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/*
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/debugfs.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 "parse-options.h"
22
23#include <sys/mman.h>
24
25#include <linux/bitops.h>
26#include <linux/hash.h>
27
28#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
29#define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
30
31void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
32 struct thread_map *threads)
33{
34 int i;
35
36 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
37 INIT_HLIST_HEAD(&evlist->heads[i]);
38 INIT_LIST_HEAD(&evlist->entries);
39 perf_evlist__set_maps(evlist, cpus, threads);
40 evlist->workload.pid = -1;
41}
42
43struct perf_evlist *perf_evlist__new(void)
44{
45 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
46
47 if (evlist != NULL)
48 perf_evlist__init(evlist, NULL, NULL);
49
50 return evlist;
51}
52
53struct perf_evlist *perf_evlist__new_default(void)
54{
55 struct perf_evlist *evlist = perf_evlist__new();
56
57 if (evlist && perf_evlist__add_default(evlist)) {
58 perf_evlist__delete(evlist);
59 evlist = NULL;
60 }
61
62 return evlist;
63}
64
65/**
66 * perf_evlist__set_id_pos - set the positions of event ids.
67 * @evlist: selected event list
68 *
69 * Events with compatible sample types all have the same id_pos
70 * and is_pos. For convenience, put a copy on evlist.
71 */
72void perf_evlist__set_id_pos(struct perf_evlist *evlist)
73{
74 struct perf_evsel *first = perf_evlist__first(evlist);
75
76 evlist->id_pos = first->id_pos;
77 evlist->is_pos = first->is_pos;
78}
79
80static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
81{
82 struct perf_evsel *evsel;
83
84 evlist__for_each(evlist, evsel)
85 perf_evsel__calc_id_pos(evsel);
86
87 perf_evlist__set_id_pos(evlist);
88}
89
90static void perf_evlist__purge(struct perf_evlist *evlist)
91{
92 struct perf_evsel *pos, *n;
93
94 evlist__for_each_safe(evlist, n, pos) {
95 list_del_init(&pos->node);
96 perf_evsel__delete(pos);
97 }
98
99 evlist->nr_entries = 0;
100}
101
102void perf_evlist__exit(struct perf_evlist *evlist)
103{
104 zfree(&evlist->mmap);
105 zfree(&evlist->pollfd);
106}
107
108void perf_evlist__delete(struct perf_evlist *evlist)
109{
110 perf_evlist__munmap(evlist);
111 perf_evlist__close(evlist);
112 cpu_map__delete(evlist->cpus);
113 thread_map__delete(evlist->threads);
114 evlist->cpus = NULL;
115 evlist->threads = NULL;
116 perf_evlist__purge(evlist);
117 perf_evlist__exit(evlist);
118 free(evlist);
119}
120
121void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
122{
123 list_add_tail(&entry->node, &evlist->entries);
124 entry->idx = evlist->nr_entries;
125
126 if (!evlist->nr_entries++)
127 perf_evlist__set_id_pos(evlist);
128}
129
130void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
131 struct list_head *list,
132 int nr_entries)
133{
134 bool set_id_pos = !evlist->nr_entries;
135
136 list_splice_tail(list, &evlist->entries);
137 evlist->nr_entries += nr_entries;
138 if (set_id_pos)
139 perf_evlist__set_id_pos(evlist);
140}
141
142void __perf_evlist__set_leader(struct list_head *list)
143{
144 struct perf_evsel *evsel, *leader;
145
146 leader = list_entry(list->next, struct perf_evsel, node);
147 evsel = list_entry(list->prev, struct perf_evsel, node);
148
149 leader->nr_members = evsel->idx - leader->idx + 1;
150
151 __evlist__for_each(list, evsel) {
152 evsel->leader = leader;
153 }
154}
155
156void perf_evlist__set_leader(struct perf_evlist *evlist)
157{
158 if (evlist->nr_entries) {
159 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
160 __perf_evlist__set_leader(&evlist->entries);
161 }
162}
163
164int perf_evlist__add_default(struct perf_evlist *evlist)
165{
166 struct perf_event_attr attr = {
167 .type = PERF_TYPE_HARDWARE,
168 .config = PERF_COUNT_HW_CPU_CYCLES,
169 };
170 struct perf_evsel *evsel;
171
172 event_attr_init(&attr);
173
174 evsel = perf_evsel__new(&attr);
175 if (evsel == NULL)
176 goto error;
177
178 /* use strdup() because free(evsel) assumes name is allocated */
179 evsel->name = strdup("cycles");
180 if (!evsel->name)
181 goto error_free;
182
183 perf_evlist__add(evlist, evsel);
184 return 0;
