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 "util/env.h"
37#include "util/intel-tpebs.h"
38#include <signal.h>
39#include <unistd.h>
40#include <sched.h>
41#include <stdlib.h>
42
43#include "parse-events.h"
44#include <subcmd/parse-options.h>
45
46#include <fcntl.h>
47#include <sys/ioctl.h>
48#include <sys/mman.h>
49#include <sys/prctl.h>
50#include <sys/timerfd.h>
51#include <sys/wait.h>
52
53#include <linux/bitops.h>
54#include <linux/hash.h>
55#include <linux/log2.h>
56#include <linux/err.h>
57#include <linux/string.h>
58#include <linux/time64.h>
59#include <linux/zalloc.h>
60#include <perf/evlist.h>
61#include <perf/evsel.h>
62#include <perf/cpumap.h>
63#include <perf/mmap.h>
64
65#include <internal/xyarray.h>
66
67#ifdef LACKS_SIGQUEUE_PROTOTYPE
68int sigqueue(pid_t pid, int sig, const union sigval value);
69#endif
70
71#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
72#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
73
74void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
75 struct perf_thread_map *threads)
76{
77 perf_evlist__init(&evlist->core);
78 perf_evlist__set_maps(&evlist->core, cpus, threads);
79 evlist->workload.pid = -1;
80 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
81 evlist->ctl_fd.fd = -1;
82 evlist->ctl_fd.ack = -1;
83 evlist->ctl_fd.pos = -1;
84 evlist->nr_br_cntr = -1;
85}
86
87struct evlist *evlist__new(void)
88{
89 struct evlist *evlist = zalloc(sizeof(*evlist));
90
91 if (evlist != NULL)
92 evlist__init(evlist, NULL, NULL);
93
94 return evlist;
95}
96
97struct evlist *evlist__new_default(void)
98{
99 struct evlist *evlist = evlist__new();
100 bool can_profile_kernel;
101 int err;
102
103 if (!evlist)
104 return NULL;
105
106 can_profile_kernel = perf_event_paranoid_check(1);
107 err = parse_event(evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
108 if (err) {
109 evlist__delete(evlist);
110 return NULL;
111 }
112
113 if (evlist->core.nr_entries > 1) {
114 struct evsel *evsel;
115
116 evlist__for_each_entry(evlist, evsel)
117 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false);
118 }
119
120 return evlist;
121}
122
123struct evlist *evlist__new_dummy(void)
124{
125 struct evlist *evlist = evlist__new();
126
127 if (evlist && evlist__add_dummy(evlist)) {
128 evlist__delete(evlist);
129 evlist = NULL;
130 }
131
132 return evlist;
133}
134
135/**
136 * evlist__set_id_pos - set the positions of event ids.
137 * @evlist: selected event list
138 *
139 * Events with compatible sample types all have the same id_pos
140 * and is_pos. For convenience, put a copy on evlist.
141 */
142void evlist__set_id_pos(struct evlist *evlist)
143{
144 struct evsel *first = evlist__first(evlist);
145
146 evlist->id_pos = first->id_pos;
147 evlist->is_pos = first->is_pos;
148}
149
150static void evlist__update_id_pos(struct evlist *evlist)
151{
152 struct evsel *evsel;
153
154 evlist__for_each_entry(evlist, evsel)
155 evsel__calc_id_pos(evsel);
156
157 evlist__set_id_pos(evlist);
158}
159
160static void evlist__purge(struct evlist *evlist)
161{
162 struct evsel *pos, *n;
163
164 evlist__for_each_entry_safe(evlist, n, pos) {
165 list_del_init(&pos->core.node);
166 pos->evlist = NULL;
167 evsel__delete(pos);
168 }
169
170 evlist->core.nr_entries = 0;
171}
172
173void evlist__exit(struct evlist *evlist)
174{
175 event_enable_timer__exit(&evlist->eet);
176 zfree(&evlist->mmap);
177 zfree(&evlist->overwrite_mmap);
178 perf_evlist__exit(&evlist->core);
179}
180
181void evlist__delete(struct evlist *evlist)
182{
183 if (evlist == NULL)
184 return;
185
186 tpebs_delete();
187 evlist__free_stats(evlist);
188 evlist__munmap(evlist);
189 evlist__close(evlist);
190 evlist__purge(evlist);
191 evlist__exit(evlist);
192 free(evlist);
193}
194
195void evlist__add(struct evlist *evlist, struct evsel *entry)
196{
197 perf_evlist__add(&evlist->core, &entry->core);
198 entry->evlist = evlist;
199 entry->tracking = !entry->core.idx;
200
201 if (evlist->core.nr_entries == 1)
202 evlist__set_id_pos(evlist);
203}
204
205void evlist__remove(struct evlist *evlist, struct evsel *evsel)
206{
207 evsel->evlist = NULL;
208 perf_evlist__remove(&evlist->core, &evsel->core);
209}
210
211void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
212{
213 while (!list_empty(list)) {
214 struct evsel *evsel, *temp, *leader = NULL;
215
216 __evlist__for_each_entry_safe(list, temp, evsel) {
217 list_del_init(&evsel->core.node);
218 evlist__add(evlist, evsel);
219 leader = evsel;
220 break;
221 }
222
223 __evlist__for_each_entry_safe(list, temp, evsel) {
224 if (evsel__has_leader(evsel, leader)) {
225 list_del_init(&evsel->core.node);
226 evlist__add(evlist, evsel);
227 }
228 }
229 }
230}
231
232int __evlist__set_tracepoints_handlers(struct evlist *evlist,
233 const struct evsel_str_handler *assocs, size_t nr_assocs)
234{
235 size_t i;
236 int err;
237
238 for (i = 0; i < nr_assocs; i++) {
239 // Adding a handler for an event not in this evlist, just ignore it.
240 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
241 if (evsel == NULL)
242 continue;
243
244 err = -EEXIST;
245 if (evsel->handler != NULL)
246 goto out;
247 evsel->handler = assocs[i].handler;
248 }
249
250 err = 0;
251out:
252 return err;
253}
254
255static void evlist__set_leader(struct evlist *evlist)
256{
257 perf_evlist__set_leader(&evlist->core);
258}
259
260static struct evsel *evlist__dummy_event(struct evlist *evlist)
261{
262 struct perf_event_attr attr = {
263 .type = PERF_TYPE_SOFTWARE,
264 .config = PERF_COUNT_SW_DUMMY,
265 .size = sizeof(attr), /* to capture ABI version */
266 /* Avoid frequency mode for dummy events to avoid associated timers. */
267 .freq = 0,
268 .sample_period = 1,
269 };
270
271 return evsel__new_idx(&attr, evlist->core.nr_entries);
272}
273
274int evlist__add_dummy(struct evlist *evlist)
275{
276 struct evsel *evsel = evlist__dummy_event(evlist);
277
278 if (evsel == NULL)
279 return -ENOMEM;
280
281 evlist__add(evlist, evsel);
282 return 0;
283}
284
285struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
286{
287 struct evsel *evsel = evlist__dummy_event(evlist);
288
289 if (!evsel)
290 return NULL;
291
292 evsel->core.attr.exclude_kernel = 1;
293 evsel->core.attr.exclude_guest = 1;
294 evsel->core.attr.exclude_hv = 1;
295 evsel->core.system_wide = system_wide;
296 evsel->no_aux_samples = true;
297 evsel->name = strdup("dummy:u");
298
299 evlist__add(evlist, evsel);
300 return evsel;
301}
302
303#ifdef HAVE_LIBTRACEEVENT
304struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
305{
306 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0,
307 /*format=*/true);
308
309 if (IS_ERR(evsel))
310 return evsel;
311
312 evsel__set_sample_bit(evsel, CPU);
313 evsel__set_sample_bit(evsel, TIME);
314
315 evsel->core.system_wide = system_wide;
316 evsel->no_aux_samples = true;
317
318 evlist__add(evlist, evsel);
319 return evsel;
320}
321#endif
322
323struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
324{
325 struct evsel *evsel;
326
327 evlist__for_each_entry(evlist, evsel) {
328 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
329 (strcmp(evsel->name, name) == 0))
330 return evsel;
331 }
332
333 return NULL;
334}
335
336#ifdef HAVE_LIBTRACEEVENT
337int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
338{
339 struct evsel *evsel = evsel__newtp(sys, name);
340
341 if (IS_ERR(evsel))
342 return -1;
343
344 evsel->handler = handler;
345 evlist__add(evlist, evsel);
346 return 0;
347}
348#endif
349
350struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
351{
352 struct evlist_cpu_iterator itr = {
353 .container = evlist,
354 .evsel = NULL,
355 .cpu_map_idx = 0,
356 .evlist_cpu_map_idx = 0,
357 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
358 .cpu = (struct perf_cpu){ .cpu = -1},
359 .affinity = affinity,
360 };
361
362 if (evlist__empty(evlist)) {
363 /* Ensure the empty list doesn't iterate. */
364 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
365 } else {
366 itr.evsel = evlist__first(evlist);
367 if (itr.affinity) {
368 itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
369 affinity__set(itr.affinity, itr.cpu.cpu);
370 itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
371 /*
372 * If this CPU isn't in the evsel's cpu map then advance
373 * through the list.
