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"
25 #include "parse-events.h"
26 #include <subcmd/parse-options.h>
29 #include <sys/ioctl.h>
32 #include <linux/bitops.h>
33 #include <linux/hash.h>
34 #include <linux/log2.h>
35 #include <linux/err.h>
37 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
38 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
40 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
41 struct thread_map *threads)
45 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
46 INIT_HLIST_HEAD(&evlist->heads[i]);
47 INIT_LIST_HEAD(&evlist->entries);
48 perf_evlist__set_maps(evlist, cpus, threads);
49 fdarray__init(&evlist->pollfd, 64);
50 evlist->workload.pid = -1;
51 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
54 struct perf_evlist *perf_evlist__new(void)
56 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
59 perf_evlist__init(evlist, NULL, NULL);
64 struct perf_evlist *perf_evlist__new_default(void)
66 struct perf_evlist *evlist = perf_evlist__new();
68 if (evlist && perf_evlist__add_default(evlist)) {
69 perf_evlist__delete(evlist);
76 struct perf_evlist *perf_evlist__new_dummy(void)
78 struct perf_evlist *evlist = perf_evlist__new();
80 if (evlist && perf_evlist__add_dummy(evlist)) {
81 perf_evlist__delete(evlist);
89 * perf_evlist__set_id_pos - set the positions of event ids.
90 * @evlist: selected event list
92 * Events with compatible sample types all have the same id_pos
93 * and is_pos. For convenience, put a copy on evlist.
95 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
97 struct perf_evsel *first = perf_evlist__first(evlist);
99 evlist->id_pos = first->id_pos;
100 evlist->is_pos = first->is_pos;
103 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
105 struct perf_evsel *evsel;
107 evlist__for_each_entry(evlist, evsel)
108 perf_evsel__calc_id_pos(evsel);
110 perf_evlist__set_id_pos(evlist);
113 static void perf_evlist__purge(struct perf_evlist *evlist)
115 struct perf_evsel *pos, *n;
117 evlist__for_each_entry_safe(evlist, n, pos) {
118 list_del_init(&pos->node);
120 perf_evsel__delete(pos);
123 evlist->nr_entries = 0;
126 void perf_evlist__exit(struct perf_evlist *evlist)
128 zfree(&evlist->mmap);
129 zfree(&evlist->overwrite_mmap);
130 fdarray__exit(&evlist->pollfd);
133 void perf_evlist__delete(struct perf_evlist *evlist)
138 perf_evlist__munmap(evlist);
139 perf_evlist__close(evlist);
140 cpu_map__put(evlist->cpus);
141 thread_map__put(evlist->threads);
143 evlist->threads = NULL;
144 perf_evlist__purge(evlist);
145 perf_evlist__exit(evlist);
149 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
150 struct perf_evsel *evsel)
153 * We already have cpus for evsel (via PMU sysfs) so
154 * keep it, if there's no target cpu list defined.
