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/fs.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>
27 #include <linux/log2.h>
29 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
30 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
33 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
35 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
36 struct thread_map *threads)
40 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
41 INIT_HLIST_HEAD(&evlist->heads[i]);
42 INIT_LIST_HEAD(&evlist->entries);
43 perf_evlist__set_maps(evlist, cpus, threads);
44 fdarray__init(&evlist->pollfd, 64);
45 evlist->workload.pid = -1;
48 struct perf_evlist *perf_evlist__new(void)
50 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
53 perf_evlist__init(evlist, NULL, NULL);
58 struct perf_evlist *perf_evlist__new_default(void)
60 struct perf_evlist *evlist = perf_evlist__new();
62 if (evlist && perf_evlist__add_default(evlist)) {
63 perf_evlist__delete(evlist);
71 * perf_evlist__set_id_pos - set the positions of event ids.
72 * @evlist: selected event list
74 * Events with compatible sample types all have the same id_pos
75 * and is_pos. For convenience, put a copy on evlist.
77 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
79 struct perf_evsel *first = perf_evlist__first(evlist);
81 evlist->id_pos = first->id_pos;
82 evlist->is_pos = first->is_pos;
85 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
87 struct perf_evsel *evsel;
89 evlist__for_each(evlist, evsel)
90 perf_evsel__calc_id_pos(evsel);
92 perf_evlist__set_id_pos(evlist);
95 static void perf_evlist__purge(struct perf_evlist *evlist)
97 struct perf_evsel *pos, *n;
99 evlist__for_each_safe(evlist, n, pos) {
100 list_del_init(&pos->node);
102 perf_evsel__delete(pos);
105 evlist->nr_entries = 0;
108 void perf_evlist__exit(struct perf_evlist *evlist)
110 zfree(&evlist->mmap);
111 fdarray__exit(&evlist->pollfd);
114 void perf_evlist__delete(struct perf_evlist *evlist)
116 perf_evlist__munmap(evlist);
117 perf_evlist__close(evlist);
118 cpu_map__put(evlist->cpus);
119 thread_map__put(evlist->threads);
121 evlist->threads = NULL;
122 perf_evlist__purge(evlist);
123 perf_evlist__exit(evlist);
127 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
129 entry->evlist = evlist;
130 list_add_tail(&entry->node, &evlist->entries);
131 entry->idx = evlist->nr_entries;
132 entry->tracking = !entry->idx;
134 if (!evlist->nr_entries++)
135 perf_evlist__set_id_pos(evlist);
138 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
139 struct list_head *list,
142 bool set_id_pos = !evlist->nr_entries;
144 list_splice_tail(list, &evlist->entries);
145 evlist->nr_entries += nr_entries;
147 perf_evlist__set_id_pos(evlist);
150 void __perf_evlist__set_leader(struct list_head *list)
152 struct perf_evsel *evsel, *leader;
154 leader = list_entry(list->next, struct perf_evsel, node);
155 evsel = list_entry(list->prev, struct perf_evsel, node);
157 leader->nr_members = evsel->idx - leader->idx + 1;
159 __evlist__for_each(list, evsel) {
160 evsel->leader = leader;
164 void perf_evlist__set_leader(struct perf_evlist *evlist)
166 if (evlist->nr_entries) {
167 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
168 __perf_evlist__set_leader(&evlist->entries);
172 int perf_evlist__add_default(struct perf_evlist *evlist)
174 struct perf_event_attr attr = {
175 .type = PERF_TYPE_HARDWARE,
176 .config = PERF_COUNT_HW_CPU_CYCLES,
178 struct perf_evsel *evsel;
180 event_attr_init(&attr);
182 evsel = perf_evsel__new(&attr);
186 /* use strdup() because free(evsel) assumes name is allocated */
187 evsel->name = strdup("cycles");
191 perf_evlist__add(evlist, evsel);
194 perf_evsel__delete(evsel);
199 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
200 struct perf_event_attr *attrs, size_t nr_attrs)
202 struct perf_evsel *evsel, *n;
206 for (i = 0; i < nr_attrs; i++) {
207 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
209 goto out_delete_partial_list;
210 list_add_tail(&evsel->node, &head);
213 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
217 out_delete_partial_list:
218 __evlist__for_each_safe(&head, n, evsel)
219 perf_evsel__delete(evsel);
223 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
