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>
15 #include "thread_map.h"
22 #include "bpf-event.h"
26 #include "parse-events.h"
27 #include <subcmd/parse-options.h>
30 #include <sys/ioctl.h>
33 #include <linux/bitops.h>
34 #include <linux/hash.h>
35 #include <linux/log2.h>
36 #include <linux/err.h>
38 #ifdef LACKS_SIGQUEUE_PROTOTYPE
39 int sigqueue(pid_t pid, int sig, const union sigval value);
42 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
43 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
45 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
46 struct thread_map *threads)
50 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
51 INIT_HLIST_HEAD(&evlist->heads[i]);
52 INIT_LIST_HEAD(&evlist->entries);
53 perf_evlist__set_maps(evlist, cpus, threads);
54 fdarray__init(&evlist->pollfd, 64);
55 evlist->workload.pid = -1;
56 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
59 struct perf_evlist *perf_evlist__new(void)
61 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
64 perf_evlist__init(evlist, NULL, NULL);
69 struct perf_evlist *perf_evlist__new_default(void)
71 struct perf_evlist *evlist = perf_evlist__new();
73 if (evlist && perf_evlist__add_default(evlist)) {
74 perf_evlist__delete(evlist);
81 struct perf_evlist *perf_evlist__new_dummy(void)
83 struct perf_evlist *evlist = perf_evlist__new();
85 if (evlist && perf_evlist__add_dummy(evlist)) {
86 perf_evlist__delete(evlist);
94 * perf_evlist__set_id_pos - set the positions of event ids.
95 * @evlist: selected event list
97 * Events with compatible sample types all have the same id_pos
98 * and is_pos. For convenience, put a copy on evlist.
100 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
102 struct perf_evsel *first = perf_evlist__first(evlist);
104 evlist->id_pos = first->id_pos;
105 evlist->is_pos = first->is_pos;
108 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
110 struct perf_evsel *evsel;
112 evlist__for_each_entry(evlist, evsel)
113 perf_evsel__calc_id_pos(evsel);
115 perf_evlist__set_id_pos(evlist);
118 static void perf_evlist__purge(struct perf_evlist *evlist)
120 struct perf_evsel *pos, *n;
122 evlist__for_each_entry_safe(evlist, n, pos) {
123 list_del_init(&pos->node);
125 perf_evsel__delete(pos);
128 evlist->nr_entries = 0;
131 void perf_evlist__exit(struct perf_evlist *evlist)
133 zfree(&evlist->mmap);
134 zfree(&evlist->overwrite_mmap);
135 fdarray__exit(&evlist->pollfd);
138 void perf_evlist__delete(struct perf_evlist *evlist)
143 perf_evlist__munmap(evlist);
144 perf_evlist__close(evlist);
145 cpu_map__put(evlist->cpus);
146 thread_map__put(evlist->threads);
148 evlist->threads = NULL;
149 perf_evlist__purge(evlist);
150 perf_evlist__exit(evlist);
154 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
155 struct perf_evsel *evsel)
158 * We already have cpus for evsel (via PMU sysfs) so
159 * keep it, if there's no target cpu list defined.
