xref: /linux/tools/perf/builtin-sched.c (revision 273b281fa22c293963ee3e6eec418f5dda2dbc83)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/util.h"
5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
9 
10 #include "util/parse-options.h"
11 #include "util/trace-event.h"
12 
13 #include "util/debug.h"
14 #include "util/data_map.h"
15 
16 #include <sys/prctl.h>
17 
18 #include <semaphore.h>
19 #include <pthread.h>
20 #include <math.h>
21 
22 static char			const *input_name = "perf.data";
23 
24 static struct perf_header	*header;
25 static u64			sample_type;
26 
27 static char			default_sort_order[] = "avg, max, switch, runtime";
28 static char			*sort_order = default_sort_order;
29 
30 static int			profile_cpu = -1;
31 
32 #define PR_SET_NAME		15               /* Set process name */
33 #define MAX_CPUS		4096
34 
35 static u64			run_measurement_overhead;
36 static u64			sleep_measurement_overhead;
37 
38 #define COMM_LEN		20
39 #define SYM_LEN			129
40 
41 #define MAX_PID			65536
42 
43 static unsigned long		nr_tasks;
44 
45 struct sched_atom;
46 
47 struct task_desc {
48 	unsigned long		nr;
49 	unsigned long		pid;
50 	char			comm[COMM_LEN];
51 
52 	unsigned long		nr_events;
53 	unsigned long		curr_event;
54 	struct sched_atom	**atoms;
55 
56 	pthread_t		thread;
57 	sem_t			sleep_sem;
58 
59 	sem_t			ready_for_work;
60 	sem_t			work_done_sem;
61 
62 	u64			cpu_usage;
63 };
64 
65 enum sched_event_type {
66 	SCHED_EVENT_RUN,
67 	SCHED_EVENT_SLEEP,
68 	SCHED_EVENT_WAKEUP,
69 	SCHED_EVENT_MIGRATION,
70 };
71 
72 struct sched_atom {
73 	enum sched_event_type	type;
74 	u64			timestamp;
75 	u64			duration;
76 	unsigned long		nr;
77 	int			specific_wait;
78 	sem_t			*wait_sem;
79 	struct task_desc	*wakee;
80 };
81 
82 static struct task_desc		*pid_to_task[MAX_PID];
83 
84 static struct task_desc		**tasks;
85 
86 static pthread_mutex_t		start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
87 static u64			start_time;
88 
89 static pthread_mutex_t		work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
90 
91 static unsigned long		nr_run_events;
92 static unsigned long		nr_sleep_events;
93 static unsigned long		nr_wakeup_events;
94 
95 static unsigned long		nr_sleep_corrections;
96 static unsigned long		nr_run_events_optimized;
97 
98 static unsigned long		targetless_wakeups;
99 static unsigned long		multitarget_wakeups;
100 
101 static u64			cpu_usage;
102 static u64			runavg_cpu_usage;
103 static u64			parent_cpu_usage;
104 static u64			runavg_parent_cpu_usage;
105 
106 static unsigned long		nr_runs;
107 static u64			sum_runtime;
108 static u64			sum_fluct;
109 static u64			run_avg;
110 
111 static unsigned long		replay_repeat = 10;
112 static unsigned long		nr_timestamps;
113 static unsigned long		nr_unordered_timestamps;
114 static unsigned long		nr_state_machine_bugs;
115 static unsigned long		nr_context_switch_bugs;
116 static unsigned long		nr_events;
117 static unsigned long		nr_lost_chunks;
118 static unsigned long		nr_lost_events;
119 
120 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
121 
122 enum thread_state {
123 	THREAD_SLEEPING = 0,
124 	THREAD_WAIT_CPU,
125 	THREAD_SCHED_IN,
126 	THREAD_IGNORE
127 };
128 
129 struct work_atom {
130 	struct list_head	list;
131 	enum thread_state	state;
132 	u64			sched_out_time;
133 	u64			wake_up_time;
134 	u64			sched_in_time;
135 	u64			runtime;
136 };
137 
138 struct work_atoms {
139 	struct list_head	work_list;
140 	struct thread		*thread;
141 	struct rb_node		node;
142 	u64			max_lat;
143 	u64			max_lat_at;
144 	u64			total_lat;
145 	u64			nb_atoms;
146 	u64			total_runtime;
147 };
148 
149 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
150 
151 static struct rb_root		atom_root, sorted_atom_root;
152 
153 static u64			all_runtime;
154 static u64			all_count;
155 
156 
157 static u64 get_nsecs(void)
158 {
159 	struct timespec ts;
160 
161 	clock_gettime(CLOCK_MONOTONIC, &ts);
162 
163 	return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
164 }
165 
166 static void burn_nsecs(u64 nsecs)
167 {
168 	u64 T0 = get_nsecs(), T1;
169 
170 	do {
171 		T1 = get_nsecs();
172 	} while (T1 + run_measurement_overhead < T0 + nsecs);
173 }
174 
175 static void sleep_nsecs(u64 nsecs)
176 {
177 	struct timespec ts;
178 
179 	ts.tv_nsec = nsecs % 999999999;
180 	ts.tv_sec = nsecs / 999999999;
181 
182 	nanosleep(&ts, NULL);
183 }
184 
185 static void calibrate_run_measurement_overhead(void)
186 {
187 	u64 T0, T1, delta, min_delta = 1000000000ULL;
188 	int i;
189 
190 	for (i = 0; i < 10; i++) {
191 		T0 = get_nsecs();
192 		burn_nsecs(0);
193 		T1 = get_nsecs();
194 		delta = T1-T0;
195 		min_delta = min(min_delta, delta);
196 	}
197 	run_measurement_overhead = min_delta;
198 
199 	printf("run measurement overhead: %Ld nsecs\n", min_delta);
200 }
201 
202 static void calibrate_sleep_measurement_overhead(void)
203 {
204 	u64 T0, T1, delta, min_delta = 1000000000ULL;
205 	int i;
206 
207 	for (i = 0; i < 10; i++) {
208 		T0 = get_nsecs();
209 		sleep_nsecs(10000);
210 		T1 = get_nsecs();
211 		delta = T1-T0;
212 		min_delta = min(min_delta, delta);
213 	}
214 	min_delta -= 10000;
215 	sleep_measurement_overhead = min_delta;
216 
217 	printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
218 }
219 
220 static struct sched_atom *
221 get_new_event(struct task_desc *task, u64 timestamp)
222 {
223 	struct sched_atom *event = zalloc(sizeof(*event));
224 	unsigned long idx = task->nr_events;
225 	size_t size;
226 
227 	event->timestamp = timestamp;
228 	event->nr = idx;
229 
230 	task->nr_events++;
231 	size = sizeof(struct sched_atom *) * task->nr_events;
232 	task->atoms = realloc(task->atoms, size);
233 	BUG_ON(!task->atoms);
234 
235 	task->atoms[idx] = event;
236 
237 	return event;
238 }
239 
240 static struct sched_atom *last_event(struct task_desc *task)
241 {
242 	if (!task->nr_events)
243 		return NULL;
244 
245 	return task->atoms[task->nr_events - 1];
246 }
247 
248 static void
249 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
250 {
251 	struct sched_atom *event, *curr_event = last_event(task);
252 
253 	/*
254 	 * optimize an existing RUN event by merging this one
255 	 * to it:
256 	 */
257 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
258 		nr_run_events_optimized++;
259 		curr_event->duration += duration;
260 		return;
261 	}
262 