224 struct perf_event_attr *attrs, size_t nr_attrs)
228 for (i = 0; i < nr_attrs; i++)
229 event_attr_init(attrs + i);
231 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
235 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
237 struct perf_evsel *evsel;
239 evlist__for_each(evlist, evsel) {
240 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
241 (int)evsel->attr.config == id)
249 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
252 struct perf_evsel *evsel;
254 evlist__for_each(evlist, evsel) {
255 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
256 (strcmp(evsel->name, name) == 0))
263 int perf_evlist__add_newtp(struct perf_evlist *evlist,
264 const char *sys, const char *name, void *handler)
266 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
271 evsel->handler = handler;
272 perf_evlist__add(evlist, evsel);
276 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
277 struct perf_evsel *evsel)
279 if (evsel->system_wide)
282 return thread_map__nr(evlist->threads);
285 void perf_evlist__disable(struct perf_evlist *evlist)
288 struct perf_evsel *pos;
289 int nr_cpus = cpu_map__nr(evlist->cpus);
292 for (cpu = 0; cpu < nr_cpus; cpu++) {
293 evlist__for_each(evlist, pos) {
294 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
296 nr_threads = perf_evlist__nr_threads(evlist, pos);
297 for (thread = 0; thread < nr_threads; thread++)
298 ioctl(FD(pos, cpu, thread),
299 PERF_EVENT_IOC_DISABLE, 0);
303 evlist->enabled = false;
306 void perf_evlist__enable(struct perf_evlist *evlist)
309 struct perf_evsel *pos;
310 int nr_cpus = cpu_map__nr(evlist->cpus);
313 for (cpu = 0; cpu < nr_cpus; cpu++) {
314 evlist__for_each(evlist, pos) {
315 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
317 nr_threads = perf_evlist__nr_threads(evlist, pos);
318 for (thread = 0; thread < nr_threads; thread++)
319 ioctl(FD(pos, cpu, thread),
320 PERF_EVENT_IOC_ENABLE, 0);
324 evlist->enabled = true;
327 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
329 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
332 int perf_evlist__disable_event(struct perf_evlist *evlist,
333 struct perf_evsel *evsel)
335 int cpu, thread, err;
336 int nr_cpus = cpu_map__nr(evlist->cpus);
337 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
342 for (cpu = 0; cpu < nr_cpus; cpu++) {
343 for (thread = 0; thread < nr_threads; thread++) {
344 err = ioctl(FD(evsel, cpu, thread),
345 PERF_EVENT_IOC_DISABLE, 0);
353 int perf_evlist__enable_event(struct perf_evlist *evlist,
354 struct perf_evsel *evsel)
356 int cpu, thread, err;
357 int nr_cpus = cpu_map__nr(evlist->cpus);
358 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
363 for (cpu = 0; cpu < nr_cpus; cpu++) {
364 for (thread = 0; thread < nr_threads; thread++) {
365 err = ioctl(FD(evsel, cpu, thread),
366 PERF_EVENT_IOC_ENABLE, 0);
374 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
375 struct perf_evsel *evsel, int cpu)
378 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
383 for (thread = 0; thread < nr_threads; thread++) {
384 err = ioctl(FD(evsel, cpu, thread),
385 PERF_EVENT_IOC_ENABLE, 0);
392 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
393 struct perf_evsel *evsel,
397 int nr_cpus = cpu_map__nr(evlist->cpus);
402 for (cpu = 0; cpu < nr_cpus; cpu++) {
403 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
410 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
411 struct perf_evsel *evsel, int idx)
413 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
416 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
418 return perf_evlist__enable_event_thread(evlist, evsel, idx);
421 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
423 int nr_cpus = cpu_map__nr(evlist->cpus);
424 int nr_threads = thread_map__nr(evlist->threads);
426 struct perf_evsel *evsel;
428 evlist__for_each(evlist, evsel) {
429 if (evsel->system_wide)
432 nfds += nr_cpus * nr_threads;
435 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
436 fdarray__grow(&evlist->pollfd, nfds) < 0)
442 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
444 int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
446 * Save the idx so that when we filter out fds POLLHUP'ed we can
447 * close the associated evlist->mmap[] entry.
