xref: /openbmc/linux/tools/perf/builtin-sched.c (revision 7c2435ef)
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[MAX_CPUS];
197 	struct thread	 *curr_thread[MAX_CPUS];
198 	char		 next_shortname1;
199 	char		 next_shortname2;
200 	unsigned int	 replay_repeat;
201 	unsigned long	 nr_run_events;
202 	unsigned long	 nr_sleep_events;
203 	unsigned long	 nr_wakeup_events;
204 	unsigned long	 nr_sleep_corrections;
205 	unsigned long	 nr_run_events_optimized;
206 	unsigned long	 targetless_wakeups;
207 	unsigned long	 multitarget_wakeups;
208 	unsigned long	 nr_runs;
209 	unsigned long	 nr_timestamps;
210 	unsigned long	 nr_unordered_timestamps;
211 	unsigned long	 nr_context_switch_bugs;
212 	unsigned long	 nr_events;
213 	unsigned long	 nr_lost_chunks;
214 	unsigned long	 nr_lost_events;
215 	u64		 run_measurement_overhead;
216 	u64		 sleep_measurement_overhead;
217 	u64		 start_time;
218 	u64		 cpu_usage;
219 	u64		 runavg_cpu_usage;
220 	u64		 parent_cpu_usage;
221 	u64		 runavg_parent_cpu_usage;
222 	u64		 sum_runtime;
223 	u64		 sum_fluct;
224 	u64		 run_avg;
225 	u64		 all_runtime;
226 	u64		 all_count;
227 	u64		 cpu_last_switched[MAX_CPUS];
228 	struct rb_root_cached atom_root, sorted_atom_root, merged_atom_root;
229 	struct list_head sort_list, cmp_pid;
230 	bool force;
231 	bool skip_merge;
232 	struct perf_sched_map map;
233 
234 	/* options for timehist command */
235 	bool		summary;
236 	bool		summary_only;
237 	bool		idle_hist;
238 	bool		show_callchain;
239 	unsigned int	max_stack;
240 	bool		show_cpu_visual;
241 	bool		show_wakeups;
242 	bool		show_next;
243 	bool		show_migrations;
244 	bool		show_state;
245 	u64		skipped_samples;
246 	const char	*time_str;
247 	struct perf_time_interval ptime;
248 	struct perf_time_interval hist_time;
249 	volatile bool   thread_funcs_exit;
250 };
251 
252 /* per thread run time data */
253 struct thread_runtime {
254 	u64 last_time;      /* time of previous sched in/out event */
255 	u64 dt_run;         /* run time */
256 	u64 dt_sleep;       /* time between CPU access by sleep (off cpu) */
257 	u64 dt_iowait;      /* time between CPU access by iowait (off cpu) */
258 	u64 dt_preempt;     /* time between CPU access by preempt (off cpu) */
259 	u64 dt_delay;       /* time between wakeup and sched-in */
260 	u64 ready_to_run;   /* time of wakeup */
261 
262 	struct stats run_stats;
263 	u64 total_run_time;
264 	u64 total_sleep_time;
265 	u64 total_iowait_time;
266 	u64 total_preempt_time;
267 	u64 total_delay_time;
268 
269 	int last_state;
270 
271 	char shortname[3];
272 	bool comm_changed;
273 
274 	u64 migrations;
275 };
276 
277 /* per event run time data */
278 struct evsel_runtime {
279 	u64 *last_time; /* time this event was last seen per cpu */
280 	u32 ncpu;       /* highest cpu slot allocated */
281 };
282 
283 /* per cpu idle time data */
284 struct idle_thread_runtime {
285 	struct thread_runtime	tr;
286 	struct thread		*last_thread;
287 	struct rb_root_cached	sorted_root;
288 	struct callchain_root	callchain;
289 	struct callchain_cursor	cursor;
290 };
291 
292 /* track idle times per cpu */
293 static struct thread **idle_threads;
294 static int idle_max_cpu;
295 static char idle_comm[] = "<idle>";
296 
297 static u64 get_nsecs(void)
298 {
299 	struct timespec ts;
300 
301 	clock_gettime(CLOCK_MONOTONIC, &ts);
302 
303 	return ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec;
304 }
305 
306 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
307 {
308 	u64 T0 = get_nsecs(), T1;
309 
310 	do {
311 		T1 = get_nsecs();
312 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
313 }
314 
315 static void sleep_nsecs(u64 nsecs)
316 {
317 	struct timespec ts;
318 
319 	ts.tv_nsec = nsecs % 999999999;
320 	ts.tv_sec = nsecs / 999999999;
321 
322 	nanosleep(&ts, NULL);
323 }
324 
325 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
326 {
327 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
328 	int i;
329 
330 	for (i = 0; i < 10; i++) {
331 		T0 = get_nsecs();
332 		burn_nsecs(sched, 0);
333 		T1 = get_nsecs();
334 		delta = T1-T0;
335 		min_delta = min(min_delta, delta);
336 	}
337 	sched->run_measurement_overhead = min_delta;
338 
339 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
340 }
341 
342 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
343 {
344 	u64 T0, T1, delta, min_delta = NSEC_PER_SEC;
345 	int i;
346 
347 	for (i = 0; i < 10; i++) {
348 		T0 = get_nsecs();
349 		sleep_nsecs(10000);
350 		T1 = get_nsecs();
351 		delta = T1-T0;
352 		min_delta = min(min_delta, delta);
353 	}
354 	min_delta -= 10000;
355 	sched->sleep_measurement_overhead = min_delta;
356 
357 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
358 }
359 
360 static struct sched_atom *
361 get_new_event(struct task_desc *task, u64 timestamp)
362 {
363 	struct sched_atom *event = zalloc(sizeof(*event));
364 	unsigned long idx = task->nr_events;
365 	size_t size;
366 
367 	event->timestamp = timestamp;
368 	event->nr = idx;
369 
370 	task->nr_events++;
371 	size = sizeof(struct sched_atom *) * task->nr_events;
372 	task->atoms = realloc(task->atoms, size);
373 	BUG_ON(!task->atoms);
374 
375 	task->atoms[idx] = event;
376 
377 	return event;
378 }
379 
380 static struct sched_atom *last_event(struct task_desc *task)
381 {
382 	if (!task->nr_events)
383 		return NULL;
384 
385 	return task->atoms[task->nr_events - 1];
386 }
387 
388 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
389 				u64 timestamp, u64 duration)
390 {
391 	struct sched_atom *event, *curr_event = last_event(task);
392 
393 	/*
394 	 * optimize an existing RUN event by merging this one
395 	 * to it:
396 	 */
397 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
398 		sched->nr_run_events_optimized++;
399 		curr_event->duration += duration;
400 		return;
401 	}
402 
403 	event = get_new_event(task, timestamp);
404 
405 	event->type = SCHED_EVENT_RUN;
406 	event->duration = duration;
407 
408 	sched->nr_run_events++;
409 }
410 
411 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
412 				   u64 timestamp, struct task_desc *wakee)
413 {
414 	struct sched_atom *event, *wakee_event;
415 
416 	event = get_new_event(task, timestamp);
417 	event->type = SCHED_EVENT_WAKEUP;
418 	event->wakee = wakee;
419 
420 	wakee_event = last_event(wakee);
421 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
422 		sched->targetless_wakeups++;
423 		return;
424 	}
425 	if (wakee_event->wait_sem) {
426 		sched->multitarget_wakeups++;
427 		return;
428 	}
429 
430 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
431 	sem_init(wakee_event->wait_sem, 0, 0);
432 	wakee_event->specific_wait = 1;
433 	event->wait_sem = wakee_event->wait_sem;
434 
435 	sched->nr_wakeup_events++;
436 }
437 
438 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
439 				  u64 timestamp, u64 task_state __maybe_unused)
440 {
441 	struct sched_atom *event = get_new_event(task, timestamp);
442 
443 	event->type = SCHED_EVENT_SLEEP;
444 
445 	sched->nr_sleep_events++;
446 }
447 
448 static struct task_desc *register_pid(struct perf_sched *sched,
449 				      unsigned long pid, const char *comm)
450 {
451 	struct task_desc *task;
452 	static int pid_max;
453 
454 	if (sched->pid_to_task == NULL) {
455 		if (sysctl__read_int("kernel/pid_max", &pid_max) < 0)
456 			pid_max = MAX_PID;
457 		BUG_ON((sched->pid_to_task = calloc(pid_max, sizeof(struct task_desc *))) == NULL);
458 	}
459 	if (pid >= (unsigned long)pid_max) {
460 		BUG_ON((sched->pid_to_task = realloc(sched->pid_to_task, (pid + 1) *
461 			sizeof(struct task_desc *))) == NULL);
462 		while (pid >= (unsigned long)pid_max)
463 			sched->pid_to_task[pid_max++] = NULL;
464 	}
465 
466 	task = sched->pid_to_task[pid];
467 
468 	if (task)
469 		return task;
470 
471 	task = zalloc(sizeof(*task));
472 	task->pid = pid;
473 	task->nr = sched->nr_tasks;
474 	strcpy(task->comm, comm);
475 	/*
476 	 * every task starts in sleeping state - this gets ignored
477 	 * if there's no wakeup pointing to this sleep state:
478 	 */
479 	add_sched_event_sleep(sched, task, 0, 0);
480 
481 	sched->pid_to_task[pid] = task;
482 	sched->nr_tasks++;
483 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_desc *));
484 	BUG_ON(!sched->tasks);
485 	sched->tasks[task->nr] = task;
486 
487 	if (verbose > 0)
488 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
489 
490 	return task;
491 }
492 
493 
494 static void print_task_traces(struct perf_sched *sched)
495 {
496 	struct task_desc *task;
497 	unsigned long i;
498 
499 	for (i = 0; i < sched->nr_tasks; i++) {
500 		task = sched->tasks[i];
501 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
502 			task->nr, task->comm, task->pid, task->nr_events);
503 	}
504 }
505 
506 static void add_cross_task_wakeups(struct perf_sched *sched)
507 {
508 	struct task_desc *task1, *task2;
509 	unsigned long i, j;
510 
511 	for (i = 0; i < sched->nr_tasks; i++) {
512 		task1 = sched->tasks[i];
513 		j = i + 1;
514 		if (j == sched->nr_tasks)
515 			j = 0;
516 		task2 = sched->tasks[j];
517 		add_sched_event_wakeup(sched, task1, 0, task2);
518 	}
519 }
520 
521 static void perf_sched__process_event(struct perf_sched *sched,
522 				      struct sched_atom *atom)
523 {
524 	int ret = 0;
525 
526 	switch (atom->type) {
527 		case SCHED_EVENT_RUN:
528 			burn_nsecs(sched, atom->duration);
529 			break;
530 		case SCHED_EVENT_SLEEP:
531 			if (atom->wait_sem)
532 				ret = sem_wait(atom->wait_sem);
533 			BUG_ON(ret);
534 			break;
535 		case SCHED_EVENT_WAKEUP:
536 			if (atom->wait_sem)
537 				ret = sem_post(atom->wait_sem);
538 			BUG_ON(ret);
539 			break;
540 		case SCHED_EVENT_MIGRATION:
541 			break;
542 		default:
543 			BUG_ON(1);
544 	}
545 }
546 
547 static u64 get_cpu_usage_nsec_parent(void)
548 {
549 	struct rusage ru;
550 	u64 sum;
551 	int err;
552 
553 	err = getrusage(RUSAGE_SELF, &ru);
554 	BUG_ON(err);
555 
556 	sum =  ru.ru_utime.tv_sec * NSEC_PER_SEC + ru.ru_utime.tv_usec * NSEC_PER_USEC;
557 	sum += ru.ru_stime.tv_sec * NSEC_PER_SEC + ru.ru_stime.tv_usec * NSEC_PER_USEC;
558 
559 	return sum;
560 }
561 
562 static int self_open_counters(struct perf_sched *sched, unsigned long cur_task)
563 {
564 	struct perf_event_attr attr;
565 	char sbuf[STRERR_BUFSIZE], info[STRERR_BUFSIZE];
566 	int fd;
567 	struct rlimit limit;
568 	bool need_privilege = false;
569 
570 	memset(&attr, 0, sizeof(attr));
571 
572 	attr.type = PERF_TYPE_SOFTWARE;
573 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
574 
575 force_again:
576 	fd = sys_perf_event_open(&attr, 0, -1, -1,
577 				 perf_event_open_cloexec_flag());
578 
579 	if (fd < 0) {
580 		if (errno == EMFILE) {
581 			if (sched->force) {
582 				BUG_ON(getrlimit(RLIMIT_NOFILE, &limit) == -1);
583 				limit.rlim_cur += sched->nr_tasks - cur_task;
584 				if (limit.rlim_cur > limit.rlim_max) {
585 					limit.rlim_max = limit.rlim_cur;
586 					need_privilege = true;
587 				}
588 				if (setrlimit(RLIMIT_NOFILE, &limit) == -1) {
589 					if (need_privilege && errno == EPERM)
590 						strcpy(info, "Need privilege\n");
591 				} else
592 					goto force_again;
593 			} else
594 				strcpy(info, "Have a try with -f option\n");
595 		}
596 		pr_err("Error: sys_perf_event_open() syscall returned "
597 		       "with %d (%s)\n%s", fd,
598 		       str_error_r(errno, sbuf, sizeof(sbuf)), info);
599 		exit(EXIT_FAILURE);
600 	}
601 	return fd;
602 }
603 
604 static u64 get_cpu_usage_nsec_self(int fd)
605 {
606 	u64 runtime;
607 	int ret;
608 
609 	ret = read(fd, &runtime, sizeof(runtime));
610 	BUG_ON(ret != sizeof(runtime));
611 
612 	return runtime;
613 }
614 
615 struct sched_thread_parms {
616 	struct task_desc  *task;
617 	struct perf_sched *sched;
618 	int fd;
619 };
620 
621 static void *thread_func(void *ctx)
622 {
623 	struct sched_thread_parms *parms = ctx;
624 	struct task_desc *this_task = parms->task;
625 	struct perf_sched *sched = parms->sched;
626 	u64 cpu_usage_0, cpu_usage_1;
627 	unsigned long i, ret;
628 	char comm2[22];
629 	int fd = parms->fd;
630 
631 	zfree(&parms);
632 
633 	sprintf(comm2, ":%s", this_task->comm);
634 	prctl(PR_SET_NAME, comm2);
635 	if (fd < 0)
636 		return NULL;
637 
638 	while (!sched->thread_funcs_exit) {
639 		ret = sem_post(&this_task->ready_for_work);
640 		BUG_ON(ret);
641 		mutex_lock(&sched->start_work_mutex);
642 		mutex_unlock(&sched->start_work_mutex);
643 
644 		cpu_usage_0 = get_cpu_usage_nsec_self(fd);
645 
646 		for (i = 0; i < this_task->nr_events; i++) {
647 			this_task->curr_event = i;
648 			perf_sched__process_event(sched, this_task->atoms[i]);
649 		}
650 
651 		cpu_usage_1 = get_cpu_usage_nsec_self(fd);
652 		this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
653 		ret = sem_post(&this_task->work_done_sem);
654 		BUG_ON(ret);
655 
656 		mutex_lock(&sched->work_done_wait_mutex);
657 		mutex_unlock(&sched->work_done_wait_mutex);
658 	}
659 	return NULL;
660 }
661 
662 static void create_tasks(struct perf_sched *sched)
663 	EXCLUSIVE_LOCK_FUNCTION(sched->start_work_mutex)
664 	EXCLUSIVE_LOCK_FUNCTION(sched->work_done_wait_mutex)
665 {
666 	struct task_desc *task;
667 	pthread_attr_t attr;
668 	unsigned long i;
669 	int err;
670 
671 	err = pthread_attr_init(&attr);
672 	BUG_ON(err);
673 	err = pthread_attr_setstacksize(&attr,
674 			(size_t) max(16 * 1024, (int)PTHREAD_STACK_MIN));
675 	BUG_ON(err);
676 	mutex_lock(&sched->start_work_mutex);
677 	mutex_lock(&sched->work_done_wait_mutex);
678 	for (i = 0; i < sched->nr_tasks; i++) {
679 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
