1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
13 #include "thread_map.h"
20 #include "bpf-event.h"
24 #include "parse-events.h"
25 #include <subcmd/parse-options.h>
28 #include <sys/ioctl.h>
31 #include <linux/bitops.h>
32 #include <linux/hash.h>
33 #include <linux/log2.h>
34 #include <linux/err.h>
35 #include <linux/zalloc.h>
36 #include <perf/evlist.h>
37 #include <perf/evsel.h>
38 #include <perf/cpumap.h>
40 #include <internal/xyarray.h>
42 #ifdef LACKS_SIGQUEUE_PROTOTYPE
43 int sigqueue(pid_t pid, int sig, const union sigval value);
46 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
47 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
49 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
50 struct perf_thread_map *threads)
54 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
55 INIT_HLIST_HEAD(&evlist->heads[i]);
56 perf_evlist__init(&evlist->core);
57 perf_evlist__set_maps(&evlist->core, cpus, threads);
58 fdarray__init(&evlist->pollfd, 64);
59 evlist->workload.pid = -1;
60 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
63 struct evlist *evlist__new(void)
65 struct evlist *evlist = zalloc(sizeof(*evlist));
68 evlist__init(evlist, NULL, NULL);
73 struct evlist *perf_evlist__new_default(void)
75 struct evlist *evlist = evlist__new();
77 if (evlist && perf_evlist__add_default(evlist)) {
78 evlist__delete(evlist);
85 struct evlist *perf_evlist__new_dummy(void)
87 struct evlist *evlist = evlist__new();
89 if (evlist && perf_evlist__add_dummy(evlist)) {
90 evlist__delete(evlist);
98 * perf_evlist__set_id_pos - set the positions of event ids.
99 * @evlist: selected event list
101 * Events with compatible sample types all have the same id_pos
102 * and is_pos. For convenience, put a copy on evlist.
104 void perf_evlist__set_id_pos(struct evlist *evlist)
106 struct evsel *first = perf_evlist__first(evlist);
108 evlist->id_pos = first->id_pos;
109 evlist->is_pos = first->is_pos;
112 static void perf_evlist__update_id_pos(struct evlist *evlist)
116 evlist__for_each_entry(evlist, evsel)
117 perf_evsel__calc_id_pos(evsel);
119 perf_evlist__set_id_pos(evlist);
122 static void perf_evlist__purge(struct evlist *evlist)
124 struct evsel *pos, *n;
126 evlist__for_each_entry_safe(evlist, n, pos) {
127 list_del_init(&pos->core.node);
132 evlist->core.nr_entries = 0;
135 void perf_evlist__exit(struct evlist *evlist)
137 zfree(&evlist->mmap);
138 zfree(&evlist->overwrite_mmap);
139 fdarray__exit(&evlist->pollfd);
142 void evlist__delete(struct evlist *evlist)
147 perf_evlist__munmap(evlist);
148 evlist__close(evlist);
149 perf_cpu_map__put(evlist->core.cpus);
150 perf_thread_map__put(evlist->core.threads);
151 evlist->core.cpus = NULL;
152 evlist->core.threads = NULL;
153 perf_evlist__purge(evlist);
154 perf_evlist__exit(evlist);
158 void evlist__add(struct evlist *evlist, struct evsel *entry)
160 entry->evlist = evlist;
161 entry->idx = evlist->core.nr_entries;
162 entry->tracking = !entry->idx;
164 perf_evlist__add(&evlist->core, &entry->core);
166 if (evlist->core.nr_entries == 1)
167 perf_evlist__set_id_pos(evlist);
170 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
172 evsel->evlist = NULL;
173 perf_evlist__remove(&evlist->core, &evsel->core);
176 void perf_evlist__splice_list_tail(struct evlist *evlist,
177 struct list_head *list)
179 struct evsel *evsel, *temp;
181 __evlist__for_each_entry_safe(list, temp, evsel) {
182 list_del_init(&evsel->core.node);
183 evlist__add(evlist, evsel);
187 void __perf_evlist__set_leader(struct list_head *list)
189 struct evsel *evsel, *leader;
191 leader = list_entry(list->next, struct evsel, core.node);
192 evsel = list_entry(list->prev, struct evsel, core.node);
194 leader->core.nr_members = evsel->idx - leader->idx + 1;
196 __evlist__for_each_entry(list, evsel) {
197 evsel->leader = leader;
201 void perf_evlist__set_leader(struct evlist *evlist)
203 if (evlist->core.nr_entries) {
