2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
10 #include <api/fs/debugfs.h>
13 #include "thread_map.h"
20 #include "parse-events.h"
21 #include "parse-options.h"
25 #include <linux/bitops.h>
26 #include <linux/hash.h>
28 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
29 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
31 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
32 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
34 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
35 struct thread_map *threads)
39 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
40 INIT_HLIST_HEAD(&evlist->heads[i]);
41 INIT_LIST_HEAD(&evlist->entries);
42 perf_evlist__set_maps(evlist, cpus, threads);
43 fdarray__init(&evlist->pollfd, 64);
44 evlist->workload.pid = -1;
47 struct perf_evlist *perf_evlist__new(void)
49 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
52 perf_evlist__init(evlist, NULL, NULL);
57 struct perf_evlist *perf_evlist__new_default(void)
59 struct perf_evlist *evlist = perf_evlist__new();
61 if (evlist && perf_evlist__add_default(evlist)) {
62 perf_evlist__delete(evlist);
70 * perf_evlist__set_id_pos - set the positions of event ids.
71 * @evlist: selected event list
73 * Events with compatible sample types all have the same id_pos
74 * and is_pos. For convenience, put a copy on evlist.
76 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
78 struct perf_evsel *first = perf_evlist__first(evlist);
80 evlist->id_pos = first->id_pos;
81 evlist->is_pos = first->is_pos;
84 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
86 struct perf_evsel *evsel;
88 evlist__for_each(evlist, evsel)
89 perf_evsel__calc_id_pos(evsel);
91 perf_evlist__set_id_pos(evlist);
94 static void perf_evlist__purge(struct perf_evlist *evlist)
96 struct perf_evsel *pos, *n;
98 evlist__for_each_safe(evlist, n, pos) {
99 list_del_init(&pos->node);
100 perf_evsel__delete(pos);
103 evlist->nr_entries = 0;
106 void perf_evlist__exit(struct perf_evlist *evlist)
108 zfree(&evlist->mmap);
109 fdarray__exit(&evlist->pollfd);
112 void perf_evlist__delete(struct perf_evlist *evlist)
114 perf_evlist__munmap(evlist);
115 perf_evlist__close(evlist);
116 cpu_map__delete(evlist->cpus);
117 thread_map__delete(evlist->threads);
119 evlist->threads = NULL;
120 perf_evlist__purge(evlist);
121 perf_evlist__exit(evlist);
125 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
127 list_add_tail(&entry->node, &evlist->entries);
128 entry->idx = evlist->nr_entries;
129 entry->tracking = !entry->idx;
131 if (!evlist->nr_entries++)
132 perf_evlist__set_id_pos(evlist);
135 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
136 struct list_head *list,
139 bool set_id_pos = !evlist->nr_entries;
141 list_splice_tail(list, &evlist->entries);
142 evlist->nr_entries += nr_entries;
144 perf_evlist__set_id_pos(evlist);
147 void __perf_evlist__set_leader(struct list_head *list)
149 struct perf_evsel *evsel, *leader;
151 leader = list_entry(list->next, struct perf_evsel, node);
152 evsel = list_entry(list->prev, struct perf_evsel, node);
154 leader->nr_members = evsel->idx - leader->idx + 1;
156 __evlist__for_each(list, evsel) {
157 evsel->leader = leader;
161 void perf_evlist__set_leader(struct perf_evlist *evlist)
163 if (evlist->nr_entries) {
164 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
165 __perf_evlist__set_leader(&evlist->entries);
169 int perf_evlist__add_default(struct perf_evlist *evlist)
171 struct perf_event_attr attr = {
172 .type = PERF_TYPE_HARDWARE,
173 .config = PERF_COUNT_HW_CPU_CYCLES,
175 struct perf_evsel *evsel;
177 event_attr_init(&attr);
179 evsel = perf_evsel__new(&attr);
183 /* use strdup() because free(evsel) assumes name is allocated */
184 evsel->name = strdup("cycles");
188 perf_evlist__add(evlist, evsel);
191 perf_evsel__delete(evsel);
196 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
197 struct perf_event_attr *attrs, size_t nr_attrs)
199 struct perf_evsel *evsel, *n;
203 for (i = 0; i < nr_attrs; i++) {
204 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
206 goto out_delete_partial_list;
207 list_add_tail(&evsel->node, &head);
210 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
214 out_delete_partial_list:
