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>
28 #include <linux/err.h>
30 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
31 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
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)
36 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
37 struct thread_map *threads)
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;
49 struct perf_evlist *perf_evlist__new(void)
51 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
54 perf_evlist__init(evlist, NULL, NULL);
59 struct perf_evlist *perf_evlist__new_default(void)
61 struct perf_evlist *evlist = perf_evlist__new();
63 if (evlist && perf_evlist__add_default(evlist)) {
64 perf_evlist__delete(evlist);
72 * perf_evlist__set_id_pos - set the positions of event ids.
73 * @evlist: selected event list
75 * Events with compatible sample types all have the same id_pos
76 * and is_pos. For convenience, put a copy on evlist.
78 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
80 struct perf_evsel *first = perf_evlist__first(evlist);
82 evlist->id_pos = first->id_pos;
83 evlist->is_pos = first->is_pos;
86 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
88 struct perf_evsel *evsel;
90 evlist__for_each(evlist, evsel)
91 perf_evsel__calc_id_pos(evsel);
93 perf_evlist__set_id_pos(evlist);
96 static void perf_evlist__purge(struct perf_evlist *evlist)
98 struct perf_evsel *pos, *n;
100 evlist__for_each_safe(evlist, n, pos) {
101 list_del_init(&pos->node);
103 perf_evsel__delete(pos);
106 evlist->nr_entries = 0;
109 void perf_evlist__exit(struct perf_evlist *evlist)
111 zfree(&evlist->mmap);
112 fdarray__exit(&evlist->pollfd);
115 void perf_evlist__delete(struct perf_evlist *evlist)
117 perf_evlist__munmap(evlist);
118 perf_evlist__close(evlist);
119 cpu_map__put(evlist->cpus);
120 thread_map__put(evlist->threads);
122 evlist->threads = NULL;
123 perf_evlist__purge(evlist);
124 perf_evlist__exit(evlist);
128 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
130 entry->evlist = evlist;
131 list_add_tail(&entry->node, &evlist->entries);
132 entry->idx = evlist->nr_entries;
133 entry->tracking = !entry->idx;
135 if (!evlist->nr_entries++)
136 perf_evlist__set_id_pos(evlist);
139 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
140 struct list_head *list,
143 bool set_id_pos = !evlist->nr_entries;
145 list_splice_tail(list, &evlist->entries);
146 evlist->nr_entries += nr_entries;
148 perf_evlist__set_id_pos(evlist);
151 void __perf_evlist__set_leader(struct list_head *list)
153 struct perf_evsel *evsel, *leader;
155 leader = list_entry(list->next, struct perf_evsel, node);
156 evsel = list_entry(list->prev, struct perf_evsel, node);
158 leader->nr_members = evsel->idx - leader->idx + 1;
160 __evlist__for_each(list, evsel) {
161 evsel->leader = leader;
165 void perf_evlist__set_leader(struct perf_evlist *evlist)
167 if (evlist->nr_entries) {
168 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
169 __perf_evlist__set_leader(&evlist->entries);
173 int perf_evlist__add_default(struct perf_evlist *evlist)
175 struct perf_event_attr attr = {
176 .type = PERF_TYPE_HARDWARE,
177 .config = PERF_COUNT_HW_CPU_CYCLES,
179 struct perf_evsel *evsel;
181 event_attr_init(&attr);
183 evsel = perf_evsel__new(&attr);
187 /* use strdup() because free(evsel) assumes name is allocated */
188 evsel->name = strdup("cycles");
192 perf_evlist__add(evlist, evsel);
195 perf_evsel__delete(evsel);
200 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
201 struct perf_event_attr *attrs, size_t nr_attrs)
203 struct perf_evsel *evsel, *n;
207 for (i = 0; i < nr_attrs; i++) {
208 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
210 goto out_delete_partial_list;
211 list_add_tail(&evsel->node, &head);
214 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
218 out_delete_partial_list:
219 __evlist__for_each_safe(&head, n, evsel)
220 perf_evsel__delete(evsel);
224 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
225 struct perf_event_attr *attrs, size_t nr_attrs)
229 for (i = 0; i < nr_attrs; i++)
230 event_attr_init(attrs + i);