185error_free:
186 perf_evsel__delete(evsel);
187error:
188 return -ENOMEM;
189}
190
191static int perf_evlist__add_attrs(struct perf_evlist *evlist,
192 struct perf_event_attr *attrs, size_t nr_attrs)
193{
194 struct perf_evsel *evsel, *n;
195 LIST_HEAD(head);
196 size_t i;
197
198 for (i = 0; i < nr_attrs; i++) {
199 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
200 if (evsel == NULL)
201 goto out_delete_partial_list;
202 list_add_tail(&evsel->node, &head);
203 }
204
205 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
206
207 return 0;
208
209out_delete_partial_list:
210 __evlist__for_each_safe(&head, n, evsel)
211 perf_evsel__delete(evsel);
212 return -1;
213}
214
215int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
216 struct perf_event_attr *attrs, size_t nr_attrs)
217{
218 size_t i;
219
220 for (i = 0; i < nr_attrs; i++)
221 event_attr_init(attrs + i);
222
223 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
224}
225
226struct perf_evsel *
227perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
228{
229 struct perf_evsel *evsel;
230
231 evlist__for_each(evlist, evsel) {
232 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
233 (int)evsel->attr.config == id)
234 return evsel;
235 }
236
237 return NULL;
238}
239
240struct perf_evsel *
241perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
242 const char *name)
243{
244 struct perf_evsel *evsel;
245
246 evlist__for_each(evlist, evsel) {
247 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
248 (strcmp(evsel->name, name) == 0))
249 return evsel;
250 }
251
252 return NULL;
253}
254
255int perf_evlist__add_newtp(struct perf_evlist *evlist,
256 const char *sys, const char *name, void *handler)
257{
258 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
259
260 if (evsel == NULL)
261 return -1;
262
263 evsel->handler = handler;
264 perf_evlist__add(evlist, evsel);
265 return 0;
266}
267
268void perf_evlist__disable(struct perf_evlist *evlist)
269{
270 int cpu, thread;
271 struct perf_evsel *pos;
272 int nr_cpus = cpu_map__nr(evlist->cpus);
273 int nr_threads = thread_map__nr(evlist->threads);
274
275 for (cpu = 0; cpu < nr_cpus; cpu++) {
276 evlist__for_each(evlist, pos) {
277 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
278 continue;
279 for (thread = 0; thread < nr_threads; thread++)
280 ioctl(FD(pos, cpu, thread),
281 PERF_EVENT_IOC_DISABLE, 0);
282 }
283 }
284}
285
286void perf_evlist__enable(struct perf_evlist *evlist)
287{
288 int cpu, thread;
289 struct perf_evsel *pos;
290 int nr_cpus = cpu_map__nr(evlist->cpus);
291 int nr_threads = thread_map__nr(evlist->threads);
292
293 for (cpu = 0; cpu < nr_cpus; cpu++) {
294 evlist__for_each(evlist, pos) {
295 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
296 continue;
297 for (thread = 0; thread < nr_threads; thread++)
298 ioctl(FD(pos, cpu, thread),
299 PERF_EVENT_IOC_ENABLE, 0);
300 }
301 }
302}
303
304int perf_evlist__disable_event(struct perf_evlist *evlist,
305 struct perf_evsel *evsel)
306{
307 int cpu, thread, err;
308
309 if (!evsel->fd)
310 return 0;
311
312 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
313 for (thread = 0; thread < evlist->threads->nr; thread++) {
314 err = ioctl(FD(evsel, cpu, thread),
315 PERF_EVENT_IOC_DISABLE, 0);
316 if (err)
317 return err;
318 }
319 }
320 return 0;
321}
322
323int perf_evlist__enable_event(struct perf_evlist *evlist,
324 struct perf_evsel *evsel)
325{
326 int cpu, thread, err;
327
328 if (!evsel->fd)
329 return -EINVAL;
330
331 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
332 for (thread = 0; thread < evlist->threads->nr; thread++) {
333 err = ioctl(FD(evsel, cpu, thread),
334 PERF_EVENT_IOC_ENABLE, 0);
335 if (err)
336 return err;
337 }
338 }
339 return 0;
340}
341
342static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
343{
344 int nr_cpus = cpu_map__nr(evlist->cpus);
345 int nr_threads = thread_map__nr(evlist->threads);
346 int nfds = nr_cpus * nr_threads * evlist->nr_entries;
347 evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
348 return evlist->pollfd != NULL ? 0 : -ENOMEM;
349}
350
351void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
352{
353 fcntl(fd, F_SETFL, O_NONBLOCK);
