xref: /openbmc/linux/tools/perf/util/evlist.c (revision fe1f61b37ffada9fc7ec2c9d4ca5376b5a797dbc)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4  *
5  * Parts came from builtin-{top,stat,record}.c, see those files for further
6  * copyright notes.
7  */
8 #include <api/fs/fs.h>
9 #include <errno.h>
10 #include <inttypes.h>
11 #include <poll.h>
12 #include "cpumap.h"
13 #include "thread_map.h"
14 #include "target.h"
15 #include "evlist.h"
16 #include "evsel.h"
17 #include "debug.h"
18 #include "units.h"
19 #include "asm/bug.h"
20 #include "bpf-event.h"
21 #include <signal.h>
22 #include <unistd.h>
23 
24 #include "parse-events.h"
25 #include <subcmd/parse-options.h>
26 
27 #include <fcntl.h>
28 #include <sys/ioctl.h>
29 #include <sys/mman.h>
30 
31 #include <linux/bitops.h>
32 #include <linux/hash.h>
33 #include <linux/log2.h>
34 #include <linux/err.h>
35 #include <linux/zalloc.h>
36 #include <perf/evlist.h>
37 #include <perf/cpumap.h>
38 
39 #ifdef LACKS_SIGQUEUE_PROTOTYPE
40 int sigqueue(pid_t pid, int sig, const union sigval value);
41 #endif
42 
43 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
44 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
45 
46 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
47 		  struct perf_thread_map *threads)
48 {
49 	int i;
50 
51 	for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
52 		INIT_HLIST_HEAD(&evlist->heads[i]);
53 	perf_evlist__init(&evlist->core);
54 	perf_evlist__set_maps(evlist, cpus, threads);
55 	fdarray__init(&evlist->pollfd, 64);
56 	evlist->workload.pid = -1;
57 	evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
58 }
59 
60 struct evlist *evlist__new(void)
61 {
62 	struct evlist *evlist = zalloc(sizeof(*evlist));
63 
64 	if (evlist != NULL)
65 		evlist__init(evlist, NULL, NULL);
66 
67 	return evlist;
68 }
69 
70 struct evlist *perf_evlist__new_default(void)
71 {
72 	struct evlist *evlist = evlist__new();
73 
74 	if (evlist && perf_evlist__add_default(evlist)) {
75 		evlist__delete(evlist);
76 		evlist = NULL;
77 	}
78 
79 	return evlist;
80 }
81 
82 struct evlist *perf_evlist__new_dummy(void)
83 {
84 	struct evlist *evlist = evlist__new();
85 
86 	if (evlist && perf_evlist__add_dummy(evlist)) {
87 		evlist__delete(evlist);
88 		evlist = NULL;
89 	}
90 
91 	return evlist;
92 }
93 
94 /**
95  * perf_evlist__set_id_pos - set the positions of event ids.
96  * @evlist: selected event list
97  *
98  * Events with compatible sample types all have the same id_pos
99  * and is_pos.  For convenience, put a copy on evlist.
100  */
101 void perf_evlist__set_id_pos(struct evlist *evlist)
102 {
103 	struct evsel *first = perf_evlist__first(evlist);
104 
105 	evlist->id_pos = first->id_pos;
106 	evlist->is_pos = first->is_pos;
107 }
108 
109 static void perf_evlist__update_id_pos(struct evlist *evlist)
110 {
111 	struct evsel *evsel;
112 
113 	evlist__for_each_entry(evlist, evsel)
114 		perf_evsel__calc_id_pos(evsel);
115 
116 	perf_evlist__set_id_pos(evlist);
117 }
118 
119 static void perf_evlist__purge(struct evlist *evlist)
120 {
121 	struct evsel *pos, *n;
122 
123 	evlist__for_each_entry_safe(evlist, n, pos) {
124 		list_del_init(&pos->core.node);
125 		pos->evlist = NULL;
126 		evsel__delete(pos);
127 	}
128 
129 	evlist->core.nr_entries = 0;
130 }
131 
132 void perf_evlist__exit(struct evlist *evlist)
133 {
134 	zfree(&evlist->mmap);
135 	zfree(&evlist->overwrite_mmap);
136 	fdarray__exit(&evlist->pollfd);
137 }
138 
139 void evlist__delete(struct evlist *evlist)
140 {
141 	if (evlist == NULL)
142 		return;
143 
144 	perf_evlist__munmap(evlist);
145 	evlist__close(evlist);
146 	perf_cpu_map__put(evlist->cpus);
147 	perf_thread_map__put(evlist->threads);
148 	evlist->cpus = NULL;
149 	evlist->threads = NULL;
150 	perf_evlist__purge(evlist);
151 	perf_evlist__exit(evlist);
152 	free(evlist);
153 }
154 
155 static void __perf_evlist__propagate_maps(struct evlist *evlist,
156 					  struct evsel *evsel)
157 {
158 	/*
159 	 * We already have cpus for evsel (via PMU sysfs) so
160 	 * keep it, if there's no target cpu list defined.
161 	 */
162 	if (!evsel->core.own_cpus || evlist->has_user_cpus) {
163 		perf_cpu_map__put(evsel->core.cpus);
164 		evsel->core.cpus = perf_cpu_map__get(evlist->cpus);
165 	} else if (evsel->core.cpus != evsel->core.own_cpus) {
166 		perf_cpu_map__put(evsel->core.cpus);
167 		evsel->core.cpus = perf_cpu_map__get(evsel->core.own_cpus);
168 	}
169 
170 	perf_thread_map__put(evsel->threads);
171 	evsel->threads = perf_thread_map__get(evlist->threads);
172 }
173 
174 static void perf_evlist__propagate_maps(struct evlist *evlist)
175 {
176 	struct evsel *evsel;
177 
178 	evlist__for_each_entry(evlist, evsel)
179 		__perf_evlist__propagate_maps(evlist, evsel);
180 }
181 
182 void evlist__add(struct evlist *evlist, struct evsel *entry)
183 {
184 	entry->evlist = evlist;
185 	entry->idx = evlist->core.nr_entries;
186 	entry->tracking = !entry->idx;
187 
188 	perf_evlist__add(&evlist->core, &entry->core);
189 
190 	if (evlist->core.nr_entries == 1)
191 		perf_evlist__set_id_pos(evlist);
192 
193 	__perf_evlist__propagate_maps(evlist, entry);
194 }
195 
196 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
197 {
198 	evsel->evlist = NULL;
199 	perf_evlist__remove(&evlist->core, &evsel->core);
200 }
201 
202 void perf_evlist__splice_list_tail(struct evlist *evlist,
203 				   struct list_head *list)
204 {
205 	struct evsel *evsel, *temp;
206 
207 	__evlist__for_each_entry_safe(list, temp, evsel) {
208 		list_del_init(&evsel->core.node);
209 		evlist__add(evlist, evsel);
210 	}
211 }
212 
213 void __perf_evlist__set_leader(struct list_head *list)
214 {
215 	struct evsel *evsel, *leader;
216 
217 	leader = list_entry(list->next, struct evsel, core.node);
218 	evsel = list_entry(list->prev, struct evsel, core.node);
219 
220 	leader->nr_members = evsel->idx - leader->idx + 1;