232 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
236 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
238 struct perf_evsel *evsel;
240 evlist__for_each(evlist, evsel) {
241 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
242 (int)evsel->attr.config == id)
250 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
253 struct perf_evsel *evsel;
255 evlist__for_each(evlist, evsel) {
256 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
257 (strcmp(evsel->name, name) == 0))
264 int perf_evlist__add_newtp(struct perf_evlist *evlist,
265 const char *sys, const char *name, void *handler)
267 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
272 evsel->handler = handler;
273 perf_evlist__add(evlist, evsel);
277 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
278 struct perf_evsel *evsel)
280 if (evsel->system_wide)
283 return thread_map__nr(evlist->threads);
286 void perf_evlist__disable(struct perf_evlist *evlist)
289 struct perf_evsel *pos;
290 int nr_cpus = cpu_map__nr(evlist->cpus);
293 for (cpu = 0; cpu < nr_cpus; cpu++) {
294 evlist__for_each(evlist, pos) {
295 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
297 nr_threads = perf_evlist__nr_threads(evlist, pos);
298 for (thread = 0; thread < nr_threads; thread++)
299 ioctl(FD(pos, cpu, thread),
300 PERF_EVENT_IOC_DISABLE, 0);
304 evlist->enabled = false;
307 void perf_evlist__enable(struct perf_evlist *evlist)
310 struct perf_evsel *pos;
311 int nr_cpus = cpu_map__nr(evlist->cpus);
314 for (cpu = 0; cpu < nr_cpus; cpu++) {
315 evlist__for_each(evlist, pos) {
316 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
318 nr_threads = perf_evlist__nr_threads(evlist, pos);
319 for (thread = 0; thread < nr_threads; thread++)
320 ioctl(FD(pos, cpu, thread),
321 PERF_EVENT_IOC_ENABLE, 0);
325 evlist->enabled = true;
328 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
330 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
333 int perf_evlist__disable_event(struct perf_evlist *evlist,
334 struct perf_evsel *evsel)
336 int cpu, thread, err;
337 int nr_cpus = cpu_map__nr(evlist->cpus);
338 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
343 for (cpu = 0; cpu < nr_cpus; cpu++) {
344 for (thread = 0; thread < nr_threads; thread++) {
345 err = ioctl(FD(evsel, cpu, thread),
346 PERF_EVENT_IOC_DISABLE, 0);
354 int perf_evlist__enable_event(struct perf_evlist *evlist,
355 struct perf_evsel *evsel)
357 int cpu, thread, err;
358 int nr_cpus = cpu_map__nr(evlist->cpus);
359 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
364 for (cpu = 0; cpu < nr_cpus; cpu++) {
365 for (thread = 0; thread < nr_threads; thread++) {
366 err = ioctl(FD(evsel, cpu, thread),
367 PERF_EVENT_IOC_ENABLE, 0);
375 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
376 struct perf_evsel *evsel, int cpu)
379 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
384 for (thread = 0; thread < nr_threads; thread++) {
385 err = ioctl(FD(evsel, cpu, thread),
386 PERF_EVENT_IOC_ENABLE, 0);
393 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
394 struct perf_evsel *evsel,
398 int nr_cpus = cpu_map__nr(evlist->cpus);
403 for (cpu = 0; cpu < nr_cpus; cpu++) {
404 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
411 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
412 struct perf_evsel *evsel, int idx)
414 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
417 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
419 return perf_evlist__enable_event_thread(evlist, evsel, idx);
422 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
424 int nr_cpus = cpu_map__nr(evlist->cpus);
425 int nr_threads = thread_map__nr(evlist->threads);
427 struct perf_evsel *evsel;
429 evlist__for_each(evlist, evsel) {
430 if (evsel->system_wide)
433 nfds += nr_cpus * nr_threads;
436 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
437 fdarray__grow(&evlist->pollfd, nfds) < 0)
443 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
445 int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
447 * Save the idx so that when we filter out fds POLLHUP'ed we can
448 * close the associated evlist->mmap[] entry.