374 */
375 if (itr.cpu_map_idx == -1)
376 evlist_cpu_iterator__next(&itr);
377 }
378 }
379 return itr;
380}
381
382void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
383{
384 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
385 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
386 evlist_cpu_itr->cpu_map_idx =
387 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
388 evlist_cpu_itr->cpu);
389 if (evlist_cpu_itr->cpu_map_idx != -1)
390 return;
391 }
392 evlist_cpu_itr->evlist_cpu_map_idx++;
393 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
394 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
395 evlist_cpu_itr->cpu =
396 perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
397 evlist_cpu_itr->evlist_cpu_map_idx);
398 if (evlist_cpu_itr->affinity)
399 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
400 evlist_cpu_itr->cpu_map_idx =
401 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
402 evlist_cpu_itr->cpu);
403 /*
404 * If this CPU isn't in the evsel's cpu map then advance through
405 * the list.
406 */
407 if (evlist_cpu_itr->cpu_map_idx == -1)
408 evlist_cpu_iterator__next(evlist_cpu_itr);
409 }
410}
411
412bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
413{
414 return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
415}
416
417static int evsel__strcmp(struct evsel *pos, char *evsel_name)
418{
419 if (!evsel_name)
420 return 0;
421 if (evsel__is_dummy_event(pos))
422 return 1;
423 return !evsel__name_is(pos, evsel_name);
424}
425
426static int evlist__is_enabled(struct evlist *evlist)
427{
428 struct evsel *pos;
429
430 evlist__for_each_entry(evlist, pos) {
431 if (!evsel__is_group_leader(pos) || !pos->core.fd)
432 continue;
433 /* If at least one event is enabled, evlist is enabled. */
434 if (!pos->disabled)
435 return true;
436 }
437 return false;
438}
439
440static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
441{
442 struct evsel *pos;
443 struct evlist_cpu_iterator evlist_cpu_itr;
444 struct affinity saved_affinity, *affinity = NULL;
445 bool has_imm = false;
446
447 // See explanation in evlist__close()
448 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
449 if (affinity__setup(&saved_affinity) < 0)
450 return;
451 affinity = &saved_affinity;
452 }
453
454 /* Disable 'immediate' events last */
455 for (int imm = 0; imm <= 1; imm++) {
456 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
457 pos = evlist_cpu_itr.evsel;
458 if (evsel__strcmp(pos, evsel_name))
459 continue;
460 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
461 continue;
462 if (excl_dummy && evsel__is_dummy_event(pos))
463 continue;
464 if (pos->immediate)
465 has_imm = true;
466 if (pos->immediate != imm)
467 continue;
468 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
469 }
470 if (!has_imm)
471 break;
472 }
473
474 affinity__cleanup(affinity);
475 evlist__for_each_entry(evlist, pos) {
476 if (evsel__strcmp(pos, evsel_name))
477 continue;
478 if (!evsel__is_group_leader(pos) || !pos->core.fd)
479 continue;
480 if (excl_dummy && evsel__is_dummy_event(pos))
481 continue;
482 pos->disabled = true;
483 }
484
485 /*
486 * If we disabled only single event, we need to check
487 * the enabled state of the evlist manually.
488 */
489 if (evsel_name)
490 evlist->enabled = evlist__is_enabled(evlist);
491 else
492 evlist->enabled = false;
493}
494
495void evlist__disable(struct evlist *evlist)
496{
497 __evlist__disable(evlist, NULL, false);
498}
499
500void evlist__disable_non_dummy(struct evlist *evlist)
501{
502 __evlist__disable(evlist, NULL, true);
503}
504
505void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
506{
507 __evlist__disable(evlist, evsel_name, false);
508}
509
510static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
511{
512 struct evsel *pos;
513 struct evlist_cpu_iterator evlist_cpu_itr;
514 struct affinity saved_affinity, *affinity = NULL;
515
516 // See explanation in evlist__close()
517 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
518 if (affinity__setup(&saved_affinity) < 0)
519 return;
520 affinity = &saved_affinity;
521 }
522
523 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
524 pos = evlist_cpu_itr.evsel;
525 if (evsel__strcmp(pos, evsel_name))
526 continue;
527 if (!evsel__is_group_leader(pos) || !pos->core.fd)
528 continue;
529 if (excl_dummy && evsel__is_dummy_event(pos))
530 continue;
531 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
532 }
533 affinity__cleanup(affinity);
534 evlist__for_each_entry(evlist, pos) {
535 if (evsel__strcmp(pos, evsel_name))
536 continue;
537 if (!evsel__is_group_leader(pos) || !pos->core.fd)
538 continue;
539 if (excl_dummy && evsel__is_dummy_event(pos))
540 continue;
541 pos->disabled = false;
542 }
543
544 /*
545 * Even single event sets the 'enabled' for evlist,
546 * so the toggle can work properly and toggle to
547 * 'disabled' state.
548 */
549 evlist->enabled = true;
550}
551
552void evlist__enable(struct evlist *evlist)
553{
554 __evlist__enable(evlist, NULL, false);
555}
556
557void evlist__enable_non_dummy(struct evlist *evlist)
558{
559 __evlist__enable(evlist, NULL, true);
560}
561
562void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
563{
564 __evlist__enable(evlist, evsel_name, false);
565}
566
567void evlist__toggle_enable(struct evlist *evlist)
568{
569 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
570}
571
572int evlist__add_pollfd(struct evlist *evlist, int fd)
573{
574 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
575}
576
577int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
578{
579 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
580}
581
582#ifdef HAVE_EVENTFD_SUPPORT
583int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
584{
585 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
586 fdarray_flag__nonfilterable |
587 fdarray_flag__non_perf_event);
588}
589#endif
590
591int evlist__poll(struct evlist *evlist, int timeout)
592{
593 return perf_evlist__poll(&evlist->core, timeout);
594}
595
596struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
597{
598 struct hlist_head *head;
599 struct perf_sample_id *sid;
600 int hash;
601
602 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
603 head = &evlist->core.heads[hash];
604
605 hlist_for_each_entry(sid, head, node)
606 if (sid->id == id)
607 return sid;
608
609 return NULL;
610}
611
612struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
613{
614 struct perf_sample_id *sid;
615
616 if (evlist->core.nr_entries == 1 || !id)
617 return evlist__first(evlist);
618
619 sid = evlist__id2sid(evlist, id);
620 if (sid)
621 return container_of(sid->evsel, struct evsel, core);
622
623 if (!evlist__sample_id_all(evlist))
624 return evlist__first(evlist);
625
626 return NULL;
627}
628
629struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
630{
631 struct perf_sample_id *sid;
632
633 if (!id)
634 return NULL;
635
636 sid = evlist__id2sid(evlist, id);
637 if (sid)
638 return container_of(sid->evsel, struct evsel, core);
639
640 return NULL;
641}
642
643static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
644{
645 const __u64 *array = event->sample.array;
646 ssize_t n;
647
648 n = (event->header.size - sizeof(event->header)) >> 3;
649
650 if (event->header.type == PERF_RECORD_SAMPLE) {
651 if (evlist->id_pos >= n)
652 return -1;
653 *id = array[evlist->id_pos];
654 } else {
655 if (evlist->is_pos > n)
656 return -1;
657 n -= evlist->is_pos;
658 *id = array[n];
659 }
660 return 0;
661}
662
663struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
664{
665 struct evsel *first = evlist__first(evlist);
666 struct hlist_head *head;
667 struct perf_sample_id *sid;
668 int hash;
669 u64 id;
670
671 if (evlist->core.nr_entries == 1)
672 return first;
673
674 if (!first->core.attr.sample_id_all &&
675 event->header.type != PERF_RECORD_SAMPLE)
676 return first;
677
678 if (evlist__event2id(evlist, event, &id))
679 return NULL;
680
681 /* Synthesized events have an id of zero */
682 if (!id)
683 return first;
684
685 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
686 head = &evlist->core.heads[hash];
687
688 hlist_for_each_entry(sid, head, node) {
689 if (sid->id == id)
690 return container_of(sid->evsel, struct evsel, core);
691 }
692 return NULL;
693}
694
695static int evlist__set_paused(struct evlist *evlist, bool value)
696{
697 int i;
698
699 if (!evlist->overwrite_mmap)
700 return 0;
701
702 for (i = 0; i < evlist->core.nr_mmaps; i++) {
703 int fd = evlist->overwrite_mmap[i].core.fd;
704 int err;
705
706 if (fd < 0)
707 continue;
708 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
709 if (err)
710 return err;
711 }
712 return 0;
713}
714
715static int evlist__pause(struct evlist *evlist)
716{
717 return evlist__set_paused(evlist, true);
718}
719
720static int evlist__resume(struct evlist *evlist)
721{
722 return evlist__set_paused(evlist, false);
723}
724
725static void evlist__munmap_nofree(struct evlist *evlist)
726{
727 int i;
728
729 if (evlist->mmap)
730 for (i = 0; i < evlist->core.nr_mmaps; i++)
731 perf_mmap__munmap(&evlist->mmap[i].core);
732
733 if (evlist->overwrite_mmap)
734 for (i = 0; i < evlist->core.nr_mmaps; i++)
735 perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
736}
737
738void evlist__munmap(struct evlist *evlist)
739{
740 evlist__munmap_nofree(evlist);
741 zfree(&evlist->mmap);
742 zfree(&evlist->overwrite_mmap);
743}
744
745static void perf_mmap__unmap_cb(struct perf_mmap *map)
746{
747 struct mmap *m = container_of(map, struct mmap, core);
748
749 mmap__munmap(m);
750}
751
752static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
753 bool overwrite)
754{
755 int i;
756 struct mmap *map;
757
758 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
759 if (!map)
760 return NULL;
761
762 for (i = 0; i < evlist->core.nr_mmaps; i++) {
763 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
764
765 /*
766 * When the perf_mmap() call is made we grab one refcount, plus
767 * one extra to let perf_mmap__consume() get the last
768 * events after all real references (perf_mmap__get()) are
769 * dropped.