221 
222 	__evlist__for_each_entry(list, evsel) {
223 		evsel->leader = leader;
224 	}
225 }
226 
227 void perf_evlist__set_leader(struct evlist *evlist)
228 {
229 	if (evlist->core.nr_entries) {
230 		evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
231 		__perf_evlist__set_leader(&evlist->core.entries);
232 	}
233 }
234 
235 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
236 {
237 	struct evsel *evsel = perf_evsel__new_cycles(precise);
238 
239 	if (evsel == NULL)
240 		return -ENOMEM;
241 
242 	evlist__add(evlist, evsel);
243 	return 0;
244 }
245 
246 int perf_evlist__add_dummy(struct evlist *evlist)
247 {
248 	struct perf_event_attr attr = {
249 		.type	= PERF_TYPE_SOFTWARE,
250 		.config = PERF_COUNT_SW_DUMMY,
251 		.size	= sizeof(attr), /* to capture ABI version */
252 	};
253 	struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
254 
255 	if (evsel == NULL)
256 		return -ENOMEM;
257 
258 	evlist__add(evlist, evsel);
259 	return 0;
260 }
261 
262 static int evlist__add_attrs(struct evlist *evlist,
263 				  struct perf_event_attr *attrs, size_t nr_attrs)
264 {
265 	struct evsel *evsel, *n;
266 	LIST_HEAD(head);
267 	size_t i;
268 
269 	for (i = 0; i < nr_attrs; i++) {
270 		evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
271 		if (evsel == NULL)
272 			goto out_delete_partial_list;
273 		list_add_tail(&evsel->core.node, &head);
274 	}
275 
276 	perf_evlist__splice_list_tail(evlist, &head);
277 
278 	return 0;
279 
280 out_delete_partial_list:
281 	__evlist__for_each_entry_safe(&head, n, evsel)
282 		evsel__delete(evsel);
283 	return -1;
284 }
285 
286 int __perf_evlist__add_default_attrs(struct evlist *evlist,
287 				     struct perf_event_attr *attrs, size_t nr_attrs)
288 {
289 	size_t i;
290 
291 	for (i = 0; i < nr_attrs; i++)
292 		event_attr_init(attrs + i);
293 
294 	return evlist__add_attrs(evlist, attrs, nr_attrs);
295 }
296 
297 struct evsel *
298 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
299 {
300 	struct evsel *evsel;
301 
302 	evlist__for_each_entry(evlist, evsel) {
303 		if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
304 		    (int)evsel->core.attr.config == id)
305 			return evsel;
306 	}
307 
308 	return NULL;
309 }
310 
311 struct evsel *
312 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
313 				     const char *name)
314 {
315 	struct evsel *evsel;
316 
317 	evlist__for_each_entry(evlist, evsel) {
318 		if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
319 		    (strcmp(evsel->name, name) == 0))
320 			return evsel;
321 	}
322 
323 	return NULL;
324 }
325 
326 int perf_evlist__add_newtp(struct evlist *evlist,
327 			   const char *sys, const char *name, void *handler)
328 {
329 	struct evsel *evsel = perf_evsel__newtp(sys, name);
330 
331 	if (IS_ERR(evsel))
332 		return -1;
333 
334 	evsel->handler = handler;
335 	evlist__add(evlist, evsel);
336 	return 0;
337 }
338 
339 static int perf_evlist__nr_threads(struct evlist *evlist,
340 				   struct evsel *evsel)
341 {
342 	if (evsel->system_wide)
343 		return 1;
344 	else
345 		return thread_map__nr(evlist->threads);
346 }
347 
348 void evlist__disable(struct evlist *evlist)
349 {
350 	struct evsel *pos;
351 
352 	evlist__for_each_entry(evlist, pos) {
353 		if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
354 			continue;
355 		evsel__disable(pos);
356 	}
357 
358 	evlist->enabled = false;
359 }
360 
361 void evlist__enable(struct evlist *evlist)
362 {
363 	struct evsel *pos;
364 
365 	evlist__for_each_entry(evlist, pos) {
366 		if (!perf_evsel__is_group_leader(pos) || !pos->fd)
367 			continue;
368 		evsel__enable(pos);
369 	}
370 
371 	evlist->enabled = true;
372 }
373 
374 void perf_evlist__toggle_enable(struct evlist *evlist)
375 {
376 	(evlist->enabled ? evlist__disable : evlist__enable)(evlist);
377 }
378 
379 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
380 					 struct evsel *evsel, int cpu)
381 {
382 	int thread;
383 	int nr_threads = perf_evlist__nr_threads(evlist, evsel);
384 
385 	if (!evsel->fd)
386 		return -EINVAL;
387 
388 	for (thread = 0; thread < nr_threads; thread++) {
389 		int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
390 		if (err)
391 			return err;
392 	}
393 	return 0;
394 }
395 
396 static int perf_evlist__enable_event_thread(struct evlist *evlist,
397 					    struct evsel *evsel,
398 					    int thread)
399 {
400 	int cpu;
401 	int nr_cpus = cpu_map__nr(evlist->cpus);
402 
403 	if (!evsel->fd)
404 		return -EINVAL;
405 
406 	for (cpu = 0; cpu < nr_cpus; cpu++) {
407 		int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
408 		if (err)
409 			return err;
410 	}
411 	return 0;
412 }
413 
414 int perf_evlist__enable_event_idx(struct evlist *evlist,
415 				  struct evsel *evsel, int idx)
416 {
417 	bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
418 
419 	if (per_cpu_mmaps)
420 		return perf_evlist__enable_event_cpu(evlist, evsel, idx);
421 	else
422 		return perf_evlist__enable_event_thread(evlist, evsel, idx);
423 }
424 
425 int perf_evlist__alloc_pollfd(struct evlist *evlist)
426 {
427 	int nr_cpus = cpu_map__nr(evlist->cpus);
428 	int nr_threads = thread_map__nr(evlist->threads);
429 	int nfds = 0;
430 	struct evsel *evsel;
431 
432 	evlist__for_each_entry(evlist, evsel) {
433 		if (evsel->system_wide)
434 			nfds += nr_cpus;
435 		else
436 			nfds += nr_cpus * nr_threads;
437 	}
438 
439 	if (fdarray__available_entries(&evlist->pollfd) < nfds &&
440 	    fdarray__grow(&evlist->pollfd, nfds) < 0)
441 		return -ENOMEM;
442 
443 	return 0;
444 }
445 
446 static int __perf_evlist__add_pollfd(struct evlist *evlist, int fd,
447 				     struct perf_mmap *map, short revent)
448 {
449 	int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
450 	/*
451 	 * Save the idx so that when we filter out fds POLLHUP'ed we can
452 	 * close the associated evlist->mmap[] entry.