451 evlist->pollfd.priv[pos].idx = idx;
453 fcntl(fd, F_SETFL, O_NONBLOCK);
459 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
461 return __perf_evlist__add_pollfd(evlist, fd, -1);
464 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
466 struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
468 perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
471 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
473 return fdarray__filter(&evlist->pollfd, revents_and_mask,
474 perf_evlist__munmap_filtered);
477 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
479 return fdarray__poll(&evlist->pollfd, timeout);
482 static void perf_evlist__id_hash(struct perf_evlist *evlist,
483 struct perf_evsel *evsel,
484 int cpu, int thread, u64 id)
487 struct perf_sample_id *sid = SID(evsel, cpu, thread);
491 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
492 hlist_add_head(&sid->node, &evlist->heads[hash]);
495 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
496 int cpu, int thread, u64 id)
498 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
499 evsel->id[evsel->ids++] = id;
502 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
503 struct perf_evsel *evsel,
504 int cpu, int thread, int fd)
506 u64 read_data[4] = { 0, };
507 int id_idx = 1; /* The first entry is the counter value */
511 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
518 /* Legacy way to get event id.. All hail to old kernels! */
521 * This way does not work with group format read, so bail
524 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
527 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
528 read(fd, &read_data, sizeof(read_data)) == -1)
531 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
533 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
536 id = read_data[id_idx];
539 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
543 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
544 struct perf_evsel *evsel, int idx, int cpu,
547 struct perf_sample_id *sid = SID(evsel, cpu, thread);
549 if (evlist->cpus && cpu >= 0)
550 sid->cpu = evlist->cpus->map[cpu];
553 if (!evsel->system_wide && evlist->threads && thread >= 0)
554 sid->tid = thread_map__pid(evlist->threads, thread);
559 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
561 struct hlist_head *head;
562 struct perf_sample_id *sid;
565 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
566 head = &evlist->heads[hash];
568 hlist_for_each_entry(sid, head, node)
575 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
577 struct perf_sample_id *sid;
579 if (evlist->nr_entries == 1 || !id)
580 return perf_evlist__first(evlist);
582 sid = perf_evlist__id2sid(evlist, id);
586 if (!perf_evlist__sample_id_all(evlist))
587 return perf_evlist__first(evlist);
592 static int perf_evlist__event2id(struct perf_evlist *evlist,
593 union perf_event *event, u64 *id)
595 const u64 *array = event->sample.array;
598 n = (event->header.size - sizeof(event->header)) >> 3;
600 if (event->header.type == PERF_RECORD_SAMPLE) {
601 if (evlist->id_pos >= n)
603 *id = array[evlist->id_pos];
605 if (evlist->is_pos > n)
613 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
614 union perf_event *event)
616 struct perf_evsel *first = perf_evlist__first(evlist);
617 struct hlist_head *head;
618 struct perf_sample_id *sid;
622 if (evlist->nr_entries == 1)
625 if (!first->attr.sample_id_all &&
626 event->header.type != PERF_RECORD_SAMPLE)
629 if (perf_evlist__event2id(evlist, event, &id))
632 /* Synthesized events have an id of zero */
636 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
637 head = &evlist->heads[hash];
639 hlist_for_each_entry(sid, head, node) {
646 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
648 struct perf_mmap *md = &evlist->mmap[idx];
651 unsigned char *data = md->base + page_size;
652 union perf_event *event = NULL;
655 * Check if event was unmapped due to a POLLHUP/POLLERR.
657 if (!atomic_read(&md->refcnt))
660 head = perf_mmap__read_head(md);
661 if (evlist->overwrite) {
663 * If we're further behind than half the buffer, there's a chance
664 * the writer will bite our tail and mess up the samples under us.
666 * If we somehow ended up ahead of the head, we got messed up.
