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 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
129 struct perf_evsel *evsel)
132 * We already have cpus for evsel (via PMU sysfs) so
133 * keep it, if there's no target cpu list defined.
135 if (!evsel->own_cpus || evlist->has_user_cpus) {
136 cpu_map__put(evsel->cpus);
137 evsel->cpus = cpu_map__get(evlist->cpus);
138 } else if (evsel->cpus != evsel->own_cpus) {
139 cpu_map__put(evsel->cpus);
140 evsel->cpus = cpu_map__get(evsel->own_cpus);
143 thread_map__put(evsel->threads);
144 evsel->threads = thread_map__get(evlist->threads);
147 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
149 struct perf_evsel *evsel;
151 evlist__for_each(evlist, evsel)
152 __perf_evlist__propagate_maps(evlist, evsel);
155 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
157 entry->evlist = evlist;
158 list_add_tail(&entry->node, &evlist->entries);
159 entry->idx = evlist->nr_entries;
160 entry->tracking = !entry->idx;
162 if (!evlist->nr_entries++)
163 perf_evlist__set_id_pos(evlist);
165 __perf_evlist__propagate_maps(evlist, entry);
168 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
169 struct list_head *list)
171 struct perf_evsel *evsel, *temp;
173 __evlist__for_each_safe(list, temp, evsel) {
174 list_del_init(&evsel->node);
175 perf_evlist__add(evlist, evsel);
179 void __perf_evlist__set_leader(struct list_head *list)
181 struct perf_evsel *evsel, *leader;
183 leader = list_entry(list->next, struct perf_evsel, node);
184 evsel = list_entry(list->prev, struct perf_evsel, node);
186 leader->nr_members = evsel->idx - leader->idx + 1;
188 __evlist__for_each(list, evsel) {
189 evsel->leader = leader;
193 void perf_evlist__set_leader(struct perf_evlist *evlist)
195 if (evlist->nr_entries) {
196 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
197 __perf_evlist__set_leader(&evlist->entries);
201 int perf_evlist__add_default(struct perf_evlist *evlist)
203 struct perf_event_attr attr = {
204 .type = PERF_TYPE_HARDWARE,
205 .config = PERF_COUNT_HW_CPU_CYCLES,
207 struct perf_evsel *evsel;
209 event_attr_init(&attr);
211 evsel = perf_evsel__new(&attr);
215 /* use strdup() because free(evsel) assumes name is allocated */
216 evsel->name = strdup("cycles");
220 perf_evlist__add(evlist, evsel);
223 perf_evsel__delete(evsel);
228 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
229 struct perf_event_attr *attrs, size_t nr_attrs)
231 struct perf_evsel *evsel, *n;
235 for (i = 0; i < nr_attrs; i++) {
236 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
238 goto out_delete_partial_list;
239 list_add_tail(&evsel->node, &head);
242 perf_evlist__splice_list_tail(evlist, &head);
246 out_delete_partial_list:
247 __evlist__for_each_safe(&head, n, evsel)
248 perf_evsel__delete(evsel);
252 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
253 struct perf_event_attr *attrs, size_t nr_attrs)
257 for (i = 0; i < nr_attrs; i++)
258 event_attr_init(attrs + i);
260 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
264 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
266 struct perf_evsel *evsel;
268 evlist__for_each(evlist, evsel) {
269 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
270 (int)evsel->attr.config == id)
278 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
281 struct perf_evsel *evsel;
283 evlist__for_each(evlist, evsel) {
284 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
285 (strcmp(evsel->name, name) == 0))
292 int perf_evlist__add_newtp(struct perf_evlist *evlist,
293 const char *sys, const char *name, void *handler)
295 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
300 evsel->handler = handler;
301 perf_evlist__add(evlist, evsel);
305 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
306 struct perf_evsel *evsel)
308 if (evsel->system_wide)
311 return thread_map__nr(evlist->threads);
314 void perf_evlist__disable(struct perf_evlist *evlist)
317 struct perf_evsel *pos;
318 int nr_cpus = cpu_map__nr(evlist->cpus);
321 for (cpu = 0; cpu < nr_cpus; cpu++) {
322 evlist__for_each(evlist, pos) {
323 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
325 nr_threads = perf_evlist__nr_threads(evlist, pos);
326 for (thread = 0; thread < nr_threads; thread++)
327 ioctl(FD(pos, cpu, thread),
328 PERF_EVENT_IOC_DISABLE, 0);
332 evlist->enabled = false;
335 void perf_evlist__enable(struct perf_evlist *evlist)
338 struct perf_evsel *pos;
339 int nr_cpus = cpu_map__nr(evlist->cpus);
342 for (cpu = 0; cpu < nr_cpus; cpu++) {
343 evlist__for_each(evlist, pos) {
344 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
346 nr_threads = perf_evlist__nr_threads(evlist, pos);
347 for (thread = 0; thread < nr_threads; thread++)
348 ioctl(FD(pos, cpu, thread),
349 PERF_EVENT_IOC_ENABLE, 0);
353 evlist->enabled = true;
356 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
358 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