453 	 */
454 	if (pos >= 0) {
455 		evlist->pollfd.priv[pos].ptr = map;
456 
457 		fcntl(fd, F_SETFL, O_NONBLOCK);
458 	}
459 
460 	return pos;
461 }
462 
463 int perf_evlist__add_pollfd(struct evlist *evlist, int fd)
464 {
465 	return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
466 }
467 
468 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
469 					 void *arg __maybe_unused)
470 {
471 	struct perf_mmap *map = fda->priv[fd].ptr;
472 
473 	if (map)
474 		perf_mmap__put(map);
475 }
476 
477 int perf_evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
478 {
479 	return fdarray__filter(&evlist->pollfd, revents_and_mask,
480 			       perf_evlist__munmap_filtered, NULL);
481 }
482 
483 int perf_evlist__poll(struct evlist *evlist, int timeout)
484 {
485 	return fdarray__poll(&evlist->pollfd, timeout);
486 }
487 
488 static void perf_evlist__id_hash(struct evlist *evlist,
489 				 struct evsel *evsel,
490 				 int cpu, int thread, u64 id)
491 {
492 	int hash;
493 	struct perf_sample_id *sid = SID(evsel, cpu, thread);
494 
495 	sid->id = id;
496 	sid->evsel = evsel;
497 	hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
498 	hlist_add_head(&sid->node, &evlist->heads[hash]);
499 }
500 
501 void perf_evlist__id_add(struct evlist *evlist, struct evsel *evsel,
502 			 int cpu, int thread, u64 id)
503 {
504 	perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
505 	evsel->id[evsel->ids++] = id;
506 }
507 
508 int perf_evlist__id_add_fd(struct evlist *evlist,
509 			   struct evsel *evsel,
510 			   int cpu, int thread, int fd)
511 {
512 	u64 read_data[4] = { 0, };
513 	int id_idx = 1; /* The first entry is the counter value */
514 	u64 id;
515 	int ret;
516 
517 	ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
518 	if (!ret)
519 		goto add;
520 
521 	if (errno != ENOTTY)
522 		return -1;
523 
524 	/* Legacy way to get event id.. All hail to old kernels! */
525 
526 	/*
527 	 * This way does not work with group format read, so bail
528 	 * out in that case.
529 	 */
530 	if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
531 		return -1;
532 
533 	if (!(evsel->core.attr.read_format & PERF_FORMAT_ID) ||
534 	    read(fd, &read_data, sizeof(read_data)) == -1)
535 		return -1;
536 
537 	if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
538 		++id_idx;
539 	if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
540 		++id_idx;
541 
542 	id = read_data[id_idx];
543 
544  add:
545 	perf_evlist__id_add(evlist, evsel, cpu, thread, id);
546 	return 0;
547 }
548 
549 static void perf_evlist__set_sid_idx(struct evlist *evlist,
550 				     struct evsel *evsel, int idx, int cpu,
551 				     int thread)
552 {
553 	struct perf_sample_id *sid = SID(evsel, cpu, thread);
554 	sid->idx = idx;
555 	if (evlist->cpus && cpu >= 0)
556 		sid->cpu = evlist->cpus->map[cpu];
557 	else
558 		sid->cpu = -1;
559 	if (!evsel->system_wide && evlist->threads && thread >= 0)
560 		sid->tid = thread_map__pid(evlist->threads, thread);
561 	else
562 		sid->tid = -1;
563 }
564 
565 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
566 {
567 	struct hlist_head *head;
568 	struct perf_sample_id *sid;
569 	int hash;
570 
571 	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
572 	head = &evlist->heads[hash];
573 
574 	hlist_for_each_entry(sid, head, node)
575 		if (sid->id == id)
576 			return sid;
577 
578 	return NULL;
579 }
580 
581 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
582 {
583 	struct perf_sample_id *sid;
584 
585 	if (evlist->core.nr_entries == 1 || !id)
586 		return perf_evlist__first(evlist);
587 
588 	sid = perf_evlist__id2sid(evlist, id);
589 	if (sid)
590 		return sid->evsel;
591 
592 	if (!perf_evlist__sample_id_all(evlist))
593 		return perf_evlist__first(evlist);
594 
595 	return NULL;
596 }
597 
598 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
599 						u64 id)
600 {
601 	struct perf_sample_id *sid;
602 
603 	if (!id)
604 		return NULL;
605 
606 	sid = perf_evlist__id2sid(evlist, id);
607 	if (sid)
608 		return sid->evsel;
609 
610 	return NULL;
611 }
612 
613 static int perf_evlist__event2id(struct evlist *evlist,
614 				 union perf_event *event, u64 *id)
615 {
616 	const u64 *array = event->sample.array;
617 	ssize_t n;
618 
619 	n = (event->header.size - sizeof(event->header)) >> 3;
620 
621 	if (event->header.type == PERF_RECORD_SAMPLE) {
622 		if (evlist->id_pos >= n)
623 			return -1;
624 		*id = array[evlist->id_pos];
625 	} else {
626 		if (evlist->is_pos > n)
627 			return -1;
628 		n -= evlist->is_pos;
629 		*id = array[n];
630 	}
631 	return 0;
632 }
633 
634 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
635 					    union perf_event *event)
636 {
637 	struct evsel *first = perf_evlist__first(evlist);
638 	struct hlist_head *head;
639 	struct perf_sample_id *sid;
640 	int hash;
641 	u64 id;
642 
643 	if (evlist->core.nr_entries == 1)
644 		return first;
645 
646 	if (!first->core.attr.sample_id_all &&
647 	    event->header.type != PERF_RECORD_SAMPLE)
648 		return first;
649 
650 	if (perf_evlist__event2id(evlist, event, &id))
651 		return NULL;
652 
653 	/* Synthesized events have an id of zero */
654 	if (!id)
655 		return first;
656 
657 	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
658 	head = &evlist->heads[hash];
659 
660 	hlist_for_each_entry(sid, head, node) {
661 		if (sid->id == id)
662 			return sid->evsel;
663 	}
664 	return NULL;
665 }
666 
667 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
668 {
669 	int i;
670 
671 	if (!evlist->overwrite_mmap)
672 		return 0;
673 
674 	for (i = 0; i < evlist->nr_mmaps; i++) {
675 		int fd = evlist->overwrite_mmap[i].fd;
676 		int err;
677 
678 		if (fd < 0)
679 			continue;
680 		err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
681 		if (err)
682 			return err;
683 	}
684 	return 0;
685 }
686 
687 static int perf_evlist__pause(struct evlist *evlist)
688 {
689 	return perf_evlist__set_paused(evlist, true);
690 }
691 
692 static int perf_evlist__resume(struct evlist *evlist)
693 {
694 	return perf_evlist__set_paused(evlist, false);
695 }
696 
697 static void perf_evlist__munmap_nofree(struct evlist *evlist)
698 {
699 	int i;
700 
701 	if (evlist->mmap)
702 		for (i = 0; i < evlist->nr_mmaps; i++)
703 			perf_mmap__munmap(&evlist->mmap[i]);
704 
705 	if (evlist->overwrite_mmap)
706 		for (i = 0; i < evlist->nr_mmaps; i++)
707 			perf_mmap__munmap(&evlist->overwrite_mmap[i]);
708 }
709 
710 void perf_evlist__munmap(struct evlist *evlist)
711 {
712 	perf_evlist__munmap_nofree(evlist);
713 	zfree(&evlist->mmap);
714 	zfree(&evlist->overwrite_mmap);
715 }
716 
717 static struct perf_mmap *perf_evlist__alloc_mmap(struct evlist *evlist,
718 						 bool overwrite)
719 {
720 	int i;
721 	struct perf_mmap *map;
722 
723 	evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
724 	if (cpu_map__empty(evlist->cpus))
725 		evlist->nr_mmaps = thread_map__nr(evlist->threads);
726 	map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
727 	if (!map)
728 		return NULL;
729 
730 	for (i = 0; i < evlist->nr_mmaps; i++) {
731 		map[i].fd = -1;
732 		map[i].overwrite = overwrite;
733 		/*
734 		 * When the perf_mmap() call is made we grab one refcount, plus
735 		 * one extra to let perf_mmap__consume() get the last
736 		 * events after all real references (perf_mmap__get()) are
737 		 * dropped.