770 *
771 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
772 * thus does perf_mmap__get() on it.
773 */
774 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
775 }
776
777 return map;
778}
779
780static void
781perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
782 struct perf_evsel *_evsel,
783 struct perf_mmap_param *_mp,
784 int idx)
785{
786 struct evlist *evlist = container_of(_evlist, struct evlist, core);
787 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
788 struct evsel *evsel = container_of(_evsel, struct evsel, core);
789
790 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
791}
792
793static struct perf_mmap*
794perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
795{
796 struct evlist *evlist = container_of(_evlist, struct evlist, core);
797 struct mmap *maps;
798
799 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
800
801 if (!maps) {
802 maps = evlist__alloc_mmap(evlist, overwrite);
803 if (!maps)
804 return NULL;
805
806 if (overwrite) {
807 evlist->overwrite_mmap = maps;
808 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
809 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
810 } else {
811 evlist->mmap = maps;
812 }
813 }
814
815 return &maps[idx].core;
816}
817
818static int
819perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
820 int output, struct perf_cpu cpu)
821{
822 struct mmap *map = container_of(_map, struct mmap, core);
823 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
824
825 return mmap__mmap(map, mp, output, cpu);
826}
827
828unsigned long perf_event_mlock_kb_in_pages(void)
829{
830 unsigned long pages;
831 int max;
832
833 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
834 /*
835 * Pick a once upon a time good value, i.e. things look
836 * strange since we can't read a sysctl value, but lets not
837 * die yet...
838 */
839 max = 512;
840 } else {
841 max -= (page_size / 1024);
842 }
843
844 pages = (max * 1024) / page_size;
845 if (!is_power_of_2(pages))
846 pages = rounddown_pow_of_two(pages);
847
848 return pages;
849}
850
851size_t evlist__mmap_size(unsigned long pages)
852{
853 if (pages == UINT_MAX)
854 pages = perf_event_mlock_kb_in_pages();
855 else if (!is_power_of_2(pages))
856 return 0;
857
858 return (pages + 1) * page_size;
859}
860
861static long parse_pages_arg(const char *str, unsigned long min,
862 unsigned long max)
863{
864 unsigned long pages, val;
865 static struct parse_tag tags[] = {
866 { .tag = 'B', .mult = 1 },
867 { .tag = 'K', .mult = 1 << 10 },
868 { .tag = 'M', .mult = 1 << 20 },
869 { .tag = 'G', .mult = 1 << 30 },
870 { .tag = 0 },
871 };
872
873 if (str == NULL)
874 return -EINVAL;
875
876 val = parse_tag_value(str, tags);
877 if (val != (unsigned long) -1) {
878 /* we got file size value */
879 pages = PERF_ALIGN(val, page_size) / page_size;
880 } else {
881 /* we got pages count value */
882 char *eptr;
883 pages = strtoul(str, &eptr, 10);
884 if (*eptr != '\0')
885 return -EINVAL;
886 }
887
888 if (pages == 0 && min == 0) {
889 /* leave number of pages at 0 */
890 } else if (!is_power_of_2(pages)) {
891 char buf[100];
892
893 /* round pages up to next power of 2 */
894 pages = roundup_pow_of_two(pages);
895 if (!pages)
896 return -EINVAL;
897
898 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
899 pr_info("rounding mmap pages size to %s (%lu pages)\n",
900 buf, pages);
901 }
902
903 if (pages > max)
904 return -EINVAL;
905
906 return pages;
907}
908
909int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
910{
911 unsigned long max = UINT_MAX;
912 long pages;
913
914 if (max > SIZE_MAX / page_size)
915 max = SIZE_MAX / page_size;
916
917 pages = parse_pages_arg(str, 1, max);
918 if (pages < 0) {
919 pr_err("Invalid argument for --mmap_pages/-m\n");
920 return -1;
921 }
922
923 *mmap_pages = pages;
924 return 0;
925}
926
927int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
928{
929 return __evlist__parse_mmap_pages(opt->value, str);
930}
931
932/**
933 * evlist__mmap_ex - Create mmaps to receive events.
934 * @evlist: list of events
935 * @pages: map length in pages
936 * @overwrite: overwrite older events?
937 * @auxtrace_pages - auxtrace map length in pages
938 * @auxtrace_overwrite - overwrite older auxtrace data?
939 *
940 * If @overwrite is %false the user needs to signal event consumption using
941 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
942 * automatically.
943 *
944 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
945 * consumption using auxtrace_mmap__write_tail().
946 *
947 * Return: %0 on success, negative error code otherwise.
948 */
949int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
950 unsigned int auxtrace_pages,
951 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
952 int comp_level)
953{
954 /*
955 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
956 * Its value is decided by evsel's write_backward.
957 * So &mp should not be passed through const pointer.
958 */
959 struct mmap_params mp = {
960 .nr_cblocks = nr_cblocks,
961 .affinity = affinity,
962 .flush = flush,
963 .comp_level = comp_level
964 };
965 struct perf_evlist_mmap_ops ops = {
966 .idx = perf_evlist__mmap_cb_idx,
967 .get = perf_evlist__mmap_cb_get,
968 .mmap = perf_evlist__mmap_cb_mmap,
969 };
970
971 evlist->core.mmap_len = evlist__mmap_size(pages);
972 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
973
974 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
975 auxtrace_pages, auxtrace_overwrite);
976
977 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
978}
979
980int evlist__mmap(struct evlist *evlist, unsigned int pages)
981{
982 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
983}
984
985int evlist__create_maps(struct evlist *evlist, struct target *target)
986{
987 bool all_threads = (target->per_thread && target->system_wide);
988 struct perf_cpu_map *cpus;
989 struct perf_thread_map *threads;
990
991 /*
992 * If specify '-a' and '--per-thread' to perf record, perf record
993 * will override '--per-thread'. target->per_thread = false and
994 * target->system_wide = true.
995 *
996 * If specify '--per-thread' only to perf record,
997 * target->per_thread = true and target->system_wide = false.
998 *
999 * So target->per_thread && target->system_wide is false.
1000 * For perf record, thread_map__new_str doesn't call
1001 * thread_map__new_all_cpus. That will keep perf record's
1002 * current behavior.
1003 *
1004 * For perf stat, it allows the case that target->per_thread and
1005 * target->system_wide are all true. It means to collect system-wide
1006 * per-thread data. thread_map__new_str will call
1007 * thread_map__new_all_cpus to enumerate all threads.