361 int perf_evlist__disable_event(struct perf_evlist *evlist,
362 struct perf_evsel *evsel)
364 int cpu, thread, err;
365 int nr_cpus = cpu_map__nr(evlist->cpus);
366 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
371 for (cpu = 0; cpu < nr_cpus; cpu++) {
372 for (thread = 0; thread < nr_threads; thread++) {
373 err = ioctl(FD(evsel, cpu, thread),
374 PERF_EVENT_IOC_DISABLE, 0);
382 int perf_evlist__enable_event(struct perf_evlist *evlist,
383 struct perf_evsel *evsel)
385 int cpu, thread, err;
386 int nr_cpus = cpu_map__nr(evlist->cpus);
387 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
392 for (cpu = 0; cpu < nr_cpus; cpu++) {
393 for (thread = 0; thread < nr_threads; thread++) {
394 err = ioctl(FD(evsel, cpu, thread),
395 PERF_EVENT_IOC_ENABLE, 0);
403 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
404 struct perf_evsel *evsel, int cpu)
407 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
412 for (thread = 0; thread < nr_threads; thread++) {
413 err = ioctl(FD(evsel, cpu, thread),
414 PERF_EVENT_IOC_ENABLE, 0);
421 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
422 struct perf_evsel *evsel,
426 int nr_cpus = cpu_map__nr(evlist->cpus);
431 for (cpu = 0; cpu < nr_cpus; cpu++) {
432 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
439 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
440 struct perf_evsel *evsel, int idx)
442 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
445 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
447 return perf_evlist__enable_event_thread(evlist, evsel, idx);
450 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
452 int nr_cpus = cpu_map__nr(evlist->cpus);
453 int nr_threads = thread_map__nr(evlist->threads);
455 struct perf_evsel *evsel;
457 evlist__for_each(evlist, evsel) {
458 if (evsel->system_wide)
461 nfds += nr_cpus * nr_threads;
464 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
465 fdarray__grow(&evlist->pollfd, nfds) < 0)
471 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
473 int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
475 * Save the idx so that when we filter out fds POLLHUP'ed we can
476 * close the associated evlist->mmap[] entry.
479 evlist->pollfd.priv[pos].idx = idx;
481 fcntl(fd, F_SETFL, O_NONBLOCK);
487 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
489 return __perf_evlist__add_pollfd(evlist, fd, -1);
492 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
494 struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
496 perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
499 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
501 return fdarray__filter(&evlist->pollfd, revents_and_mask,
502 perf_evlist__munmap_filtered);
505 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
507 return fdarray__poll(&evlist->pollfd, timeout);
510 static void perf_evlist__id_hash(struct perf_evlist *evlist,
511 struct perf_evsel *evsel,
512 int cpu, int thread, u64 id)
515 struct perf_sample_id *sid = SID(evsel, cpu, thread);
519 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
520 hlist_add_head(&sid->node, &evlist->heads[hash]);
523 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
524 int cpu, int thread, u64 id)
526 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
527 evsel->id[evsel->ids++] = id;
530 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
531 struct perf_evsel *evsel,
532 int cpu, int thread, int fd)
534 u64 read_data[4] = { 0, };
535 int id_idx = 1; /* The first entry is the counter value */
539 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
546 /* Legacy way to get event id.. All hail to old kernels! */
549 * This way does not work with group format read, so bail
552 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
555 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
556 read(fd, &read_data, sizeof(read_data)) == -1)
559 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
561 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
564 id = read_data[id_idx];
567 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
571 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
572 struct perf_evsel *evsel, int idx, int cpu,
575 struct perf_sample_id *sid = SID(evsel, cpu, thread);
577 if (evlist->cpus && cpu >= 0)
578 sid->cpu = evlist->cpus->map[cpu];
581 if (!evsel->system_wide && evlist->threads && thread >= 0)
582 sid->tid = thread_map__pid(evlist->threads, thread);
587 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
589 struct hlist_head *head;
590 struct perf_sample_id *sid;
593 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
594 head = &evlist->heads[hash];
596 hlist_for_each_entry(sid, head, node)
603 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
605 struct perf_sample_id *sid;
607 if (evlist->nr_entries == 1 || !id)
608 return perf_evlist__first(evlist);
610 sid = perf_evlist__id2sid(evlist, id);
614 if (!perf_evlist__sample_id_all(evlist))
