Loading...
1/*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9#include "util.h"
10#include <api/fs/fs.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 <unistd.h>
19
20#include "parse-events.h"
21#include <subcmd/parse-options.h>
22
23#include <sys/mman.h>
24
25#include <linux/bitops.h>
26#include <linux/hash.h>
27#include <linux/log2.h>
28#include <linux/err.h>
29
30static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
31static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
32
33#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
35
36void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
37 struct thread_map *threads)
38{
39 int i;
40
41 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
42 INIT_HLIST_HEAD(&evlist->heads[i]);
43 INIT_LIST_HEAD(&evlist->entries);
44 perf_evlist__set_maps(evlist, cpus, threads);
45 fdarray__init(&evlist->pollfd, 64);
46 evlist->workload.pid = -1;
47}
48
49struct perf_evlist *perf_evlist__new(void)
50{
51 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
52
53 if (evlist != NULL)
54 perf_evlist__init(evlist, NULL, NULL);
55
56 return evlist;
57}
58
59struct perf_evlist *perf_evlist__new_default(void)
60{
61 struct perf_evlist *evlist = perf_evlist__new();
62
63 if (evlist && perf_evlist__add_default(evlist)) {
64 perf_evlist__delete(evlist);
65 evlist = NULL;
66 }
67
68 return evlist;
69}
70
71struct perf_evlist *perf_evlist__new_dummy(void)
72{
73 struct perf_evlist *evlist = perf_evlist__new();
74
75 if (evlist && perf_evlist__add_dummy(evlist)) {
76 perf_evlist__delete(evlist);
77 evlist = NULL;
78 }
79
80 return evlist;
81}
82
83/**
84 * perf_evlist__set_id_pos - set the positions of event ids.
85 * @evlist: selected event list
86 *
87 * Events with compatible sample types all have the same id_pos
88 * and is_pos. For convenience, put a copy on evlist.
89 */
90void perf_evlist__set_id_pos(struct perf_evlist *evlist)
91{
92 struct perf_evsel *first = perf_evlist__first(evlist);
93
94 evlist->id_pos = first->id_pos;
95 evlist->is_pos = first->is_pos;
96}
97
98static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
99{
100 struct perf_evsel *evsel;
101
102 evlist__for_each(evlist, evsel)
103 perf_evsel__calc_id_pos(evsel);
104
105 perf_evlist__set_id_pos(evlist);
106}
107
108static void perf_evlist__purge(struct perf_evlist *evlist)
109{
110 struct perf_evsel *pos, *n;
111
112 evlist__for_each_safe(evlist, n, pos) {
113 list_del_init(&pos->node);
114 pos->evlist = NULL;
115 perf_evsel__delete(pos);
116 }
117
118 evlist->nr_entries = 0;
119}
120
121void perf_evlist__exit(struct perf_evlist *evlist)
122{
123 zfree(&evlist->mmap);
124 fdarray__exit(&evlist->pollfd);
125}
126
127void perf_evlist__delete(struct perf_evlist *evlist)
128{
129 perf_evlist__munmap(evlist);
130 perf_evlist__close(evlist);
131 cpu_map__put(evlist->cpus);
132 thread_map__put(evlist->threads);
133 evlist->cpus = NULL;
134 evlist->threads = NULL;
135 perf_evlist__purge(evlist);
136 perf_evlist__exit(evlist);
137 free(evlist);
138}
139
140static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
141 struct perf_evsel *evsel)
142{
143 /*
144 * We already have cpus for evsel (via PMU sysfs) so
145 * keep it, if there's no target cpu list defined.
146 */
147 if (!evsel->own_cpus || evlist->has_user_cpus) {
148 cpu_map__put(evsel->cpus);
149 evsel->cpus = cpu_map__get(evlist->cpus);
150 } else if (evsel->cpus != evsel->own_cpus) {
151 cpu_map__put(evsel->cpus);
152 evsel->cpus = cpu_map__get(evsel->own_cpus);
153 }
154
155 thread_map__put(evsel->threads);
156 evsel->threads = thread_map__get(evlist->threads);
157}
158
159static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
160{
161 struct perf_evsel *evsel;
162
163 evlist__for_each(evlist, evsel)
164 __perf_evlist__propagate_maps(evlist, evsel);
165}
166
167void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
168{
169 entry->evlist = evlist;
170 list_add_tail(&entry->node, &evlist->entries);
171 entry->idx = evlist->nr_entries;
172 entry->tracking = !entry->idx;
173
174 if (!evlist->nr_entries++)
175 perf_evlist__set_id_pos(evlist);
176
177 __perf_evlist__propagate_maps(evlist, entry);
178}
179
180void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
181{
182 evsel->evlist = NULL;
183 list_del_init(&evsel->node);
184 evlist->nr_entries -= 1;
185}
186
187void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
188 struct list_head *list)
189{
190 struct perf_evsel *evsel, *temp;
191
192 __evlist__for_each_safe(list, temp, evsel) {
193 list_del_init(&evsel->node);
194 perf_evlist__add(evlist, evsel);
195 }
196}
197
198void __perf_evlist__set_leader(struct list_head *list)
199{
200 struct perf_evsel *evsel, *leader;
201
202 leader = list_entry(list->next, struct perf_evsel, node);
203 evsel = list_entry(list->prev, struct perf_evsel, node);
204
205 leader->nr_members = evsel->idx - leader->idx + 1;
206
207 __evlist__for_each(list, evsel) {
208 evsel->leader = leader;
209 }
210}
211
212void perf_evlist__set_leader(struct perf_evlist *evlist)
213{
214 if (evlist->nr_entries) {
215 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
216 __perf_evlist__set_leader(&evlist->entries);
217 }
218}
219
220void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
221{
222 attr->precise_ip = 3;
223
224 while (attr->precise_ip != 0) {
225 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
226 if (fd != -1) {
227 close(fd);
228 break;
229 }
230 --attr->precise_ip;
231 }
232}
233
234int perf_evlist__add_default(struct perf_evlist *evlist)
235{
236 struct perf_event_attr attr = {
237 .type = PERF_TYPE_HARDWARE,
238 .config = PERF_COUNT_HW_CPU_CYCLES,
239 };
240 struct perf_evsel *evsel;
241
242 event_attr_init(&attr);
243
244 perf_event_attr__set_max_precise_ip(&attr);
245
246 evsel = perf_evsel__new(&attr);
247 if (evsel == NULL)
248 goto error;
249
250 /* use asprintf() because free(evsel) assumes name is allocated */
251 if (asprintf(&evsel->name, "cycles%.*s",
252 attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
253 goto error_free;
254
255 perf_evlist__add(evlist, evsel);
256 return 0;
257error_free:
258 perf_evsel__delete(evsel);
259error:
260 return -ENOMEM;
261}
262
263int perf_evlist__add_dummy(struct perf_evlist *evlist)
264{
265 struct perf_event_attr attr = {
266 .type = PERF_TYPE_SOFTWARE,
267 .config = PERF_COUNT_SW_DUMMY,
268 .size = sizeof(attr), /* to capture ABI version */
269 };
270 struct perf_evsel *evsel = perf_evsel__new(&attr);
271
272 if (evsel == NULL)
273 return -ENOMEM;
274
275 perf_evlist__add(evlist, evsel);
276 return 0;
277}
278
279static int perf_evlist__add_attrs(struct perf_evlist *evlist,
280 struct perf_event_attr *attrs, size_t nr_attrs)
281{
282 struct perf_evsel *evsel, *n;
283 LIST_HEAD(head);
284 size_t i;
285
286 for (i = 0; i < nr_attrs; i++) {
287 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
288 if (evsel == NULL)
289 goto out_delete_partial_list;
290 list_add_tail(&evsel->node, &head);
291 }
292
293 perf_evlist__splice_list_tail(evlist, &head);
294
295 return 0;
296
297out_delete_partial_list:
298 __evlist__for_each_safe(&head, n, evsel)
299 perf_evsel__delete(evsel);
300 return -1;
301}
302
303int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
304 struct perf_event_attr *attrs, size_t nr_attrs)
305{
306 size_t i;
307
308 for (i = 0; i < nr_attrs; i++)
309 event_attr_init(attrs + i);
310
311 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
312}
313
314struct perf_evsel *
315perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
316{
317 struct perf_evsel *evsel;
318
319 evlist__for_each(evlist, evsel) {
320 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
321 (int)evsel->attr.config == id)
322 return evsel;
323 }
324
325 return NULL;
326}
327
328struct perf_evsel *
329perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
330 const char *name)
331{
332 struct perf_evsel *evsel;
333
334 evlist__for_each(evlist, evsel) {
335 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
336 (strcmp(evsel->name, name) == 0))
337 return evsel;
338 }
339
340 return NULL;
341}
342
343int perf_evlist__add_newtp(struct perf_evlist *evlist,
344 const char *sys, const char *name, void *handler)
345{
346 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
347
348 if (IS_ERR(evsel))
349 return -1;
350
351 evsel->handler = handler;
352 perf_evlist__add(evlist, evsel);
353 return 0;
354}
355
356static int perf_evlist__nr_threads(struct perf_evlist *evlist,
357 struct perf_evsel *evsel)
358{
359 if (evsel->system_wide)
360 return 1;
361 else
362 return thread_map__nr(evlist->threads);
363}
364
365void perf_evlist__disable(struct perf_evlist *evlist)
366{
367 struct perf_evsel *pos;
368
369 evlist__for_each(evlist, pos) {
370 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
371 continue;
372 perf_evsel__disable(pos);
373 }
374
375 evlist->enabled = false;
376}
377
378void perf_evlist__enable(struct perf_evlist *evlist)
379{
380 struct perf_evsel *pos;
381
382 evlist__for_each(evlist, pos) {
383 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
384 continue;
385 perf_evsel__enable(pos);
386 }
387
388 evlist->enabled = true;
389}
390
391void perf_evlist__toggle_enable(struct perf_evlist *evlist)
392{
393 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
394}
395
396static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
397 struct perf_evsel *evsel, int cpu)
398{
399 int thread, err;
400 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
401
402 if (!evsel->fd)
403 return -EINVAL;
404
405 for (thread = 0; thread < nr_threads; thread++) {
406 err = ioctl(FD(evsel, cpu, thread),
407 PERF_EVENT_IOC_ENABLE, 0);
408 if (err)
409 return err;
410 }
411 return 0;
412}
413
414static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
415 struct perf_evsel *evsel,
416 int thread)
417{
418 int cpu, err;
419 int nr_cpus = cpu_map__nr(evlist->cpus);
420
421 if (!evsel->fd)
422 return -EINVAL;
423
424 for (cpu = 0; cpu < nr_cpus; cpu++) {
425 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
426 if (err)
427 return err;
428 }
429 return 0;
430}
431
432int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
433 struct perf_evsel *evsel, int idx)
434{
435 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
436
437 if (per_cpu_mmaps)
438 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
439 else
440 return perf_evlist__enable_event_thread(evlist, evsel, idx);
441}
442
443int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
444{
445 int nr_cpus = cpu_map__nr(evlist->cpus);
446 int nr_threads = thread_map__nr(evlist->threads);
447 int nfds = 0;
448 struct perf_evsel *evsel;
449
450 evlist__for_each(evlist, evsel) {
451 if (evsel->system_wide)
452 nfds += nr_cpus;
453 else
454 nfds += nr_cpus * nr_threads;
455 }
456
457 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
458 fdarray__grow(&evlist->pollfd, nfds) < 0)
459 return -ENOMEM;
460
461 return 0;
462}
463
464static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
465{
466 int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
467 /*
468 * Save the idx so that when we filter out fds POLLHUP'ed we can
469 * close the associated evlist->mmap[] entry.
470 */
471 if (pos >= 0) {
472 evlist->pollfd.priv[pos].idx = idx;
473
474 fcntl(fd, F_SETFL, O_NONBLOCK);
475 }
476
477 return pos;
478}
479
480int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
481{
482 return __perf_evlist__add_pollfd(evlist, fd, -1);
483}
484
485static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
486{
487 struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
488
489 perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
490}
491
492int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
493{
494 return fdarray__filter(&evlist->pollfd, revents_and_mask,
495 perf_evlist__munmap_filtered);
496}
497
498int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
499{
500 return fdarray__poll(&evlist->pollfd, timeout);
501}
502
503static void perf_evlist__id_hash(struct perf_evlist *evlist,
504 struct perf_evsel *evsel,
505 int cpu, int thread, u64 id)
506{
507 int hash;
508 struct perf_sample_id *sid = SID(evsel, cpu, thread);
509
510 sid->id = id;
511 sid->evsel = evsel;
512 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
513 hlist_add_head(&sid->node, &evlist->heads[hash]);
514}
515
516void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
517 int cpu, int thread, u64 id)
518{
519 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
520 evsel->id[evsel->ids++] = id;
521}
522
523int perf_evlist__id_add_fd(struct perf_evlist *evlist,
524 struct perf_evsel *evsel,
525 int cpu, int thread, int fd)
526{
527 u64 read_data[4] = { 0, };
528 int id_idx = 1; /* The first entry is the counter value */
529 u64 id;
530 int ret;
531
532 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
533 if (!ret)
534 goto add;
535
536 if (errno != ENOTTY)
537 return -1;
538
539 /* Legacy way to get event id.. All hail to old kernels! */
540
541 /*
542 * This way does not work with group format read, so bail
543 * out in that case.