668 * In either case, truncate and restart at head.
670 int diff = head - old;
671 if (diff > md->mask / 2 || diff < 0) {
672 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
675 * head points to a known good entry, start there.
684 event = (union perf_event *)&data[old & md->mask];
685 size = event->header.size;
688 * Event straddles the mmap boundary -- header should always
689 * be inside due to u64 alignment of output.
691 if ((old & md->mask) + size != ((old + size) & md->mask)) {
692 unsigned int offset = old;
693 unsigned int len = min(sizeof(*event), size), cpy;
694 void *dst = md->event_copy;
697 cpy = min(md->mask + 1 - (offset & md->mask), len);
698 memcpy(dst, &data[offset & md->mask], cpy);
704 event = (union perf_event *) md->event_copy;
715 static bool perf_mmap__empty(struct perf_mmap *md)
717 return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
720 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
722 atomic_inc(&evlist->mmap[idx].refcnt);
725 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
727 BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
729 if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
730 __perf_evlist__munmap(evlist, idx);
733 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
735 struct perf_mmap *md = &evlist->mmap[idx];
737 if (!evlist->overwrite) {
740 perf_mmap__write_tail(md, old);
743 if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
744 perf_evlist__mmap_put(evlist, idx);
747 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
748 struct auxtrace_mmap_params *mp __maybe_unused,
749 void *userpg __maybe_unused,
750 int fd __maybe_unused)
755 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
759 void __weak auxtrace_mmap_params__init(
760 struct auxtrace_mmap_params *mp __maybe_unused,
761 off_t auxtrace_offset __maybe_unused,
762 unsigned int auxtrace_pages __maybe_unused,
763 bool auxtrace_overwrite __maybe_unused)
767 void __weak auxtrace_mmap_params__set_idx(
768 struct auxtrace_mmap_params *mp __maybe_unused,
769 struct perf_evlist *evlist __maybe_unused,
770 int idx __maybe_unused,
771 bool per_cpu __maybe_unused)
775 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
777 if (evlist->mmap[idx].base != NULL) {
778 munmap(evlist->mmap[idx].base, evlist->mmap_len);
779 evlist->mmap[idx].base = NULL;
780 atomic_set(&evlist->mmap[idx].refcnt, 0);
782 auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
785 void perf_evlist__munmap(struct perf_evlist *evlist)
789 if (evlist->mmap == NULL)
792 for (i = 0; i < evlist->nr_mmaps; i++)
793 __perf_evlist__munmap(evlist, i);
795 zfree(&evlist->mmap);
798 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
800 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
801 if (cpu_map__empty(evlist->cpus))
802 evlist->nr_mmaps = thread_map__nr(evlist->threads);
803 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
804 return evlist->mmap != NULL ? 0 : -ENOMEM;
810 struct auxtrace_mmap_params auxtrace_mp;
813 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
814 struct mmap_params *mp, int fd)
817 * The last one will be done at perf_evlist__mmap_consume(), so that we
818 * make sure we don't prevent tools from consuming every last event in
821 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
822 * anymore, but the last events for it are still in the ring buffer,
823 * waiting to be consumed.
825 * Tools can chose to ignore this at their own discretion, but the
826 * evlist layer can't just drop it when filtering events in
827 * perf_evlist__filter_pollfd().
829 atomic_set(&evlist->mmap[idx].refcnt, 2);
830 evlist->mmap[idx].prev = 0;
831 evlist->mmap[idx].mask = mp->mask;
832 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
834 if (evlist->mmap[idx].base == MAP_FAILED) {
835 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
837 evlist->mmap[idx].base = NULL;
841 if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
842 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
848 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
849 struct mmap_params *mp, int cpu,
850 int thread, int *output)
852 struct perf_evsel *evsel;
854 evlist__for_each(evlist, evsel) {
857 if (evsel->system_wide && thread)
860 fd = FD(evsel, cpu, thread);
864 if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
867 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
870 perf_evlist__mmap_get(evlist, idx);
874 * The system_wide flag causes a selected event to be opened
875 * always without a pid. Consequently it will never get a
876 * POLLHUP, but it is used for tracking in combination with
877 * other events, so it should not need to be polled anyway.