615 return perf_evlist__first(evlist);
620 static int perf_evlist__event2id(struct perf_evlist *evlist,
621 union perf_event *event, u64 *id)
623 const u64 *array = event->sample.array;
626 n = (event->header.size - sizeof(event->header)) >> 3;
628 if (event->header.type == PERF_RECORD_SAMPLE) {
629 if (evlist->id_pos >= n)
631 *id = array[evlist->id_pos];
633 if (evlist->is_pos > n)
641 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
642 union perf_event *event)
644 struct perf_evsel *first = perf_evlist__first(evlist);
645 struct hlist_head *head;
646 struct perf_sample_id *sid;
650 if (evlist->nr_entries == 1)
653 if (!first->attr.sample_id_all &&
654 event->header.type != PERF_RECORD_SAMPLE)
657 if (perf_evlist__event2id(evlist, event, &id))
660 /* Synthesized events have an id of zero */
664 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
665 head = &evlist->heads[hash];
667 hlist_for_each_entry(sid, head, node) {
674 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
676 struct perf_mmap *md = &evlist->mmap[idx];
679 unsigned char *data = md->base + page_size;
680 union perf_event *event = NULL;
683 * Check if event was unmapped due to a POLLHUP/POLLERR.
685 if (!atomic_read(&md->refcnt))
688 head = perf_mmap__read_head(md);
689 if (evlist->overwrite) {
691 * If we're further behind than half the buffer, there's a chance
692 * the writer will bite our tail and mess up the samples under us.
694 * If we somehow ended up ahead of the head, we got messed up.
696 * In either case, truncate and restart at head.
698 int diff = head - old;
699 if (diff > md->mask / 2 || diff < 0) {
700 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
703 * head points to a known good entry, start there.
712 event = (union perf_event *)&data[old & md->mask];
713 size = event->header.size;
716 * Event straddles the mmap boundary -- header should always
717 * be inside due to u64 alignment of output.
719 if ((old & md->mask) + size != ((old + size) & md->mask)) {
720 unsigned int offset = old;
721 unsigned int len = min(sizeof(*event), size), cpy;
722 void *dst = md->event_copy;
725 cpy = min(md->mask + 1 - (offset & md->mask), len);
726 memcpy(dst, &data[offset & md->mask], cpy);
732 event = (union perf_event *) md->event_copy;
743 static bool perf_mmap__empty(struct perf_mmap *md)
745 return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
748 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
750 atomic_inc(&evlist->mmap[idx].refcnt);
753 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
755 BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
757 if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
758 __perf_evlist__munmap(evlist, idx);
761 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
763 struct perf_mmap *md = &evlist->mmap[idx];
765 if (!evlist->overwrite) {
768 perf_mmap__write_tail(md, old);
771 if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
772 perf_evlist__mmap_put(evlist, idx);
775 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
776 struct auxtrace_mmap_params *mp __maybe_unused,
777 void *userpg __maybe_unused,
778 int fd __maybe_unused)
783 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
787 void __weak auxtrace_mmap_params__init(
788 struct auxtrace_mmap_params *mp __maybe_unused,
789 off_t auxtrace_offset __maybe_unused,
790 unsigned int auxtrace_pages __maybe_unused,
791 bool auxtrace_overwrite __maybe_unused)
795 void __weak auxtrace_mmap_params__set_idx(
796 struct auxtrace_mmap_params *mp __maybe_unused,
797 struct perf_evlist *evlist __maybe_unused,
798 int idx __maybe_unused,
799 bool per_cpu __maybe_unused)
803 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
805 if (evlist->mmap[idx].base != NULL) {
806 munmap(evlist->mmap[idx].base, evlist->mmap_len);
807 evlist->mmap[idx].base = NULL;
808 atomic_set(&evlist->mmap[idx].refcnt, 0);
810 auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
813 void perf_evlist__munmap(struct perf_evlist *evlist)
817 if (evlist->mmap == NULL)
820 for (i = 0; i < evlist->nr_mmaps; i++)
821 __perf_evlist__munmap(evlist, i);
823 zfree(&evlist->mmap);
826 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
828 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
829 if (cpu_map__empty(evlist->cpus))
830 evlist->nr_mmaps = thread_map__nr(evlist->threads);
831 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
832 return evlist->mmap != NULL ? 0 : -ENOMEM;
838 struct auxtrace_mmap_params auxtrace_mp;
841 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
842 struct mmap_params *mp, int fd)
845 * The last one will be done at perf_evlist__mmap_consume(), so that we
846 * make sure we don't prevent tools from consuming every last event in
849 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
850 * anymore, but the last events for it are still in the ring buffer,
851 * waiting to be consumed.