204 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
205 __perf_evlist__set_leader(&evlist->core.entries);
209 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
211 struct evsel *evsel = perf_evsel__new_cycles(precise);
216 evlist__add(evlist, evsel);
220 int perf_evlist__add_dummy(struct evlist *evlist)
222 struct perf_event_attr attr = {
223 .type = PERF_TYPE_SOFTWARE,
224 .config = PERF_COUNT_SW_DUMMY,
225 .size = sizeof(attr), /* to capture ABI version */
227 struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
232 evlist__add(evlist, evsel);
236 static int evlist__add_attrs(struct evlist *evlist,
237 struct perf_event_attr *attrs, size_t nr_attrs)
239 struct evsel *evsel, *n;
243 for (i = 0; i < nr_attrs; i++) {
244 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
246 goto out_delete_partial_list;
247 list_add_tail(&evsel->core.node, &head);
250 perf_evlist__splice_list_tail(evlist, &head);
254 out_delete_partial_list:
255 __evlist__for_each_entry_safe(&head, n, evsel)
256 evsel__delete(evsel);
260 int __perf_evlist__add_default_attrs(struct evlist *evlist,
261 struct perf_event_attr *attrs, size_t nr_attrs)
265 for (i = 0; i < nr_attrs; i++)
266 event_attr_init(attrs + i);
268 return evlist__add_attrs(evlist, attrs, nr_attrs);
272 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
276 evlist__for_each_entry(evlist, evsel) {
277 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
278 (int)evsel->core.attr.config == id)
286 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
291 evlist__for_each_entry(evlist, evsel) {
292 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
293 (strcmp(evsel->name, name) == 0))
300 int perf_evlist__add_newtp(struct evlist *evlist,
301 const char *sys, const char *name, void *handler)
303 struct evsel *evsel = perf_evsel__newtp(sys, name);
308 evsel->handler = handler;
309 evlist__add(evlist, evsel);
313 static int perf_evlist__nr_threads(struct evlist *evlist,
316 if (evsel->system_wide)
319 return perf_thread_map__nr(evlist->core.threads);
322 void evlist__disable(struct evlist *evlist)
326 evlist__for_each_entry(evlist, pos) {
327 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
332 evlist->enabled = false;
335 void evlist__enable(struct evlist *evlist)
339 evlist__for_each_entry(evlist, pos) {
340 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
345 evlist->enabled = true;
348 void perf_evlist__toggle_enable(struct evlist *evlist)
350 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
353 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
354 struct evsel *evsel, int cpu)
357 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
362 for (thread = 0; thread < nr_threads; thread++) {
363 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
370 static int perf_evlist__enable_event_thread(struct evlist *evlist,
375 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
380 for (cpu = 0; cpu < nr_cpus; cpu++) {
381 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
388 int perf_evlist__enable_event_idx(struct evlist *evlist,
389 struct evsel *evsel, int idx)
391 bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
394 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
396 return perf_evlist__enable_event_thread(evlist, evsel, idx);
399 int perf_evlist__alloc_pollfd(struct evlist *evlist)
401 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
402 int nr_threads = perf_thread_map__nr(evlist->core.threads);
406 evlist__for_each_entry(evlist, evsel) {
407 if (evsel->system_wide)
410 nfds += nr_cpus * nr_threads;
413 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
414 fdarray__grow(&evlist->pollfd, nfds) < 0)
420 static int __perf_evlist__add_pollfd(struct evlist *evlist, int fd,
421 struct perf_mmap *map, short revent)
423 int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
425 * Save the idx so that when we filter out fds POLLHUP'ed we can
426 * close the associated evlist->mmap[] entry.