156 if (!evsel->own_cpus || evlist->has_user_cpus) {
157 cpu_map__put(evsel->cpus);
158 evsel->cpus = cpu_map__get(evlist->cpus);
159 } else if (evsel->cpus != evsel->own_cpus) {
160 cpu_map__put(evsel->cpus);
161 evsel->cpus = cpu_map__get(evsel->own_cpus);
164 thread_map__put(evsel->threads);
165 evsel->threads = thread_map__get(evlist->threads);
168 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
170 struct perf_evsel *evsel;
172 evlist__for_each_entry(evlist, evsel)
173 __perf_evlist__propagate_maps(evlist, evsel);
176 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
178 entry->evlist = evlist;
179 list_add_tail(&entry->node, &evlist->entries);
180 entry->idx = evlist->nr_entries;
181 entry->tracking = !entry->idx;
183 if (!evlist->nr_entries++)
184 perf_evlist__set_id_pos(evlist);
186 __perf_evlist__propagate_maps(evlist, entry);
189 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
191 evsel->evlist = NULL;
192 list_del_init(&evsel->node);
193 evlist->nr_entries -= 1;
196 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
197 struct list_head *list)
199 struct perf_evsel *evsel, *temp;
201 __evlist__for_each_entry_safe(list, temp, evsel) {
202 list_del_init(&evsel->node);
203 perf_evlist__add(evlist, evsel);
207 void __perf_evlist__set_leader(struct list_head *list)
209 struct perf_evsel *evsel, *leader;
211 leader = list_entry(list->next, struct perf_evsel, node);
212 evsel = list_entry(list->prev, struct perf_evsel, node);
214 leader->nr_members = evsel->idx - leader->idx + 1;
216 __evlist__for_each_entry(list, evsel) {
217 evsel->leader = leader;
221 void perf_evlist__set_leader(struct perf_evlist *evlist)
223 if (evlist->nr_entries) {
224 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
225 __perf_evlist__set_leader(&evlist->entries);
229 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
231 attr->precise_ip = 3;
233 while (attr->precise_ip != 0) {
234 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
243 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
245 struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
250 perf_evlist__add(evlist, evsel);
254 int perf_evlist__add_dummy(struct perf_evlist *evlist)
256 struct perf_event_attr attr = {
257 .type = PERF_TYPE_SOFTWARE,
258 .config = PERF_COUNT_SW_DUMMY,
259 .size = sizeof(attr), /* to capture ABI version */
261 struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
266 perf_evlist__add(evlist, evsel);
270 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
271 struct perf_event_attr *attrs, size_t nr_attrs)
273 struct perf_evsel *evsel, *n;
277 for (i = 0; i < nr_attrs; i++) {
278 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
280 goto out_delete_partial_list;
281 list_add_tail(&evsel->node, &head);
284 perf_evlist__splice_list_tail(evlist, &head);
288 out_delete_partial_list:
289 __evlist__for_each_entry_safe(&head, n, evsel)
290 perf_evsel__delete(evsel);
294 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
295 struct perf_event_attr *attrs, size_t nr_attrs)
299 for (i = 0; i < nr_attrs; i++)
300 event_attr_init(attrs + i);
302 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
306 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
308 struct perf_evsel *evsel;
310 evlist__for_each_entry(evlist, evsel) {
311 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
312 (int)evsel->attr.config == id)
320 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
323 struct perf_evsel *evsel;
325 evlist__for_each_entry(evlist, evsel) {
326 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
327 (strcmp(evsel->name, name) == 0))
334 int perf_evlist__add_newtp(struct perf_evlist *evlist,
335 const char *sys, const char *name, void *handler)
337 struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
342 evsel->handler = handler;
343 perf_evlist__add(evlist, evsel);
347 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
348 struct perf_evsel *evsel)
350 if (evsel->system_wide)
353 return thread_map__nr(evlist->threads);
356 void perf_evlist__disable(struct perf_evlist *evlist)
358 struct perf_evsel *pos;
360 evlist__for_each_entry(evlist, pos) {
361 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
363 perf_evsel__disable(pos);
366 evlist->enabled = false;
369 void perf_evlist__enable(struct perf_evlist *evlist)
371 struct perf_evsel *pos;
373 evlist__for_each_entry(evlist, pos) {
374 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
376 perf_evsel__enable(pos);
379 evlist->enabled = true;
382 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
384 (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
387 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
388 struct perf_evsel *evsel, int cpu)
391 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
396 for (thread = 0; thread < nr_threads; thread++) {
397 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
404 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
405 struct perf_evsel *evsel,
409 int nr_cpus = cpu_map__nr(evlist->cpus);
414 for (cpu = 0; cpu < nr_cpus; cpu++) {
415 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
422 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
423 struct perf_evsel *evsel, int idx)
425 bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
428 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
430 return perf_evlist__enable_event_thread(evlist, evsel, idx);
433 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
435 int nr_cpus = cpu_map__nr(evlist->cpus);
436 int nr_threads = thread_map__nr(evlist->threads);
438 struct perf_evsel *evsel;
440 evlist__for_each_entry(evlist, evsel) {
441 if (evsel->system_wide)
444 nfds += nr_cpus * nr_threads;
447 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
448 fdarray__grow(&evlist->pollfd, nfds) < 0)
454 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
455 struct perf_mmap *map, short revent)
457 int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
459 * Save the idx so that when we filter out fds POLLHUP'ed we can
460 * close the associated evlist->mmap[] entry.