354 evlist->pollfd[evlist->nr_fds].fd = fd;
355 evlist->pollfd[evlist->nr_fds].events = POLLIN;
356 evlist->nr_fds++;
357}
358
359static void perf_evlist__id_hash(struct perf_evlist *evlist,
360 struct perf_evsel *evsel,
361 int cpu, int thread, u64 id)
362{
363 int hash;
364 struct perf_sample_id *sid = SID(evsel, cpu, thread);
365
366 sid->id = id;
367 sid->evsel = evsel;
368 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
369 hlist_add_head(&sid->node, &evlist->heads[hash]);
370}
371
372void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
373 int cpu, int thread, u64 id)
374{
375 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
376 evsel->id[evsel->ids++] = id;
377}
378
379static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
380 struct perf_evsel *evsel,
381 int cpu, int thread, int fd)
382{
383 u64 read_data[4] = { 0, };
384 int id_idx = 1; /* The first entry is the counter value */
385 u64 id;
386 int ret;
387
388 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
389 if (!ret)
390 goto add;
391
392 if (errno != ENOTTY)
393 return -1;
394
395 /* Legacy way to get event id.. All hail to old kernels! */
396
397 /*
398 * This way does not work with group format read, so bail
399 * out in that case.
400 */
401 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
402 return -1;
403
404 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
405 read(fd, &read_data, sizeof(read_data)) == -1)
406 return -1;
407
408 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
409 ++id_idx;
410 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
411 ++id_idx;
412
413 id = read_data[id_idx];
414
415 add:
416 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
417 return 0;
418}
419
420struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
421{
422 struct hlist_head *head;
423 struct perf_sample_id *sid;
424 int hash;
425
426 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
427 head = &evlist->heads[hash];
428
429 hlist_for_each_entry(sid, head, node)
430 if (sid->id == id)
431 return sid;
432
433 return NULL;
434}
435
436struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
437{
438 struct perf_sample_id *sid;
439
440 if (evlist->nr_entries == 1)
441 return perf_evlist__first(evlist);
442
443 sid = perf_evlist__id2sid(evlist, id);
444 if (sid)
445 return sid->evsel;
446
447 if (!perf_evlist__sample_id_all(evlist))
448 return perf_evlist__first(evlist);
449
450 return NULL;
451}
452
453static int perf_evlist__event2id(struct perf_evlist *evlist,
454 union perf_event *event, u64 *id)
455{
456 const u64 *array = event->sample.array;
457 ssize_t n;
458
459 n = (event->header.size - sizeof(event->header)) >> 3;
460
461 if (event->header.type == PERF_RECORD_SAMPLE) {
462 if (evlist->id_pos >= n)
463 return -1;
464 *id = array[evlist->id_pos];
465 } else {
466 if (evlist->is_pos > n)
467 return -1;
468 n -= evlist->is_pos;
469 *id = array[n];
470 }
471 return 0;
472}
473
474static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
475 union perf_event *event)
476{
477 struct perf_evsel *first = perf_evlist__first(evlist);
478 struct hlist_head *head;
479 struct perf_sample_id *sid;
480 int hash;
481 u64 id;
482
483 if (evlist->nr_entries == 1)
484 return first;
485
486 if (!first->attr.sample_id_all &&
487 event->header.type != PERF_RECORD_SAMPLE)
488 return first;
489
490 if (perf_evlist__event2id(evlist, event, &id))
491 return NULL;
492
493 /* Synthesized events have an id of zero */
494 if (!id)
495 return first;
496
497 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
498 head = &evlist->heads[hash];
499
500 hlist_for_each_entry(sid, head, node) {
501 if (sid->id == id)
502 return sid->evsel;
503 }
504 return NULL;
505}
506
507union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
508{
509 struct perf_mmap *md = &evlist->mmap[idx];
510 unsigned int head = perf_mmap__read_head(md);
511 unsigned int old = md->prev;
512 unsigned char *data = md->base + page_size;
513 union perf_event *event = NULL;
514
515 if (evlist->overwrite) {
516 /*
517 * If we're further behind than half the buffer, there's a chance
518 * the writer will bite our tail and mess up the samples under us.