680 		BUG_ON(parms == NULL);
681 		parms->task = task = sched->tasks[i];
682 		parms->sched = sched;
683 		parms->fd = self_open_counters(sched, i);
684 		sem_init(&task->sleep_sem, 0, 0);
685 		sem_init(&task->ready_for_work, 0, 0);
686 		sem_init(&task->work_done_sem, 0, 0);
687 		task->curr_event = 0;
688 		err = pthread_create(&task->thread, &attr, thread_func, parms);
689 		BUG_ON(err);
690 	}
691 }
692 
693 static void destroy_tasks(struct perf_sched *sched)
694 	UNLOCK_FUNCTION(sched->start_work_mutex)
695 	UNLOCK_FUNCTION(sched->work_done_wait_mutex)
696 {
697 	struct task_desc *task;
698 	unsigned long i;
699 	int err;
700 
701 	mutex_unlock(&sched->start_work_mutex);
702 	mutex_unlock(&sched->work_done_wait_mutex);
703 	/* Get rid of threads so they won't be upset by mutex destrunction */
704 	for (i = 0; i < sched->nr_tasks; i++) {
705 		task = sched->tasks[i];
706 		err = pthread_join(task->thread, NULL);
707 		BUG_ON(err);
708 		sem_destroy(&task->sleep_sem);
709 		sem_destroy(&task->ready_for_work);
710 		sem_destroy(&task->work_done_sem);
711 	}
712 }
713 
714 static void wait_for_tasks(struct perf_sched *sched)
715 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
716 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
717 {
718 	u64 cpu_usage_0, cpu_usage_1;
719 	struct task_desc *task;
720 	unsigned long i, ret;
721 
722 	sched->start_time = get_nsecs();
723 	sched->cpu_usage = 0;
724 	mutex_unlock(&sched->work_done_wait_mutex);
725 
726 	for (i = 0; i < sched->nr_tasks; i++) {
727 		task = sched->tasks[i];
728 		ret = sem_wait(&task->ready_for_work);
729 		BUG_ON(ret);
730 		sem_init(&task->ready_for_work, 0, 0);
731 	}
732 	mutex_lock(&sched->work_done_wait_mutex);
733 
734 	cpu_usage_0 = get_cpu_usage_nsec_parent();
735 
736 	mutex_unlock(&sched->start_work_mutex);
737 
738 	for (i = 0; i < sched->nr_tasks; i++) {
739 		task = sched->tasks[i];
740 		ret = sem_wait(&task->work_done_sem);
741 		BUG_ON(ret);
742 		sem_init(&task->work_done_sem, 0, 0);
743 		sched->cpu_usage += task->cpu_usage;
744 		task->cpu_usage = 0;
745 	}
746 
747 	cpu_usage_1 = get_cpu_usage_nsec_parent();
748 	if (!sched->runavg_cpu_usage)
749 		sched->runavg_cpu_usage = sched->cpu_usage;
750 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * (sched->replay_repeat - 1) + sched->cpu_usage) / sched->replay_repeat;
751 
752 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
753 	if (!sched->runavg_parent_cpu_usage)
754 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
755 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * (sched->replay_repeat - 1) +
756 					 sched->parent_cpu_usage)/sched->replay_repeat;
757 
758 	mutex_lock(&sched->start_work_mutex);
759 
760 	for (i = 0; i < sched->nr_tasks; i++) {
761 		task = sched->tasks[i];
762 		sem_init(&task->sleep_sem, 0, 0);
763 		task->curr_event = 0;
764 	}
765 }
766 
767 static void run_one_test(struct perf_sched *sched)
768 	EXCLUSIVE_LOCKS_REQUIRED(sched->work_done_wait_mutex)
769 	EXCLUSIVE_LOCKS_REQUIRED(sched->start_work_mutex)
770 {
771 	u64 T0, T1, delta, avg_delta, fluct;
772 
773 	T0 = get_nsecs();
774 	wait_for_tasks(sched);
775 	T1 = get_nsecs();
776 
777 	delta = T1 - T0;
778 	sched->sum_runtime += delta;
779 	sched->nr_runs++;
780 
781 	avg_delta = sched->sum_runtime / sched->nr_runs;
782 	if (delta < avg_delta)
783 		fluct = avg_delta - delta;
784 	else
785 		fluct = delta - avg_delta;
786 	sched->sum_fluct += fluct;
787 	if (!sched->run_avg)
788 		sched->run_avg = delta;
789 	sched->run_avg = (sched->run_avg * (sched->replay_repeat - 1) + delta) / sched->replay_repeat;
790 
791 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / NSEC_PER_MSEC);
792 
793 	printf("ravg: %0.2f, ", (double)sched->run_avg / NSEC_PER_MSEC);
794 
795 	printf("cpu: %0.2f / %0.2f",
796 		(double)sched->cpu_usage / NSEC_PER_MSEC, (double)sched->runavg_cpu_usage / NSEC_PER_MSEC);
797 
798 #if 0
799 	/*
800 	 * rusage statistics done by the parent, these are less
801 	 * accurate than the sched->sum_exec_runtime based statistics:
802 	 */
803 	printf(" [%0.2f / %0.2f]",
804 		(double)sched->parent_cpu_usage / NSEC_PER_MSEC,
805 		(double)sched->runavg_parent_cpu_usage / NSEC_PER_MSEC);
806 #endif
807 
808 	printf("\n");
809 
810 	if (sched->nr_sleep_corrections)
811 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
812 	sched->nr_sleep_corrections = 0;
813 }
814 
815 static void test_calibrations(struct perf_sched *sched)
816 {
817 	u64 T0, T1;
818 
819 	T0 = get_nsecs();
820 	burn_nsecs(sched, NSEC_PER_MSEC);
821 	T1 = get_nsecs();
822 
823 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
824 
825 	T0 = get_nsecs();
826 	sleep_nsecs(NSEC_PER_MSEC);
827 	T1 = get_nsecs();
828 
829 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
830 }
831 
832 static int
833 replay_wakeup_event(struct perf_sched *sched,
834 		    struct evsel *evsel, struct perf_sample *sample,
835 		    struct machine *machine __maybe_unused)
836 {
837 	const char *comm = evsel__strval(evsel, sample, "comm");
838 	const u32 pid	 = evsel__intval(evsel, sample, "pid");
839 	struct task_desc *waker, *wakee;
840 
841 	if (verbose > 0) {
842 		printf("sched_wakeup event %p\n", evsel);
843 
844 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
845 	}
846 
847 	waker = register_pid(sched, sample->tid, "<unknown>");
848 	wakee = register_pid(sched, pid, comm);
849 
850 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
851 	return 0;
852 }
853 
854 static int replay_switch_event(struct perf_sched *sched,
855 			       struct evsel *evsel,
856 			       struct perf_sample *sample,
857 			       struct machine *machine __maybe_unused)
858 {
859 	const char *prev_comm  = evsel__strval(evsel, sample, "prev_comm"),
860 		   *next_comm  = evsel__strval(evsel, sample, "next_comm");
861 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
862 		  next_pid = evsel__intval(evsel, sample, "next_pid");
863 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
864 	struct task_desc *prev, __maybe_unused *next;
865 	u64 timestamp0, timestamp = sample->time;
866 	int cpu = sample->cpu;
867 	s64 delta;
868 
869 	if (verbose > 0)
870 		printf("sched_switch event %p\n", evsel);
871 
872 	if (cpu >= MAX_CPUS || cpu < 0)
873 		return 0;
874 
875 	timestamp0 = sched->cpu_last_switched[cpu];
876 	if (timestamp0)
877 		delta = timestamp - timestamp0;
878 	else
879 		delta = 0;
880 
881 	if (delta < 0) {
882 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
883 		return -1;
884 	}
885 
886 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
887 		 prev_comm, prev_pid, next_comm, next_pid, delta);
888 
889 	prev = register_pid(sched, prev_pid, prev_comm);
890 	next = register_pid(sched, next_pid, next_comm);
891 
892 	sched->cpu_last_switched[cpu] = timestamp;
893 
894 	add_sched_event_run(sched, prev, timestamp, delta);
895 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
896 
897 	return 0;
898 }
899 
900 static int replay_fork_event(struct perf_sched *sched,
901 			     union perf_event *event,
902 			     struct machine *machine)
903 {
904 	struct thread *child, *parent;
905 
906 	child = machine__findnew_thread(machine, event->fork.pid,
907 					event->fork.tid);
908 	parent = machine__findnew_thread(machine, event->fork.ppid,
909 					 event->fork.ptid);
910 
911 	if (child == NULL || parent == NULL) {
912 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
913 				 child, parent);
914 		goto out_put;
915 	}
916 
917 	if (verbose > 0) {
918 		printf("fork event\n");
919 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
920 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
921 	}
922 
923 	register_pid(sched, parent->tid, thread__comm_str(parent));
924 	register_pid(sched, child->tid, thread__comm_str(child));
925 out_put:
926 	thread__put(child);
927 	thread__put(parent);
928 	return 0;
929 }
930 
931 struct sort_dimension {
932 	const char		*name;
933 	sort_fn_t		cmp;
934 	struct list_head	list;
935 };
936 
937 /*
938  * handle runtime stats saved per thread
939  */
940 static struct thread_runtime *thread__init_runtime(struct thread *thread)
941 {
942 	struct thread_runtime *r;
943 
944 	r = zalloc(sizeof(struct thread_runtime));
945 	if (!r)
946 		return NULL;
947 
948 	init_stats(&r->run_stats);
949 	thread__set_priv(thread, r);
950 
951 	return r;
952 }
953 
954 static struct thread_runtime *thread__get_runtime(struct thread *thread)
955 {
956 	struct thread_runtime *tr;
957 
958 	tr = thread__priv(thread);
959 	if (tr == NULL) {
960 		tr = thread__init_runtime(thread);
961 		if (tr == NULL)
962 			pr_debug("Failed to malloc memory for runtime data.\n");
963 	}
964 
965 	return tr;
966 }
967 
968 static int
969 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
970 {
971 	struct sort_dimension *sort;
972 	int ret = 0;
973 
974 	BUG_ON(list_empty(list));
975 
976 	list_for_each_entry(sort, list, list) {
977 		ret = sort->cmp(l, r);
978 		if (ret)
979 			return ret;
980 	}
981 
982 	return ret;
983 }
984 
985 static struct work_atoms *
986 thread_atoms_search(struct rb_root_cached *root, struct thread *thread,
987 			 struct list_head *sort_list)
988 {
989 	struct rb_node *node = root->rb_root.rb_node;
990 	struct work_atoms key = { .thread = thread };
991 
992 	while (node) {
993 		struct work_atoms *atoms;
994 		int cmp;
995 
996 		atoms = container_of(node, struct work_atoms, node);
997 
998 		cmp = thread_lat_cmp(sort_list, &key, atoms);
999 		if (cmp > 0)
1000 			node = node->rb_left;
1001 		else if (cmp < 0)
1002 			node = node->rb_right;
1003 		else {
1004 			BUG_ON(thread != atoms->thread);
1005 			return atoms;
1006 		}
1007 	}
1008 	return NULL;
1009 }
1010 
1011 static void
1012 __thread_latency_insert(struct rb_root_cached *root, struct work_atoms *data,
1013 			 struct list_head *sort_list)
1014 {
1015 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
1016 	bool leftmost = true;
1017 
1018 	while (*new) {
1019 		struct work_atoms *this;
1020 		int cmp;
1021 
1022 		this = container_of(*new, struct work_atoms, node);
1023 		parent = *new;
1024 
1025 		cmp = thread_lat_cmp(sort_list, data, this);
1026 
1027 		if (cmp > 0)
1028 			new = &((*new)->rb_left);
1029 		else {
1030 			new = &((*new)->rb_right);
1031 			leftmost = false;
1032 		}
1033 	}
1034 
1035 	rb_link_node(&data->node, parent, new);
1036 	rb_insert_color_cached(&data->node, root, leftmost);
1037 }
1038 
1039 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
1040 {
1041 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
1042 	if (!atoms) {
1043 		pr_err("No memory at %s\n", __func__);
1044 		return -1;
1045 	}
1046 
1047 	atoms->thread = thread__get(thread);
1048 	INIT_LIST_HEAD(&atoms->work_list);
1049 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
1050 	return 0;
1051 }
1052 
1053 static char sched_out_state(u64 prev_state)
1054 {
1055 	const char *str = TASK_STATE_TO_CHAR_STR;
1056 
1057 	return str[prev_state];
1058 }
1059 
1060 static int
1061 add_sched_out_event(struct work_atoms *atoms,
1062 		    char run_state,
1063 		    u64 timestamp)
1064 {
1065 	struct work_atom *atom = zalloc(sizeof(*atom));
1066 	if (!atom) {
1067 		pr_err("Non memory at %s", __func__);
1068 		return -1;
1069 	}
1070 
1071 	atom->sched_out_time = timestamp;
1072 
1073 	if (run_state == 'R') {
1074 		atom->state = THREAD_WAIT_CPU;
1075 		atom->wake_up_time = atom->sched_out_time;
1076 	}
1077 
1078 	list_add_tail(&atom->list, &atoms->work_list);
1079 	return 0;
1080 }
1081 
1082 static void
1083 add_runtime_event(struct work_atoms *atoms, u64 delta,
1084 		  u64 timestamp __maybe_unused)
1085 {
1086 	struct work_atom *atom;
1087 
1088 	BUG_ON(list_empty(&atoms->work_list));
1089 
1090 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1091 
1092 	atom->runtime += delta;
1093 	atoms->total_runtime += delta;
1094 }
1095 
1096 static void
1097 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1098 {
1099 	struct work_atom *atom;
1100 	u64 delta;
1101 
1102 	if (list_empty(&atoms->work_list))
1103 		return;
1104 
1105 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1106 
1107 	if (atom->state != THREAD_WAIT_CPU)
1108 		return;
1109 
1110 	if (timestamp < atom->wake_up_time) {
1111 		atom->state = THREAD_IGNORE;
1112 		return;
1113 	}
1114 
1115 	atom->state = THREAD_SCHED_IN;
1116 	atom->sched_in_time = timestamp;
1117 
1118 	delta = atom->sched_in_time - atom->wake_up_time;
1119 	atoms->total_lat += delta;
1120 	if (delta > atoms->max_lat) {
1121 		atoms->max_lat = delta;
1122 		atoms->max_lat_start = atom->wake_up_time;
1123 		atoms->max_lat_end = timestamp;
1124 	}
1125 	atoms->nb_atoms++;
1126 }
1127 
1128 static int latency_switch_event(struct perf_sched *sched,
1129 				struct evsel *evsel,
1130 				struct perf_sample *sample,
1131 				struct machine *machine)
1132 {
1133 	const u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1134 		  next_pid = evsel__intval(evsel, sample, "next_pid");
1135 	const u64 prev_state = evsel__intval(evsel, sample, "prev_state");
1136 	struct work_atoms *out_events, *in_events;
1137 	struct thread *sched_out, *sched_in;
1138 	u64 timestamp0, timestamp = sample->time;
1139 	int cpu = sample->cpu, err = -1;
1140 	s64 delta;
1141 
1142 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1143 
1144 	timestamp0 = sched->cpu_last_switched[cpu];
1145 	sched->cpu_last_switched[cpu] = timestamp;
1146 	if (timestamp0)
1147 		delta = timestamp - timestamp0;
1148 	else
1149 		delta = 0;
1150 
1151 	if (delta < 0) {
1152 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1153 		return -1;
