xref: /openbmc/linux/tools/perf/builtin-sched.c (revision af958a38)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/util.h"
5 #include "util/evlist.h"
6 #include "util/cache.h"
7 #include "util/evsel.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/tool.h"
13 #include "util/cloexec.h"
14 
15 #include "util/parse-options.h"
16 #include "util/trace-event.h"
17 
18 #include "util/debug.h"
19 
20 #include <sys/prctl.h>
21 #include <sys/resource.h>
22 
23 #include <semaphore.h>
24 #include <pthread.h>
25 #include <math.h>
26 
27 #define PR_SET_NAME		15               /* Set process name */
28 #define MAX_CPUS		4096
29 #define COMM_LEN		20
30 #define SYM_LEN			129
31 #define MAX_PID			65536
32 
33 struct sched_atom;
34 
35 struct task_desc {
36 	unsigned long		nr;
37 	unsigned long		pid;
38 	char			comm[COMM_LEN];
39 
40 	unsigned long		nr_events;
41 	unsigned long		curr_event;
42 	struct sched_atom	**atoms;
43 
44 	pthread_t		thread;
45 	sem_t			sleep_sem;
46 
47 	sem_t			ready_for_work;
48 	sem_t			work_done_sem;
49 
50 	u64			cpu_usage;
51 };
52 
53 enum sched_event_type {
54 	SCHED_EVENT_RUN,
55 	SCHED_EVENT_SLEEP,
56 	SCHED_EVENT_WAKEUP,
57 	SCHED_EVENT_MIGRATION,
58 };
59 
60 struct sched_atom {
61 	enum sched_event_type	type;
62 	int			specific_wait;
63 	u64			timestamp;
64 	u64			duration;
65 	unsigned long		nr;
66 	sem_t			*wait_sem;
67 	struct task_desc	*wakee;
68 };
69 
70 #define TASK_STATE_TO_CHAR_STR "RSDTtZXxKWP"
71 
72 enum thread_state {
73 	THREAD_SLEEPING = 0,
74 	THREAD_WAIT_CPU,
75 	THREAD_SCHED_IN,
76 	THREAD_IGNORE
77 };
78 
79 struct work_atom {
80 	struct list_head	list;
81 	enum thread_state	state;
82 	u64			sched_out_time;
83 	u64			wake_up_time;
84 	u64			sched_in_time;
85 	u64			runtime;
86 };
87 
88 struct work_atoms {
89 	struct list_head	work_list;
90 	struct thread		*thread;
91 	struct rb_node		node;
92 	u64			max_lat;
93 	u64			max_lat_at;
94 	u64			total_lat;
95 	u64			nb_atoms;
96 	u64			total_runtime;
97 };
98 
99 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
100 
101 struct perf_sched;
102 
103 struct trace_sched_handler {
104 	int (*switch_event)(struct perf_sched *sched, struct perf_evsel *evsel,
105 			    struct perf_sample *sample, struct machine *machine);
106 
107 	int (*runtime_event)(struct perf_sched *sched, struct perf_evsel *evsel,
108 			     struct perf_sample *sample, struct machine *machine);
109 
110 	int (*wakeup_event)(struct perf_sched *sched, struct perf_evsel *evsel,
111 			    struct perf_sample *sample, struct machine *machine);
112 
113 	/* PERF_RECORD_FORK event, not sched_process_fork tracepoint */
114 	int (*fork_event)(struct perf_sched *sched, union perf_event *event,
115 			  struct machine *machine);
116 
117 	int (*migrate_task_event)(struct perf_sched *sched,
118 				  struct perf_evsel *evsel,
119 				  struct perf_sample *sample,
120 				  struct machine *machine);
121 };
122 
123 struct perf_sched {
124 	struct perf_tool tool;
125 	const char	 *sort_order;
126 	unsigned long	 nr_tasks;
127 	struct task_desc *pid_to_task[MAX_PID];
128 	struct task_desc **tasks;
129 	const struct trace_sched_handler *tp_handler;
130 	pthread_mutex_t	 start_work_mutex;
131 	pthread_mutex_t	 work_done_wait_mutex;
132 	int		 profile_cpu;
133 /*
134  * Track the current task - that way we can know whether there's any
135  * weird events, such as a task being switched away that is not current.
136  */
137 	int		 max_cpu;
138 	u32		 curr_pid[MAX_CPUS];
139 	struct thread	 *curr_thread[MAX_CPUS];
140 	char		 next_shortname1;
141 	char		 next_shortname2;
142 	unsigned int	 replay_repeat;
143 	unsigned long	 nr_run_events;
144 	unsigned long	 nr_sleep_events;
145 	unsigned long	 nr_wakeup_events;
146 	unsigned long	 nr_sleep_corrections;
147 	unsigned long	 nr_run_events_optimized;
148 	unsigned long	 targetless_wakeups;
149 	unsigned long	 multitarget_wakeups;
150 	unsigned long	 nr_runs;
151 	unsigned long	 nr_timestamps;
152 	unsigned long	 nr_unordered_timestamps;
153 	unsigned long	 nr_context_switch_bugs;
154 	unsigned long	 nr_events;
155 	unsigned long	 nr_lost_chunks;
156 	unsigned long	 nr_lost_events;
157 	u64		 run_measurement_overhead;
158 	u64		 sleep_measurement_overhead;
159 	u64		 start_time;
160 	u64		 cpu_usage;
161 	u64		 runavg_cpu_usage;
162 	u64		 parent_cpu_usage;
163 	u64		 runavg_parent_cpu_usage;
164 	u64		 sum_runtime;
165 	u64		 sum_fluct;
166 	u64		 run_avg;
167 	u64		 all_runtime;
168 	u64		 all_count;
169 	u64		 cpu_last_switched[MAX_CPUS];
170 	struct rb_root	 atom_root, sorted_atom_root;
171 	struct list_head sort_list, cmp_pid;
172 };
173 
174 static u64 get_nsecs(void)
175 {
176 	struct timespec ts;
177 
178 	clock_gettime(CLOCK_MONOTONIC, &ts);
179 
180 	return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
181 }
182 
183 static void burn_nsecs(struct perf_sched *sched, u64 nsecs)
184 {
185 	u64 T0 = get_nsecs(), T1;
186 
187 	do {
188 		T1 = get_nsecs();
189 	} while (T1 + sched->run_measurement_overhead < T0 + nsecs);
190 }
191 
192 static void sleep_nsecs(u64 nsecs)
193 {
194 	struct timespec ts;
195 
196 	ts.tv_nsec = nsecs % 999999999;
197 	ts.tv_sec = nsecs / 999999999;
198 
199 	nanosleep(&ts, NULL);
200 }
201 
202 static void calibrate_run_measurement_overhead(struct perf_sched *sched)
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 		burn_nsecs(sched, 0);
210 		T1 = get_nsecs();
211 		delta = T1-T0;
212 		min_delta = min(min_delta, delta);
213 	}
214 	sched->run_measurement_overhead = min_delta;
215 
216 	printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
217 }
218 
219 static void calibrate_sleep_measurement_overhead(struct perf_sched *sched)
220 {
221 	u64 T0, T1, delta, min_delta = 1000000000ULL;
222 	int i;
223 
224 	for (i = 0; i < 10; i++) {
225 		T0 = get_nsecs();
226 		sleep_nsecs(10000);
227 		T1 = get_nsecs();
228 		delta = T1-T0;
229 		min_delta = min(min_delta, delta);
230 	}
231 	min_delta -= 10000;
232 	sched->sleep_measurement_overhead = min_delta;
233 
234 	printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
235 }
236 
237 static struct sched_atom *
238 get_new_event(struct task_desc *task, u64 timestamp)
239 {
240 	struct sched_atom *event = zalloc(sizeof(*event));
241 	unsigned long idx = task->nr_events;
242 	size_t size;
243 
244 	event->timestamp = timestamp;