544 */
545 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
546 return -1;
547
548 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
549 read(fd, &read_data, sizeof(read_data)) == -1)
550 return -1;
551
552 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
553 ++id_idx;
554 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
555 ++id_idx;
556
557 id = read_data[id_idx];
558
559 add:
560 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
561 return 0;
562}
563
564static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
565 struct perf_evsel *evsel, int idx, int cpu,
566 int thread)
567{
568 struct perf_sample_id *sid = SID(evsel, cpu, thread);
569 sid->idx = idx;
570 if (evlist->cpus && cpu >= 0)
571 sid->cpu = evlist->cpus->map[cpu];
572 else
573 sid->cpu = -1;
574 if (!evsel->system_wide && evlist->threads && thread >= 0)
575 sid->tid = thread_map__pid(evlist->threads, thread);
576 else
577 sid->tid = -1;
578}
579
580struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
581{
582 struct hlist_head *head;
583 struct perf_sample_id *sid;
584 int hash;
585
586 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
587 head = &evlist->heads[hash];
588
589 hlist_for_each_entry(sid, head, node)
590 if (sid->id == id)
591 return sid;
592
593 return NULL;
594}
595
596struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
597{
598 struct perf_sample_id *sid;
599
600 if (evlist->nr_entries == 1 || !id)
601 return perf_evlist__first(evlist);
602
603 sid = perf_evlist__id2sid(evlist, id);
604 if (sid)
605 return sid->evsel;
606
607 if (!perf_evlist__sample_id_all(evlist))
608 return perf_evlist__first(evlist);
609
610 return NULL;
611}
612
613struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
614 u64 id)
615{
616 struct perf_sample_id *sid;
617
618 if (!id)
619 return NULL;
620
621 sid = perf_evlist__id2sid(evlist, id);
622 if (sid)
623 return sid->evsel;
624
625 return NULL;
626}
627
628static int perf_evlist__event2id(struct perf_evlist *evlist,
629 union perf_event *event, u64 *id)
630{
631 const u64 *array = event->sample.array;
632 ssize_t n;
633
634 n = (event->header.size - sizeof(event->header)) >> 3;
635
636 if (event->header.type == PERF_RECORD_SAMPLE) {
637 if (evlist->id_pos >= n)
638 return -1;
639 *id = array[evlist->id_pos];
640 } else {
641 if (evlist->is_pos > n)
642 return -1;
643 n -= evlist->is_pos;
644 *id = array[n];
645 }
646 return 0;
647}
648
649static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
650 union perf_event *event)
651{
652 struct perf_evsel *first = perf_evlist__first(evlist);
653 struct hlist_head *head;
654 struct perf_sample_id *sid;
655 int hash;
656 u64 id;
657
658 if (evlist->nr_entries == 1)
659 return first;
660
661 if (!first->attr.sample_id_all &&
662 event->header.type != PERF_RECORD_SAMPLE)
663 return first;
664
665 if (perf_evlist__event2id(evlist, event, &id))
666 return NULL;
667
668 /* Synthesized events have an id of zero */
669 if (!id)
670 return first;
671
672 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
673 head = &evlist->heads[hash];
674
675 hlist_for_each_entry(sid, head, node) {
676 if (sid->id == id)
677 return sid->evsel;
678 }
679 return NULL;
680}
681
682union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
683{
684 struct perf_mmap *md = &evlist->mmap[idx];
685 u64 head;
686 u64 old = md->prev;
687 unsigned char *data = md->base + page_size;
688 union perf_event *event = NULL;
689
690 /*
691 * Check if event was unmapped due to a POLLHUP/POLLERR.
692 */
693 if (!atomic_read(&md->refcnt))
694 return NULL;
695
696 head = perf_mmap__read_head(md);
697 if (evlist->overwrite) {
698 /*
699 * If we're further behind than half the buffer, there's a chance
700 * the writer will bite our tail and mess up the samples under us.
701 *
702 * If we somehow ended up ahead of the head, we got messed up.
703 *
704 * In either case, truncate and restart at head.
705 */
706 int diff = head - old;
707 if (diff > md->mask / 2 || diff < 0) {
708 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
709
710 /*
711 * head points to a known good entry, start there.
712 */
713 old = head;
714 }
715 }
716
717 if (old != head) {
718 size_t size;
719
720 event = (union perf_event *)&data[old & md->mask];
721 size = event->header.size;
722
723 /*
724 * Event straddles the mmap boundary -- header should always
725 * be inside due to u64 alignment of output.
726 */
727 if ((old & md->mask) + size != ((old + size) & md->mask)) {
728 unsigned int offset = old;
729 unsigned int len = min(sizeof(*event), size), cpy;
730 void *dst = md->event_copy;
731
732 do {
733 cpy = min(md->mask + 1 - (offset & md->mask), len);
734 memcpy(dst, &data[offset & md->mask], cpy);
735 offset += cpy;
736 dst += cpy;
737 len -= cpy;
738 } while (len);
739
740 event = (union perf_event *) md->event_copy;
741 }
742
743 old += size;
744 }
745
746 md->prev = old;
747
748 return event;
749}
750
751static bool perf_mmap__empty(struct perf_mmap *md)
752{
753 return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
754}
755
756static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
757{
758 atomic_inc(&evlist->mmap[idx].refcnt);
759}
760
761static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
762{
763 BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
764
765 if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
766 __perf_evlist__munmap(evlist, idx);
767}
768
769void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
770{
771 struct perf_mmap *md = &evlist->mmap[idx];
772
773 if (!evlist->overwrite) {
774 u64 old = md->prev;
775
776 perf_mmap__write_tail(md, old);
777 }
778
779 if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
780 perf_evlist__mmap_put(evlist, idx);
781}
782
783int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
784 struct auxtrace_mmap_params *mp __maybe_unused,
785 void *userpg __maybe_unused,
786 int fd __maybe_unused)
787{
788 return 0;
789}
790
791void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
792{
793}
794
795void __weak auxtrace_mmap_params__init(
796 struct auxtrace_mmap_params *mp __maybe_unused,
797 off_t auxtrace_offset __maybe_unused,
798 unsigned int auxtrace_pages __maybe_unused,
799 bool auxtrace_overwrite __maybe_unused)
800{
801}
802
803void __weak auxtrace_mmap_params__set_idx(
804 struct auxtrace_mmap_params *mp __maybe_unused,
805 struct perf_evlist *evlist __maybe_unused,
806 int idx __maybe_unused,
807 bool per_cpu __maybe_unused)
808{
809}
810
811static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
812{
813 if (evlist->mmap[idx].base != NULL) {
814 munmap(evlist->mmap[idx].base, evlist->mmap_len);
815 evlist->mmap[idx].base = NULL;
816 atomic_set(&evlist->mmap[idx].refcnt, 0);
817 }
818 auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
819}
820
821void perf_evlist__munmap(struct perf_evlist *evlist)
822{
823 int i;
824
825 if (evlist->mmap == NULL)
826 return;
827
828 for (i = 0; i < evlist->nr_mmaps; i++)
829 __perf_evlist__munmap(evlist, i);
830
831 zfree(&evlist->mmap);
832}
833
834static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
835{
836 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
837 if (cpu_map__empty(evlist->cpus))
838 evlist->nr_mmaps = thread_map__nr(evlist->threads);
839 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
840 return evlist->mmap != NULL ? 0 : -ENOMEM;
841}
842
843struct mmap_params {
844 int prot;
845 int mask;
846 struct auxtrace_mmap_params auxtrace_mp;
847};
848
849static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
850 struct mmap_params *mp, int fd)
851{
852 /*
853 * The last one will be done at perf_evlist__mmap_consume(), so that we
854 * make sure we don't prevent tools from consuming every last event in
855 * the ring buffer.
856 *
857 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
858 * anymore, but the last events for it are still in the ring buffer,
859 * waiting to be consumed.
860 *
861 * Tools can chose to ignore this at their own discretion, but the
862 * evlist layer can't just drop it when filtering events in
863 * perf_evlist__filter_pollfd().
864 */
865 atomic_set(&evlist->mmap[idx].refcnt, 2);
866 evlist->mmap[idx].prev = 0;
867 evlist->mmap[idx].mask = mp->mask;
868 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
869 MAP_SHARED, fd, 0);
870 if (evlist->mmap[idx].base == MAP_FAILED) {
871 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
872 errno);
873 evlist->mmap[idx].base = NULL;
874 return -1;
875 }
876
877 if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
878 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
879 return -1;
880
881 return 0;
882}
883
884static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
885 struct mmap_params *mp, int cpu,
886 int thread, int *output)
887{
888 struct perf_evsel *evsel;
889
890 evlist__for_each(evlist, evsel) {
891 int fd;
892
893 if (evsel->system_wide && thread)
894 continue;
895
896 fd = FD(evsel, cpu, thread);
897
898 if (*output == -1) {
899 *output = fd;
900 if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
901 return -1;
902 } else {
903 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
904 return -1;
905
906 perf_evlist__mmap_get(evlist, idx);
907 }
908
909 /*
910 * The system_wide flag causes a selected event to be opened
911 * always without a pid. Consequently it will never get a
912 * POLLHUP, but it is used for tracking in combination with
913 * other events, so it should not need to be polled anyway.
914 * Therefore don't add it for polling.
915 */
916 if (!evsel->system_wide &&
917 __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
918 perf_evlist__mmap_put(evlist, idx);
919 return -1;
920 }
921
922 if (evsel->attr.read_format & PERF_FORMAT_ID) {
923 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
924 fd) < 0)
925 return -1;
926 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
927 thread);
928 }
929 }
930
931 return 0;
932}
933
934static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
935 struct mmap_params *mp)
936{
937 int cpu, thread;
938 int nr_cpus = cpu_map__nr(evlist->cpus);
939 int nr_threads = thread_map__nr(evlist->threads);
940
941 pr_debug2("perf event ring buffer mmapped per cpu\n");
942 for (cpu = 0; cpu < nr_cpus; cpu++) {
943 int output = -1;
944
945 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
946 true);
947
948 for (thread = 0; thread < nr_threads; thread++) {
949 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
950 thread, &output))
951 goto out_unmap;
952 }
953 }
954
955 return 0;
956
957out_unmap:
958 for (cpu = 0; cpu < nr_cpus; cpu++)
959 __perf_evlist__munmap(evlist, cpu);
960 return -1;
961}
962
963static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
964 struct mmap_params *mp)
965{
966 int thread;
967 int nr_threads = thread_map__nr(evlist->threads);
968
969 pr_debug2("perf event ring buffer mmapped per thread\n");
970 for (thread = 0; thread < nr_threads; thread++) {
971 int output = -1;
972
973 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
974 false);
975
976 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
977 &output))
978 goto out_unmap;
979 }
980
981 return 0;
982
983out_unmap:
984 for (thread = 0; thread < nr_threads; thread++)
985 __perf_evlist__munmap(evlist, thread);
986 return -1;
987}
988
989static size_t perf_evlist__mmap_size(unsigned long pages)
990{
991 if (pages == UINT_MAX) {
992 int max;
993
994 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
995 /*
996 * Pick a once upon a time good value, i.e. things look
997 * strange since we can't read a sysctl value, but lets not
998 * die yet...
999 */
1000 max = 512;
1001 } else {
1002 max -= (page_size / 1024);
1003 }
1004
1005 pages = (max * 1024) / page_size;
1006 if (!is_power_of_2(pages))
1007 pages = rounddown_pow_of_two(pages);
1008 } else if (!is_power_of_2(pages))
1009 return 0;
1010
1011 return (pages + 1) * page_size;
1012}
1013
1014static long parse_pages_arg(const char *str, unsigned long min,
1015 unsigned long max)
1016{
1017 unsigned long pages, val;
1018 static struct parse_tag tags[] = {
1019 { .tag = 'B', .mult = 1 },
1020 { .tag = 'K', .mult = 1 << 10 },
1021 { .tag = 'M', .mult = 1 << 20 },
1022 { .tag = 'G', .mult = 1 << 30 },
1023 { .tag = 0 },
1024 };
1025
1026 if (str == NULL)
1027 return -EINVAL;
1028
1029 val = parse_tag_value(str, tags);
1030 if (val != (unsigned long) -1) {
1031 /* we got file size value */
1032 pages = PERF_ALIGN(val, page_size) / page_size;
1033 } else {
1034 /* we got pages count value */
1035 char *eptr;
1036 pages = strtoul(str, &eptr, 10);
1037 if (*eptr != '\0')
1038 return -EINVAL;
1039 }
1040
1041 if (pages == 0 && min == 0) {
1042 /* leave number of pages at 0 */
1043 } else if (!is_power_of_2(pages)) {
1044 /* round pages up to next power of 2 */
1045 pages = roundup_pow_of_two(pages);
1046 if (!pages)
1047 return -EINVAL;
1048 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1049 pages * page_size, pages);
1050 }
1051
1052 if (pages > max)
1053 return -EINVAL;
1054
1055 return pages;
1056}
1057
1058int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1059{
1060 unsigned long max = UINT_MAX;
1061 long pages;
1062
1063 if (max > SIZE_MAX / page_size)
1064 max = SIZE_MAX / page_size;
1065
1066 pages = parse_pages_arg(str, 1, max);
1067 if (pages < 0) {
1068 pr_err("Invalid argument for --mmap_pages/-m\n");
1069 return -1;
1070 }
1071
1072 *mmap_pages = pages;
1073 return 0;
1074}
1075
1076int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1077 int unset __maybe_unused)
1078{
1079 return __perf_evlist__parse_mmap_pages(opt->value, str);
1080}
1081
1082/**
1083 * perf_evlist__mmap_ex - Create mmaps to receive events.
1084 * @evlist: list of events
1085 * @pages: map length in pages
1086 * @overwrite: overwrite older events?
1087 * @auxtrace_pages - auxtrace map length in pages
1088 * @auxtrace_overwrite - overwrite older auxtrace data?
1089 *
1090 * If @overwrite is %false the user needs to signal event consumption using
1091 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1092 * automatically.
1093 *
1094 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1095 * consumption using auxtrace_mmap__write_tail().
1096 *
1097 * Return: %0 on success, negative error code otherwise.