1154 	}
1155 
1156 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
1157 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1158 	if (sched_out == NULL || sched_in == NULL)
1159 		goto out_put;
1160 
1161 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1162 	if (!out_events) {
1163 		if (thread_atoms_insert(sched, sched_out))
1164 			goto out_put;
1165 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
1166 		if (!out_events) {
1167 			pr_err("out-event: Internal tree error");
1168 			goto out_put;
1169 		}
1170 	}
1171 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
1172 		return -1;
1173 
1174 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1175 	if (!in_events) {
1176 		if (thread_atoms_insert(sched, sched_in))
1177 			goto out_put;
1178 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
1179 		if (!in_events) {
1180 			pr_err("in-event: Internal tree error");
1181 			goto out_put;
1182 		}
1183 		/*
1184 		 * Take came in we have not heard about yet,
1185 		 * add in an initial atom in runnable state:
1186 		 */
1187 		if (add_sched_out_event(in_events, 'R', timestamp))
1188 			goto out_put;
1189 	}
1190 	add_sched_in_event(in_events, timestamp);
1191 	err = 0;
1192 out_put:
1193 	thread__put(sched_out);
1194 	thread__put(sched_in);
1195 	return err;
1196 }
1197 
1198 static int latency_runtime_event(struct perf_sched *sched,
1199 				 struct evsel *evsel,
1200 				 struct perf_sample *sample,
1201 				 struct machine *machine)
1202 {
1203 	const u32 pid	   = evsel__intval(evsel, sample, "pid");
1204 	const u64 runtime  = evsel__intval(evsel, sample, "runtime");
1205 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
1206 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1207 	u64 timestamp = sample->time;
1208 	int cpu = sample->cpu, err = -1;
1209 
1210 	if (thread == NULL)
1211 		return -1;
1212 
1213 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1214 	if (!atoms) {
1215 		if (thread_atoms_insert(sched, thread))
1216 			goto out_put;
1217 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
1218 		if (!atoms) {
1219 			pr_err("in-event: Internal tree error");
1220 			goto out_put;
1221 		}
1222 		if (add_sched_out_event(atoms, 'R', timestamp))
1223 			goto out_put;
1224 	}
1225 
1226 	add_runtime_event(atoms, runtime, timestamp);
1227 	err = 0;
1228 out_put:
1229 	thread__put(thread);
1230 	return err;
1231 }
1232 
1233 static int latency_wakeup_event(struct perf_sched *sched,
1234 				struct evsel *evsel,
1235 				struct perf_sample *sample,
1236 				struct machine *machine)
1237 {
1238 	const u32 pid	  = evsel__intval(evsel, sample, "pid");
1239 	struct work_atoms *atoms;
1240 	struct work_atom *atom;
1241 	struct thread *wakee;
1242 	u64 timestamp = sample->time;
1243 	int err = -1;
1244 
1245 	wakee = machine__findnew_thread(machine, -1, pid);
1246 	if (wakee == NULL)
1247 		return -1;
1248 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1249 	if (!atoms) {
1250 		if (thread_atoms_insert(sched, wakee))
1251 			goto out_put;
1252 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1253 		if (!atoms) {
1254 			pr_err("wakeup-event: Internal tree error");
1255 			goto out_put;
1256 		}
1257 		if (add_sched_out_event(atoms, 'S', timestamp))
1258 			goto out_put;
1259 	}
1260 
1261 	BUG_ON(list_empty(&atoms->work_list));
1262 
1263 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1264 
1265 	/*
1266 	 * As we do not guarantee the wakeup event happens when
1267 	 * task is out of run queue, also may happen when task is
1268 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1269 	 * then we should not set the ->wake_up_time when wake up a
1270 	 * task which is on run queue.
1271 	 *
1272 	 * You WILL be missing events if you've recorded only
1273 	 * one CPU, or are only looking at only one, so don't
1274 	 * skip in this case.
1275 	 */
1276 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1277 		goto out_ok;
1278 
1279 	sched->nr_timestamps++;
1280 	if (atom->sched_out_time > timestamp) {
1281 		sched->nr_unordered_timestamps++;
1282 		goto out_ok;
1283 	}
1284 
1285 	atom->state = THREAD_WAIT_CPU;
1286 	atom->wake_up_time = timestamp;
1287 out_ok:
1288 	err = 0;
1289 out_put:
1290 	thread__put(wakee);
1291 	return err;
1292 }
1293 
1294 static int latency_migrate_task_event(struct perf_sched *sched,
1295 				      struct evsel *evsel,
1296 				      struct perf_sample *sample,
1297 				      struct machine *machine)
1298 {
1299 	const u32 pid = evsel__intval(evsel, sample, "pid");
1300 	u64 timestamp = sample->time;
1301 	struct work_atoms *atoms;
1302 	struct work_atom *atom;
1303 	struct thread *migrant;
1304 	int err = -1;
1305 
1306 	/*
1307 	 * Only need to worry about migration when profiling one CPU.
1308 	 */
1309 	if (sched->profile_cpu == -1)
1310 		return 0;
1311 
1312 	migrant = machine__findnew_thread(machine, -1, pid);
1313 	if (migrant == NULL)
1314 		return -1;
1315 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1316 	if (!atoms) {
1317 		if (thread_atoms_insert(sched, migrant))
1318 			goto out_put;
1319 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1320 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1321 		if (!atoms) {
1322 			pr_err("migration-event: Internal tree error");
1323 			goto out_put;
1324 		}
1325 		if (add_sched_out_event(atoms, 'R', timestamp))
1326 			goto out_put;
1327 	}
1328 
1329 	BUG_ON(list_empty(&atoms->work_list));
1330 
1331 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1332 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1333 
1334 	sched->nr_timestamps++;
1335 
1336 	if (atom->sched_out_time > timestamp)
1337 		sched->nr_unordered_timestamps++;
1338 	err = 0;
1339 out_put:
1340 	thread__put(migrant);
1341 	return err;
1342 }
1343 
1344 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1345 {
1346 	int i;
1347 	int ret;
1348 	u64 avg;
1349 	char max_lat_start[32], max_lat_end[32];
1350 
1351 	if (!work_list->nb_atoms)
1352 		return;
1353 	/*
1354 	 * Ignore idle threads:
1355 	 */
1356 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1357 		return;
1358 
1359 	sched->all_runtime += work_list->total_runtime;
1360 	sched->all_count   += work_list->nb_atoms;
1361 
1362 	if (work_list->num_merged > 1)
1363 		ret = printf("  %s:(%d) ", thread__comm_str(work_list->thread), work_list->num_merged);
1364 	else
1365 		ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1366 
1367 	for (i = 0; i < 24 - ret; i++)
1368 		printf(" ");
1369 
1370 	avg = work_list->total_lat / work_list->nb_atoms;
1371 	timestamp__scnprintf_usec(work_list->max_lat_start, max_lat_start, sizeof(max_lat_start));
1372 	timestamp__scnprintf_usec(work_list->max_lat_end, max_lat_end, sizeof(max_lat_end));
1373 
1374 	printf("|%11.3f ms |%9" PRIu64 " | avg:%8.3f ms | max:%8.3f ms | max start: %12s s | max end: %12s s\n",
1375 	      (double)work_list->total_runtime / NSEC_PER_MSEC,
1376 		 work_list->nb_atoms, (double)avg / NSEC_PER_MSEC,
1377 		 (double)work_list->max_lat / NSEC_PER_MSEC,
1378 		 max_lat_start, max_lat_end);
1379 }
1380 
1381 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1382 {
1383 	if (l->thread == r->thread)
1384 		return 0;
1385 	if (l->thread->tid < r->thread->tid)
1386 		return -1;
1387 	if (l->thread->tid > r->thread->tid)
1388 		return 1;
1389 	return (int)(l->thread - r->thread);
1390 }
1391 
1392 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1393 {
1394 	u64 avgl, avgr;
1395 
1396 	if (!l->nb_atoms)
1397 		return -1;
1398 
1399 	if (!r->nb_atoms)
1400 		return 1;
1401 
1402 	avgl = l->total_lat / l->nb_atoms;
1403 	avgr = r->total_lat / r->nb_atoms;
1404 
1405 	if (avgl < avgr)
1406 		return -1;
1407 	if (avgl > avgr)
1408 		return 1;
1409 
1410 	return 0;
1411 }
1412 
1413 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1414 {
1415 	if (l->max_lat < r->max_lat)
1416 		return -1;
1417 	if (l->max_lat > r->max_lat)
1418 		return 1;
1419 
1420 	return 0;
1421 }
1422 
1423 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1424 {
1425 	if (l->nb_atoms < r->nb_atoms)
1426 		return -1;
1427 	if (l->nb_atoms > r->nb_atoms)
1428 		return 1;
1429 
1430 	return 0;
1431 }
1432 
1433 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1434 {
1435 	if (l->total_runtime < r->total_runtime)
1436 		return -1;
1437 	if (l->total_runtime > r->total_runtime)
1438 		return 1;
1439 
1440 	return 0;
1441 }
1442 
1443 static int sort_dimension__add(const char *tok, struct list_head *list)
1444 {
1445 	size_t i;
1446 	static struct sort_dimension avg_sort_dimension = {
1447 		.name = "avg",
1448 		.cmp  = avg_cmp,
1449 	};
1450 	static struct sort_dimension max_sort_dimension = {
1451 		.name = "max",
1452 		.cmp  = max_cmp,
1453 	};
1454 	static struct sort_dimension pid_sort_dimension = {
1455 		.name = "pid",
1456 		.cmp  = pid_cmp,
1457 	};
1458 	static struct sort_dimension runtime_sort_dimension = {
1459 		.name = "runtime",
1460 		.cmp  = runtime_cmp,
1461 	};
1462 	static struct sort_dimension switch_sort_dimension = {
1463 		.name = "switch",
1464 		.cmp  = switch_cmp,
1465 	};
1466 	struct sort_dimension *available_sorts[] = {
1467 		&pid_sort_dimension,
1468 		&avg_sort_dimension,
1469 		&max_sort_dimension,
1470 		&switch_sort_dimension,
1471 		&runtime_sort_dimension,
1472 	};
1473 
1474 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1475 		if (!strcmp(available_sorts[i]->name, tok)) {
1476 			list_add_tail(&available_sorts[i]->list, list);
1477 
1478 			return 0;
1479 		}
1480 	}
1481 
1482 	return -1;
1483 }
1484 
1485 static void perf_sched__sort_lat(struct perf_sched *sched)
1486 {
1487 	struct rb_node *node;
1488 	struct rb_root_cached *root = &sched->atom_root;
1489 again:
1490 	for (;;) {
1491 		struct work_atoms *data;
1492 		node = rb_first_cached(root);
1493 		if (!node)
1494 			break;
1495 
1496 		rb_erase_cached(node, root);
1497 		data = rb_entry(node, struct work_atoms, node);
1498 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1499 	}
1500 	if (root == &sched->atom_root) {
1501 		root = &sched->merged_atom_root;
1502 		goto again;
1503 	}
1504 }
1505 
1506 static int process_sched_wakeup_event(struct perf_tool *tool,
1507 				      struct evsel *evsel,
1508 				      struct perf_sample *sample,
1509 				      struct machine *machine)
1510 {
1511 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1512 
1513 	if (sched->tp_handler->wakeup_event)
1514 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1515 
1516 	return 0;
1517 }
1518 
1519 static int process_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
1520 				      struct evsel *evsel __maybe_unused,
1521 				      struct perf_sample *sample __maybe_unused,
1522 				      struct machine *machine __maybe_unused)
1523 {
1524 	return 0;
1525 }
1526 
1527 union map_priv {
1528 	void	*ptr;
1529 	bool	 color;
1530 };
1531 
1532 static bool thread__has_color(struct thread *thread)
1533 {
1534 	union map_priv priv = {
1535 		.ptr = thread__priv(thread),
1536 	};
1537 
1538 	return priv.color;
1539 }
1540 
1541 static struct thread*
1542 map__findnew_thread(struct perf_sched *sched, struct machine *machine, pid_t pid, pid_t tid)
1543 {
1544 	struct thread *thread = machine__findnew_thread(machine, pid, tid);
1545 	union map_priv priv = {
1546 		.color = false,
1547 	};
1548 
1549 	if (!sched->map.color_pids || !thread || thread__priv(thread))
1550 		return thread;
1551 
1552 	if (thread_map__has(sched->map.color_pids, tid))
1553 		priv.color = true;
1554 
1555 	thread__set_priv(thread, priv.ptr);
1556 	return thread;
1557 }
1558 
1559 static int map_switch_event(struct perf_sched *sched, struct evsel *evsel,
1560 			    struct perf_sample *sample, struct machine *machine)
1561 {
1562 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
1563 	struct thread *sched_in;
1564 	struct thread_runtime *tr;
1565 	int new_shortname;
1566 	u64 timestamp0, timestamp = sample->time;
1567 	s64 delta;
1568 	int i;
1569 	struct perf_cpu this_cpu = {
1570 		.cpu = sample->cpu,
1571 	};
1572 	int cpus_nr;
1573 	bool new_cpu = false;
1574 	const char *color = PERF_COLOR_NORMAL;
1575 	char stimestamp[32];
1576 
1577 	BUG_ON(this_cpu.cpu >= MAX_CPUS || this_cpu.cpu < 0);
1578 
1579 	if (this_cpu.cpu > sched->max_cpu.cpu)
1580 		sched->max_cpu = this_cpu;
1581 
1582 	if (sched->map.comp) {
1583 		cpus_nr = bitmap_weight(sched->map.comp_cpus_mask, MAX_CPUS);
1584 		if (!__test_and_set_bit(this_cpu.cpu, sched->map.comp_cpus_mask)) {
1585 			sched->map.comp_cpus[cpus_nr++] = this_cpu;
1586 			new_cpu = true;
1587 		}
1588 	} else
1589 		cpus_nr = sched->max_cpu.cpu;
1590 
1591 	timestamp0 = sched->cpu_last_switched[this_cpu.cpu];
1592 	sched->cpu_last_switched[this_cpu.cpu] = timestamp;
1593 	if (timestamp0)
1594 		delta = timestamp - timestamp0;
1595 	else
1596 		delta = 0;
1597 
1598 	if (delta < 0) {
1599 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1600 		return -1;
1601 	}
1602 
1603 	sched_in = map__findnew_thread(sched, machine, -1, next_pid);
1604 	if (sched_in == NULL)
1605 		return -1;
1606 
1607 	tr = thread__get_runtime(sched_in);
1608 	if (tr == NULL) {
1609 		thread__put(sched_in);
1610 		return -1;
1611 	}
1612 
1613 	sched->curr_thread[this_cpu.cpu] = thread__get(sched_in);
1614 
1615 	printf("  ");
1616 
1617 	new_shortname = 0;
1618 	if (!tr->shortname[0]) {
1619 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1620 			/*
1621 			 * Don't allocate a letter-number for swapper:0
1622 			 * as a shortname. Instead, we use '.' for it.