245 	event->nr = idx;
246 
247 	task->nr_events++;
248 	size = sizeof(struct sched_atom *) * task->nr_events;
249 	task->atoms = realloc(task->atoms, size);
250 	BUG_ON(!task->atoms);
251 
252 	task->atoms[idx] = event;
253 
254 	return event;
255 }
256 
257 static struct sched_atom *last_event(struct task_desc *task)
258 {
259 	if (!task->nr_events)
260 		return NULL;
261 
262 	return task->atoms[task->nr_events - 1];
263 }
264 
265 static void add_sched_event_run(struct perf_sched *sched, struct task_desc *task,
266 				u64 timestamp, u64 duration)
267 {
268 	struct sched_atom *event, *curr_event = last_event(task);
269 
270 	/*
271 	 * optimize an existing RUN event by merging this one
272 	 * to it:
273 	 */
274 	if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
275 		sched->nr_run_events_optimized++;
276 		curr_event->duration += duration;
277 		return;
278 	}
279 
280 	event = get_new_event(task, timestamp);
281 
282 	event->type = SCHED_EVENT_RUN;
283 	event->duration = duration;
284 
285 	sched->nr_run_events++;
286 }
287 
288 static void add_sched_event_wakeup(struct perf_sched *sched, struct task_desc *task,
289 				   u64 timestamp, struct task_desc *wakee)
290 {
291 	struct sched_atom *event, *wakee_event;
292 
293 	event = get_new_event(task, timestamp);
294 	event->type = SCHED_EVENT_WAKEUP;
295 	event->wakee = wakee;
296 
297 	wakee_event = last_event(wakee);
298 	if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
299 		sched->targetless_wakeups++;
300 		return;
301 	}
302 	if (wakee_event->wait_sem) {
303 		sched->multitarget_wakeups++;
304 		return;
305 	}
306 
307 	wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
308 	sem_init(wakee_event->wait_sem, 0, 0);
309 	wakee_event->specific_wait = 1;
310 	event->wait_sem = wakee_event->wait_sem;
311 
312 	sched->nr_wakeup_events++;
313 }
314 
315 static void add_sched_event_sleep(struct perf_sched *sched, struct task_desc *task,
316 				  u64 timestamp, u64 task_state __maybe_unused)
317 {
318 	struct sched_atom *event = get_new_event(task, timestamp);
319 
320 	event->type = SCHED_EVENT_SLEEP;
321 
322 	sched->nr_sleep_events++;
323 }
324 
325 static struct task_desc *register_pid(struct perf_sched *sched,
326 				      unsigned long pid, const char *comm)
327 {
328 	struct task_desc *task;
329 
330 	BUG_ON(pid >= MAX_PID);
331 
332 	task = sched->pid_to_task[pid];
333 
334 	if (task)
335 		return task;
336 
337 	task = zalloc(sizeof(*task));
338 	task->pid = pid;
339 	task->nr = sched->nr_tasks;
340 	strcpy(task->comm, comm);
341 	/*
342 	 * every task starts in sleeping state - this gets ignored
343 	 * if there's no wakeup pointing to this sleep state:
344 	 */
345 	add_sched_event_sleep(sched, task, 0, 0);
346 
347 	sched->pid_to_task[pid] = task;
348 	sched->nr_tasks++;
349 	sched->tasks = realloc(sched->tasks, sched->nr_tasks * sizeof(struct task_task *));
350 	BUG_ON(!sched->tasks);
351 	sched->tasks[task->nr] = task;
352 
353 	if (verbose)
354 		printf("registered task #%ld, PID %ld (%s)\n", sched->nr_tasks, pid, comm);
355 
356 	return task;
357 }
358 
359 
360 static void print_task_traces(struct perf_sched *sched)
361 {
362 	struct task_desc *task;
363 	unsigned long i;
364 
365 	for (i = 0; i < sched->nr_tasks; i++) {
366 		task = sched->tasks[i];
367 		printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
368 			task->nr, task->comm, task->pid, task->nr_events);
369 	}
370 }
371 
372 static void add_cross_task_wakeups(struct perf_sched *sched)
373 {
374 	struct task_desc *task1, *task2;
375 	unsigned long i, j;
376 
377 	for (i = 0; i < sched->nr_tasks; i++) {
378 		task1 = sched->tasks[i];
379 		j = i + 1;
380 		if (j == sched->nr_tasks)
381 			j = 0;
382 		task2 = sched->tasks[j];
383 		add_sched_event_wakeup(sched, task1, 0, task2);
384 	}
385 }
386 
387 static void perf_sched__process_event(struct perf_sched *sched,
388 				      struct sched_atom *atom)
389 {
390 	int ret = 0;
391 
392 	switch (atom->type) {
393 		case SCHED_EVENT_RUN:
394 			burn_nsecs(sched, atom->duration);
395 			break;
396 		case SCHED_EVENT_SLEEP:
397 			if (atom->wait_sem)
398 				ret = sem_wait(atom->wait_sem);
399 			BUG_ON(ret);
400 			break;
401 		case SCHED_EVENT_WAKEUP:
402 			if (atom->wait_sem)
403 				ret = sem_post(atom->wait_sem);
404 			BUG_ON(ret);
405 			break;
406 		case SCHED_EVENT_MIGRATION:
407 			break;
408 		default:
409 			BUG_ON(1);
410 	}
411 }
412 
413 static u64 get_cpu_usage_nsec_parent(void)
414 {
415 	struct rusage ru;
416 	u64 sum;
417 	int err;
418 
419 	err = getrusage(RUSAGE_SELF, &ru);
420 	BUG_ON(err);
421 
422 	sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
423 	sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
424 
425 	return sum;
426 }
427 
428 static int self_open_counters(void)
429 {
430 	struct perf_event_attr attr;
431 	int fd;
432 
433 	memset(&attr, 0, sizeof(attr));
434 
435 	attr.type = PERF_TYPE_SOFTWARE;
436 	attr.config = PERF_COUNT_SW_TASK_CLOCK;
437 
438 	fd = sys_perf_event_open(&attr, 0, -1, -1,
439 				 perf_event_open_cloexec_flag());
440 
441 	if (fd < 0)
442 		pr_err("Error: sys_perf_event_open() syscall returned "
443 		       "with %d (%s)\n", fd, strerror(errno));
444 	return fd;
445 }
446 
447 static u64 get_cpu_usage_nsec_self(int fd)
448 {
449 	u64 runtime;
450 	int ret;
451 
452 	ret = read(fd, &runtime, sizeof(runtime));
453 	BUG_ON(ret != sizeof(runtime));
454 
455 	return runtime;
456 }
457 
458 struct sched_thread_parms {
459 	struct task_desc  *task;
460 	struct perf_sched *sched;
461 };
462 
463 static void *thread_func(void *ctx)
464 {
465 	struct sched_thread_parms *parms = ctx;
466 	struct task_desc *this_task = parms->task;
467 	struct perf_sched *sched = parms->sched;
468 	u64 cpu_usage_0, cpu_usage_1;
469 	unsigned long i, ret;
470 	char comm2[22];
471 	int fd;
472 
473 	zfree(&parms);
474 
475 	sprintf(comm2, ":%s", this_task->comm);
476 	prctl(PR_SET_NAME, comm2);
477 	fd = self_open_counters();
478 	if (fd < 0)
479 		return NULL;
480 again:
481 	ret = sem_post(&this_task->ready_for_work);
482 	BUG_ON(ret);
483 	ret = pthread_mutex_lock(&sched->start_work_mutex);
484 	BUG_ON(ret);
485 	ret = pthread_mutex_unlock(&sched->start_work_mutex);
486 	BUG_ON(ret);
487 
488 	cpu_usage_0 = get_cpu_usage_nsec_self(fd);
489 