519 *
520 * If we somehow ended up ahead of the head, we got messed up.
521 *
522 * In either case, truncate and restart at head.
523 */
524 int diff = head - old;
525 if (diff > md->mask / 2 || diff < 0) {
526 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
527
528 /*
529 * head points to a known good entry, start there.
530 */
531 old = head;
532 }
533 }
534
535 if (old != head) {
536 size_t size;
537
538 event = (union perf_event *)&data[old & md->mask];
539 size = event->header.size;
540
541 /*
542 * Event straddles the mmap boundary -- header should always
543 * be inside due to u64 alignment of output.
544 */
545 if ((old & md->mask) + size != ((old + size) & md->mask)) {
546 unsigned int offset = old;
547 unsigned int len = min(sizeof(*event), size), cpy;
548 void *dst = md->event_copy;
549
550 do {
551 cpy = min(md->mask + 1 - (offset & md->mask), len);
552 memcpy(dst, &data[offset & md->mask], cpy);
553 offset += cpy;
554 dst += cpy;
555 len -= cpy;
556 } while (len);
557
558 event = (union perf_event *) md->event_copy;
559 }
560
561 old += size;
562 }
563
564 md->prev = old;
565
566 return event;
567}
568
569void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
570{
571 if (!evlist->overwrite) {
572 struct perf_mmap *md = &evlist->mmap[idx];
573 unsigned int old = md->prev;
574
575 perf_mmap__write_tail(md, old);
576 }
577}
578
579static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
580{
581 if (evlist->mmap[idx].base != NULL) {
582 munmap(evlist->mmap[idx].base, evlist->mmap_len);
583 evlist->mmap[idx].base = NULL;
584 }
585}
586
587void perf_evlist__munmap(struct perf_evlist *evlist)
588{
589 int i;
590
591 if (evlist->mmap == NULL)
592 return;
593
594 for (i = 0; i < evlist->nr_mmaps; i++)
595 __perf_evlist__munmap(evlist, i);
596
597 zfree(&evlist->mmap);
598}
599
600static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
601{
602 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
603 if (cpu_map__empty(evlist->cpus))
604 evlist->nr_mmaps = thread_map__nr(evlist->threads);
605 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
606 return evlist->mmap != NULL ? 0 : -ENOMEM;
607}
608
609static int __perf_evlist__mmap(struct perf_evlist *evlist,
610 int idx, int prot, int mask, int fd)
611{
612 evlist->mmap[idx].prev = 0;
613 evlist->mmap[idx].mask = mask;
614 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
615 MAP_SHARED, fd, 0);
616 if (evlist->mmap[idx].base == MAP_FAILED) {
617 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
618 errno);
619 evlist->mmap[idx].base = NULL;
620 return -1;
621 }
622
623 perf_evlist__add_pollfd(evlist, fd);
624 return 0;
625}
626
627static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
628 int prot, int mask, int cpu, int thread,
629 int *output)
630{
631 struct perf_evsel *evsel;
632
633 evlist__for_each(evlist, evsel) {
634 int fd = FD(evsel, cpu, thread);
635
636 if (*output == -1) {
637 *output = fd;
638 if (__perf_evlist__mmap(evlist, idx, prot, mask,
639 *output) < 0)
640 return -1;
641 } else {
642 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
643 return -1;
644 }
645
646 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
647 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
648 return -1;