853 * Tools can chose to ignore this at their own discretion, but the
854 * evlist layer can't just drop it when filtering events in
855 * perf_evlist__filter_pollfd().
857 atomic_set(&evlist->mmap[idx].refcnt, 2);
858 evlist->mmap[idx].prev = 0;
859 evlist->mmap[idx].mask = mp->mask;
860 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
862 if (evlist->mmap[idx].base == MAP_FAILED) {
863 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
865 evlist->mmap[idx].base = NULL;
869 if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
870 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
876 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
877 struct mmap_params *mp, int cpu,
878 int thread, int *output)
880 struct perf_evsel *evsel;
882 evlist__for_each(evlist, evsel) {
885 if (evsel->system_wide && thread)
888 fd = FD(evsel, cpu, thread);
892 if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
895 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
898 perf_evlist__mmap_get(evlist, idx);
902 * The system_wide flag causes a selected event to be opened
903 * always without a pid. Consequently it will never get a
904 * POLLHUP, but it is used for tracking in combination with
905 * other events, so it should not need to be polled anyway.
906 * Therefore don't add it for polling.
908 if (!evsel->system_wide &&
909 __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
910 perf_evlist__mmap_put(evlist, idx);
914 if (evsel->attr.read_format & PERF_FORMAT_ID) {
915 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
918 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
926 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
927 struct mmap_params *mp)
930 int nr_cpus = cpu_map__nr(evlist->cpus);
931 int nr_threads = thread_map__nr(evlist->threads);
933 pr_debug2("perf event ring buffer mmapped per cpu\n");
934 for (cpu = 0; cpu < nr_cpus; cpu++) {
937 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
940 for (thread = 0; thread < nr_threads; thread++) {
941 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
950 for (cpu = 0; cpu < nr_cpus; cpu++)
951 __perf_evlist__munmap(evlist, cpu);
955 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
956 struct mmap_params *mp)
959 int nr_threads = thread_map__nr(evlist->threads);
961 pr_debug2("perf event ring buffer mmapped per thread\n");
962 for (thread = 0; thread < nr_threads; thread++) {
965 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
968 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
976 for (thread = 0; thread < nr_threads; thread++)
977 __perf_evlist__munmap(evlist, thread);
981 static size_t perf_evlist__mmap_size(unsigned long pages)
983 if (pages == UINT_MAX) {
986 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
988 * Pick a once upon a time good value, i.e. things look
989 * strange since we can't read a sysctl value, but lets not
994 max -= (page_size / 1024);
997 pages = (max * 1024) / page_size;
998 if (!is_power_of_2(pages))
999 pages = rounddown_pow_of_two(pages);
1000 } else if (!is_power_of_2(pages))
1003 return (pages + 1) * page_size;
1006 static long parse_pages_arg(const char *str, unsigned long min,
1009 unsigned long pages, val;
1010 static struct parse_tag tags[] = {
1011 { .tag = 'B', .mult = 1 },
1012 { .tag = 'K', .mult = 1 << 10 },
1013 { .tag = 'M', .mult = 1 << 20 },
1014 { .tag = 'G', .mult = 1 << 30 },
1021 val = parse_tag_value(str, tags);
1022 if (val != (unsigned long) -1) {
1023 /* we got file size value */
1024 pages = PERF_ALIGN(val, page_size) / page_size;
1026 /* we got pages count value */
1028 pages = strtoul(str, &eptr, 10);
1033 if (pages == 0 && min == 0) {
1034 /* leave number of pages at 0 */
1035 } else if (!is_power_of_2(pages)) {
1036 /* round pages up to next power of 2 */
1037 pages = roundup_pow_of_two(pages);
1040 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1041 pages * page_size, pages);
1050 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1052 unsigned long max = UINT_MAX;
1055 if (max > SIZE_MAX / page_size)
1056 max = SIZE_MAX / page_size;
1058 pages = parse_pages_arg(str, 1, max);
1060 pr_err("Invalid argument for --mmap_pages/-m\n");
1064 *mmap_pages = pages;
1068 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1069 int unset __maybe_unused)
1071 return __perf_evlist__parse_mmap_pages(opt->value, str);
1075 * perf_evlist__mmap_ex - Create mmaps to receive events.