429 evlist->pollfd.priv[pos].ptr = map;
431 fcntl(fd, F_SETFL, O_NONBLOCK);
437 int perf_evlist__add_pollfd(struct evlist *evlist, int fd)
439 return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
442 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
443 void *arg __maybe_unused)
445 struct perf_mmap *map = fda->priv[fd].ptr;
451 int perf_evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
453 return fdarray__filter(&evlist->pollfd, revents_and_mask,
454 perf_evlist__munmap_filtered, NULL);
457 int perf_evlist__poll(struct evlist *evlist, int timeout)
459 return fdarray__poll(&evlist->pollfd, timeout);
462 static void perf_evlist__id_hash(struct evlist *evlist,
464 int cpu, int thread, u64 id)
467 struct perf_sample_id *sid = SID(evsel, cpu, thread);
471 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
472 hlist_add_head(&sid->node, &evlist->heads[hash]);
475 void perf_evlist__id_add(struct evlist *evlist, struct evsel *evsel,
476 int cpu, int thread, u64 id)
478 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
479 evsel->id[evsel->ids++] = id;
482 int perf_evlist__id_add_fd(struct evlist *evlist,
484 int cpu, int thread, int fd)
486 u64 read_data[4] = { 0, };
487 int id_idx = 1; /* The first entry is the counter value */
491 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
498 /* Legacy way to get event id.. All hail to old kernels! */
501 * This way does not work with group format read, so bail
504 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
507 if (!(evsel->core.attr.read_format & PERF_FORMAT_ID) ||
508 read(fd, &read_data, sizeof(read_data)) == -1)
511 if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
513 if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
516 id = read_data[id_idx];
519 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
523 static void perf_evlist__set_sid_idx(struct evlist *evlist,
524 struct evsel *evsel, int idx, int cpu,
527 struct perf_sample_id *sid = SID(evsel, cpu, thread);
529 if (evlist->core.cpus && cpu >= 0)
530 sid->cpu = evlist->core.cpus->map[cpu];
533 if (!evsel->system_wide && evlist->core.threads && thread >= 0)
534 sid->tid = perf_thread_map__pid(evlist->core.threads, thread);
539 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
541 struct hlist_head *head;
542 struct perf_sample_id *sid;
545 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
546 head = &evlist->heads[hash];
548 hlist_for_each_entry(sid, head, node)
555 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
557 struct perf_sample_id *sid;
559 if (evlist->core.nr_entries == 1 || !id)
560 return perf_evlist__first(evlist);
562 sid = perf_evlist__id2sid(evlist, id);
566 if (!perf_evlist__sample_id_all(evlist))
567 return perf_evlist__first(evlist);
572 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
575 struct perf_sample_id *sid;
580 sid = perf_evlist__id2sid(evlist, id);
587 static int perf_evlist__event2id(struct evlist *evlist,
588 union perf_event *event, u64 *id)
590 const u64 *array = event->sample.array;
593 n = (event->header.size - sizeof(event->header)) >> 3;
595 if (event->header.type == PERF_RECORD_SAMPLE) {
596 if (evlist->id_pos >= n)
598 *id = array[evlist->id_pos];
600 if (evlist->is_pos > n)
608 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
609 union perf_event *event)
611 struct evsel *first = perf_evlist__first(evlist);
612 struct hlist_head *head;
613 struct perf_sample_id *sid;