649 }
650
651 return 0;
652}
653
654static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot,
655 int mask)
656{
657 int cpu, thread;
658 int nr_cpus = cpu_map__nr(evlist->cpus);
659 int nr_threads = thread_map__nr(evlist->threads);
660
661 pr_debug2("perf event ring buffer mmapped per cpu\n");
662 for (cpu = 0; cpu < nr_cpus; cpu++) {
663 int output = -1;
664
665 for (thread = 0; thread < nr_threads; thread++) {
666 if (perf_evlist__mmap_per_evsel(evlist, cpu, prot, mask,
667 cpu, thread, &output))
668 goto out_unmap;
669 }
670 }
671
672 return 0;
673
674out_unmap:
675 for (cpu = 0; cpu < nr_cpus; cpu++)
676 __perf_evlist__munmap(evlist, cpu);
677 return -1;
678}
679
680static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot,
681 int mask)
682{
683 int thread;
684 int nr_threads = thread_map__nr(evlist->threads);
685
686 pr_debug2("perf event ring buffer mmapped per thread\n");
687 for (thread = 0; thread < nr_threads; thread++) {
688 int output = -1;
689
690 if (perf_evlist__mmap_per_evsel(evlist, thread, prot, mask, 0,
691 thread, &output))
692 goto out_unmap;
693 }
694
695 return 0;
696
697out_unmap:
698 for (thread = 0; thread < nr_threads; thread++)
699 __perf_evlist__munmap(evlist, thread);
700 return -1;
701}
702
703static size_t perf_evlist__mmap_size(unsigned long pages)
704{
705 /* 512 kiB: default amount of unprivileged mlocked memory */
706 if (pages == UINT_MAX)
707 pages = (512 * 1024) / page_size;
708 else if (!is_power_of_2(pages))
709 return 0;
710
711 return (pages + 1) * page_size;
712}
713
714static long parse_pages_arg(const char *str, unsigned long min,
715 unsigned long max)
716{
717 unsigned long pages, val;
718 static struct parse_tag tags[] = {
719 { .tag = 'B', .mult = 1 },
720 { .tag = 'K', .mult = 1 << 10 },
721 { .tag = 'M', .mult = 1 << 20 },
722 { .tag = 'G', .mult = 1 << 30 },
723 { .tag = 0 },
724 };
725
726 if (str == NULL)
727 return -EINVAL;
728
729 val = parse_tag_value(str, tags);
730 if (val != (unsigned long) -1) {
731 /* we got file size value */
732 pages = PERF_ALIGN(val, page_size) / page_size;
733 } else {
734 /* we got pages count value */
735 char *eptr;
736 pages = strtoul(str, &eptr, 10);
737 if (*eptr != '\0')
738 return -EINVAL;
739 }
740
741 if (pages == 0 && min == 0) {
742 /* leave number of pages at 0 */
743 } else if (!is_power_of_2(pages)) {
744 /* round pages up to next power of 2 */
745 pages = next_pow2_l(pages);
746 if (!pages)
747 return -EINVAL;
748 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
749 pages * page_size, pages);
750 }
751
752 if (pages > max)
753 return -EINVAL;
754
755 return pages;
756}
757
758int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
759 int unset __maybe_unused)
760{
761 unsigned int *mmap_pages = opt->value;
762 unsigned long max = UINT_MAX;
763 long pages;
764
765 if (max > SIZE_MAX / page_size)
766 max = SIZE_MAX / page_size;
767
768 pages = parse_pages_arg(str, 1, max);
769 if (pages < 0) {
770 pr_err("Invalid argument for --mmap_pages/-m\n");
771 return -1;
772 }
773
774 *mmap_pages = pages;
775 return 0;
776}
777
778/**
779 * perf_evlist__mmap - Create mmaps to receive events.
780 * @evlist: list of events
781 * @pages: map length in pages
782 * @overwrite: overwrite older events?