1623 			 */
1624 			tr->shortname[0] = '.';
1625 			tr->shortname[1] = ' ';
1626 		} else {
1627 			tr->shortname[0] = sched->next_shortname1;
1628 			tr->shortname[1] = sched->next_shortname2;
1629 
1630 			if (sched->next_shortname1 < 'Z') {
1631 				sched->next_shortname1++;
1632 			} else {
1633 				sched->next_shortname1 = 'A';
1634 				if (sched->next_shortname2 < '9')
1635 					sched->next_shortname2++;
1636 				else
1637 					sched->next_shortname2 = '0';
1638 			}
1639 		}
1640 		new_shortname = 1;
1641 	}
1642 
1643 	for (i = 0; i < cpus_nr; i++) {
1644 		struct perf_cpu cpu = {
1645 			.cpu = sched->map.comp ? sched->map.comp_cpus[i].cpu : i,
1646 		};
1647 		struct thread *curr_thread = sched->curr_thread[cpu.cpu];
1648 		struct thread_runtime *curr_tr;
1649 		const char *pid_color = color;
1650 		const char *cpu_color = color;
1651 
1652 		if (curr_thread && thread__has_color(curr_thread))
1653 			pid_color = COLOR_PIDS;
1654 
1655 		if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, cpu))
1656 			continue;
1657 
1658 		if (sched->map.color_cpus && perf_cpu_map__has(sched->map.color_cpus, cpu))
1659 			cpu_color = COLOR_CPUS;
1660 
1661 		if (cpu.cpu != this_cpu.cpu)
1662 			color_fprintf(stdout, color, " ");
1663 		else
1664 			color_fprintf(stdout, cpu_color, "*");
1665 
1666 		if (sched->curr_thread[cpu.cpu]) {
1667 			curr_tr = thread__get_runtime(sched->curr_thread[cpu.cpu]);
1668 			if (curr_tr == NULL) {
1669 				thread__put(sched_in);
1670 				return -1;
1671 			}
1672 			color_fprintf(stdout, pid_color, "%2s ", curr_tr->shortname);
1673 		} else
1674 			color_fprintf(stdout, color, "   ");
1675 	}
1676 
1677 	if (sched->map.cpus && !perf_cpu_map__has(sched->map.cpus, this_cpu))
1678 		goto out;
1679 
1680 	timestamp__scnprintf_usec(timestamp, stimestamp, sizeof(stimestamp));
1681 	color_fprintf(stdout, color, "  %12s secs ", stimestamp);
1682 	if (new_shortname || tr->comm_changed || (verbose > 0 && sched_in->tid)) {
1683 		const char *pid_color = color;
1684 
1685 		if (thread__has_color(sched_in))
1686 			pid_color = COLOR_PIDS;
1687 
1688 		color_fprintf(stdout, pid_color, "%s => %s:%d",
1689 		       tr->shortname, thread__comm_str(sched_in), sched_in->tid);
1690 		tr->comm_changed = false;
1691 	}
1692 
1693 	if (sched->map.comp && new_cpu)
1694 		color_fprintf(stdout, color, " (CPU %d)", this_cpu);
1695 
1696 out:
1697 	color_fprintf(stdout, color, "\n");
1698 
1699 	thread__put(sched_in);
1700 
1701 	return 0;
1702 }
1703 
1704 static int process_sched_switch_event(struct perf_tool *tool,
1705 				      struct evsel *evsel,
1706 				      struct perf_sample *sample,
1707 				      struct machine *machine)
1708 {
1709 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1710 	int this_cpu = sample->cpu, err = 0;
1711 	u32 prev_pid = evsel__intval(evsel, sample, "prev_pid"),
1712 	    next_pid = evsel__intval(evsel, sample, "next_pid");
1713 
1714 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1715 		/*
1716 		 * Are we trying to switch away a PID that is
1717 		 * not current?
1718 		 */
1719 		if (sched->curr_pid[this_cpu] != prev_pid)
1720 			sched->nr_context_switch_bugs++;
1721 	}
1722 
1723 	if (sched->tp_handler->switch_event)
1724 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1725 
1726 	sched->curr_pid[this_cpu] = next_pid;
1727 	return err;
1728 }
1729 
1730 static int process_sched_runtime_event(struct perf_tool *tool,
1731 				       struct evsel *evsel,
1732 				       struct perf_sample *sample,
1733 				       struct machine *machine)
1734 {
1735 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1736 
1737 	if (sched->tp_handler->runtime_event)
1738 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1739 
1740 	return 0;
1741 }
1742 
1743 static int perf_sched__process_fork_event(struct perf_tool *tool,
1744 					  union perf_event *event,
1745 					  struct perf_sample *sample,
1746 					  struct machine *machine)
1747 {
1748 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1749 
1750 	/* run the fork event through the perf machinery */
1751 	perf_event__process_fork(tool, event, sample, machine);
1752 
1753 	/* and then run additional processing needed for this command */
1754 	if (sched->tp_handler->fork_event)
1755 		return sched->tp_handler->fork_event(sched, event, machine);
1756 
1757 	return 0;
1758 }
1759 
1760 static int process_sched_migrate_task_event(struct perf_tool *tool,
1761 					    struct evsel *evsel,
1762 					    struct perf_sample *sample,
1763 					    struct machine *machine)
1764 {
1765 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1766 
1767 	if (sched->tp_handler->migrate_task_event)
1768 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1769 
1770 	return 0;
1771 }
1772 
1773 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1774 				  struct evsel *evsel,
1775 				  struct perf_sample *sample,
1776 				  struct machine *machine);
1777 
1778 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1779 						 union perf_event *event __maybe_unused,
1780 						 struct perf_sample *sample,
1781 						 struct evsel *evsel,
1782 						 struct machine *machine)
1783 {
1784 	int err = 0;
1785 
1786 	if (evsel->handler != NULL) {
1787 		tracepoint_handler f = evsel->handler;
1788 		err = f(tool, evsel, sample, machine);
1789 	}
1790 
1791 	return err;
1792 }
1793 
1794 static int perf_sched__process_comm(struct perf_tool *tool __maybe_unused,
1795 				    union perf_event *event,
1796 				    struct perf_sample *sample,
1797 				    struct machine *machine)
1798 {
1799 	struct thread *thread;
1800 	struct thread_runtime *tr;
1801 	int err;
1802 
1803 	err = perf_event__process_comm(tool, event, sample, machine);
1804 	if (err)
1805 		return err;
1806 
1807 	thread = machine__find_thread(machine, sample->pid, sample->tid);
1808 	if (!thread) {
1809 		pr_err("Internal error: can't find thread\n");
1810 		return -1;
1811 	}
1812 
1813 	tr = thread__get_runtime(thread);
1814 	if (tr == NULL) {
1815 		thread__put(thread);
1816 		return -1;
1817 	}
1818 
1819 	tr->comm_changed = true;
1820 	thread__put(thread);
1821 
1822 	return 0;
1823 }
1824 
1825 static int perf_sched__read_events(struct perf_sched *sched)
1826 {
1827 	struct evsel_str_handler handlers[] = {
1828 		{ "sched:sched_switch",	      process_sched_switch_event, },
1829 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1830 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1831 		{ "sched:sched_waking",	      process_sched_wakeup_event, },
1832 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1833 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1834 	};
1835 	struct perf_session *session;
1836 	struct perf_data data = {
1837 		.path  = input_name,
1838 		.mode  = PERF_DATA_MODE_READ,
1839 		.force = sched->force,
1840 	};
1841 	int rc = -1;
1842 
1843 	session = perf_session__new(&data, &sched->tool);
1844 	if (IS_ERR(session)) {
1845 		pr_debug("Error creating perf session");
1846 		return PTR_ERR(session);
1847 	}
1848 
1849 	symbol__init(&session->header.env);
1850 
1851 	/* prefer sched_waking if it is captured */
1852 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
1853 		handlers[2].handler = process_sched_wakeup_ignore;
1854 
1855 	if (perf_session__set_tracepoints_handlers(session, handlers))
1856 		goto out_delete;
1857 
1858 	if (perf_session__has_traces(session, "record -R")) {
1859 		int err = perf_session__process_events(session);
1860 		if (err) {
1861 			pr_err("Failed to process events, error %d", err);
1862 			goto out_delete;
1863 		}
1864 
1865 		sched->nr_events      = session->evlist->stats.nr_events[0];
1866 		sched->nr_lost_events = session->evlist->stats.total_lost;
1867 		sched->nr_lost_chunks = session->evlist->stats.nr_events[PERF_RECORD_LOST];
1868 	}
1869 
1870 	rc = 0;
1871 out_delete:
1872 	perf_session__delete(session);
1873 	return rc;
1874 }
1875 
1876 /*
1877  * scheduling times are printed as msec.usec
1878  */
1879 static inline void print_sched_time(unsigned long long nsecs, int width)
1880 {
1881 	unsigned long msecs;
1882 	unsigned long usecs;
1883 
1884 	msecs  = nsecs / NSEC_PER_MSEC;
1885 	nsecs -= msecs * NSEC_PER_MSEC;
1886 	usecs  = nsecs / NSEC_PER_USEC;
1887 	printf("%*lu.%03lu ", width, msecs, usecs);
1888 }
1889 
1890 /*
1891  * returns runtime data for event, allocating memory for it the
1892  * first time it is used.
1893  */
1894 static struct evsel_runtime *evsel__get_runtime(struct evsel *evsel)
1895 {
1896 	struct evsel_runtime *r = evsel->priv;
1897 
1898 	if (r == NULL) {
1899 		r = zalloc(sizeof(struct evsel_runtime));
1900 		evsel->priv = r;
1901 	}
1902 
1903 	return r;
1904 }
1905 
1906 /*
1907  * save last time event was seen per cpu
1908  */
1909 static void evsel__save_time(struct evsel *evsel, u64 timestamp, u32 cpu)
1910 {
1911 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1912 
1913 	if (r == NULL)
1914 		return;
1915 
1916 	if ((cpu >= r->ncpu) || (r->last_time == NULL)) {
1917 		int i, n = __roundup_pow_of_two(cpu+1);
1918 		void *p = r->last_time;
1919 
1920 		p = realloc(r->last_time, n * sizeof(u64));
1921 		if (!p)
1922 			return;
1923 
1924 		r->last_time = p;
1925 		for (i = r->ncpu; i < n; ++i)
1926 			r->last_time[i] = (u64) 0;
1927 
1928 		r->ncpu = n;
1929 	}
1930 
1931 	r->last_time[cpu] = timestamp;
1932 }
1933 
1934 /* returns last time this event was seen on the given cpu */
1935 static u64 evsel__get_time(struct evsel *evsel, u32 cpu)
1936 {
1937 	struct evsel_runtime *r = evsel__get_runtime(evsel);
1938 
1939 	if ((r == NULL) || (r->last_time == NULL) || (cpu >= r->ncpu))
1940 		return 0;
1941 
1942 	return r->last_time[cpu];
1943 }
1944 
1945 static int comm_width = 30;
1946 
1947 static char *timehist_get_commstr(struct thread *thread)
1948 {
1949 	static char str[32];
1950 	const char *comm = thread__comm_str(thread);
1951 	pid_t tid = thread->tid;
1952 	pid_t pid = thread->pid_;
1953 	int n;
1954 
1955 	if (pid == 0)
1956 		n = scnprintf(str, sizeof(str), "%s", comm);
1957 
1958 	else if (tid != pid)
1959 		n = scnprintf(str, sizeof(str), "%s[%d/%d]", comm, tid, pid);
1960 
1961 	else
1962 		n = scnprintf(str, sizeof(str), "%s[%d]", comm, tid);
1963 
1964 	if (n > comm_width)
1965 		comm_width = n;
1966 
1967 	return str;
1968 }
1969 
1970 static void timehist_header(struct perf_sched *sched)
1971 {
1972 	u32 ncpus = sched->max_cpu.cpu + 1;
1973 	u32 i, j;
1974 
1975 	printf("%15s %6s ", "time", "cpu");
1976 
1977 	if (sched->show_cpu_visual) {
1978 		printf(" ");
1979 		for (i = 0, j = 0; i < ncpus; ++i) {
1980 			printf("%x", j++);
1981 			if (j > 15)
1982 				j = 0;
1983 		}
1984 		printf(" ");
1985 	}
1986 
1987 	printf(" %-*s  %9s  %9s  %9s", comm_width,
1988 		"task name", "wait time", "sch delay", "run time");
1989 
1990 	if (sched->show_state)
1991 		printf("  %s", "state");
1992 
1993 	printf("\n");
1994 
1995 	/*
1996 	 * units row
1997 	 */
1998 	printf("%15s %-6s ", "", "");
1999 
2000 	if (sched->show_cpu_visual)
2001 		printf(" %*s ", ncpus, "");
2002 
2003 	printf(" %-*s  %9s  %9s  %9s", comm_width,
2004 	       "[tid/pid]", "(msec)", "(msec)", "(msec)");
2005 
2006 	if (sched->show_state)
2007 		printf("  %5s", "");
2008 
2009 	printf("\n");
2010 
2011 	/*
2012 	 * separator
2013 	 */
2014 	printf("%.15s %.6s ", graph_dotted_line, graph_dotted_line);
2015 
2016 	if (sched->show_cpu_visual)
2017 		printf(" %.*s ", ncpus, graph_dotted_line);
2018 
2019 	printf(" %.*s  %.9s  %.9s  %.9s", comm_width,
2020 		graph_dotted_line, graph_dotted_line, graph_dotted_line,
2021 		graph_dotted_line);
2022 
2023 	if (sched->show_state)
2024 		printf("  %.5s", graph_dotted_line);
2025 
2026 	printf("\n");
2027 }
2028 
2029 static char task_state_char(struct thread *thread, int state)
2030 {
2031 	static const char state_to_char[] = TASK_STATE_TO_CHAR_STR;
2032 	unsigned bit = state ? ffs(state) : 0;
2033 
2034 	/* 'I' for idle */
2035 	if (thread->tid == 0)
2036 		return 'I';
2037 
2038 	return bit < sizeof(state_to_char) - 1 ? state_to_char[bit] : '?';
2039 }
2040 
2041 static void timehist_print_sample(struct perf_sched *sched,
2042 				  struct evsel *evsel,
2043 				  struct perf_sample *sample,
2044 				  struct addr_location *al,
2045 				  struct thread *thread,
2046 				  u64 t, int state)
2047 {
2048 	struct thread_runtime *tr = thread__priv(thread);
2049 	const char *next_comm = evsel__strval(evsel, sample, "next_comm");
2050 	const u32 next_pid = evsel__intval(evsel, sample, "next_pid");
2051 	u32 max_cpus = sched->max_cpu.cpu + 1;
2052 	char tstr[64];
2053 	char nstr[30];
2054 	u64 wait_time;
2055 
2056 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap))
2057 		return;
2058 
2059 	timestamp__scnprintf_usec(t, tstr, sizeof(tstr));
2060 	printf("%15s [%04d] ", tstr, sample->cpu);
2061 
2062 	if (sched->show_cpu_visual) {
2063 		u32 i;
2064 		char c;
2065 
2066 		printf(" ");
2067 		for (i = 0; i < max_cpus; ++i) {
2068 			/* flag idle times with 'i'; others are sched events */
2069 			if (i == sample->cpu)
2070 				c = (thread->tid == 0) ? 'i' : 's';
2071 			else
2072 				c = ' ';
2073 			printf("%c", c);
2074 		}
2075 		printf(" ");
2076 	}
2077 
2078 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2079 