215 __evlist__for_each_safe(&head, n, evsel)
216 perf_evsel__delete(evsel);
220 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
221 struct perf_event_attr *attrs, size_t nr_attrs)
225 for (i = 0; i < nr_attrs; i++)
226 event_attr_init(attrs + i);
228 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
232 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
234 struct perf_evsel *evsel;
236 evlist__for_each(evlist, evsel) {
237 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
238 (int)evsel->attr.config == id)
246 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
249 struct perf_evsel *evsel;
251 evlist__for_each(evlist, evsel) {
252 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
253 (strcmp(evsel->name, name) == 0))
260 int perf_evlist__add_newtp(struct perf_evlist *evlist,
261 const char *sys, const char *name, void *handler)
263 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
268 evsel->handler = handler;
269 perf_evlist__add(evlist, evsel);
273 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
274 struct perf_evsel *evsel)
276 if (evsel->system_wide)
279 return thread_map__nr(evlist->threads);
282 void perf_evlist__disable(struct perf_evlist *evlist)
285 struct perf_evsel *pos;
286 int nr_cpus = cpu_map__nr(evlist->cpus);
289 for (cpu = 0; cpu < nr_cpus; cpu++) {
290 evlist__for_each(evlist, pos) {
291 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
293 nr_threads = perf_evlist__nr_threads(evlist, pos);
294 for (thread = 0; thread < nr_threads; thread++)
295 ioctl(FD(pos, cpu, thread),
296 PERF_EVENT_IOC_DISABLE, 0);
301 void perf_evlist__enable(struct perf_evlist *evlist)
304 struct perf_evsel *pos;
305 int nr_cpus = cpu_map__nr(evlist->cpus);
308 for (cpu = 0; cpu < nr_cpus; cpu++) {
309 evlist__for_each(evlist, pos) {
310 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
312 nr_threads = perf_evlist__nr_threads(evlist, pos);
313 for (thread = 0; thread < nr_threads; thread++)
314 ioctl(FD(pos, cpu, thread),
315 PERF_EVENT_IOC_ENABLE, 0);
320 int perf_evlist__disable_event(struct perf_evlist *evlist,
321 struct perf_evsel *evsel)
323 int cpu, thread, err;
324 int nr_cpus = cpu_map__nr(evlist->cpus);
325 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
330 for (cpu = 0; cpu < nr_cpus; cpu++) {
331 for (thread = 0; thread < nr_threads; thread++) {
332 err = ioctl(FD(evsel, cpu, thread),
333 PERF_EVENT_IOC_DISABLE, 0);
341 int perf_evlist__enable_event(struct perf_evlist *evlist,
342 struct perf_evsel *evsel)
344 int cpu, thread, err;
345 int nr_cpus = cpu_map__nr(evlist->cpus);
346 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
351 for (cpu = 0; cpu < nr_cpus; cpu++) {
352 for (thread = 0; thread < nr_threads; thread++) {
353 err = ioctl(FD(evsel, cpu, thread),
354 PERF_EVENT_IOC_ENABLE, 0);
362 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
363 struct perf_evsel *evsel, int cpu)
366 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
371 for (thread = 0; thread < nr_threads; thread++) {
372 err = ioctl(FD(evsel, cpu, thread),
373 PERF_EVENT_IOC_ENABLE, 0);
380 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
381 struct perf_evsel *evsel,
385 int nr_cpus = cpu_map__nr(evlist->cpus);
390 for (cpu = 0; cpu < nr_cpus; cpu++) {
391 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
398 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
399 struct perf_evsel *evsel, int idx)
401 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
404 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
406 return perf_evlist__enable_event_thread(evlist, evsel, idx);
409 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
411 int nr_cpus = cpu_map__nr(evlist->cpus);
412 int nr_threads = thread_map__nr(evlist->threads);
414 struct perf_evsel *evsel;
416 list_for_each_entry(evsel, &evlist->entries, node) {
417 if (evsel->system_wide)
420 nfds += nr_cpus * nr_threads;
423 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
424 fdarray__grow(&evlist->pollfd, nfds) < 0)
430 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
432 int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
434 * Save the idx so that when we filter out fds POLLHUP'ed we can
435 * close the associated evlist->mmap[] entry.