738 		 *
739 		 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
740 		 * thus does perf_mmap__get() on it.
741 		 */
742 		refcount_set(&map[i].refcnt, 0);
743 	}
744 	return map;
745 }
746 
747 static bool
748 perf_evlist__should_poll(struct evlist *evlist __maybe_unused,
749 			 struct evsel *evsel)
750 {
751 	if (evsel->core.attr.write_backward)
752 		return false;
753 	return true;
754 }
755 
756 static int perf_evlist__mmap_per_evsel(struct evlist *evlist, int idx,
757 				       struct mmap_params *mp, int cpu_idx,
758 				       int thread, int *_output, int *_output_overwrite)
759 {
760 	struct evsel *evsel;
761 	int revent;
762 	int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
763 
764 	evlist__for_each_entry(evlist, evsel) {
765 		struct perf_mmap *maps = evlist->mmap;
766 		int *output = _output;
767 		int fd;
768 		int cpu;
769 
770 		mp->prot = PROT_READ | PROT_WRITE;
771 		if (evsel->core.attr.write_backward) {
772 			output = _output_overwrite;
773 			maps = evlist->overwrite_mmap;
774 
775 			if (!maps) {
776 				maps = perf_evlist__alloc_mmap(evlist, true);
777 				if (!maps)
778 					return -1;
779 				evlist->overwrite_mmap = maps;
780 				if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
781 					perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
782 			}
783 			mp->prot &= ~PROT_WRITE;
784 		}
785 
786 		if (evsel->system_wide && thread)
787 			continue;
788 
789 		cpu = cpu_map__idx(evsel->core.cpus, evlist_cpu);
790 		if (cpu == -1)
791 			continue;
792 
793 		fd = FD(evsel, cpu, thread);
794 
795 		if (*output == -1) {
796 			*output = fd;
797 
798 			if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
799 				return -1;
800 		} else {
801 			if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
802 				return -1;
803 
804 			perf_mmap__get(&maps[idx]);
805 		}
806 
807 		revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
808 
809 		/*
810 		 * The system_wide flag causes a selected event to be opened
811 		 * always without a pid.  Consequently it will never get a
812 		 * POLLHUP, but it is used for tracking in combination with
813 		 * other events, so it should not need to be polled anyway.
814 		 * Therefore don't add it for polling.
815 		 */
816 		if (!evsel->system_wide &&
817 		    __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
818 			perf_mmap__put(&maps[idx]);
819 			return -1;
820 		}
821 
822 		if (evsel->core.attr.read_format & PERF_FORMAT_ID) {
823 			if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
824 						   fd) < 0)
825 				return -1;
826 			perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
827 						 thread);
828 		}
829 	}
830 
831 	return 0;
832 }
833 
834 static int perf_evlist__mmap_per_cpu(struct evlist *evlist,
835 				     struct mmap_params *mp)
836 {
837 	int cpu, thread;
838 	int nr_cpus = cpu_map__nr(evlist->cpus);
839 	int nr_threads = thread_map__nr(evlist->threads);
840 
841 	pr_debug2("perf event ring buffer mmapped per cpu\n");
842 	for (cpu = 0; cpu < nr_cpus; cpu++) {
843 		int output = -1;
844 		int output_overwrite = -1;
845 
846 		auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
847 					      true);
848 
849 		for (thread = 0; thread < nr_threads; thread++) {
850 			if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
851 							thread, &output, &output_overwrite))
852 				goto out_unmap;
853 		}
854 	}
855 
856 	return 0;
857 
858 out_unmap:
859 	perf_evlist__munmap_nofree(evlist);
860 	return -1;
861 }
862 
863 static int perf_evlist__mmap_per_thread(struct evlist *evlist,
864 					struct mmap_params *mp)
865 {
866 	int thread;
867 	int nr_threads = thread_map__nr(evlist->threads);
868 
869 	pr_debug2("perf event ring buffer mmapped per thread\n");
870 	for (thread = 0; thread < nr_threads; thread++) {
871 		int output = -1;
872 		int output_overwrite = -1;
873 
874 		auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
875 					      false);
876 
877 		if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
878 						&output, &output_overwrite))
879 			goto out_unmap;
880 	}
881 
882 	return 0;
883 
884 out_unmap:
885 	perf_evlist__munmap_nofree(evlist);
886 	return -1;
887 }
888 
889 unsigned long perf_event_mlock_kb_in_pages(void)
890 {
891 	unsigned long pages;
892 	int max;
893 
894 	if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
895 		/*
896 		 * Pick a once upon a time good value, i.e. things look
897 		 * strange since we can't read a sysctl value, but lets not
898 		 * die yet...
899 		 */
900 		max = 512;
901 	} else {
902 		max -= (page_size / 1024);
903 	}
904 
905 	pages = (max * 1024) / page_size;
906 	if (!is_power_of_2(pages))
907 		pages = rounddown_pow_of_two(pages);
908 
909 	return pages;
910 }
911 
912 size_t perf_evlist__mmap_size(unsigned long pages)
913 {
914 	if (pages == UINT_MAX)
915 		pages = perf_event_mlock_kb_in_pages();
916 	else if (!is_power_of_2(pages))
917 		return 0;
918 
919 	return (pages + 1) * page_size;
920 }
921 
922 static long parse_pages_arg(const char *str, unsigned long min,
923 			    unsigned long max)
924 {
925 	unsigned long pages, val;
926 	static struct parse_tag tags[] = {
927 		{ .tag  = 'B', .mult = 1       },
928 		{ .tag  = 'K', .mult = 1 << 10 },
929 		{ .tag  = 'M', .mult = 1 << 20 },
930 		{ .tag  = 'G', .mult = 1 << 30 },
931 		{ .tag  = 0 },
932 	};
933 
934 	if (str == NULL)
935 		return -EINVAL;
936 
937 	val = parse_tag_value(str, tags);
938 	if (val != (unsigned long) -1) {
939 		/* we got file size value */
940 		pages = PERF_ALIGN(val, page_size) / page_size;
941 	} else {
942 		/* we got pages count value */
943 		char *eptr;
944 		pages = strtoul(str, &eptr, 10);
945 		if (*eptr != '\0')
946 			return -EINVAL;
947 	}
948 
949 	if (pages == 0 && min == 0) {
950 		/* leave number of pages at 0 */
951 	} else if (!is_power_of_2(pages)) {
952 		char buf[100];
953 
954 		/* round pages up to next power of 2 */
955 		pages = roundup_pow_of_two(pages);
956 		if (!pages)
957 			return -EINVAL;
958 
959 		unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
960 		pr_info("rounding mmap pages size to %s (%lu pages)\n",
961 			buf, pages);
962 	}
963 
964 	if (pages > max)
965 		return -EINVAL;
966 
967 	return pages;
968 }
969 
970 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
971 {
972 	unsigned long max = UINT_MAX;
973 	long pages;
974 
975 	if (max > SIZE_MAX / page_size)
976 		max = SIZE_MAX / page_size;
977 
978 	pages = parse_pages_arg(str, 1, max);
979 	if (pages < 0) {
980 		pr_err("Invalid argument for --mmap_pages/-m\n");
981 		return -1;
982 	}
983 
984 	*mmap_pages = pages;
985 	return 0;
986 }
987 
988 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
989 				  int unset __maybe_unused)
990 {
991 	return __perf_evlist__parse_mmap_pages(opt->value, str);
992 }
993 
994 /**
995  * perf_evlist__mmap_ex - Create mmaps to receive events.