1008 */
1009 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1010 all_threads);
1011
1012 if (!threads)
1013 return -1;
1014
1015 if (target__uses_dummy_map(target) && !evlist__has_bpf_output(evlist))
1016 cpus = perf_cpu_map__new_any_cpu();
1017 else
1018 cpus = perf_cpu_map__new(target->cpu_list);
1019
1020 if (!cpus)
1021 goto out_delete_threads;
1022
1023 evlist->core.has_user_cpus = !!target->cpu_list;
1024
1025 perf_evlist__set_maps(&evlist->core, cpus, threads);
1026
1027 /* as evlist now has references, put count here */
1028 perf_cpu_map__put(cpus);
1029 perf_thread_map__put(threads);
1030
1031 return 0;
1032
1033out_delete_threads:
1034 perf_thread_map__put(threads);
1035 return -1;
1036}
1037
1038int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel,
1039 struct target *target)
1040{
1041 struct evsel *evsel;
1042 int err = 0;
1043
1044 evlist__for_each_entry(evlist, evsel) {
1045 /*
1046 * filters only work for tracepoint event, which doesn't have cpu limit.
1047 * So evlist and evsel should always be same.
1048 */
1049 if (evsel->filter) {
1050 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1051 if (err) {
1052 *err_evsel = evsel;
1053 break;
1054 }
1055 }
1056
1057 /*
1058 * non-tracepoint events can have BPF filters.
1059 */
1060 if (!list_empty(&evsel->bpf_filters)) {
1061 err = perf_bpf_filter__prepare(evsel, target);
1062 if (err) {
1063 *err_evsel = evsel;
1064 break;
1065 }
1066 }
1067 }
1068
1069 return err;
1070}
1071
1072int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1073{
1074 struct evsel *evsel;
1075 int err = 0;
1076
1077 if (filter == NULL)
1078 return -1;
1079
1080 evlist__for_each_entry(evlist, evsel) {
1081 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1082 continue;
1083
1084 err = evsel__set_filter(evsel, filter);
1085 if (err)
1086 break;
1087 }
1088
1089 return err;
1090}
1091
1092int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1093{
1094 struct evsel *evsel;
1095 int err = 0;
1096
1097 if (filter == NULL)
1098 return -1;
1099
1100 evlist__for_each_entry(evlist, evsel) {
1101 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1102 continue;
1103
1104 err = evsel__append_tp_filter(evsel, filter);
1105 if (err)
1106 break;
1107 }
1108
1109 return err;
1110}
1111
1112char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1113{
1114 char *filter;
1115 size_t i;
1116
1117 for (i = 0; i < npids; ++i) {
1118 if (i == 0) {
1119 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1120 return NULL;
1121 } else {
1122 char *tmp;
1123
1124 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1125 goto out_free;
1126
1127 free(filter);
1128 filter = tmp;
1129 }
1130 }
1131
1132 return filter;
1133out_free:
1134 free(filter);
1135 return NULL;
1136}
1137
1138int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1139{
1140 char *filter = asprintf__tp_filter_pids(npids, pids);
1141 int ret = evlist__set_tp_filter(evlist, filter);
1142
1143 free(filter);
1144 return ret;
1145}
1146
1147int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1148{
1149 char *filter = asprintf__tp_filter_pids(npids, pids);
1150 int ret = evlist__append_tp_filter(evlist, filter);
1151
1152 free(filter);
1153 return ret;
1154}
1155
1156int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1157{
1158 return evlist__append_tp_filter_pids(evlist, 1, &pid);
1159}
1160
1161bool evlist__valid_sample_type(struct evlist *evlist)
1162{
1163 struct evsel *pos;
1164
1165 if (evlist->core.nr_entries == 1)
1166 return true;
1167
1168 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1169 return false;
1170
1171 evlist__for_each_entry(evlist, pos) {
1172 if (pos->id_pos != evlist->id_pos ||
1173 pos->is_pos != evlist->is_pos)
1174 return false;
1175 }
1176
1177 return true;
1178}
1179
1180u64 __evlist__combined_sample_type(struct evlist *evlist)
1181{
1182 struct evsel *evsel;
1183
1184 if (evlist->combined_sample_type)
1185 return evlist->combined_sample_type;
1186
1187 evlist__for_each_entry(evlist, evsel)
1188 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1189
1190 return evlist->combined_sample_type;
1191}
1192
1193u64 evlist__combined_sample_type(struct evlist *evlist)
1194{
1195 evlist->combined_sample_type = 0;
1196 return __evlist__combined_sample_type(evlist);
1197}
1198
1199u64 evlist__combined_branch_type(struct evlist *evlist)
1200{
1201 struct evsel *evsel;
1202 u64 branch_type = 0;
1203
1204 evlist__for_each_entry(evlist, evsel)
1205 branch_type |= evsel->core.attr.branch_sample_type;
1206 return branch_type;
1207}
1208
1209static struct evsel *
1210evlist__find_dup_event_from_prev(struct evlist *evlist, struct evsel *event)
1211{
1212 struct evsel *pos;
1213
1214 evlist__for_each_entry(evlist, pos) {
1215 if (event == pos)
1216 break;
1217 if ((pos->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) &&
1218 !strcmp(pos->name, event->name))
1219 return pos;
1220 }
1221 return NULL;
1222}
1223
1224#define MAX_NR_ABBR_NAME (26 * 11)
1225
1226/*
1227 * The abbr name is from A to Z9. If the number of event
1228 * which requires the branch counter > MAX_NR_ABBR_NAME,
1229 * return NA.
1230 */
1231static void evlist__new_abbr_name(char *name)
1232{
1233 static int idx;
1234 int i = idx / 26;
1235
1236 if (idx >= MAX_NR_ABBR_NAME) {
1237 name[0] = 'N';
1238 name[1] = 'A';
1239 name[2] = '\0';
1240 return;
1241 }
1242
1243 name[0] = 'A' + (idx % 26);
1244
1245 if (!i)
1246 name[1] = '\0';
1247 else {
1248 name[1] = '0' + i - 1;
1249 name[2] = '\0';
1250 }
1251
1252 idx++;
1253}
1254
1255void evlist__update_br_cntr(struct evlist *evlist)
1256{
1257 struct evsel *evsel, *dup;
1258 int i = 0;
1259
1260 evlist__for_each_entry(evlist, evsel) {
1261 if (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) {
1262 evsel->br_cntr_idx = i++;
1263 evsel__leader(evsel)->br_cntr_nr++;
1264
1265 dup = evlist__find_dup_event_from_prev(evlist, evsel);
1266 if (dup)
1267 memcpy(evsel->abbr_name, dup->abbr_name, 3 * sizeof(char));
1268 else
1269 evlist__new_abbr_name(evsel->abbr_name);
1270 }
1271 }
1272 evlist->nr_br_cntr = i;
1273}
1274
1275bool evlist__valid_read_format(struct evlist *evlist)
1276{
1277 struct evsel *first = evlist__first(evlist), *pos = first;
1278 u64 read_format = first->core.attr.read_format;
1279 u64 sample_type = first->core.attr.sample_type;
1280
1281 evlist__for_each_entry(evlist, pos) {
1282 if (read_format != pos->core.attr.read_format) {
1283 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1284 read_format, (u64)pos->core.attr.read_format);
1285 }
1286 }
1287
1288 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1289 if ((sample_type & PERF_SAMPLE_READ) &&
1290 !(read_format & PERF_FORMAT_ID)) {
1291 return false;
1292 }
1293
1294 return true;
1295}
1296
1297u16 evlist__id_hdr_size(struct evlist *evlist)
1298{
1299 struct evsel *first = evlist__first(evlist);
1300
1301 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1302}
1303
1304bool evlist__valid_sample_id_all(struct evlist *evlist)
1305{
1306 struct evsel *first = evlist__first(evlist), *pos = first;
1307
1308 evlist__for_each_entry_continue(evlist, pos) {
1309 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1310 return false;
1311 }
1312
1313 return true;
1314}
1315
1316bool evlist__sample_id_all(struct evlist *evlist)
1317{
1318 struct evsel *first = evlist__first(evlist);
1319 return first->core.attr.sample_id_all;
1320}
1321
1322void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1323{
1324 evlist->selected = evsel;
1325}
1326
1327void evlist__close(struct evlist *evlist)
1328{
1329 struct evsel *evsel;
1330 struct evlist_cpu_iterator evlist_cpu_itr;
1331 struct affinity affinity;
1332
1333 /*
1334 * With perf record core.user_requested_cpus is usually NULL.
1335 * Use the old method to handle this for now.