878 * Therefore don't add it for polling.
880 if (!evsel->system_wide &&
881 __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
882 perf_evlist__mmap_put(evlist, idx);
886 if (evsel->attr.read_format & PERF_FORMAT_ID) {
887 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
890 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
898 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
899 struct mmap_params *mp)
902 int nr_cpus = cpu_map__nr(evlist->cpus);
903 int nr_threads = thread_map__nr(evlist->threads);
905 pr_debug2("perf event ring buffer mmapped per cpu\n");
906 for (cpu = 0; cpu < nr_cpus; cpu++) {
909 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
912 for (thread = 0; thread < nr_threads; thread++) {
913 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
922 for (cpu = 0; cpu < nr_cpus; cpu++)
923 __perf_evlist__munmap(evlist, cpu);
927 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
928 struct mmap_params *mp)
931 int nr_threads = thread_map__nr(evlist->threads);
933 pr_debug2("perf event ring buffer mmapped per thread\n");
934 for (thread = 0; thread < nr_threads; thread++) {
937 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
940 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
948 for (thread = 0; thread < nr_threads; thread++)
949 __perf_evlist__munmap(evlist, thread);
953 static size_t perf_evlist__mmap_size(unsigned long pages)
955 if (pages == UINT_MAX) {
958 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
960 * Pick a once upon a time good value, i.e. things look
961 * strange since we can't read a sysctl value, but lets not
966 max -= (page_size / 1024);
969 pages = (max * 1024) / page_size;
970 if (!is_power_of_2(pages))
971 pages = rounddown_pow_of_two(pages);
972 } else if (!is_power_of_2(pages))
975 return (pages + 1) * page_size;
978 static long parse_pages_arg(const char *str, unsigned long min,
981 unsigned long pages, val;
982 static struct parse_tag tags[] = {
983 { .tag = 'B', .mult = 1 },
984 { .tag = 'K', .mult = 1 << 10 },
985 { .tag = 'M', .mult = 1 << 20 },
986 { .tag = 'G', .mult = 1 << 30 },
993 val = parse_tag_value(str, tags);
994 if (val != (unsigned long) -1) {
995 /* we got file size value */
996 pages = PERF_ALIGN(val, page_size) / page_size;
998 /* we got pages count value */
1000 pages = strtoul(str, &eptr, 10);
1005 if (pages == 0 && min == 0) {
1006 /* leave number of pages at 0 */
1007 } else if (!is_power_of_2(pages)) {
1008 /* round pages up to next power of 2 */
1009 pages = roundup_pow_of_two(pages);
1012 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1013 pages * page_size, pages);
1022 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1024 unsigned long max = UINT_MAX;
1027 if (max > SIZE_MAX / page_size)
1028 max = SIZE_MAX / page_size;
1030 pages = parse_pages_arg(str, 1, max);
1032 pr_err("Invalid argument for --mmap_pages/-m\n");
1036 *mmap_pages = pages;
1040 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1041 int unset __maybe_unused)
1043 return __perf_evlist__parse_mmap_pages(opt->value, str);
1047 * perf_evlist__mmap_ex - Create mmaps to receive events.
1048 * @evlist: list of events
1049 * @pages: map length in pages
1050 * @overwrite: overwrite older events?
1051 * @auxtrace_pages - auxtrace map length in pages
1052 * @auxtrace_overwrite - overwrite older auxtrace data?
1054 * If @overwrite is %false the user needs to signal event consumption using
1055 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1058 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1059 * consumption using auxtrace_mmap__write_tail().
1061 * Return: %0 on success, negative error code otherwise.