617 if (evlist->core.nr_entries == 1)
620 if (!first->core.attr.sample_id_all &&
621 event->header.type != PERF_RECORD_SAMPLE)
624 if (perf_evlist__event2id(evlist, event, &id))
627 /* Synthesized events have an id of zero */
631 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
632 head = &evlist->heads[hash];
634 hlist_for_each_entry(sid, head, node) {
641 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
645 if (!evlist->overwrite_mmap)
648 for (i = 0; i < evlist->nr_mmaps; i++) {
649 int fd = evlist->overwrite_mmap[i].fd;
654 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
661 static int perf_evlist__pause(struct evlist *evlist)
663 return perf_evlist__set_paused(evlist, true);
666 static int perf_evlist__resume(struct evlist *evlist)
668 return perf_evlist__set_paused(evlist, false);
671 static void perf_evlist__munmap_nofree(struct evlist *evlist)
676 for (i = 0; i < evlist->nr_mmaps; i++)
677 perf_mmap__munmap(&evlist->mmap[i]);
679 if (evlist->overwrite_mmap)
680 for (i = 0; i < evlist->nr_mmaps; i++)
681 perf_mmap__munmap(&evlist->overwrite_mmap[i]);
684 void perf_evlist__munmap(struct evlist *evlist)
686 perf_evlist__munmap_nofree(evlist);
687 zfree(&evlist->mmap);
688 zfree(&evlist->overwrite_mmap);
691 static struct perf_mmap *perf_evlist__alloc_mmap(struct evlist *evlist,
695 struct perf_mmap *map;
697 evlist->nr_mmaps = perf_cpu_map__nr(evlist->core.cpus);
698 if (perf_cpu_map__empty(evlist->core.cpus))
699 evlist->nr_mmaps = perf_thread_map__nr(evlist->core.threads);
700 map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
704 for (i = 0; i < evlist->nr_mmaps; i++) {
706 map[i].overwrite = overwrite;
708 * When the perf_mmap() call is made we grab one refcount, plus
709 * one extra to let perf_mmap__consume() get the last
710 * events after all real references (perf_mmap__get()) are
713 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
714 * thus does perf_mmap__get() on it.
716 refcount_set(&map[i].refcnt, 0);
722 perf_evlist__should_poll(struct evlist *evlist __maybe_unused,
725 if (evsel->core.attr.write_backward)
730 static int perf_evlist__mmap_per_evsel(struct evlist *evlist, int idx,
731 struct mmap_params *mp, int cpu_idx,
732 int thread, int *_output, int *_output_overwrite)
736 int evlist_cpu = cpu_map__cpu(evlist->core.cpus, cpu_idx);
738 evlist__for_each_entry(evlist, evsel) {
739 struct perf_mmap *maps = evlist->mmap;
740 int *output = _output;
744 mp->prot = PROT_READ | PROT_WRITE;
745 if (evsel->core.attr.write_backward) {
746 output = _output_overwrite;
747 maps = evlist->overwrite_mmap;
750 maps = perf_evlist__alloc_mmap(evlist, true);
753 evlist->overwrite_mmap = maps;
754 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
755 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
757 mp->prot &= ~PROT_WRITE;
760 if (evsel->system_wide && thread)
763 cpu = perf_cpu_map__idx(evsel->core.cpus, evlist_cpu);
767 fd = FD(evsel, cpu, thread);
772 if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
775 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
778 perf_mmap__get(&maps[idx]);
781 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
784 * The system_wide flag causes a selected event to be opened
785 * always without a pid. Consequently it will never get a
786 * POLLHUP, but it is used for tracking in combination with
787 * other events, so it should not need to be polled anyway.
788 * Therefore don't add it for polling.