996  * @evlist: list of events
997  * @pages: map length in pages
998  * @overwrite: overwrite older events?
999  * @auxtrace_pages - auxtrace map length in pages
1000  * @auxtrace_overwrite - overwrite older auxtrace data?
1001  *
1002  * If @overwrite is %false the user needs to signal event consumption using
1003  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1004  * automatically.
1005  *
1006  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1007  * consumption using auxtrace_mmap__write_tail().
1008  *
1009  * Return: %0 on success, negative error code otherwise.
1010  */
1011 int perf_evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
1012 			 unsigned int auxtrace_pages,
1013 			 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1014 			 int comp_level)
1015 {
1016 	struct evsel *evsel;
1017 	const struct perf_cpu_map *cpus = evlist->cpus;
1018 	const struct perf_thread_map *threads = evlist->threads;
1019 	/*
1020 	 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1021 	 * Its value is decided by evsel's write_backward.
1022 	 * So &mp should not be passed through const pointer.
1023 	 */
1024 	struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
1025 				  .comp_level = comp_level };
1026 
1027 	if (!evlist->mmap)
1028 		evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1029 	if (!evlist->mmap)
1030 		return -ENOMEM;
1031 
1032 	if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1033 		return -ENOMEM;
1034 
1035 	evlist->mmap_len = perf_evlist__mmap_size(pages);
1036 	pr_debug("mmap size %zuB\n", evlist->mmap_len);
1037 	mp.mask = evlist->mmap_len - page_size - 1;
1038 
1039 	auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1040 				   auxtrace_pages, auxtrace_overwrite);
1041 
1042 	evlist__for_each_entry(evlist, evsel) {
1043 		if ((evsel->core.attr.read_format & PERF_FORMAT_ID) &&
1044 		    evsel->sample_id == NULL &&
1045 		    perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1046 			return -ENOMEM;
1047 	}
1048 
1049 	if (cpu_map__empty(cpus))
1050 		return perf_evlist__mmap_per_thread(evlist, &mp);
1051 
1052 	return perf_evlist__mmap_per_cpu(evlist, &mp);
1053 }
1054 
1055 int perf_evlist__mmap(struct evlist *evlist, unsigned int pages)
1056 {
1057 	return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1058 }
1059 
1060 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
1061 {
1062 	bool all_threads = (target->per_thread && target->system_wide);
1063 	struct perf_cpu_map *cpus;
1064 	struct perf_thread_map *threads;
1065 
1066 	/*
1067 	 * If specify '-a' and '--per-thread' to perf record, perf record
1068 	 * will override '--per-thread'. target->per_thread = false and
1069 	 * target->system_wide = true.
1070 	 *
1071 	 * If specify '--per-thread' only to perf record,
1072 	 * target->per_thread = true and target->system_wide = false.
1073 	 *
1074 	 * So target->per_thread && target->system_wide is false.
1075 	 * For perf record, thread_map__new_str doesn't call
1076 	 * thread_map__new_all_cpus. That will keep perf record's
1077 	 * current behavior.
1078 	 *
1079 	 * For perf stat, it allows the case that target->per_thread and
1080 	 * target->system_wide are all true. It means to collect system-wide
1081 	 * per-thread data. thread_map__new_str will call
1082 	 * thread_map__new_all_cpus to enumerate all threads.
1083 	 */
1084 	threads = thread_map__new_str(target->pid, target->tid, target->uid,
1085 				      all_threads);
1086 
1087 	if (!threads)
1088 		return -1;
1089 
1090 	if (target__uses_dummy_map(target))
1091 		cpus = perf_cpu_map__dummy_new();
1092 	else
1093 		cpus = perf_cpu_map__new(target->cpu_list);
1094 
1095 	if (!cpus)
1096 		goto out_delete_threads;
1097 
1098 	evlist->has_user_cpus = !!target->cpu_list;
1099 
1100 	perf_evlist__set_maps(evlist, cpus, threads);
1101 
1102 	return 0;
1103 
1104 out_delete_threads:
1105 	perf_thread_map__put(threads);
1106 	return -1;
1107 }
1108 
1109 void perf_evlist__set_maps(struct evlist *evlist, struct perf_cpu_map *cpus,
1110 			   struct perf_thread_map *threads)
1111 {
1112 	/*
1113 	 * Allow for the possibility that one or another of the maps isn't being
1114 	 * changed i.e. don't put it.  Note we are assuming the maps that are
1115 	 * being applied are brand new and evlist is taking ownership of the
1116 	 * original reference count of 1.  If that is not the case it is up to
1117 	 * the caller to increase the reference count.
1118 	 */
1119 	if (cpus != evlist->cpus) {
1120 		perf_cpu_map__put(evlist->cpus);
1121 		evlist->cpus = perf_cpu_map__get(cpus);
1122 	}
1123 
1124 	if (threads != evlist->threads) {
1125 		perf_thread_map__put(evlist->threads);
1126 		evlist->threads = perf_thread_map__get(threads);
1127 	}
1128 
1129 	perf_evlist__propagate_maps(evlist);
1130 }
1131 
1132 void __perf_evlist__set_sample_bit(struct evlist *evlist,
1133 				   enum perf_event_sample_format bit)
1134 {
1135 	struct evsel *evsel;
1136 
1137 	evlist__for_each_entry(evlist, evsel)
1138 		__perf_evsel__set_sample_bit(evsel, bit);
1139 }
1140 
1141 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
1142 				     enum perf_event_sample_format bit)
1143 {
1144 	struct evsel *evsel;
1145 
1146 	evlist__for_each_entry(evlist, evsel)
1147 		__perf_evsel__reset_sample_bit(evsel, bit);
1148 }
1149 
1150 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1151 {
1152 	struct evsel *evsel;
1153 	int err = 0;
1154 
1155 	evlist__for_each_entry(evlist, evsel) {
1156 		if (evsel->filter == NULL)
1157 			continue;
1158 
1159 		/*
1160 		 * filters only work for tracepoint event, which doesn't have cpu limit.
1161 		 * So evlist and evsel should always be same.