450 evlist->pollfd.priv[pos].idx = idx;
452 fcntl(fd, F_SETFL, O_NONBLOCK);
458 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
460 return __perf_evlist__add_pollfd(evlist, fd, -1);
463 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
465 struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
467 perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
470 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
472 return fdarray__filter(&evlist->pollfd, revents_and_mask,
473 perf_evlist__munmap_filtered);
476 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
478 return fdarray__poll(&evlist->pollfd, timeout);
481 static void perf_evlist__id_hash(struct perf_evlist *evlist,
482 struct perf_evsel *evsel,
483 int cpu, int thread, u64 id)
486 struct perf_sample_id *sid = SID(evsel, cpu, thread);
490 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
491 hlist_add_head(&sid->node, &evlist->heads[hash]);
494 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
495 int cpu, int thread, u64 id)
497 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
498 evsel->id[evsel->ids++] = id;
501 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
502 struct perf_evsel *evsel,
503 int cpu, int thread, int fd)
505 u64 read_data[4] = { 0, };
506 int id_idx = 1; /* The first entry is the counter value */
510 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
517 /* Legacy way to get event id.. All hail to old kernels! */
520 * This way does not work with group format read, so bail
523 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
526 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
527 read(fd, &read_data, sizeof(read_data)) == -1)
530 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
532 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
535 id = read_data[id_idx];
538 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
542 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
543 struct perf_evsel *evsel, int idx, int cpu,
546 struct perf_sample_id *sid = SID(evsel, cpu, thread);
548 if (evlist->cpus && cpu >= 0)
549 sid->cpu = evlist->cpus->map[cpu];
552 if (!evsel->system_wide && evlist->threads && thread >= 0)
553 sid->tid = thread_map__pid(evlist->threads, thread);
558 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
560 struct hlist_head *head;
561 struct perf_sample_id *sid;
564 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
565 head = &evlist->heads[hash];
567 hlist_for_each_entry(sid, head, node)
574 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
576 struct perf_sample_id *sid;
578 if (evlist->nr_entries == 1 || !id)
579 return perf_evlist__first(evlist);
581 sid = perf_evlist__id2sid(evlist, id);
585 if (!perf_evlist__sample_id_all(evlist))
586 return perf_evlist__first(evlist);
591 static int perf_evlist__event2id(struct perf_evlist *evlist,
592 union perf_event *event, u64 *id)
594 const u64 *array = event->sample.array;
597 n = (event->header.size - sizeof(event->header)) >> 3;
599 if (event->header.type == PERF_RECORD_SAMPLE) {
600 if (evlist->id_pos >= n)
602 *id = array[evlist->id_pos];
604 if (evlist->is_pos > n)
612 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
613 union perf_event *event)
615 struct perf_evsel *first = perf_evlist__first(evlist);
616 struct hlist_head *head;
617 struct perf_sample_id *sid;
621 if (evlist->nr_entries == 1)
624 if (!first->attr.sample_id_all &&
625 event->header.type != PERF_RECORD_SAMPLE)
628 if (perf_evlist__event2id(evlist, event, &id))
631 /* Synthesized events have an id of zero */
635 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
636 head = &evlist->heads[hash];
638 hlist_for_each_entry(sid, head, node) {
645 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
647 struct perf_mmap *md = &evlist->mmap[idx];
650 unsigned char *data = md->base + page_size;
651 union perf_event *event = NULL;
654 * Check if event was unmapped due to a POLLHUP/POLLERR.