263 	event = get_new_event(task, timestamp);
264 
265 	event->type = SCHED_EVENT_RUN;
266 	event->duration = duration;
267 
268 	nr_run_events++;
269 }
270 
271 static void
272 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
273 		       struct task_desc *wakee)
274 {
275 	struct sched_atom *event, *wakee_event;
276 
277 	event = get_new_event(task, timestamp);
278 	event->type = SCHED_EVENT_WAKEUP;
279 	event->wakee = wakee;
280 
281 	wakee_event = last_event(wakee);
282 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
283 		targetless_wakeups++;
284 		return;
285 	}
286 	if (wakee_event->wait_sem) {
287 		multitarget_wakeups++;
288 		return;
289 	}
290 
291 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
292 	sem_init(wakee_event->wait_sem, 0, 0);
293 	wakee_event->specific_wait = 1;
294 	event->wait_sem = wakee_event->wait_sem;
295 
296 	nr_wakeup_events++;
297 }
298 
299 static void
300 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
301 		      u64 task_state __used)
302 {
303 	struct sched_atom *event = get_new_event(task, timestamp);
304 
305 	event->type = SCHED_EVENT_SLEEP;
306 
307 	nr_sleep_events++;
308 }
309 
310 static struct task_desc *register_pid(unsigned long pid, const char *comm)
311 {
312 	struct task_desc *task;
313 
314 	BUG_ON(pid >= MAX_PID);
315 
316 	task = pid_to_task[pid];
317 
318 	if (task)
319 		return task;
320 
321 	task = zalloc(sizeof(*task));
322 	task->pid = pid;
323 	task->nr = nr_tasks;
324 	strcpy(task->comm, comm);
325 	/*
326 	 * every task starts in sleeping state - this gets ignored
327 	 * if there's no wakeup pointing to this sleep state:
328 	 */
329 	add_sched_event_sleep(task, 0, 0);
330 
331 	pid_to_task[pid] = task;
332 	nr_tasks++;
333 	tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
334 	BUG_ON(!tasks);
335 	tasks[task->nr] = task;
336 
337 	if (verbose)
338 		printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
339 
340 	return task;
341 }
342 
343 
344 static void print_task_traces(void)
345 {
346 	struct task_desc *task;
347 	unsigned long i;
348 
349 	for (i = 0; i < nr_tasks; i++) {
350 		task = tasks[i];
351 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
352 			task->nr, task->comm, task->pid, task->nr_events);
353 	}
354 }
355 
356 static void add_cross_task_wakeups(void)
357 {
358 	struct task_desc *task1, *task2;
359 	unsigned long i, j;
360 
361 	for (i = 0; i < nr_tasks; i++) {
362 		task1 = tasks[i];
363 		j = i + 1;
364 		if (j == nr_tasks)
365 			j = 0;
366 		task2 = tasks[j];
367 		add_sched_event_wakeup(task1, 0, task2);
368 	}
369 }
370 
371 static void
372 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
373 {
374 	int ret = 0;
375 	u64 now;
376 	long long delta;
377 
378 	now = get_nsecs();
379 	delta = start_time + atom->timestamp - now;
380 
381 	switch (atom->type) {
382 		case SCHED_EVENT_RUN:
383 			burn_nsecs(atom->duration);
384 			break;
385 		case SCHED_EVENT_SLEEP:
386 			if (atom->wait_sem)
387 				ret = sem_wait(atom->wait_sem);
388 			BUG_ON(ret);
389 			break;
390 		case SCHED_EVENT_WAKEUP:
391 			if (atom->wait_sem)
392 				ret = sem_post(atom->wait_sem);
393 			BUG_ON(ret);
394 			break;
395 		case SCHED_EVENT_MIGRATION:
396 			break;
397 		default:
398 			BUG_ON(1);
399 	}
400 }
401 
402 static u64 get_cpu_usage_nsec_parent(void)
403 {
404 	struct rusage ru;
405 	u64 sum;
406 	int err;
407 
408 	err = getrusage(RUSAGE_SELF, &ru);
409 	BUG_ON(err);
410 
411 	sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
412 	sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
413 
414 	return sum;
415 }
416 
417 static int self_open_counters(void)
418 {
419 	struct perf_event_attr attr;
420 	int fd;
421 
422 	memset(&attr, 0, sizeof(attr));
423 
424 	attr.type = PERF_TYPE_SOFTWARE;
425 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
426 
427 	fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
428 
429 	if (fd < 0)
430 		die("Error: sys_perf_event_open() syscall returned"
431 		    "with %d (%s)\n", fd, strerror(errno));
432 	return fd;
433 }
434 
435 static u64 get_cpu_usage_nsec_self(int fd)
436 {
437 	u64 runtime;
438 	int ret;
439 
440 	ret = read(fd, &runtime, sizeof(runtime));
441 	BUG_ON(ret != sizeof(runtime));
442 
443 	return runtime;
444 }
445 
446 static void *thread_func(void *ctx)
447 {
448 	struct task_desc *this_task = ctx;
449 	u64 cpu_usage_0, cpu_usage_1;
450 	unsigned long i, ret;
451 	char comm2[22];
452 	int fd;
453 
454 	sprintf(comm2, ":%s", this_task->comm);
455 	prctl(PR_SET_NAME, comm2);
456 	fd = self_open_counters();
457 
458 again:
459 	ret = sem_post(&this_task->ready_for_work);
460 	BUG_ON(ret);
461 	ret = pthread_mutex_lock(&start_work_mutex);
462 	BUG_ON(ret);
463 	ret = pthread_mutex_unlock(&start_work_mutex);
464 	BUG_ON(ret);
465 
466 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
467 
468 	for (i = 0; i < this_task->nr_events; i++) {
469 		this_task->curr_event = i;
470 		process_sched_event(this_task, this_task->atoms[i]);
471 	}
472 
473 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
474 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
475 	ret = sem_post(&this_task->work_done_sem);
476 	BUG_ON(ret);
477 
478 	ret = pthread_mutex_lock(&work_done_wait_mutex);
479 	BUG_ON(ret);
480 	ret = pthread_mutex_unlock(&work_done_wait_mutex);
481 	BUG_ON(ret);
482 
483 	goto again;
484 }
485 
486 static void create_tasks(void)
487 {
488 	struct task_desc *task;
489 	pthread_attr_t attr;
490 	unsigned long i;
491 	int err;
492 
493 	err = pthread_attr_init(&attr);
494 	BUG_ON(err);
495 	err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
496 	BUG_ON(err);
497 	err = pthread_mutex_lock(&start_work_mutex);
498 	BUG_ON(err);
499 	err = pthread_mutex_lock(&work_done_wait_mutex);
500 	BUG_ON(err);
501 	for (i = 0; i < nr_tasks; i++) {
502 		task = tasks[i];
503 		sem_init(&task->sleep_sem, 0, 0);
504 		sem_init(&task->ready_for_work, 0, 0);
505 		sem_init(&task->work_done_sem, 0, 0);
506 		task->curr_event = 0;
507 		err = pthread_create(&task->thread, &attr, thread_func, task);
508 		BUG_ON(err);
509 	}
510 }
511 
512 static void wait_for_tasks(void)
513 {
514 	u64 cpu_usage_0, cpu_usage_1;
515 	struct task_desc *task;
516 	unsigned long i, ret;
517 
518 	start_time = get_nsecs();
519 	cpu_usage = 0;
520 	pthread_mutex_unlock(&work_done_wait_mutex);
521 
522 	for (i = 0; i < nr_tasks; i++) {