1336 */
1337 if (!evlist->core.user_requested_cpus ||
1338 cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1339 evlist__for_each_entry_reverse(evlist, evsel)
1340 evsel__close(evsel);
1341 return;
1342 }
1343
1344 if (affinity__setup(&affinity) < 0)
1345 return;
1346
1347 evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1348 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1349 evlist_cpu_itr.cpu_map_idx);
1350 }
1351
1352 affinity__cleanup(&affinity);
1353 evlist__for_each_entry_reverse(evlist, evsel) {
1354 perf_evsel__free_fd(&evsel->core);
1355 perf_evsel__free_id(&evsel->core);
1356 }
1357 perf_evlist__reset_id_hash(&evlist->core);
1358}
1359
1360static int evlist__create_syswide_maps(struct evlist *evlist)
1361{
1362 struct perf_cpu_map *cpus;
1363 struct perf_thread_map *threads;
1364
1365 /*
1366 * Try reading /sys/devices/system/cpu/online to get
1367 * an all cpus map.
1368 *
1369 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1370 * code needs an overhaul to properly forward the
1371 * error, and we may not want to do that fallback to a
1372 * default cpu identity map :-\
1373 */
1374 cpus = perf_cpu_map__new_online_cpus();
1375 if (!cpus)
1376 goto out;
1377
1378 threads = perf_thread_map__new_dummy();
1379 if (!threads)
1380 goto out_put;
1381
1382 perf_evlist__set_maps(&evlist->core, cpus, threads);
1383
1384 perf_thread_map__put(threads);
1385out_put:
1386 perf_cpu_map__put(cpus);
1387out:
1388 return -ENOMEM;
1389}
1390
1391int evlist__open(struct evlist *evlist)
1392{
1393 struct evsel *evsel;
1394 int err;
1395
1396 /*
1397 * Default: one fd per CPU, all threads, aka systemwide
1398 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1399 */
1400 if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1401 err = evlist__create_syswide_maps(evlist);
1402 if (err < 0)
1403 goto out_err;
1404 }
1405
1406 evlist__update_id_pos(evlist);
1407
1408 evlist__for_each_entry(evlist, evsel) {
1409 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1410 if (err < 0)
1411 goto out_err;
1412 }
1413
1414 return 0;
1415out_err:
1416 evlist__close(evlist);
1417 errno = -err;
1418 return err;
1419}
1420
1421int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1422 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1423{
1424 int child_ready_pipe[2], go_pipe[2];
1425 char bf;
1426
1427 evlist->workload.cork_fd = -1;
1428
1429 if (pipe(child_ready_pipe) < 0) {
1430 perror("failed to create 'ready' pipe");
1431 return -1;
1432 }
1433
1434 if (pipe(go_pipe) < 0) {
1435 perror("failed to create 'go' pipe");
1436 goto out_close_ready_pipe;
1437 }
1438
1439 evlist->workload.pid = fork();
1440 if (evlist->workload.pid < 0) {
1441 perror("failed to fork");
1442 goto out_close_pipes;
1443 }
1444
1445 if (!evlist->workload.pid) {
1446 int ret;
1447
1448 if (pipe_output)
1449 dup2(2, 1);
1450
1451 signal(SIGTERM, SIG_DFL);
1452
1453 close(child_ready_pipe[0]);
1454 close(go_pipe[1]);
1455 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1456
1457 /*
1458 * Change the name of this process not to confuse --exclude-perf users
1459 * that sees 'perf' in the window up to the execvp() and thinks that
1460 * perf samples are not being excluded.
1461 */
1462 prctl(PR_SET_NAME, "perf-exec");
1463
1464 /*
1465 * Tell the parent we're ready to go
1466 */
1467 close(child_ready_pipe[1]);
1468
1469 /*
1470 * Wait until the parent tells us to go.
1471 */
1472 ret = read(go_pipe[0], &bf, 1);
1473 /*
1474 * The parent will ask for the execvp() to be performed by
1475 * writing exactly one byte, in workload.cork_fd, usually via
1476 * evlist__start_workload().
1477 *
1478 * For cancelling the workload without actually running it,
1479 * the parent will just close workload.cork_fd, without writing
1480 * anything, i.e. read will return zero and we just exit()
1481 * here (See evlist__cancel_workload()).
1482 */
1483 if (ret != 1) {
1484 if (ret == -1)
1485 perror("unable to read pipe");
1486 exit(ret);
1487 }
1488
1489 execvp(argv[0], (char **)argv);
1490
1491 if (exec_error) {
1492 union sigval val;
1493
1494 val.sival_int = errno;
1495 if (sigqueue(getppid(), SIGUSR1, val))
1496 perror(argv[0]);
1497 } else
1498 perror(argv[0]);
1499 exit(-1);
1500 }
1501
1502 if (exec_error) {
1503 struct sigaction act = {
1504 .sa_flags = SA_SIGINFO,
1505 .sa_sigaction = exec_error,
1506 };
1507 sigaction(SIGUSR1, &act, NULL);
1508 }
1509
1510 if (target__none(target)) {
1511 if (evlist->core.threads == NULL) {
1512 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1513 __func__, __LINE__);
1514 goto out_close_pipes;
1515 }
1516 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1517 }
1518
1519 close(child_ready_pipe[1]);
1520 close(go_pipe[0]);
1521 /*
1522 * wait for child to settle
1523 */
1524 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1525 perror("unable to read pipe");
1526 goto out_close_pipes;
1527 }
1528
1529 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1530 evlist->workload.cork_fd = go_pipe[1];
1531 close(child_ready_pipe[0]);
1532 return 0;
1533
1534out_close_pipes:
1535 close(go_pipe[0]);
1536 close(go_pipe[1]);
1537out_close_ready_pipe:
1538 close(child_ready_pipe[0]);
1539 close(child_ready_pipe[1]);
1540 return -1;
1541}
1542
1543int evlist__start_workload(struct evlist *evlist)
1544{
1545 if (evlist->workload.cork_fd >= 0) {
1546 char bf = 0;
1547 int ret;
1548 /*
1549 * Remove the cork, let it rip!
1550 */
1551 ret = write(evlist->workload.cork_fd, &bf, 1);
1552 if (ret < 0)
1553 perror("unable to write to pipe");
1554
1555 close(evlist->workload.cork_fd);
1556 evlist->workload.cork_fd = -1;
1557 return ret;
1558 }
1559
1560 return 0;
1561}
1562
1563void evlist__cancel_workload(struct evlist *evlist)
1564{
1565 int status;
1566
1567 if (evlist->workload.cork_fd >= 0) {
1568 close(evlist->workload.cork_fd);
1569 evlist->workload.cork_fd = -1;
1570 waitpid(evlist->workload.pid, &status, WNOHANG);
1571 }
1572}
1573
1574int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1575{
1576 struct evsel *evsel = evlist__event2evsel(evlist, event);
1577 int ret;
1578
1579 if (!evsel)
1580 return -EFAULT;
1581 ret = evsel__parse_sample(evsel, event, sample);
1582 if (ret)
1583 return ret;
1584 if (perf_guest && sample->id) {
1585 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1586
1587 if (sid) {
1588 sample->machine_pid = sid->machine_pid;
1589 sample->vcpu = sid->vcpu.cpu;
1590 }
1591 }
1592 return 0;
1593}
1594
1595int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1596{
1597 struct evsel *evsel = evlist__event2evsel(evlist, event);
1598
1599 if (!evsel)
1600 return -EFAULT;
1601 return evsel__parse_sample_timestamp(evsel, event, timestamp);
1602}
1603
1604int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1605{
1606 int printed, value;
1607 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1608
1609 switch (err) {
1610 case EACCES:
1611 case EPERM:
1612 printed = scnprintf(buf, size,
1613 "Error:\t%s.\n"
1614 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1615
1616 value = perf_event_paranoid();
1617
1618 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1619
1620 if (value >= 2) {