1076 * @evlist: list of events
1077 * @pages: map length in pages
1078 * @overwrite: overwrite older events?
1079 * @auxtrace_pages - auxtrace map length in pages
1080 * @auxtrace_overwrite - overwrite older auxtrace data?
1082 * If @overwrite is %false the user needs to signal event consumption using
1083 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1086 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1087 * consumption using auxtrace_mmap__write_tail().
1089 * Return: %0 on success, negative error code otherwise.
1091 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1092 bool overwrite, unsigned int auxtrace_pages,
1093 bool auxtrace_overwrite)
1095 struct perf_evsel *evsel;
1096 const struct cpu_map *cpus = evlist->cpus;
1097 const struct thread_map *threads = evlist->threads;
1098 struct mmap_params mp = {
1099 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1102 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1105 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1108 evlist->overwrite = overwrite;
1109 evlist->mmap_len = perf_evlist__mmap_size(pages);
1110 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1111 mp.mask = evlist->mmap_len - page_size - 1;
1113 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1114 auxtrace_pages, auxtrace_overwrite);
1116 evlist__for_each(evlist, evsel) {
1117 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1118 evsel->sample_id == NULL &&
1119 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1123 if (cpu_map__empty(cpus))
1124 return perf_evlist__mmap_per_thread(evlist, &mp);
1126 return perf_evlist__mmap_per_cpu(evlist, &mp);
1129 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1132 return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1135 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1137 struct cpu_map *cpus;
1138 struct thread_map *threads;
1140 threads = thread_map__new_str(target->pid, target->tid, target->uid);
1145 if (target__uses_dummy_map(target))
1146 cpus = cpu_map__dummy_new();
1148 cpus = cpu_map__new(target->cpu_list);
1151 goto out_delete_threads;
1153 evlist->has_user_cpus = !!target->cpu_list;
1155 perf_evlist__set_maps(evlist, cpus, threads);
1160 thread_map__put(threads);
1164 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1165 struct thread_map *threads)
1168 * Allow for the possibility that one or another of the maps isn't being
1169 * changed i.e. don't put it. Note we are assuming the maps that are
1170 * being applied are brand new and evlist is taking ownership of the
1171 * original reference count of 1. If that is not the case it is up to
1172 * the caller to increase the reference count.
1174 if (cpus != evlist->cpus) {
1175 cpu_map__put(evlist->cpus);
1176 evlist->cpus = cpus;
1179 if (threads != evlist->threads) {
1180 thread_map__put(evlist->threads);
1181 evlist->threads = threads;
1184 perf_evlist__propagate_maps(evlist);
1187 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1189 struct perf_evsel *evsel;
1191 const int ncpus = cpu_map__nr(evlist->cpus),
1192 nthreads = thread_map__nr(evlist->threads);
1194 evlist__for_each(evlist, evsel) {
1195 if (evsel->filter == NULL)
1199 * filters only work for tracepoint event, which doesn't have cpu limit.
1200 * So evlist and evsel should always be same.