523 		task = tasks[i];
524 		ret = sem_wait(&task->ready_for_work);
525 		BUG_ON(ret);
526 		sem_init(&task->ready_for_work, 0, 0);
527 	}
528 	ret = pthread_mutex_lock(&work_done_wait_mutex);
529 	BUG_ON(ret);
530 
531 	cpu_usage_0 = get_cpu_usage_nsec_parent();
532 
533 	pthread_mutex_unlock(&start_work_mutex);
534 
535 	for (i = 0; i < nr_tasks; i++) {
536 		task = tasks[i];
537 		ret = sem_wait(&task->work_done_sem);
538 		BUG_ON(ret);
539 		sem_init(&task->work_done_sem, 0, 0);
540 		cpu_usage += task->cpu_usage;
541 		task->cpu_usage = 0;
542 	}
543 
544 	cpu_usage_1 = get_cpu_usage_nsec_parent();
545 	if (!runavg_cpu_usage)
546 		runavg_cpu_usage = cpu_usage;
547 	runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
548 
549 	parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
550 	if (!runavg_parent_cpu_usage)
551 		runavg_parent_cpu_usage = parent_cpu_usage;
552 	runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
553 				   parent_cpu_usage)/10;
554 
555 	ret = pthread_mutex_lock(&start_work_mutex);
556 	BUG_ON(ret);
557 
558 	for (i = 0; i < nr_tasks; i++) {
559 		task = tasks[i];
560 		sem_init(&task->sleep_sem, 0, 0);
561 		task->curr_event = 0;
562 	}
563 }
564 
565 static void run_one_test(void)
566 {
567 	u64 T0, T1, delta, avg_delta, fluct, std_dev;
568 
569 	T0 = get_nsecs();
570 	wait_for_tasks();
571 	T1 = get_nsecs();
572 
573 	delta = T1 - T0;
574 	sum_runtime += delta;
575 	nr_runs++;
576 
577 	avg_delta = sum_runtime / nr_runs;
578 	if (delta < avg_delta)
579 		fluct = avg_delta - delta;
580 	else
581 		fluct = delta - avg_delta;
582 	sum_fluct += fluct;
583 	std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
584 	if (!run_avg)
585 		run_avg = delta;
586 	run_avg = (run_avg*9 + delta)/10;
587 
588 	printf("#%-3ld: %0.3f, ",
589 		nr_runs, (double)delta/1000000.0);
590 
591 	printf("ravg: %0.2f, ",
592 		(double)run_avg/1e6);
593 
594 	printf("cpu: %0.2f / %0.2f",
595 		(double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
596 
597 #if 0
598 	/*
599 	 * rusage statistics done by the parent, these are less
600 	 * accurate than the sum_exec_runtime based statistics:
601 	 */
602 	printf(" [%0.2f / %0.2f]",
603 		(double)parent_cpu_usage/1e6,
604 		(double)runavg_parent_cpu_usage/1e6);
605 #endif
606 
607 	printf("\n");
608 
609 	if (nr_sleep_corrections)
610 		printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
611 	nr_sleep_corrections = 0;
612 }
613 
614 static void test_calibrations(void)
615 {
616 	u64 T0, T1;
617 
618 	T0 = get_nsecs();
619 	burn_nsecs(1e6);
620 	T1 = get_nsecs();
621 
622 	printf("the run test took %Ld nsecs\n", T1-T0);
623 
624 	T0 = get_nsecs();
625 	sleep_nsecs(1e6);
626 	T1 = get_nsecs();
627 
628 	printf("the sleep test took %Ld nsecs\n", T1-T0);
629 }
630 
631 #define FILL_FIELD(ptr, field, event, data)	\
632 	ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
633 
634 #define FILL_ARRAY(ptr, array, event, data)			\
635 do {								\
636 	void *__array = raw_field_ptr(event, #array, data);	\
637 	memcpy(ptr.array, __array, sizeof(ptr.array));	\
638 } while(0)
639 
640 #define FILL_COMMON_FIELDS(ptr, event, data)			\
641 do {								\
642 	FILL_FIELD(ptr, common_type, event, data);		\
643 	FILL_FIELD(ptr, common_flags, event, data);		\
644 	FILL_FIELD(ptr, common_preempt_count, event, data);	\
645 	FILL_FIELD(ptr, common_pid, event, data);		\
646 	FILL_FIELD(ptr, common_tgid, event, data);		\
647 } while (0)
648 
649 
650 
651 struct trace_switch_event {
652 	u32 size;
653 
654 	u16 common_type;
655 	u8 common_flags;
656 	u8 common_preempt_count;
657 	u32 common_pid;
658 	u32 common_tgid;
659 
660 	char prev_comm[16];
661 	u32 prev_pid;
662 	u32 prev_prio;
663 	u64 prev_state;
664 	char next_comm[16];
665 	u32 next_pid;
666 	u32 next_prio;
667 };
668 
669 struct trace_runtime_event {
670 	u32 size;
671 
672 	u16 common_type;
673 	u8 common_flags;
674 	u8 common_preempt_count;
675 	u32 common_pid;
676 	u32 common_tgid;
677 
678 	char comm[16];
679 	u32 pid;
680 	u64 runtime;
681 	u64 vruntime;
682 };
683 
684 struct trace_wakeup_event {
685 	u32 size;
686 
687 	u16 common_type;
688 	u8 common_flags;
689 	u8 common_preempt_count;
690 	u32 common_pid;
691 	u32 common_tgid;
692 
693 	char comm[16];
694 	u32 pid;
695 
696 	u32 prio;
697 	u32 success;
698 	u32 cpu;
699 };
700 
701 struct trace_fork_event {
702 	u32 size;
703 
704 	u16 common_type;
705 	u8 common_flags;
706 	u8 common_preempt_count;
707 	u32 common_pid;
708 	u32 common_tgid;
709 
710 	char parent_comm[16];
711 	u32 parent_pid;
712 	char child_comm[16];
713 	u32 child_pid;
714 };
715 
716 struct trace_migrate_task_event {
717 	u32 size;
718 
719 	u16 common_type;
720 	u8 common_flags;
721 	u8 common_preempt_count;
722 	u32 common_pid;
723 	u32 common_tgid;
724 
725 	char comm[16];
726 	u32 pid;
727 
728 	u32 prio;
729 	u32 cpu;
730 };
731 
732 struct trace_sched_handler {
733 	void (*switch_event)(struct trace_switch_event *,
734 			     struct event *,
735 			     int cpu,
736 			     u64 timestamp,
737 			     struct thread *thread);
738 
739 	void (*runtime_event)(struct trace_runtime_event *,
740 			      struct event *,
741 			      int cpu,
742 			      u64 timestamp,
743 			      struct thread *thread);
744 
745 	void (*wakeup_event)(struct trace_wakeup_event *,
746 			     struct event *,
747 			     int cpu,
748 			     u64 timestamp,
749 			     struct thread *thread);
750 
751 	void (*fork_event)(struct trace_fork_event *,
752 			   struct event *,
753 			   int cpu,
754 			   u64 timestamp,
755 			   struct thread *thread);
756 
757 	void (*migrate_task_event)(struct trace_migrate_task_event *,
758 			   struct event *,
759 			   int cpu,
760 			   u64 timestamp,
761 			   struct thread *thread);
762 };
763 
764 
765 static void
766 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
767 		    struct event *event,
768 		    int cpu __used,
769 		    u64 timestamp __used,
770 		    struct thread *thread __used)
771 {
772 	struct task_desc *waker, *wakee;
773 
774 	if (verbose) {
775 		printf("sched_wakeup event %p\n", event);
776 
777 		printf(" ... pid %d woke up %s/%d\n",
778 			wakeup_event->common_pid,
779 			wakeup_event->comm,
780 			wakeup_event->pid);
781 	}
782 
783 	waker = register_pid(wakeup_event->common_pid, "<unknown>");
784 	wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