790 if (!evsel->system_wide &&
791 __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
792 perf_mmap__put(&maps[idx]);
796 if (evsel->core.attr.read_format & PERF_FORMAT_ID) {
797 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
800 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
808 static int perf_evlist__mmap_per_cpu(struct evlist *evlist,
809 struct mmap_params *mp)
812 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
813 int nr_threads = perf_thread_map__nr(evlist->core.threads);
815 pr_debug2("perf event ring buffer mmapped per cpu\n");
816 for (cpu = 0; cpu < nr_cpus; cpu++) {
818 int output_overwrite = -1;
820 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
823 for (thread = 0; thread < nr_threads; thread++) {
824 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
825 thread, &output, &output_overwrite))
833 perf_evlist__munmap_nofree(evlist);
837 static int perf_evlist__mmap_per_thread(struct evlist *evlist,
838 struct mmap_params *mp)
841 int nr_threads = perf_thread_map__nr(evlist->core.threads);
843 pr_debug2("perf event ring buffer mmapped per thread\n");
844 for (thread = 0; thread < nr_threads; thread++) {
846 int output_overwrite = -1;
848 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
851 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
852 &output, &output_overwrite))
859 perf_evlist__munmap_nofree(evlist);
863 unsigned long perf_event_mlock_kb_in_pages(void)
868 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
870 * Pick a once upon a time good value, i.e. things look
871 * strange since we can't read a sysctl value, but lets not
876 max -= (page_size / 1024);
879 pages = (max * 1024) / page_size;
880 if (!is_power_of_2(pages))
881 pages = rounddown_pow_of_two(pages);
886 size_t perf_evlist__mmap_size(unsigned long pages)
888 if (pages == UINT_MAX)
889 pages = perf_event_mlock_kb_in_pages();
890 else if (!is_power_of_2(pages))
893 return (pages + 1) * page_size;
896 static long parse_pages_arg(const char *str, unsigned long min,
899 unsigned long pages, val;
900 static struct parse_tag tags[] = {
901 { .tag = 'B', .mult = 1 },
902 { .tag = 'K', .mult = 1 << 10 },
903 { .tag = 'M', .mult = 1 << 20 },
904 { .tag = 'G', .mult = 1 << 30 },
911 val = parse_tag_value(str, tags);
912 if (val != (unsigned long) -1) {
913 /* we got file size value */
914 pages = PERF_ALIGN(val, page_size) / page_size;
916 /* we got pages count value */
918 pages = strtoul(str, &eptr, 10);
923 if (pages == 0 && min == 0) {
924 /* leave number of pages at 0 */
925 } else if (!is_power_of_2(pages)) {
928 /* round pages up to next power of 2 */
929 pages = roundup_pow_of_two(pages);
933 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
934 pr_info("rounding mmap pages size to %s (%lu pages)\n",
944 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
946 unsigned long max = UINT_MAX;
949 if (max > SIZE_MAX / page_size)
950 max = SIZE_MAX / page_size;
952 pages = parse_pages_arg(str, 1, max);
954 pr_err("Invalid argument for --mmap_pages/-m\n");
962 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
963 int unset __maybe_unused)
965 return __perf_evlist__parse_mmap_pages(opt->value, str);
969 * perf_evlist__mmap_ex - Create mmaps to receive events.
970 * @evlist: list of events
971 * @pages: map length in pages
972 * @overwrite: overwrite older events?
973 * @auxtrace_pages - auxtrace map length in pages
974 * @auxtrace_overwrite - overwrite older auxtrace data?
976 * If @overwrite is %false the user needs to signal event consumption using
977 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
980 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
981 * consumption using auxtrace_mmap__write_tail().
983 * Return: %0 on success, negative error code otherwise.
985 int perf_evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
986 unsigned int auxtrace_pages,
987 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
991 const struct perf_cpu_map *cpus = evlist->core.cpus;
992 const struct perf_thread_map *threads = evlist->core.threads;
994 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
995 * Its value is decided by evsel's write_backward.
996 * So &mp should not be passed through const pointer.