1098 */
1099int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1100 bool overwrite, unsigned int auxtrace_pages,
1101 bool auxtrace_overwrite)
1102{
1103 struct perf_evsel *evsel;
1104 const struct cpu_map *cpus = evlist->cpus;
1105 const struct thread_map *threads = evlist->threads;
1106 struct mmap_params mp = {
1107 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1108 };
1109
1110 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1111 return -ENOMEM;
1112
1113 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1114 return -ENOMEM;
1115
1116 evlist->overwrite = overwrite;
1117 evlist->mmap_len = perf_evlist__mmap_size(pages);
1118 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1119 mp.mask = evlist->mmap_len - page_size - 1;
1120
1121 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1122 auxtrace_pages, auxtrace_overwrite);
1123
1124 evlist__for_each(evlist, evsel) {
1125 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1126 evsel->sample_id == NULL &&
1127 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1128 return -ENOMEM;
1129 }
1130
1131 if (cpu_map__empty(cpus))
1132 return perf_evlist__mmap_per_thread(evlist, &mp);
1133
1134 return perf_evlist__mmap_per_cpu(evlist, &mp);
1135}
1136
1137int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1138 bool overwrite)
1139{
1140 return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1141}
1142
1143int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1144{
1145 struct cpu_map *cpus;
1146 struct thread_map *threads;
1147
1148 threads = thread_map__new_str(target->pid, target->tid, target->uid);
1149
1150 if (!threads)
1151 return -1;
1152
1153 if (target__uses_dummy_map(target))
1154 cpus = cpu_map__dummy_new();
1155 else
1156 cpus = cpu_map__new(target->cpu_list);
1157
1158 if (!cpus)
1159 goto out_delete_threads;
1160
1161 evlist->has_user_cpus = !!target->cpu_list;
1162
1163 perf_evlist__set_maps(evlist, cpus, threads);
1164
1165 return 0;
1166
1167out_delete_threads:
1168 thread_map__put(threads);
1169 return -1;
1170}
1171
1172void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1173 struct thread_map *threads)
1174{
1175 /*
1176 * Allow for the possibility that one or another of the maps isn't being
1177 * changed i.e. don't put it. Note we are assuming the maps that are
1178 * being applied are brand new and evlist is taking ownership of the
1179 * original reference count of 1. If that is not the case it is up to
1180 * the caller to increase the reference count.
1181 */
1182 if (cpus != evlist->cpus) {
1183 cpu_map__put(evlist->cpus);
1184 evlist->cpus = cpu_map__get(cpus);
1185 }
1186
1187 if (threads != evlist->threads) {
1188 thread_map__put(evlist->threads);
1189 evlist->threads = thread_map__get(threads);
1190 }
1191
1192 perf_evlist__propagate_maps(evlist);
1193}
1194
1195int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1196{
1197 struct perf_evsel *evsel;
1198 int err = 0;
1199 const int ncpus = cpu_map__nr(evlist->cpus),
1200 nthreads = thread_map__nr(evlist->threads);
1201
1202 evlist__for_each(evlist, evsel) {
1203 if (evsel->filter == NULL)
1204 continue;
1205
1206 /*
1207 * filters only work for tracepoint event, which doesn't have cpu limit.
1208 * So evlist and evsel should always be same.
1209 */
1210 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1211 if (err) {
1212 *err_evsel = evsel;
1213 break;
1214 }
1215 }
1216
1217 return err;
1218}
1219
1220int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1221{
1222 struct perf_evsel *evsel;
1223 int err = 0;
1224
1225 evlist__for_each(evlist, evsel) {
1226 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1227 continue;
1228
1229 err = perf_evsel__set_filter(evsel, filter);
1230 if (err)
1231 break;
1232 }
1233
1234 return err;
1235}
1236
1237int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1238{
1239 char *filter;
1240 int ret = -1;
1241 size_t i;
1242
1243 for (i = 0; i < npids; ++i) {
1244 if (i == 0) {
1245 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1246 return -1;
1247 } else {
1248 char *tmp;
1249
1250 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1251 goto out_free;
1252
1253 free(filter);
1254 filter = tmp;
1255 }
1256 }
1257
1258 ret = perf_evlist__set_filter(evlist, filter);
1259out_free:
1260 free(filter);
1261 return ret;
1262}
1263
1264int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1265{
1266 return perf_evlist__set_filter_pids(evlist, 1, &pid);
1267}
1268
1269bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1270{
1271 struct perf_evsel *pos;
1272
1273 if (evlist->nr_entries == 1)
1274 return true;
1275
1276 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1277 return false;
1278
1279 evlist__for_each(evlist, pos) {
1280 if (pos->id_pos != evlist->id_pos ||
1281 pos->is_pos != evlist->is_pos)
1282 return false;
1283 }
1284
1285 return true;
1286}
1287
1288u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1289{
1290 struct perf_evsel *evsel;
1291
1292 if (evlist->combined_sample_type)
1293 return evlist->combined_sample_type;
1294
1295 evlist__for_each(evlist, evsel)
1296 evlist->combined_sample_type |= evsel->attr.sample_type;
1297
1298 return evlist->combined_sample_type;
1299}
1300
1301u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1302{
1303 evlist->combined_sample_type = 0;
1304 return __perf_evlist__combined_sample_type(evlist);
1305}
1306
1307u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1308{
1309 struct perf_evsel *evsel;
1310 u64 branch_type = 0;
1311
1312 evlist__for_each(evlist, evsel)
1313 branch_type |= evsel->attr.branch_sample_type;
1314 return branch_type;
1315}
1316
1317bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1318{
1319 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1320 u64 read_format = first->attr.read_format;
1321 u64 sample_type = first->attr.sample_type;
1322
1323 evlist__for_each(evlist, pos) {
1324 if (read_format != pos->attr.read_format)
1325 return false;
1326 }
1327
1328 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1329 if ((sample_type & PERF_SAMPLE_READ) &&
1330 !(read_format & PERF_FORMAT_ID)) {
1331 return false;
1332 }
1333
1334 return true;
1335}
1336
1337u64 perf_evlist__read_format(struct perf_evlist *evlist)
1338{
1339 struct perf_evsel *first = perf_evlist__first(evlist);
1340 return first->attr.read_format;
1341}
1342
1343u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1344{
1345 struct perf_evsel *first = perf_evlist__first(evlist);
1346 struct perf_sample *data;
1347 u64 sample_type;
1348 u16 size = 0;
1349
1350 if (!first->attr.sample_id_all)
1351 goto out;
1352
1353 sample_type = first->attr.sample_type;
1354
1355 if (sample_type & PERF_SAMPLE_TID)
1356 size += sizeof(data->tid) * 2;
1357
1358 if (sample_type & PERF_SAMPLE_TIME)
1359 size += sizeof(data->time);
1360
1361 if (sample_type & PERF_SAMPLE_ID)
1362 size += sizeof(data->id);
1363
1364 if (sample_type & PERF_SAMPLE_STREAM_ID)
1365 size += sizeof(data->stream_id);
1366
1367 if (sample_type & PERF_SAMPLE_CPU)
1368 size += sizeof(data->cpu) * 2;
1369
1370 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1371 size += sizeof(data->id);
1372out:
1373 return size;
1374}
1375
1376bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1377{
1378 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1379
1380 evlist__for_each_continue(evlist, pos) {
1381 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1382 return false;
1383 }
1384
1385 return true;
1386}
1387
1388bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1389{
1390 struct perf_evsel *first = perf_evlist__first(evlist);
1391 return first->attr.sample_id_all;
1392}
1393
1394void perf_evlist__set_selected(struct perf_evlist *evlist,
1395 struct perf_evsel *evsel)
1396{
1397 evlist->selected = evsel;
1398}
1399
1400void perf_evlist__close(struct perf_evlist *evlist)
1401{
1402 struct perf_evsel *evsel;
1403 int ncpus = cpu_map__nr(evlist->cpus);
1404 int nthreads = thread_map__nr(evlist->threads);
1405 int n;
1406
1407 evlist__for_each_reverse(evlist, evsel) {
1408 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1409 perf_evsel__close(evsel, n, nthreads);
1410 }
1411}
1412
1413static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1414{
1415 struct cpu_map *cpus;
1416 struct thread_map *threads;
1417 int err = -ENOMEM;
1418
1419 /*
1420 * Try reading /sys/devices/system/cpu/online to get
1421 * an all cpus map.
1422 *
1423 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1424 * code needs an overhaul to properly forward the
1425 * error, and we may not want to do that fallback to a
1426 * default cpu identity map :-\
1427 */
1428 cpus = cpu_map__new(NULL);
1429 if (!cpus)
1430 goto out;
1431
1432 threads = thread_map__new_dummy();
1433 if (!threads)
1434 goto out_put;
1435
1436 perf_evlist__set_maps(evlist, cpus, threads);
1437out:
1438 return err;
1439out_put:
1440 cpu_map__put(cpus);
1441 goto out;
1442}
1443
1444int perf_evlist__open(struct perf_evlist *evlist)
1445{
1446 struct perf_evsel *evsel;
1447 int err;
1448
1449 /*
1450 * Default: one fd per CPU, all threads, aka systemwide
1451 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1452 */
1453 if (evlist->threads == NULL && evlist->cpus == NULL) {
1454 err = perf_evlist__create_syswide_maps(evlist);
1455 if (err < 0)
1456 goto out_err;
1457 }
1458
1459 perf_evlist__update_id_pos(evlist);
1460
1461 evlist__for_each(evlist, evsel) {
1462 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1463 if (err < 0)
1464 goto out_err;
1465 }
1466
1467 return 0;
1468out_err:
1469 perf_evlist__close(evlist);
1470 errno = -err;
1471 return err;
1472}
1473
1474int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1475 const char *argv[], bool pipe_output,
1476 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1477{
1478 int child_ready_pipe[2], go_pipe[2];
1479 char bf;
1480
1481 if (pipe(child_ready_pipe) < 0) {
1482 perror("failed to create 'ready' pipe");
1483 return -1;
1484 }
1485
1486 if (pipe(go_pipe) < 0) {
1487 perror("failed to create 'go' pipe");
1488 goto out_close_ready_pipe;
1489 }
1490
1491 evlist->workload.pid = fork();
1492 if (evlist->workload.pid < 0) {
1493 perror("failed to fork");
1494 goto out_close_pipes;
1495 }
1496
1497 if (!evlist->workload.pid) {
1498 int ret;
1499
1500 if (pipe_output)
1501 dup2(2, 1);
1502
1503 signal(SIGTERM, SIG_DFL);
1504
1505 close(child_ready_pipe[0]);
1506 close(go_pipe[1]);
1507 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1508
1509 /*
1510 * Tell the parent we're ready to go
1511 */
1512 close(child_ready_pipe[1]);
1513
1514 /*
1515 * Wait until the parent tells us to go.
1516 */
1517 ret = read(go_pipe[0], &bf, 1);
1518 /*
1519 * The parent will ask for the execvp() to be performed by
1520 * writing exactly one byte, in workload.cork_fd, usually via
1521 * perf_evlist__start_workload().
1522 *
1523 * For cancelling the workload without actually running it,
1524 * the parent will just close workload.cork_fd, without writing
1525 * anything, i.e. read will return zero and we just exit()
1526 * here.
1527 */
1528 if (ret != 1) {
1529 if (ret == -1)
1530 perror("unable to read pipe");
1531 exit(ret);
1532 }
1533
1534 execvp(argv[0], (char **)argv);
1535
1536 if (exec_error) {
1537 union sigval val;
1538
1539 val.sival_int = errno;
1540 if (sigqueue(getppid(), SIGUSR1, val))
1541 perror(argv[0]);
1542 } else
1543 perror(argv[0]);
1544 exit(-1);
1545 }
1546
1547 if (exec_error) {
1548 struct sigaction act = {
1549 .sa_flags = SA_SIGINFO,
1550 .sa_sigaction = exec_error,
1551 };
1552 sigaction(SIGUSR1, &act, NULL);
1553 }
1554
1555 if (target__none(target)) {
1556 if (evlist->threads == NULL) {
1557 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1558 __func__, __LINE__);
1559 goto out_close_pipes;
1560 }
1561 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1562 }
1563
1564 close(child_ready_pipe[1]);
1565 close(go_pipe[0]);
1566 /*
1567 * wait for child to settle
1568 */
1569 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1570 perror("unable to read pipe");
1571 goto out_close_pipes;
1572 }
1573
1574 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1575 evlist->workload.cork_fd = go_pipe[1];
1576 close(child_ready_pipe[0]);
1577 return 0;
1578
1579out_close_pipes:
1580 close(go_pipe[0]);
1581 close(go_pipe[1]);
1582out_close_ready_pipe:
1583 close(child_ready_pipe[0]);
1584 close(child_ready_pipe[1]);
1585 return -1;
1586}
1587
1588int perf_evlist__start_workload(struct perf_evlist *evlist)
1589{
1590 if (evlist->workload.cork_fd > 0) {
1591 char bf = 0;
1592 int ret;
1593 /*
1594 * Remove the cork, let it rip!