656 if (!atomic_read(&md->refcnt))
659 head = perf_mmap__read_head(md);
660 if (evlist->overwrite) {
662 * If we're further behind than half the buffer, there's a chance
663 * the writer will bite our tail and mess up the samples under us.
665 * If we somehow ended up ahead of the head, we got messed up.
667 * In either case, truncate and restart at head.
669 int diff = head - old;
670 if (diff > md->mask / 2 || diff < 0) {
671 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
674 * head points to a known good entry, start there.
683 event = (union perf_event *)&data[old & md->mask];
684 size = event->header.size;
687 * Event straddles the mmap boundary -- header should always
688 * be inside due to u64 alignment of output.
690 if ((old & md->mask) + size != ((old + size) & md->mask)) {
691 unsigned int offset = old;
692 unsigned int len = min(sizeof(*event), size), cpy;
693 void *dst = md->event_copy;
696 cpy = min(md->mask + 1 - (offset & md->mask), len);
697 memcpy(dst, &data[offset & md->mask], cpy);
703 event = (union perf_event *) md->event_copy;
714 static bool perf_mmap__empty(struct perf_mmap *md)
716 return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
719 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
721 atomic_inc(&evlist->mmap[idx].refcnt);
724 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
726 BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
728 if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
729 __perf_evlist__munmap(evlist, idx);
732 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
734 struct perf_mmap *md = &evlist->mmap[idx];
736 if (!evlist->overwrite) {
739 perf_mmap__write_tail(md, old);
742 if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
743 perf_evlist__mmap_put(evlist, idx);
746 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
747 struct auxtrace_mmap_params *mp __maybe_unused,
748 void *userpg __maybe_unused,
749 int fd __maybe_unused)
754 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
758 void __weak auxtrace_mmap_params__init(
759 struct auxtrace_mmap_params *mp __maybe_unused,
760 off_t auxtrace_offset __maybe_unused,
761 unsigned int auxtrace_pages __maybe_unused,
762 bool auxtrace_overwrite __maybe_unused)
766 void __weak auxtrace_mmap_params__set_idx(
767 struct auxtrace_mmap_params *mp __maybe_unused,
768 struct perf_evlist *evlist __maybe_unused,
769 int idx __maybe_unused,
770 bool per_cpu __maybe_unused)
774 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
776 if (evlist->mmap[idx].base != NULL) {
777 munmap(evlist->mmap[idx].base, evlist->mmap_len);
778 evlist->mmap[idx].base = NULL;
779 atomic_set(&evlist->mmap[idx].refcnt, 0);
781 auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
784 void perf_evlist__munmap(struct perf_evlist *evlist)
788 if (evlist->mmap == NULL)
791 for (i = 0; i < evlist->nr_mmaps; i++)
792 __perf_evlist__munmap(evlist, i);
794 zfree(&evlist->mmap);
797 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
799 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
800 if (cpu_map__empty(evlist->cpus))
801 evlist->nr_mmaps = thread_map__nr(evlist->threads);
802 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
803 return evlist->mmap != NULL ? 0 : -ENOMEM;
809 struct auxtrace_mmap_params auxtrace_mp;
812 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
813 struct mmap_params *mp, int fd)
816 * The last one will be done at perf_evlist__mmap_consume(), so that we
817 * make sure we don't prevent tools from consuming every last event in
820 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
821 * anymore, but the last events for it are still in the ring buffer,
822 * waiting to be consumed.
824 * Tools can chose to ignore this at their own discretion, but the
825 * evlist layer can't just drop it when filtering events in
826 * perf_evlist__filter_pollfd().
828 atomic_set(&evlist->mmap[idx].refcnt, 2);
829 evlist->mmap[idx].prev = 0;
830 evlist->mmap[idx].mask = mp->mask;
831 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
833 if (evlist->mmap[idx].base == MAP_FAILED) {
834 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
836 evlist->mmap[idx].base = NULL;
840 if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
841 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
847 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
848 struct mmap_params *mp, int cpu,
849 int thread, int *output)
851 struct perf_evsel *evsel;
853 evlist__for_each(evlist, evsel) {
856 if (evsel->system_wide && thread)
859 fd = FD(evsel, cpu, thread);
863 if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
866 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
869 perf_evlist__mmap_get(evlist, idx);
873 * The system_wide flag causes a selected event to be opened
874 * always without a pid. Consequently it will never get a
875 * POLLHUP, but it is used for tracking in combination with
876 * other events, so it should not need to be polled anyway.