490 	for (i = 0; i < this_task->nr_events; i++) {
491 		this_task->curr_event = i;
492 		perf_sched__process_event(sched, this_task->atoms[i]);
493 	}
494 
495 	cpu_usage_1 = get_cpu_usage_nsec_self(fd);
496 	this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
497 	ret = sem_post(&this_task->work_done_sem);
498 	BUG_ON(ret);
499 
500 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
501 	BUG_ON(ret);
502 	ret = pthread_mutex_unlock(&sched->work_done_wait_mutex);
503 	BUG_ON(ret);
504 
505 	goto again;
506 }
507 
508 static void create_tasks(struct perf_sched *sched)
509 {
510 	struct task_desc *task;
511 	pthread_attr_t attr;
512 	unsigned long i;
513 	int err;
514 
515 	err = pthread_attr_init(&attr);
516 	BUG_ON(err);
517 	err = pthread_attr_setstacksize(&attr,
518 			(size_t) max(16 * 1024, PTHREAD_STACK_MIN));
519 	BUG_ON(err);
520 	err = pthread_mutex_lock(&sched->start_work_mutex);
521 	BUG_ON(err);
522 	err = pthread_mutex_lock(&sched->work_done_wait_mutex);
523 	BUG_ON(err);
524 	for (i = 0; i < sched->nr_tasks; i++) {
525 		struct sched_thread_parms *parms = malloc(sizeof(*parms));
526 		BUG_ON(parms == NULL);
527 		parms->task = task = sched->tasks[i];
528 		parms->sched = sched;
529 		sem_init(&task->sleep_sem, 0, 0);
530 		sem_init(&task->ready_for_work, 0, 0);
531 		sem_init(&task->work_done_sem, 0, 0);
532 		task->curr_event = 0;
533 		err = pthread_create(&task->thread, &attr, thread_func, parms);
534 		BUG_ON(err);
535 	}
536 }
537 
538 static void wait_for_tasks(struct perf_sched *sched)
539 {
540 	u64 cpu_usage_0, cpu_usage_1;
541 	struct task_desc *task;
542 	unsigned long i, ret;
543 
544 	sched->start_time = get_nsecs();
545 	sched->cpu_usage = 0;
546 	pthread_mutex_unlock(&sched->work_done_wait_mutex);
547 
548 	for (i = 0; i < sched->nr_tasks; i++) {
549 		task = sched->tasks[i];
550 		ret = sem_wait(&task->ready_for_work);
551 		BUG_ON(ret);
552 		sem_init(&task->ready_for_work, 0, 0);
553 	}
554 	ret = pthread_mutex_lock(&sched->work_done_wait_mutex);
555 	BUG_ON(ret);
556 
557 	cpu_usage_0 = get_cpu_usage_nsec_parent();
558 
559 	pthread_mutex_unlock(&sched->start_work_mutex);
560 
561 	for (i = 0; i < sched->nr_tasks; i++) {
562 		task = sched->tasks[i];
563 		ret = sem_wait(&task->work_done_sem);
564 		BUG_ON(ret);
565 		sem_init(&task->work_done_sem, 0, 0);
566 		sched->cpu_usage += task->cpu_usage;
567 		task->cpu_usage = 0;
568 	}
569 
570 	cpu_usage_1 = get_cpu_usage_nsec_parent();
571 	if (!sched->runavg_cpu_usage)
572 		sched->runavg_cpu_usage = sched->cpu_usage;
573 	sched->runavg_cpu_usage = (sched->runavg_cpu_usage * 9 + sched->cpu_usage) / 10;
574 
575 	sched->parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
576 	if (!sched->runavg_parent_cpu_usage)
577 		sched->runavg_parent_cpu_usage = sched->parent_cpu_usage;
578 	sched->runavg_parent_cpu_usage = (sched->runavg_parent_cpu_usage * 9 +
579 					 sched->parent_cpu_usage)/10;
580 
581 	ret = pthread_mutex_lock(&sched->start_work_mutex);
582 	BUG_ON(ret);
583 
584 	for (i = 0; i < sched->nr_tasks; i++) {
585 		task = sched->tasks[i];
586 		sem_init(&task->sleep_sem, 0, 0);
587 		task->curr_event = 0;
588 	}
589 }
590 
591 static void run_one_test(struct perf_sched *sched)
592 {
593 	u64 T0, T1, delta, avg_delta, fluct;
594 
595 	T0 = get_nsecs();
596 	wait_for_tasks(sched);
597 	T1 = get_nsecs();
598 
599 	delta = T1 - T0;
600 	sched->sum_runtime += delta;
601 	sched->nr_runs++;
602 
603 	avg_delta = sched->sum_runtime / sched->nr_runs;
604 	if (delta < avg_delta)
605 		fluct = avg_delta - delta;
606 	else
607 		fluct = delta - avg_delta;
608 	sched->sum_fluct += fluct;
609 	if (!sched->run_avg)
610 		sched->run_avg = delta;
611 	sched->run_avg = (sched->run_avg * 9 + delta) / 10;
612 
613 	printf("#%-3ld: %0.3f, ", sched->nr_runs, (double)delta / 1000000.0);
614 
615 	printf("ravg: %0.2f, ", (double)sched->run_avg / 1e6);
616 
617 	printf("cpu: %0.2f / %0.2f",
618 		(double)sched->cpu_usage / 1e6, (double)sched->runavg_cpu_usage / 1e6);
619 
620 #if 0
621 	/*
622 	 * rusage statistics done by the parent, these are less
623 	 * accurate than the sched->sum_exec_runtime based statistics:
624 	 */
625 	printf(" [%0.2f / %0.2f]",
626 		(double)sched->parent_cpu_usage/1e6,
627 		(double)sched->runavg_parent_cpu_usage/1e6);
628 #endif
629 
630 	printf("\n");
631 
632 	if (sched->nr_sleep_corrections)
633 		printf(" (%ld sleep corrections)\n", sched->nr_sleep_corrections);
634 	sched->nr_sleep_corrections = 0;
635 }
636 
637 static void test_calibrations(struct perf_sched *sched)
638 {
639 	u64 T0, T1;
640 
641 	T0 = get_nsecs();
642 	burn_nsecs(sched, 1e6);
643 	T1 = get_nsecs();
644 
645 	printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
646 
647 	T0 = get_nsecs();
648 	sleep_nsecs(1e6);
649 	T1 = get_nsecs();
650 
651 	printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
652 }
653 
654 static int
655 replay_wakeup_event(struct perf_sched *sched,
656 		    struct perf_evsel *evsel, struct perf_sample *sample,
657 		    struct machine *machine __maybe_unused)
658 {
659 	const char *comm = perf_evsel__strval(evsel, sample, "comm");
660 	const u32 pid	 = perf_evsel__intval(evsel, sample, "pid");
661 	struct task_desc *waker, *wakee;
662 
663 	if (verbose) {
664 		printf("sched_wakeup event %p\n", evsel);
665 
666 		printf(" ... pid %d woke up %s/%d\n", sample->tid, comm, pid);
667 	}
668 
669 	waker = register_pid(sched, sample->tid, "<unknown>");
670 	wakee = register_pid(sched, pid, comm);
671 
672 	add_sched_event_wakeup(sched, waker, sample->time, wakee);
673 	return 0;
674 }
675 
676 static int replay_switch_event(struct perf_sched *sched,
677 			       struct perf_evsel *evsel,
678 			       struct perf_sample *sample,
679 			       struct machine *machine __maybe_unused)
680 {
681 	const char *prev_comm  = perf_evsel__strval(evsel, sample, "prev_comm"),
682 		   *next_comm  = perf_evsel__strval(evsel, sample, "next_comm");
683 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
684 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
685 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
686 	struct task_desc *prev, __maybe_unused *next;
687 	u64 timestamp0, timestamp = sample->time;
688 	int cpu = sample->cpu;
689 	s64 delta;
690 
691 	if (verbose)
692 		printf("sched_switch event %p\n", evsel);