2080 	wait_time = tr->dt_sleep + tr->dt_iowait + tr->dt_preempt;
2081 	print_sched_time(wait_time, 6);
2082 
2083 	print_sched_time(tr->dt_delay, 6);
2084 	print_sched_time(tr->dt_run, 6);
2085 
2086 	if (sched->show_state)
2087 		printf(" %5c ", task_state_char(thread, state));
2088 
2089 	if (sched->show_next) {
2090 		snprintf(nstr, sizeof(nstr), "next: %s[%d]", next_comm, next_pid);
2091 		printf(" %-*s", comm_width, nstr);
2092 	}
2093 
2094 	if (sched->show_wakeups && !sched->show_next)
2095 		printf("  %-*s", comm_width, "");
2096 
2097 	if (thread->tid == 0)
2098 		goto out;
2099 
2100 	if (sched->show_callchain)
2101 		printf("  ");
2102 
2103 	sample__fprintf_sym(sample, al, 0,
2104 			    EVSEL__PRINT_SYM | EVSEL__PRINT_ONELINE |
2105 			    EVSEL__PRINT_CALLCHAIN_ARROW |
2106 			    EVSEL__PRINT_SKIP_IGNORED,
2107 			    &callchain_cursor, symbol_conf.bt_stop_list,  stdout);
2108 
2109 out:
2110 	printf("\n");
2111 }
2112 
2113 /*
2114  * Explanation of delta-time stats:
2115  *
2116  *            t = time of current schedule out event
2117  *        tprev = time of previous sched out event
2118  *                also time of schedule-in event for current task
2119  *    last_time = time of last sched change event for current task
2120  *                (i.e, time process was last scheduled out)
2121  * ready_to_run = time of wakeup for current task
2122  *
2123  * -----|------------|------------|------------|------
2124  *    last         ready        tprev          t
2125  *    time         to run
2126  *
2127  *      |-------- dt_wait --------|
2128  *                   |- dt_delay -|-- dt_run --|
2129  *
2130  *   dt_run = run time of current task
2131  *  dt_wait = time between last schedule out event for task and tprev
2132  *            represents time spent off the cpu
2133  * dt_delay = time between wakeup and schedule-in of task
2134  */
2135 
2136 static void timehist_update_runtime_stats(struct thread_runtime *r,
2137 					 u64 t, u64 tprev)
2138 {
2139 	r->dt_delay   = 0;
2140 	r->dt_sleep   = 0;
2141 	r->dt_iowait  = 0;
2142 	r->dt_preempt = 0;
2143 	r->dt_run     = 0;
2144 
2145 	if (tprev) {
2146 		r->dt_run = t - tprev;
2147 		if (r->ready_to_run) {
2148 			if (r->ready_to_run > tprev)
2149 				pr_debug("time travel: wakeup time for task > previous sched_switch event\n");
2150 			else
2151 				r->dt_delay = tprev - r->ready_to_run;
2152 		}
2153 
2154 		if (r->last_time > tprev)
2155 			pr_debug("time travel: last sched out time for task > previous sched_switch event\n");
2156 		else if (r->last_time) {
2157 			u64 dt_wait = tprev - r->last_time;
2158 
2159 			if (r->last_state == TASK_RUNNING)
2160 				r->dt_preempt = dt_wait;
2161 			else if (r->last_state == TASK_UNINTERRUPTIBLE)
2162 				r->dt_iowait = dt_wait;
2163 			else
2164 				r->dt_sleep = dt_wait;
2165 		}
2166 	}
2167 
2168 	update_stats(&r->run_stats, r->dt_run);
2169 
2170 	r->total_run_time     += r->dt_run;
2171 	r->total_delay_time   += r->dt_delay;
2172 	r->total_sleep_time   += r->dt_sleep;
2173 	r->total_iowait_time  += r->dt_iowait;
2174 	r->total_preempt_time += r->dt_preempt;
2175 }
2176 
2177 static bool is_idle_sample(struct perf_sample *sample,
2178 			   struct evsel *evsel)
2179 {
2180 	/* pid 0 == swapper == idle task */
2181 	if (strcmp(evsel__name(evsel), "sched:sched_switch") == 0)
2182 		return evsel__intval(evsel, sample, "prev_pid") == 0;
2183 
2184 	return sample->pid == 0;
2185 }
2186 
2187 static void save_task_callchain(struct perf_sched *sched,
2188 				struct perf_sample *sample,
2189 				struct evsel *evsel,
2190 				struct machine *machine)
2191 {
2192 	struct callchain_cursor *cursor = &callchain_cursor;
2193 	struct thread *thread;
2194 
2195 	/* want main thread for process - has maps */
2196 	thread = machine__findnew_thread(machine, sample->pid, sample->pid);
2197 	if (thread == NULL) {
2198 		pr_debug("Failed to get thread for pid %d.\n", sample->pid);
2199 		return;
2200 	}
2201 
2202 	if (!sched->show_callchain || sample->callchain == NULL)
2203 		return;
2204 
2205 	if (thread__resolve_callchain(thread, cursor, evsel, sample,
2206 				      NULL, NULL, sched->max_stack + 2) != 0) {
2207 		if (verbose > 0)
2208 			pr_err("Failed to resolve callchain. Skipping\n");
2209 
2210 		return;
2211 	}
2212 
2213 	callchain_cursor_commit(cursor);
2214 
2215 	while (true) {
2216 		struct callchain_cursor_node *node;
2217 		struct symbol *sym;
2218 
2219 		node = callchain_cursor_current(cursor);
2220 		if (node == NULL)
2221 			break;
2222 
2223 		sym = node->ms.sym;
2224 		if (sym) {
2225 			if (!strcmp(sym->name, "schedule") ||
2226 			    !strcmp(sym->name, "__schedule") ||
2227 			    !strcmp(sym->name, "preempt_schedule"))
2228 				sym->ignore = 1;
2229 		}
2230 
2231 		callchain_cursor_advance(cursor);
2232 	}
2233 }
2234 
2235 static int init_idle_thread(struct thread *thread)
2236 {
2237 	struct idle_thread_runtime *itr;
2238 
2239 	thread__set_comm(thread, idle_comm, 0);
2240 
2241 	itr = zalloc(sizeof(*itr));
2242 	if (itr == NULL)
2243 		return -ENOMEM;
2244 
2245 	init_stats(&itr->tr.run_stats);
2246 	callchain_init(&itr->callchain);
2247 	callchain_cursor_reset(&itr->cursor);
2248 	thread__set_priv(thread, itr);
2249 
2250 	return 0;
2251 }
2252 
2253 /*
2254  * Track idle stats per cpu by maintaining a local thread
2255  * struct for the idle task on each cpu.
2256  */
2257 static int init_idle_threads(int ncpu)
2258 {
2259 	int i, ret;
2260 
2261 	idle_threads = zalloc(ncpu * sizeof(struct thread *));
2262 	if (!idle_threads)
2263 		return -ENOMEM;
2264 
2265 	idle_max_cpu = ncpu;
2266 
2267 	/* allocate the actual thread struct if needed */
2268 	for (i = 0; i < ncpu; ++i) {
2269 		idle_threads[i] = thread__new(0, 0);
2270 		if (idle_threads[i] == NULL)
2271 			return -ENOMEM;
2272 
2273 		ret = init_idle_thread(idle_threads[i]);
2274 		if (ret < 0)
2275 			return ret;
2276 	}
2277 
2278 	return 0;
2279 }
2280 
2281 static void free_idle_threads(void)
2282 {
2283 	int i;
2284 
2285 	if (idle_threads == NULL)
2286 		return;
2287 
2288 	for (i = 0; i < idle_max_cpu; ++i) {
2289 		if ((idle_threads[i]))
2290 			thread__delete(idle_threads[i]);
2291 	}
2292 
2293 	free(idle_threads);
2294 }
2295 
2296 static struct thread *get_idle_thread(int cpu)
2297 {
2298 	/*
2299 	 * expand/allocate array of pointers to local thread
2300 	 * structs if needed
2301 	 */
2302 	if ((cpu >= idle_max_cpu) || (idle_threads == NULL)) {
2303 		int i, j = __roundup_pow_of_two(cpu+1);
2304 		void *p;
2305 
2306 		p = realloc(idle_threads, j * sizeof(struct thread *));
2307 		if (!p)
2308 			return NULL;
2309 
2310 		idle_threads = (struct thread **) p;
2311 		for (i = idle_max_cpu; i < j; ++i)
2312 			idle_threads[i] = NULL;
2313 
2314 		idle_max_cpu = j;
2315 	}
2316 
2317 	/* allocate a new thread struct if needed */
2318 	if (idle_threads[cpu] == NULL) {
2319 		idle_threads[cpu] = thread__new(0, 0);
2320 
2321 		if (idle_threads[cpu]) {
2322 			if (init_idle_thread(idle_threads[cpu]) < 0)
2323 				return NULL;
2324 		}
2325 	}
2326 
2327 	return idle_threads[cpu];
2328 }
2329 
2330 static void save_idle_callchain(struct perf_sched *sched,
2331 				struct idle_thread_runtime *itr,
2332 				struct perf_sample *sample)
2333 {
2334 	if (!sched->show_callchain || sample->callchain == NULL)
2335 		return;
2336 
2337 	callchain_cursor__copy(&itr->cursor, &callchain_cursor);
2338 }
2339 
2340 static struct thread *timehist_get_thread(struct perf_sched *sched,
2341 					  struct perf_sample *sample,
2342 					  struct machine *machine,
2343 					  struct evsel *evsel)
2344 {
2345 	struct thread *thread;
2346 
2347 	if (is_idle_sample(sample, evsel)) {
2348 		thread = get_idle_thread(sample->cpu);
2349 		if (thread == NULL)
2350 			pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2351 
2352 	} else {
2353 		/* there were samples with tid 0 but non-zero pid */
2354 		thread = machine__findnew_thread(machine, sample->pid,
2355 						 sample->tid ?: sample->pid);
2356 		if (thread == NULL) {
2357 			pr_debug("Failed to get thread for tid %d. skipping sample.\n",
2358 				 sample->tid);
2359 		}
2360 
2361 		save_task_callchain(sched, sample, evsel, machine);
2362 		if (sched->idle_hist) {
2363 			struct thread *idle;
2364 			struct idle_thread_runtime *itr;
2365 
2366 			idle = get_idle_thread(sample->cpu);
2367 			if (idle == NULL) {
2368 				pr_err("Failed to get idle thread for cpu %d.\n", sample->cpu);
2369 				return NULL;
2370 			}
2371 
2372 			itr = thread__priv(idle);
2373 			if (itr == NULL)
2374 				return NULL;
2375 
2376 			itr->last_thread = thread;
2377 
2378 			/* copy task callchain when entering to idle */
2379 			if (evsel__intval(evsel, sample, "next_pid") == 0)
2380 				save_idle_callchain(sched, itr, sample);
2381 		}
2382 	}
2383 
2384 	return thread;
2385 }
2386 
2387 static bool timehist_skip_sample(struct perf_sched *sched,
2388 				 struct thread *thread,
2389 				 struct evsel *evsel,
2390 				 struct perf_sample *sample)
2391 {
2392 	bool rc = false;
2393 
2394 	if (thread__is_filtered(thread)) {
2395 		rc = true;
2396 		sched->skipped_samples++;
2397 	}
2398 
2399 	if (sched->idle_hist) {
2400 		if (strcmp(evsel__name(evsel), "sched:sched_switch"))
2401 			rc = true;
2402 		else if (evsel__intval(evsel, sample, "prev_pid") != 0 &&
2403 			 evsel__intval(evsel, sample, "next_pid") != 0)
2404 			rc = true;
2405 	}
2406 
2407 	return rc;
2408 }
2409 
2410 static void timehist_print_wakeup_event(struct perf_sched *sched,
2411 					struct evsel *evsel,
2412 					struct perf_sample *sample,
2413 					struct machine *machine,
2414 					struct thread *awakened)
2415 {
2416 	struct thread *thread;
2417 	char tstr[64];
2418 
2419 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2420 	if (thread == NULL)
2421 		return;
2422 
2423 	/* show wakeup unless both awakee and awaker are filtered */
2424 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2425 	    timehist_skip_sample(sched, awakened, evsel, sample)) {
2426 		return;
2427 	}
2428 
2429 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2430 	printf("%15s [%04d] ", tstr, sample->cpu);
2431 	if (sched->show_cpu_visual)
2432 		printf(" %*s ", sched->max_cpu.cpu + 1, "");
2433 
2434 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2435 
2436 	/* dt spacer */
2437 	printf("  %9s  %9s  %9s ", "", "", "");
2438 
2439 	printf("awakened: %s", timehist_get_commstr(awakened));
2440 
2441 	printf("\n");
2442 }
2443 
2444 static int timehist_sched_wakeup_ignore(struct perf_tool *tool __maybe_unused,
2445 					union perf_event *event __maybe_unused,
2446 					struct evsel *evsel __maybe_unused,
2447 					struct perf_sample *sample __maybe_unused,
2448 					struct machine *machine __maybe_unused)
2449 {
2450 	return 0;
2451 }
2452 
2453 static int timehist_sched_wakeup_event(struct perf_tool *tool,
2454 				       union perf_event *event __maybe_unused,
2455 				       struct evsel *evsel,
2456 				       struct perf_sample *sample,
2457 				       struct machine *machine)
2458 {
2459 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2460 	struct thread *thread;
2461 	struct thread_runtime *tr = NULL;
2462 	/* want pid of awakened task not pid in sample */
2463 	const u32 pid = evsel__intval(evsel, sample, "pid");
2464 
2465 	thread = machine__findnew_thread(machine, 0, pid);
2466 	if (thread == NULL)
2467 		return -1;
2468 
2469 	tr = thread__get_runtime(thread);
2470 	if (tr == NULL)
2471 		return -1;
2472 
2473 	if (tr->ready_to_run == 0)
2474 		tr->ready_to_run = sample->time;
2475 
2476 	/* show wakeups if requested */
2477 	if (sched->show_wakeups &&
2478 	    !perf_time__skip_sample(&sched->ptime, sample->time))
2479 		timehist_print_wakeup_event(sched, evsel, sample, machine, thread);
2480 
2481 	return 0;
2482 }
2483 
2484 static void timehist_print_migration_event(struct perf_sched *sched,
2485 					struct evsel *evsel,
2486 					struct perf_sample *sample,
2487 					struct machine *machine,
2488 					struct thread *migrated)
2489 {
2490 	struct thread *thread;
2491 	char tstr[64];
2492 	u32 max_cpus;
2493 	u32 ocpu, dcpu;
2494 
2495 	if (sched->summary_only)
2496 		return;
2497 
2498 	max_cpus = sched->max_cpu.cpu + 1;
2499 	ocpu = evsel__intval(evsel, sample, "orig_cpu");
2500 	dcpu = evsel__intval(evsel, sample, "dest_cpu");
2501 
2502 	thread = machine__findnew_thread(machine, sample->pid, sample->tid);
2503 	if (thread == NULL)
2504 		return;
2505 
2506 	if (timehist_skip_sample(sched, thread, evsel, sample) &&
2507 	    timehist_skip_sample(sched, migrated, evsel, sample)) {
2508 		return;
2509 	}
2510 
2511 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2512 	printf("%15s [%04d] ", tstr, sample->cpu);
2513 
2514 	if (sched->show_cpu_visual) {
2515 		u32 i;
2516 		char c;
2517 
2518 		printf("  ");
2519 		for (i = 0; i < max_cpus; ++i) {
2520 			c = (i == sample->cpu) ? 'm' : ' ';
2521 			printf("%c", c);
2522 		}
2523 		printf("  ");
2524 	}
2525 
2526 	printf(" %-*s ", comm_width, timehist_get_commstr(thread));
2527 
2528 	/* dt spacer */
2529 	printf("  %9s  %9s  %9s ", "", "", "");
2530 
2531 	printf("migrated: %s", timehist_get_commstr(migrated));