1595 */
1596 ret = write(evlist->workload.cork_fd, &bf, 1);
1597 if (ret < 0)
1598 perror("enable to write to pipe");
1599
1600 close(evlist->workload.cork_fd);
1601 return ret;
1602 }
1603
1604 return 0;
1605}
1606
1607int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1608 struct perf_sample *sample)
1609{
1610 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1611
1612 if (!evsel)
1613 return -EFAULT;
1614 return perf_evsel__parse_sample(evsel, event, sample);
1615}
1616
1617size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1618{
1619 struct perf_evsel *evsel;
1620 size_t printed = 0;
1621
1622 evlist__for_each(evlist, evsel) {
1623 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1624 perf_evsel__name(evsel));
1625 }
1626
1627 return printed + fprintf(fp, "\n");
1628}
1629
1630int perf_evlist__strerror_open(struct perf_evlist *evlist,
1631 int err, char *buf, size_t size)
1632{
1633 int printed, value;
1634 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1635
1636 switch (err) {
1637 case EACCES:
1638 case EPERM:
1639 printed = scnprintf(buf, size,
1640 "Error:\t%s.\n"
1641 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1642
1643 value = perf_event_paranoid();
1644
1645 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1646
1647 if (value >= 2) {
1648 printed += scnprintf(buf + printed, size - printed,
1649 "For your workloads it needs to be <= 1\nHint:\t");
1650 }
1651 printed += scnprintf(buf + printed, size - printed,
1652 "For system wide tracing it needs to be set to -1.\n");
1653
1654 printed += scnprintf(buf + printed, size - printed,
1655 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1656 "Hint:\tThe current value is %d.", value);
1657 break;
1658 case EINVAL: {
1659 struct perf_evsel *first = perf_evlist__first(evlist);
1660 int max_freq;
1661
1662 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1663 goto out_default;
1664
1665 if (first->attr.sample_freq < (u64)max_freq)
1666 goto out_default;
1667
1668 printed = scnprintf(buf, size,
1669 "Error:\t%s.\n"
1670 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1671 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1672 emsg, max_freq, first->attr.sample_freq);
1673 break;
1674 }
1675 default:
1676out_default:
1677 scnprintf(buf, size, "%s", emsg);
1678 break;
1679 }
1680
1681 return 0;
1682}
1683
1684int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1685{
1686 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1687 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1688
1689 switch (err) {
1690 case EPERM:
1691 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1692 printed += scnprintf(buf + printed, size - printed,
1693 "Error:\t%s.\n"
1694 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1695 "Hint:\tTried using %zd kB.\n",
1696 emsg, pages_max_per_user, pages_attempted);
1697
1698 if (pages_attempted >= pages_max_per_user) {
1699 printed += scnprintf(buf + printed, size - printed,
1700 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1701 pages_max_per_user + pages_attempted);
1702 }
1703
1704 printed += scnprintf(buf + printed, size - printed,
1705 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1706 break;
1707 default:
1708 scnprintf(buf, size, "%s", emsg);
1709 break;
1710 }
1711
1712 return 0;
1713}
1714
1715void perf_evlist__to_front(struct perf_evlist *evlist,
1716 struct perf_evsel *move_evsel)
1717{
1718 struct perf_evsel *evsel, *n;
1719 LIST_HEAD(move);
1720
1721 if (move_evsel == perf_evlist__first(evlist))
1722 return;
1723
1724 evlist__for_each_safe(evlist, n, evsel) {
1725 if (evsel->leader == move_evsel->leader)
1726 list_move_tail(&evsel->node, &move);
1727 }
1728
1729 list_splice(&move, &evlist->entries);
1730}
1731
1732void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1733 struct perf_evsel *tracking_evsel)
1734{
1735 struct perf_evsel *evsel;
1736
1737 if (tracking_evsel->tracking)
1738 return;
1739
1740 evlist__for_each(evlist, evsel) {
1741 if (evsel != tracking_evsel)
1742 evsel->tracking = false;
1743 }
1744
1745 tracking_evsel->tracking = true;
1746}
1747
1748struct perf_evsel *
1749perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1750 const char *str)
1751{
1752 struct perf_evsel *evsel;
1753
1754 evlist__for_each(evlist, evsel) {
1755 if (!evsel->name)
1756 continue;
1757 if (strcmp(str, evsel->name) == 0)
1758 return evsel;
1759 }
1760
1761 return NULL;
1762}
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 "util/mmap.h"
14#include "thread_map.h"
15#include "target.h"
16#include "evlist.h"
17#include "evsel.h"
18#include "record.h"
19#include "debug.h"
20#include "units.h"
21#include "bpf_counter.h"
22#include <internal/lib.h> // page_size
23#include "affinity.h"
24#include "../perf.h"
25#include "asm/bug.h"
26#include "bpf-event.h"
27#include "util/event.h"
28#include "util/string2.h"
29#include "util/perf_api_probe.h"
30#include "util/evsel_fprintf.h"
31#include "util/pmu.h"
32#include "util/sample.h"
33#include "util/bpf-filter.h"
34#include "util/stat.h"
35#include "util/util.h"
36#include <signal.h>
37#include <unistd.h>
38#include <sched.h>
39#include <stdlib.h>
40
41#include "parse-events.h"
42#include <subcmd/parse-options.h>
43
44#include <fcntl.h>
45#include <sys/ioctl.h>
46#include <sys/mman.h>
47#include <sys/prctl.h>
48#include <sys/timerfd.h>
49
50#include <linux/bitops.h>
51#include <linux/hash.h>
52#include <linux/log2.h>
53#include <linux/err.h>
54#include <linux/string.h>
55#include <linux/time64.h>
56#include <linux/zalloc.h>
57#include <perf/evlist.h>
58#include <perf/evsel.h>
59#include <perf/cpumap.h>
60#include <perf/mmap.h>
61
62#include <internal/xyarray.h>
63
64#ifdef LACKS_SIGQUEUE_PROTOTYPE
65int sigqueue(pid_t pid, int sig, const union sigval value);
66#endif
67
68#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
69#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
70
71void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
72 struct perf_thread_map *threads)
73{
74 perf_evlist__init(&evlist->core);
75 perf_evlist__set_maps(&evlist->core, cpus, threads);
76 evlist->workload.pid = -1;
77 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
78 evlist->ctl_fd.fd = -1;
79 evlist->ctl_fd.ack = -1;
80 evlist->ctl_fd.pos = -1;
81}
82
83struct evlist *evlist__new(void)
84{
85 struct evlist *evlist = zalloc(sizeof(*evlist));
86
87 if (evlist != NULL)
88 evlist__init(evlist, NULL, NULL);
89
90 return evlist;
91}
92
93struct evlist *evlist__new_default(void)
94{
95 struct evlist *evlist = evlist__new();
96 bool can_profile_kernel;
97 int err;
98
99 if (!evlist)
100 return NULL;
101
102 can_profile_kernel = perf_event_paranoid_check(1);
103 err = parse_event(evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
104 if (err) {
105 evlist__delete(evlist);
106 return NULL;
107 }
108
109 if (evlist->core.nr_entries > 1) {
110 struct evsel *evsel;
111
112 evlist__for_each_entry(evlist, evsel)
113 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false);
114 }
115
116 return evlist;
117}
118
119struct evlist *evlist__new_dummy(void)
120{
121 struct evlist *evlist = evlist__new();
122
123 if (evlist && evlist__add_dummy(evlist)) {
124 evlist__delete(evlist);
125 evlist = NULL;
126 }
127
128 return evlist;
129}
130
131/**
132 * evlist__set_id_pos - set the positions of event ids.
133 * @evlist: selected event list
134 *
135 * Events with compatible sample types all have the same id_pos
136 * and is_pos. For convenience, put a copy on evlist.
137 */
138void evlist__set_id_pos(struct evlist *evlist)
139{
140 struct evsel *first = evlist__first(evlist);
141
142 evlist->id_pos = first->id_pos;
143 evlist->is_pos = first->is_pos;
144}
145
146static void evlist__update_id_pos(struct evlist *evlist)
147{
148 struct evsel *evsel;
149
150 evlist__for_each_entry(evlist, evsel)
151 evsel__calc_id_pos(evsel);
152
153 evlist__set_id_pos(evlist);
154}
155
156static void evlist__purge(struct evlist *evlist)
157{
158 struct evsel *pos, *n;
159
160 evlist__for_each_entry_safe(evlist, n, pos) {
161 list_del_init(&pos->core.node);
162 pos->evlist = NULL;
163 evsel__delete(pos);
164 }
165
166 evlist->core.nr_entries = 0;
167}
168
169void evlist__exit(struct evlist *evlist)
170{
171 event_enable_timer__exit(&evlist->eet);
172 zfree(&evlist->mmap);
173 zfree(&evlist->overwrite_mmap);
174 perf_evlist__exit(&evlist->core);
175}
176
177void evlist__delete(struct evlist *evlist)
178{
179 if (evlist == NULL)
180 return;
181
182 evlist__free_stats(evlist);
183 evlist__munmap(evlist);
184 evlist__close(evlist);
185 evlist__purge(evlist);
186 evlist__exit(evlist);
187 free(evlist);
188}
189
190void evlist__add(struct evlist *evlist, struct evsel *entry)
191{
192 perf_evlist__add(&evlist->core, &entry->core);
193 entry->evlist = evlist;
194 entry->tracking = !entry->core.idx;
195
196 if (evlist->core.nr_entries == 1)
197 evlist__set_id_pos(evlist);
198}
199
200void evlist__remove(struct evlist *evlist, struct evsel *evsel)
201{
202 evsel->evlist = NULL;
203 perf_evlist__remove(&evlist->core, &evsel->core);
204}
205
206void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
207{
208 while (!list_empty(list)) {
209 struct evsel *evsel, *temp, *leader = NULL;
210
211 __evlist__for_each_entry_safe(list, temp, evsel) {
212 list_del_init(&evsel->core.node);
213 evlist__add(evlist, evsel);
214 leader = evsel;
215 break;
216 }
217
218 __evlist__for_each_entry_safe(list, temp, evsel) {
219 if (evsel__has_leader(evsel, leader)) {
220 list_del_init(&evsel->core.node);
221 evlist__add(evlist, evsel);
222 }
223 }
224 }
225}
226
227int __evlist__set_tracepoints_handlers(struct evlist *evlist,
228 const struct evsel_str_handler *assocs, size_t nr_assocs)
229{
230 size_t i;
231 int err;
232
233 for (i = 0; i < nr_assocs; i++) {
234 // Adding a handler for an event not in this evlist, just ignore it.
235 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
236 if (evsel == NULL)
237 continue;
238
239 err = -EEXIST;
240 if (evsel->handler != NULL)
241 goto out;
242 evsel->handler = assocs[i].handler;
243 }
244
245 err = 0;
246out:
247 return err;
248}
249
250static void evlist__set_leader(struct evlist *evlist)
251{
252 perf_evlist__set_leader(&evlist->core);
253}
254
255static struct evsel *evlist__dummy_event(struct evlist *evlist)
256{
257 struct perf_event_attr attr = {
258 .type = PERF_TYPE_SOFTWARE,
259 .config = PERF_COUNT_SW_DUMMY,
260 .size = sizeof(attr), /* to capture ABI version */
261 /* Avoid frequency mode for dummy events to avoid associated timers. */
262 .freq = 0,
263 .sample_period = 1,
264 };
265
266 return evsel__new_idx(&attr, evlist->core.nr_entries);
267}
268
269int evlist__add_dummy(struct evlist *evlist)
270{
271 struct evsel *evsel = evlist__dummy_event(evlist);
272
273 if (evsel == NULL)
274 return -ENOMEM;
275
276 evlist__add(evlist, evsel);
277 return 0;
278}
279
280struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
281{
282 struct evsel *evsel = evlist__dummy_event(evlist);
283
284 if (!evsel)
285 return NULL;
286
287 evsel->core.attr.exclude_kernel = 1;
288 evsel->core.attr.exclude_guest = 1;
289 evsel->core.attr.exclude_hv = 1;
290 evsel->core.system_wide = system_wide;
291 evsel->no_aux_samples = true;
292 evsel->name = strdup("dummy:u");
293
294 evlist__add(evlist, evsel);
295 return evsel;
296}
297
298#ifdef HAVE_LIBTRACEEVENT
299struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
300{
301 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0);
302
303 if (IS_ERR(evsel))
304 return evsel;
305
306 evsel__set_sample_bit(evsel, CPU);
307 evsel__set_sample_bit(evsel, TIME);
308
309 evsel->core.system_wide = system_wide;
310 evsel->no_aux_samples = true;
311
312 evlist__add(evlist, evsel);
313 return evsel;
314}
315#endif
316
317int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
318{
319 struct evsel *evsel, *n;
320 LIST_HEAD(head);
321 size_t i;
322
323 for (i = 0; i < nr_attrs; i++) {
324 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
325 if (evsel == NULL)
326 goto out_delete_partial_list;
327 list_add_tail(&evsel->core.node, &head);
328 }
329
330 evlist__splice_list_tail(evlist, &head);
331
332 return 0;
333
334out_delete_partial_list:
335 __evlist__for_each_entry_safe(&head, n, evsel)
336 evsel__delete(evsel);
337 return -1;
338}
339
340int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
341{
342 size_t i;
343
344 for (i = 0; i < nr_attrs; i++)
345 event_attr_init(attrs + i);
346
347 return evlist__add_attrs(evlist, attrs, nr_attrs);
348}
349
350__weak int arch_evlist__add_default_attrs(struct evlist *evlist,
351 struct perf_event_attr *attrs,
352 size_t nr_attrs)
353{
354 if (!nr_attrs)
355 return 0;
356
357 return __evlist__add_default_attrs(evlist, attrs, nr_attrs);
358}
359
360struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
361{
362 struct evsel *evsel;
363
364 evlist__for_each_entry(evlist, evsel) {
365 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
366 (int)evsel->core.attr.config == id)
367 return evsel;
368 }
369
370 return NULL;
371}
372
373struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
374{
375 struct evsel *evsel;
376
377 evlist__for_each_entry(evlist, evsel) {
378 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
379 (strcmp(evsel->name, name) == 0))
380 return evsel;
381 }
382
383 return NULL;
384}
385
386#ifdef HAVE_LIBTRACEEVENT
387int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
388{
389 struct evsel *evsel = evsel__newtp(sys, name);
390
391 if (IS_ERR(evsel))
392 return -1;
393
394 evsel->handler = handler;
395 evlist__add(evlist, evsel);
396 return 0;
397}
398#endif
399
400struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
401{
402 struct evlist_cpu_iterator itr = {
403 .container = evlist,
404 .evsel = NULL,
405 .cpu_map_idx = 0,
406 .evlist_cpu_map_idx = 0,
407 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
408 .cpu = (struct perf_cpu){ .cpu = -1},
409 .affinity = affinity,
410 };
411
412 if (evlist__empty(evlist)) {
413 /* Ensure the empty list doesn't iterate. */
414 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
415 } else {
416 itr.evsel = evlist__first(evlist);
417 if (itr.affinity) {
418 itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
419 affinity__set(itr.affinity, itr.cpu.cpu);
420 itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
421 /*
422 * If this CPU isn't in the evsel's cpu map then advance
423 * through the list.