693 
694 	if (cpu >= MAX_CPUS || cpu < 0)
695 		return 0;
696 
697 	timestamp0 = sched->cpu_last_switched[cpu];
698 	if (timestamp0)
699 		delta = timestamp - timestamp0;
700 	else
701 		delta = 0;
702 
703 	if (delta < 0) {
704 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
705 		return -1;
706 	}
707 
708 	pr_debug(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
709 		 prev_comm, prev_pid, next_comm, next_pid, delta);
710 
711 	prev = register_pid(sched, prev_pid, prev_comm);
712 	next = register_pid(sched, next_pid, next_comm);
713 
714 	sched->cpu_last_switched[cpu] = timestamp;
715 
716 	add_sched_event_run(sched, prev, timestamp, delta);
717 	add_sched_event_sleep(sched, prev, timestamp, prev_state);
718 
719 	return 0;
720 }
721 
722 static int replay_fork_event(struct perf_sched *sched,
723 			     union perf_event *event,
724 			     struct machine *machine)
725 {
726 	struct thread *child, *parent;
727 
728 	child = machine__findnew_thread(machine, event->fork.pid,
729 					event->fork.tid);
730 	parent = machine__findnew_thread(machine, event->fork.ppid,
731 					 event->fork.ptid);
732 
733 	if (child == NULL || parent == NULL) {
734 		pr_debug("thread does not exist on fork event: child %p, parent %p\n",
735 				 child, parent);
736 		return 0;
737 	}
738 
739 	if (verbose) {
740 		printf("fork event\n");
741 		printf("... parent: %s/%d\n", thread__comm_str(parent), parent->tid);
742 		printf("...  child: %s/%d\n", thread__comm_str(child), child->tid);
743 	}
744 
745 	register_pid(sched, parent->tid, thread__comm_str(parent));
746 	register_pid(sched, child->tid, thread__comm_str(child));
747 	return 0;
748 }
749 
750 struct sort_dimension {
751 	const char		*name;
752 	sort_fn_t		cmp;
753 	struct list_head	list;
754 };
755 
756 static int
757 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
758 {
759 	struct sort_dimension *sort;
760 	int ret = 0;
761 
762 	BUG_ON(list_empty(list));
763 
764 	list_for_each_entry(sort, list, list) {
765 		ret = sort->cmp(l, r);
766 		if (ret)
767 			return ret;
768 	}
769 
770 	return ret;
771 }
772 
773 static struct work_atoms *
774 thread_atoms_search(struct rb_root *root, struct thread *thread,
775 			 struct list_head *sort_list)
776 {
777 	struct rb_node *node = root->rb_node;
778 	struct work_atoms key = { .thread = thread };
779 
780 	while (node) {
781 		struct work_atoms *atoms;
782 		int cmp;
783 
784 		atoms = container_of(node, struct work_atoms, node);
785 
786 		cmp = thread_lat_cmp(sort_list, &key, atoms);
787 		if (cmp > 0)
788 			node = node->rb_left;
789 		else if (cmp < 0)
790 			node = node->rb_right;
791 		else {
792 			BUG_ON(thread != atoms->thread);
793 			return atoms;
794 		}
795 	}
796 	return NULL;
797 }
798 
799 static void
800 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
801 			 struct list_head *sort_list)
802 {
803 	struct rb_node **new = &(root->rb_node), *parent = NULL;
804 
805 	while (*new) {
806 		struct work_atoms *this;
807 		int cmp;
808 
809 		this = container_of(*new, struct work_atoms, node);
810 		parent = *new;
811 
812 		cmp = thread_lat_cmp(sort_list, data, this);
813 
814 		if (cmp > 0)
815 			new = &((*new)->rb_left);
816 		else
817 			new = &((*new)->rb_right);
818 	}
819 
820 	rb_link_node(&data->node, parent, new);
821 	rb_insert_color(&data->node, root);
822 }
823 
824 static int thread_atoms_insert(struct perf_sched *sched, struct thread *thread)
825 {
826 	struct work_atoms *atoms = zalloc(sizeof(*atoms));
827 	if (!atoms) {
828 		pr_err("No memory at %s\n", __func__);
829 		return -1;
830 	}
831 
832 	atoms->thread = thread;
833 	INIT_LIST_HEAD(&atoms->work_list);
834 	__thread_latency_insert(&sched->atom_root, atoms, &sched->cmp_pid);
835 	return 0;
836 }
837 
838 static char sched_out_state(u64 prev_state)
839 {
840 	const char *str = TASK_STATE_TO_CHAR_STR;
841 
842 	return str[prev_state];
843 }
844 
845 static int
846 add_sched_out_event(struct work_atoms *atoms,
847 		    char run_state,
848 		    u64 timestamp)
849 {
850 	struct work_atom *atom = zalloc(sizeof(*atom));
851 	if (!atom) {
852 		pr_err("Non memory at %s", __func__);
853 		return -1;
854 	}
855 
856 	atom->sched_out_time = timestamp;
857 
858 	if (run_state == 'R') {
859 		atom->state = THREAD_WAIT_CPU;
860 		atom->wake_up_time = atom->sched_out_time;
861 	}
862 
863 	list_add_tail(&atom->list, &atoms->work_list);
864 	return 0;
865 }
866 
867 static void
868 add_runtime_event(struct work_atoms *atoms, u64 delta,
869 		  u64 timestamp __maybe_unused)
870 {
871 	struct work_atom *atom;
872 
873 	BUG_ON(list_empty(&atoms->work_list));
874 
875 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
876 
877 	atom->runtime += delta;
878 	atoms->total_runtime += delta;
879 }
880 
881 static void
882 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
883 {
884 	struct work_atom *atom;
885 	u64 delta;
886 
887 	if (list_empty(&atoms->work_list))
888 		return;
889 
890 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
891 
892 	if (atom->state != THREAD_WAIT_CPU)
893 		return;
894 
895 	if (timestamp < atom->wake_up_time) {
896 		atom->state = THREAD_IGNORE;
897 		return;
898 	}
899 
900 	atom->state = THREAD_SCHED_IN;
901 	atom->sched_in_time = timestamp;
902 
903 	delta = atom->sched_in_time - atom->wake_up_time;
904 	atoms->total_lat += delta;
905 	if (delta > atoms->max_lat) {
906 		atoms->max_lat = delta;
907 		atoms->max_lat_at = timestamp;
908 	}
909 	atoms->nb_atoms++;
910 }
911 
912 static int latency_switch_event(struct perf_sched *sched,
913 				struct perf_evsel *evsel,
914 				struct perf_sample *sample,
915 				struct machine *machine)
916 {
917 	const u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
918 		  next_pid = perf_evsel__intval(evsel, sample, "next_pid");
919 	const u64 prev_state = perf_evsel__intval(evsel, sample, "prev_state");
920 	struct work_atoms *out_events, *in_events;
921 	struct thread *sched_out, *sched_in;
922 	u64 timestamp0, timestamp = sample->time;
923 	int cpu = sample->cpu;
924 	s64 delta;
925 
926 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
927 
928 	timestamp0 = sched->cpu_last_switched[cpu];
929 	sched->cpu_last_switched[cpu] = timestamp;
930 	if (timestamp0)
931 		delta = timestamp - timestamp0;
932 	else
933 		delta = 0;
934 