161 if (!evsel->own_cpus || evlist->has_user_cpus) {
162 cpu_map__put(evsel->cpus);
163 evsel->cpus = cpu_map__get(evlist->cpus);
164 } else if (evsel->cpus != evsel->own_cpus) {
165 cpu_map__put(evsel->cpus);
166 evsel->cpus = cpu_map__get(evsel->own_cpus);
169 thread_map__put(evsel->threads);
170 evsel->threads = thread_map__get(evlist->threads);
173 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
175 struct perf_evsel *evsel;
177 evlist__for_each_entry(evlist, evsel)
178 __perf_evlist__propagate_maps(evlist, evsel);
181 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
183 entry->evlist = evlist;
184 list_add_tail(&entry->node, &evlist->entries);
185 entry->idx = evlist->nr_entries;
186 entry->tracking = !entry->idx;
188 if (!evlist->nr_entries++)
189 perf_evlist__set_id_pos(evlist);
191 __perf_evlist__propagate_maps(evlist, entry);
194 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
196 evsel->evlist = NULL;
197 list_del_init(&evsel->node);
198 evlist->nr_entries -= 1;
201 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
202 struct list_head *list)
204 struct perf_evsel *evsel, *temp;
206 __evlist__for_each_entry_safe(list, temp, evsel) {
207 list_del_init(&evsel->node);
208 perf_evlist__add(evlist, evsel);
212 void __perf_evlist__set_leader(struct list_head *list)
214 struct perf_evsel *evsel, *leader;
216 leader = list_entry(list->next, struct perf_evsel, node);
217 evsel = list_entry(list->prev, struct perf_evsel, node);
219 leader->nr_members = evsel->idx - leader->idx + 1;
221 __evlist__for_each_entry(list, evsel) {
222 evsel->leader = leader;
226 void perf_evlist__set_leader(struct perf_evlist *evlist)
228 if (evlist->nr_entries) {
229 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
230 __perf_evlist__set_leader(&evlist->entries);
234 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
236 struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
241 perf_evlist__add(evlist, evsel);
245 int perf_evlist__add_dummy(struct perf_evlist *evlist)
247 struct perf_event_attr attr = {
248 .type = PERF_TYPE_SOFTWARE,
249 .config = PERF_COUNT_SW_DUMMY,
250 .size = sizeof(attr), /* to capture ABI version */
252 struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
257 perf_evlist__add(evlist, evsel);
261 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
262 struct perf_event_attr *attrs, size_t nr_attrs)
264 struct perf_evsel *evsel, *n;
268 for (i = 0; i < nr_attrs; i++) {
269 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
271 goto out_delete_partial_list;
272 list_add_tail(&evsel->node, &head);
275 perf_evlist__splice_list_tail(evlist, &head);
279 out_delete_partial_list:
280 __evlist__for_each_entry_safe(&head, n, evsel)
281 perf_evsel__delete(evsel);
285 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
286 struct perf_event_attr *attrs, size_t nr_attrs)
290 for (i = 0; i < nr_attrs; i++)
291 event_attr_init(attrs + i);
293 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
297 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
299 struct perf_evsel *evsel;
301 evlist__for_each_entry(evlist, evsel) {
302 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
303 (int)evsel->attr.config == id)
311 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
314 struct perf_evsel *evsel;
316 evlist__for_each_entry(evlist, evsel) {
317 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
318 (strcmp(evsel->name, name) == 0))
325 int perf_evlist__add_newtp(struct perf_evlist *evlist,
326 const char *sys, const char *name, void *handler)
328 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
333 evsel->handler = handler;
334 perf_evlist__add(evlist, evsel);
338 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
339 struct perf_evsel *evsel)
341 if (evsel->system_wide)
344 return thread_map__nr(evlist->threads);
347 void perf_evlist__disable(struct perf_evlist *evlist)
349 struct perf_evsel *pos;
351 evlist__for_each_entry(evlist, pos) {
352 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
354 perf_evsel__disable(pos);
357 evlist->enabled = false;
360 void perf_evlist__enable(struct perf_evlist *evlist)
362 struct perf_evsel *pos;
364 evlist__for_each_entry(evlist, pos) {
365 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
367 perf_evsel__enable(pos);
370 evlist->enabled = true;
373 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
375 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
378 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
379 struct perf_evsel *evsel, int cpu)
382 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
387 for (thread = 0; thread < nr_threads; thread++) {
388 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
395 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
396 struct perf_evsel *evsel,
400 int nr_cpus = cpu_map__nr(evlist->cpus);
405 for (cpu = 0; cpu < nr_cpus; cpu++) {
406 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
413 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
414 struct perf_evsel *evsel, int idx)
416 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
419 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
421 return perf_evlist__enable_event_thread(evlist, evsel, idx);
424 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
426 int nr_cpus = cpu_map__nr(evlist->cpus);
427 int nr_threads = thread_map__nr(evlist->threads);
429 struct perf_evsel *evsel;
431 evlist__for_each_entry(evlist, evsel) {
432 if (evsel->system_wide)
435 nfds += nr_cpus * nr_threads;
438 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
439 fdarray__grow(&evlist->pollfd, nfds) < 0)
445 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
446 struct perf_mmap *map, short revent)
448 int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
450 * Save the idx so that when we filter out fds POLLHUP'ed we can
451 * close the associated evlist->mmap[] entry.