424 */
425 if (itr.cpu_map_idx == -1)
426 evlist_cpu_iterator__next(&itr);
427 }
428 }
429 return itr;
430}
431
432void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
433{
434 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
435 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
436 evlist_cpu_itr->cpu_map_idx =
437 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
438 evlist_cpu_itr->cpu);
439 if (evlist_cpu_itr->cpu_map_idx != -1)
440 return;
441 }
442 evlist_cpu_itr->evlist_cpu_map_idx++;
443 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
444 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
445 evlist_cpu_itr->cpu =
446 perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
447 evlist_cpu_itr->evlist_cpu_map_idx);
448 if (evlist_cpu_itr->affinity)
449 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
450 evlist_cpu_itr->cpu_map_idx =
451 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
452 evlist_cpu_itr->cpu);
453 /*
454 * If this CPU isn't in the evsel's cpu map then advance through
455 * the list.
456 */
457 if (evlist_cpu_itr->cpu_map_idx == -1)
458 evlist_cpu_iterator__next(evlist_cpu_itr);
459 }
460}
461
462bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
463{
464 return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
465}
466
467static int evsel__strcmp(struct evsel *pos, char *evsel_name)
468{
469 if (!evsel_name)
470 return 0;
471 if (evsel__is_dummy_event(pos))
472 return 1;
473 return !evsel__name_is(pos, evsel_name);
474}
475
476static int evlist__is_enabled(struct evlist *evlist)
477{
478 struct evsel *pos;
479
480 evlist__for_each_entry(evlist, pos) {
481 if (!evsel__is_group_leader(pos) || !pos->core.fd)
482 continue;
483 /* If at least one event is enabled, evlist is enabled. */
484 if (!pos->disabled)
485 return true;
486 }
487 return false;
488}
489
490static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
491{
492 struct evsel *pos;
493 struct evlist_cpu_iterator evlist_cpu_itr;
494 struct affinity saved_affinity, *affinity = NULL;
495 bool has_imm = false;
496
497 // See explanation in evlist__close()
498 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
499 if (affinity__setup(&saved_affinity) < 0)
500 return;
501 affinity = &saved_affinity;
502 }
503
504 /* Disable 'immediate' events last */
505 for (int imm = 0; imm <= 1; imm++) {
506 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
507 pos = evlist_cpu_itr.evsel;
508 if (evsel__strcmp(pos, evsel_name))
509 continue;
510 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
511 continue;
512 if (excl_dummy && evsel__is_dummy_event(pos))
513 continue;
514 if (pos->immediate)
515 has_imm = true;
516 if (pos->immediate != imm)
517 continue;
518 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
519 }
520 if (!has_imm)
521 break;
522 }
523
524 affinity__cleanup(affinity);
525 evlist__for_each_entry(evlist, pos) {
526 if (evsel__strcmp(pos, evsel_name))
527 continue;
528 if (!evsel__is_group_leader(pos) || !pos->core.fd)
529 continue;
530 if (excl_dummy && evsel__is_dummy_event(pos))
531 continue;
532 pos->disabled = true;
533 }
534
535 /*
536 * If we disabled only single event, we need to check
537 * the enabled state of the evlist manually.
538 */
539 if (evsel_name)
540 evlist->enabled = evlist__is_enabled(evlist);
541 else
542 evlist->enabled = false;
543}
544
545void evlist__disable(struct evlist *evlist)
546{
547 __evlist__disable(evlist, NULL, false);
548}
549
550void evlist__disable_non_dummy(struct evlist *evlist)
551{
552 __evlist__disable(evlist, NULL, true);
553}
554
555void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
556{
557 __evlist__disable(evlist, evsel_name, false);
558}
559
560static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
561{
562 struct evsel *pos;
563 struct evlist_cpu_iterator evlist_cpu_itr;
564 struct affinity saved_affinity, *affinity = NULL;
565
566 // See explanation in evlist__close()
567 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
568 if (affinity__setup(&saved_affinity) < 0)
569 return;
570 affinity = &saved_affinity;
571 }
572
573 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
574 pos = evlist_cpu_itr.evsel;
575 if (evsel__strcmp(pos, evsel_name))
576 continue;
577 if (!evsel__is_group_leader(pos) || !pos->core.fd)
578 continue;
579 if (excl_dummy && evsel__is_dummy_event(pos))
580 continue;
581 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
582 }
583 affinity__cleanup(affinity);
584 evlist__for_each_entry(evlist, pos) {
585 if (evsel__strcmp(pos, evsel_name))
586 continue;
587 if (!evsel__is_group_leader(pos) || !pos->core.fd)
588 continue;
589 if (excl_dummy && evsel__is_dummy_event(pos))
590 continue;
591 pos->disabled = false;
592 }
593
594 /*
595 * Even single event sets the 'enabled' for evlist,
596 * so the toggle can work properly and toggle to
597 * 'disabled' state.
598 */
599 evlist->enabled = true;
600}
601
602void evlist__enable(struct evlist *evlist)
603{
604 __evlist__enable(evlist, NULL, false);
605}
606
607void evlist__enable_non_dummy(struct evlist *evlist)
608{
609 __evlist__enable(evlist, NULL, true);
610}
611
612void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
613{
614 __evlist__enable(evlist, evsel_name, false);
615}
616
617void evlist__toggle_enable(struct evlist *evlist)
618{
619 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
620}
621
622int evlist__add_pollfd(struct evlist *evlist, int fd)
623{
624 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
625}
626
627int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
628{
629 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
630}
631
632#ifdef HAVE_EVENTFD_SUPPORT
633int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
634{
635 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
636 fdarray_flag__nonfilterable |
637 fdarray_flag__non_perf_event);
638}
639#endif
640
641int evlist__poll(struct evlist *evlist, int timeout)
642{
643 return perf_evlist__poll(&evlist->core, timeout);
644}
645
646struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
647{
648 struct hlist_head *head;
649 struct perf_sample_id *sid;
650 int hash;
651
652 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
653 head = &evlist->core.heads[hash];
654
655 hlist_for_each_entry(sid, head, node)
656 if (sid->id == id)
657 return sid;
658
659 return NULL;
660}
661
662struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
663{
664 struct perf_sample_id *sid;
665
666 if (evlist->core.nr_entries == 1 || !id)
667 return evlist__first(evlist);
668
669 sid = evlist__id2sid(evlist, id);
670 if (sid)
671 return container_of(sid->evsel, struct evsel, core);
672
673 if (!evlist__sample_id_all(evlist))
674 return evlist__first(evlist);
675
676 return NULL;
677}
678
679struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
680{
681 struct perf_sample_id *sid;
682
683 if (!id)
684 return NULL;
685
686 sid = evlist__id2sid(evlist, id);
687 if (sid)
688 return container_of(sid->evsel, struct evsel, core);
689
690 return NULL;
691}
692
693static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
694{
695 const __u64 *array = event->sample.array;
696 ssize_t n;
697
698 n = (event->header.size - sizeof(event->header)) >> 3;
699
700 if (event->header.type == PERF_RECORD_SAMPLE) {
701 if (evlist->id_pos >= n)
702 return -1;
703 *id = array[evlist->id_pos];
704 } else {
705 if (evlist->is_pos > n)
706 return -1;
707 n -= evlist->is_pos;
708 *id = array[n];
709 }
710 return 0;
711}
712
713struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
714{
715 struct evsel *first = evlist__first(evlist);
716 struct hlist_head *head;
717 struct perf_sample_id *sid;
718 int hash;
719 u64 id;
720
721 if (evlist->core.nr_entries == 1)
722 return first;
723
724 if (!first->core.attr.sample_id_all &&
725 event->header.type != PERF_RECORD_SAMPLE)
726 return first;
727
728 if (evlist__event2id(evlist, event, &id))
729 return NULL;
730
731 /* Synthesized events have an id of zero */
732 if (!id)
733 return first;
734
735 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
736 head = &evlist->core.heads[hash];
737
738 hlist_for_each_entry(sid, head, node) {
739 if (sid->id == id)
740 return container_of(sid->evsel, struct evsel, core);
741 }
742 return NULL;
743}
744
745static int evlist__set_paused(struct evlist *evlist, bool value)
746{
747 int i;
748
749 if (!evlist->overwrite_mmap)
750 return 0;
751
752 for (i = 0; i < evlist->core.nr_mmaps; i++) {
753 int fd = evlist->overwrite_mmap[i].core.fd;
754 int err;
755
756 if (fd < 0)
757 continue;
758 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
759 if (err)
760 return err;
761 }
762 return 0;
763}
764
765static int evlist__pause(struct evlist *evlist)
766{
767 return evlist__set_paused(evlist, true);
768}
769
770static int evlist__resume(struct evlist *evlist)
771{
772 return evlist__set_paused(evlist, false);
773}
774
775static void evlist__munmap_nofree(struct evlist *evlist)
776{
777 int i;
778
779 if (evlist->mmap)
780 for (i = 0; i < evlist->core.nr_mmaps; i++)
781 perf_mmap__munmap(&evlist->mmap[i].core);
782
783 if (evlist->overwrite_mmap)
784 for (i = 0; i < evlist->core.nr_mmaps; i++)
785 perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
786}
787
788void evlist__munmap(struct evlist *evlist)
789{
790 evlist__munmap_nofree(evlist);
791 zfree(&evlist->mmap);
792 zfree(&evlist->overwrite_mmap);
793}
794
795static void perf_mmap__unmap_cb(struct perf_mmap *map)
796{
797 struct mmap *m = container_of(map, struct mmap, core);
798
799 mmap__munmap(m);
800}
801
802static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
803 bool overwrite)
804{
805 int i;
806 struct mmap *map;
807
808 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
809 if (!map)
810 return NULL;
811
812 for (i = 0; i < evlist->core.nr_mmaps; i++) {
813 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
814
815 /*
816 * When the perf_mmap() call is made we grab one refcount, plus
817 * one extra to let perf_mmap__consume() get the last
818 * events after all real references (perf_mmap__get()) are
819 * dropped.
820 *
821 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
822 * thus does perf_mmap__get() on it.
823 */
824 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
825 }
826
827 return map;
828}
829
830static void
831perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
832 struct perf_evsel *_evsel,
833 struct perf_mmap_param *_mp,
834 int idx)
835{
836 struct evlist *evlist = container_of(_evlist, struct evlist, core);
837 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
838 struct evsel *evsel = container_of(_evsel, struct evsel, core);
839
840 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
841}
842
843static struct perf_mmap*
844perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
845{
846 struct evlist *evlist = container_of(_evlist, struct evlist, core);
847 struct mmap *maps;
848
849 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
850
851 if (!maps) {
852 maps = evlist__alloc_mmap(evlist, overwrite);
853 if (!maps)
854 return NULL;
855
856 if (overwrite) {
857 evlist->overwrite_mmap = maps;
858 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
859 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
860 } else {
861 evlist->mmap = maps;
862 }
863 }
864
865 return &maps[idx].core;
866}
867
868static int
869perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
870 int output, struct perf_cpu cpu)
871{
872 struct mmap *map = container_of(_map, struct mmap, core);
873 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
874
875 return mmap__mmap(map, mp, output, cpu);
876}
877
878unsigned long perf_event_mlock_kb_in_pages(void)
879{
880 unsigned long pages;
881 int max;
882
883 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
884 /*
885 * Pick a once upon a time good value, i.e. things look
886 * strange since we can't read a sysctl value, but lets not
887 * die yet...
888 */
889 max = 512;
890 } else {
891 max -= (page_size / 1024);
892 }
893
894 pages = (max * 1024) / page_size;
895 if (!is_power_of_2(pages))
896 pages = rounddown_pow_of_two(pages);
897
898 return pages;
899}
900
901size_t evlist__mmap_size(unsigned long pages)
902{
903 if (pages == UINT_MAX)
904 pages = perf_event_mlock_kb_in_pages();
905 else if (!is_power_of_2(pages))
906 return 0;
907
908 return (pages + 1) * page_size;
909}
910
911static long parse_pages_arg(const char *str, unsigned long min,
912 unsigned long max)
913{
914 unsigned long pages, val;
915 static struct parse_tag tags[] = {
916 { .tag = 'B', .mult = 1 },
917 { .tag = 'K', .mult = 1 << 10 },
918 { .tag = 'M', .mult = 1 << 20 },
919 { .tag = 'G', .mult = 1 << 30 },
920 { .tag = 0 },
921 };
922
923 if (str == NULL)
924 return -EINVAL;
925
926 val = parse_tag_value(str, tags);
927 if (val != (unsigned long) -1) {
928 /* we got file size value */
929 pages = PERF_ALIGN(val, page_size) / page_size;
930 } else {
931 /* we got pages count value */
932 char *eptr;
933 pages = strtoul(str, &eptr, 10);
934 if (*eptr != '\0')
935 return -EINVAL;
936 }
937
938 if (pages == 0 && min == 0) {
939 /* leave number of pages at 0 */
940 } else if (!is_power_of_2(pages)) {
941 char buf[100];
942
943 /* round pages up to next power of 2 */
944 pages = roundup_pow_of_two(pages);
945 if (!pages)
946 return -EINVAL;
947
948 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
949 pr_info("rounding mmap pages size to %s (%lu pages)\n",
950 buf, pages);
951 }
952
953 if (pages > max)
954 return -EINVAL;
955
956 return pages;
957}
958
959int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
960{
961 unsigned long max = UINT_MAX;
962 long pages;
963
964 if (max > SIZE_MAX / page_size)
965 max = SIZE_MAX / page_size;
966
967 pages = parse_pages_arg(str, 1, max);
968 if (pages < 0) {
969 pr_err("Invalid argument for --mmap_pages/-m\n");
970 return -1;
971 }
972
973 *mmap_pages = pages;
974 return 0;
975}
976
977int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
978{
979 return __evlist__parse_mmap_pages(opt->value, str);
980}
981
982/**
983 * evlist__mmap_ex - Create mmaps to receive events.
984 * @evlist: list of events
985 * @pages: map length in pages
986 * @overwrite: overwrite older events?
987 * @auxtrace_pages - auxtrace map length in pages
988 * @auxtrace_overwrite - overwrite older auxtrace data?
989 *
990 * If @overwrite is %false the user needs to signal event consumption using
991 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
992 * automatically.
993 *
994 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
995 * consumption using auxtrace_mmap__write_tail().
996 *
997 * Return: %0 on success, negative error code otherwise.