783 *
784 * If @overwrite is %false the user needs to signal event consumption using
785 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
786 * automatically.
787 *
788 * Return: %0 on success, negative error code otherwise.
789 */
790int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
791 bool overwrite)
792{
793 struct perf_evsel *evsel;
794 const struct cpu_map *cpus = evlist->cpus;
795 const struct thread_map *threads = evlist->threads;
796 int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
797
798 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
799 return -ENOMEM;
800
801 if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
802 return -ENOMEM;
803
804 evlist->overwrite = overwrite;
805 evlist->mmap_len = perf_evlist__mmap_size(pages);
806 pr_debug("mmap size %zuB\n", evlist->mmap_len);
807 mask = evlist->mmap_len - page_size - 1;
808
809 evlist__for_each(evlist, evsel) {
810 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
811 evsel->sample_id == NULL &&
812 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
813 return -ENOMEM;
814 }
815
816 if (cpu_map__empty(cpus))
817 return perf_evlist__mmap_per_thread(evlist, prot, mask);
818
819 return perf_evlist__mmap_per_cpu(evlist, prot, mask);
820}
821
822int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
823{
824 evlist->threads = thread_map__new_str(target->pid, target->tid,
825 target->uid);
826
827 if (evlist->threads == NULL)
828 return -1;
829
830 if (target__uses_dummy_map(target))
831 evlist->cpus = cpu_map__dummy_new();
832 else
833 evlist->cpus = cpu_map__new(target->cpu_list);
834
835 if (evlist->cpus == NULL)
836 goto out_delete_threads;
837
838 return 0;
839
840out_delete_threads:
841 thread_map__delete(evlist->threads);
842 return -1;
843}
844
845int perf_evlist__apply_filters(struct perf_evlist *evlist)
846{
847 struct perf_evsel *evsel;
848 int err = 0;
849 const int ncpus = cpu_map__nr(evlist->cpus),
850 nthreads = thread_map__nr(evlist->threads);
851
852 evlist__for_each(evlist, evsel) {
853 if (evsel->filter == NULL)
854 continue;
855
856 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
857 if (err)
858 break;
859 }
860
861 return err;
862}
863
864int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
865{
866 struct perf_evsel *evsel;
867 int err = 0;
868 const int ncpus = cpu_map__nr(evlist->cpus),
869 nthreads = thread_map__nr(evlist->threads);
870
871 evlist__for_each(evlist, evsel) {
872 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
873 if (err)
874 break;
875 }
876
877 return err;
878}
879
880bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
881{
882 struct perf_evsel *pos;
883
884 if (evlist->nr_entries == 1)
885 return true;
886
887 if (evlist->id_pos < 0 || evlist->is_pos < 0)
888 return false;
889
890 evlist__for_each(evlist, pos) {
891 if (pos->id_pos != evlist->id_pos ||
892 pos->is_pos != evlist->is_pos)
893 return false;
894 }
895
896 return true;
897}
898
899u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
900{
901 struct perf_evsel *evsel;
902
903 if (evlist->combined_sample_type)
904 return evlist->combined_sample_type;
905
906 evlist__for_each(evlist, evsel)
907 evlist->combined_sample_type |= evsel->attr.sample_type;
908
909 return evlist->combined_sample_type;
910}
911
912u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
913{
914 evlist->combined_sample_type = 0;
915 return __perf_evlist__combined_sample_type(evlist);
916}
917
918bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
919{
920 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
921 u64 read_format = first->attr.read_format;
922 u64 sample_type = first->attr.sample_type;