454 evlist->pollfd.priv[pos].ptr = map;
456 fcntl(fd, F_SETFL, O_NONBLOCK);
462 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
464 return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
467 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
468 void *arg __maybe_unused)
470 struct perf_mmap *map = fda->priv[fd].ptr;
476 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
478 return fdarray__filter(&evlist->pollfd, revents_and_mask,
479 perf_evlist__munmap_filtered, NULL);
482 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
484 return fdarray__poll(&evlist->pollfd, timeout);
487 static void perf_evlist__id_hash(struct perf_evlist *evlist,
488 struct perf_evsel *evsel,
489 int cpu, int thread, u64 id)
492 struct perf_sample_id *sid = SID(evsel, cpu, thread);
496 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
497 hlist_add_head(&sid->node, &evlist->heads[hash]);
500 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
501 int cpu, int thread, u64 id)
503 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
504 evsel->id[evsel->ids++] = id;
507 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
508 struct perf_evsel *evsel,
509 int cpu, int thread, int fd)
511 u64 read_data[4] = { 0, };
512 int id_idx = 1; /* The first entry is the counter value */
516 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
523 /* Legacy way to get event id.. All hail to old kernels! */
526 * This way does not work with group format read, so bail
529 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
532 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
533 read(fd, &read_data, sizeof(read_data)) == -1)
536 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
538 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
541 id = read_data[id_idx];
544 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
548 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
549 struct perf_evsel *evsel, int idx, int cpu,
552 struct perf_sample_id *sid = SID(evsel, cpu, thread);
554 if (evlist->cpus && cpu >= 0)
555 sid->cpu = evlist->cpus->map[cpu];
558 if (!evsel->system_wide && evlist->threads && thread >= 0)
559 sid->tid = thread_map__pid(evlist->threads, thread);
564 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
566 struct hlist_head *head;
567 struct perf_sample_id *sid;
570 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
571 head = &evlist->heads[hash];
573 hlist_for_each_entry(sid, head, node)
580 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
582 struct perf_sample_id *sid;
584 if (evlist->nr_entries == 1 || !id)
585 return perf_evlist__first(evlist);
587 sid = perf_evlist__id2sid(evlist, id);
591 if (!perf_evlist__sample_id_all(evlist))
592 return perf_evlist__first(evlist);
597 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
600 struct perf_sample_id *sid;
605 sid = perf_evlist__id2sid(evlist, id);
612 static int perf_evlist__event2id(struct perf_evlist *evlist,
613 union perf_event *event, u64 *id)
615 const u64 *array = event->sample.array;
618 n = (event->header.size - sizeof(event->header)) >> 3;
620 if (event->header.type == PERF_RECORD_SAMPLE) {
621 if (evlist->id_pos >= n)
623 *id = array[evlist->id_pos];
625 if (evlist->is_pos > n)
633 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
634 union perf_event *event)
636 struct perf_evsel *first = perf_evlist__first(evlist);
637 struct hlist_head *head;
638 struct perf_sample_id *sid;
642 if (evlist->nr_entries == 1)
645 if (!first->attr.sample_id_all &&
646 event->header.type != PERF_RECORD_SAMPLE)
649 if (perf_evlist__event2id(evlist, event, &id))
652 /* Synthesized events have an id of zero */
656 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
657 head = &evlist->heads[hash];
659 hlist_for_each_entry(sid, head, node) {
666 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
670 if (!evlist->overwrite_mmap)
673 for (i = 0; i < evlist->nr_mmaps; i++) {
674 int fd = evlist->overwrite_mmap[i].fd;
679 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
686 static int perf_evlist__pause(struct perf_evlist *evlist)
688 return perf_evlist__set_paused(evlist, true);
691 static int perf_evlist__resume(struct perf_evlist *evlist)
693 return perf_evlist__set_paused(evlist, false);
696 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
701 for (i = 0; i < evlist->nr_mmaps; i++)
702 perf_mmap__munmap(&evlist->mmap[i]);
704 if (evlist->overwrite_mmap)
705 for (i = 0; i < evlist->nr_mmaps; i++)
706 perf_mmap__munmap(&evlist->overwrite_mmap[i]);
709 void perf_evlist__munmap(struct perf_evlist *evlist)
711 perf_evlist__munmap_nofree(evlist);
712 zfree(&evlist->mmap);
713 zfree(&evlist->overwrite_mmap);
716 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
720 struct perf_mmap *map;
722 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
723 if (cpu_map__empty(evlist->cpus))
724 evlist->nr_mmaps = thread_map__nr(evlist->threads);
725 map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
729 for (i = 0; i < evlist->nr_mmaps; i++) {
731 map[i].overwrite = overwrite;
733 * When the perf_mmap() call is made we grab one refcount, plus
734 * one extra to let perf_mmap__consume() get the last
735 * events after all real references (perf_mmap__get()) are
738 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
739 * thus does perf_mmap__get() on it.