438 evlist->pollfd.priv[pos].idx = idx;
440 fcntl(fd, F_SETFL, O_NONBLOCK);
446 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
448 return __perf_evlist__add_pollfd(evlist, fd, -1);
451 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
453 struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
455 perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
458 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
460 return fdarray__filter(&evlist->pollfd, revents_and_mask,
461 perf_evlist__munmap_filtered);
464 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
466 return fdarray__poll(&evlist->pollfd, timeout);
469 static void perf_evlist__id_hash(struct perf_evlist *evlist,
470 struct perf_evsel *evsel,
471 int cpu, int thread, u64 id)
474 struct perf_sample_id *sid = SID(evsel, cpu, thread);
478 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
479 hlist_add_head(&sid->node, &evlist->heads[hash]);
482 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
483 int cpu, int thread, u64 id)
485 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
486 evsel->id[evsel->ids++] = id;
489 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
490 struct perf_evsel *evsel,
491 int cpu, int thread, int fd)
493 u64 read_data[4] = { 0, };
494 int id_idx = 1; /* The first entry is the counter value */
498 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
505 /* Legacy way to get event id.. All hail to old kernels! */
508 * This way does not work with group format read, so bail
511 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
514 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
515 read(fd, &read_data, sizeof(read_data)) == -1)
518 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
520 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
523 id = read_data[id_idx];
526 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
530 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
532 struct hlist_head *head;
533 struct perf_sample_id *sid;
536 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
537 head = &evlist->heads[hash];
539 hlist_for_each_entry(sid, head, node)
546 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
548 struct perf_sample_id *sid;
550 if (evlist->nr_entries == 1)
551 return perf_evlist__first(evlist);
553 sid = perf_evlist__id2sid(evlist, id);
557 if (!perf_evlist__sample_id_all(evlist))
558 return perf_evlist__first(evlist);
563 static int perf_evlist__event2id(struct perf_evlist *evlist,
564 union perf_event *event, u64 *id)
566 const u64 *array = event->sample.array;
569 n = (event->header.size - sizeof(event->header)) >> 3;
571 if (event->header.type == PERF_RECORD_SAMPLE) {
572 if (evlist->id_pos >= n)
574 *id = array[evlist->id_pos];
576 if (evlist->is_pos > n)
584 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
585 union perf_event *event)
587 struct perf_evsel *first = perf_evlist__first(evlist);
588 struct hlist_head *head;
589 struct perf_sample_id *sid;
593 if (evlist->nr_entries == 1)
596 if (!first->attr.sample_id_all &&
597 event->header.type != PERF_RECORD_SAMPLE)
600 if (perf_evlist__event2id(evlist, event, &id))
603 /* Synthesized events have an id of zero */
607 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
608 head = &evlist->heads[hash];
610 hlist_for_each_entry(sid, head, node) {
617 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
619 struct perf_mmap *md = &evlist->mmap[idx];
620 unsigned int head = perf_mmap__read_head(md);
621 unsigned int old = md->prev;
622 unsigned char *data = md->base + page_size;
623 union perf_event *event = NULL;
625 if (evlist->overwrite) {
627 * If we're further behind than half the buffer, there's a chance
628 * the writer will bite our tail and mess up the samples under us.
630 * If we somehow ended up ahead of the head, we got messed up.
632 * In either case, truncate and restart at head.
634 int diff = head - old;
635 if (diff > md->mask / 2 || diff < 0) {
636 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
639 * head points to a known good entry, start there.
648 event = (union perf_event *)&data[old & md->mask];
649 size = event->header.size;
652 * Event straddles the mmap boundary -- header should always
653 * be inside due to u64 alignment of output.