2532 	printf(" cpu %d => %d", ocpu, dcpu);
2533 
2534 	printf("\n");
2535 }
2536 
2537 static int timehist_migrate_task_event(struct perf_tool *tool,
2538 				       union perf_event *event __maybe_unused,
2539 				       struct evsel *evsel,
2540 				       struct perf_sample *sample,
2541 				       struct machine *machine)
2542 {
2543 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2544 	struct thread *thread;
2545 	struct thread_runtime *tr = NULL;
2546 	/* want pid of migrated task not pid in sample */
2547 	const u32 pid = evsel__intval(evsel, sample, "pid");
2548 
2549 	thread = machine__findnew_thread(machine, 0, pid);
2550 	if (thread == NULL)
2551 		return -1;
2552 
2553 	tr = thread__get_runtime(thread);
2554 	if (tr == NULL)
2555 		return -1;
2556 
2557 	tr->migrations++;
2558 
2559 	/* show migrations if requested */
2560 	timehist_print_migration_event(sched, evsel, sample, machine, thread);
2561 
2562 	return 0;
2563 }
2564 
2565 static int timehist_sched_change_event(struct perf_tool *tool,
2566 				       union perf_event *event,
2567 				       struct evsel *evsel,
2568 				       struct perf_sample *sample,
2569 				       struct machine *machine)
2570 {
2571 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2572 	struct perf_time_interval *ptime = &sched->ptime;
2573 	struct addr_location al;
2574 	struct thread *thread;
2575 	struct thread_runtime *tr = NULL;
2576 	u64 tprev, t = sample->time;
2577 	int rc = 0;
2578 	int state = evsel__intval(evsel, sample, "prev_state");
2579 
2580 	if (machine__resolve(machine, &al, sample) < 0) {
2581 		pr_err("problem processing %d event. skipping it\n",
2582 		       event->header.type);
2583 		rc = -1;
2584 		goto out;
2585 	}
2586 
2587 	thread = timehist_get_thread(sched, sample, machine, evsel);
2588 	if (thread == NULL) {
2589 		rc = -1;
2590 		goto out;
2591 	}
2592 
2593 	if (timehist_skip_sample(sched, thread, evsel, sample))
2594 		goto out;
2595 
2596 	tr = thread__get_runtime(thread);
2597 	if (tr == NULL) {
2598 		rc = -1;
2599 		goto out;
2600 	}
2601 
2602 	tprev = evsel__get_time(evsel, sample->cpu);
2603 
2604 	/*
2605 	 * If start time given:
2606 	 * - sample time is under window user cares about - skip sample
2607 	 * - tprev is under window user cares about  - reset to start of window
2608 	 */
2609 	if (ptime->start && ptime->start > t)
2610 		goto out;
2611 
2612 	if (tprev && ptime->start > tprev)
2613 		tprev = ptime->start;
2614 
2615 	/*
2616 	 * If end time given:
2617 	 * - previous sched event is out of window - we are done
2618 	 * - sample time is beyond window user cares about - reset it
2619 	 *   to close out stats for time window interest
2620 	 */
2621 	if (ptime->end) {
2622 		if (tprev > ptime->end)
2623 			goto out;
2624 
2625 		if (t > ptime->end)
2626 			t = ptime->end;
2627 	}
2628 
2629 	if (!sched->idle_hist || thread->tid == 0) {
2630 		if (!cpu_list || test_bit(sample->cpu, cpu_bitmap))
2631 			timehist_update_runtime_stats(tr, t, tprev);
2632 
2633 		if (sched->idle_hist) {
2634 			struct idle_thread_runtime *itr = (void *)tr;
2635 			struct thread_runtime *last_tr;
2636 
2637 			BUG_ON(thread->tid != 0);
2638 
2639 			if (itr->last_thread == NULL)
2640 				goto out;
2641 
2642 			/* add current idle time as last thread's runtime */
2643 			last_tr = thread__get_runtime(itr->last_thread);
2644 			if (last_tr == NULL)
2645 				goto out;
2646 
2647 			timehist_update_runtime_stats(last_tr, t, tprev);
2648 			/*
2649 			 * remove delta time of last thread as it's not updated
2650 			 * and otherwise it will show an invalid value next
2651 			 * time.  we only care total run time and run stat.
2652 			 */
2653 			last_tr->dt_run = 0;
2654 			last_tr->dt_delay = 0;
2655 			last_tr->dt_sleep = 0;
2656 			last_tr->dt_iowait = 0;
2657 			last_tr->dt_preempt = 0;
2658 
2659 			if (itr->cursor.nr)
2660 				callchain_append(&itr->callchain, &itr->cursor, t - tprev);
2661 
2662 			itr->last_thread = NULL;
2663 		}
2664 	}
2665 
2666 	if (!sched->summary_only)
2667 		timehist_print_sample(sched, evsel, sample, &al, thread, t, state);
2668 
2669 out:
2670 	if (sched->hist_time.start == 0 && t >= ptime->start)
2671 		sched->hist_time.start = t;
2672 	if (ptime->end == 0 || t <= ptime->end)
2673 		sched->hist_time.end = t;
2674 
2675 	if (tr) {
2676 		/* time of this sched_switch event becomes last time task seen */
2677 		tr->last_time = sample->time;
2678 
2679 		/* last state is used to determine where to account wait time */
2680 		tr->last_state = state;
2681 
2682 		/* sched out event for task so reset ready to run time */
2683 		tr->ready_to_run = 0;
2684 	}
2685 
2686 	evsel__save_time(evsel, sample->time, sample->cpu);
2687 
2688 	return rc;
2689 }
2690 
2691 static int timehist_sched_switch_event(struct perf_tool *tool,
2692 			     union perf_event *event,
2693 			     struct evsel *evsel,
2694 			     struct perf_sample *sample,
2695 			     struct machine *machine __maybe_unused)
2696 {
2697 	return timehist_sched_change_event(tool, event, evsel, sample, machine);
2698 }
2699 
2700 static int process_lost(struct perf_tool *tool __maybe_unused,
2701 			union perf_event *event,
2702 			struct perf_sample *sample,
2703 			struct machine *machine __maybe_unused)
2704 {
2705 	char tstr[64];
2706 
2707 	timestamp__scnprintf_usec(sample->time, tstr, sizeof(tstr));
2708 	printf("%15s ", tstr);
2709 	printf("lost %" PRI_lu64 " events on cpu %d\n", event->lost.lost, sample->cpu);
2710 
2711 	return 0;
2712 }
2713 
2714 
2715 static void print_thread_runtime(struct thread *t,
2716 				 struct thread_runtime *r)
2717 {
2718 	double mean = avg_stats(&r->run_stats);
2719 	float stddev;
2720 
2721 	printf("%*s   %5d  %9" PRIu64 " ",
2722 	       comm_width, timehist_get_commstr(t), t->ppid,
2723 	       (u64) r->run_stats.n);
2724 
2725 	print_sched_time(r->total_run_time, 8);
2726 	stddev = rel_stddev_stats(stddev_stats(&r->run_stats), mean);
2727 	print_sched_time(r->run_stats.min, 6);
2728 	printf(" ");
2729 	print_sched_time((u64) mean, 6);
2730 	printf(" ");
2731 	print_sched_time(r->run_stats.max, 6);
2732 	printf("  ");
2733 	printf("%5.2f", stddev);
2734 	printf("   %5" PRIu64, r->migrations);
2735 	printf("\n");
2736 }
2737 
2738 static void print_thread_waittime(struct thread *t,
2739 				  struct thread_runtime *r)
2740 {
2741 	printf("%*s   %5d  %9" PRIu64 " ",
2742 	       comm_width, timehist_get_commstr(t), t->ppid,
2743 	       (u64) r->run_stats.n);
2744 
2745 	print_sched_time(r->total_run_time, 8);
2746 	print_sched_time(r->total_sleep_time, 6);
2747 	printf(" ");
2748 	print_sched_time(r->total_iowait_time, 6);
2749 	printf(" ");
2750 	print_sched_time(r->total_preempt_time, 6);
2751 	printf(" ");
2752 	print_sched_time(r->total_delay_time, 6);
2753 	printf("\n");
2754 }
2755 
2756 struct total_run_stats {
2757 	struct perf_sched *sched;
2758 	u64  sched_count;
2759 	u64  task_count;
2760 	u64  total_run_time;
2761 };
2762 
2763 static int __show_thread_runtime(struct thread *t, void *priv)
2764 {
2765 	struct total_run_stats *stats = priv;
2766 	struct thread_runtime *r;
2767 
2768 	if (thread__is_filtered(t))
2769 		return 0;
2770 
2771 	r = thread__priv(t);
2772 	if (r && r->run_stats.n) {
2773 		stats->task_count++;
2774 		stats->sched_count += r->run_stats.n;
2775 		stats->total_run_time += r->total_run_time;
2776 
2777 		if (stats->sched->show_state)
2778 			print_thread_waittime(t, r);
2779 		else
2780 			print_thread_runtime(t, r);
2781 	}
2782 
2783 	return 0;
2784 }
2785 
2786 static int show_thread_runtime(struct thread *t, void *priv)
2787 {
2788 	if (t->dead)
2789 		return 0;
2790 
2791 	return __show_thread_runtime(t, priv);
2792 }
2793 
2794 static int show_deadthread_runtime(struct thread *t, void *priv)
2795 {
2796 	if (!t->dead)
2797 		return 0;
2798 
2799 	return __show_thread_runtime(t, priv);
2800 }
2801 
2802 static size_t callchain__fprintf_folded(FILE *fp, struct callchain_node *node)
2803 {
2804 	const char *sep = " <- ";
2805 	struct callchain_list *chain;
2806 	size_t ret = 0;
2807 	char bf[1024];
2808 	bool first;
2809 
2810 	if (node == NULL)
2811 		return 0;
2812 
2813 	ret = callchain__fprintf_folded(fp, node->parent);
2814 	first = (ret == 0);
2815 
2816 	list_for_each_entry(chain, &node->val, list) {
2817 		if (chain->ip >= PERF_CONTEXT_MAX)
2818 			continue;
2819 		if (chain->ms.sym && chain->ms.sym->ignore)
2820 			continue;
2821 		ret += fprintf(fp, "%s%s", first ? "" : sep,
2822 			       callchain_list__sym_name(chain, bf, sizeof(bf),
2823 							false));
2824 		first = false;
2825 	}
2826 
2827 	return ret;
2828 }
2829 
2830 static size_t timehist_print_idlehist_callchain(struct rb_root_cached *root)
2831 {
2832 	size_t ret = 0;
2833 	FILE *fp = stdout;
2834 	struct callchain_node *chain;
2835 	struct rb_node *rb_node = rb_first_cached(root);
2836 
2837 	printf("  %16s  %8s  %s\n", "Idle time (msec)", "Count", "Callchains");
2838 	printf("  %.16s  %.8s  %.50s\n", graph_dotted_line, graph_dotted_line,
2839 	       graph_dotted_line);
2840 
2841 	while (rb_node) {
2842 		chain = rb_entry(rb_node, struct callchain_node, rb_node);
2843 		rb_node = rb_next(rb_node);
2844 
2845 		ret += fprintf(fp, "  ");
2846 		print_sched_time(chain->hit, 12);
2847 		ret += 16;  /* print_sched_time returns 2nd arg + 4 */
2848 		ret += fprintf(fp, " %8d  ", chain->count);
2849 		ret += callchain__fprintf_folded(fp, chain);
2850 		ret += fprintf(fp, "\n");
2851 	}
2852 
2853 	return ret;
2854 }
2855 
2856 static void timehist_print_summary(struct perf_sched *sched,
2857 				   struct perf_session *session)
2858 {
2859 	struct machine *m = &session->machines.host;
2860 	struct total_run_stats totals;
2861 	u64 task_count;
2862 	struct thread *t;
2863 	struct thread_runtime *r;
2864 	int i;
2865 	u64 hist_time = sched->hist_time.end - sched->hist_time.start;
2866 
2867 	memset(&totals, 0, sizeof(totals));
2868 	totals.sched = sched;
2869 
2870 	if (sched->idle_hist) {
2871 		printf("\nIdle-time summary\n");
2872 		printf("%*s  parent  sched-out  ", comm_width, "comm");
2873 		printf("  idle-time   min-idle    avg-idle    max-idle  stddev  migrations\n");
2874 	} else if (sched->show_state) {
2875 		printf("\nWait-time summary\n");
2876 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2877 		printf("   run-time      sleep      iowait     preempt       delay\n");
2878 	} else {
2879 		printf("\nRuntime summary\n");
2880 		printf("%*s  parent   sched-in  ", comm_width, "comm");
2881 		printf("   run-time    min-run     avg-run     max-run  stddev  migrations\n");
2882 	}
2883 	printf("%*s            (count)  ", comm_width, "");
2884 	printf("     (msec)     (msec)      (msec)      (msec)       %s\n",
2885 	       sched->show_state ? "(msec)" : "%");
2886 	printf("%.117s\n", graph_dotted_line);
2887 
2888 	machine__for_each_thread(m, show_thread_runtime, &totals);
2889 	task_count = totals.task_count;
2890 	if (!task_count)
2891 		printf("<no still running tasks>\n");
2892 
2893 	printf("\nTerminated tasks:\n");
2894 	machine__for_each_thread(m, show_deadthread_runtime, &totals);
2895 	if (task_count == totals.task_count)
2896 		printf("<no terminated tasks>\n");
2897 
2898 	/* CPU idle stats not tracked when samples were skipped */
2899 	if (sched->skipped_samples && !sched->idle_hist)
2900 		return;
2901 
2902 	printf("\nIdle stats:\n");
2903 	for (i = 0; i < idle_max_cpu; ++i) {
2904 		if (cpu_list && !test_bit(i, cpu_bitmap))
2905 			continue;
2906 
2907 		t = idle_threads[i];
2908 		if (!t)
2909 			continue;
2910 
2911 		r = thread__priv(t);
2912 		if (r && r->run_stats.n) {
2913 			totals.sched_count += r->run_stats.n;
2914 			printf("    CPU %2d idle for ", i);
2915 			print_sched_time(r->total_run_time, 6);
2916 			printf(" msec  (%6.2f%%)\n", 100.0 * r->total_run_time / hist_time);
2917 		} else
2918 			printf("    CPU %2d idle entire time window\n", i);
2919 	}
2920 
2921 	if (sched->idle_hist && sched->show_callchain) {
2922 		callchain_param.mode  = CHAIN_FOLDED;
2923 		callchain_param.value = CCVAL_PERIOD;
2924 
2925 		callchain_register_param(&callchain_param);
2926 
2927 		printf("\nIdle stats by callchain:\n");
2928 		for (i = 0; i < idle_max_cpu; ++i) {
2929 			struct idle_thread_runtime *itr;
2930 
2931 			t = idle_threads[i];
2932 			if (!t)
2933 				continue;
2934 
2935 			itr = thread__priv(t);
2936 			if (itr == NULL)
2937 				continue;
2938 
2939 			callchain_param.sort(&itr->sorted_root.rb_root, &itr->callchain,
2940 					     0, &callchain_param);
2941 
2942 			printf("  CPU %2d:", i);
2943 			print_sched_time(itr->tr.total_run_time, 6);
2944 			printf(" msec\n");
2945 			timehist_print_idlehist_callchain(&itr->sorted_root);
2946 			printf("\n");
2947 		}
2948 	}
2949 
2950 	printf("\n"
2951 	       "    Total number of unique tasks: %" PRIu64 "\n"
2952 	       "Total number of context switches: %" PRIu64 "\n",
2953 	       totals.task_count, totals.sched_count);
2954 
2955 	printf("           Total run time (msec): ");
2956 	print_sched_time(totals.total_run_time, 2);
2957 	printf("\n");
2958 
2959 	printf("    Total scheduling time (msec): ");
2960 	print_sched_time(hist_time, 2);
2961 	printf(" (x %d)\n", sched->max_cpu.cpu);