1162 		 */
1163 		err = evsel__apply_filter(evsel, evsel->filter);
1164 		if (err) {
1165 			*err_evsel = evsel;
1166 			break;
1167 		}
1168 	}
1169 
1170 	return err;
1171 }
1172 
1173 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1174 {
1175 	struct evsel *evsel;
1176 	int err = 0;
1177 
1178 	evlist__for_each_entry(evlist, evsel) {
1179 		if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1180 			continue;
1181 
1182 		err = perf_evsel__set_filter(evsel, filter);
1183 		if (err)
1184 			break;
1185 	}
1186 
1187 	return err;
1188 }
1189 
1190 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1191 {
1192 	char *filter;
1193 	int ret = -1;
1194 	size_t i;
1195 
1196 	for (i = 0; i < npids; ++i) {
1197 		if (i == 0) {
1198 			if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1199 				return -1;
1200 		} else {
1201 			char *tmp;
1202 
1203 			if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1204 				goto out_free;
1205 
1206 			free(filter);
1207 			filter = tmp;
1208 		}
1209 	}
1210 
1211 	ret = perf_evlist__set_tp_filter(evlist, filter);
1212 out_free:
1213 	free(filter);
1214 	return ret;
1215 }
1216 
1217 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1218 {
1219 	return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1220 }
1221 
1222 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1223 {
1224 	struct evsel *pos;
1225 
1226 	if (evlist->core.nr_entries == 1)
1227 		return true;
1228 
1229 	if (evlist->id_pos < 0 || evlist->is_pos < 0)
1230 		return false;
1231 
1232 	evlist__for_each_entry(evlist, pos) {
1233 		if (pos->id_pos != evlist->id_pos ||
1234 		    pos->is_pos != evlist->is_pos)
1235 			return false;
1236 	}
1237 
1238 	return true;
1239 }
1240 
1241 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1242 {
1243 	struct evsel *evsel;
1244 
1245 	if (evlist->combined_sample_type)
1246 		return evlist->combined_sample_type;
1247 
1248 	evlist__for_each_entry(evlist, evsel)
1249 		evlist->combined_sample_type |= evsel->core.attr.sample_type;
1250 
1251 	return evlist->combined_sample_type;
1252 }
1253 
1254 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1255 {
1256 	evlist->combined_sample_type = 0;
1257 	return __perf_evlist__combined_sample_type(evlist);
1258 }
1259 
1260 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1261 {
1262 	struct evsel *evsel;
1263 	u64 branch_type = 0;
1264 
1265 	evlist__for_each_entry(evlist, evsel)
1266 		branch_type |= evsel->core.attr.branch_sample_type;
1267 	return branch_type;
1268 }
1269 
1270 bool perf_evlist__valid_read_format(struct evlist *evlist)
1271 {
1272 	struct evsel *first = perf_evlist__first(evlist), *pos = first;
1273 	u64 read_format = first->core.attr.read_format;
1274 	u64 sample_type = first->core.attr.sample_type;
1275 
1276 	evlist__for_each_entry(evlist, pos) {
1277 		if (read_format != pos->core.attr.read_format)
1278 			return false;
1279 	}
1280 
1281 	/* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1282 	if ((sample_type & PERF_SAMPLE_READ) &&
1283 	    !(read_format & PERF_FORMAT_ID)) {
1284 		return false;
1285 	}
1286 
1287 	return true;
1288 }
1289 
1290 u64 perf_evlist__read_format(struct evlist *evlist)
1291 {
1292 	struct evsel *first = perf_evlist__first(evlist);
1293 	return first->core.attr.read_format;
1294 }
1295 
1296 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1297 {
1298 	struct evsel *first = perf_evlist__first(evlist);
1299 	struct perf_sample *data;
1300 	u64 sample_type;
1301 	u16 size = 0;
1302 
1303 	if (!first->core.attr.sample_id_all)
1304 		goto out;
1305 
1306 	sample_type = first->core.attr.sample_type;
1307 
1308 	if (sample_type & PERF_SAMPLE_TID)
1309 		size += sizeof(data->tid) * 2;
1310 
1311        if (sample_type & PERF_SAMPLE_TIME)
1312 		size += sizeof(data->time);
1313 
1314 	if (sample_type & PERF_SAMPLE_ID)
1315 		size += sizeof(data->id);
1316 
1317 	if (sample_type & PERF_SAMPLE_STREAM_ID)
1318 		size += sizeof(data->stream_id);
1319 
1320 	if (sample_type & PERF_SAMPLE_CPU)
1321 		size += sizeof(data->cpu) * 2;
1322 
1323 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
1324 		size += sizeof(data->id);
1325 out:
1326 	return size;
1327 }
1328 
1329 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1330 {
1331 	struct evsel *first = perf_evlist__first(evlist), *pos = first;
1332 
1333 	evlist__for_each_entry_continue(evlist, pos) {
1334 		if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1335 			return false;
1336 	}
1337 
1338 	return true;
1339 }
1340 
1341 bool perf_evlist__sample_id_all(struct evlist *evlist)
1342 {
1343 	struct evsel *first = perf_evlist__first(evlist);
1344 	return first->core.attr.sample_id_all;
1345 }
1346 
1347 void perf_evlist__set_selected(struct evlist *evlist,
1348 			       struct evsel *evsel)
1349 {
1350 	evlist->selected = evsel;
1351 }
1352 
1353 void evlist__close(struct evlist *evlist)
1354 {
1355 	struct evsel *evsel;
1356 
1357 	evlist__for_each_entry_reverse(evlist, evsel)
1358 		perf_evsel__close(evsel);
1359 }
1360 
1361 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1362 {
1363 	struct perf_cpu_map *cpus;
1364 	struct perf_thread_map *threads;
1365 	int err = -ENOMEM;
1366 
1367 	/*
1368 	 * Try reading /sys/devices/system/cpu/online to get
1369 	 * an all cpus map.
1370 	 *
1371 	 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1372 	 * code needs an overhaul to properly forward the
1373 	 * error, and we may not want to do that fallback to a
1374 	 * default cpu identity map :-\
1375 	 */
1376 	cpus = perf_cpu_map__new(NULL);
1377 	if (!cpus)
1378 		goto out;
1379 
1380 	threads = perf_thread_map__new_dummy();
1381 	if (!threads)
1382 		goto out_put;
1383 
1384 	perf_evlist__set_maps(evlist, cpus, threads);
1385 out:
1386 	return err;
1387 out_put:
1388 	perf_cpu_map__put(cpus);
1389 	goto out;
1390 }
1391 
1392 int evlist__open(struct evlist *evlist)
1393 {
1394 	struct evsel *evsel;
1395 	int err;
1396 
1397 	/*
1398 	 * Default: one fd per CPU, all threads, aka systemwide
1399 	 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1400 	 */
1401 	if (evlist->threads == NULL && evlist->cpus == NULL) {
1402 		err = perf_evlist__create_syswide_maps(evlist);
1403 		if (err < 0)
1404 			goto out_err;
1405 	}
1406 
1407 	perf_evlist__update_id_pos(evlist);
1408 
1409 	evlist__for_each_entry(evlist, evsel) {
1410 		err = evsel__open(evsel, evsel->core.cpus, evsel->threads);
1411 		if (err < 0)
1412 			goto out_err;
1413 	}
1414 
1415 	return 0;
1416 out_err:
1417 	evlist__close(evlist);
1418 	errno = -err;
1419 	return err;
1420 }
1421 
1422 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1423 				  const char *argv[], bool pipe_output,
1424 				  void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1425 {
1426 	int child_ready_pipe[2], go_pipe[2];
1427 	char bf;
1428 
1429 	if (pipe(child_ready_pipe) < 0) {
1430 		perror("failed to create 'ready' pipe");
1431 		return -1;
1432 	}
1433 
1434 	if (pipe(go_pipe) < 0) {
1435 		perror("failed to create 'go' pipe");
1436 		goto out_close_ready_pipe;
1437 	}
1438 
1439 	evlist->workload.pid = fork();
1440 	if (evlist->workload.pid < 0) {
1441 		perror("failed to fork");
1442 		goto out_close_pipes;
1443 	}
1444 
1445 	if (!evlist->workload.pid) {
1446 		int ret;
1447 
1448 		if (pipe_output)
1449 			dup2(2, 1);
1450 
1451 		signal(SIGTERM, SIG_DFL);
1452 
1453 		close(child_ready_pipe[0]);
1454 		close(go_pipe[1]);
1455 		fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1456 
1457 		/*
1458 		 * Tell the parent we're ready to go
1459 		 */
1460 		close(child_ready_pipe[1]);
1461 
1462 		/*
1463 		 * Wait until the parent tells us to go.