877 * Therefore don't add it for polling.
879 if (!evsel->system_wide &&
880 __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
881 perf_evlist__mmap_put(evlist, idx);
885 if (evsel->attr.read_format & PERF_FORMAT_ID) {
886 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
889 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
897 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
898 struct mmap_params *mp)
901 int nr_cpus = cpu_map__nr(evlist->cpus);
902 int nr_threads = thread_map__nr(evlist->threads);
904 pr_debug2("perf event ring buffer mmapped per cpu\n");
905 for (cpu = 0; cpu < nr_cpus; cpu++) {
908 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
911 for (thread = 0; thread < nr_threads; thread++) {
912 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
921 for (cpu = 0; cpu < nr_cpus; cpu++)
922 __perf_evlist__munmap(evlist, cpu);
926 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
927 struct mmap_params *mp)
930 int nr_threads = thread_map__nr(evlist->threads);
932 pr_debug2("perf event ring buffer mmapped per thread\n");
933 for (thread = 0; thread < nr_threads; thread++) {
936 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
939 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
947 for (thread = 0; thread < nr_threads; thread++)
948 __perf_evlist__munmap(evlist, thread);
952 static size_t perf_evlist__mmap_size(unsigned long pages)
954 if (pages == UINT_MAX) {
957 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
959 * Pick a once upon a time good value, i.e. things look
960 * strange since we can't read a sysctl value, but lets not
965 max -= (page_size / 1024);
968 pages = (max * 1024) / page_size;
969 if (!is_power_of_2(pages))
970 pages = rounddown_pow_of_two(pages);
971 } else if (!is_power_of_2(pages))
974 return (pages + 1) * page_size;
977 static long parse_pages_arg(const char *str, unsigned long min,
980 unsigned long pages, val;
981 static struct parse_tag tags[] = {
982 { .tag = 'B', .mult = 1 },
983 { .tag = 'K', .mult = 1 << 10 },
984 { .tag = 'M', .mult = 1 << 20 },
985 { .tag = 'G', .mult = 1 << 30 },
992 val = parse_tag_value(str, tags);
993 if (val != (unsigned long) -1) {
994 /* we got file size value */
995 pages = PERF_ALIGN(val, page_size) / page_size;
997 /* we got pages count value */
999 pages = strtoul(str, &eptr, 10);
1004 if (pages == 0 && min == 0) {
1005 /* leave number of pages at 0 */
1006 } else if (!is_power_of_2(pages)) {
1007 /* round pages up to next power of 2 */
1008 pages = roundup_pow_of_two(pages);
1011 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1012 pages * page_size, pages);
1021 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1023 unsigned long max = UINT_MAX;
1026 if (max > SIZE_MAX / page_size)
1027 max = SIZE_MAX / page_size;
1029 pages = parse_pages_arg(str, 1, max);
1031 pr_err("Invalid argument for --mmap_pages/-m\n");
1035 *mmap_pages = pages;
1039 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1040 int unset __maybe_unused)
1042 return __perf_evlist__parse_mmap_pages(opt->value, str);
1046 * perf_evlist__mmap_ex - Create mmaps to receive events.
1047 * @evlist: list of events
1048 * @pages: map length in pages
1049 * @overwrite: overwrite older events?
1050 * @auxtrace_pages - auxtrace map length in pages
1051 * @auxtrace_overwrite - overwrite older auxtrace data?
1053 * If @overwrite is %false the user needs to signal event consumption using
1054 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1057 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1058 * consumption using auxtrace_mmap__write_tail().
1060 * Return: %0 on success, negative error code otherwise.