2962 }
2963 
2964 typedef int (*sched_handler)(struct perf_tool *tool,
2965 			  union perf_event *event,
2966 			  struct evsel *evsel,
2967 			  struct perf_sample *sample,
2968 			  struct machine *machine);
2969 
2970 static int perf_timehist__process_sample(struct perf_tool *tool,
2971 					 union perf_event *event,
2972 					 struct perf_sample *sample,
2973 					 struct evsel *evsel,
2974 					 struct machine *machine)
2975 {
2976 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
2977 	int err = 0;
2978 	struct perf_cpu this_cpu = {
2979 		.cpu = sample->cpu,
2980 	};
2981 
2982 	if (this_cpu.cpu > sched->max_cpu.cpu)
2983 		sched->max_cpu = this_cpu;
2984 
2985 	if (evsel->handler != NULL) {
2986 		sched_handler f = evsel->handler;
2987 
2988 		err = f(tool, event, evsel, sample, machine);
2989 	}
2990 
2991 	return err;
2992 }
2993 
2994 static int timehist_check_attr(struct perf_sched *sched,
2995 			       struct evlist *evlist)
2996 {
2997 	struct evsel *evsel;
2998 	struct evsel_runtime *er;
2999 
3000 	list_for_each_entry(evsel, &evlist->core.entries, core.node) {
3001 		er = evsel__get_runtime(evsel);
3002 		if (er == NULL) {
3003 			pr_err("Failed to allocate memory for evsel runtime data\n");
3004 			return -1;
3005 		}
3006 
3007 		if (sched->show_callchain && !evsel__has_callchain(evsel)) {
3008 			pr_info("Samples do not have callchains.\n");
3009 			sched->show_callchain = 0;
3010 			symbol_conf.use_callchain = 0;
3011 		}
3012 	}
3013 
3014 	return 0;
3015 }
3016 
3017 static int perf_sched__timehist(struct perf_sched *sched)
3018 {
3019 	struct evsel_str_handler handlers[] = {
3020 		{ "sched:sched_switch",       timehist_sched_switch_event, },
3021 		{ "sched:sched_wakeup",	      timehist_sched_wakeup_event, },
3022 		{ "sched:sched_waking",       timehist_sched_wakeup_event, },
3023 		{ "sched:sched_wakeup_new",   timehist_sched_wakeup_event, },
3024 	};
3025 	const struct evsel_str_handler migrate_handlers[] = {
3026 		{ "sched:sched_migrate_task", timehist_migrate_task_event, },
3027 	};
3028 	struct perf_data data = {
3029 		.path  = input_name,
3030 		.mode  = PERF_DATA_MODE_READ,
3031 		.force = sched->force,
3032 	};
3033 
3034 	struct perf_session *session;
3035 	struct evlist *evlist;
3036 	int err = -1;
3037 
3038 	/*
3039 	 * event handlers for timehist option
3040 	 */
3041 	sched->tool.sample	 = perf_timehist__process_sample;
3042 	sched->tool.mmap	 = perf_event__process_mmap;
3043 	sched->tool.comm	 = perf_event__process_comm;
3044 	sched->tool.exit	 = perf_event__process_exit;
3045 	sched->tool.fork	 = perf_event__process_fork;
3046 	sched->tool.lost	 = process_lost;
3047 	sched->tool.attr	 = perf_event__process_attr;
3048 	sched->tool.tracing_data = perf_event__process_tracing_data;
3049 	sched->tool.build_id	 = perf_event__process_build_id;
3050 
3051 	sched->tool.ordered_events = true;
3052 	sched->tool.ordering_requires_timestamps = true;
3053 
3054 	symbol_conf.use_callchain = sched->show_callchain;
3055 
3056 	session = perf_session__new(&data, &sched->tool);
3057 	if (IS_ERR(session))
3058 		return PTR_ERR(session);
3059 
3060 	if (cpu_list) {
3061 		err = perf_session__cpu_bitmap(session, cpu_list, cpu_bitmap);
3062 		if (err < 0)
3063 			goto out;
3064 	}
3065 
3066 	evlist = session->evlist;
3067 
3068 	symbol__init(&session->header.env);
3069 
3070 	if (perf_time__parse_str(&sched->ptime, sched->time_str) != 0) {
3071 		pr_err("Invalid time string\n");
3072 		return -EINVAL;
3073 	}
3074 
3075 	if (timehist_check_attr(sched, evlist) != 0)
3076 		goto out;
3077 
3078 	setup_pager();
3079 
3080 	/* prefer sched_waking if it is captured */
3081 	if (evlist__find_tracepoint_by_name(session->evlist, "sched:sched_waking"))
3082 		handlers[1].handler = timehist_sched_wakeup_ignore;
3083 
3084 	/* setup per-evsel handlers */
3085 	if (perf_session__set_tracepoints_handlers(session, handlers))
3086 		goto out;
3087 
3088 	/* sched_switch event at a minimum needs to exist */
3089 	if (!evlist__find_tracepoint_by_name(session->evlist, "sched:sched_switch")) {
3090 		pr_err("No sched_switch events found. Have you run 'perf sched record'?\n");
3091 		goto out;
3092 	}
3093 
3094 	if (sched->show_migrations &&
3095 	    perf_session__set_tracepoints_handlers(session, migrate_handlers))
3096 		goto out;
3097 
3098 	/* pre-allocate struct for per-CPU idle stats */
3099 	sched->max_cpu.cpu = session->header.env.nr_cpus_online;
3100 	if (sched->max_cpu.cpu == 0)
3101 		sched->max_cpu.cpu = 4;
3102 	if (init_idle_threads(sched->max_cpu.cpu))
3103 		goto out;
3104 
3105 	/* summary_only implies summary option, but don't overwrite summary if set */
3106 	if (sched->summary_only)
3107 		sched->summary = sched->summary_only;
3108 
3109 	if (!sched->summary_only)
3110 		timehist_header(sched);
3111 
3112 	err = perf_session__process_events(session);
3113 	if (err) {
3114 		pr_err("Failed to process events, error %d", err);
3115 		goto out;
3116 	}
3117 
3118 	sched->nr_events      = evlist->stats.nr_events[0];
3119 	sched->nr_lost_events = evlist->stats.total_lost;
3120 	sched->nr_lost_chunks = evlist->stats.nr_events[PERF_RECORD_LOST];
3121 
3122 	if (sched->summary)
3123 		timehist_print_summary(sched, session);
3124 
3125 out:
3126 	free_idle_threads();
3127 	perf_session__delete(session);
3128 
3129 	return err;
3130 }
3131 
3132 
3133 static void print_bad_events(struct perf_sched *sched)
3134 {
3135 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
3136 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
3137 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
3138 			sched->nr_unordered_timestamps, sched->nr_timestamps);
3139 	}
3140 	if (sched->nr_lost_events && sched->nr_events) {
3141 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
3142 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
3143 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
3144 	}
3145 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
3146 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
3147 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
3148 			sched->nr_context_switch_bugs, sched->nr_timestamps);
3149 		if (sched->nr_lost_events)
3150 			printf(" (due to lost events?)");
3151 		printf("\n");
3152 	}
3153 }
3154 
3155 static void __merge_work_atoms(struct rb_root_cached *root, struct work_atoms *data)
3156 {
3157 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
3158 	struct work_atoms *this;
3159 	const char *comm = thread__comm_str(data->thread), *this_comm;
3160 	bool leftmost = true;
3161 
3162 	while (*new) {
3163 		int cmp;
3164 
3165 		this = container_of(*new, struct work_atoms, node);
3166 		parent = *new;
3167 
3168 		this_comm = thread__comm_str(this->thread);
3169 		cmp = strcmp(comm, this_comm);
3170 		if (cmp > 0) {
3171 			new = &((*new)->rb_left);
3172 		} else if (cmp < 0) {
3173 			new = &((*new)->rb_right);
3174 			leftmost = false;
3175 		} else {
3176 			this->num_merged++;
3177 			this->total_runtime += data->total_runtime;
3178 			this->nb_atoms += data->nb_atoms;
3179 			this->total_lat += data->total_lat;
3180 			list_splice(&data->work_list, &this->work_list);
3181 			if (this->max_lat < data->max_lat) {
3182 				this->max_lat = data->max_lat;
3183 				this->max_lat_start = data->max_lat_start;
3184 				this->max_lat_end = data->max_lat_end;
3185 			}
3186 			zfree(&data);
3187 			return;
3188 		}
3189 	}
3190 
3191 	data->num_merged++;
3192 	rb_link_node(&data->node, parent, new);
3193 	rb_insert_color_cached(&data->node, root, leftmost);
3194 }
3195 
3196 static void perf_sched__merge_lat(struct perf_sched *sched)
3197 {
3198 	struct work_atoms *data;
3199 	struct rb_node *node;
3200 
3201 	if (sched->skip_merge)
3202 		return;
3203 
3204 	while ((node = rb_first_cached(&sched->atom_root))) {
3205 		rb_erase_cached(node, &sched->atom_root);
3206 		data = rb_entry(node, struct work_atoms, node);
3207 		__merge_work_atoms(&sched->merged_atom_root, data);
3208 	}
3209 }
3210 
3211 static int perf_sched__lat(struct perf_sched *sched)
3212 {
3213 	struct rb_node *next;
3214 
3215 	setup_pager();
3216 
3217 	if (perf_sched__read_events(sched))
3218 		return -1;
3219 
3220 	perf_sched__merge_lat(sched);
3221 	perf_sched__sort_lat(sched);
3222 
3223 	printf("\n -------------------------------------------------------------------------------------------------------------------------------------------\n");
3224 	printf("  Task                  |   Runtime ms  | Switches | Avg delay ms    | Max delay ms    | Max delay start           | Max delay end          |\n");
3225 	printf(" -------------------------------------------------------------------------------------------------------------------------------------------\n");
3226 
3227 	next = rb_first_cached(&sched->sorted_atom_root);
3228 
3229 	while (next) {
3230 		struct work_atoms *work_list;
3231 
3232 		work_list = rb_entry(next, struct work_atoms, node);
3233 		output_lat_thread(sched, work_list);
3234 		next = rb_next(next);
3235 		thread__zput(work_list->thread);
3236 	}
3237 
3238 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
3239 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
3240 		(double)sched->all_runtime / NSEC_PER_MSEC, sched->all_count);
3241 
3242 	printf(" ---------------------------------------------------\n");
3243 
3244 	print_bad_events(sched);
3245 	printf("\n");
3246 
3247 	return 0;
3248 }
3249 
3250 static int setup_map_cpus(struct perf_sched *sched)
3251 {
3252 	struct perf_cpu_map *map;
3253 
3254 	sched->max_cpu.cpu  = sysconf(_SC_NPROCESSORS_CONF);
3255 
3256 	if (sched->map.comp) {
3257 		sched->map.comp_cpus = zalloc(sched->max_cpu.cpu * sizeof(int));
3258 		if (!sched->map.comp_cpus)
3259 			return -1;
3260 	}
3261 
3262 	if (!sched->map.cpus_str)
3263 		return 0;
3264 
3265 	map = perf_cpu_map__new(sched->map.cpus_str);
3266 	if (!map) {
3267 		pr_err("failed to get cpus map from %s\n", sched->map.cpus_str);
3268 		return -1;
3269 	}
3270 
3271 	sched->map.cpus = map;
3272 	return 0;
3273 }
3274 
3275 static int setup_color_pids(struct perf_sched *sched)
3276 {
3277 	struct perf_thread_map *map;
3278 
3279 	if (!sched->map.color_pids_str)
3280 		return 0;
3281 
3282 	map = thread_map__new_by_tid_str(sched->map.color_pids_str);
3283 	if (!map) {
3284 		pr_err("failed to get thread map from %s\n", sched->map.color_pids_str);
3285 		return -1;
3286 	}
3287 
3288 	sched->map.color_pids = map;
3289 	return 0;
3290 }
3291 
3292 static int setup_color_cpus(struct perf_sched *sched)
3293 {
3294 	struct perf_cpu_map *map;
3295 
3296 	if (!sched->map.color_cpus_str)
3297 		return 0;
3298 
3299 	map = perf_cpu_map__new(sched->map.color_cpus_str);
3300 	if (!map) {
3301 		pr_err("failed to get thread map from %s\n", sched->map.color_cpus_str);
3302 		return -1;
3303 	}
3304 
3305 	sched->map.color_cpus = map;
3306 	return 0;
3307 }
3308 
3309 static int perf_sched__map(struct perf_sched *sched)
3310 {
3311 	if (setup_map_cpus(sched))
3312 		return -1;
3313 
3314 	if (setup_color_pids(sched))
3315 		return -1;
3316 
3317 	if (setup_color_cpus(sched))
3318 		return -1;
3319 
3320 	setup_pager();
3321 	if (perf_sched__read_events(sched))
3322 		return -1;
3323 	print_bad_events(sched);
3324 	return 0;
3325 }
3326 
3327 static int perf_sched__replay(struct perf_sched *sched)
3328 {
3329 	unsigned long i;
3330 
3331 	calibrate_run_measurement_overhead(sched);
3332 	calibrate_sleep_measurement_overhead(sched);
3333 
3334 	test_calibrations(sched);
3335 
3336 	if (perf_sched__read_events(sched))
3337 		return -1;
3338 
3339 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
3340 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
3341 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
3342 
3343 	if (sched->targetless_wakeups)
3344 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
3345 	if (sched->multitarget_wakeups)
3346 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
3347 	if (sched->nr_run_events_optimized)
3348 		printf("run atoms optimized: %ld\n",
3349 			sched->nr_run_events_optimized);
3350 
3351 	print_task_traces(sched);
3352 	add_cross_task_wakeups(sched);
3353 
3354 	sched->thread_funcs_exit = false;
3355 	create_tasks(sched);
3356 	printf("------------------------------------------------------------\n");
3357 	for (i = 0; i < sched->replay_repeat; i++)
3358 		run_one_test(sched);
3359 
3360 	sched->thread_funcs_exit = true;
3361 	destroy_tasks(sched);
3362 	return 0;
3363 }
3364 
3365 static void setup_sorting(struct perf_sched *sched, const struct option *options,
3366 			  const char * const usage_msg[])
3367 {
3368 	char *tmp, *tok, *str = strdup(sched->sort_order);
3369 
3370 	for (tok = strtok_r(str, ", ", &tmp);
3371 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
3372 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
3373 			usage_with_options_msg(usage_msg, options,
3374 					"Unknown --sort key: `%s'", tok);
3375 		}
3376 	}
3377 
3378 	free(str);
3379 
3380 	sort_dimension__add("pid", &sched->cmp_pid);
3381 }
3382 
3383 static bool schedstat_events_exposed(void)
3384 {
3385 	/*
3386 	 * Select "sched:sched_stat_wait" event to check
3387 	 * whether schedstat tracepoints are exposed.