998 struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
999 .comp_level = comp_level };
1002 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1006 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1009 evlist->mmap_len = perf_evlist__mmap_size(pages);
1010 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1011 mp.mask = evlist->mmap_len - page_size - 1;
1013 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1014 auxtrace_pages, auxtrace_overwrite);
1016 evlist__for_each_entry(evlist, evsel) {
1017 if ((evsel->core.attr.read_format & PERF_FORMAT_ID) &&
1018 evsel->sample_id == NULL &&
1019 perf_evsel__alloc_id(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
1023 if (perf_cpu_map__empty(cpus))
1024 return perf_evlist__mmap_per_thread(evlist, &mp);
1026 return perf_evlist__mmap_per_cpu(evlist, &mp);
1029 int perf_evlist__mmap(struct evlist *evlist, unsigned int pages)
1031 return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1034 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
1036 bool all_threads = (target->per_thread && target->system_wide);
1037 struct perf_cpu_map *cpus;
1038 struct perf_thread_map *threads;
1041 * If specify '-a' and '--per-thread' to perf record, perf record
1042 * will override '--per-thread'. target->per_thread = false and
1043 * target->system_wide = true.
1045 * If specify '--per-thread' only to perf record,
1046 * target->per_thread = true and target->system_wide = false.
1048 * So target->per_thread && target->system_wide is false.
1049 * For perf record, thread_map__new_str doesn't call
1050 * thread_map__new_all_cpus. That will keep perf record's
1053 * For perf stat, it allows the case that target->per_thread and
1054 * target->system_wide are all true. It means to collect system-wide
1055 * per-thread data. thread_map__new_str will call
1056 * thread_map__new_all_cpus to enumerate all threads.
1058 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1064 if (target__uses_dummy_map(target))
1065 cpus = perf_cpu_map__dummy_new();
1067 cpus = perf_cpu_map__new(target->cpu_list);
1070 goto out_delete_threads;
1072 evlist->core.has_user_cpus = !!target->cpu_list;
1074 perf_evlist__set_maps(&evlist->core, cpus, threads);
1079 perf_thread_map__put(threads);
1083 void __perf_evlist__set_sample_bit(struct evlist *evlist,
1084 enum perf_event_sample_format bit)
1086 struct evsel *evsel;
1088 evlist__for_each_entry(evlist, evsel)
1089 __perf_evsel__set_sample_bit(evsel, bit);
1092 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
1093 enum perf_event_sample_format bit)
1095 struct evsel *evsel;
1097 evlist__for_each_entry(evlist, evsel)
1098 __perf_evsel__reset_sample_bit(evsel, bit);
1101 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1103 struct evsel *evsel;
1106 evlist__for_each_entry(evlist, evsel) {
1107 if (evsel->filter == NULL)
1111 * filters only work for tracepoint event, which doesn't have cpu limit.
1112 * So evlist and evsel should always be same.
1114 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1124 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1126 struct evsel *evsel;
1129 evlist__for_each_entry(evlist, evsel) {
1130 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1133 err = perf_evsel__set_filter(evsel, filter);
1141 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1147 for (i = 0; i < npids; ++i) {
1149 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1154 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1162 ret = perf_evlist__set_tp_filter(evlist, filter);
1168 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1170 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1173 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1177 if (evlist->core.nr_entries == 1)
1180 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1183 evlist__for_each_entry(evlist, pos) {
1184 if (pos->id_pos != evlist->id_pos ||
1185 pos->is_pos != evlist->is_pos)
1192 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1194 struct evsel *evsel;
1196 if (evlist->combined_sample_type)
1197 return evlist->combined_sample_type;
1199 evlist__for_each_entry(evlist, evsel)
1200 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1202 return evlist->combined_sample_type;
1205 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1207 evlist->combined_sample_type = 0;