998 */
999int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
1000 unsigned int auxtrace_pages,
1001 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1002 int comp_level)
1003{
1004 /*
1005 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1006 * Its value is decided by evsel's write_backward.
1007 * So &mp should not be passed through const pointer.
1008 */
1009 struct mmap_params mp = {
1010 .nr_cblocks = nr_cblocks,
1011 .affinity = affinity,
1012 .flush = flush,
1013 .comp_level = comp_level
1014 };
1015 struct perf_evlist_mmap_ops ops = {
1016 .idx = perf_evlist__mmap_cb_idx,
1017 .get = perf_evlist__mmap_cb_get,
1018 .mmap = perf_evlist__mmap_cb_mmap,
1019 };
1020
1021 evlist->core.mmap_len = evlist__mmap_size(pages);
1022 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
1023
1024 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
1025 auxtrace_pages, auxtrace_overwrite);
1026
1027 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
1028}
1029
1030int evlist__mmap(struct evlist *evlist, unsigned int pages)
1031{
1032 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1033}
1034
1035int evlist__create_maps(struct evlist *evlist, struct target *target)
1036{
1037 bool all_threads = (target->per_thread && target->system_wide);
1038 struct perf_cpu_map *cpus;
1039 struct perf_thread_map *threads;
1040
1041 /*
1042 * If specify '-a' and '--per-thread' to perf record, perf record
1043 * will override '--per-thread'. target->per_thread = false and
1044 * target->system_wide = true.
1045 *
1046 * If specify '--per-thread' only to perf record,
1047 * target->per_thread = true and target->system_wide = false.
1048 *
1049 * So target->per_thread && target->system_wide is false.
1050 * For perf record, thread_map__new_str doesn't call
1051 * thread_map__new_all_cpus. That will keep perf record's
1052 * current behavior.
1053 *
1054 * For perf stat, it allows the case that target->per_thread and
1055 * target->system_wide are all true. It means to collect system-wide
1056 * per-thread data. thread_map__new_str will call
1057 * thread_map__new_all_cpus to enumerate all threads.
1058 */
1059 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1060 all_threads);
1061
1062 if (!threads)
1063 return -1;
1064
1065 if (target__uses_dummy_map(target))
1066 cpus = perf_cpu_map__new_any_cpu();
1067 else
1068 cpus = perf_cpu_map__new(target->cpu_list);
1069
1070 if (!cpus)
1071 goto out_delete_threads;
1072
1073 evlist->core.has_user_cpus = !!target->cpu_list;
1074
1075 perf_evlist__set_maps(&evlist->core, cpus, threads);
1076
1077 /* as evlist now has references, put count here */
1078 perf_cpu_map__put(cpus);
1079 perf_thread_map__put(threads);
1080
1081 return 0;
1082
1083out_delete_threads:
1084 perf_thread_map__put(threads);
1085 return -1;
1086}
1087
1088int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1089{
1090 struct evsel *evsel;
1091 int err = 0;
1092
1093 evlist__for_each_entry(evlist, evsel) {
1094 /*
1095 * filters only work for tracepoint event, which doesn't have cpu limit.
1096 * So evlist and evsel should always be same.
1097 */
1098 if (evsel->filter) {
1099 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1100 if (err) {
1101 *err_evsel = evsel;
1102 break;
1103 }
1104 }
1105
1106 /*
1107 * non-tracepoint events can have BPF filters.
1108 */
1109 if (!list_empty(&evsel->bpf_filters)) {
1110 err = perf_bpf_filter__prepare(evsel);
1111 if (err) {
1112 *err_evsel = evsel;
1113 break;
1114 }
1115 }
1116 }
1117
1118 return err;
1119}
1120
1121int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1122{
1123 struct evsel *evsel;
1124 int err = 0;
1125
1126 if (filter == NULL)
1127 return -1;
1128
1129 evlist__for_each_entry(evlist, evsel) {
1130 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1131 continue;
1132
1133 err = evsel__set_filter(evsel, filter);
1134 if (err)
1135 break;
1136 }
1137
1138 return err;
1139}
1140
1141int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1142{
1143 struct evsel *evsel;
1144 int err = 0;
1145
1146 if (filter == NULL)
1147 return -1;
1148
1149 evlist__for_each_entry(evlist, evsel) {
1150 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1151 continue;
1152
1153 err = evsel__append_tp_filter(evsel, filter);
1154 if (err)
1155 break;
1156 }
1157
1158 return err;
1159}
1160
1161char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1162{
1163 char *filter;
1164 size_t i;
1165
1166 for (i = 0; i < npids; ++i) {
1167 if (i == 0) {
1168 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1169 return NULL;
1170 } else {
1171 char *tmp;
1172
1173 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1174 goto out_free;
1175
1176 free(filter);
1177 filter = tmp;
1178 }
1179 }
1180
1181 return filter;
1182out_free:
1183 free(filter);
1184 return NULL;
1185}
1186
1187int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1188{
1189 char *filter = asprintf__tp_filter_pids(npids, pids);
1190 int ret = evlist__set_tp_filter(evlist, filter);
1191
1192 free(filter);
1193 return ret;
1194}
1195
1196int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1197{
1198 return evlist__set_tp_filter_pids(evlist, 1, &pid);
1199}
1200
1201int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1202{
1203 char *filter = asprintf__tp_filter_pids(npids, pids);
1204 int ret = evlist__append_tp_filter(evlist, filter);
1205
1206 free(filter);
1207 return ret;
1208}
1209
1210int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1211{
1212 return evlist__append_tp_filter_pids(evlist, 1, &pid);
1213}
1214
1215bool evlist__valid_sample_type(struct evlist *evlist)
1216{
1217 struct evsel *pos;
1218
1219 if (evlist->core.nr_entries == 1)
1220 return true;
1221
1222 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1223 return false;
1224
1225 evlist__for_each_entry(evlist, pos) {
1226 if (pos->id_pos != evlist->id_pos ||
1227 pos->is_pos != evlist->is_pos)
1228 return false;
1229 }
1230
1231 return true;
1232}
1233
1234u64 __evlist__combined_sample_type(struct evlist *evlist)
1235{
1236 struct evsel *evsel;
1237
1238 if (evlist->combined_sample_type)
1239 return evlist->combined_sample_type;
1240
1241 evlist__for_each_entry(evlist, evsel)
1242 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1243
1244 return evlist->combined_sample_type;
1245}
1246
1247u64 evlist__combined_sample_type(struct evlist *evlist)
1248{
1249 evlist->combined_sample_type = 0;
1250 return __evlist__combined_sample_type(evlist);
1251}
1252
1253u64 evlist__combined_branch_type(struct evlist *evlist)
1254{
1255 struct evsel *evsel;
1256 u64 branch_type = 0;
1257
1258 evlist__for_each_entry(evlist, evsel)
1259 branch_type |= evsel->core.attr.branch_sample_type;
1260 return branch_type;
1261}
1262
1263bool evlist__valid_read_format(struct evlist *evlist)
1264{
1265 struct evsel *first = evlist__first(evlist), *pos = first;
1266 u64 read_format = first->core.attr.read_format;
1267 u64 sample_type = first->core.attr.sample_type;
1268
1269 evlist__for_each_entry(evlist, pos) {
1270 if (read_format != pos->core.attr.read_format) {
1271 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1272 read_format, (u64)pos->core.attr.read_format);
1273 }
1274 }
1275
1276 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1277 if ((sample_type & PERF_SAMPLE_READ) &&
1278 !(read_format & PERF_FORMAT_ID)) {
1279 return false;
1280 }
1281
1282 return true;
1283}
1284
1285u16 evlist__id_hdr_size(struct evlist *evlist)
1286{
1287 struct evsel *first = evlist__first(evlist);
1288
1289 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1290}
1291
1292bool evlist__valid_sample_id_all(struct evlist *evlist)
1293{
1294 struct evsel *first = evlist__first(evlist), *pos = first;
1295
1296 evlist__for_each_entry_continue(evlist, pos) {
1297 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1298 return false;
1299 }
1300
1301 return true;
1302}
1303
1304bool evlist__sample_id_all(struct evlist *evlist)
1305{
1306 struct evsel *first = evlist__first(evlist);
1307 return first->core.attr.sample_id_all;
1308}
1309
1310void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1311{
1312 evlist->selected = evsel;
1313}
1314
1315void evlist__close(struct evlist *evlist)
1316{
1317 struct evsel *evsel;
1318 struct evlist_cpu_iterator evlist_cpu_itr;
1319 struct affinity affinity;
1320
1321 /*
1322 * With perf record core.user_requested_cpus is usually NULL.
1323 * Use the old method to handle this for now.
1324 */
1325 if (!evlist->core.user_requested_cpus ||
1326 cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1327 evlist__for_each_entry_reverse(evlist, evsel)
1328 evsel__close(evsel);
1329 return;
1330 }
1331
1332 if (affinity__setup(&affinity) < 0)
1333 return;
1334
1335 evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1336 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1337 evlist_cpu_itr.cpu_map_idx);
1338 }
1339
1340 affinity__cleanup(&affinity);
1341 evlist__for_each_entry_reverse(evlist, evsel) {
1342 perf_evsel__free_fd(&evsel->core);
1343 perf_evsel__free_id(&evsel->core);
1344 }
1345 perf_evlist__reset_id_hash(&evlist->core);
1346}
1347
1348static int evlist__create_syswide_maps(struct evlist *evlist)
1349{
1350 struct perf_cpu_map *cpus;
1351 struct perf_thread_map *threads;
1352
1353 /*
1354 * Try reading /sys/devices/system/cpu/online to get
1355 * an all cpus map.
1356 *
1357 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1358 * code needs an overhaul to properly forward the
1359 * error, and we may not want to do that fallback to a
1360 * default cpu identity map :-\
1361 */
1362 cpus = perf_cpu_map__new_online_cpus();
1363 if (!cpus)
1364 goto out;
1365
1366 threads = perf_thread_map__new_dummy();
1367 if (!threads)
1368 goto out_put;
1369
1370 perf_evlist__set_maps(&evlist->core, cpus, threads);
1371
1372 perf_thread_map__put(threads);
1373out_put:
1374 perf_cpu_map__put(cpus);
1375out:
1376 return -ENOMEM;
1377}
1378
1379int evlist__open(struct evlist *evlist)
1380{
1381 struct evsel *evsel;
1382 int err;
1383
1384 /*
1385 * Default: one fd per CPU, all threads, aka systemwide
1386 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1387 */
1388 if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1389 err = evlist__create_syswide_maps(evlist);
1390 if (err < 0)
1391 goto out_err;
1392 }
1393
1394 evlist__update_id_pos(evlist);
1395
1396 evlist__for_each_entry(evlist, evsel) {
1397 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1398 if (err < 0)
1399 goto out_err;
1400 }
1401
1402 return 0;
1403out_err:
1404 evlist__close(evlist);
1405 errno = -err;
1406 return err;
1407}
1408
1409int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1410 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1411{
1412 int child_ready_pipe[2], go_pipe[2];
1413 char bf;
1414
1415 if (pipe(child_ready_pipe) < 0) {
1416 perror("failed to create 'ready' pipe");
1417 return -1;
1418 }
1419
1420 if (pipe(go_pipe) < 0) {
1421 perror("failed to create 'go' pipe");
1422 goto out_close_ready_pipe;
1423 }
1424
1425 evlist->workload.pid = fork();
1426 if (evlist->workload.pid < 0) {
1427 perror("failed to fork");
1428 goto out_close_pipes;
1429 }
1430
1431 if (!evlist->workload.pid) {
1432 int ret;
1433
1434 if (pipe_output)
1435 dup2(2, 1);
1436
1437 signal(SIGTERM, SIG_DFL);
1438
1439 close(child_ready_pipe[0]);
1440 close(go_pipe[1]);
1441 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1442
1443 /*
1444 * Change the name of this process not to confuse --exclude-perf users
1445 * that sees 'perf' in the window up to the execvp() and thinks that
1446 * perf samples are not being excluded.
1447 */
1448 prctl(PR_SET_NAME, "perf-exec");
1449
1450 /*
1451 * Tell the parent we're ready to go
1452 */
1453 close(child_ready_pipe[1]);
1454
1455 /*
1456 * Wait until the parent tells us to go.
1457 */
1458 ret = read(go_pipe[0], &bf, 1);
1459 /*
1460 * The parent will ask for the execvp() to be performed by
1461 * writing exactly one byte, in workload.cork_fd, usually via
1462 * evlist__start_workload().
1463 *
1464 * For cancelling the workload without actually running it,
1465 * the parent will just close workload.cork_fd, without writing
1466 * anything, i.e. read will return zero and we just exit()
1467 * here.
1468 */
1469 if (ret != 1) {
1470 if (ret == -1)
1471 perror("unable to read pipe");
1472 exit(ret);
1473 }
1474
1475 execvp(argv[0], (char **)argv);
1476
1477 if (exec_error) {
1478 union sigval val;
1479
1480 val.sival_int = errno;
1481 if (sigqueue(getppid(), SIGUSR1, val))
1482 perror(argv[0]);
1483 } else
1484 perror(argv[0]);
1485 exit(-1);
1486 }
1487
1488 if (exec_error) {
1489 struct sigaction act = {
1490 .sa_flags = SA_SIGINFO,
1491 .sa_sigaction = exec_error,
1492 };
1493 sigaction(SIGUSR1, &act, NULL);
1494 }
1495
1496 if (target__none(target)) {
1497 if (evlist->core.threads == NULL) {
1498 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1499 __func__, __LINE__);
1500 goto out_close_pipes;
1501 }
1502 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1503 }
1504
1505 close(child_ready_pipe[1]);
1506 close(go_pipe[0]);
1507 /*
1508 * wait for child to settle
1509 */
1510 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1511 perror("unable to read pipe");
1512 goto out_close_pipes;
1513 }
1514
1515 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1516 evlist->workload.cork_fd = go_pipe[1];
1517 close(child_ready_pipe[0]);
1518 return 0;
1519
1520out_close_pipes:
1521 close(go_pipe[0]);
1522 close(go_pipe[1]);
1523out_close_ready_pipe:
1524 close(child_ready_pipe[0]);
1525 close(child_ready_pipe[1]);
1526 return -1;
1527}
1528
1529int evlist__start_workload(struct evlist *evlist)
1530{
1531 if (evlist->workload.cork_fd > 0) {
1532 char bf = 0;
1533 int ret;
1534 /*
1535 * Remove the cork, let it rip!