935 	if (delta < 0) {
936 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
937 		return -1;
938 	}
939 
940 	sched_out = machine__findnew_thread(machine, -1, prev_pid);
941 	sched_in = machine__findnew_thread(machine, -1, next_pid);
942 
943 	out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
944 	if (!out_events) {
945 		if (thread_atoms_insert(sched, sched_out))
946 			return -1;
947 		out_events = thread_atoms_search(&sched->atom_root, sched_out, &sched->cmp_pid);
948 		if (!out_events) {
949 			pr_err("out-event: Internal tree error");
950 			return -1;
951 		}
952 	}
953 	if (add_sched_out_event(out_events, sched_out_state(prev_state), timestamp))
954 		return -1;
955 
956 	in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
957 	if (!in_events) {
958 		if (thread_atoms_insert(sched, sched_in))
959 			return -1;
960 		in_events = thread_atoms_search(&sched->atom_root, sched_in, &sched->cmp_pid);
961 		if (!in_events) {
962 			pr_err("in-event: Internal tree error");
963 			return -1;
964 		}
965 		/*
966 		 * Take came in we have not heard about yet,
967 		 * add in an initial atom in runnable state:
968 		 */
969 		if (add_sched_out_event(in_events, 'R', timestamp))
970 			return -1;
971 	}
972 	add_sched_in_event(in_events, timestamp);
973 
974 	return 0;
975 }
976 
977 static int latency_runtime_event(struct perf_sched *sched,
978 				 struct perf_evsel *evsel,
979 				 struct perf_sample *sample,
980 				 struct machine *machine)
981 {
982 	const u32 pid	   = perf_evsel__intval(evsel, sample, "pid");
983 	const u64 runtime  = perf_evsel__intval(evsel, sample, "runtime");
984 	struct thread *thread = machine__findnew_thread(machine, -1, pid);
985 	struct work_atoms *atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
986 	u64 timestamp = sample->time;
987 	int cpu = sample->cpu;
988 
989 	BUG_ON(cpu >= MAX_CPUS || cpu < 0);
990 	if (!atoms) {
991 		if (thread_atoms_insert(sched, thread))
992 			return -1;
993 		atoms = thread_atoms_search(&sched->atom_root, thread, &sched->cmp_pid);
994 		if (!atoms) {
995 			pr_err("in-event: Internal tree error");
996 			return -1;
997 		}
998 		if (add_sched_out_event(atoms, 'R', timestamp))
999 			return -1;
1000 	}
1001 
1002 	add_runtime_event(atoms, runtime, timestamp);
1003 	return 0;
1004 }
1005 
1006 static int latency_wakeup_event(struct perf_sched *sched,
1007 				struct perf_evsel *evsel,
1008 				struct perf_sample *sample,
1009 				struct machine *machine)
1010 {
1011 	const u32 pid	  = perf_evsel__intval(evsel, sample, "pid");
1012 	struct work_atoms *atoms;
1013 	struct work_atom *atom;
1014 	struct thread *wakee;
1015 	u64 timestamp = sample->time;
1016 
1017 	wakee = machine__findnew_thread(machine, -1, pid);
1018 	atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1019 	if (!atoms) {
1020 		if (thread_atoms_insert(sched, wakee))
1021 			return -1;
1022 		atoms = thread_atoms_search(&sched->atom_root, wakee, &sched->cmp_pid);
1023 		if (!atoms) {
1024 			pr_err("wakeup-event: Internal tree error");
1025 			return -1;
1026 		}
1027 		if (add_sched_out_event(atoms, 'S', timestamp))
1028 			return -1;
1029 	}
1030 
1031 	BUG_ON(list_empty(&atoms->work_list));
1032 
1033 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1034 
1035 	/*
1036 	 * As we do not guarantee the wakeup event happens when
1037 	 * task is out of run queue, also may happen when task is
1038 	 * on run queue and wakeup only change ->state to TASK_RUNNING,
1039 	 * then we should not set the ->wake_up_time when wake up a
1040 	 * task which is on run queue.
1041 	 *
1042 	 * You WILL be missing events if you've recorded only
1043 	 * one CPU, or are only looking at only one, so don't
1044 	 * skip in this case.
1045 	 */
1046 	if (sched->profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1047 		return 0;
1048 
1049 	sched->nr_timestamps++;
1050 	if (atom->sched_out_time > timestamp) {
1051 		sched->nr_unordered_timestamps++;
1052 		return 0;
1053 	}
1054 
1055 	atom->state = THREAD_WAIT_CPU;
1056 	atom->wake_up_time = timestamp;
1057 	return 0;
1058 }
1059 
1060 static int latency_migrate_task_event(struct perf_sched *sched,
1061 				      struct perf_evsel *evsel,
1062 				      struct perf_sample *sample,
1063 				      struct machine *machine)
1064 {
1065 	const u32 pid = perf_evsel__intval(evsel, sample, "pid");
1066 	u64 timestamp = sample->time;
1067 	struct work_atoms *atoms;
1068 	struct work_atom *atom;
1069 	struct thread *migrant;
1070 
1071 	/*
1072 	 * Only need to worry about migration when profiling one CPU.
1073 	 */
1074 	if (sched->profile_cpu == -1)
1075 		return 0;
1076 
1077 	migrant = machine__findnew_thread(machine, -1, pid);
1078 	atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1079 	if (!atoms) {
1080 		if (thread_atoms_insert(sched, migrant))
1081 			return -1;
1082 		register_pid(sched, migrant->tid, thread__comm_str(migrant));
1083 		atoms = thread_atoms_search(&sched->atom_root, migrant, &sched->cmp_pid);
1084 		if (!atoms) {
1085 			pr_err("migration-event: Internal tree error");
1086 			return -1;
1087 		}
1088 		if (add_sched_out_event(atoms, 'R', timestamp))
1089 			return -1;
1090 	}
1091 
1092 	BUG_ON(list_empty(&atoms->work_list));
1093 
1094 	atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1095 	atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1096 
1097 	sched->nr_timestamps++;
1098 
1099 	if (atom->sched_out_time > timestamp)
1100 		sched->nr_unordered_timestamps++;
1101 
1102 	return 0;
1103 }
1104 
1105 static void output_lat_thread(struct perf_sched *sched, struct work_atoms *work_list)
1106 {
1107 	int i;
1108 	int ret;
1109 	u64 avg;
1110 
1111 	if (!work_list->nb_atoms)
1112 		return;
1113 	/*
1114 	 * Ignore idle threads:
1115 	 */
1116 	if (!strcmp(thread__comm_str(work_list->thread), "swapper"))
1117 		return;
1118 
1119 	sched->all_runtime += work_list->total_runtime;
1120 	sched->all_count   += work_list->nb_atoms;
1121 
1122 	ret = printf("  %s:%d ", thread__comm_str(work_list->thread), work_list->thread->tid);
1123 
1124 	for (i = 0; i < 24 - ret; i++)
1125 		printf(" ");
1126 
1127 	avg = work_list->total_lat / work_list->nb_atoms;
1128 
1129 	printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %13.6f s\n",
1130 	      (double)work_list->total_runtime / 1e6,