785 
786 	add_sched_event_wakeup(waker, timestamp, wakee);
787 }
788 
789 static u64 cpu_last_switched[MAX_CPUS];
790 
791 static void
792 replay_switch_event(struct trace_switch_event *switch_event,
793 		    struct event *event,
794 		    int cpu,
795 		    u64 timestamp,
796 		    struct thread *thread __used)
797 {
798 	struct task_desc *prev, *next;
799 	u64 timestamp0;
800 	s64 delta;
801 
802 	if (verbose)
803 		printf("sched_switch event %p\n", event);
804 
805 	if (cpu >= MAX_CPUS || cpu < 0)
806 		return;
807 
808 	timestamp0 = cpu_last_switched[cpu];
809 	if (timestamp0)
810 		delta = timestamp - timestamp0;
811 	else
812 		delta = 0;
813 
814 	if (delta < 0)
815 		die("hm, delta: %Ld < 0 ?\n", delta);
816 
817 	if (verbose) {
818 		printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
819 			switch_event->prev_comm, switch_event->prev_pid,
820 			switch_event->next_comm, switch_event->next_pid,
821 			delta);
822 	}
823 
824 	prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
825 	next = register_pid(switch_event->next_pid, switch_event->next_comm);
826 
827 	cpu_last_switched[cpu] = timestamp;
828 
829 	add_sched_event_run(prev, timestamp, delta);
830 	add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
831 }
832 
833 
834 static void
835 replay_fork_event(struct trace_fork_event *fork_event,
836 		  struct event *event,
837 		  int cpu __used,
838 		  u64 timestamp __used,
839 		  struct thread *thread __used)
840 {
841 	if (verbose) {
842 		printf("sched_fork event %p\n", event);
843 		printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
844 		printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
845 	}
846 	register_pid(fork_event->parent_pid, fork_event->parent_comm);
847 	register_pid(fork_event->child_pid, fork_event->child_comm);
848 }
849 
850 static struct trace_sched_handler replay_ops  = {
851 	.wakeup_event		= replay_wakeup_event,
852 	.switch_event		= replay_switch_event,
853 	.fork_event		= replay_fork_event,
854 };
855 
856 struct sort_dimension {
857 	const char		*name;
858 	sort_fn_t		cmp;
859 	struct list_head	list;
860 };
861 
862 static LIST_HEAD(cmp_pid);
863 
864 static int
865 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
866 {
867 	struct sort_dimension *sort;
868 	int ret = 0;
869 
870 	BUG_ON(list_empty(list));
871 
872 	list_for_each_entry(sort, list, list) {
873 		ret = sort->cmp(l, r);
874 		if (ret)
875 			return ret;
876 	}
877 
878 	return ret;
879 }
880 
881 static struct work_atoms *
882 thread_atoms_search(struct rb_root *root, struct thread *thread,
883 			 struct list_head *sort_list)
884 {
885 	struct rb_node *node = root->rb_node;
886 	struct work_atoms key = { .thread = thread };
887 
888 	while (node) {
889 		struct work_atoms *atoms;
890 		int cmp;
891 
892 		atoms = container_of(node, struct work_atoms, node);
893 
894 		cmp = thread_lat_cmp(sort_list, &key, atoms);
895 		if (cmp > 0)
896 			node = node->rb_left;
897 		else if (cmp < 0)
898 			node = node->rb_right;
899 		else {
900 			BUG_ON(thread != atoms->thread);
901 			return atoms;
902 		}
903 	}
904 	return NULL;
905 }
906 
907 static void
908 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
909 			 struct list_head *sort_list)
910 {
911 	struct rb_node **new = &(root->rb_node), *parent = NULL;
912 
913 	while (*new) {
914 		struct work_atoms *this;
915 		int cmp;
916 
917 		this = container_of(*new, struct work_atoms, node);
918 		parent = *new;
919 
920 		cmp = thread_lat_cmp(sort_list, data, this);
921 
922 		if (cmp > 0)
923 			new = &((*new)->rb_left);
924 		else
925 			new = &((*new)->rb_right);
926 	}
927 
928 	rb_link_node(&data->node, parent, new);
929 	rb_insert_color(&data->node, root);
930 }
931 
932 static void thread_atoms_insert(struct thread *thread)
933 {
934 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
935 	if (!atoms)
936 		die("No memory");
937 
938 	atoms->thread = thread;
939 	INIT_LIST_HEAD(&atoms->work_list);
940 	__thread_latency_insert(&atom_root, atoms, &cmp_pid);
941 }
942 
943 static void
944 latency_fork_event(struct trace_fork_event *fork_event __used,
945 		   struct event *event __used,
946 		   int cpu __used,
947 		   u64 timestamp __used,
948 		   struct thread *thread __used)
949 {
950 	/* should insert the newcomer */
951 }
952 
953 __used
954 static char sched_out_state(struct trace_switch_event *switch_event)
955 {
956 	const char *str = TASK_STATE_TO_CHAR_STR;
957 
958 	return str[switch_event->prev_state];
959 }
960 
961 static void
962 add_sched_out_event(struct work_atoms *atoms,
963 		    char run_state,
964 		    u64 timestamp)
965 {
966 	struct work_atom *atom = zalloc(sizeof(*atom));
967 	if (!atom)
968 		die("Non memory");
969 
970 	atom->sched_out_time = timestamp;
971 
972 	if (run_state == 'R') {
973 		atom->state = THREAD_WAIT_CPU;
974 		atom->wake_up_time = atom->sched_out_time;
975 	}
976 
977 	list_add_tail(&atom->list, &atoms->work_list);
978 }
979 
980 static void
981 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
982 {
983 	struct work_atom *atom;
984 
985 	BUG_ON(list_empty(&atoms->work_list));
986 
987 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
988 
989 	atom->runtime += delta;
990 	atoms->total_runtime += delta;
991 }
992 
993 static void
994 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
995 {
996 	struct work_atom *atom;
997 	u64 delta;
998 
999 	if (list_empty(&atoms->work_list))
1000 		return;
1001 
1002 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1003 
1004 	if (atom->state != THREAD_WAIT_CPU)
1005 		return;
1006 
1007 	if (timestamp < atom->wake_up_time) {
1008 		atom->state = THREAD_IGNORE;
1009 		return;
1010 	}
1011 
1012 	atom->state = THREAD_SCHED_IN;
1013 	atom->sched_in_time = timestamp;
1014 
1015 	delta = atom->sched_in_time - atom->wake_up_time;
1016 	atoms->total_lat += delta;
1017 	if (delta > atoms->max_lat) {
1018 		atoms->max_lat = delta;
1019 		atoms->max_lat_at = timestamp;
1020 	}
1021 	atoms->nb_atoms++;
1022 }
1023 
1024 static void
1025 latency_switch_event(struct trace_switch_event *switch_event,
1026 		     struct event *event __used,
1027 		     int cpu,
1028 		     u64 timestamp,
1029 		     struct thread *thread __used)
1030 {
1031 	struct work_atoms *out_events, *in_events;
1032 	struct thread *sched_out, *sched_in;
1033 	u64 timestamp0;