1062 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1063 bool overwrite, unsigned int auxtrace_pages,
1064 bool auxtrace_overwrite)
1066 struct perf_evsel *evsel;
1067 const struct cpu_map *cpus = evlist->cpus;
1068 const struct thread_map *threads = evlist->threads;
1069 struct mmap_params mp = {
1070 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1073 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1076 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1079 evlist->overwrite = overwrite;
1080 evlist->mmap_len = perf_evlist__mmap_size(pages);
1081 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1082 mp.mask = evlist->mmap_len - page_size - 1;
1084 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1085 auxtrace_pages, auxtrace_overwrite);
1087 evlist__for_each(evlist, evsel) {
1088 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1089 evsel->sample_id == NULL &&
1090 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1094 if (cpu_map__empty(cpus))
1095 return perf_evlist__mmap_per_thread(evlist, &mp);
1097 return perf_evlist__mmap_per_cpu(evlist, &mp);
1100 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1103 return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1106 static int perf_evlist__propagate_maps(struct perf_evlist *evlist,
1109 struct perf_evsel *evsel;
1111 evlist__for_each(evlist, evsel) {
1113 * We already have cpus for evsel (via PMU sysfs) so
1114 * keep it, if there's no target cpu list defined.
1116 if (evsel->cpus && has_user_cpus)
1117 cpu_map__put(evsel->cpus);
1119 if (!evsel->cpus || has_user_cpus)
1120 evsel->cpus = cpu_map__get(evlist->cpus);
1122 evsel->threads = thread_map__get(evlist->threads);
1124 if ((evlist->cpus && !evsel->cpus) ||
1125 (evlist->threads && !evsel->threads))
1132 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1134 evlist->threads = thread_map__new_str(target->pid, target->tid,
1137 if (evlist->threads == NULL)
1140 if (target__uses_dummy_map(target))
1141 evlist->cpus = cpu_map__dummy_new();
1143 evlist->cpus = cpu_map__new(target->cpu_list);
1145 if (evlist->cpus == NULL)
1146 goto out_delete_threads;
1148 return perf_evlist__propagate_maps(evlist, !!target->cpu_list);
1151 thread_map__put(evlist->threads);
1152 evlist->threads = NULL;
1156 int perf_evlist__set_maps(struct perf_evlist *evlist,
1157 struct cpu_map *cpus,
1158 struct thread_map *threads)
1161 cpu_map__put(evlist->cpus);
1163 evlist->cpus = cpus;
1165 if (evlist->threads)
1166 thread_map__put(evlist->threads);
1168 evlist->threads = threads;
1170 return perf_evlist__propagate_maps(evlist, false);
1173 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1175 struct perf_evsel *evsel;
1177 const int ncpus = cpu_map__nr(evlist->cpus),
1178 nthreads = thread_map__nr(evlist->threads);
1180 evlist__for_each(evlist, evsel) {
1181 if (evsel->filter == NULL)
1185 * filters only work for tracepoint event, which doesn't have cpu limit.
1186 * So evlist and evsel should always be same.