923
924 evlist__for_each(evlist, pos) {
925 if (read_format != pos->attr.read_format)
926 return false;
927 }
928
929 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
930 if ((sample_type & PERF_SAMPLE_READ) &&
931 !(read_format & PERF_FORMAT_ID)) {
932 return false;
933 }
934
935 return true;
936}
937
938u64 perf_evlist__read_format(struct perf_evlist *evlist)
939{
940 struct perf_evsel *first = perf_evlist__first(evlist);
941 return first->attr.read_format;
942}
943
944u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
945{
946 struct perf_evsel *first = perf_evlist__first(evlist);
947 struct perf_sample *data;
948 u64 sample_type;
949 u16 size = 0;
950
951 if (!first->attr.sample_id_all)
952 goto out;
953
954 sample_type = first->attr.sample_type;
955
956 if (sample_type & PERF_SAMPLE_TID)
957 size += sizeof(data->tid) * 2;
958
959 if (sample_type & PERF_SAMPLE_TIME)
960 size += sizeof(data->time);
961
962 if (sample_type & PERF_SAMPLE_ID)
963 size += sizeof(data->id);
964
965 if (sample_type & PERF_SAMPLE_STREAM_ID)
966 size += sizeof(data->stream_id);
967
968 if (sample_type & PERF_SAMPLE_CPU)
969 size += sizeof(data->cpu) * 2;
970
971 if (sample_type & PERF_SAMPLE_IDENTIFIER)
972 size += sizeof(data->id);
973out:
974 return size;
975}
976
977bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
978{
979 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
980
981 evlist__for_each_continue(evlist, pos) {
982 if (first->attr.sample_id_all != pos->attr.sample_id_all)
983 return false;
984 }
985
986 return true;
987}
988
989bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
990{
991 struct perf_evsel *first = perf_evlist__first(evlist);
992 return first->attr.sample_id_all;
993}
994
995void perf_evlist__set_selected(struct perf_evlist *evlist,
996 struct perf_evsel *evsel)
997{
998 evlist->selected = evsel;
999}
1000
1001void perf_evlist__close(struct perf_evlist *evlist)
1002{
1003 struct perf_evsel *evsel;
1004 int ncpus = cpu_map__nr(evlist->cpus);
1005 int nthreads = thread_map__nr(evlist->threads);
1006 int n;
1007
1008 evlist__for_each_reverse(evlist, evsel) {
1009 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1010 perf_evsel__close(evsel, n, nthreads);
1011 }
1012}
1013
1014int perf_evlist__open(struct perf_evlist *evlist)
1015{
1016 struct perf_evsel *evsel;
1017 int err;
1018
1019 perf_evlist__update_id_pos(evlist);
1020
1021 evlist__for_each(evlist, evsel) {
1022 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1023 if (err < 0)
1024 goto out_err;
1025 }
1026
1027 return 0;
1028out_err:
1029 perf_evlist__close(evlist);
1030 errno = -err;
1031 return err;
1032}
1033
1034int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1035 const char *argv[], bool pipe_output,
1036 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1037{
1038 int child_ready_pipe[2], go_pipe[2];
1039 char bf;
1040
1041 if (pipe(child_ready_pipe) < 0) {
1042 perror("failed to create 'ready' pipe");
1043 return -1;
1044 }
1045
1046 if (pipe(go_pipe) < 0) {
1047 perror("failed to create 'go' pipe");
1048 goto out_close_ready_pipe;
1049 }
1050
1051 evlist->workload.pid = fork();
1052 if (evlist->workload.pid < 0) {
1053 perror("failed to fork");
1054 goto out_close_pipes;
1055 }
1056
1057 if (!evlist->workload.pid) {
1058 if (pipe_output)
1059 dup2(2, 1);
1060
1061 signal(SIGTERM, SIG_DFL);
1062
1063 close(child_ready_pipe[0]);
1064 close(go_pipe[1]);
1065 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1066
1067 /*
1068 * Tell the parent we're ready to go
1069 */
1070 close(child_ready_pipe[1]);
1071
1072 /*
1073 * Wait until the parent tells us to go.