1202 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1212 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1214 struct perf_evsel *evsel;
1217 evlist__for_each(evlist, evsel) {
1218 err = perf_evsel__set_filter(evsel, filter);
1226 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1232 for (i = 0; i < npids; ++i) {
1234 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1239 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1247 ret = perf_evlist__set_filter(evlist, filter);
1253 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1255 return perf_evlist__set_filter_pids(evlist, 1, &pid);
1258 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1260 struct perf_evsel *pos;
1262 if (evlist->nr_entries == 1)
1265 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1268 evlist__for_each(evlist, pos) {
1269 if (pos->id_pos != evlist->id_pos ||
1270 pos->is_pos != evlist->is_pos)
1277 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1279 struct perf_evsel *evsel;
1281 if (evlist->combined_sample_type)
1282 return evlist->combined_sample_type;
1284 evlist__for_each(evlist, evsel)
1285 evlist->combined_sample_type |= evsel->attr.sample_type;
1287 return evlist->combined_sample_type;
1290 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1292 evlist->combined_sample_type = 0;
1293 return __perf_evlist__combined_sample_type(evlist);
1296 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1298 struct perf_evsel *evsel;
1299 u64 branch_type = 0;
1301 evlist__for_each(evlist, evsel)
1302 branch_type |= evsel->attr.branch_sample_type;
1306 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1308 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1309 u64 read_format = first->attr.read_format;
1310 u64 sample_type = first->attr.sample_type;
1312 evlist__for_each(evlist, pos) {
1313 if (read_format != pos->attr.read_format)
1317 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1318 if ((sample_type & PERF_SAMPLE_READ) &&
1319 !(read_format & PERF_FORMAT_ID)) {
1326 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1328 struct perf_evsel *first = perf_evlist__first(evlist);
1329 return first->attr.read_format;
1332 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1334 struct perf_evsel *first = perf_evlist__first(evlist);
1335 struct perf_sample *data;
1339 if (!first->attr.sample_id_all)
1342 sample_type = first->attr.sample_type;
1344 if (sample_type & PERF_SAMPLE_TID)
1345 size += sizeof(data->tid) * 2;
1347 if (sample_type & PERF_SAMPLE_TIME)
1348 size += sizeof(data->time);
1350 if (sample_type & PERF_SAMPLE_ID)
1351 size += sizeof(data->id);
1353 if (sample_type & PERF_SAMPLE_STREAM_ID)
1354 size += sizeof(data->stream_id);
1356 if (sample_type & PERF_SAMPLE_CPU)
1357 size += sizeof(data->cpu) * 2;
1359 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1360 size += sizeof(data->id);
1365 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1367 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1369 evlist__for_each_continue(evlist, pos) {
1370 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1377 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1379 struct perf_evsel *first = perf_evlist__first(evlist);
1380 return first->attr.sample_id_all;
1383 void perf_evlist__set_selected(struct perf_evlist *evlist,
1384 struct perf_evsel *evsel)
1386 evlist->selected = evsel;
1389 void perf_evlist__close(struct perf_evlist *evlist)
1391 struct perf_evsel *evsel;
1392 int ncpus = cpu_map__nr(evlist->cpus);
1393 int nthreads = thread_map__nr(evlist->threads);
1396 evlist__for_each_reverse(evlist, evsel) {
1397 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1398 perf_evsel__close(evsel, n, nthreads);
1402 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1404 struct cpu_map *cpus;
1405 struct thread_map *threads;
1409 * Try reading /sys/devices/system/cpu/online to get
1412 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1413 * code needs an overhaul to properly forward the
1414 * error, and we may not want to do that fallback to a
1415 * default cpu identity map :-\
1417 cpus = cpu_map__new(NULL);
1421 threads = thread_map__new_dummy();
1425 perf_evlist__set_maps(evlist, cpus, threads);
1433 int perf_evlist__open(struct perf_evlist *evlist)
1435 struct perf_evsel *evsel;
1439 * Default: one fd per CPU, all threads, aka systemwide
1440 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1442 if (evlist->threads == NULL && evlist->cpus == NULL) {
1443 err = perf_evlist__create_syswide_maps(evlist);
1448 perf_evlist__update_id_pos(evlist);
1450 evlist__for_each(evlist, evsel) {
1451 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1458 perf_evlist__close(evlist);
1463 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1464 const char *argv[], bool pipe_output,
1465 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1467 int child_ready_pipe[2], go_pipe[2];
1470 if (pipe(child_ready_pipe) < 0) {
1471 perror("failed to create 'ready' pipe");
1475 if (pipe(go_pipe) < 0) {
1476 perror("failed to create 'go' pipe");
1477 goto out_close_ready_pipe;
1480 evlist->workload.pid = fork();
1481 if (evlist->workload.pid < 0) {
1482 perror("failed to fork");
1483 goto out_close_pipes;
1486 if (!evlist->workload.pid) {
1492 signal(SIGTERM, SIG_DFL);
1494 close(child_ready_pipe[0]);
1496 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1499 * Tell the parent we're ready to go
1501 close(child_ready_pipe[1]);
1504 * Wait until the parent tells us to go.