1034 	s64 delta;
1035 
1036 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1037 
1038 	timestamp0 = cpu_last_switched[cpu];
1039 	cpu_last_switched[cpu] = timestamp;
1040 	if (timestamp0)
1041 		delta = timestamp - timestamp0;
1042 	else
1043 		delta = 0;
1044 
1045 	if (delta < 0)
1046 		die("hm, delta: %Ld < 0 ?\n", delta);
1047 
1048 
1049 	sched_out = threads__findnew(switch_event->prev_pid);
1050 	sched_in = threads__findnew(switch_event->next_pid);
1051 
1052 	out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1053 	if (!out_events) {
1054 		thread_atoms_insert(sched_out);
1055 		out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1056 		if (!out_events)
1057 			die("out-event: Internal tree error");
1058 	}
1059 	add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1060 
1061 	in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1062 	if (!in_events) {
1063 		thread_atoms_insert(sched_in);
1064 		in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1065 		if (!in_events)
1066 			die("in-event: Internal tree error");
1067 		/*
1068 		 * Take came in we have not heard about yet,
1069 		 * add in an initial atom in runnable state:
1070 		 */
1071 		add_sched_out_event(in_events, 'R', timestamp);
1072 	}
1073 	add_sched_in_event(in_events, timestamp);
1074 }
1075 
1076 static void
1077 latency_runtime_event(struct trace_runtime_event *runtime_event,
1078 		     struct event *event __used,
1079 		     int cpu,
1080 		     u64 timestamp,
1081 		     struct thread *this_thread __used)
1082 {
1083 	struct thread *thread = threads__findnew(runtime_event->pid);
1084 	struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1085 
1086 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1087 	if (!atoms) {
1088 		thread_atoms_insert(thread);
1089 		atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1090 		if (!atoms)
1091 			die("in-event: Internal tree error");
1092 		add_sched_out_event(atoms, 'R', timestamp);
1093 	}
1094 
1095 	add_runtime_event(atoms, runtime_event->runtime, timestamp);
1096 }
1097 
1098 static void
1099 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1100 		     struct event *__event __used,
1101 		     int cpu __used,
1102 		     u64 timestamp,
1103 		     struct thread *thread __used)
1104 {
1105 	struct work_atoms *atoms;
1106 	struct work_atom *atom;
1107 	struct thread *wakee;
1108 
1109 	/* Note for later, it may be interesting to observe the failing cases */
1110 	if (!wakeup_event->success)
1111 		return;
1112 
1113 	wakee = threads__findnew(wakeup_event->pid);
1114 	atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1115 	if (!atoms) {
1116 		thread_atoms_insert(wakee);
1117 		atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1118 		if (!atoms)
1119 			die("wakeup-event: Internal tree error");
1120 		add_sched_out_event(atoms, 'S', timestamp);
1121 	}
1122 
1123 	BUG_ON(list_empty(&atoms->work_list));
1124 
1125 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1126 
1127 	/*
1128 	 * You WILL be missing events if you've recorded only
1129 	 * one CPU, or are only looking at only one, so don't
1130 	 * make useless noise.
1131 	 */
1132 	if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1133 		nr_state_machine_bugs++;
1134 
1135 	nr_timestamps++;
1136 	if (atom->sched_out_time > timestamp) {
1137 		nr_unordered_timestamps++;
1138 		return;
1139 	}
1140 
1141 	atom->state = THREAD_WAIT_CPU;
1142 	atom->wake_up_time = timestamp;
1143 }
1144 
1145 static void
1146 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1147 		     struct event *__event __used,
1148 		     int cpu __used,
1149 		     u64 timestamp,
1150 		     struct thread *thread __used)
1151 {
1152 	struct work_atoms *atoms;
1153 	struct work_atom *atom;
1154 	struct thread *migrant;
1155 
1156 	/*
1157 	 * Only need to worry about migration when profiling one CPU.
1158 	 */
1159 	if (profile_cpu == -1)
1160 		return;
1161 
1162 	migrant = threads__findnew(migrate_task_event->pid);
1163 	atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1164 	if (!atoms) {
1165 		thread_atoms_insert(migrant);
1166 		register_pid(migrant->pid, migrant->comm);
1167 		atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1168 		if (!atoms)
1169 			die("migration-event: Internal tree error");
1170 		add_sched_out_event(atoms, 'R', timestamp);
1171 	}
1172 
1173 	BUG_ON(list_empty(&atoms->work_list));
1174 
1175 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1176 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1177 
1178 	nr_timestamps++;
1179 
1180 	if (atom->sched_out_time > timestamp)
1181 		nr_unordered_timestamps++;
1182 }
1183 
1184 static struct trace_sched_handler lat_ops  = {
1185 	.wakeup_event		= latency_wakeup_event,
1186 	.switch_event		= latency_switch_event,
1187 	.runtime_event		= latency_runtime_event,
1188 	.fork_event		= latency_fork_event,
1189 	.migrate_task_event	= latency_migrate_task_event,
1190 };
1191 
1192 static void output_lat_thread(struct work_atoms *work_list)
1193 {
1194 	int i;
1195 	int ret;
1196 	u64 avg;
1197 
1198 	if (!work_list->nb_atoms)
1199 		return;
1200 	/*
1201 	 * Ignore idle threads:
1202 	 */
1203 	if (!strcmp(work_list->thread->comm, "swapper"))
1204 		return;
1205 
1206 	all_runtime += work_list->total_runtime;
1207 	all_count += work_list->nb_atoms;
1208 
1209 	ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1210 
1211 	for (i = 0; i < 24 - ret; i++)
1212 		printf(" ");
1213 
1214 	avg = work_list->total_lat / work_list->nb_atoms;
1215 
1216 	printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1217 	      (double)work_list->total_runtime / 1e6,
1218 		 work_list->nb_atoms, (double)avg / 1e6,
1219 		 (double)work_list->max_lat / 1e6,
1220 		 (double)work_list->max_lat_at / 1e9);
1221 }
1222 
1223 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1224 {
1225 	if (l->thread->pid < r->thread->pid)
1226 		return -1;
1227 	if (l->thread->pid > r->thread->pid)
1228 		return 1;
1229 
1230 	return 0;
1231 }
1232 
1233 static struct sort_dimension pid_sort_dimension = {
1234 	.name			= "pid",
1235 	.cmp			= pid_cmp,
1236 };
1237 
1238 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1239 {
1240 	u64 avgl, avgr;
1241 
1242 	if (!l->nb_atoms)
1243 		return -1;
1244 
1245 	if (!r->nb_atoms)
1246 		return 1;
1247 
1248 	avgl = l->total_lat / l->nb_atoms;