655 if ((old & md->mask) + size != ((old + size) & md->mask)) {
656 unsigned int offset = old;
657 unsigned int len = min(sizeof(*event), size), cpy;
658 void *dst = md->event_copy;
661 cpy = min(md->mask + 1 - (offset & md->mask), len);
662 memcpy(dst, &data[offset & md->mask], cpy);
668 event = (union perf_event *) md->event_copy;
679 static bool perf_mmap__empty(struct perf_mmap *md)
681 return perf_mmap__read_head(md) != md->prev;
684 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
686 ++evlist->mmap[idx].refcnt;
689 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
691 BUG_ON(evlist->mmap[idx].refcnt == 0);
693 if (--evlist->mmap[idx].refcnt == 0)
694 __perf_evlist__munmap(evlist, idx);
697 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
699 struct perf_mmap *md = &evlist->mmap[idx];
701 if (!evlist->overwrite) {
702 unsigned int old = md->prev;
704 perf_mmap__write_tail(md, old);
707 if (md->refcnt == 1 && perf_mmap__empty(md))
708 perf_evlist__mmap_put(evlist, idx);
711 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
713 if (evlist->mmap[idx].base != NULL) {
714 munmap(evlist->mmap[idx].base, evlist->mmap_len);
715 evlist->mmap[idx].base = NULL;
716 evlist->mmap[idx].refcnt = 0;
720 void perf_evlist__munmap(struct perf_evlist *evlist)
724 if (evlist->mmap == NULL)
727 for (i = 0; i < evlist->nr_mmaps; i++)
728 __perf_evlist__munmap(evlist, i);
730 zfree(&evlist->mmap);
733 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
735 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
736 if (cpu_map__empty(evlist->cpus))
737 evlist->nr_mmaps = thread_map__nr(evlist->threads);
738 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
739 return evlist->mmap != NULL ? 0 : -ENOMEM;
747 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
748 struct mmap_params *mp, int fd)
751 * The last one will be done at perf_evlist__mmap_consume(), so that we
752 * make sure we don't prevent tools from consuming every last event in
755 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
756 * anymore, but the last events for it are still in the ring buffer,
757 * waiting to be consumed.
759 * Tools can chose to ignore this at their own discretion, but the
760 * evlist layer can't just drop it when filtering events in
761 * perf_evlist__filter_pollfd().
763 evlist->mmap[idx].refcnt = 2;
764 evlist->mmap[idx].prev = 0;
765 evlist->mmap[idx].mask = mp->mask;
766 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
768 if (evlist->mmap[idx].base == MAP_FAILED) {
769 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
771 evlist->mmap[idx].base = NULL;
778 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
779 struct mmap_params *mp, int cpu,
780 int thread, int *output)
782 struct perf_evsel *evsel;
784 evlist__for_each(evlist, evsel) {
787 if (evsel->system_wide && thread)
790 fd = FD(evsel, cpu, thread);
794 if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
797 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
800 perf_evlist__mmap_get(evlist, idx);
803 if (__perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
804 perf_evlist__mmap_put(evlist, idx);
808 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
809 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
816 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
817 struct mmap_params *mp)
820 int nr_cpus = cpu_map__nr(evlist->cpus);
821 int nr_threads = thread_map__nr(evlist->threads);
823 pr_debug2("perf event ring buffer mmapped per cpu\n");
824 for (cpu = 0; cpu < nr_cpus; cpu++) {
827 for (thread = 0; thread < nr_threads; thread++) {
828 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
837 for (cpu = 0; cpu < nr_cpus; cpu++)
838 __perf_evlist__munmap(evlist, cpu);
842 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
843 struct mmap_params *mp)
846 int nr_threads = thread_map__nr(evlist->threads);
848 pr_debug2("perf event ring buffer mmapped per thread\n");
849 for (thread = 0; thread < nr_threads; thread++) {
852 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
860 for (thread = 0; thread < nr_threads; thread++)
861 __perf_evlist__munmap(evlist, thread);
865 static size_t perf_evlist__mmap_size(unsigned long pages)