1063 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1064 bool overwrite, unsigned int auxtrace_pages,
1065 bool auxtrace_overwrite)
1067 struct perf_evsel *evsel;
1068 const struct cpu_map *cpus = evlist->cpus;
1069 const struct thread_map *threads = evlist->threads;
1070 struct mmap_params mp = {
1071 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1074 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1077 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1080 evlist->overwrite = overwrite;
1081 evlist->mmap_len = perf_evlist__mmap_size(pages);
1082 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1083 mp.mask = evlist->mmap_len - page_size - 1;
1085 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1086 auxtrace_pages, auxtrace_overwrite);
1088 evlist__for_each(evlist, evsel) {
1089 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1090 evsel->sample_id == NULL &&
1091 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1095 if (cpu_map__empty(cpus))
1096 return perf_evlist__mmap_per_thread(evlist, &mp);
1098 return perf_evlist__mmap_per_cpu(evlist, &mp);
1101 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1104 return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1107 static int perf_evlist__propagate_maps(struct perf_evlist *evlist,
1110 struct perf_evsel *evsel;
1112 evlist__for_each(evlist, evsel) {
1114 * We already have cpus for evsel (via PMU sysfs) so
1115 * keep it, if there's no target cpu list defined.
1117 if (evsel->cpus && has_user_cpus)
1118 cpu_map__put(evsel->cpus);
1120 if (!evsel->cpus || has_user_cpus)
1121 evsel->cpus = cpu_map__get(evlist->cpus);
1123 evsel->threads = thread_map__get(evlist->threads);
1125 if ((evlist->cpus && !evsel->cpus) ||
1126 (evlist->threads && !evsel->threads))
1133 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1135 evlist->threads = thread_map__new_str(target->pid, target->tid,
1138 if (evlist->threads == NULL)
1141 if (target__uses_dummy_map(target))
1142 evlist->cpus = cpu_map__dummy_new();
1144 evlist->cpus = cpu_map__new(target->cpu_list);
1146 if (evlist->cpus == NULL)
1147 goto out_delete_threads;
1149 return perf_evlist__propagate_maps(evlist, !!target->cpu_list);
1152 thread_map__put(evlist->threads);
1153 evlist->threads = NULL;
1157 int perf_evlist__set_maps(struct perf_evlist *evlist,
1158 struct cpu_map *cpus,
1159 struct thread_map *threads)
1162 cpu_map__put(evlist->cpus);
1164 evlist->cpus = cpus;
1166 if (evlist->threads)
1167 thread_map__put(evlist->threads);
1169 evlist->threads = threads;
1171 return perf_evlist__propagate_maps(evlist, false);
1174 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1176 struct perf_evsel *evsel;
1178 const int ncpus = cpu_map__nr(evlist->cpus),
1179 nthreads = thread_map__nr(evlist->threads);
1181 evlist__for_each(evlist, evsel) {
1182 if (evsel->filter == NULL)
1186 * filters only work for tracepoint event, which doesn't have cpu limit.
1187 * So evlist and evsel should always be same.