1131 		 work_list->nb_atoms, (double)avg / 1e6,
1132 		 (double)work_list->max_lat / 1e6,
1133 		 (double)work_list->max_lat_at / 1e9);
1134 }
1135 
1136 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1137 {
1138 	if (l->thread->tid < r->thread->tid)
1139 		return -1;
1140 	if (l->thread->tid > r->thread->tid)
1141 		return 1;
1142 
1143 	return 0;
1144 }
1145 
1146 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1147 {
1148 	u64 avgl, avgr;
1149 
1150 	if (!l->nb_atoms)
1151 		return -1;
1152 
1153 	if (!r->nb_atoms)
1154 		return 1;
1155 
1156 	avgl = l->total_lat / l->nb_atoms;
1157 	avgr = r->total_lat / r->nb_atoms;
1158 
1159 	if (avgl < avgr)
1160 		return -1;
1161 	if (avgl > avgr)
1162 		return 1;
1163 
1164 	return 0;
1165 }
1166 
1167 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1168 {
1169 	if (l->max_lat < r->max_lat)
1170 		return -1;
1171 	if (l->max_lat > r->max_lat)
1172 		return 1;
1173 
1174 	return 0;
1175 }
1176 
1177 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1178 {
1179 	if (l->nb_atoms < r->nb_atoms)
1180 		return -1;
1181 	if (l->nb_atoms > r->nb_atoms)
1182 		return 1;
1183 
1184 	return 0;
1185 }
1186 
1187 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1188 {
1189 	if (l->total_runtime < r->total_runtime)
1190 		return -1;
1191 	if (l->total_runtime > r->total_runtime)
1192 		return 1;
1193 
1194 	return 0;
1195 }
1196 
1197 static int sort_dimension__add(const char *tok, struct list_head *list)
1198 {
1199 	size_t i;
1200 	static struct sort_dimension avg_sort_dimension = {
1201 		.name = "avg",
1202 		.cmp  = avg_cmp,
1203 	};
1204 	static struct sort_dimension max_sort_dimension = {
1205 		.name = "max",
1206 		.cmp  = max_cmp,
1207 	};
1208 	static struct sort_dimension pid_sort_dimension = {
1209 		.name = "pid",
1210 		.cmp  = pid_cmp,
1211 	};
1212 	static struct sort_dimension runtime_sort_dimension = {
1213 		.name = "runtime",
1214 		.cmp  = runtime_cmp,
1215 	};
1216 	static struct sort_dimension switch_sort_dimension = {
1217 		.name = "switch",
1218 		.cmp  = switch_cmp,
1219 	};
1220 	struct sort_dimension *available_sorts[] = {
1221 		&pid_sort_dimension,
1222 		&avg_sort_dimension,
1223 		&max_sort_dimension,
1224 		&switch_sort_dimension,
1225 		&runtime_sort_dimension,
1226 	};
1227 
1228 	for (i = 0; i < ARRAY_SIZE(available_sorts); i++) {
1229 		if (!strcmp(available_sorts[i]->name, tok)) {
1230 			list_add_tail(&available_sorts[i]->list, list);
1231 
1232 			return 0;
1233 		}
1234 	}
1235 
1236 	return -1;
1237 }
1238 
1239 static void perf_sched__sort_lat(struct perf_sched *sched)
1240 {
1241 	struct rb_node *node;
1242 
1243 	for (;;) {
1244 		struct work_atoms *data;
1245 		node = rb_first(&sched->atom_root);
1246 		if (!node)
1247 			break;
1248 
1249 		rb_erase(node, &sched->atom_root);
1250 		data = rb_entry(node, struct work_atoms, node);
1251 		__thread_latency_insert(&sched->sorted_atom_root, data, &sched->sort_list);
1252 	}
1253 }
1254 
1255 static int process_sched_wakeup_event(struct perf_tool *tool,
1256 				      struct perf_evsel *evsel,
1257 				      struct perf_sample *sample,
1258 				      struct machine *machine)
1259 {
1260 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1261 
1262 	if (sched->tp_handler->wakeup_event)
1263 		return sched->tp_handler->wakeup_event(sched, evsel, sample, machine);
1264 
1265 	return 0;
1266 }
1267 
1268 static int map_switch_event(struct perf_sched *sched, struct perf_evsel *evsel,
1269 			    struct perf_sample *sample, struct machine *machine)
1270 {
1271 	const u32 next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1272 	struct thread *sched_in;
1273 	int new_shortname;
1274 	u64 timestamp0, timestamp = sample->time;
1275 	s64 delta;
1276 	int cpu, this_cpu = sample->cpu;
1277 
1278 	BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1279 
1280 	if (this_cpu > sched->max_cpu)
1281 		sched->max_cpu = this_cpu;
1282 
1283 	timestamp0 = sched->cpu_last_switched[this_cpu];
1284 	sched->cpu_last_switched[this_cpu] = timestamp;
1285 	if (timestamp0)
1286 		delta = timestamp - timestamp0;
1287 	else
1288 		delta = 0;
1289 
1290 	if (delta < 0) {
1291 		pr_err("hm, delta: %" PRIu64 " < 0 ?\n", delta);
1292 		return -1;
1293 	}
1294 
1295 	sched_in = machine__findnew_thread(machine, -1, next_pid);
1296 
1297 	sched->curr_thread[this_cpu] = sched_in;
1298 
1299 	printf("  ");
1300 
1301 	new_shortname = 0;
1302 	if (!sched_in->shortname[0]) {
1303 		if (!strcmp(thread__comm_str(sched_in), "swapper")) {
1304 			/*
1305 			 * Don't allocate a letter-number for swapper:0
1306 			 * as a shortname. Instead, we use '.' for it.
1307 			 */
1308 			sched_in->shortname[0] = '.';
1309 			sched_in->shortname[1] = ' ';
1310 		} else {
1311 			sched_in->shortname[0] = sched->next_shortname1;
1312 			sched_in->shortname[1] = sched->next_shortname2;
1313 
1314 			if (sched->next_shortname1 < 'Z') {
1315 				sched->next_shortname1++;
1316 			} else {
1317 				sched->next_shortname1 = 'A';
1318 				if (sched->next_shortname2 < '9')
1319 					sched->next_shortname2++;
1320 				else
1321 					sched->next_shortname2 = '0';
1322 			}
1323 		}
1324 		new_shortname = 1;
1325 	}
1326 
1327 	for (cpu = 0; cpu <= sched->max_cpu; cpu++) {
1328 		if (cpu != this_cpu)
1329 			printf(" ");
1330 		else
1331 			printf("*");
1332 
1333 		if (sched->curr_thread[cpu])
1334 			printf("%2s ", sched->curr_thread[cpu]->shortname);
1335 		else
1336 			printf("   ");
1337 	}
1338 
1339 	printf("  %12.6f secs ", (double)timestamp/1e9);
1340 	if (new_shortname) {
1341 		printf("%s => %s:%d\n",
1342 		       sched_in->shortname, thread__comm_str(sched_in), sched_in->tid);
1343 	} else {
1344 		printf("\n");
1345 	}
1346 
1347 	return 0;
1348 }
1349 
1350 static int process_sched_switch_event(struct perf_tool *tool,
1351 				      struct perf_evsel *evsel,
1352 				      struct perf_sample *sample,
1353 				      struct machine *machine)
1354 {
1355 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1356 	int this_cpu = sample->cpu, err = 0;
1357 	u32 prev_pid = perf_evsel__intval(evsel, sample, "prev_pid"),
1358 	    next_pid = perf_evsel__intval(evsel, sample, "next_pid");
1359 
1360 	if (sched->curr_pid[this_cpu] != (u32)-1) {
1361 		/*
1362 		 * Are we trying to switch away a PID that is
1363 		 * not current?