867 /* 512 kiB: default amount of unprivileged mlocked memory */
868 if (pages == UINT_MAX)
869 pages = (512 * 1024) / page_size;
870 else if (!is_power_of_2(pages))
873 return (pages + 1) * page_size;
876 static long parse_pages_arg(const char *str, unsigned long min,
879 unsigned long pages, val;
880 static struct parse_tag tags[] = {
881 { .tag = 'B', .mult = 1 },
882 { .tag = 'K', .mult = 1 << 10 },
883 { .tag = 'M', .mult = 1 << 20 },
884 { .tag = 'G', .mult = 1 << 30 },
891 val = parse_tag_value(str, tags);
892 if (val != (unsigned long) -1) {
893 /* we got file size value */
894 pages = PERF_ALIGN(val, page_size) / page_size;
896 /* we got pages count value */
898 pages = strtoul(str, &eptr, 10);
903 if (pages == 0 && min == 0) {
904 /* leave number of pages at 0 */
905 } else if (!is_power_of_2(pages)) {
906 /* round pages up to next power of 2 */
907 pages = next_pow2_l(pages);
910 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
911 pages * page_size, pages);
920 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
921 int unset __maybe_unused)
923 unsigned int *mmap_pages = opt->value;
924 unsigned long max = UINT_MAX;
927 if (max > SIZE_MAX / page_size)
928 max = SIZE_MAX / page_size;
930 pages = parse_pages_arg(str, 1, max);
932 pr_err("Invalid argument for --mmap_pages/-m\n");
941 * perf_evlist__mmap - Create mmaps to receive events.
942 * @evlist: list of events
943 * @pages: map length in pages
944 * @overwrite: overwrite older events?
946 * If @overwrite is %false the user needs to signal event consumption using
947 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
950 * Return: %0 on success, negative error code otherwise.
952 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
955 struct perf_evsel *evsel;
956 const struct cpu_map *cpus = evlist->cpus;
957 const struct thread_map *threads = evlist->threads;
958 struct mmap_params mp = {
959 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
962 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
965 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
968 evlist->overwrite = overwrite;
969 evlist->mmap_len = perf_evlist__mmap_size(pages);
970 pr_debug("mmap size %zuB\n", evlist->mmap_len);
971 mp.mask = evlist->mmap_len - page_size - 1;
973 evlist__for_each(evlist, evsel) {
974 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
975 evsel->sample_id == NULL &&
976 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
980 if (cpu_map__empty(cpus))
981 return perf_evlist__mmap_per_thread(evlist, &mp);
983 return perf_evlist__mmap_per_cpu(evlist, &mp);
986 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
988 evlist->threads = thread_map__new_str(target->pid, target->tid,
991 if (evlist->threads == NULL)
994 if (target__uses_dummy_map(target))
995 evlist->cpus = cpu_map__dummy_new();
997 evlist->cpus = cpu_map__new(target->cpu_list);
999 if (evlist->cpus == NULL)
1000 goto out_delete_threads;
1005 thread_map__delete(evlist->threads);
1006 evlist->threads = NULL;
1010 int perf_evlist__apply_filters(struct perf_evlist *evlist)
1012 struct perf_evsel *evsel;
1014 const int ncpus = cpu_map__nr(evlist->cpus),
1015 nthreads = thread_map__nr(evlist->threads);
1017 evlist__for_each(evlist, evsel) {
1018 if (evsel->filter == NULL)
1021 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1029 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1031 struct perf_evsel *evsel;
1033 const int ncpus = cpu_map__nr(evlist->cpus),
1034 nthreads = thread_map__nr(evlist->threads);
1036 evlist__for_each(evlist, evsel) {
1037 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1045 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1047 struct perf_evsel *pos;
1049 if (evlist->nr_entries == 1)
1052 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1055 evlist__for_each(evlist, pos) {
1056 if (pos->id_pos != evlist->id_pos ||
1057 pos->is_pos != evlist->is_pos)
1064 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1066 struct perf_evsel *evsel;
1068 if (evlist->combined_sample_type)
1069 return evlist->combined_sample_type;
1071 evlist__for_each(evlist, evsel)
1072 evlist->combined_sample_type |= evsel->attr.sample_type;