1188 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1198 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1200 struct perf_evsel *evsel;
1203 evlist__for_each(evlist, evsel) {
1204 err = perf_evsel__set_filter(evsel, filter);
1212 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1218 for (i = 0; i < npids; ++i) {
1220 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1225 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1233 ret = perf_evlist__set_filter(evlist, filter);
1239 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1241 return perf_evlist__set_filter_pids(evlist, 1, &pid);
1244 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1246 struct perf_evsel *pos;
1248 if (evlist->nr_entries == 1)
1251 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1254 evlist__for_each(evlist, pos) {
1255 if (pos->id_pos != evlist->id_pos ||
1256 pos->is_pos != evlist->is_pos)
1263 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1265 struct perf_evsel *evsel;
1267 if (evlist->combined_sample_type)
1268 return evlist->combined_sample_type;
1270 evlist__for_each(evlist, evsel)
1271 evlist->combined_sample_type |= evsel->attr.sample_type;
1273 return evlist->combined_sample_type;
1276 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1278 evlist->combined_sample_type = 0;
1279 return __perf_evlist__combined_sample_type(evlist);
1282 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1284 struct perf_evsel *evsel;
1285 u64 branch_type = 0;
1287 evlist__for_each(evlist, evsel)
1288 branch_type |= evsel->attr.branch_sample_type;
1292 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1294 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1295 u64 read_format = first->attr.read_format;
1296 u64 sample_type = first->attr.sample_type;
1298 evlist__for_each(evlist, pos) {
1299 if (read_format != pos->attr.read_format)
1303 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1304 if ((sample_type & PERF_SAMPLE_READ) &&
1305 !(read_format & PERF_FORMAT_ID)) {
1312 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1314 struct perf_evsel *first = perf_evlist__first(evlist);
1315 return first->attr.read_format;
1318 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1320 struct perf_evsel *first = perf_evlist__first(evlist);
1321 struct perf_sample *data;
1325 if (!first->attr.sample_id_all)
1328 sample_type = first->attr.sample_type;
1330 if (sample_type & PERF_SAMPLE_TID)
1331 size += sizeof(data->tid) * 2;
1333 if (sample_type & PERF_SAMPLE_TIME)
1334 size += sizeof(data->time);
1336 if (sample_type & PERF_SAMPLE_ID)
1337 size += sizeof(data->id);
1339 if (sample_type & PERF_SAMPLE_STREAM_ID)
1340 size += sizeof(data->stream_id);
1342 if (sample_type & PERF_SAMPLE_CPU)
1343 size += sizeof(data->cpu) * 2;
1345 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1346 size += sizeof(data->id);
1351 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1353 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1355 evlist__for_each_continue(evlist, pos) {
1356 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1363 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1365 struct perf_evsel *first = perf_evlist__first(evlist);
1366 return first->attr.sample_id_all;
1369 void perf_evlist__set_selected(struct perf_evlist *evlist,
1370 struct perf_evsel *evsel)
1372 evlist->selected = evsel;
1375 void perf_evlist__close(struct perf_evlist *evlist)
1377 struct perf_evsel *evsel;
1378 int ncpus = cpu_map__nr(evlist->cpus);
1379 int nthreads = thread_map__nr(evlist->threads);
1382 evlist__for_each_reverse(evlist, evsel) {
1383 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1384 perf_evsel__close(evsel, n, nthreads);
1388 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1393 * Try reading /sys/devices/system/cpu/online to get
1396 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1397 * code needs an overhaul to properly forward the
1398 * error, and we may not want to do that fallback to a
1399 * default cpu identity map :-\
1401 evlist->cpus = cpu_map__new(NULL);
1402 if (evlist->cpus == NULL)
1405 evlist->threads = thread_map__new_dummy();
1406 if (evlist->threads == NULL)
1413 cpu_map__put(evlist->cpus);
1414 evlist->cpus = NULL;
1418 int perf_evlist__open(struct perf_evlist *evlist)
1420 struct perf_evsel *evsel;
1424 * Default: one fd per CPU, all threads, aka systemwide
1425 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1427 if (evlist->threads == NULL && evlist->cpus == NULL) {
1428 err = perf_evlist__create_syswide_maps(evlist);
1433 perf_evlist__update_id_pos(evlist);
1435 evlist__for_each(evlist, evsel) {
1436 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1443 perf_evlist__close(evlist);
1448 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1449 const char *argv[], bool pipe_output,
1450 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1452 int child_ready_pipe[2], go_pipe[2];
1455 if (pipe(child_ready_pipe) < 0) {
1456 perror("failed to create 'ready' pipe");
1460 if (pipe(go_pipe) < 0) {
1461 perror("failed to create 'go' pipe");
1462 goto out_close_ready_pipe;
1465 evlist->workload.pid = fork();
1466 if (evlist->workload.pid < 0) {
1467 perror("failed to fork");
1468 goto out_close_pipes;
1471 if (!evlist->workload.pid) {
1477 signal(SIGTERM, SIG_DFL);
1479 close(child_ready_pipe[0]);
1481 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1484 * Tell the parent we're ready to go
1486 close(child_ready_pipe[1]);
1489 * Wait until the parent tells us to go.