1364 		 */
1365 		if (sched->curr_pid[this_cpu] != prev_pid)
1366 			sched->nr_context_switch_bugs++;
1367 	}
1368 
1369 	if (sched->tp_handler->switch_event)
1370 		err = sched->tp_handler->switch_event(sched, evsel, sample, machine);
1371 
1372 	sched->curr_pid[this_cpu] = next_pid;
1373 	return err;
1374 }
1375 
1376 static int process_sched_runtime_event(struct perf_tool *tool,
1377 				       struct perf_evsel *evsel,
1378 				       struct perf_sample *sample,
1379 				       struct machine *machine)
1380 {
1381 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1382 
1383 	if (sched->tp_handler->runtime_event)
1384 		return sched->tp_handler->runtime_event(sched, evsel, sample, machine);
1385 
1386 	return 0;
1387 }
1388 
1389 static int perf_sched__process_fork_event(struct perf_tool *tool,
1390 					  union perf_event *event,
1391 					  struct perf_sample *sample,
1392 					  struct machine *machine)
1393 {
1394 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1395 
1396 	/* run the fork event through the perf machineruy */
1397 	perf_event__process_fork(tool, event, sample, machine);
1398 
1399 	/* and then run additional processing needed for this command */
1400 	if (sched->tp_handler->fork_event)
1401 		return sched->tp_handler->fork_event(sched, event, machine);
1402 
1403 	return 0;
1404 }
1405 
1406 static int process_sched_migrate_task_event(struct perf_tool *tool,
1407 					    struct perf_evsel *evsel,
1408 					    struct perf_sample *sample,
1409 					    struct machine *machine)
1410 {
1411 	struct perf_sched *sched = container_of(tool, struct perf_sched, tool);
1412 
1413 	if (sched->tp_handler->migrate_task_event)
1414 		return sched->tp_handler->migrate_task_event(sched, evsel, sample, machine);
1415 
1416 	return 0;
1417 }
1418 
1419 typedef int (*tracepoint_handler)(struct perf_tool *tool,
1420 				  struct perf_evsel *evsel,
1421 				  struct perf_sample *sample,
1422 				  struct machine *machine);
1423 
1424 static int perf_sched__process_tracepoint_sample(struct perf_tool *tool __maybe_unused,
1425 						 union perf_event *event __maybe_unused,
1426 						 struct perf_sample *sample,
1427 						 struct perf_evsel *evsel,
1428 						 struct machine *machine)
1429 {
1430 	int err = 0;
1431 
1432 	evsel->hists.stats.total_period += sample->period;
1433 	hists__inc_nr_samples(&evsel->hists, true);
1434 
1435 	if (evsel->handler != NULL) {
1436 		tracepoint_handler f = evsel->handler;
1437 		err = f(tool, evsel, sample, machine);
1438 	}
1439 
1440 	return err;
1441 }
1442 
1443 static int perf_sched__read_events(struct perf_sched *sched,
1444 				   struct perf_session **psession)
1445 {
1446 	const struct perf_evsel_str_handler handlers[] = {
1447 		{ "sched:sched_switch",	      process_sched_switch_event, },
1448 		{ "sched:sched_stat_runtime", process_sched_runtime_event, },
1449 		{ "sched:sched_wakeup",	      process_sched_wakeup_event, },
1450 		{ "sched:sched_wakeup_new",   process_sched_wakeup_event, },
1451 		{ "sched:sched_migrate_task", process_sched_migrate_task_event, },
1452 	};
1453 	struct perf_session *session;
1454 	struct perf_data_file file = {
1455 		.path = input_name,
1456 		.mode = PERF_DATA_MODE_READ,
1457 	};
1458 
1459 	session = perf_session__new(&file, false, &sched->tool);
1460 	if (session == NULL) {
1461 		pr_debug("No Memory for session\n");
1462 		return -1;
1463 	}
1464 
1465 	if (perf_session__set_tracepoints_handlers(session, handlers))
1466 		goto out_delete;
1467 
1468 	if (perf_session__has_traces(session, "record -R")) {
1469 		int err = perf_session__process_events(session, &sched->tool);
1470 		if (err) {
1471 			pr_err("Failed to process events, error %d", err);
1472 			goto out_delete;
1473 		}
1474 
1475 		sched->nr_events      = session->stats.nr_events[0];
1476 		sched->nr_lost_events = session->stats.total_lost;
1477 		sched->nr_lost_chunks = session->stats.nr_events[PERF_RECORD_LOST];
1478 	}
1479 
1480 	if (psession)
1481 		*psession = session;
1482 	else
1483 		perf_session__delete(session);
1484 
1485 	return 0;
1486 
1487 out_delete:
1488 	perf_session__delete(session);
1489 	return -1;
1490 }
1491 
1492 static void print_bad_events(struct perf_sched *sched)
1493 {
1494 	if (sched->nr_unordered_timestamps && sched->nr_timestamps) {
1495 		printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1496 			(double)sched->nr_unordered_timestamps/(double)sched->nr_timestamps*100.0,
1497 			sched->nr_unordered_timestamps, sched->nr_timestamps);
1498 	}
1499 	if (sched->nr_lost_events && sched->nr_events) {
1500 		printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1501 			(double)sched->nr_lost_events/(double)sched->nr_events * 100.0,
1502 			sched->nr_lost_events, sched->nr_events, sched->nr_lost_chunks);
1503 	}
1504 	if (sched->nr_context_switch_bugs && sched->nr_timestamps) {
1505 		printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1506 			(double)sched->nr_context_switch_bugs/(double)sched->nr_timestamps*100.0,
1507 			sched->nr_context_switch_bugs, sched->nr_timestamps);
1508 		if (sched->nr_lost_events)
1509 			printf(" (due to lost events?)");
1510 		printf("\n");
1511 	}
1512 }
1513 
1514 static int perf_sched__lat(struct perf_sched *sched)
1515 {
1516 	struct rb_node *next;
1517 	struct perf_session *session;
1518 
1519 	setup_pager();
1520 
1521 	/* save session -- references to threads are held in work_list */
1522 	if (perf_sched__read_events(sched, &session))
1523 		return -1;
1524 
1525 	perf_sched__sort_lat(sched);
1526 
1527 	printf("\n -----------------------------------------------------------------------------------------------------------------\n");
1528 	printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at       |\n");
1529 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
1530 
1531 	next = rb_first(&sched->sorted_atom_root);
1532 
1533 	while (next) {
1534 		struct work_atoms *work_list;
1535 
1536 		work_list = rb_entry(next, struct work_atoms, node);
1537 		output_lat_thread(sched, work_list);
1538 		next = rb_next(next);
1539 	}
1540 
1541 	printf(" -----------------------------------------------------------------------------------------------------------------\n");
1542 	printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
1543 		(double)sched->all_runtime / 1e6, sched->all_count);
1544 
1545 	printf(" ---------------------------------------------------\n");
1546 
1547 	print_bad_events(sched);
1548 	printf("\n");
1549 
1550 	perf_session__delete(session);
1551 	return 0;
1552 }
1553 
1554 static int perf_sched__map(struct perf_sched *sched)
1555 {
1556 	sched->max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1557 
1558 	setup_pager();
1559 	if (perf_sched__read_events(sched, NULL))
1560 		return -1;
1561 	print_bad_events(sched);
1562 	return 0;
1563 }
1564 
1565 static int perf_sched__replay(struct perf_sched *sched)
1566 {
1567 	unsigned long i;
1568 
1569 	calibrate_run_measurement_overhead(sched);
1570 	calibrate_sleep_measurement_overhead(sched);
1571 
1572 	test_calibrations(sched);
1573 
1574 	if (perf_sched__read_events(sched, NULL))