741 refcount_set(&map[i].refcnt, 0);
747 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
748 struct perf_evsel *evsel)
750 if (evsel->attr.write_backward)
755 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
756 struct mmap_params *mp, int cpu_idx,
757 int thread, int *_output, int *_output_overwrite)
759 struct perf_evsel *evsel;
761 int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
763 evlist__for_each_entry(evlist, evsel) {
764 struct perf_mmap *maps = evlist->mmap;
765 int *output = _output;
769 mp->prot = PROT_READ | PROT_WRITE;
770 if (evsel->attr.write_backward) {
771 output = _output_overwrite;
772 maps = evlist->overwrite_mmap;
775 maps = perf_evlist__alloc_mmap(evlist, true);
778 evlist->overwrite_mmap = maps;
779 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
780 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
782 mp->prot &= ~PROT_WRITE;
785 if (evsel->system_wide && thread)
788 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
792 fd = FD(evsel, cpu, thread);
797 if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
800 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
803 perf_mmap__get(&maps[idx]);
806 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
809 * The system_wide flag causes a selected event to be opened
810 * always without a pid. Consequently it will never get a
811 * POLLHUP, but it is used for tracking in combination with
812 * other events, so it should not need to be polled anyway.
813 * Therefore don't add it for polling.
815 if (!evsel->system_wide &&
816 __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
817 perf_mmap__put(&maps[idx]);
821 if (evsel->attr.read_format & PERF_FORMAT_ID) {
822 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
825 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
833 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
834 struct mmap_params *mp)
837 int nr_cpus = cpu_map__nr(evlist->cpus);
838 int nr_threads = thread_map__nr(evlist->threads);
840 pr_debug2("perf event ring buffer mmapped per cpu\n");
841 for (cpu = 0; cpu < nr_cpus; cpu++) {
843 int output_overwrite = -1;
845 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
848 for (thread = 0; thread < nr_threads; thread++) {
849 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
850 thread, &output, &output_overwrite))
858 perf_evlist__munmap_nofree(evlist);
862 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
863 struct mmap_params *mp)
866 int nr_threads = thread_map__nr(evlist->threads);
868 pr_debug2("perf event ring buffer mmapped per thread\n");
869 for (thread = 0; thread < nr_threads; thread++) {
871 int output_overwrite = -1;
873 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
876 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
877 &output, &output_overwrite))
884 perf_evlist__munmap_nofree(evlist);
888 unsigned long perf_event_mlock_kb_in_pages(void)
893 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
895 * Pick a once upon a time good value, i.e. things look
896 * strange since we can't read a sysctl value, but lets not
901 max -= (page_size / 1024);
904 pages = (max * 1024) / page_size;
905 if (!is_power_of_2(pages))
906 pages = rounddown_pow_of_two(pages);
911 size_t perf_evlist__mmap_size(unsigned long pages)
913 if (pages == UINT_MAX)
914 pages = perf_event_mlock_kb_in_pages();
915 else if (!is_power_of_2(pages))
918 return (pages + 1) * page_size;
921 static long parse_pages_arg(const char *str, unsigned long min,
924 unsigned long pages, val;
925 static struct parse_tag tags[] = {
926 { .tag = 'B', .mult = 1 },
927 { .tag = 'K', .mult = 1 << 10 },
928 { .tag = 'M', .mult = 1 << 20 },
929 { .tag = 'G', .mult = 1 << 30 },
936 val = parse_tag_value(str, tags);
937 if (val != (unsigned long) -1) {
938 /* we got file size value */
939 pages = PERF_ALIGN(val, page_size) / page_size;
941 /* we got pages count value */
943 pages = strtoul(str, &eptr, 10);
948 if (pages == 0 && min == 0) {
949 /* leave number of pages at 0 */
950 } else if (!is_power_of_2(pages)) {
953 /* round pages up to next power of 2 */
954 pages = roundup_pow_of_two(pages);
958 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
959 pr_info("rounding mmap pages size to %s (%lu pages)\n",
969 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
971 unsigned long max = UINT_MAX;
974 if (max > SIZE_MAX / page_size)
975 max = SIZE_MAX / page_size;
977 pages = parse_pages_arg(str, 1, max);
979 pr_err("Invalid argument for --mmap_pages/-m\n");
987 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
988 int unset __maybe_unused)
990 return __perf_evlist__parse_mmap_pages(opt->value, str);
994 * perf_evlist__mmap_ex - Create mmaps to receive events.