1621 printed += scnprintf(buf + printed, size - printed,
1622 "For your workloads it needs to be <= 1\nHint:\t");
1623 }
1624 printed += scnprintf(buf + printed, size - printed,
1625 "For system wide tracing it needs to be set to -1.\n");
1626
1627 printed += scnprintf(buf + printed, size - printed,
1628 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1629 "Hint:\tThe current value is %d.", value);
1630 break;
1631 case EINVAL: {
1632 struct evsel *first = evlist__first(evlist);
1633 int max_freq;
1634
1635 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1636 goto out_default;
1637
1638 if (first->core.attr.sample_freq < (u64)max_freq)
1639 goto out_default;
1640
1641 printed = scnprintf(buf, size,
1642 "Error:\t%s.\n"
1643 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1644 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1645 emsg, max_freq, first->core.attr.sample_freq);
1646 break;
1647 }
1648 default:
1649out_default:
1650 scnprintf(buf, size, "%s", emsg);
1651 break;
1652 }
1653
1654 return 0;
1655}
1656
1657int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1658{
1659 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1660 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1661
1662 switch (err) {
1663 case EPERM:
1664 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1665 printed += scnprintf(buf + printed, size - printed,
1666 "Error:\t%s.\n"
1667 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1668 "Hint:\tTried using %zd kB.\n",
1669 emsg, pages_max_per_user, pages_attempted);
1670
1671 if (pages_attempted >= pages_max_per_user) {
1672 printed += scnprintf(buf + printed, size - printed,
1673 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1674 pages_max_per_user + pages_attempted);
1675 }
1676
1677 printed += scnprintf(buf + printed, size - printed,
1678 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1679 break;
1680 default:
1681 scnprintf(buf, size, "%s", emsg);
1682 break;
1683 }
1684
1685 return 0;
1686}
1687
1688void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1689{
1690 struct evsel *evsel, *n;
1691 LIST_HEAD(move);
1692
1693 if (move_evsel == evlist__first(evlist))
1694 return;
1695
1696 evlist__for_each_entry_safe(evlist, n, evsel) {
1697 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1698 list_move_tail(&evsel->core.node, &move);
1699 }
1700
1701 list_splice(&move, &evlist->core.entries);
1702}
1703
1704struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1705{
1706 struct evsel *evsel;
1707
1708 evlist__for_each_entry(evlist, evsel) {
1709 if (evsel->tracking)
1710 return evsel;
1711 }
1712
1713 return evlist__first(evlist);
1714}
1715
1716void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1717{
1718 struct evsel *evsel;
1719
1720 if (tracking_evsel->tracking)
1721 return;
1722
1723 evlist__for_each_entry(evlist, evsel) {
1724 if (evsel != tracking_evsel)
1725 evsel->tracking = false;
1726 }
1727
1728 tracking_evsel->tracking = true;
1729}
1730
1731struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide)
1732{
1733 struct evsel *evsel;
1734
1735 evsel = evlist__get_tracking_event(evlist);
1736 if (!evsel__is_dummy_event(evsel)) {
1737 evsel = evlist__add_aux_dummy(evlist, system_wide);
1738 if (!evsel)
1739 return NULL;
1740
1741 evlist__set_tracking_event(evlist, evsel);
1742 } else if (system_wide) {
1743 perf_evlist__go_system_wide(&evlist->core, &evsel->core);
1744 }
1745
1746 return evsel;
1747}
1748
1749struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1750{
1751 struct evsel *evsel;
1752
1753 evlist__for_each_entry(evlist, evsel) {
1754 if (!evsel->name)
1755 continue;
1756 if (evsel__name_is(evsel, str))
1757 return evsel;
1758 }
1759
1760 return NULL;
1761}
1762
1763void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1764{
1765 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1766 enum action {
1767 NONE,
1768 PAUSE,
1769 RESUME,
1770 } action = NONE;
1771
1772 if (!evlist->overwrite_mmap)
1773 return;
1774
1775 switch (old_state) {
1776 case BKW_MMAP_NOTREADY: {
1777 if (state != BKW_MMAP_RUNNING)
1778 goto state_err;
1779 break;
1780 }
1781 case BKW_MMAP_RUNNING: {
1782 if (state != BKW_MMAP_DATA_PENDING)
1783 goto state_err;
1784 action = PAUSE;
1785 break;
1786 }
1787 case BKW_MMAP_DATA_PENDING: {
1788 if (state != BKW_MMAP_EMPTY)
1789 goto state_err;
1790 break;
1791 }
1792 case BKW_MMAP_EMPTY: {
1793 if (state != BKW_MMAP_RUNNING)
1794 goto state_err;
1795 action = RESUME;
1796 break;
1797 }
1798 default:
1799 WARN_ONCE(1, "Shouldn't get there\n");
1800 }
1801
1802 evlist->bkw_mmap_state = state;
1803
1804 switch (action) {
1805 case PAUSE:
1806 evlist__pause(evlist);
1807 break;
1808 case RESUME:
1809 evlist__resume(evlist);
1810 break;
1811 case NONE:
1812 default:
1813 break;
1814 }
1815
1816state_err:
1817 return;
1818}
1819
1820bool evlist__exclude_kernel(struct evlist *evlist)
1821{
1822 struct evsel *evsel;
1823
1824 evlist__for_each_entry(evlist, evsel) {
1825 if (!evsel->core.attr.exclude_kernel)
1826 return false;
1827 }
1828
1829 return true;
1830}
1831
1832/*
1833 * Events in data file are not collect in groups, but we still want
1834 * the group display. Set the artificial group and set the leader's
1835 * forced_leader flag to notify the display code.
1836 */
1837void evlist__force_leader(struct evlist *evlist)
1838{
1839 if (evlist__nr_groups(evlist) == 0) {
1840 struct evsel *leader = evlist__first(evlist);
1841
1842 evlist__set_leader(evlist);
1843 leader->forced_leader = true;
1844 }
1845}
1846
1847struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1848{
1849 struct evsel *c2, *leader;
1850 bool is_open = true;
1851
1852 leader = evsel__leader(evsel);
1853
1854 pr_debug("Weak group for %s/%d failed\n",
1855 leader->name, leader->core.nr_members);
1856
1857 /*
1858 * for_each_group_member doesn't work here because it doesn't
1859 * include the first entry.
1860 */
1861 evlist__for_each_entry(evsel_list, c2) {
1862 if (c2 == evsel)
1863 is_open = false;
1864 if (evsel__has_leader(c2, leader)) {
1865 if (is_open && close)
1866 perf_evsel__close(&c2->core);
1867 /*
1868 * We want to close all members of the group and reopen
1869 * them. Some events, like Intel topdown, require being
1870 * in a group and so keep these in the group.
1871 */
1872 evsel__remove_from_group(c2, leader);
1873
1874 /*
1875 * Set this for all former members of the group
1876 * to indicate they get reopened.
1877 */
1878 c2->reset_group = true;
1879 }
1880 }
1881 /* Reset the leader count if all entries were removed. */
1882 if (leader->core.nr_members == 1)
1883 leader->core.nr_members = 0;
1884 return leader;
1885}
1886
1887static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1888{
1889 char *s, *p;
1890 int ret = 0, fd;
1891
1892 if (strncmp(str, "fifo:", 5))
1893 return -EINVAL;
1894
1895 str += 5;
1896 if (!*str || *str == ',')
1897 return -EINVAL;
1898
1899 s = strdup(str);
1900 if (!s)
1901 return -ENOMEM;
1902
1903 p = strchr(s, ',');
1904 if (p)
1905 *p = '\0';
1906
1907 /*
1908 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1909 * end of a FIFO to be repeatedly opened and closed.