1189 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1199 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1201 struct perf_evsel *evsel;
1204 evlist__for_each(evlist, evsel) {
1205 err = perf_evsel__set_filter(evsel, filter);
1213 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1219 for (i = 0; i < npids; ++i) {
1221 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1226 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1234 ret = perf_evlist__set_filter(evlist, filter);
1240 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1242 return perf_evlist__set_filter_pids(evlist, 1, &pid);
1245 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1247 struct perf_evsel *pos;
1249 if (evlist->nr_entries == 1)
1252 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1255 evlist__for_each(evlist, pos) {
1256 if (pos->id_pos != evlist->id_pos ||
1257 pos->is_pos != evlist->is_pos)
1264 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1266 struct perf_evsel *evsel;
1268 if (evlist->combined_sample_type)
1269 return evlist->combined_sample_type;
1271 evlist__for_each(evlist, evsel)
1272 evlist->combined_sample_type |= evsel->attr.sample_type;
1274 return evlist->combined_sample_type;
1277 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1279 evlist->combined_sample_type = 0;
1280 return __perf_evlist__combined_sample_type(evlist);
1283 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1285 struct perf_evsel *evsel;
1286 u64 branch_type = 0;
1288 evlist__for_each(evlist, evsel)
1289 branch_type |= evsel->attr.branch_sample_type;
1293 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1295 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1296 u64 read_format = first->attr.read_format;
1297 u64 sample_type = first->attr.sample_type;
1299 evlist__for_each(evlist, pos) {
1300 if (read_format != pos->attr.read_format)
1304 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1305 if ((sample_type & PERF_SAMPLE_READ) &&
1306 !(read_format & PERF_FORMAT_ID)) {
1313 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1315 struct perf_evsel *first = perf_evlist__first(evlist);
1316 return first->attr.read_format;
1319 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1321 struct perf_evsel *first = perf_evlist__first(evlist);
1322 struct perf_sample *data;
1326 if (!first->attr.sample_id_all)
1329 sample_type = first->attr.sample_type;
1331 if (sample_type & PERF_SAMPLE_TID)
1332 size += sizeof(data->tid) * 2;
1334 if (sample_type & PERF_SAMPLE_TIME)
1335 size += sizeof(data->time);
1337 if (sample_type & PERF_SAMPLE_ID)
1338 size += sizeof(data->id);
1340 if (sample_type & PERF_SAMPLE_STREAM_ID)
1341 size += sizeof(data->stream_id);
1343 if (sample_type & PERF_SAMPLE_CPU)
1344 size += sizeof(data->cpu) * 2;
1346 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1347 size += sizeof(data->id);
1352 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1354 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1356 evlist__for_each_continue(evlist, pos) {
1357 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1364 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1366 struct perf_evsel *first = perf_evlist__first(evlist);
1367 return first->attr.sample_id_all;
1370 void perf_evlist__set_selected(struct perf_evlist *evlist,
1371 struct perf_evsel *evsel)
1373 evlist->selected = evsel;
1376 void perf_evlist__close(struct perf_evlist *evlist)
1378 struct perf_evsel *evsel;
1379 int ncpus = cpu_map__nr(evlist->cpus);
1380 int nthreads = thread_map__nr(evlist->threads);
1383 evlist__for_each_reverse(evlist, evsel) {
1384 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1385 perf_evsel__close(evsel, n, nthreads);
1389 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1394 * Try reading /sys/devices/system/cpu/online to get
1397 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1398 * code needs an overhaul to properly forward the
1399 * error, and we may not want to do that fallback to a
1400 * default cpu identity map :-\
1402 evlist->cpus = cpu_map__new(NULL);
1403 if (evlist->cpus == NULL)
1406 evlist->threads = thread_map__new_dummy();
1407 if (evlist->threads == NULL)
1414 cpu_map__put(evlist->cpus);
1415 evlist->cpus = NULL;
1419 int perf_evlist__open(struct perf_evlist *evlist)
1421 struct perf_evsel *evsel;
1425 * Default: one fd per CPU, all threads, aka systemwide
1426 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1428 if (evlist->threads == NULL && evlist->cpus == NULL) {
1429 err = perf_evlist__create_syswide_maps(evlist);
1434 perf_evlist__update_id_pos(evlist);
1436 evlist__for_each(evlist, evsel) {
1437 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1444 perf_evlist__close(evlist);
1449 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1450 const char *argv[], bool pipe_output,
1451 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1453 int child_ready_pipe[2], go_pipe[2];
1456 if (pipe(child_ready_pipe) < 0) {
1457 perror("failed to create 'ready' pipe");
1461 if (pipe(go_pipe) < 0) {
1462 perror("failed to create 'go' pipe");
1463 goto out_close_ready_pipe;
1466 evlist->workload.pid = fork();
1467 if (evlist->workload.pid < 0) {
1468 perror("failed to fork");
1469 goto out_close_pipes;
1472 if (!evlist->workload.pid) {
1478 signal(SIGTERM, SIG_DFL);
1480 close(child_ready_pipe[0]);
1482 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1485 * Tell the parent we're ready to go
1487 close(child_ready_pipe[1]);
1490 * Wait until the parent tells us to go.