1575 		return -1;
1576 
1577 	printf("nr_run_events:        %ld\n", sched->nr_run_events);
1578 	printf("nr_sleep_events:      %ld\n", sched->nr_sleep_events);
1579 	printf("nr_wakeup_events:     %ld\n", sched->nr_wakeup_events);
1580 
1581 	if (sched->targetless_wakeups)
1582 		printf("target-less wakeups:  %ld\n", sched->targetless_wakeups);
1583 	if (sched->multitarget_wakeups)
1584 		printf("multi-target wakeups: %ld\n", sched->multitarget_wakeups);
1585 	if (sched->nr_run_events_optimized)
1586 		printf("run atoms optimized: %ld\n",
1587 			sched->nr_run_events_optimized);
1588 
1589 	print_task_traces(sched);
1590 	add_cross_task_wakeups(sched);
1591 
1592 	create_tasks(sched);
1593 	printf("------------------------------------------------------------\n");
1594 	for (i = 0; i < sched->replay_repeat; i++)
1595 		run_one_test(sched);
1596 
1597 	return 0;
1598 }
1599 
1600 static void setup_sorting(struct perf_sched *sched, const struct option *options,
1601 			  const char * const usage_msg[])
1602 {
1603 	char *tmp, *tok, *str = strdup(sched->sort_order);
1604 
1605 	for (tok = strtok_r(str, ", ", &tmp);
1606 			tok; tok = strtok_r(NULL, ", ", &tmp)) {
1607 		if (sort_dimension__add(tok, &sched->sort_list) < 0) {
1608 			error("Unknown --sort key: `%s'", tok);
1609 			usage_with_options(usage_msg, options);
1610 		}
1611 	}
1612 
1613 	free(str);
1614 
1615 	sort_dimension__add("pid", &sched->cmp_pid);
1616 }
1617 
1618 static int __cmd_record(int argc, const char **argv)
1619 {
1620 	unsigned int rec_argc, i, j;
1621 	const char **rec_argv;
1622 	const char * const record_args[] = {
1623 		"record",
1624 		"-a",
1625 		"-R",
1626 		"-m", "1024",
1627 		"-c", "1",
1628 		"-e", "sched:sched_switch",
1629 		"-e", "sched:sched_stat_wait",
1630 		"-e", "sched:sched_stat_sleep",
1631 		"-e", "sched:sched_stat_iowait",
1632 		"-e", "sched:sched_stat_runtime",
1633 		"-e", "sched:sched_process_fork",
1634 		"-e", "sched:sched_wakeup",
1635 		"-e", "sched:sched_wakeup_new",
1636 		"-e", "sched:sched_migrate_task",
1637 	};
1638 
1639 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1640 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1641 
1642 	if (rec_argv == NULL)
1643 		return -ENOMEM;
1644 
1645 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1646 		rec_argv[i] = strdup(record_args[i]);
1647 
1648 	for (j = 1; j < (unsigned int)argc; j++, i++)
1649 		rec_argv[i] = argv[j];
1650 
1651 	BUG_ON(i != rec_argc);
1652 
1653 	return cmd_record(i, rec_argv, NULL);
1654 }
1655 
1656 int cmd_sched(int argc, const char **argv, const char *prefix __maybe_unused)
1657 {
1658 	const char default_sort_order[] = "avg, max, switch, runtime";
1659 	struct perf_sched sched = {
1660 		.tool = {
1661 			.sample		 = perf_sched__process_tracepoint_sample,
1662 			.comm		 = perf_event__process_comm,
1663 			.lost		 = perf_event__process_lost,
1664 			.fork		 = perf_sched__process_fork_event,
1665 			.ordered_samples = true,
1666 		},
1667 		.cmp_pid	      = LIST_HEAD_INIT(sched.cmp_pid),
1668 		.sort_list	      = LIST_HEAD_INIT(sched.sort_list),
1669 		.start_work_mutex     = PTHREAD_MUTEX_INITIALIZER,
1670 		.work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER,
1671 		.sort_order	      = default_sort_order,
1672 		.replay_repeat	      = 10,
1673 		.profile_cpu	      = -1,
1674 		.next_shortname1      = 'A',
1675 		.next_shortname2      = '0',
1676 	};
1677 	const struct option latency_options[] = {
1678 	OPT_STRING('s', "sort", &sched.sort_order, "key[,key2...]",
1679 		   "sort by key(s): runtime, switch, avg, max"),
1680 	OPT_INCR('v', "verbose", &verbose,
1681 		    "be more verbose (show symbol address, etc)"),
1682 	OPT_INTEGER('C', "CPU", &sched.profile_cpu,
1683 		    "CPU to profile on"),
1684 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1685 		    "dump raw trace in ASCII"),
1686 	OPT_END()
1687 	};
1688 	const struct option replay_options[] = {
1689 	OPT_UINTEGER('r', "repeat", &sched.replay_repeat,
1690 		     "repeat the workload replay N times (-1: infinite)"),
1691 	OPT_INCR('v', "verbose", &verbose,
1692 		    "be more verbose (show symbol address, etc)"),
1693 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1694 		    "dump raw trace in ASCII"),
1695 	OPT_END()
1696 	};
1697 	const struct option sched_options[] = {
1698 	OPT_STRING('i', "input", &input_name, "file",
1699 		    "input file name"),
1700 	OPT_INCR('v', "verbose", &verbose,
1701 		    "be more verbose (show symbol address, etc)"),
1702 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1703 		    "dump raw trace in ASCII"),
1704 	OPT_END()
1705 	};
1706 	const char * const latency_usage[] = {
1707 		"perf sched latency [<options>]",
1708 		NULL
1709 	};
1710 	const char * const replay_usage[] = {
1711 		"perf sched replay [<options>]",
1712 		NULL
1713 	};
1714 	const char *const sched_subcommands[] = { "record", "latency", "map",
1715 						  "replay", "script", NULL };
1716 	const char *sched_usage[] = {
1717 		NULL,
1718 		NULL
1719 	};
1720 	struct trace_sched_handler lat_ops  = {
1721 		.wakeup_event	    = latency_wakeup_event,
1722 		.switch_event	    = latency_switch_event,
1723 		.runtime_event	    = latency_runtime_event,
1724 		.migrate_task_event = latency_migrate_task_event,
1725 	};
1726 	struct trace_sched_handler map_ops  = {
1727 		.switch_event	    = map_switch_event,
1728 	};
1729 	struct trace_sched_handler replay_ops  = {
1730 		.wakeup_event	    = replay_wakeup_event,
1731 		.switch_event	    = replay_switch_event,
1732 		.fork_event	    = replay_fork_event,
1733 	};
1734 	unsigned int i;
1735 
1736 	for (i = 0; i < ARRAY_SIZE(sched.curr_pid); i++)
1737 		sched.curr_pid[i] = -1;
1738 
1739 	argc = parse_options_subcommand(argc, argv, sched_options, sched_subcommands,
1740 					sched_usage, PARSE_OPT_STOP_AT_NON_OPTION);
1741 	if (!argc)
1742 		usage_with_options(sched_usage, sched_options);
1743 
1744 	/*
1745 	 * Aliased to 'perf script' for now:
1746 	 */
1747 	if (!strcmp(argv[0], "script"))
1748 		return cmd_script(argc, argv, prefix);
1749 
1750 	symbol__init();
1751 	if (!strncmp(argv[0], "rec", 3)) {
1752 		return __cmd_record(argc, argv);
1753 	} else if (!strncmp(argv[0], "lat", 3)) {
1754 		sched.tp_handler = &lat_ops;
1755 		if (argc > 1) {
1756 			argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1757 			if (argc)
1758 				usage_with_options(latency_usage, latency_options);
1759 		}
1760 		setup_sorting(&sched, latency_options, latency_usage);
1761 		return perf_sched__lat(&sched);
1762 	} else if (!strcmp(argv[0], "map")) {
1763 		sched.tp_handler = &map_ops;
1764 		setup_sorting(&sched, latency_options, latency_usage);
1765 		return perf_sched__map(&sched);
1766 	} else if (!strncmp(argv[0], "rep", 3)) {
1767 		sched.tp_handler = &replay_ops;
1768 		if (argc) {
1769 			argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1770 			if (argc)
1771 				usage_with_options(replay_usage, replay_options);
1772 		}
1773 		return perf_sched__replay(&sched);
1774 	} else {
1775 		usage_with_options(sched_usage, sched_options);
1776 	}
1777 
1778 	return 0;
1779 }
1780