463 evlist->pollfd.priv[pos].ptr = map;
465 fcntl(fd, F_SETFL, O_NONBLOCK);
471 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
473 return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
476 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
477 void *arg __maybe_unused)
479 struct perf_mmap *map = fda->priv[fd].ptr;
485 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
487 return fdarray__filter(&evlist->pollfd, revents_and_mask,
488 perf_evlist__munmap_filtered, NULL);
491 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
493 return fdarray__poll(&evlist->pollfd, timeout);
496 static void perf_evlist__id_hash(struct perf_evlist *evlist,
497 struct perf_evsel *evsel,
498 int cpu, int thread, u64 id)
501 struct perf_sample_id *sid = SID(evsel, cpu, thread);
505 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
506 hlist_add_head(&sid->node, &evlist->heads[hash]);
509 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
510 int cpu, int thread, u64 id)
512 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
513 evsel->id[evsel->ids++] = id;
516 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
517 struct perf_evsel *evsel,
518 int cpu, int thread, int fd)
520 u64 read_data[4] = { 0, };
521 int id_idx = 1; /* The first entry is the counter value */
525 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
532 /* Legacy way to get event id.. All hail to old kernels! */
535 * This way does not work with group format read, so bail
538 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
541 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
542 read(fd, &read_data, sizeof(read_data)) == -1)
545 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
547 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
550 id = read_data[id_idx];
553 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
557 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
558 struct perf_evsel *evsel, int idx, int cpu,
561 struct perf_sample_id *sid = SID(evsel, cpu, thread);
563 if (evlist->cpus && cpu >= 0)
564 sid->cpu = evlist->cpus->map[cpu];
567 if (!evsel->system_wide && evlist->threads && thread >= 0)
568 sid->tid = thread_map__pid(evlist->threads, thread);
573 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
575 struct hlist_head *head;
576 struct perf_sample_id *sid;
579 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
580 head = &evlist->heads[hash];
582 hlist_for_each_entry(sid, head, node)
589 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
591 struct perf_sample_id *sid;
593 if (evlist->nr_entries == 1 || !id)
594 return perf_evlist__first(evlist);
596 sid = perf_evlist__id2sid(evlist, id);
600 if (!perf_evlist__sample_id_all(evlist))
601 return perf_evlist__first(evlist);
606 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
609 struct perf_sample_id *sid;
614 sid = perf_evlist__id2sid(evlist, id);
621 static int perf_evlist__event2id(struct perf_evlist *evlist,
622 union perf_event *event, u64 *id)
624 const u64 *array = event->sample.array;
627 n = (event->header.size - sizeof(event->header)) >> 3;
629 if (event->header.type == PERF_RECORD_SAMPLE) {
630 if (evlist->id_pos >= n)
632 *id = array[evlist->id_pos];
634 if (evlist->is_pos > n)
642 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
643 union perf_event *event)
645 struct perf_evsel *first = perf_evlist__first(evlist);
646 struct hlist_head *head;
647 struct perf_sample_id *sid;
651 if (evlist->nr_entries == 1)
654 if (!first->attr.sample_id_all &&
655 event->header.type != PERF_RECORD_SAMPLE)
658 if (perf_evlist__event2id(evlist, event, &id))
661 /* Synthesized events have an id of zero */
665 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
666 head = &evlist->heads[hash];
668 hlist_for_each_entry(sid, head, node) {
675 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
679 if (!evlist->overwrite_mmap)
682 for (i = 0; i < evlist->nr_mmaps; i++) {
683 int fd = evlist->overwrite_mmap[i].fd;
688 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
695 static int perf_evlist__pause(struct perf_evlist *evlist)
697 return perf_evlist__set_paused(evlist, true);
700 static int perf_evlist__resume(struct perf_evlist *evlist)
702 return perf_evlist__set_paused(evlist, false);
705 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
710 for (i = 0; i < evlist->nr_mmaps; i++)
711 perf_mmap__munmap(&evlist->mmap[i]);
713 if (evlist->overwrite_mmap)
714 for (i = 0; i < evlist->nr_mmaps; i++)
715 perf_mmap__munmap(&evlist->overwrite_mmap[i]);
718 void perf_evlist__munmap(struct perf_evlist *evlist)
720 perf_evlist__munmap_nofree(evlist);
721 zfree(&evlist->mmap);
722 zfree(&evlist->overwrite_mmap);
725 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
729 struct perf_mmap *map;
731 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
732 if (cpu_map__empty(evlist->cpus))
733 evlist->nr_mmaps = thread_map__nr(evlist->threads);
734 map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
738 for (i = 0; i < evlist->nr_mmaps; i++) {
740 map[i].overwrite = overwrite;
742 * When the perf_mmap() call is made we grab one refcount, plus
743 * one extra to let perf_mmap__consume() get the last
744 * events after all real references (perf_mmap__get()) are
747 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
748 * thus does perf_mmap__get() on it.