1208 return __perf_evlist__combined_sample_type(evlist);
1211 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1213 struct evsel *evsel;
1214 u64 branch_type = 0;
1216 evlist__for_each_entry(evlist, evsel)
1217 branch_type |= evsel->core.attr.branch_sample_type;
1221 bool perf_evlist__valid_read_format(struct evlist *evlist)
1223 struct evsel *first = perf_evlist__first(evlist), *pos = first;
1224 u64 read_format = first->core.attr.read_format;
1225 u64 sample_type = first->core.attr.sample_type;
1227 evlist__for_each_entry(evlist, pos) {
1228 if (read_format != pos->core.attr.read_format)
1232 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1233 if ((sample_type & PERF_SAMPLE_READ) &&
1234 !(read_format & PERF_FORMAT_ID)) {
1241 u64 perf_evlist__read_format(struct evlist *evlist)
1243 struct evsel *first = perf_evlist__first(evlist);
1244 return first->core.attr.read_format;
1247 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1249 struct evsel *first = perf_evlist__first(evlist);
1250 struct perf_sample *data;
1254 if (!first->core.attr.sample_id_all)
1257 sample_type = first->core.attr.sample_type;
1259 if (sample_type & PERF_SAMPLE_TID)
1260 size += sizeof(data->tid) * 2;
1262 if (sample_type & PERF_SAMPLE_TIME)
1263 size += sizeof(data->time);
1265 if (sample_type & PERF_SAMPLE_ID)
1266 size += sizeof(data->id);
1268 if (sample_type & PERF_SAMPLE_STREAM_ID)
1269 size += sizeof(data->stream_id);
1271 if (sample_type & PERF_SAMPLE_CPU)
1272 size += sizeof(data->cpu) * 2;
1274 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1275 size += sizeof(data->id);
1280 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1282 struct evsel *first = perf_evlist__first(evlist), *pos = first;
1284 evlist__for_each_entry_continue(evlist, pos) {
1285 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1292 bool perf_evlist__sample_id_all(struct evlist *evlist)
1294 struct evsel *first = perf_evlist__first(evlist);
1295 return first->core.attr.sample_id_all;
1298 void perf_evlist__set_selected(struct evlist *evlist,
1299 struct evsel *evsel)
1301 evlist->selected = evsel;
1304 void evlist__close(struct evlist *evlist)
1306 struct evsel *evsel;
1308 evlist__for_each_entry_reverse(evlist, evsel)
1309 evsel__close(evsel);
1312 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1314 struct perf_cpu_map *cpus;
1315 struct perf_thread_map *threads;
1319 * Try reading /sys/devices/system/cpu/online to get
1322 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1323 * code needs an overhaul to properly forward the
1324 * error, and we may not want to do that fallback to a
1325 * default cpu identity map :-\
1327 cpus = perf_cpu_map__new(NULL);
1331 threads = perf_thread_map__new_dummy();
1335 perf_evlist__set_maps(&evlist->core, cpus, threads);
1339 perf_cpu_map__put(cpus);
1343 int evlist__open(struct evlist *evlist)
1345 struct evsel *evsel;
1349 * Default: one fd per CPU, all threads, aka systemwide
1350 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1352 if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1353 err = perf_evlist__create_syswide_maps(evlist);
1358 perf_evlist__update_id_pos(evlist);
1360 evlist__for_each_entry(evlist, evsel) {
1361 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1368 evlist__close(evlist);
1373 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1374 const char *argv[], bool pipe_output,
1375 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1377 int child_ready_pipe[2], go_pipe[2];
1380 if (pipe(child_ready_pipe) < 0) {
1381 perror("failed to create 'ready' pipe");
1385 if (pipe(go_pipe) < 0) {
1386 perror("failed to create 'go' pipe");
1387 goto out_close_ready_pipe;
1390 evlist->workload.pid = fork();
1391 if (evlist->workload.pid < 0) {
1392 perror("failed to fork");
1393 goto out_close_pipes;
1396 if (!evlist->workload.pid) {
1402 signal(SIGTERM, SIG_DFL);
1404 close(child_ready_pipe[0]);
1406 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1409 * Tell the parent we're ready to go
1411 close(child_ready_pipe[1]);
1414 * Wait until the parent tells us to go.
1416 ret = read(go_pipe[0], &bf, 1);
1418 * The parent will ask for the execvp() to be performed by
1419 * writing exactly one byte, in workload.cork_fd, usually via
1420 * perf_evlist__start_workload().