1910 */
1911 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1912 if (fd < 0) {
1913 pr_err("Failed to open '%s'\n", s);
1914 ret = -errno;
1915 goto out_free;
1916 }
1917 *ctl_fd = fd;
1918 *ctl_fd_close = true;
1919
1920 if (p && *++p) {
1921 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1922 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1923 if (fd < 0) {
1924 pr_err("Failed to open '%s'\n", p);
1925 ret = -errno;
1926 goto out_free;
1927 }
1928 *ctl_fd_ack = fd;
1929 }
1930
1931out_free:
1932 free(s);
1933 return ret;
1934}
1935
1936int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1937{
1938 char *comma = NULL, *endptr = NULL;
1939
1940 *ctl_fd_close = false;
1941
1942 if (strncmp(str, "fd:", 3))
1943 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1944
1945 *ctl_fd = strtoul(&str[3], &endptr, 0);
1946 if (endptr == &str[3])
1947 return -EINVAL;
1948
1949 comma = strchr(str, ',');
1950 if (comma) {
1951 if (endptr != comma)
1952 return -EINVAL;
1953
1954 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1955 if (endptr == comma + 1 || *endptr != '\0')
1956 return -EINVAL;
1957 }
1958
1959 return 0;
1960}
1961
1962void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1963{
1964 if (*ctl_fd_close) {
1965 *ctl_fd_close = false;
1966 close(ctl_fd);
1967 if (ctl_fd_ack >= 0)
1968 close(ctl_fd_ack);
1969 }
1970}
1971
1972int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1973{
1974 if (fd == -1) {
1975 pr_debug("Control descriptor is not initialized\n");
1976 return 0;
1977 }
1978
1979 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1980 fdarray_flag__nonfilterable |
1981 fdarray_flag__non_perf_event);
1982 if (evlist->ctl_fd.pos < 0) {
1983 evlist->ctl_fd.pos = -1;
1984 pr_err("Failed to add ctl fd entry: %m\n");
1985 return -1;
1986 }
1987
1988 evlist->ctl_fd.fd = fd;
1989 evlist->ctl_fd.ack = ack;
1990
1991 return 0;
1992}
1993
1994bool evlist__ctlfd_initialized(struct evlist *evlist)
1995{
1996 return evlist->ctl_fd.pos >= 0;
1997}
1998
1999int evlist__finalize_ctlfd(struct evlist *evlist)
2000{
2001 struct pollfd *entries = evlist->core.pollfd.entries;
2002
2003 if (!evlist__ctlfd_initialized(evlist))
2004 return 0;
2005
2006 entries[evlist->ctl_fd.pos].fd = -1;
2007 entries[evlist->ctl_fd.pos].events = 0;
2008 entries[evlist->ctl_fd.pos].revents = 0;
2009
2010 evlist->ctl_fd.pos = -1;
2011 evlist->ctl_fd.ack = -1;
2012 evlist->ctl_fd.fd = -1;
2013
2014 return 0;
2015}
2016
2017static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
2018 char *cmd_data, size_t data_size)
2019{
2020 int err;
2021 char c;
2022 size_t bytes_read = 0;
2023
2024 *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2025 memset(cmd_data, 0, data_size);
2026 data_size--;
2027
2028 do {
2029 err = read(evlist->ctl_fd.fd, &c, 1);
2030 if (err > 0) {
2031 if (c == '\n' || c == '\0')
2032 break;
2033 cmd_data[bytes_read++] = c;
2034 if (bytes_read == data_size)
2035 break;
2036 continue;
2037 } else if (err == -1) {
2038 if (errno == EINTR)
2039 continue;
2040 if (errno == EAGAIN || errno == EWOULDBLOCK)
2041 err = 0;
2042 else
2043 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
2044 }
2045 break;
2046 } while (1);
2047
2048 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
2049 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
2050
2051 if (bytes_read > 0) {
2052 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
2053 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
2054 *cmd = EVLIST_CTL_CMD_ENABLE;
2055 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
2056 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
2057 *cmd = EVLIST_CTL_CMD_DISABLE;
2058 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2059 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2060 *cmd = EVLIST_CTL_CMD_SNAPSHOT;
2061 pr_debug("is snapshot\n");
2062 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2063 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2064 *cmd = EVLIST_CTL_CMD_EVLIST;
2065 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2066 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2067 *cmd = EVLIST_CTL_CMD_STOP;
2068 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2069 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2070 *cmd = EVLIST_CTL_CMD_PING;
2071 }
2072 }
2073
2074 return bytes_read ? (int)bytes_read : err;
2075}
2076
2077int evlist__ctlfd_ack(struct evlist *evlist)
2078{
2079 int err;
2080
2081 if (evlist->ctl_fd.ack == -1)
2082 return 0;
2083
2084 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2085 sizeof(EVLIST_CTL_CMD_ACK_TAG));
2086 if (err == -1)
2087 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2088
2089 return err;
2090}
2091
2092static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2093{
2094 char *data = cmd_data + cmd_size;
2095
2096 /* no argument */
2097 if (!*data)
2098 return 0;
2099
2100 /* there's argument */
2101 if (*data == ' ') {
2102 *arg = data + 1;
2103 return 1;
2104 }
2105
2106 /* malformed */
2107 return -1;
2108}
2109
2110static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2111{
2112 struct evsel *evsel;
2113 char *name;
2114 int err;
2115
2116 err = get_cmd_arg(cmd_data,
2117 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2118 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2119 &name);
2120 if (err < 0) {
2121 pr_info("failed: wrong command\n");
2122 return -1;
2123 }
2124
2125 if (err) {
2126 evsel = evlist__find_evsel_by_str(evlist, name);
2127 if (evsel) {
2128 if (enable)
2129 evlist__enable_evsel(evlist, name);
2130 else
2131 evlist__disable_evsel(evlist, name);
2132 pr_info("Event %s %s\n", evsel->name,
2133 enable ? "enabled" : "disabled");
2134 } else {
2135 pr_info("failed: can't find '%s' event\n", name);
2136 }
2137 } else {
2138 if (enable) {
2139 evlist__enable(evlist);
2140 pr_info(EVLIST_ENABLED_MSG);
2141 } else {
2142 evlist__disable(evlist);
2143 pr_info(EVLIST_DISABLED_MSG);
2144 }
2145 }
2146
2147 return 0;
2148}
2149
2150static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2151{
2152 struct perf_attr_details details = { .verbose = false, };
2153 struct evsel *evsel;
2154 char *arg;
2155 int err;
2156
2157 err = get_cmd_arg(cmd_data,
2158 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2159 &arg);
2160 if (err < 0) {
2161 pr_info("failed: wrong command\n");
2162 return -1;
2163 }
2164
2165 if (err) {
2166 if (!strcmp(arg, "-v")) {
2167 details.verbose = true;
2168 } else if (!strcmp(arg, "-g")) {
2169 details.event_group = true;
2170 } else if (!strcmp(arg, "-F")) {
2171 details.freq = true;
2172 } else {
2173 pr_info("failed: wrong command\n");
2174 return -1;
2175 }
2176 }
2177
2178 evlist__for_each_entry(evlist, evsel)
2179 evsel__fprintf(evsel, &details, stderr);
2180
2181 return 0;
2182}
2183
2184int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2185{
2186 int err = 0;
2187 char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2188 int ctlfd_pos = evlist->ctl_fd.pos;
2189 struct pollfd *entries = evlist->core.pollfd.entries;
2190
2191 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2192 return 0;
2193
2194 if (entries[ctlfd_pos].revents & POLLIN) {
2195 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2196 EVLIST_CTL_CMD_MAX_LEN);
2197 if (err > 0) {
2198 switch (*cmd) {
2199 case EVLIST_CTL_CMD_ENABLE:
2200 case EVLIST_CTL_CMD_DISABLE:
2201 err = evlist__ctlfd_enable(evlist, cmd_data,
2202 *cmd == EVLIST_CTL_CMD_ENABLE);
2203 break;
2204 case EVLIST_CTL_CMD_EVLIST:
2205 err = evlist__ctlfd_list(evlist, cmd_data);
2206 break;
2207 case EVLIST_CTL_CMD_SNAPSHOT:
2208 case EVLIST_CTL_CMD_STOP:
2209 case EVLIST_CTL_CMD_PING:
2210 break;
2211 case EVLIST_CTL_CMD_ACK:
2212 case EVLIST_CTL_CMD_UNSUPPORTED:
2213 default:
2214 pr_debug("ctlfd: unsupported %d\n", *cmd);
2215 break;
2216 }
2217 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2218 *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2219 evlist__ctlfd_ack(evlist);
2220 }
2221 }
2222
2223 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2224 evlist__finalize_ctlfd(evlist);
2225 else
2226 entries[ctlfd_pos].revents = 0;
2227
2228 return err;
2229}
2230
2231/**
2232 * struct event_enable_time - perf record -D/--delay single time range.
2233 * @start: start of time range to enable events in milliseconds
2234 * @end: end of time range to enable events in milliseconds
2235 *
2236 * N.B. this structure is also accessed as an array of int.
2237 */
2238struct event_enable_time {
2239 int start;
2240 int end;
2241};
2242
2243static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2244{
2245 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2246 int ret, start, end, n;
2247
2248 ret = sscanf(str, fmt, &start, &end, &n);
2249 if (ret != 2 || end <= start)
2250 return -EINVAL;
2251 if (range) {
2252 range->start = start;
2253 range->end = end;
2254 }
2255 return n;
2256}
2257
2258static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2259{
2260 int incr = !!range;
2261 bool first = true;
2262 ssize_t ret, cnt;
2263
2264 for (cnt = 0; *str; cnt++) {
2265 ret = parse_event_enable_time(str, range, first);
2266 if (ret < 0)
2267 return ret;
2268 /* Check no overlap */
2269 if (!first && range && range->start <= range[-1].end)
2270 return -EINVAL;
2271 str += ret;
2272 range += incr;
2273 first = false;
2274 }
2275 return cnt;
2276}
2277
2278/**
2279 * struct event_enable_timer - control structure for perf record -D/--delay.
2280 * @evlist: event list
2281 * @times: time ranges that events are enabled (N.B. this is also accessed as an
2282 * array of int)
2283 * @times_cnt: number of time ranges
2284 * @timerfd: timer file descriptor
2285 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2286 * @times_step: current position in (int *)@times)[],
2287 * refer event_enable_timer__process()
2288 *
2289 * Note, this structure is only used when there are time ranges, not when there
2290 * is only an initial delay.