750 refcount_set(&map[i].refcnt, 0);
756 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
757 struct perf_evsel *evsel)
759 if (evsel->attr.write_backward)
764 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
765 struct mmap_params *mp, int cpu_idx,
766 int thread, int *_output, int *_output_overwrite)
768 struct perf_evsel *evsel;
770 int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
772 evlist__for_each_entry(evlist, evsel) {
773 struct perf_mmap *maps = evlist->mmap;
774 int *output = _output;
778 mp->prot = PROT_READ | PROT_WRITE;
779 if (evsel->attr.write_backward) {
780 output = _output_overwrite;
781 maps = evlist->overwrite_mmap;
784 maps = perf_evlist__alloc_mmap(evlist, true);
787 evlist->overwrite_mmap = maps;
788 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
789 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
791 mp->prot &= ~PROT_WRITE;
794 if (evsel->system_wide && thread)
797 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
801 fd = FD(evsel, cpu, thread);
806 if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
809 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
812 perf_mmap__get(&maps[idx]);
815 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
818 * The system_wide flag causes a selected event to be opened
819 * always without a pid. Consequently it will never get a
820 * POLLHUP, but it is used for tracking in combination with
821 * other events, so it should not need to be polled anyway.
822 * Therefore don't add it for polling.
824 if (!evsel->system_wide &&
825 __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
826 perf_mmap__put(&maps[idx]);
830 if (evsel->attr.read_format & PERF_FORMAT_ID) {
831 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
834 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
842 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
843 struct mmap_params *mp)
846 int nr_cpus = cpu_map__nr(evlist->cpus);
847 int nr_threads = thread_map__nr(evlist->threads);
849 pr_debug2("perf event ring buffer mmapped per cpu\n");
850 for (cpu = 0; cpu < nr_cpus; cpu++) {
852 int output_overwrite = -1;
854 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
857 for (thread = 0; thread < nr_threads; thread++) {
858 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
859 thread, &output, &output_overwrite))
867 perf_evlist__munmap_nofree(evlist);
871 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
872 struct mmap_params *mp)
875 int nr_threads = thread_map__nr(evlist->threads);
877 pr_debug2("perf event ring buffer mmapped per thread\n");
878 for (thread = 0; thread < nr_threads; thread++) {
880 int output_overwrite = -1;
882 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
885 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
886 &output, &output_overwrite))
893 perf_evlist__munmap_nofree(evlist);
897 unsigned long perf_event_mlock_kb_in_pages(void)
902 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
904 * Pick a once upon a time good value, i.e. things look
905 * strange since we can't read a sysctl value, but lets not
910 max -= (page_size / 1024);
913 pages = (max * 1024) / page_size;
914 if (!is_power_of_2(pages))
915 pages = rounddown_pow_of_two(pages);
920 size_t perf_evlist__mmap_size(unsigned long pages)
922 if (pages == UINT_MAX)
923 pages = perf_event_mlock_kb_in_pages();
924 else if (!is_power_of_2(pages))
927 return (pages + 1) * page_size;
930 static long parse_pages_arg(const char *str, unsigned long min,
933 unsigned long pages, val;
934 static struct parse_tag tags[] = {
935 { .tag = 'B', .mult = 1 },
936 { .tag = 'K', .mult = 1 << 10 },
937 { .tag = 'M', .mult = 1 << 20 },
938 { .tag = 'G', .mult = 1 << 30 },
945 val = parse_tag_value(str, tags);
946 if (val != (unsigned long) -1) {
947 /* we got file size value */
948 pages = PERF_ALIGN(val, page_size) / page_size;
950 /* we got pages count value */
952 pages = strtoul(str, &eptr, 10);
957 if (pages == 0 && min == 0) {
958 /* leave number of pages at 0 */
959 } else if (!is_power_of_2(pages)) {
962 /* round pages up to next power of 2 */