995 * @evlist: list of events
996 * @pages: map length in pages
997 * @overwrite: overwrite older events?
998 * @auxtrace_pages - auxtrace map length in pages
999 * @auxtrace_overwrite - overwrite older auxtrace data?
1001 * If @overwrite is %false the user needs to signal event consumption using
1002 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1005 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1006 * consumption using auxtrace_mmap__write_tail().
1008 * Return: %0 on success, negative error code otherwise.
1010 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1011 unsigned int auxtrace_pages,
1012 bool auxtrace_overwrite, int nr_cblocks, int affinity)
1014 struct perf_evsel *evsel;
1015 const struct cpu_map *cpus = evlist->cpus;
1016 const struct thread_map *threads = evlist->threads;
1018 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1019 * Its value is decided by evsel's write_backward.
1020 * So &mp should not be passed through const pointer.
1022 struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity };
1025 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1029 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1032 evlist->mmap_len = perf_evlist__mmap_size(pages);
1033 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1034 mp.mask = evlist->mmap_len - page_size - 1;
1036 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1037 auxtrace_pages, auxtrace_overwrite);
1039 evlist__for_each_entry(evlist, evsel) {
1040 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1041 evsel->sample_id == NULL &&
1042 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1046 if (cpu_map__empty(cpus))
1047 return perf_evlist__mmap_per_thread(evlist, &mp);
1049 return perf_evlist__mmap_per_cpu(evlist, &mp);
1052 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1054 return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS);
1057 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1059 bool all_threads = (target->per_thread && target->system_wide);
1060 struct cpu_map *cpus;
1061 struct thread_map *threads;
1064 * If specify '-a' and '--per-thread' to perf record, perf record
1065 * will override '--per-thread'. target->per_thread = false and
1066 * target->system_wide = true.
1068 * If specify '--per-thread' only to perf record,
1069 * target->per_thread = true and target->system_wide = false.
1071 * So target->per_thread && target->system_wide is false.
1072 * For perf record, thread_map__new_str doesn't call
1073 * thread_map__new_all_cpus. That will keep perf record's
1076 * For perf stat, it allows the case that target->per_thread and
1077 * target->system_wide are all true. It means to collect system-wide
1078 * per-thread data. thread_map__new_str will call
1079 * thread_map__new_all_cpus to enumerate all threads.
1081 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1087 if (target__uses_dummy_map(target))
1088 cpus = cpu_map__dummy_new();
1090 cpus = cpu_map__new(target->cpu_list);
1093 goto out_delete_threads;
1095 evlist->has_user_cpus = !!target->cpu_list;
1097 perf_evlist__set_maps(evlist, cpus, threads);
1102 thread_map__put(threads);
1106 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1107 struct thread_map *threads)
1110 * Allow for the possibility that one or another of the maps isn't being
1111 * changed i.e. don't put it. Note we are assuming the maps that are
1112 * being applied are brand new and evlist is taking ownership of the
1113 * original reference count of 1. If that is not the case it is up to
1114 * the caller to increase the reference count.
1116 if (cpus != evlist->cpus) {
1117 cpu_map__put(evlist->cpus);
1118 evlist->cpus = cpu_map__get(cpus);
1121 if (threads != evlist->threads) {
1122 thread_map__put(evlist->threads);
1123 evlist->threads = thread_map__get(threads);
1126 perf_evlist__propagate_maps(evlist);
1129 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1130 enum perf_event_sample_format bit)
1132 struct perf_evsel *evsel;
1134 evlist__for_each_entry(evlist, evsel)
1135 __perf_evsel__set_sample_bit(evsel, bit);
1138 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1139 enum perf_event_sample_format bit)
1141 struct perf_evsel *evsel;
1143 evlist__for_each_entry(evlist, evsel)
1144 __perf_evsel__reset_sample_bit(evsel, bit);
1147 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1149 struct perf_evsel *evsel;
1152 evlist__for_each_entry(evlist, evsel) {
1153 if (evsel->filter == NULL)
1157 * filters only work for tracepoint event, which doesn't have cpu limit.
1158 * So evlist and evsel should always be same.