1074 */
1075 if (read(go_pipe[0], &bf, 1) == -1)
1076 perror("unable to read pipe");
1077
1078 execvp(argv[0], (char **)argv);
1079
1080 if (exec_error) {
1081 union sigval val;
1082
1083 val.sival_int = errno;
1084 if (sigqueue(getppid(), SIGUSR1, val))
1085 perror(argv[0]);
1086 } else
1087 perror(argv[0]);
1088 exit(-1);
1089 }
1090
1091 if (exec_error) {
1092 struct sigaction act = {
1093 .sa_flags = SA_SIGINFO,
1094 .sa_sigaction = exec_error,
1095 };
1096 sigaction(SIGUSR1, &act, NULL);
1097 }
1098
1099 if (target__none(target))
1100 evlist->threads->map[0] = evlist->workload.pid;
1101
1102 close(child_ready_pipe[1]);
1103 close(go_pipe[0]);
1104 /*
1105 * wait for child to settle
1106 */
1107 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1108 perror("unable to read pipe");
1109 goto out_close_pipes;
1110 }
1111
1112 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1113 evlist->workload.cork_fd = go_pipe[1];
1114 close(child_ready_pipe[0]);
1115 return 0;
1116
1117out_close_pipes:
1118 close(go_pipe[0]);
1119 close(go_pipe[1]);
1120out_close_ready_pipe:
1121 close(child_ready_pipe[0]);
1122 close(child_ready_pipe[1]);
1123 return -1;
1124}
1125
1126int perf_evlist__start_workload(struct perf_evlist *evlist)
1127{
1128 if (evlist->workload.cork_fd > 0) {
1129 char bf = 0;
1130 int ret;
1131 /*
1132 * Remove the cork, let it rip!
1133 */
1134 ret = write(evlist->workload.cork_fd, &bf, 1);
1135 if (ret < 0)
1136 perror("enable to write to pipe");
1137
1138 close(evlist->workload.cork_fd);
1139 return ret;
1140 }
1141
1142 return 0;
1143}
1144
1145int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1146 struct perf_sample *sample)
1147{
1148 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1149
1150 if (!evsel)
1151 return -EFAULT;
1152 return perf_evsel__parse_sample(evsel, event, sample);
1153}
1154
1155size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1156{
1157 struct perf_evsel *evsel;
1158 size_t printed = 0;
1159
1160 evlist__for_each(evlist, evsel) {
1161 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1162 perf_evsel__name(evsel));
1163 }
1164
1165 return printed + fprintf(fp, "\n");
1166}
1167
1168int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1169 int err, char *buf, size_t size)
1170{
1171 char sbuf[128];
1172
1173 switch (err) {
1174 case ENOENT:
1175 scnprintf(buf, size, "%s",
1176 "Error:\tUnable to find debugfs\n"
1177 "Hint:\tWas your kernel was compiled with debugfs support?\n"
1178 "Hint:\tIs the debugfs filesystem mounted?\n"
1179 "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1180 break;
1181 case EACCES:
1182 scnprintf(buf, size,
1183 "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1184 "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1185 debugfs_mountpoint, debugfs_mountpoint);
1186 break;
1187 default:
1188 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1189 break;
1190 }
1191
1192 return 0;
1193}
1194
1195int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1196 int err, char *buf, size_t size)
1197{
1198 int printed, value;
1199 char sbuf[128], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1200
1201 switch (err) {
1202 case EACCES:
1203 case EPERM:
1204 printed = scnprintf(buf, size,
1205 "Error:\t%s.\n"
1206 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1207
1208 value = perf_event_paranoid();
1209
1210 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1211
1212 if (value >= 2) {
1213 printed += scnprintf(buf + printed, size - printed,
1214 "For your workloads it needs to be <= 1\nHint:\t");
1215 }
1216 printed += scnprintf(buf + printed, size - printed,
1217 "For system wide tracing it needs to be set to -1");
1218
1219 printed += scnprintf(buf + printed, size - printed,
1220 ".\nHint:\tThe current value is %d.", value);
1221 break;
1222 default:
1223 scnprintf(buf, size, "%s", emsg);
1224 break;
1225 }
1226
1227 return 0;
1228}
1229
1230void perf_evlist__to_front(struct perf_evlist *evlist,
1231 struct perf_evsel *move_evsel)
1232{
1233 struct perf_evsel *evsel, *n;
1234 LIST_HEAD(move);
1235
1236 if (move_evsel == perf_evlist__first(evlist))
1237 return;
1238
1239 evlist__for_each_safe(evlist, n, evsel) {
1240 if (evsel->leader == move_evsel->leader)
1241 list_move_tail(&evsel->node, &move);
1242 }
1243
1244 list_splice(&move, &evlist->entries);
1245}