3388 	 */
3389 	return IS_ERR(trace_event__tp_format("sched", "sched_stat_wait")) ?
3390 		false : true;
3391 }
3392 
3393 static int __cmd_record(int argc, const char **argv)
3394 {
3395 	unsigned int rec_argc, i, j;
3396 	char **rec_argv;
3397 	const char **rec_argv_copy;
3398 	const char * const record_args[] = {
3399 		"record",
3400 		"-a",
3401 		"-R",
3402 		"-m", "1024",
3403 		"-c", "1",
3404 		"-e", "sched:sched_switch",
3405 		"-e", "sched:sched_stat_runtime",
3406 		"-e", "sched:sched_process_fork",
3407 		"-e", "sched:sched_wakeup_new",
3408 		"-e", "sched:sched_migrate_task",
3409 	};
3410 
3411 	/*
3412 	 * The tracepoints trace_sched_stat_{wait, sleep, iowait}
3413 	 * are not exposed to user if CONFIG_SCHEDSTATS is not set,
3414 	 * to prevent "perf sched record" execution failure, determine
3415 	 * whether to record schedstat events according to actual situation.
3416 	 */
3417 	const char * const schedstat_args[] = {
3418 		"-e", "sched:sched_stat_wait",
3419 		"-e", "sched:sched_stat_sleep",
3420 		"-e", "sched:sched_stat_iowait",
3421 	};
3422 	unsigned int schedstat_argc = schedstat_events_exposed() ?
3423 		ARRAY_SIZE(schedstat_args) : 0;
3424 
3425 	struct tep_event *waking_event;
3426 	int ret;
3427 
3428 	/*
3429 	 * +2 for either "-e", "sched:sched_wakeup" or
3430 	 * "-e", "sched:sched_waking"
3431 	 */
3432 	rec_argc = ARRAY_SIZE(record_args) + 2 + schedstat_argc + argc - 1;
3433 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
3434 	if (rec_argv == NULL)
3435 		return -ENOMEM;
3436 	rec_argv_copy = calloc(rec_argc + 1, sizeof(char *));
3437 	if (rec_argv_copy == NULL) {
3438 		free(rec_argv);
3439 		return -ENOMEM;
3440 	}
3441 
3442 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
3443 		rec_argv[i] = strdup(record_args[i]);
3444 
3445 	rec_argv[i++] = strdup("-e");
3446 	waking_event = trace_event__tp_format("sched", "sched_waking");
3447 	if (!IS_ERR(waking_event))
3448 		rec_argv[i++] = strdup("sched:sched_waking");
3449 	else
3450 		rec_argv[i++] = strdup("sched:sched_wakeup");
3451 
3452 	for (j = 0; j < schedstat_argc; j++)
3453 		rec_argv[i++] = strdup(schedstat_args[j]);
3454 
3455 	for (j = 1; j < (unsigned int)argc; j++, i++)
3456 		rec_argv[i] = strdup(argv[j]);
3457 
3458 	BUG_ON(i != rec_argc);
3459 
3460 	memcpy(rec_argv_copy, rec_argv, sizeof(char *) * rec_argc);
3461 	ret = cmd_record(rec_argc, rec_argv_copy);
3462 
3463 	for (i = 0; i < rec_argc; i++)
3464 		free(rec_argv[i]);
3465 	free(rec_argv);
3466 	free(rec_argv_copy);
3467 
3468 	return ret;
3469 }
3470 
3471 int cmd_sched(int argc, const char **argv)
3472 {
3473 	static const char default_sort_order[] = "avg, max, switch, runtime";
3474 	struct perf_sched sched = {
3475 		.tool = {
3476 			.sample		 = perf_sched__process_tracepoint_sample,
3477 			.comm		 = perf_sched__process_comm,
3478 			.namespaces	 = perf_event__process_namespaces,
3479 			.lost		 = perf_event__process_lost,
3480 			.fork		 = perf_sched__process_fork_event,
3481 			.ordered_events = true,
3482 		},
3483 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
3484 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
3485 		.sort_order	      = default_sort_order,
3486 		.replay_repeat	      = 10,
3487 		.profile_cpu	      = -1,
3488 		.next_shortname1      = 'A',
3489 		.next_shortname2      = '0',
3490 		.skip_merge           = 0,
3491 		.show_callchain	      = 1,
3492 		.max_stack            = 5,
3493 	};
3494 	const struct option sched_options[] = {
3495 	OPT_STRING('i', "input", &input_name, "file",
3496 		    "input file name"),
3497 	OPT_INCR('v', "verbose", &verbose,
3498 		    "be more verbose (show symbol address, etc)"),
3499 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
3500 		    "dump raw trace in ASCII"),
3501 	OPT_BOOLEAN('f', "force", &sched.force, "don't complain, do it"),
3502 	OPT_END()
3503 	};
3504 	const struct option latency_options[] = {
3505 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
3506 		   "sort by key(s): runtime, switch, avg, max"),
3507 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
3508 		    "CPU to profile on"),
3509 	OPT_BOOLEAN('p', "pids", &sched.skip_merge,
3510 		    "latency stats per pid instead of per comm"),
3511 	OPT_PARENT(sched_options)
3512 	};
3513 	const struct option replay_options[] = {
3514 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
3515 		     "repeat the workload replay N times (-1: infinite)"),
3516 	OPT_PARENT(sched_options)
3517 	};
3518 	const struct option map_options[] = {
3519 	OPT_BOOLEAN(0, "compact", &sched.map.comp,
3520 		    "map output in compact mode"),
3521 	OPT_STRING(0, "color-pids", &sched.map.color_pids_str, "pids",
3522 		   "highlight given pids in map"),
3523 	OPT_STRING(0, "color-cpus", &sched.map.color_cpus_str, "cpus",
3524                     "highlight given CPUs in map"),
3525 	OPT_STRING(0, "cpus", &sched.map.cpus_str, "cpus",
3526                     "display given CPUs in map"),
3527 	OPT_PARENT(sched_options)
3528 	};
3529 	const struct option timehist_options[] = {
3530 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
3531 		   "file", "vmlinux pathname"),
3532 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
3533 		   "file", "kallsyms pathname"),
3534 	OPT_BOOLEAN('g', "call-graph", &sched.show_callchain,
3535 		    "Display call chains if present (default on)"),
3536 	OPT_UINTEGER(0, "max-stack", &sched.max_stack,
3537 		   "Maximum number of functions to display backtrace."),
3538 	OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
3539 		    "Look for files with symbols relative to this directory"),
3540 	OPT_BOOLEAN('s', "summary", &sched.summary_only,
3541 		    "Show only syscall summary with statistics"),
3542 	OPT_BOOLEAN('S', "with-summary", &sched.summary,
3543 		    "Show all syscalls and summary with statistics"),
3544 	OPT_BOOLEAN('w', "wakeups", &sched.show_wakeups, "Show wakeup events"),
3545 	OPT_BOOLEAN('n', "next", &sched.show_next, "Show next task"),
3546 	OPT_BOOLEAN('M', "migrations", &sched.show_migrations, "Show migration events"),
3547 	OPT_BOOLEAN('V', "cpu-visual", &sched.show_cpu_visual, "Add CPU visual"),
3548 	OPT_BOOLEAN('I', "idle-hist", &sched.idle_hist, "Show idle events only"),
3549 	OPT_STRING(0, "time", &sched.time_str, "str",
3550 		   "Time span for analysis (start,stop)"),
3551 	OPT_BOOLEAN(0, "state", &sched.show_state, "Show task state when sched-out"),
3552 	OPT_STRING('p', "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
3553 		   "analyze events only for given process id(s)"),
3554 	OPT_STRING('t', "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
3555 		   "analyze events only for given thread id(s)"),
3556 	OPT_STRING('C', "cpu", &cpu_list, "cpu", "list of cpus to profile"),
3557 	OPT_PARENT(sched_options)
3558 	};
3559 
3560 	const char * const latency_usage[] = {
3561 		"perf sched latency [<options>]",
3562 		NULL
3563 	};
3564 	const char * const replay_usage[] = {
3565 		"perf sched replay [<options>]",
3566 		NULL
3567 	};
3568 	const char * const map_usage[] = {
3569 		"perf sched map [<options>]",
3570 		NULL
3571 	};
3572 	const char * const timehist_usage[] = {
3573 		"perf sched timehist [<options>]",
3574 		NULL
3575 	};
3576 	const char *const sched_subcommands[] = { "record", "latency", "map",
3577 						  "replay", "script",
3578 						  "timehist", NULL };
3579 	const char *sched_usage[] = {
3580 		NULL,
3581 		NULL
3582 	};
3583 	struct trace_sched_handler lat_ops  = {
3584 		.wakeup_event	    = latency_wakeup_event,
3585 		.switch_event	    = latency_switch_event,
3586 		.runtime_event	    = latency_runtime_event,
3587 		.migrate_task_event = latency_migrate_task_event,
3588 	};
3589 	struct trace_sched_handler map_ops  = {
3590 		.switch_event	    = map_switch_event,
3591 	};
3592 	struct trace_sched_handler replay_ops  = {
3593 		.wakeup_event	    = replay_wakeup_event,
3594 		.switch_event	    = replay_switch_event,
3595 		.fork_event	    = replay_fork_event,
3596 	};
3597 	unsigned int i;
3598 	int ret = 0;
3599 
3600 	mutex_init(&sched.start_work_mutex);
3601 	mutex_init(&sched.work_done_wait_mutex);
3602 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
3603 		sched.curr_pid[i] = -1;
3604 
3605 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
3606 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3607 	if (!argc)
3608 		usage_with_options(sched_usage, sched_options);
3609 
3610 	/*
3611 	 * Aliased to 'perf script' for now:
3612 	 */
3613 	if (!strcmp(argv[0], "script")) {
3614 		ret = cmd_script(argc, argv);
3615 	} else if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
3616 		ret = __cmd_record(argc, argv);
3617 	} else if (strlen(argv[0]) > 2 && strstarts("latency", argv[0])) {
3618 		sched.tp_handler = &lat_ops;
3619 		if (argc > 1) {
3620 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
3621 			if (argc)
3622 				usage_with_options(latency_usage, latency_options);
3623 		}
3624 		setup_sorting(&sched, latency_options, latency_usage);
3625 		ret = perf_sched__lat(&sched);
3626 	} else if (!strcmp(argv[0], "map")) {
3627 		if (argc) {
3628 			argc = parse_options(argc, argv, map_options, map_usage, 0);
3629 			if (argc)
3630 				usage_with_options(map_usage, map_options);
3631 		}
3632 		sched.tp_handler = &map_ops;
3633 		setup_sorting(&sched, latency_options, latency_usage);
3634 		ret = perf_sched__map(&sched);
3635 	} else if (strlen(argv[0]) > 2 && strstarts("replay", argv[0])) {
3636 		sched.tp_handler = &replay_ops;
3637 		if (argc) {
3638 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
3639 			if (argc)
3640 				usage_with_options(replay_usage, replay_options);
3641 		}
3642 		ret = perf_sched__replay(&sched);
3643 	} else if (!strcmp(argv[0], "timehist")) {
3644 		if (argc) {
3645 			argc = parse_options(argc, argv, timehist_options,
3646 					     timehist_usage, 0);
3647 			if (argc)
3648 				usage_with_options(timehist_usage, timehist_options);
3649 		}
3650 		if ((sched.show_wakeups || sched.show_next) &&
3651 		    sched.summary_only) {
3652 			pr_err(" Error: -s and -[n|w] are mutually exclusive.\n");
3653 			parse_options_usage(timehist_usage, timehist_options, "s", true);
3654 			if (sched.show_wakeups)
3655 				parse_options_usage(NULL, timehist_options, "w", true);
3656 			if (sched.show_next)
3657 				parse_options_usage(NULL, timehist_options, "n", true);
3658 			ret = -EINVAL;
3659 			goto out;
3660 		}
3661 		ret = symbol__validate_sym_arguments();
3662 		if (ret)
3663 			goto out;
3664 
3665 		ret = perf_sched__timehist(&sched);
3666 	} else {
3667 		usage_with_options(sched_usage, sched_options);
3668 	}
3669 
3670 out:
3671 	mutex_destroy(&sched.start_work_mutex);
3672 	mutex_destroy(&sched.work_done_wait_mutex);
3673 
3674 	return ret;
3675 }
3676