option timehist_options[] = {
3529 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3530 		   "file", "vmlinux pathname"),
3531 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3532 		   "file", "kallsyms pathname"),
3533 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3534 		    "Display call chains if present (default on)"),
3535 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3536 		   "Maximum number of functions to display backtrace."),
3537 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3538 		    "Look for files with symbols relative to this directory"),
3539 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3540 		    "Show only syscall summary with statistics"),
3541 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3542 		    "Show all syscalls and summary with statistics"),
3543 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3544 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3545 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3546 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3547 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3548 	OPT_STRING(0, "time", &sched.time_str, "str",
3549 		   "Time span for analysis (start,stop)"),
3550 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3551 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3552 		   "analyze events only for given process id(s)"),
3553 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3554 		   "analyze events only for given thread id(s)"),
3555 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3556 	OPT_PARENT(sched_options)
3557 	};
3558 
3559 	const char * const latency_usage[] = {
3560 		"perf sched latency [<options>]",
3561 		NULL
3562 	};
3563 	const char * const replay_usage[] = {
3564 		"perf sched replay [<options>]",
3565 		NULL
3566 	};
3567 	const char * const map_usage[] = {
3568 		"perf sched map [<options>]",
3569 		NULL
3570 	};
3571 	const char * const timehist_usage[] = {
3572 		"perf sched timehist [<options>]",
3573 		NULL
3574 	};
3575 	const char *const sched_subcommands[] = { "record", "latency", "map",
3576 						  "replay", "script",
3577 						  "timehist", NULL };
3578 	const char *sched_usage[] = {
3579 		NULL,
3580 		NULL
3581 	};
3582 	struct trace_sched_handler lat_ops  = {
3583 		.wakeup_event	    = latency_wakeup_event,
3584 		.switch_event	    = latency_switch_event,
3585 		.runtime_event	    = latency_runtime_event,
3586 		.migrate_task_event = latency_migrate_task_event,
3587 	};
3588 	struct trace_sched_handler map_ops  = {
3589 		.switch_event	    = map_switch_event,
3590 	};
3591 	struct trace_sched_handler replay_ops  = {
3592 		.wakeup_event	    = replay_wakeup_event,
3593 		.switch_event	    = replay_switch_event,
3594 		.fork_event	    = replay_fork_event,
3595 	};
3596 	unsigned int i;
3597 	int ret = 0;
3598 
3599 	mutex_init(&sched.start_work_mutex);
3600 	mutex_init(&sched.work_done_wait_mutex);
3601 	sched.curr_thread = calloc(MAX_CPUS, sizeof(*sched.curr_thread));
3602 	if (!sched.curr_thread) {
3603 		ret = -ENOMEM;
3604 		goto out;
3605 	}
3606 	sched.cpu_last_switched = calloc(MAX_CPUS, sizeof(*sched.cpu_last_switched));
3607 	if (!sched.cpu_last_switched) {
3608 		ret = -ENOMEM;
3609 		goto out;
3610 	}
3611 	sched.curr_pid = malloc(MAX_CPUS * sizeof(*sched.curr_pid));
3612 	if (!sched.curr_pid) {
3613 		ret = -ENOMEM;
3614 		goto out;
3615 	}
3616 	for (i = 0; i < MAX_CPUS; i++)
3617 		sched.curr_pid[i] = -1;
3618 
3619 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3620 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3621 	if (!argc)
3622 		usage_with_options(sched_usage, sched_options);
3623 
3624 	/*
3625 	 * Aliased to 'perf script' for now:
3626 	 */
3627 	if (!strcmp(argv[0], "script")) {
3628 		ret = cmd_script(argc, argv);
3629 	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3630 		ret = __cmd_record(argc, argv);
3631 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3632 		sched.tp_handler = &lat_ops;
3633 		if (argc > 1) {
3634 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3635 			if (argc)
3636 				usage_with_options(latency_usage, latency_options);
3637 		}
3638 		setup_sorting(&sched, latency_options, latency_usage);
3639 		ret = perf_sched__lat(&sched);
3640 	} else if (!strcmp(argv[0], "map")) {
3641 		if (argc) {
3642 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3643 			if (argc)
3644 				usage_with_options(map_usage, map_options);
3645 		}
3646 		sched.tp_handler = &map_ops;
3647 		setup_sorting(&sched, latency_options, latency_usage);
3648 		ret = perf_sched__map(&sched);
3649 	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3650 		sched.tp_handler = &replay_ops;
3651 		if (argc) {
3652 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3653 			if (argc)
3654 				usage_with_options(replay_usage, replay_options);
3655 		}
3656 		ret = perf_sched__replay(&sched);
3657 	} else if (!strcmp(argv[0], "timehist")) {
3658 		if (argc) {
3659 			argc = parse_options(argc, argv, timehist_options,
3660 					     timehist_usage, 0);
3661 			if (argc)
3662 				usage_with_options(timehist_usage, timehist_options);
3663 		}
3664 		if ((sched.show_wakeups || sched.show_next) &&
3665 		    sched.summary_only) {
3666 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3667 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3668 			if (sched.show_wakeups)
3669 				parse_options_usage(NULL, timehist_options, "w", true);
3670 			if (sched.show_next)
3671 				parse_options_usage(NULL, timehist_options, "n", true);
3672 			ret = -EINVAL;
3673 			goto out;
3674 		}
3675 		ret = symbol__validate_sym_arguments();
3676 		if (ret)
3677 			goto out;
3678 
3679 		ret = perf_sched__timehist(&sched);
3680 	} else {
3681 		usage_with_options(sched_usage, sched_options);
3682 	}
3683 
3684 out:
3685 	free(sched.curr_pid);
3686 	free(sched.cpu_last_switched);
3687 	free(sched.curr_thread);
3688 	mutex_destroy(&sched.start_work_mutex);
3689 	mutex_destroy(&sched.work_done_wait_mutex);
3690 
3691 	return ret;
3692 }
3693