1536 */
1537 ret = write(evlist->workload.cork_fd, &bf, 1);
1538 if (ret < 0)
1539 perror("unable to write to pipe");
1540
1541 close(evlist->workload.cork_fd);
1542 return ret;
1543 }
1544
1545 return 0;
1546}
1547
1548int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1549{
1550 struct evsel *evsel = evlist__event2evsel(evlist, event);
1551 int ret;
1552
1553 if (!evsel)
1554 return -EFAULT;
1555 ret = evsel__parse_sample(evsel, event, sample);
1556 if (ret)
1557 return ret;
1558 if (perf_guest && sample->id) {
1559 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1560
1561 if (sid) {
1562 sample->machine_pid = sid->machine_pid;
1563 sample->vcpu = sid->vcpu.cpu;
1564 }
1565 }
1566 return 0;
1567}
1568
1569int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1570{
1571 struct evsel *evsel = evlist__event2evsel(evlist, event);
1572
1573 if (!evsel)
1574 return -EFAULT;
1575 return evsel__parse_sample_timestamp(evsel, event, timestamp);
1576}
1577
1578int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1579{
1580 int printed, value;
1581 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1582
1583 switch (err) {
1584 case EACCES:
1585 case EPERM:
1586 printed = scnprintf(buf, size,
1587 "Error:\t%s.\n"
1588 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1589
1590 value = perf_event_paranoid();
1591
1592 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1593
1594 if (value >= 2) {
1595 printed += scnprintf(buf + printed, size - printed,
1596 "For your workloads it needs to be <= 1\nHint:\t");
1597 }
1598 printed += scnprintf(buf + printed, size - printed,
1599 "For system wide tracing it needs to be set to -1.\n");
1600
1601 printed += scnprintf(buf + printed, size - printed,
1602 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1603 "Hint:\tThe current value is %d.", value);
1604 break;
1605 case EINVAL: {
1606 struct evsel *first = evlist__first(evlist);
1607 int max_freq;
1608
1609 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1610 goto out_default;
1611
1612 if (first->core.attr.sample_freq < (u64)max_freq)
1613 goto out_default;
1614
1615 printed = scnprintf(buf, size,
1616 "Error:\t%s.\n"
1617 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1618 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1619 emsg, max_freq, first->core.attr.sample_freq);
1620 break;
1621 }
1622 default:
1623out_default:
1624 scnprintf(buf, size, "%s", emsg);
1625 break;
1626 }
1627
1628 return 0;
1629}
1630
1631int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1632{
1633 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1634 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1635
1636 switch (err) {
1637 case EPERM:
1638 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1639 printed += scnprintf(buf + printed, size - printed,
1640 "Error:\t%s.\n"
1641 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1642 "Hint:\tTried using %zd kB.\n",
1643 emsg, pages_max_per_user, pages_attempted);
1644
1645 if (pages_attempted >= pages_max_per_user) {
1646 printed += scnprintf(buf + printed, size - printed,
1647 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1648 pages_max_per_user + pages_attempted);
1649 }
1650
1651 printed += scnprintf(buf + printed, size - printed,
1652 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1653 break;
1654 default:
1655 scnprintf(buf, size, "%s", emsg);
1656 break;
1657 }
1658
1659 return 0;
1660}
1661
1662void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1663{
1664 struct evsel *evsel, *n;
1665 LIST_HEAD(move);
1666
1667 if (move_evsel == evlist__first(evlist))
1668 return;
1669
1670 evlist__for_each_entry_safe(evlist, n, evsel) {
1671 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1672 list_move_tail(&evsel->core.node, &move);
1673 }
1674
1675 list_splice(&move, &evlist->core.entries);
1676}
1677
1678struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1679{
1680 struct evsel *evsel;
1681
1682 evlist__for_each_entry(evlist, evsel) {
1683 if (evsel->tracking)
1684 return evsel;
1685 }
1686
1687 return evlist__first(evlist);
1688}
1689
1690void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1691{
1692 struct evsel *evsel;
1693
1694 if (tracking_evsel->tracking)
1695 return;
1696
1697 evlist__for_each_entry(evlist, evsel) {
1698 if (evsel != tracking_evsel)
1699 evsel->tracking = false;
1700 }
1701
1702 tracking_evsel->tracking = true;
1703}
1704
1705struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide)
1706{
1707 struct evsel *evsel;
1708
1709 evsel = evlist__get_tracking_event(evlist);
1710 if (!evsel__is_dummy_event(evsel)) {
1711 evsel = evlist__add_aux_dummy(evlist, system_wide);
1712 if (!evsel)
1713 return NULL;
1714
1715 evlist__set_tracking_event(evlist, evsel);
1716 } else if (system_wide) {
1717 perf_evlist__go_system_wide(&evlist->core, &evsel->core);
1718 }
1719
1720 return evsel;
1721}
1722
1723struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1724{
1725 struct evsel *evsel;
1726
1727 evlist__for_each_entry(evlist, evsel) {
1728 if (!evsel->name)
1729 continue;
1730 if (evsel__name_is(evsel, str))
1731 return evsel;
1732 }
1733
1734 return NULL;
1735}
1736
1737void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1738{
1739 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1740 enum action {
1741 NONE,
1742 PAUSE,
1743 RESUME,
1744 } action = NONE;
1745
1746 if (!evlist->overwrite_mmap)
1747 return;
1748
1749 switch (old_state) {
1750 case BKW_MMAP_NOTREADY: {
1751 if (state != BKW_MMAP_RUNNING)
1752 goto state_err;
1753 break;
1754 }
1755 case BKW_MMAP_RUNNING: {
1756 if (state != BKW_MMAP_DATA_PENDING)
1757 goto state_err;
1758 action = PAUSE;
1759 break;
1760 }
1761 case BKW_MMAP_DATA_PENDING: {
1762 if (state != BKW_MMAP_EMPTY)
1763 goto state_err;
1764 break;
1765 }
1766 case BKW_MMAP_EMPTY: {
1767 if (state != BKW_MMAP_RUNNING)
1768 goto state_err;
1769 action = RESUME;
1770 break;
1771 }
1772 default:
1773 WARN_ONCE(1, "Shouldn't get there\n");
1774 }
1775
1776 evlist->bkw_mmap_state = state;
1777
1778 switch (action) {
1779 case PAUSE:
1780 evlist__pause(evlist);
1781 break;
1782 case RESUME:
1783 evlist__resume(evlist);
1784 break;
1785 case NONE:
1786 default:
1787 break;
1788 }
1789
1790state_err:
1791 return;
1792}
1793
1794bool evlist__exclude_kernel(struct evlist *evlist)
1795{
1796 struct evsel *evsel;
1797
1798 evlist__for_each_entry(evlist, evsel) {
1799 if (!evsel->core.attr.exclude_kernel)
1800 return false;
1801 }
1802
1803 return true;
1804}
1805
1806/*
1807 * Events in data file are not collect in groups, but we still want
1808 * the group display. Set the artificial group and set the leader's
1809 * forced_leader flag to notify the display code.
1810 */
1811void evlist__force_leader(struct evlist *evlist)
1812{
1813 if (evlist__nr_groups(evlist) == 0) {
1814 struct evsel *leader = evlist__first(evlist);
1815
1816 evlist__set_leader(evlist);
1817 leader->forced_leader = true;
1818 }
1819}
1820
1821struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1822{
1823 struct evsel *c2, *leader;
1824 bool is_open = true;
1825
1826 leader = evsel__leader(evsel);
1827
1828 pr_debug("Weak group for %s/%d failed\n",
1829 leader->name, leader->core.nr_members);
1830
1831 /*
1832 * for_each_group_member doesn't work here because it doesn't
1833 * include the first entry.
1834 */
1835 evlist__for_each_entry(evsel_list, c2) {
1836 if (c2 == evsel)
1837 is_open = false;
1838 if (evsel__has_leader(c2, leader)) {
1839 if (is_open && close)
1840 perf_evsel__close(&c2->core);
1841 /*
1842 * We want to close all members of the group and reopen
1843 * them. Some events, like Intel topdown, require being
1844 * in a group and so keep these in the group.
1845 */
1846 evsel__remove_from_group(c2, leader);
1847
1848 /*
1849 * Set this for all former members of the group
1850 * to indicate they get reopened.
1851 */
1852 c2->reset_group = true;
1853 }
1854 }
1855 /* Reset the leader count if all entries were removed. */
1856 if (leader->core.nr_members == 1)
1857 leader->core.nr_members = 0;
1858 return leader;
1859}
1860
1861static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1862{
1863 char *s, *p;
1864 int ret = 0, fd;
1865
1866 if (strncmp(str, "fifo:", 5))
1867 return -EINVAL;
1868
1869 str += 5;
1870 if (!*str || *str == ',')
1871 return -EINVAL;
1872
1873 s = strdup(str);
1874 if (!s)
1875 return -ENOMEM;
1876
1877 p = strchr(s, ',');
1878 if (p)
1879 *p = '\0';
1880
1881 /*
1882 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1883 * end of a FIFO to be repeatedly opened and closed.
1884 */
1885 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1886 if (fd < 0) {
1887 pr_err("Failed to open '%s'\n", s);
1888 ret = -errno;
1889 goto out_free;
1890 }
1891 *ctl_fd = fd;
1892 *ctl_fd_close = true;
1893
1894 if (p && *++p) {
1895 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1896 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1897 if (fd < 0) {
1898 pr_err("Failed to open '%s'\n", p);
1899 ret = -errno;
1900 goto out_free;
1901 }
1902 *ctl_fd_ack = fd;
1903 }
1904
1905out_free:
1906 free(s);
1907 return ret;
1908}
1909
1910int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1911{
1912 char *comma = NULL, *endptr = NULL;
1913
1914 *ctl_fd_close = false;
1915
1916 if (strncmp(str, "fd:", 3))
1917 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1918
1919 *ctl_fd = strtoul(&str[3], &endptr, 0);
1920 if (endptr == &str[3])
1921 return -EINVAL;
1922
1923 comma = strchr(str, ',');
1924 if (comma) {
1925 if (endptr != comma)
1926 return -EINVAL;
1927
1928 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1929 if (endptr == comma + 1 || *endptr != '\0')
1930 return -EINVAL;
1931 }
1932
1933 return 0;
1934}
1935
1936void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1937{
1938 if (*ctl_fd_close) {
1939 *ctl_fd_close = false;
1940 close(ctl_fd);
1941 if (ctl_fd_ack >= 0)
1942 close(ctl_fd_ack);
1943 }
1944}
1945
1946int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1947{
1948 if (fd == -1) {
1949 pr_debug("Control descriptor is not initialized\n");
1950 return 0;
1951 }
1952
1953 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1954 fdarray_flag__nonfilterable |
1955 fdarray_flag__non_perf_event);
1956 if (evlist->ctl_fd.pos < 0) {
1957 evlist->ctl_fd.pos = -1;
1958 pr_err("Failed to add ctl fd entry: %m\n");
1959 return -1;
1960 }
1961
1962 evlist->ctl_fd.fd = fd;
1963 evlist->ctl_fd.ack = ack;
1964
1965 return 0;
1966}
1967
1968bool evlist__ctlfd_initialized(struct evlist *evlist)
1969{
1970 return evlist->ctl_fd.pos >= 0;
1971}
1972
1973int evlist__finalize_ctlfd(struct evlist *evlist)
1974{
1975 struct pollfd *entries = evlist->core.pollfd.entries;
1976
1977 if (!evlist__ctlfd_initialized(evlist))
1978 return 0;
1979
1980 entries[evlist->ctl_fd.pos].fd = -1;
1981 entries[evlist->ctl_fd.pos].events = 0;
1982 entries[evlist->ctl_fd.pos].revents = 0;
1983
1984 evlist->ctl_fd.pos = -1;
1985 evlist->ctl_fd.ack = -1;
1986 evlist->ctl_fd.fd = -1;
1987
1988 return 0;
1989}
1990
1991static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1992 char *cmd_data, size_t data_size)
1993{
1994 int err;
1995 char c;
1996 size_t bytes_read = 0;
1997
1998 *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1999 memset(cmd_data, 0, data_size);
2000 data_size--;
2001
2002 do {
2003 err = read(evlist->ctl_fd.fd, &c, 1);
2004 if (err > 0) {
2005 if (c == '\n' || c == '\0')
2006 break;
2007 cmd_data[bytes_read++] = c;
2008 if (bytes_read == data_size)
2009 break;
2010 continue;
2011 } else if (err == -1) {
2012 if (errno == EINTR)
2013 continue;
2014 if (errno == EAGAIN || errno == EWOULDBLOCK)
2015 err = 0;
2016 else
2017 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
2018 }
2019 break;
2020 } while (1);
2021
2022 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
2023 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
2024
2025 if (bytes_read > 0) {
2026 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
2027 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
2028 *cmd = EVLIST_CTL_CMD_ENABLE;
2029 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
2030 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
2031 *cmd = EVLIST_CTL_CMD_DISABLE;
2032 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2033 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2034 *cmd = EVLIST_CTL_CMD_SNAPSHOT;
2035 pr_debug("is snapshot\n");
2036 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2037 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2038 *cmd = EVLIST_CTL_CMD_EVLIST;
2039 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2040 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2041 *cmd = EVLIST_CTL_CMD_STOP;
2042 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2043 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2044 *cmd = EVLIST_CTL_CMD_PING;
2045 }
2046 }
2047
2048 return bytes_read ? (int)bytes_read : err;
2049}
2050
2051int evlist__ctlfd_ack(struct evlist *evlist)