1249 	avgr = r->total_lat / r->nb_atoms;
1250 
1251 	if (avgl < avgr)
1252 		return -1;
1253 	if (avgl > avgr)
1254 		return 1;
1255 
1256 	return 0;
1257 }
1258 
1259 static struct sort_dimension avg_sort_dimension = {
1260 	.name			= "avg",
1261 	.cmp			= avg_cmp,
1262 };
1263 
1264 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1265 {
1266 	if (l->max_lat < r->max_lat)
1267 		return -1;
1268 	if (l->max_lat > r->max_lat)
1269 		return 1;
1270 
1271 	return 0;
1272 }
1273 
1274 static struct sort_dimension max_sort_dimension = {
1275 	.name			= "max",
1276 	.cmp			= max_cmp,
1277 };
1278 
1279 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1280 {
1281 	if (l->nb_atoms < r->nb_atoms)
1282 		return -1;
1283 	if (l->nb_atoms > r->nb_atoms)
1284 		return 1;
1285 
1286 	return 0;
1287 }
1288 
1289 static struct sort_dimension switch_sort_dimension = {
1290 	.name			= "switch",
1291 	.cmp			= switch_cmp,
1292 };
1293 
1294 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1295 {
1296 	if (l->total_runtime < r->total_runtime)
1297 		return -1;
1298 	if (l->total_runtime > r->total_runtime)
1299 		return 1;
1300 
1301 	return 0;
1302 }
1303 
1304 static struct sort_dimension runtime_sort_dimension = {
1305 	.name			= "runtime",
1306 	.cmp			= runtime_cmp,
1307 };
1308 
1309 static struct sort_dimension *available_sorts[] = {
1310 	&pid_sort_dimension,
1311 	&avg_sort_dimension,
1312 	&max_sort_dimension,
1313 	&switch_sort_dimension,
1314 	&runtime_sort_dimension,
1315 };
1316 
1317 #define NB_AVAILABLE_SORTS	(int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1318 
1319 static LIST_HEAD(sort_list);
1320 
1321 static int sort_dimension__add(const char *tok, struct list_head *list)
1322 {
1323 	int i;
1324 
1325 	for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1326 		if (!strcmp(available_sorts[i]->name, tok)) {
1327 			list_add_tail(&available_sorts[i]->list, list);
1328 
1329 			return 0;
1330 		}
1331 	}
1332 
1333 	return -1;
1334 }
1335 
1336 static void setup_sorting(void);
1337 
1338 static void sort_lat(void)
1339 {
1340 	struct rb_node *node;
1341 
1342 	for (;;) {
1343 		struct work_atoms *data;
1344 		node = rb_first(&atom_root);
1345 		if (!node)
1346 			break;
1347 
1348 		rb_erase(node, &atom_root);
1349 		data = rb_entry(node, struct work_atoms, node);
1350 		__thread_latency_insert(&sorted_atom_root, data, &sort_list);
1351 	}
1352 }
1353 
1354 static struct trace_sched_handler *trace_handler;
1355 
1356 static void
1357 process_sched_wakeup_event(void *data,
1358 			   struct event *event,
1359 			   int cpu __used,
1360 			   u64 timestamp __used,
1361 			   struct thread *thread __used)
1362 {
1363 	struct trace_wakeup_event wakeup_event;
1364 
1365 	FILL_COMMON_FIELDS(wakeup_event, event, data);
1366 
1367 	FILL_ARRAY(wakeup_event, comm, event, data);
1368 	FILL_FIELD(wakeup_event, pid, event, data);
1369 	FILL_FIELD(wakeup_event, prio, event, data);
1370 	FILL_FIELD(wakeup_event, success, event, data);
1371 	FILL_FIELD(wakeup_event, cpu, event, data);
1372 
1373 	if (trace_handler->wakeup_event)
1374 		trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1375 }
1376 
1377 /*
1378  * Track the current task - that way we can know whether there's any
1379  * weird events, such as a task being switched away that is not current.
1380  */
1381 static int max_cpu;
1382 
1383 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1384 
1385 static struct thread *curr_thread[MAX_CPUS];
1386 
1387 static char next_shortname1 = 'A';
1388 static char next_shortname2 = '0';
1389 
1390 static void
1391 map_switch_event(struct trace_switch_event *switch_event,
1392 		 struct event *event __used,
1393 		 int this_cpu,
1394 		 u64 timestamp,
1395 		 struct thread *thread __used)
1396 {
1397 	struct thread *sched_out, *sched_in;
1398 	int new_shortname;
1399 	u64 timestamp0;
1400 	s64 delta;
1401 	int cpu;
1402 
1403 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1404 
1405 	if (this_cpu > max_cpu)
1406 		max_cpu = this_cpu;
1407 
1408 	timestamp0 = cpu_last_switched[this_cpu];
1409 	cpu_last_switched[this_cpu] = timestamp;
1410 	if (timestamp0)
1411 		delta = timestamp - timestamp0;
1412 	else
1413 		delta = 0;
1414 
1415 	if (delta < 0)
1416 		die("hm, delta: %Ld < 0 ?\n", delta);
1417 
1418 
1419 	sched_out = threads__findnew(switch_event->prev_pid);
1420 	sched_in = threads__findnew(switch_event->next_pid);
1421 
1422 	curr_thread[this_cpu] = sched_in;
1423 
1424 	printf("  ");
1425 
1426 	new_shortname = 0;
1427 	if (!sched_in->shortname[0]) {
1428 		sched_in->shortname[0] = next_shortname1;
1429 		sched_in->shortname[1] = next_shortname2;
1430 
1431 		if (next_shortname1 < 'Z') {
1432 			next_shortname1++;
1433 		} else {
1434 			next_shortname1='A';
1435 			if (next_shortname2 < '9') {
1436 				next_shortname2++;
1437 			} else {
1438 				next_shortname2='0';
1439 			}
1440 		}
1441 		new_shortname = 1;
1442 	}
1443 
1444 	for (cpu = 0; cpu <= max_cpu; cpu++) {
1445 		if (cpu != this_cpu)
1446 			printf(" ");
1447 		else
1448 			printf("*");
1449 
1450 		if (curr_thread[cpu]) {
1451 			if (curr_thread[cpu]->pid)
1452 				printf("%2s ", curr_thread[cpu]->shortname);
1453 			else
1454 				printf(".  ");
1455 		} else
1456 			printf("   ");
1457 	}
1458 
1459 	printf("  %12.6f secs ", (double)timestamp/1e9);
1460 	if (new_shortname) {
1461 		printf("%s => %s:%d\n",
1462 			sched_in->shortname, sched_in->comm, sched_in->pid);
1463 	} else {
1464 		printf("\n");
1465 	}
1466 }
1467 
1468 
1469 static void
1470 process_sched_switch_event(void *data,
1471 			   struct event *event,
1472 			   int this_cpu,
1473 			   u64 timestamp __used,
1474 			   struct thread *thread __used)
1475 {
1476 	struct trace_switch_event switch_event;
1477 
1478 	FILL_COMMON_FIELDS(switch_event, event, data);
1479 
1480 	FILL_ARRAY(switch_event, prev_comm, event, data);
1481 	FILL_FIELD(switch_event, prev_pid, event, data);
1482 	FILL_FIELD(switch_event, prev_prio, event, data);
1483 	FILL_FIELD(switch_event, prev_state, event, data);
1484 	FILL_ARRAY(switch_event, next_comm, event, data);
1485 	FILL_FIELD(switch_event, next_pid, event, data);
1486 	FILL_FIELD(switch_event, next_prio, event, data);
1487 
1488 	if (curr_pid[this_cpu] != (u32)-1) {
1489 		/*
1490 		 * Are we trying to switch away a PID that is
1491 		 * not current?