1464 		 */
1465 		ret = read(go_pipe[0], &bf, 1);
1466 		/*
1467 		 * The parent will ask for the execvp() to be performed by
1468 		 * writing exactly one byte, in workload.cork_fd, usually via
1469 		 * perf_evlist__start_workload().
1470 		 *
1471 		 * For cancelling the workload without actually running it,
1472 		 * the parent will just close workload.cork_fd, without writing
1473 		 * anything, i.e. read will return zero and we just exit()
1474 		 * here.
1475 		 */
1476 		if (ret != 1) {
1477 			if (ret == -1)
1478 				perror("unable to read pipe");
1479 			exit(ret);
1480 		}
1481 
1482 		execvp(argv[0], (char **)argv);
1483 
1484 		if (exec_error) {
1485 			union sigval val;
1486 
1487 			val.sival_int = errno;
1488 			if (sigqueue(getppid(), SIGUSR1, val))
1489 				perror(argv[0]);
1490 		} else
1491 			perror(argv[0]);
1492 		exit(-1);
1493 	}
1494 
1495 	if (exec_error) {
1496 		struct sigaction act = {
1497 			.sa_flags     = SA_SIGINFO,
1498 			.sa_sigaction = exec_error,
1499 		};
1500 		sigaction(SIGUSR1, &act, NULL);
1501 	}
1502 
1503 	if (target__none(target)) {
1504 		if (evlist->threads == NULL) {
1505 			fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1506 				__func__, __LINE__);
1507 			goto out_close_pipes;
1508 		}
1509 		perf_thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1510 	}
1511 
1512 	close(child_ready_pipe[1]);
1513 	close(go_pipe[0]);
1514 	/*
1515 	 * wait for child to settle
1516 	 */
1517 	if (read(child_ready_pipe[0], &bf, 1) == -1) {
1518 		perror("unable to read pipe");
1519 		goto out_close_pipes;
1520 	}
1521 
1522 	fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1523 	evlist->workload.cork_fd = go_pipe[1];
1524 	close(child_ready_pipe[0]);
1525 	return 0;
1526 
1527 out_close_pipes:
1528 	close(go_pipe[0]);
1529 	close(go_pipe[1]);
1530 out_close_ready_pipe:
1531 	close(child_ready_pipe[0]);
1532 	close(child_ready_pipe[1]);
1533 	return -1;
1534 }
1535 
1536 int perf_evlist__start_workload(struct evlist *evlist)
1537 {
1538 	if (evlist->workload.cork_fd > 0) {
1539 		char bf = 0;
1540 		int ret;
1541 		/*
1542 		 * Remove the cork, let it rip!
1543 		 */
1544 		ret = write(evlist->workload.cork_fd, &bf, 1);
1545 		if (ret < 0)
1546 			perror("unable to write to pipe");
1547 
1548 		close(evlist->workload.cork_fd);
1549 		return ret;
1550 	}
1551 
1552 	return 0;
1553 }
1554 
1555 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1556 			      struct perf_sample *sample)
1557 {
1558 	struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1559 
1560 	if (!evsel)
1561 		return -EFAULT;
1562 	return perf_evsel__parse_sample(evsel, event, sample);
1563 }
1564 
1565 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1566 					union perf_event *event,
1567 					u64 *timestamp)
1568 {
1569 	struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1570 
1571 	if (!evsel)
1572 		return -EFAULT;
1573 	return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1574 }
1575 
1576 size_t perf_evlist__fprintf(struct evlist *evlist, FILE *fp)
1577 {
1578 	struct evsel *evsel;
1579 	size_t printed = 0;
1580 
1581 	evlist__for_each_entry(evlist, evsel) {
1582 		printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1583 				   perf_evsel__name(evsel));
1584 	}
1585 
1586 	return printed + fprintf(fp, "\n");
1587 }
1588 
1589 int perf_evlist__strerror_open(struct evlist *evlist,
1590 			       int err, char *buf, size_t size)
1591 {
1592 	int printed, value;
1593 	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1594 
1595 	switch (err) {
1596 	case EACCES:
1597 	case EPERM:
1598 		printed = scnprintf(buf, size,
1599 				    "Error:\t%s.\n"
1600 				    "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1601 
1602 		value = perf_event_paranoid();
1603 
1604 		printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1605 
1606 		if (value >= 2) {
1607 			printed += scnprintf(buf + printed, size - printed,
1608 					     "For your workloads it needs to be <= 1\nHint:\t");
1609 		}
1610 		printed += scnprintf(buf + printed, size - printed,
1611 				     "For system wide tracing it needs to be set to -1.\n");
1612 
1613 		printed += scnprintf(buf + printed, size - printed,
1614 				    "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1615 				    "Hint:\tThe current value is %d.", value);
1616 		break;
1617 	case EINVAL: {
1618 		struct evsel *first = perf_evlist__first(evlist);
1619 		int max_freq;
1620 
1621 		if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1622 			goto out_default;
1623 
1624 		if (first->core.attr.sample_freq < (u64)max_freq)
1625 			goto out_default;
1626 
1627 		printed = scnprintf(buf, size,
1628 				    "Error:\t%s.\n"
1629 				    "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1630 				    "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1631 				    emsg, max_freq, first->core.attr.sample_freq);
1632 		break;
1633 	}
1634 	default:
1635 out_default:
1636 		scnprintf(buf, size, "%s", emsg);
1637 		break;
1638 	}
1639 
1640 	return 0;
1641 }
1642 
1643 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1644 {
1645 	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1646 	int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1647 
1648 	switch (err) {
1649 	case EPERM:
1650 		sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1651 		printed += scnprintf(buf + printed, size - printed,
1652 				     "Error:\t%s.\n"
1653 				     "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1654 				     "Hint:\tTried using %zd kB.\n",
1655 				     emsg, pages_max_per_user, pages_attempted);
1656 
1657 		if (pages_attempted >= pages_max_per_user) {
1658 			printed += scnprintf(buf + printed, size - printed,
1659 					     "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1660 					     pages_max_per_user + pages_attempted);
1661 		}
1662 
1663 		printed += scnprintf(buf + printed, size - printed,
1664 				     "Hint:\tTry using a smaller -m/--mmap-pages value.");
1665 		break;
1666 	default:
1667 		scnprintf(buf, size, "%s", emsg);
1668 		break;
1669 	}
1670 
1671 	return 0;
1672 }
1673 
1674 void perf_evlist__to_front(struct evlist *evlist,
1675 			   struct evsel *move_evsel)
1676 {
1677 	struct evsel *evsel, *n;
1678 	LIST_HEAD(move);
1679 
1680 	if (move_evsel == perf_evlist__first(evlist))
1681 		return;
1682 
1683 	evlist__for_each_entry_safe(evlist, n, evsel) {
1684 		if (evsel->leader == move_evsel->leader)
1685 			list_move_tail(&evsel->core.node, &move);
1686 	}
1687 