1074 return evlist->combined_sample_type;
1077 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1079 evlist->combined_sample_type = 0;
1080 return __perf_evlist__combined_sample_type(evlist);
1083 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1085 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1086 u64 read_format = first->attr.read_format;
1087 u64 sample_type = first->attr.sample_type;
1089 evlist__for_each(evlist, pos) {
1090 if (read_format != pos->attr.read_format)
1094 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1095 if ((sample_type & PERF_SAMPLE_READ) &&
1096 !(read_format & PERF_FORMAT_ID)) {
1103 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1105 struct perf_evsel *first = perf_evlist__first(evlist);
1106 return first->attr.read_format;
1109 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1111 struct perf_evsel *first = perf_evlist__first(evlist);
1112 struct perf_sample *data;
1116 if (!first->attr.sample_id_all)
1119 sample_type = first->attr.sample_type;
1121 if (sample_type & PERF_SAMPLE_TID)
1122 size += sizeof(data->tid) * 2;
1124 if (sample_type & PERF_SAMPLE_TIME)
1125 size += sizeof(data->time);
1127 if (sample_type & PERF_SAMPLE_ID)
1128 size += sizeof(data->id);
1130 if (sample_type & PERF_SAMPLE_STREAM_ID)
1131 size += sizeof(data->stream_id);
1133 if (sample_type & PERF_SAMPLE_CPU)
1134 size += sizeof(data->cpu) * 2;
1136 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1137 size += sizeof(data->id);
1142 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1144 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1146 evlist__for_each_continue(evlist, pos) {
1147 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1154 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1156 struct perf_evsel *first = perf_evlist__first(evlist);
1157 return first->attr.sample_id_all;
1160 void perf_evlist__set_selected(struct perf_evlist *evlist,
1161 struct perf_evsel *evsel)
1163 evlist->selected = evsel;
1166 void perf_evlist__close(struct perf_evlist *evlist)
1168 struct perf_evsel *evsel;
1169 int ncpus = cpu_map__nr(evlist->cpus);
1170 int nthreads = thread_map__nr(evlist->threads);
1173 evlist__for_each_reverse(evlist, evsel) {
1174 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1175 perf_evsel__close(evsel, n, nthreads);
1179 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1184 * Try reading /sys/devices/system/cpu/online to get
1187 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1188 * code needs an overhaul to properly forward the
1189 * error, and we may not want to do that fallback to a
1190 * default cpu identity map :-\
1192 evlist->cpus = cpu_map__new(NULL);
1193 if (evlist->cpus == NULL)
1196 evlist->threads = thread_map__new_dummy();
1197 if (evlist->threads == NULL)
1204 cpu_map__delete(evlist->cpus);
1205 evlist->cpus = NULL;
1209 int perf_evlist__open(struct perf_evlist *evlist)
1211 struct perf_evsel *evsel;
1215 * Default: one fd per CPU, all threads, aka systemwide
1216 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1218 if (evlist->threads == NULL && evlist->cpus == NULL) {
1219 err = perf_evlist__create_syswide_maps(evlist);
1224 perf_evlist__update_id_pos(evlist);
1226 evlist__for_each(evlist, evsel) {
1227 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1234 perf_evlist__close(evlist);
1239 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1240 const char *argv[], bool pipe_output,
1241 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1243 int child_ready_pipe[2], go_pipe[2];
1246 if (pipe(child_ready_pipe) < 0) {
1247 perror("failed to create 'ready' pipe");
1251 if (pipe(go_pipe) < 0) {
1252 perror("failed to create 'go' pipe");
1253 goto out_close_ready_pipe;
1256 evlist->workload.pid = fork();
1257 if (evlist->workload.pid < 0) {
1258 perror("failed to fork");
1259 goto out_close_pipes;
1262 if (!evlist->workload.pid) {
1268 signal(SIGTERM, SIG_DFL);
1270 close(child_ready_pipe[0]);
1272 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1275 * Tell the parent we're ready to go
1277 close(child_ready_pipe[1]);
1280 * Wait until the parent tells us to go.