1492 		 */
1493 		if (curr_pid[this_cpu] != switch_event.prev_pid)
1494 			nr_context_switch_bugs++;
1495 	}
1496 	if (trace_handler->switch_event)
1497 		trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1498 
1499 	curr_pid[this_cpu] = switch_event.next_pid;
1500 }
1501 
1502 static void
1503 process_sched_runtime_event(void *data,
1504 			   struct event *event,
1505 			   int cpu __used,
1506 			   u64 timestamp __used,
1507 			   struct thread *thread __used)
1508 {
1509 	struct trace_runtime_event runtime_event;
1510 
1511 	FILL_ARRAY(runtime_event, comm, event, data);
1512 	FILL_FIELD(runtime_event, pid, event, data);
1513 	FILL_FIELD(runtime_event, runtime, event, data);
1514 	FILL_FIELD(runtime_event, vruntime, event, data);
1515 
1516 	if (trace_handler->runtime_event)
1517 		trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1518 }
1519 
1520 static void
1521 process_sched_fork_event(void *data,
1522 			 struct event *event,
1523 			 int cpu __used,
1524 			 u64 timestamp __used,
1525 			 struct thread *thread __used)
1526 {
1527 	struct trace_fork_event fork_event;
1528 
1529 	FILL_COMMON_FIELDS(fork_event, event, data);
1530 
1531 	FILL_ARRAY(fork_event, parent_comm, event, data);
1532 	FILL_FIELD(fork_event, parent_pid, event, data);
1533 	FILL_ARRAY(fork_event, child_comm, event, data);
1534 	FILL_FIELD(fork_event, child_pid, event, data);
1535 
1536 	if (trace_handler->fork_event)
1537 		trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1538 }
1539 
1540 static void
1541 process_sched_exit_event(struct event *event,
1542 			 int cpu __used,
1543 			 u64 timestamp __used,
1544 			 struct thread *thread __used)
1545 {
1546 	if (verbose)
1547 		printf("sched_exit event %p\n", event);
1548 }
1549 
1550 static void
1551 process_sched_migrate_task_event(void *data,
1552 			   struct event *event,
1553 			   int cpu __used,
1554 			   u64 timestamp __used,
1555 			   struct thread *thread __used)
1556 {
1557 	struct trace_migrate_task_event migrate_task_event;
1558 
1559 	FILL_COMMON_FIELDS(migrate_task_event, event, data);
1560 
1561 	FILL_ARRAY(migrate_task_event, comm, event, data);
1562 	FILL_FIELD(migrate_task_event, pid, event, data);
1563 	FILL_FIELD(migrate_task_event, prio, event, data);
1564 	FILL_FIELD(migrate_task_event, cpu, event, data);
1565 
1566 	if (trace_handler->migrate_task_event)
1567 		trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1568 }
1569 
1570 static void
1571 process_raw_event(event_t *raw_event __used, void *data,
1572 		  int cpu, u64 timestamp, struct thread *thread)
1573 {
1574 	struct event *event;
1575 	int type;
1576 
1577 
1578 	type = trace_parse_common_type(data);
1579 	event = trace_find_event(type);
1580 
1581 	if (!strcmp(event->name, "sched_switch"))
1582 		process_sched_switch_event(data, event, cpu, timestamp, thread);
1583 	if (!strcmp(event->name, "sched_stat_runtime"))
1584 		process_sched_runtime_event(data, event, cpu, timestamp, thread);
1585 	if (!strcmp(event->name, "sched_wakeup"))
1586 		process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1587 	if (!strcmp(event->name, "sched_wakeup_new"))
1588 		process_sched_wakeup_event(data, event, cpu, timestamp, thread);
1589 	if (!strcmp(event->name, "sched_process_fork"))
1590 		process_sched_fork_event(data, event, cpu, timestamp, thread);
1591 	if (!strcmp(event->name, "sched_process_exit"))
1592 		process_sched_exit_event(event, cpu, timestamp, thread);
1593 	if (!strcmp(event->name, "sched_migrate_task"))
1594 		process_sched_migrate_task_event(data, event, cpu, timestamp, thread);
1595 }
1596 
1597 static int process_sample_event(event_t *event)
1598 {
1599 	struct sample_data data;
1600 	struct thread *thread;
1601 
1602 	if (!(sample_type & PERF_SAMPLE_RAW))
1603 		return 0;
1604 
1605 	memset(&data, 0, sizeof(data));
1606 	data.time = -1;
1607 	data.cpu = -1;
1608 	data.period = -1;
1609 
1610 	event__parse_sample(event, sample_type, &data);
1611 
1612 	dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1613 		event->header.misc,
1614 		data.pid, data.tid,
1615 		(void *)(long)data.ip,
1616 		(long long)data.period);
1617 
1618 	thread = threads__findnew(data.pid);
1619 	if (thread == NULL) {
1620 		pr_debug("problem processing %d event, skipping it.\n",
1621 			 event->header.type);
1622 		return -1;
1623 	}
1624 
1625 	dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1626 
1627 	if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1628 		return 0;
1629 
1630 	process_raw_event(event, data.raw_data, data.cpu, data.time, thread);
1631 
1632 	return 0;
1633 }
1634 
1635 static int process_lost_event(event_t *event __used)
1636 {
1637 	nr_lost_chunks++;
1638 	nr_lost_events += event->lost.lost;
1639 
1640 	return 0;
1641 }
1642 
1643 static int sample_type_check(u64 type)
1644 {
1645 	sample_type = type;
1646 
1647 	if (!(sample_type & PERF_SAMPLE_RAW)) {
1648 		fprintf(stderr,
1649 			"No trace sample to read. Did you call perf record "
1650 			"without -R?");
1651 		return -1;
1652 	}
1653 
1654 	return 0;
1655 }
1656 
1657 static struct perf_file_handler file_handler = {
1658 	.process_sample_event	= process_sample_event,
1659 	.process_comm_event	= event__process_comm,
1660 	.process_lost_event	= process_lost_event,
1661 	.sample_type_check	= sample_type_check,
1662 };
1663 
1664 static int read_events(void)
1665 {
1666 	register_idle_thread();
1667 	register_perf_file_handler(&file_handler);
1668 
1669 	return mmap_dispatch_perf_file(&header, input_name, 0, 0,
1670 				       &event__cwdlen, &event__cwd);
1671 }
1672 
1673 static void print_bad_events(void)
1674 {
1675 	if (nr_unordered_timestamps && nr_timestamps) {
1676 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1677 			(double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1678 			nr_unordered_timestamps, nr_timestamps);
1679 	}
1680 	if (nr_lost_events && nr_events) {
1681 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1682 			(double)nr_lost_events/(double)nr_events*100.0,
1683 			nr_lost_events, nr_events, nr_lost_chunks);
1684 	}
1685 	if (nr_state_machine_bugs && nr_timestamps) {
1686 		printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1687 			(double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1688 			nr_state_machine_bugs, nr_timestamps);
1689 		if (nr_lost_events)
1690 			printf(" (due to lost events?)");
1691 		printf("\n");
1692 	}
1693 	if (nr_context_switch_bugs && nr_timestamps) {
1694 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1695 			(double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1696 			nr_context_switch_bugs, nr_timestamps);
1697 		if (nr_lost_events)
1698 			printf(" (due to lost events?)");
1699 		printf("\n");
1700 	}
1701 }
1702 
1703 static void __cmd_lat(void)
1704 {