1160 err = perf_evsel__apply_filter(evsel, evsel->filter);
1170 int perf_evlist__set_tp_filter(struct perf_evlist *evlist, const char *filter)
1172 struct perf_evsel *evsel;
1175 evlist__for_each_entry(evlist, evsel) {
1176 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1179 err = perf_evsel__set_filter(evsel, filter);
1187 int perf_evlist__set_tp_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1193 for (i = 0; i < npids; ++i) {
1195 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1200 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1208 ret = perf_evlist__set_tp_filter(evlist, filter);
1214 int perf_evlist__set_tp_filter_pid(struct perf_evlist *evlist, pid_t pid)
1216 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1219 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1221 struct perf_evsel *pos;
1223 if (evlist->nr_entries == 1)
1226 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1229 evlist__for_each_entry(evlist, pos) {
1230 if (pos->id_pos != evlist->id_pos ||
1231 pos->is_pos != evlist->is_pos)
1238 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1240 struct perf_evsel *evsel;
1242 if (evlist->combined_sample_type)
1243 return evlist->combined_sample_type;
1245 evlist__for_each_entry(evlist, evsel)
1246 evlist->combined_sample_type |= evsel->attr.sample_type;
1248 return evlist->combined_sample_type;
1251 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1253 evlist->combined_sample_type = 0;
1254 return __perf_evlist__combined_sample_type(evlist);
1257 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1259 struct perf_evsel *evsel;
1260 u64 branch_type = 0;
1262 evlist__for_each_entry(evlist, evsel)
1263 branch_type |= evsel->attr.branch_sample_type;
1267 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1269 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1270 u64 read_format = first->attr.read_format;
1271 u64 sample_type = first->attr.sample_type;
1273 evlist__for_each_entry(evlist, pos) {
1274 if (read_format != pos->attr.read_format)
1278 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1279 if ((sample_type & PERF_SAMPLE_READ) &&
1280 !(read_format & PERF_FORMAT_ID)) {
1287 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1289 struct perf_evsel *first = perf_evlist__first(evlist);
1290 return first->attr.read_format;
1293 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1295 struct perf_evsel *first = perf_evlist__first(evlist);
1296 struct perf_sample *data;
1300 if (!first->attr.sample_id_all)
1303 sample_type = first->attr.sample_type;
1305 if (sample_type & PERF_SAMPLE_TID)
1306 size += sizeof(data->tid) * 2;
1308 if (sample_type & PERF_SAMPLE_TIME)
1309 size += sizeof(data->time);
1311 if (sample_type & PERF_SAMPLE_ID)
1312 size += sizeof(data->id);
1314 if (sample_type & PERF_SAMPLE_STREAM_ID)
1315 size += sizeof(data->stream_id);
1317 if (sample_type & PERF_SAMPLE_CPU)
1318 size += sizeof(data->cpu) * 2;
1320 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1321 size += sizeof(data->id);
1326 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1328 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1330 evlist__for_each_entry_continue(evlist, pos) {
1331 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1338 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1340 struct perf_evsel *first = perf_evlist__first(evlist);
1341 return first->attr.sample_id_all;
1344 void perf_evlist__set_selected(struct perf_evlist *evlist,
1345 struct perf_evsel *evsel)
1347 evlist->selected = evsel;
1350 void perf_evlist__close(struct perf_evlist *evlist)
1352 struct perf_evsel *evsel;
1354 evlist__for_each_entry_reverse(evlist, evsel)
1355 perf_evsel__close(evsel);
1358 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1360 struct cpu_map *cpus;
1361 struct thread_map *threads;
1365 * Try reading /sys/devices/system/cpu/online to get
1368 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1369 * code needs an overhaul to properly forward the
1370 * error, and we may not want to do that fallback to a
1371 * default cpu identity map :-\
1373 cpus = cpu_map__new(NULL);
1377 threads = thread_map__new_dummy();
1381 perf_evlist__set_maps(evlist, cpus, threads);
1389 int perf_evlist__open(struct perf_evlist *evlist)
1391 struct perf_evsel *evsel;
1395 * Default: one fd per CPU, all threads, aka systemwide
1396 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1398 if (evlist->threads == NULL && evlist->cpus == NULL) {
1399 err = perf_evlist__create_syswide_maps(evlist);
1404 perf_evlist__update_id_pos(evlist);
1406 evlist__for_each_entry(evlist, evsel) {
1407 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1414 perf_evlist__close(evlist);
1419 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1420 const char *argv[], bool pipe_output,
1421 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1423 int child_ready_pipe[2], go_pipe[2];
1426 if (pipe(child_ready_pipe) < 0) {
1427 perror("failed to create 'ready' pipe");
1431 if (pipe(go_pipe) < 0) {
1432 perror("failed to create 'go' pipe");
1433 goto out_close_ready_pipe;
1436 evlist->workload.pid = fork();
1437 if (evlist->workload.pid < 0) {
1438 perror("failed to fork");
1439 goto out_close_pipes;
1442 if (!evlist->workload.pid) {
1448 signal(SIGTERM, SIG_DFL);
1450 close(child_ready_pipe[0]);
1452 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1455 * Tell the parent we're ready to go
1457 close(child_ready_pipe[1]);
1460 * Wait until the parent tells us to go.