963 pages = roundup_pow_of_two(pages);
967 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
968 pr_info("rounding mmap pages size to %s (%lu pages)\n",
978 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
980 unsigned long max = UINT_MAX;
983 if (max > SIZE_MAX / page_size)
984 max = SIZE_MAX / page_size;
986 pages = parse_pages_arg(str, 1, max);
988 pr_err("Invalid argument for --mmap_pages/-m\n");
996 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
997 int unset __maybe_unused)
999 return __perf_evlist__parse_mmap_pages(opt->value, str);
1003 * perf_evlist__mmap_ex - Create mmaps to receive events.
1004 * @evlist: list of events
1005 * @pages: map length in pages
1006 * @overwrite: overwrite older events?
1007 * @auxtrace_pages - auxtrace map length in pages
1008 * @auxtrace_overwrite - overwrite older auxtrace data?
1010 * If @overwrite is %false the user needs to signal event consumption using
1011 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1014 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1015 * consumption using auxtrace_mmap__write_tail().
1017 * Return: %0 on success, negative error code otherwise.
1019 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1020 unsigned int auxtrace_pages,
1021 bool auxtrace_overwrite)
1023 struct perf_evsel *evsel;
1024 const struct cpu_map *cpus = evlist->cpus;
1025 const struct thread_map *threads = evlist->threads;
1027 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1028 * Its value is decided by evsel's write_backward.
1029 * So &mp should not be passed through const pointer.
1031 struct mmap_params mp;
1034 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1038 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1041 evlist->mmap_len = perf_evlist__mmap_size(pages);
1042 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1043 mp.mask = evlist->mmap_len - page_size - 1;
1045 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1046 auxtrace_pages, auxtrace_overwrite);
1048 evlist__for_each_entry(evlist, evsel) {
1049 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1050 evsel->sample_id == NULL &&
1051 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1055 if (cpu_map__empty(cpus))
1056 return perf_evlist__mmap_per_thread(evlist, &mp);
1058 return perf_evlist__mmap_per_cpu(evlist, &mp);
1061 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1063 return perf_evlist__mmap_ex(evlist, pages, 0, false);
1066 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1068 bool all_threads = (target->per_thread && target->system_wide);
1069 struct cpu_map *cpus;
1070 struct thread_map *threads;
1073 * If specify '-a' and '--per-thread' to perf record, perf record
1074 * will override '--per-thread'. target->per_thread = false and
1075 * target->system_wide = true.
1077 * If specify '--per-thread' only to perf record,
1078 * target->per_thread = true and target->system_wide = false.
1080 * So target->per_thread && target->system_wide is false.
1081 * For perf record, thread_map__new_str doesn't call
1082 * thread_map__new_all_cpus. That will keep perf record's
1085 * For perf stat, it allows the case that target->per_thread and
1086 * target->system_wide are all true. It means to collect system-wide
1087 * per-thread data. thread_map__new_str will call
1088 * thread_map__new_all_cpus to enumerate all threads.
1090 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1096 if (target__uses_dummy_map(target))
1097 cpus = cpu_map__dummy_new();
1099 cpus = cpu_map__new(target->cpu_list);
1102 goto out_delete_threads;
1104 evlist->has_user_cpus = !!target->cpu_list;
1106 perf_evlist__set_maps(evlist, cpus, threads);
1111 thread_map__put(threads);
1115 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1116 struct thread_map *threads)
1119 * Allow for the possibility that one or another of the maps isn't being
1120 * changed i.e. don't put it. Note we are assuming the maps that are
1121 * being applied are brand new and evlist is taking ownership of the
1122 * original reference count of 1. If that is not the case it is up to
1123 * the caller to increase the reference count.