1491 ret = read(go_pipe[0], &bf, 1);
1493 * The parent will ask for the execvp() to be performed by
1494 * writing exactly one byte, in workload.cork_fd, usually via
1495 * perf_evlist__start_workload().
1497 * For cancelling the workload without actually running it,
1498 * the parent will just close workload.cork_fd, without writing
1499 * anything, i.e. read will return zero and we just exit()
1504 perror("unable to read pipe");
1508 execvp(argv[0], (char **)argv);
1513 val.sival_int = errno;
1514 if (sigqueue(getppid(), SIGUSR1, val))
1522 struct sigaction act = {
1523 .sa_flags = SA_SIGINFO,
1524 .sa_sigaction = exec_error,
1526 sigaction(SIGUSR1, &act, NULL);
1529 if (target__none(target)) {
1530 if (evlist->threads == NULL) {
1531 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1532 __func__, __LINE__);
1533 goto out_close_pipes;
1535 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1538 close(child_ready_pipe[1]);
1541 * wait for child to settle
1543 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1544 perror("unable to read pipe");
1545 goto out_close_pipes;
1548 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1549 evlist->workload.cork_fd = go_pipe[1];
1550 close(child_ready_pipe[0]);
1556 out_close_ready_pipe:
1557 close(child_ready_pipe[0]);
1558 close(child_ready_pipe[1]);
1562 int perf_evlist__start_workload(struct perf_evlist *evlist)
1564 if (evlist->workload.cork_fd > 0) {
1568 * Remove the cork, let it rip!
1570 ret = write(evlist->workload.cork_fd, &bf, 1);
1572 perror("enable to write to pipe");
1574 close(evlist->workload.cork_fd);
1581 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1582 struct perf_sample *sample)
1584 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1588 return perf_evsel__parse_sample(evsel, event, sample);
1591 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1593 struct perf_evsel *evsel;
1596 evlist__for_each(evlist, evsel) {
1597 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1598 perf_evsel__name(evsel));
1601 return printed + fprintf(fp, "\n");
1604 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1605 int err, char *buf, size_t size)
1608 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1613 printed = scnprintf(buf, size,
1615 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1617 value = perf_event_paranoid();
1619 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1622 printed += scnprintf(buf + printed, size - printed,
1623 "For your workloads it needs to be <= 1\nHint:\t");
1625 printed += scnprintf(buf + printed, size - printed,
1626 "For system wide tracing it needs to be set to -1.\n");
1628 printed += scnprintf(buf + printed, size - printed,
1629 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1630 "Hint:\tThe current value is %d.", value);
1633 scnprintf(buf, size, "%s", emsg);
1640 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1642 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1643 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1647 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1648 printed += scnprintf(buf + printed, size - printed,
1650 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1651 "Hint:\tTried using %zd kB.\n",
1652 emsg, pages_max_per_user, pages_attempted);
1654 if (pages_attempted >= pages_max_per_user) {
1655 printed += scnprintf(buf + printed, size - printed,
1656 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1657 pages_max_per_user + pages_attempted);
1660 printed += scnprintf(buf + printed, size - printed,
1661 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1664 scnprintf(buf, size, "%s", emsg);
1671 void perf_evlist__to_front(struct perf_evlist *evlist,
1672 struct perf_evsel *move_evsel)
1674 struct perf_evsel *evsel, *n;
1677 if (move_evsel == perf_evlist__first(evlist))
1680 evlist__for_each_safe(evlist, n, evsel) {
1681 if (evsel->leader == move_evsel->leader)
1682 list_move_tail(&evsel->node, &move);
1685 list_splice(&move, &evlist->entries);
1688 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1689 struct perf_evsel *tracking_evsel)
1691 struct perf_evsel *evsel;
1693 if (tracking_evsel->tracking)
1696 evlist__for_each(evlist, evsel) {
1697 if (evsel != tracking_evsel)
1698 evsel->tracking = false;
1701 tracking_evsel->tracking = true;