2291 */
2292struct event_enable_timer {
2293 struct evlist *evlist;
2294 struct event_enable_time *times;
2295 size_t times_cnt;
2296 int timerfd;
2297 int pollfd_pos;
2298 size_t times_step;
2299};
2300
2301static int str_to_delay(const char *str)
2302{
2303 char *endptr;
2304 long d;
2305
2306 d = strtol(str, &endptr, 10);
2307 if (*endptr || d > INT_MAX || d < -1)
2308 return 0;
2309 return d;
2310}
2311
2312int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2313 const char *str, int unset)
2314{
2315 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2316 struct event_enable_timer *eet;
2317 ssize_t times_cnt;
2318 ssize_t ret;
2319 int err;
2320
2321 if (unset)
2322 return 0;
2323
2324 opts->target.initial_delay = str_to_delay(str);
2325 if (opts->target.initial_delay)
2326 return 0;
2327
2328 ret = parse_event_enable_times(str, NULL);
2329 if (ret < 0)
2330 return ret;
2331
2332 times_cnt = ret;
2333 if (times_cnt == 0)
2334 return -EINVAL;
2335
2336 eet = zalloc(sizeof(*eet));
2337 if (!eet)
2338 return -ENOMEM;
2339
2340 eet->times = calloc(times_cnt, sizeof(*eet->times));
2341 if (!eet->times) {
2342 err = -ENOMEM;
2343 goto free_eet;
2344 }
2345
2346 if (parse_event_enable_times(str, eet->times) != times_cnt) {
2347 err = -EINVAL;
2348 goto free_eet_times;
2349 }
2350
2351 eet->times_cnt = times_cnt;
2352
2353 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2354 if (eet->timerfd == -1) {
2355 err = -errno;
2356 pr_err("timerfd_create failed: %s\n", strerror(errno));
2357 goto free_eet_times;
2358 }
2359
2360 eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2361 if (eet->pollfd_pos < 0) {
2362 err = eet->pollfd_pos;
2363 goto close_timerfd;
2364 }
2365
2366 eet->evlist = evlist;
2367 evlist->eet = eet;
2368 opts->target.initial_delay = eet->times[0].start;
2369
2370 return 0;
2371
2372close_timerfd:
2373 close(eet->timerfd);
2374free_eet_times:
2375 zfree(&eet->times);
2376free_eet:
2377 free(eet);
2378 return err;
2379}
2380
2381static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2382{
2383 struct itimerspec its = {
2384 .it_value.tv_sec = ms / MSEC_PER_SEC,
2385 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2386 };
2387 int err = 0;
2388
2389 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2390 err = -errno;
2391 pr_err("timerfd_settime failed: %s\n", strerror(errno));
2392 }
2393 return err;
2394}
2395
2396int event_enable_timer__start(struct event_enable_timer *eet)
2397{
2398 int ms;
2399
2400 if (!eet)
2401 return 0;
2402
2403 ms = eet->times[0].end - eet->times[0].start;
2404 eet->times_step = 1;
2405
2406 return event_enable_timer__set_timer(eet, ms);
2407}
2408
2409int event_enable_timer__process(struct event_enable_timer *eet)
2410{
2411 struct pollfd *entries;
2412 short revents;
2413
2414 if (!eet)
2415 return 0;
2416
2417 entries = eet->evlist->core.pollfd.entries;
2418 revents = entries[eet->pollfd_pos].revents;
2419 entries[eet->pollfd_pos].revents = 0;
2420
2421 if (revents & POLLIN) {
2422 size_t step = eet->times_step;
2423 size_t pos = step / 2;
2424
2425 if (step & 1) {
2426 evlist__disable_non_dummy(eet->evlist);
2427 pr_info(EVLIST_DISABLED_MSG);
2428 if (pos >= eet->times_cnt - 1) {
2429 /* Disarm timer */
2430 event_enable_timer__set_timer(eet, 0);
2431 return 1; /* Stop */
2432 }
2433 } else {
2434 evlist__enable_non_dummy(eet->evlist);
2435 pr_info(EVLIST_ENABLED_MSG);
2436 }
2437
2438 step += 1;
2439 pos = step / 2;
2440
2441 if (pos < eet->times_cnt) {
2442 int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2443 int ms = times[step] - times[step - 1];
2444
2445 eet->times_step = step;
2446 return event_enable_timer__set_timer(eet, ms);
2447 }
2448 }
2449
2450 return 0;
2451}
2452
2453void event_enable_timer__exit(struct event_enable_timer **ep)
2454{
2455 if (!ep || !*ep)
2456 return;
2457 zfree(&(*ep)->times);
2458 zfree(ep);
2459}
2460
2461struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2462{
2463 struct evsel *evsel;
2464
2465 evlist__for_each_entry(evlist, evsel) {
2466 if (evsel->core.idx == idx)
2467 return evsel;
2468 }
2469 return NULL;
2470}
2471
2472int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
2473{
2474 struct evsel *evsel;
2475 int printed = 0;
2476
2477 evlist__for_each_entry(evlist, evsel) {
2478 if (evsel__is_dummy_event(evsel))
2479 continue;
2480 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) {
2481 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel));
2482 } else {
2483 printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : "");
2484 break;
2485 }
2486 }
2487
2488 return printed;
2489}
2490
2491void evlist__check_mem_load_aux(struct evlist *evlist)
2492{
2493 struct evsel *leader, *evsel, *pos;
2494
2495 /*
2496 * For some platforms, the 'mem-loads' event is required to use
2497 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2498 * must be the group leader. Now we disable this group before reporting
2499 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2500 * any valid memory load information.
2501 */
2502 evlist__for_each_entry(evlist, evsel) {
2503 leader = evsel__leader(evsel);
2504 if (leader == evsel)
2505 continue;
2506
2507 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2508 for_each_group_evsel(pos, leader) {
2509 evsel__set_leader(pos, pos);
2510 pos->core.nr_members = 0;
2511 }
2512 }
2513 }
2514}
2515
2516/**
2517 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs
2518 * and warn if the user CPU list is inapplicable for the event's PMU's
2519 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a
2520 * user requested CPU and so any online CPU is applicable. Core PMUs handle
2521 * events on the CPUs in their list and otherwise the event isn't supported.
2522 * @evlist: The list of events being checked.
2523 * @cpu_list: The user provided list of CPUs.
2524 */
2525void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list)
2526{
2527 struct perf_cpu_map *user_requested_cpus;
2528 struct evsel *pos;
2529
2530 if (!cpu_list)
2531 return;
2532
2533 user_requested_cpus = perf_cpu_map__new(cpu_list);
2534 if (!user_requested_cpus)
2535 return;
2536
2537 evlist__for_each_entry(evlist, pos) {
2538 struct perf_cpu_map *intersect, *to_test;
2539 const struct perf_pmu *pmu = evsel__find_pmu(pos);
2540
2541 to_test = pmu && pmu->is_core ? pmu->cpus : cpu_map__online();
2542 intersect = perf_cpu_map__intersect(to_test, user_requested_cpus);
2543 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) {
2544 char buf[128];
2545
2546 cpu_map__snprint(to_test, buf, sizeof(buf));
2547 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n",
2548 cpu_list, pmu ? pmu->name : "cpu", buf, evsel__name(pos));
2549 }
2550 perf_cpu_map__put(intersect);
2551 }
2552 perf_cpu_map__put(user_requested_cpus);
2553}
2554
2555void evlist__uniquify_name(struct evlist *evlist)
2556{
2557 char *new_name, empty_attributes[2] = ":", *attributes;
2558 struct evsel *pos;
2559
2560 if (perf_pmus__num_core_pmus() == 1)
2561 return;
2562
2563 evlist__for_each_entry(evlist, pos) {
2564 if (!evsel__is_hybrid(pos))
2565 continue;
2566
2567 if (strchr(pos->name, '/'))
2568 continue;
2569
2570 attributes = strchr(pos->name, ':');
2571 if (attributes)
2572 *attributes = '\0';
2573 else
2574 attributes = empty_attributes;
2575
2576 if (asprintf(&new_name, "%s/%s/%s", pos->pmu ? pos->pmu->name : "",
2577 pos->name, attributes + 1)) {
2578 free(pos->name);
2579 pos->name = new_name;
2580 } else {
2581 *attributes = ':';
2582 }
2583 }
2584}
2585
2586bool evlist__has_bpf_output(struct evlist *evlist)
2587{
2588 struct evsel *evsel;
2589
2590 evlist__for_each_entry(evlist, evsel) {
2591 if (evsel__is_bpf_output(evsel))
2592 return true;
2593 }
2594
2595 return false;
2596}