1125 if (cpus != evlist->cpus) {
1126 cpu_map__put(evlist->cpus);
1127 evlist->cpus = cpu_map__get(cpus);
1130 if (threads != evlist->threads) {
1131 thread_map__put(evlist->threads);
1132 evlist->threads = thread_map__get(threads);
1135 perf_evlist__propagate_maps(evlist);
1138 void __perf_evlist__set_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__set_sample_bit(evsel, bit);
1147 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1148 enum perf_event_sample_format bit)
1150 struct perf_evsel *evsel;
1152 evlist__for_each_entry(evlist, evsel)
1153 __perf_evsel__reset_sample_bit(evsel, bit);
1156 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1158 struct perf_evsel *evsel;
1161 evlist__for_each_entry(evlist, evsel) {
1162 if (evsel->filter == NULL)
1166 * filters only work for tracepoint event, which doesn't have cpu limit.
1167 * So evlist and evsel should always be same.
1169 err = perf_evsel__apply_filter(evsel, evsel->filter);
1179 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1181 struct perf_evsel *evsel;
1184 evlist__for_each_entry(evlist, evsel) {
1185 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1188 err = perf_evsel__set_filter(evsel, filter);
1196 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1202 for (i = 0; i < npids; ++i) {
1204 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1209 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1217 ret = perf_evlist__set_filter(evlist, filter);
1223 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1225 return perf_evlist__set_filter_pids(evlist, 1, &pid);
1228 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1230 struct perf_evsel *pos;
1232 if (evlist->nr_entries == 1)
1235 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1238 evlist__for_each_entry(evlist, pos) {
1239 if (pos->id_pos != evlist->id_pos ||
1240 pos->is_pos != evlist->is_pos)
1247 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1249 struct perf_evsel *evsel;
1251 if (evlist->combined_sample_type)
1252 return evlist->combined_sample_type;
1254 evlist__for_each_entry(evlist, evsel)
1255 evlist->combined_sample_type |= evsel->attr.sample_type;
1257 return evlist->combined_sample_type;
1260 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1262 evlist->combined_sample_type = 0;
1263 return __perf_evlist__combined_sample_type(evlist);
1266 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1268 struct perf_evsel *evsel;
1269 u64 branch_type = 0;
1271 evlist__for_each_entry(evlist, evsel)
1272 branch_type |= evsel->attr.branch_sample_type;
1276 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1278 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1279 u64 read_format = first->attr.read_format;
1280 u64 sample_type = first->attr.sample_type;
1282 evlist__for_each_entry(evlist, pos) {
1283 if (read_format != pos->attr.read_format)
1287 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1288 if ((sample_type & PERF_SAMPLE_READ) &&
1289 !(read_format & PERF_FORMAT_ID)) {
1296 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1298 struct perf_evsel *first = perf_evlist__first(evlist);
1299 return first->attr.read_format;
1302 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1304 struct perf_evsel *first = perf_evlist__first(evlist);
1305 struct perf_sample *data;
1309 if (!first->attr.sample_id_all)
1312 sample_type = first->attr.sample_type;
1314 if (sample_type & PERF_SAMPLE_TID)
1315 size += sizeof(data->tid) * 2;
1317 if (sample_type & PERF_SAMPLE_TIME)
1318 size += sizeof(data->time);
1320 if (sample_type & PERF_SAMPLE_ID)
1321 size += sizeof(data->id);
1323 if (sample_type & PERF_SAMPLE_STREAM_ID)
1324 size += sizeof(data->stream_id);
1326 if (sample_type & PERF_SAMPLE_CPU)
1327 size += sizeof(data->cpu) * 2;
1329 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1330 size += sizeof(data->id);