1282 ret = read(go_pipe[0], &bf, 1);
1284 * The parent will ask for the execvp() to be performed by
1285 * writing exactly one byte, in workload.cork_fd, usually via
1286 * perf_evlist__start_workload().
1288 * For cancelling the workload without actuallin running it,
1289 * the parent will just close workload.cork_fd, without writing
1290 * anything, i.e. read will return zero and we just exit()
1295 perror("unable to read pipe");
1299 execvp(argv[0], (char **)argv);
1304 val.sival_int = errno;
1305 if (sigqueue(getppid(), SIGUSR1, val))
1313 struct sigaction act = {
1314 .sa_flags = SA_SIGINFO,
1315 .sa_sigaction = exec_error,
1317 sigaction(SIGUSR1, &act, NULL);
1320 if (target__none(target)) {
1321 if (evlist->threads == NULL) {
1322 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1323 __func__, __LINE__);
1324 goto out_close_pipes;
1326 evlist->threads->map[0] = evlist->workload.pid;
1329 close(child_ready_pipe[1]);
1332 * wait for child to settle
1334 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1335 perror("unable to read pipe");
1336 goto out_close_pipes;
1339 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1340 evlist->workload.cork_fd = go_pipe[1];
1341 close(child_ready_pipe[0]);
1347 out_close_ready_pipe:
1348 close(child_ready_pipe[0]);
1349 close(child_ready_pipe[1]);
1353 int perf_evlist__start_workload(struct perf_evlist *evlist)
1355 if (evlist->workload.cork_fd > 0) {
1359 * Remove the cork, let it rip!
1361 ret = write(evlist->workload.cork_fd, &bf, 1);
1363 perror("enable to write to pipe");
1365 close(evlist->workload.cork_fd);
1372 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1373 struct perf_sample *sample)
1375 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1379 return perf_evsel__parse_sample(evsel, event, sample);
1382 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1384 struct perf_evsel *evsel;
1387 evlist__for_each(evlist, evsel) {
1388 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1389 perf_evsel__name(evsel));
1392 return printed + fprintf(fp, "\n");
1395 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1396 int err, char *buf, size_t size)
1402 scnprintf(buf, size, "%s",
1403 "Error:\tUnable to find debugfs\n"
1404 "Hint:\tWas your kernel was compiled with debugfs support?\n"
1405 "Hint:\tIs the debugfs filesystem mounted?\n"
1406 "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1409 scnprintf(buf, size,
1410 "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1411 "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1412 debugfs_mountpoint, debugfs_mountpoint);
1415 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1422 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1423 int err, char *buf, size_t size)
1426 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1431 printed = scnprintf(buf, size,
1433 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1435 value = perf_event_paranoid();
1437 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1440 printed += scnprintf(buf + printed, size - printed,
1441 "For your workloads it needs to be <= 1\nHint:\t");
1443 printed += scnprintf(buf + printed, size - printed,
1444 "For system wide tracing it needs to be set to -1.\n");
1446 printed += scnprintf(buf + printed, size - printed,
1447 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1448 "Hint:\tThe current value is %d.", value);
1451 scnprintf(buf, size, "%s", emsg);
1458 void perf_evlist__to_front(struct perf_evlist *evlist,
1459 struct perf_evsel *move_evsel)
1461 struct perf_evsel *evsel, *n;
1464 if (move_evsel == perf_evlist__first(evlist))
1467 evlist__for_each_safe(evlist, n, evsel) {
1468 if (evsel->leader == move_evsel->leader)
1469 list_move_tail(&evsel->node, &move);
1472 list_splice(&move, &evlist->entries);
1475 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1476 struct perf_evsel *tracking_evsel)
1478 struct perf_evsel *evsel;
1480 if (tracking_evsel->tracking)
1483 evlist__for_each(evlist, evsel) {
1484 if (evsel != tracking_evsel)
1485 evsel->tracking = false;
1488 tracking_evsel->tracking = true;