1705 	struct rb_node *next;
1706 
1707 	setup_pager();
1708 	read_events();
1709 	sort_lat();
1710 
1711 	printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1712 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1713 	printf(" ---------------------------------------------------------------------------------------------------------------\n");
1714 
1715 	next = rb_first(&sorted_atom_root);
1716 
1717 	while (next) {
1718 		struct work_atoms *work_list;
1719 
1720 		work_list = rb_entry(next, struct work_atoms, node);
1721 		output_lat_thread(work_list);
1722 		next = rb_next(next);
1723 	}
1724 
1725 	printf(" -----------------------------------------------------------------------------------------\n");
1726 	printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1727 		(double)all_runtime/1e6, all_count);
1728 
1729 	printf(" ---------------------------------------------------\n");
1730 
1731 	print_bad_events();
1732 	printf("\n");
1733 
1734 }
1735 
1736 static struct trace_sched_handler map_ops  = {
1737 	.wakeup_event		= NULL,
1738 	.switch_event		= map_switch_event,
1739 	.runtime_event		= NULL,
1740 	.fork_event		= NULL,
1741 };
1742 
1743 static void __cmd_map(void)
1744 {
1745 	max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1746 
1747 	setup_pager();
1748 	read_events();
1749 	print_bad_events();
1750 }
1751 
1752 static void __cmd_replay(void)
1753 {
1754 	unsigned long i;
1755 
1756 	calibrate_run_measurement_overhead();
1757 	calibrate_sleep_measurement_overhead();
1758 
1759 	test_calibrations();
1760 
1761 	read_events();
1762 
1763 	printf("nr_run_events:        %ld\n", nr_run_events);
1764 	printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1765 	printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1766 
1767 	if (targetless_wakeups)
1768 		printf("target-less wakeups:  %ld\n", targetless_wakeups);
1769 	if (multitarget_wakeups)
1770 		printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1771 	if (nr_run_events_optimized)
1772 		printf("run atoms optimized: %ld\n",
1773 			nr_run_events_optimized);
1774 
1775 	print_task_traces();
1776 	add_cross_task_wakeups();
1777 
1778 	create_tasks();
1779 	printf("------------------------------------------------------------\n");
1780 	for (i = 0; i < replay_repeat; i++)
1781 		run_one_test();
1782 }
1783 
1784 
1785 static const char * const sched_usage[] = {
1786 	"perf sched [<options>] {record|latency|map|replay|trace}",
1787 	NULL
1788 };
1789 
1790 static const struct option sched_options[] = {
1791 	OPT_STRING('i', "input", &input_name, "file",
1792 		    "input file name"),
1793 	OPT_BOOLEAN('v', "verbose", &verbose,
1794 		    "be more verbose (show symbol address, etc)"),
1795 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1796 		    "dump raw trace in ASCII"),
1797 	OPT_END()
1798 };
1799 
1800 static const char * const latency_usage[] = {
1801 	"perf sched latency [<options>]",
1802 	NULL
1803 };
1804 
1805 static const struct option latency_options[] = {
1806 	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1807 		   "sort by key(s): runtime, switch, avg, max"),
1808 	OPT_BOOLEAN('v', "verbose", &verbose,
1809 		    "be more verbose (show symbol address, etc)"),
1810 	OPT_INTEGER('C', "CPU", &profile_cpu,
1811 		    "CPU to profile on"),
1812 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1813 		    "dump raw trace in ASCII"),
1814 	OPT_END()
1815 };
1816 
1817 static const char * const replay_usage[] = {
1818 	"perf sched replay [<options>]",
1819 	NULL
1820 };
1821 
1822 static const struct option replay_options[] = {
1823 	OPT_INTEGER('r', "repeat", &replay_repeat,
1824 		    "repeat the workload replay N times (-1: infinite)"),
1825 	OPT_BOOLEAN('v', "verbose", &verbose,
1826 		    "be more verbose (show symbol address, etc)"),
1827 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1828 		    "dump raw trace in ASCII"),
1829 	OPT_END()
1830 };
1831 
1832 static void setup_sorting(void)
1833 {
1834 	char *tmp, *tok, *str = strdup(sort_order);
1835 
1836 	for (tok = strtok_r(str, ", ", &tmp);
1837 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
1838 		if (sort_dimension__add(tok, &sort_list) < 0) {
1839 			error("Unknown --sort key: `%s'", tok);
1840 			usage_with_options(latency_usage, latency_options);
1841 		}
1842 	}
1843 
1844 	free(str);
1845 
1846 	sort_dimension__add("pid", &cmp_pid);
1847 }
1848 
1849 static const char *record_args[] = {
1850 	"record",
1851 	"-a",
1852 	"-R",
1853 	"-M",
1854 	"-f",
1855 	"-m", "1024",
1856 	"-c", "1",
1857 	"-e", "sched:sched_switch:r",
1858 	"-e", "sched:sched_stat_wait:r",
1859 	"-e", "sched:sched_stat_sleep:r",
1860 	"-e", "sched:sched_stat_iowait:r",
1861 	"-e", "sched:sched_stat_runtime:r",
1862 	"-e", "sched:sched_process_exit:r",
1863 	"-e", "sched:sched_process_fork:r",
1864 	"-e", "sched:sched_wakeup:r",
1865 	"-e", "sched:sched_migrate_task:r",
1866 };
1867 
1868 static int __cmd_record(int argc, const char **argv)
1869 {
1870 	unsigned int rec_argc, i, j;
1871 	const char **rec_argv;
1872 
1873 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1874 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1875 
1876 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1877 		rec_argv[i] = strdup(record_args[i]);
1878 
1879 	for (j = 1; j < (unsigned int)argc; j++, i++)
1880 		rec_argv[i] = argv[j];
1881 
1882 	BUG_ON(i != rec_argc);
1883 
1884 	return cmd_record(i, rec_argv, NULL);
1885 }
1886 
1887 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1888 {
1889 	argc = parse_options(argc, argv, sched_options, sched_usage,
1890 			     PARSE_OPT_STOP_AT_NON_OPTION);
1891 	if (!argc)
1892 		usage_with_options(sched_usage, sched_options);
1893 
1894 	/*
1895 	 * Aliased to 'perf trace' for now:
1896 	 */
1897 	if (!strcmp(argv[0], "trace"))
1898 		return cmd_trace(argc, argv, prefix);
1899 
1900 	symbol__init(0);
1901 	if (!strncmp(argv[0], "rec", 3)) {
1902 		return __cmd_record(argc, argv);
1903 	} else if (!strncmp(argv[0], "lat", 3)) {
1904 		trace_handler = &lat_ops;
1905 		if (argc > 1) {
1906 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1907 			if (argc)
1908 				usage_with_options(latency_usage, latency_options);
1909 		}
1910 		setup_sorting();
1911 		__cmd_lat();
1912 	} else if (!strcmp(argv[0], "map")) {
1913 		trace_handler = &map_ops;
1914 		setup_sorting();
1915 		__cmd_map();
1916 	} else if (!strncmp(argv[0], "rep", 3)) {
1917 		trace_handler = &replay_ops;
1918 		if (argc) {
1919 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1920 			if (argc)
1921 				usage_with_options(replay_usage, replay_options);
1922 		}
1923 		__cmd_replay();
1924 	} else {
1925 		usage_with_options(sched_usage, sched_options);
1926 	}
1927 
1928 	return 0;
1929 }
1930