1335 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1337 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1339 evlist__for_each_entry_continue(evlist, pos) {
1340 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1347 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1349 struct perf_evsel *first = perf_evlist__first(evlist);
1350 return first->attr.sample_id_all;
1353 void perf_evlist__set_selected(struct perf_evlist *evlist,
1354 struct perf_evsel *evsel)
1356 evlist->selected = evsel;
1359 void perf_evlist__close(struct perf_evlist *evlist)
1361 struct perf_evsel *evsel;
1363 evlist__for_each_entry_reverse(evlist, evsel)
1364 perf_evsel__close(evsel);
1367 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1369 struct cpu_map *cpus;
1370 struct thread_map *threads;
1374 * Try reading /sys/devices/system/cpu/online to get
1377 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1378 * code needs an overhaul to properly forward the
1379 * error, and we may not want to do that fallback to a
1380 * default cpu identity map :-\
1382 cpus = cpu_map__new(NULL);
1386 threads = thread_map__new_dummy();
1390 perf_evlist__set_maps(evlist, cpus, threads);
1398 int perf_evlist__open(struct perf_evlist *evlist)
1400 struct perf_evsel *evsel;
1404 * Default: one fd per CPU, all threads, aka systemwide
1405 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1407 if (evlist->threads == NULL && evlist->cpus == NULL) {
1408 err = perf_evlist__create_syswide_maps(evlist);
1413 perf_evlist__update_id_pos(evlist);
1415 evlist__for_each_entry(evlist, evsel) {
1416 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1423 perf_evlist__close(evlist);
1428 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1429 const char *argv[], bool pipe_output,
1430 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1432 int child_ready_pipe[2], go_pipe[2];
1435 if (pipe(child_ready_pipe) < 0) {
1436 perror("failed to create 'ready' pipe");
1440 if (pipe(go_pipe) < 0) {
1441 perror("failed to create 'go' pipe");
1442 goto out_close_ready_pipe;
1445 evlist->workload.pid = fork();
1446 if (evlist->workload.pid < 0) {
1447 perror("failed to fork");
1448 goto out_close_pipes;
1451 if (!evlist->workload.pid) {
1457 signal(SIGTERM, SIG_DFL);
1459 close(child_ready_pipe[0]);
1461 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1464 * Tell the parent we're ready to go
1466 close(child_ready_pipe[1]);
1469 * Wait until the parent tells us to go.
1471 ret = read(go_pipe[0], &bf, 1);
1473 * The parent will ask for the execvp() to be performed by
1474 * writing exactly one byte, in workload.cork_fd, usually via
1475 * perf_evlist__start_workload().
1477 * For cancelling the workload without actually running it,
1478 * the parent will just close workload.cork_fd, without writing
1479 * anything, i.e. read will return zero and we just exit()
1484 perror("unable to read pipe");
1488 execvp(argv[0], (char **)argv);
1493 val.sival_int = errno;
1494 if (sigqueue(getppid(), SIGUSR1, val))
1502 struct sigaction act = {
1503 .sa_flags = SA_SIGINFO,
1504 .sa_sigaction = exec_error,
1506 sigaction(SIGUSR1, &act, NULL);
1509 if (target__none(target)) {
1510 if (evlist->threads == NULL) {
1511 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1512 __func__, __LINE__);
1513 goto out_close_pipes;
1515 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1518 close(child_ready_pipe[1]);
1521 * wait for child to settle
1523 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1524 perror("unable to read pipe");
1525 goto out_close_pipes;
1528 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1529 evlist->workload.cork_fd = go_pipe[1];
1530 close(child_ready_pipe[0]);
1536 out_close_ready_pipe:
1537 close(child_ready_pipe[0]);
1538 close(child_ready_pipe[1]);
1542 int perf_evlist__start_workload(struct perf_evlist *evlist)
1544 if (evlist->workload.cork_fd > 0) {
1548 * Remove the cork, let it rip!