1492 ret = read(go_pipe[0], &bf, 1);
1494 * The parent will ask for the execvp() to be performed by
1495 * writing exactly one byte, in workload.cork_fd, usually via
1496 * perf_evlist__start_workload().
1498 * For cancelling the workload without actually running it,
1499 * the parent will just close workload.cork_fd, without writing
1500 * anything, i.e. read will return zero and we just exit()
1505 perror("unable to read pipe");
1509 execvp(argv[0], (char **)argv);
1514 val.sival_int = errno;
1515 if (sigqueue(getppid(), SIGUSR1, val))
1523 struct sigaction act = {
1524 .sa_flags = SA_SIGINFO,
1525 .sa_sigaction = exec_error,
1527 sigaction(SIGUSR1, &act, NULL);
1530 if (target__none(target)) {
1531 if (evlist->threads == NULL) {
1532 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1533 __func__, __LINE__);
1534 goto out_close_pipes;
1536 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1539 close(child_ready_pipe[1]);
1542 * wait for child to settle
1544 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1545 perror("unable to read pipe");
1546 goto out_close_pipes;
1549 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1550 evlist->workload.cork_fd = go_pipe[1];
1551 close(child_ready_pipe[0]);
1557 out_close_ready_pipe:
1558 close(child_ready_pipe[0]);
1559 close(child_ready_pipe[1]);
1563 int perf_evlist__start_workload(struct perf_evlist *evlist)
1565 if (evlist->workload.cork_fd > 0) {
1569 * Remove the cork, let it rip!
1571 ret = write(evlist->workload.cork_fd, &bf, 1);
1573 perror("enable to write to pipe");
1575 close(evlist->workload.cork_fd);
1582 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1583 struct perf_sample *sample)
1585 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1589 return perf_evsel__parse_sample(evsel, event, sample);
1592 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1594 struct perf_evsel *evsel;
1597 evlist__for_each(evlist, evsel) {
1598 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1599 perf_evsel__name(evsel));
1602 return printed + fprintf(fp, "\n");
1605 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1606 int err, char *buf, size_t size)
1609 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1614 printed = scnprintf(buf, size,
1616 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1618 value = perf_event_paranoid();
1620 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1623 printed += scnprintf(buf + printed, size - printed,
1624 "For your workloads it needs to be <= 1\nHint:\t");
1626 printed += scnprintf(buf + printed, size - printed,
1627 "For system wide tracing it needs to be set to -1.\n");
1629 printed += scnprintf(buf + printed, size - printed,
1630 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1631 "Hint:\tThe current value is %d.", value);
1634 scnprintf(buf, size, "%s", emsg);
1641 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1643 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1644 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1648 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1649 printed += scnprintf(buf + printed, size - printed,
1651 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1652 "Hint:\tTried using %zd kB.\n",
1653 emsg, pages_max_per_user, pages_attempted);
1655 if (pages_attempted >= pages_max_per_user) {
1656 printed += scnprintf(buf + printed, size - printed,
1657 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1658 pages_max_per_user + pages_attempted);
1661 printed += scnprintf(buf + printed, size - printed,
1662 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1665 scnprintf(buf, size, "%s", emsg);
1672 void perf_evlist__to_front(struct perf_evlist *evlist,
1673 struct perf_evsel *move_evsel)
1675 struct perf_evsel *evsel, *n;
1678 if (move_evsel == perf_evlist__first(evlist))
1681 evlist__for_each_safe(evlist, n, evsel) {
1682 if (evsel->leader == move_evsel->leader)
1683 list_move_tail(&evsel->node, &move);
1686 list_splice(&move, &evlist->entries);
1689 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1690 struct perf_evsel *tracking_evsel)
1692 struct perf_evsel *evsel;
1694 if (tracking_evsel->tracking)
1697 evlist__for_each(evlist, evsel) {
1698 if (evsel != tracking_evsel)
1699 evsel->tracking = false;
1702 tracking_evsel->tracking = true;