1462 ret = read(go_pipe[0], &bf, 1);
1464 * The parent will ask for the execvp() to be performed by
1465 * writing exactly one byte, in workload.cork_fd, usually via
1466 * perf_evlist__start_workload().
1468 * For cancelling the workload without actually running it,
1469 * the parent will just close workload.cork_fd, without writing
1470 * anything, i.e. read will return zero and we just exit()
1475 perror("unable to read pipe");
1479 execvp(argv[0], (char **)argv);
1484 val.sival_int = errno;
1485 if (sigqueue(getppid(), SIGUSR1, val))
1493 struct sigaction act = {
1494 .sa_flags = SA_SIGINFO,
1495 .sa_sigaction = exec_error,
1497 sigaction(SIGUSR1, &act, NULL);
1500 if (target__none(target)) {
1501 if (evlist->threads == NULL) {
1502 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1503 __func__, __LINE__);
1504 goto out_close_pipes;
1506 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1509 close(child_ready_pipe[1]);
1512 * wait for child to settle
1514 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1515 perror("unable to read pipe");
1516 goto out_close_pipes;
1519 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1520 evlist->workload.cork_fd = go_pipe[1];
1521 close(child_ready_pipe[0]);
1527 out_close_ready_pipe:
1528 close(child_ready_pipe[0]);
1529 close(child_ready_pipe[1]);
1533 int perf_evlist__start_workload(struct perf_evlist *evlist)
1535 if (evlist->workload.cork_fd > 0) {
1539 * Remove the cork, let it rip!
1541 ret = write(evlist->workload.cork_fd, &bf, 1);
1543 perror("unable to write to pipe");
1545 close(evlist->workload.cork_fd);
1552 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1553 struct perf_sample *sample)
1555 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1559 return perf_evsel__parse_sample(evsel, event, sample);
1562 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1563 union perf_event *event,
1566 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1570 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1573 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1575 struct perf_evsel *evsel;
1578 evlist__for_each_entry(evlist, evsel) {
1579 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1580 perf_evsel__name(evsel));
1583 return printed + fprintf(fp, "\n");
1586 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1587 int err, char *buf, size_t size)
1590 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1595 printed = scnprintf(buf, size,
1597 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1599 value = perf_event_paranoid();
1601 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1604 printed += scnprintf(buf + printed, size - printed,
1605 "For your workloads it needs to be <= 1\nHint:\t");
1607 printed += scnprintf(buf + printed, size - printed,
1608 "For system wide tracing it needs to be set to -1.\n");
1610 printed += scnprintf(buf + printed, size - printed,
1611 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1612 "Hint:\tThe current value is %d.", value);
1615 struct perf_evsel *first = perf_evlist__first(evlist);
1618 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1621 if (first->attr.sample_freq < (u64)max_freq)
1624 printed = scnprintf(buf, size,
1626 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1627 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1628 emsg, max_freq, first->attr.sample_freq);
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 = str_error_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_entry_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_entry(evlist, evsel) {
1697 if (evsel != tracking_evsel)
1698 evsel->tracking = false;
1701 tracking_evsel->tracking = true;
1705 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1708 struct perf_evsel *evsel;
1710 evlist__for_each_entry(evlist, evsel) {
1713 if (strcmp(str, evsel->name) == 0)
1720 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1721 enum bkw_mmap_state state)
1723 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1730 if (!evlist->overwrite_mmap)
1733 switch (old_state) {
1734 case BKW_MMAP_NOTREADY: {
1735 if (state != BKW_MMAP_RUNNING)
1739 case BKW_MMAP_RUNNING: {
1740 if (state != BKW_MMAP_DATA_PENDING)
1745 case BKW_MMAP_DATA_PENDING: {
1746 if (state != BKW_MMAP_EMPTY)
1750 case BKW_MMAP_EMPTY: {
1751 if (state != BKW_MMAP_RUNNING)
1757 WARN_ONCE(1, "Shouldn't get there\n");
1760 evlist->bkw_mmap_state = state;
1764 perf_evlist__pause(evlist);
1767 perf_evlist__resume(evlist);
1778 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1780 struct perf_evsel *evsel;
1782 evlist__for_each_entry(evlist, evsel) {
1783 if (!evsel->attr.exclude_kernel)
1791 * Events in data file are not collect in groups, but we still want
1792 * the group display. Set the artificial group and set the leader's
1793 * forced_leader flag to notify the display code.