1550 ret = write(evlist->workload.cork_fd, &bf, 1);
1552 perror("unable to write to pipe");
1554 close(evlist->workload.cork_fd);
1561 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1562 struct perf_sample *sample)
1564 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1568 return perf_evsel__parse_sample(evsel, event, sample);
1571 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1572 union perf_event *event,
1575 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1579 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1582 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1584 struct perf_evsel *evsel;
1587 evlist__for_each_entry(evlist, evsel) {
1588 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1589 perf_evsel__name(evsel));
1592 return printed + fprintf(fp, "\n");
1595 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1596 int err, char *buf, size_t size)
1599 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1604 printed = scnprintf(buf, size,
1606 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1608 value = perf_event_paranoid();
1610 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1613 printed += scnprintf(buf + printed, size - printed,
1614 "For your workloads it needs to be <= 1\nHint:\t");
1616 printed += scnprintf(buf + printed, size - printed,
1617 "For system wide tracing it needs to be set to -1.\n");
1619 printed += scnprintf(buf + printed, size - printed,
1620 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1621 "Hint:\tThe current value is %d.", value);
1624 struct perf_evsel *first = perf_evlist__first(evlist);
1627 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1630 if (first->attr.sample_freq < (u64)max_freq)
1633 printed = scnprintf(buf, size,
1635 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1636 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1637 emsg, max_freq, first->attr.sample_freq);
1642 scnprintf(buf, size, "%s", emsg);
1649 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1651 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1652 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1656 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1657 printed += scnprintf(buf + printed, size - printed,
1659 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1660 "Hint:\tTried using %zd kB.\n",
1661 emsg, pages_max_per_user, pages_attempted);
1663 if (pages_attempted >= pages_max_per_user) {
1664 printed += scnprintf(buf + printed, size - printed,
1665 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1666 pages_max_per_user + pages_attempted);
1669 printed += scnprintf(buf + printed, size - printed,
1670 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1673 scnprintf(buf, size, "%s", emsg);
1680 void perf_evlist__to_front(struct perf_evlist *evlist,
1681 struct perf_evsel *move_evsel)
1683 struct perf_evsel *evsel, *n;
1686 if (move_evsel == perf_evlist__first(evlist))
1689 evlist__for_each_entry_safe(evlist, n, evsel) {
1690 if (evsel->leader == move_evsel->leader)
1691 list_move_tail(&evsel->node, &move);
1694 list_splice(&move, &evlist->entries);
1697 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1698 struct perf_evsel *tracking_evsel)
1700 struct perf_evsel *evsel;
1702 if (tracking_evsel->tracking)
1705 evlist__for_each_entry(evlist, evsel) {
1706 if (evsel != tracking_evsel)
1707 evsel->tracking = false;
1710 tracking_evsel->tracking = true;
1714 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1717 struct perf_evsel *evsel;
1719 evlist__for_each_entry(evlist, evsel) {
1722 if (strcmp(str, evsel->name) == 0)
1729 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1730 enum bkw_mmap_state state)
1732 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1739 if (!evlist->overwrite_mmap)
1742 switch (old_state) {
1743 case BKW_MMAP_NOTREADY: {
1744 if (state != BKW_MMAP_RUNNING)
1748 case BKW_MMAP_RUNNING: {
1749 if (state != BKW_MMAP_DATA_PENDING)
1754 case BKW_MMAP_DATA_PENDING: {
1755 if (state != BKW_MMAP_EMPTY)
1759 case BKW_MMAP_EMPTY: {
1760 if (state != BKW_MMAP_RUNNING)
1766 WARN_ONCE(1, "Shouldn't get there\n");
1769 evlist->bkw_mmap_state = state;
1773 perf_evlist__pause(evlist);
1776 perf_evlist__resume(evlist);
1787 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1789 struct perf_evsel *evsel;
1791 evlist__for_each_entry(evlist, evsel) {
1792 if (!evsel->attr.exclude_kernel)
1800 * Events in data file are not collect in groups, but we still want
1801 * the group display. Set the artificial group and set the leader's
1802 * forced_leader flag to notify the display code.
1804 void perf_evlist__force_leader(struct perf_evlist *evlist)
1806 if (!evlist->nr_groups) {
1807 struct perf_evsel *leader = perf_evlist__first(evlist);
1809 perf_evlist__set_leader(evlist);
1810 leader->forced_leader = true;