1688 	list_splice(&move, &evlist->core.entries);
1689 }
1690 
1691 void perf_evlist__set_tracking_event(struct evlist *evlist,
1692 				     struct evsel *tracking_evsel)
1693 {
1694 	struct evsel *evsel;
1695 
1696 	if (tracking_evsel->tracking)
1697 		return;
1698 
1699 	evlist__for_each_entry(evlist, evsel) {
1700 		if (evsel != tracking_evsel)
1701 			evsel->tracking = false;
1702 	}
1703 
1704 	tracking_evsel->tracking = true;
1705 }
1706 
1707 struct evsel *
1708 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1709 			       const char *str)
1710 {
1711 	struct evsel *evsel;
1712 
1713 	evlist__for_each_entry(evlist, evsel) {
1714 		if (!evsel->name)
1715 			continue;
1716 		if (strcmp(str, evsel->name) == 0)
1717 			return evsel;
1718 	}
1719 
1720 	return NULL;
1721 }
1722 
1723 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1724 				  enum bkw_mmap_state state)
1725 {
1726 	enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1727 	enum action {
1728 		NONE,
1729 		PAUSE,
1730 		RESUME,
1731 	} action = NONE;
1732 
1733 	if (!evlist->overwrite_mmap)
1734 		return;
1735 
1736 	switch (old_state) {
1737 	case BKW_MMAP_NOTREADY: {
1738 		if (state != BKW_MMAP_RUNNING)
1739 			goto state_err;
1740 		break;
1741 	}
1742 	case BKW_MMAP_RUNNING: {
1743 		if (state != BKW_MMAP_DATA_PENDING)
1744 			goto state_err;
1745 		action = PAUSE;
1746 		break;
1747 	}
1748 	case BKW_MMAP_DATA_PENDING: {
1749 		if (state != BKW_MMAP_EMPTY)
1750 			goto state_err;
1751 		break;
1752 	}
1753 	case BKW_MMAP_EMPTY: {
1754 		if (state != BKW_MMAP_RUNNING)
1755 			goto state_err;
1756 		action = RESUME;
1757 		break;
1758 	}
1759 	default:
1760 		WARN_ONCE(1, "Shouldn't get there\n");
1761 	}
1762 
1763 	evlist->bkw_mmap_state = state;
1764 
1765 	switch (action) {
1766 	case PAUSE:
1767 		perf_evlist__pause(evlist);
1768 		break;
1769 	case RESUME:
1770 		perf_evlist__resume(evlist);
1771 		break;
1772 	case NONE:
1773 	default:
1774 		break;
1775 	}
1776 
1777 state_err:
1778 	return;
1779 }
1780 
1781 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1782 {
1783 	struct evsel *evsel;
1784 
1785 	evlist__for_each_entry(evlist, evsel) {
1786 		if (!evsel->core.attr.exclude_kernel)
1787 			return false;
1788 	}
1789 
1790 	return true;
1791 }
1792 
1793 /*
1794  * Events in data file are not collect in groups, but we still want
1795  * the group display. Set the artificial group and set the leader's
1796  * forced_leader flag to notify the display code.
1797  */
1798 void perf_evlist__force_leader(struct evlist *evlist)
1799 {
1800 	if (!evlist->nr_groups) {
1801 		struct evsel *leader = perf_evlist__first(evlist);
1802 
1803 		perf_evlist__set_leader(evlist);
1804 		leader->forced_leader = true;
1805 	}
1806 }
1807 
1808 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1809 						 struct evsel *evsel)
1810 {
1811 	struct evsel *c2, *leader;
1812 	bool is_open = true;
1813 
1814 	leader = evsel->leader;
1815 	pr_debug("Weak group for %s/%d failed\n",
1816 			leader->name, leader->nr_members);
1817 
1818 	/*
1819 	 * for_each_group_member doesn't work here because it doesn't
1820 	 * include the first entry.
1821 	 */
1822 	evlist__for_each_entry(evsel_list, c2) {
1823 		if (c2 == evsel)
1824 			is_open = false;
1825 		if (c2->leader == leader) {
1826 			if (is_open)
1827 				perf_evsel__close(c2);
1828 			c2->leader = c2;
1829 			c2->nr_members = 0;
1830 		}
1831 	}
1832 	return leader;
1833 }
1834 
1835 int perf_evlist__add_sb_event(struct evlist **evlist,
1836 			      struct perf_event_attr *attr,
1837 			      perf_evsel__sb_cb_t cb,
1838 			      void *data)
1839 {
1840 	struct evsel *evsel;
1841 	bool new_evlist = (*evlist) == NULL;
1842 
1843 	if (*evlist == NULL)
1844 		*evlist = evlist__new();
1845 	if (*evlist == NULL)
1846 		return -1;
1847 
1848 	if (!attr->sample_id_all) {
1849 		pr_warning("enabling sample_id_all for all side band events\n");
1850 		attr->sample_id_all = 1;
1851 	}
1852 
1853 	evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1854 	if (!evsel)
1855 		goto out_err;
1856 
1857 	evsel->side_band.cb = cb;
1858 	evsel->side_band.data = data;
1859 	evlist__add(*evlist, evsel);
1860 	return 0;
1861 
1862 out_err:
1863 	if (new_evlist) {
1864 		evlist__delete(*evlist);
1865 		*evlist = NULL;
1866 	}
1867 	return -1;
1868 }
1869 
1870 static void *perf_evlist__poll_thread(void *arg)
1871 {
1872 	struct evlist *evlist = arg;
1873 	bool draining = false;
1874 	int i, done = 0;
1875 
1876 	while (!done) {
1877 		bool got_data = false;
1878 
1879 		if (evlist->thread.done)
1880 			draining = true;
1881 
1882 		if (!draining)
1883 			perf_evlist__poll(evlist, 1000);
1884 
1885 		for (i = 0; i < evlist->nr_mmaps; i++) {
1886 			struct perf_mmap *map = &evlist->mmap[i];
1887 			union perf_event *event;
1888 
1889 			if (perf_mmap__read_init(map))
1890 				continue;
1891 			while ((event = perf_mmap__read_event(map)) != NULL) {
1892 				struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1893 
1894 				if (evsel && evsel->side_band.cb)
1895 					evsel->side_band.cb(event, evsel->side_band.data);
1896 				else
1897 					pr_warning("cannot locate proper evsel for the side band event\n");
1898 
1899 				perf_mmap__consume(map);
1900 				got_data = true;
1901 			}
1902 			perf_mmap__read_done(map);
1903 		}
1904 
1905 		if (draining && !got_data)
1906 			break;
1907 	}
1908 	return NULL;
1909 }
1910 
1911 int perf_evlist__start_sb_thread(struct evlist *evlist,
1912 				 struct target *target)
1913 {
1914 	struct evsel *counter;
1915 
1916 	if (!evlist)
1917 		return 0;
1918 
1919 	if (perf_evlist__create_maps(evlist, target))
1920 		goto out_delete_evlist;
1921 
1922 	evlist__for_each_entry(evlist, counter) {
1923 		if (evsel__open(counter, evlist->cpus,
1924 				     evlist->threads) < 0)
1925 			goto out_delete_evlist;
1926 	}
1927 
1928 	if (perf_evlist__mmap(evlist, UINT_MAX))
1929 		goto out_delete_evlist;
1930 
1931 	evlist__for_each_entry(evlist, counter) {
1932 		if (evsel__enable(counter))
1933 			goto out_delete_evlist;
1934 	}
1935 
1936 	evlist->thread.done = 0;
1937 	if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1938 		goto out_delete_evlist;
1939 
1940 	return 0;
1941 
1942 out_delete_evlist:
1943 	evlist__delete(evlist);
1944 	evlist = NULL;
1945 	return -1;
1946 }
1947 
1948 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1949 {
1950 	if (!evlist)
1951 		return;
1952 	evlist->thread.done = 1;
1953 	pthread_join(evlist->thread.th, NULL);
1954 	evlist__delete(evlist);
1955 }
1956