1506 ret = read(go_pipe[0], &bf, 1);
1508 * The parent will ask for the execvp() to be performed by
1509 * writing exactly one byte, in workload.cork_fd, usually via
1510 * perf_evlist__start_workload().
1512 * For cancelling the workload without actually running it,
1513 * the parent will just close workload.cork_fd, without writing
1514 * anything, i.e. read will return zero and we just exit()
1519 perror("unable to read pipe");
1523 execvp(argv[0], (char **)argv);
1528 val.sival_int = errno;
1529 if (sigqueue(getppid(), SIGUSR1, val))
1537 struct sigaction act = {
1538 .sa_flags = SA_SIGINFO,
1539 .sa_sigaction = exec_error,
1541 sigaction(SIGUSR1, &act, NULL);
1544 if (target__none(target)) {
1545 if (evlist->threads == NULL) {
1546 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1547 __func__, __LINE__);
1548 goto out_close_pipes;
1550 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1553 close(child_ready_pipe[1]);
1556 * wait for child to settle
1558 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1559 perror("unable to read pipe");
1560 goto out_close_pipes;
1563 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1564 evlist->workload.cork_fd = go_pipe[1];
1565 close(child_ready_pipe[0]);
1571 out_close_ready_pipe:
1572 close(child_ready_pipe[0]);
1573 close(child_ready_pipe[1]);
1577 int perf_evlist__start_workload(struct perf_evlist *evlist)
1579 if (evlist->workload.cork_fd > 0) {
1583 * Remove the cork, let it rip!
1585 ret = write(evlist->workload.cork_fd, &bf, 1);
1587 perror("enable to write to pipe");
1589 close(evlist->workload.cork_fd);
1596 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1597 struct perf_sample *sample)
1599 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1603 return perf_evsel__parse_sample(evsel, event, sample);
1606 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1608 struct perf_evsel *evsel;
1611 evlist__for_each(evlist, evsel) {
1612 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1613 perf_evsel__name(evsel));
1616 return printed + fprintf(fp, "\n");
1619 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1620 int err, char *buf, size_t size)
1623 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1628 printed = scnprintf(buf, size,
1630 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1632 value = perf_event_paranoid();
1634 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1637 printed += scnprintf(buf + printed, size - printed,
1638 "For your workloads it needs to be <= 1\nHint:\t");
1640 printed += scnprintf(buf + printed, size - printed,
1641 "For system wide tracing it needs to be set to -1.\n");
1643 printed += scnprintf(buf + printed, size - printed,
1644 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1645 "Hint:\tThe current value is %d.", value);
1648 scnprintf(buf, size, "%s", emsg);
1655 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1657 char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1658 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1662 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1663 printed += scnprintf(buf + printed, size - printed,
1665 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1666 "Hint:\tTried using %zd kB.\n",
1667 emsg, pages_max_per_user, pages_attempted);
1669 if (pages_attempted >= pages_max_per_user) {
1670 printed += scnprintf(buf + printed, size - printed,
1671 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1672 pages_max_per_user + pages_attempted);
1675 printed += scnprintf(buf + printed, size - printed,
1676 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1679 scnprintf(buf, size, "%s", emsg);
1686 void perf_evlist__to_front(struct perf_evlist *evlist,
1687 struct perf_evsel *move_evsel)
1689 struct perf_evsel *evsel, *n;
1692 if (move_evsel == perf_evlist__first(evlist))
1695 evlist__for_each_safe(evlist, n, evsel) {
1696 if (evsel->leader == move_evsel->leader)
1697 list_move_tail(&evsel->node, &move);
1700 list_splice(&move, &evlist->entries);
1703 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1704 struct perf_evsel *tracking_evsel)
1706 struct perf_evsel *evsel;
1708 if (tracking_evsel->tracking)
1711 evlist__for_each(evlist, evsel) {
1712 if (evsel != tracking_evsel)
1713 evsel->tracking = false;
1716 tracking_evsel->tracking = true;