1795 void perf_evlist__force_leader(struct perf_evlist *evlist)
1797 if (!evlist->nr_groups) {
1798 struct perf_evsel *leader = perf_evlist__first(evlist);
1800 perf_evlist__set_leader(evlist);
1801 leader->forced_leader = true;
1805 struct perf_evsel *perf_evlist__reset_weak_group(struct perf_evlist *evsel_list,
1806 struct perf_evsel *evsel)
1808 struct perf_evsel *c2, *leader;
1809 bool is_open = true;
1811 leader = evsel->leader;
1812 pr_debug("Weak group for %s/%d failed\n",
1813 leader->name, leader->nr_members);
1816 * for_each_group_member doesn't work here because it doesn't
1817 * include the first entry.
1819 evlist__for_each_entry(evsel_list, c2) {
1822 if (c2->leader == leader) {
1824 perf_evsel__close(c2);
1832 int perf_evlist__add_sb_event(struct perf_evlist **evlist,
1833 struct perf_event_attr *attr,
1834 perf_evsel__sb_cb_t cb,
1837 struct perf_evsel *evsel;
1838 bool new_evlist = (*evlist) == NULL;
1840 if (*evlist == NULL)
1841 *evlist = perf_evlist__new();
1842 if (*evlist == NULL)
1845 if (!attr->sample_id_all) {
1846 pr_warning("enabling sample_id_all for all side band events\n");
1847 attr->sample_id_all = 1;
1850 evsel = perf_evsel__new_idx(attr, (*evlist)->nr_entries);
1854 evsel->side_band.cb = cb;
1855 evsel->side_band.data = data;
1856 perf_evlist__add(*evlist, evsel);
1861 perf_evlist__delete(*evlist);
1867 static void *perf_evlist__poll_thread(void *arg)
1869 struct perf_evlist *evlist = arg;
1870 bool draining = false;
1873 while (draining || !(evlist->thread.done)) {
1876 else if (evlist->thread.done)
1880 perf_evlist__poll(evlist, 1000);
1882 for (i = 0; i < evlist->nr_mmaps; i++) {
1883 struct perf_mmap *map = &evlist->mmap[i];
1884 union perf_event *event;
1886 if (perf_mmap__read_init(map))
1888 while ((event = perf_mmap__read_event(map)) != NULL) {
1889 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1891 if (evsel && evsel->side_band.cb)
1892 evsel->side_band.cb(event, evsel->side_band.data);
1894 pr_warning("cannot locate proper evsel for the side band event\n");
1896 perf_mmap__consume(map);
1898 perf_mmap__read_done(map);
1904 int perf_evlist__start_sb_thread(struct perf_evlist *evlist,
1905 struct target *target)
1907 struct perf_evsel *counter;
1912 if (perf_evlist__create_maps(evlist, target))
1913 goto out_delete_evlist;
1915 evlist__for_each_entry(evlist, counter) {
1916 if (perf_evsel__open(counter, evlist->cpus,
1917 evlist->threads) < 0)
1918 goto out_delete_evlist;
1921 if (perf_evlist__mmap(evlist, UINT_MAX))
1922 goto out_delete_evlist;
1924 evlist__for_each_entry(evlist, counter) {
1925 if (perf_evsel__enable(counter))
1926 goto out_delete_evlist;
1929 evlist->thread.done = 0;
1930 if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1931 goto out_delete_evlist;
1936 perf_evlist__delete(evlist);
1941 void perf_evlist__stop_sb_thread(struct perf_evlist *evlist)
1945 evlist->thread.done = 1;
1946 pthread_join(evlist->thread.th, NULL);
1947 perf_evlist__delete(evlist);