2052{
2053 int err;
2054
2055 if (evlist->ctl_fd.ack == -1)
2056 return 0;
2057
2058 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2059 sizeof(EVLIST_CTL_CMD_ACK_TAG));
2060 if (err == -1)
2061 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2062
2063 return err;
2064}
2065
2066static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2067{
2068 char *data = cmd_data + cmd_size;
2069
2070 /* no argument */
2071 if (!*data)
2072 return 0;
2073
2074 /* there's argument */
2075 if (*data == ' ') {
2076 *arg = data + 1;
2077 return 1;
2078 }
2079
2080 /* malformed */
2081 return -1;
2082}
2083
2084static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2085{
2086 struct evsel *evsel;
2087 char *name;
2088 int err;
2089
2090 err = get_cmd_arg(cmd_data,
2091 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2092 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2093 &name);
2094 if (err < 0) {
2095 pr_info("failed: wrong command\n");
2096 return -1;
2097 }
2098
2099 if (err) {
2100 evsel = evlist__find_evsel_by_str(evlist, name);
2101 if (evsel) {
2102 if (enable)
2103 evlist__enable_evsel(evlist, name);
2104 else
2105 evlist__disable_evsel(evlist, name);
2106 pr_info("Event %s %s\n", evsel->name,
2107 enable ? "enabled" : "disabled");
2108 } else {
2109 pr_info("failed: can't find '%s' event\n", name);
2110 }
2111 } else {
2112 if (enable) {
2113 evlist__enable(evlist);
2114 pr_info(EVLIST_ENABLED_MSG);
2115 } else {
2116 evlist__disable(evlist);
2117 pr_info(EVLIST_DISABLED_MSG);
2118 }
2119 }
2120
2121 return 0;
2122}
2123
2124static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2125{
2126 struct perf_attr_details details = { .verbose = false, };
2127 struct evsel *evsel;
2128 char *arg;
2129 int err;
2130
2131 err = get_cmd_arg(cmd_data,
2132 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2133 &arg);
2134 if (err < 0) {
2135 pr_info("failed: wrong command\n");
2136 return -1;
2137 }
2138
2139 if (err) {
2140 if (!strcmp(arg, "-v")) {
2141 details.verbose = true;
2142 } else if (!strcmp(arg, "-g")) {
2143 details.event_group = true;
2144 } else if (!strcmp(arg, "-F")) {
2145 details.freq = true;
2146 } else {
2147 pr_info("failed: wrong command\n");
2148 return -1;
2149 }
2150 }
2151
2152 evlist__for_each_entry(evlist, evsel)
2153 evsel__fprintf(evsel, &details, stderr);
2154
2155 return 0;
2156}
2157
2158int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2159{
2160 int err = 0;
2161 char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2162 int ctlfd_pos = evlist->ctl_fd.pos;
2163 struct pollfd *entries = evlist->core.pollfd.entries;
2164
2165 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2166 return 0;
2167
2168 if (entries[ctlfd_pos].revents & POLLIN) {
2169 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2170 EVLIST_CTL_CMD_MAX_LEN);
2171 if (err > 0) {
2172 switch (*cmd) {
2173 case EVLIST_CTL_CMD_ENABLE:
2174 case EVLIST_CTL_CMD_DISABLE:
2175 err = evlist__ctlfd_enable(evlist, cmd_data,
2176 *cmd == EVLIST_CTL_CMD_ENABLE);
2177 break;
2178 case EVLIST_CTL_CMD_EVLIST:
2179 err = evlist__ctlfd_list(evlist, cmd_data);
2180 break;
2181 case EVLIST_CTL_CMD_SNAPSHOT:
2182 case EVLIST_CTL_CMD_STOP:
2183 case EVLIST_CTL_CMD_PING:
2184 break;
2185 case EVLIST_CTL_CMD_ACK:
2186 case EVLIST_CTL_CMD_UNSUPPORTED:
2187 default:
2188 pr_debug("ctlfd: unsupported %d\n", *cmd);
2189 break;
2190 }
2191 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2192 *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2193 evlist__ctlfd_ack(evlist);
2194 }
2195 }
2196
2197 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2198 evlist__finalize_ctlfd(evlist);
2199 else
2200 entries[ctlfd_pos].revents = 0;
2201
2202 return err;
2203}
2204
2205/**
2206 * struct event_enable_time - perf record -D/--delay single time range.
2207 * @start: start of time range to enable events in milliseconds
2208 * @end: end of time range to enable events in milliseconds
2209 *
2210 * N.B. this structure is also accessed as an array of int.
2211 */
2212struct event_enable_time {
2213 int start;
2214 int end;
2215};
2216
2217static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2218{
2219 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2220 int ret, start, end, n;
2221
2222 ret = sscanf(str, fmt, &start, &end, &n);
2223 if (ret != 2 || end <= start)
2224 return -EINVAL;
2225 if (range) {
2226 range->start = start;
2227 range->end = end;
2228 }
2229 return n;
2230}
2231
2232static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2233{
2234 int incr = !!range;
2235 bool first = true;
2236 ssize_t ret, cnt;
2237
2238 for (cnt = 0; *str; cnt++) {
2239 ret = parse_event_enable_time(str, range, first);
2240 if (ret < 0)
2241 return ret;
2242 /* Check no overlap */
2243 if (!first && range && range->start <= range[-1].end)
2244 return -EINVAL;
2245 str += ret;
2246 range += incr;
2247 first = false;
2248 }
2249 return cnt;
2250}
2251
2252/**
2253 * struct event_enable_timer - control structure for perf record -D/--delay.
2254 * @evlist: event list
2255 * @times: time ranges that events are enabled (N.B. this is also accessed as an
2256 * array of int)
2257 * @times_cnt: number of time ranges
2258 * @timerfd: timer file descriptor
2259 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2260 * @times_step: current position in (int *)@times)[],
2261 * refer event_enable_timer__process()
2262 *
2263 * Note, this structure is only used when there are time ranges, not when there
2264 * is only an initial delay.
2265 */
2266struct event_enable_timer {
2267 struct evlist *evlist;
2268 struct event_enable_time *times;
2269 size_t times_cnt;
2270 int timerfd;
2271 int pollfd_pos;
2272 size_t times_step;
2273};
2274
2275static int str_to_delay(const char *str)
2276{
2277 char *endptr;
2278 long d;
2279
2280 d = strtol(str, &endptr, 10);
2281 if (*endptr || d > INT_MAX || d < -1)
2282 return 0;
2283 return d;
2284}
2285
2286int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2287 const char *str, int unset)
2288{
2289 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2290 struct event_enable_timer *eet;
2291 ssize_t times_cnt;
2292 ssize_t ret;
2293 int err;
2294
2295 if (unset)
2296 return 0;
2297
2298 opts->target.initial_delay = str_to_delay(str);
2299 if (opts->target.initial_delay)
2300 return 0;
2301
2302 ret = parse_event_enable_times(str, NULL);
2303 if (ret < 0)
2304 return ret;
2305
2306 times_cnt = ret;
2307 if (times_cnt == 0)
2308 return -EINVAL;
2309
2310 eet = zalloc(sizeof(*eet));
2311 if (!eet)
2312 return -ENOMEM;
2313
2314 eet->times = calloc(times_cnt, sizeof(*eet->times));
2315 if (!eet->times) {
2316 err = -ENOMEM;
2317 goto free_eet;
2318 }
2319
2320 if (parse_event_enable_times(str, eet->times) != times_cnt) {
2321 err = -EINVAL;
2322 goto free_eet_times;
2323 }
2324
2325 eet->times_cnt = times_cnt;
2326
2327 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2328 if (eet->timerfd == -1) {
2329 err = -errno;
2330 pr_err("timerfd_create failed: %s\n", strerror(errno));
2331 goto free_eet_times;
2332 }
2333
2334 eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2335 if (eet->pollfd_pos < 0) {
2336 err = eet->pollfd_pos;
2337 goto close_timerfd;
2338 }
2339
2340 eet->evlist = evlist;
2341 evlist->eet = eet;
2342 opts->target.initial_delay = eet->times[0].start;
2343
2344 return 0;
2345
2346close_timerfd:
2347 close(eet->timerfd);
2348free_eet_times:
2349 zfree(&eet->times);
2350free_eet:
2351 free(eet);
2352 return err;
2353}
2354
2355static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2356{
2357 struct itimerspec its = {
2358 .it_value.tv_sec = ms / MSEC_PER_SEC,
2359 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2360 };
2361 int err = 0;
2362
2363 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2364 err = -errno;
2365 pr_err("timerfd_settime failed: %s\n", strerror(errno));
2366 }
2367 return err;
2368}
2369
2370int event_enable_timer__start(struct event_enable_timer *eet)
2371{
2372 int ms;
2373
2374 if (!eet)
2375 return 0;
2376
2377 ms = eet->times[0].end - eet->times[0].start;
2378 eet->times_step = 1;
2379
2380 return event_enable_timer__set_timer(eet, ms);
2381}
2382
2383int event_enable_timer__process(struct event_enable_timer *eet)
2384{
2385 struct pollfd *entries;
2386 short revents;
2387
2388 if (!eet)
2389 return 0;
2390
2391 entries = eet->evlist->core.pollfd.entries;
2392 revents = entries[eet->pollfd_pos].revents;
2393 entries[eet->pollfd_pos].revents = 0;
2394
2395 if (revents & POLLIN) {
2396 size_t step = eet->times_step;
2397 size_t pos = step / 2;
2398
2399 if (step & 1) {
2400 evlist__disable_non_dummy(eet->evlist);
2401 pr_info(EVLIST_DISABLED_MSG);
2402 if (pos >= eet->times_cnt - 1) {
2403 /* Disarm timer */
2404 event_enable_timer__set_timer(eet, 0);
2405 return 1; /* Stop */
2406 }
2407 } else {
2408 evlist__enable_non_dummy(eet->evlist);
2409 pr_info(EVLIST_ENABLED_MSG);
2410 }
2411
2412 step += 1;
2413 pos = step / 2;
2414
2415 if (pos < eet->times_cnt) {
2416 int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2417 int ms = times[step] - times[step - 1];
2418
2419 eet->times_step = step;
2420 return event_enable_timer__set_timer(eet, ms);
2421 }
2422 }
2423
2424 return 0;
2425}
2426
2427void event_enable_timer__exit(struct event_enable_timer **ep)
2428{
2429 if (!ep || !*ep)
2430 return;
2431 zfree(&(*ep)->times);
2432 zfree(ep);
2433}
2434
2435struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2436{
2437 struct evsel *evsel;
2438
2439 evlist__for_each_entry(evlist, evsel) {
2440 if (evsel->core.idx == idx)
2441 return evsel;
2442 }
2443 return NULL;
2444}
2445
2446int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
2447{
2448 struct evsel *evsel;
2449 int printed = 0;
2450
2451 evlist__for_each_entry(evlist, evsel) {
2452 if (evsel__is_dummy_event(evsel))
2453 continue;
2454 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) {
2455 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel));
2456 } else {
2457 printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : "");
2458 break;
2459 }
2460 }
2461
2462 return printed;
2463}
2464
2465void evlist__check_mem_load_aux(struct evlist *evlist)
2466{
2467 struct evsel *leader, *evsel, *pos;
2468
2469 /*
2470 * For some platforms, the 'mem-loads' event is required to use
2471 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2472 * must be the group leader. Now we disable this group before reporting
2473 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2474 * any valid memory load information.
2475 */
2476 evlist__for_each_entry(evlist, evsel) {
2477 leader = evsel__leader(evsel);
2478 if (leader == evsel)
2479 continue;
2480
2481 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2482 for_each_group_evsel(pos, leader) {
2483 evsel__set_leader(pos, pos);
2484 pos->core.nr_members = 0;
2485 }
2486 }
2487 }
2488}
2489
2490/**
2491 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs
2492 * and warn if the user CPU list is inapplicable for the event's PMU's
2493 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a
2494 * user requested CPU and so any online CPU is applicable. Core PMUs handle
2495 * events on the CPUs in their list and otherwise the event isn't supported.
2496 * @evlist: The list of events being checked.
2497 * @cpu_list: The user provided list of CPUs.
2498 */
2499void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list)
2500{
2501 struct perf_cpu_map *user_requested_cpus;
2502 struct evsel *pos;
2503
2504 if (!cpu_list)
2505 return;
2506
2507 user_requested_cpus = perf_cpu_map__new(cpu_list);
2508 if (!user_requested_cpus)
2509 return;
2510
2511 evlist__for_each_entry(evlist, pos) {
2512 struct perf_cpu_map *intersect, *to_test;
2513 const struct perf_pmu *pmu = evsel__find_pmu(pos);
2514
2515 to_test = pmu && pmu->is_core ? pmu->cpus : cpu_map__online();
2516 intersect = perf_cpu_map__intersect(to_test, user_requested_cpus);
2517 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) {
2518 char buf[128];
2519
2520 cpu_map__snprint(to_test, buf, sizeof(buf));
2521 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n",
2522 cpu_list, pmu ? pmu->name : "cpu", buf, evsel__name(pos));
2523 }
2524 perf_cpu_map__put(intersect);
2525 }
2526 perf_cpu_map__put(user_requested_cpus);
2527}
2528
2529void evlist__uniquify_name(struct evlist *evlist)
2530{
2531 char *new_name, empty_attributes[2] = ":", *attributes;
2532 struct evsel *pos;
2533
2534 if (perf_pmus__num_core_pmus() == 1)
2535 return;
2536
2537 evlist__for_each_entry(evlist, pos) {
2538 if (!evsel__is_hybrid(pos))
2539 continue;
2540
2541 if (strchr(pos->name, '/'))
2542 continue;
2543
2544 attributes = strchr(pos->name, ':');
2545 if (attributes)
2546 *attributes = '\0';
2547 else
2548 attributes = empty_attributes;
2549
2550 if (asprintf(&new_name, "%s/%s/%s", pos->pmu_name, pos->name, attributes + 1)) {
2551 free(pos->name);
2552 pos->name = new_name;
2553 } else {
2554 *attributes = ':';
2555 }
2556 }
2557}