1422 * For cancelling the workload without actually running it,
1423 * the parent will just close workload.cork_fd, without writing
1424 * anything, i.e. read will return zero and we just exit()
1429 perror("unable to read pipe");
1433 execvp(argv[0], (char **)argv);
1438 val.sival_int = errno;
1439 if (sigqueue(getppid(), SIGUSR1, val))
1447 struct sigaction act = {
1448 .sa_flags = SA_SIGINFO,
1449 .sa_sigaction = exec_error,
1451 sigaction(SIGUSR1, &act, NULL);
1454 if (target__none(target)) {
1455 if (evlist->core.threads == NULL) {
1456 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1457 __func__, __LINE__);
1458 goto out_close_pipes;
1460 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1463 close(child_ready_pipe[1]);
1466 * wait for child to settle
1468 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1469 perror("unable to read pipe");
1470 goto out_close_pipes;
1473 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1474 evlist->workload.cork_fd = go_pipe[1];
1475 close(child_ready_pipe[0]);
1481 out_close_ready_pipe:
1482 close(child_ready_pipe[0]);
1483 close(child_ready_pipe[1]);
1487 int perf_evlist__start_workload(struct evlist *evlist)
1489 if (evlist->workload.cork_fd > 0) {
1493 * Remove the cork, let it rip!
1495 ret = write(evlist->workload.cork_fd, &bf, 1);
1497 perror("unable to write to pipe");
1499 close(evlist->workload.cork_fd);
1506 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1507 struct perf_sample *sample)
1509 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1513 return perf_evsel__parse_sample(evsel, event, sample);
1516 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1517 union perf_event *event,
1520 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1524 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1527 size_t perf_evlist__fprintf(struct evlist *evlist, FILE *fp)
1529 struct evsel *evsel;
1532 evlist__for_each_entry(evlist, evsel) {
1533 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1534 perf_evsel__name(evsel));
1537 return printed + fprintf(fp, "\n");
1540 int perf_evlist__strerror_open(struct evlist *evlist,
1541 int err, char *buf, size_t size)
1544 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1549 printed = scnprintf(buf, size,
1551 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1553 value = perf_event_paranoid();
1555 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1558 printed += scnprintf(buf + printed, size - printed,
1559 "For your workloads it needs to be <= 1\nHint:\t");
1561 printed += scnprintf(buf + printed, size - printed,
1562 "For system wide tracing it needs to be set to -1.\n");
1564 printed += scnprintf(buf + printed, size - printed,
1565 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1566 "Hint:\tThe current value is %d.", value);
1569 struct evsel *first = perf_evlist__first(evlist);
1572 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1575 if (first->core.attr.sample_freq < (u64)max_freq)
1578 printed = scnprintf(buf, size,
1580 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1581 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1582 emsg, max_freq, first->core.attr.sample_freq);
1587 scnprintf(buf, size, "%s", emsg);
1594 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1596 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1597 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1601 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1602 printed += scnprintf(buf + printed, size - printed,
1604 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1605 "Hint:\tTried using %zd kB.\n",
1606 emsg, pages_max_per_user, pages_attempted);
1608 if (pages_attempted >= pages_max_per_user) {
1609 printed += scnprintf(buf + printed, size - printed,
1610 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1611 pages_max_per_user + pages_attempted);
1614 printed += scnprintf(buf + printed, size - printed,
1615 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1618 scnprintf(buf, size, "%s", emsg);
1625 void perf_evlist__to_front(struct evlist *evlist,
1626 struct evsel *move_evsel)
1628 struct evsel *evsel, *n;
1631 if (move_evsel == perf_evlist__first(evlist))
1634 evlist__for_each_entry_safe(evlist, n, evsel) {
1635 if (evsel->leader == move_evsel->leader)
1636 list_move_tail(&evsel->core.node, &move);
1639 list_splice(&move, &evlist->core.entries);
1642 void perf_evlist__set_tracking_event(struct evlist *evlist,
1643 struct evsel *tracking_evsel)
1645 struct evsel *evsel;
1647 if (tracking_evsel->tracking)
1650 evlist__for_each_entry(evlist, evsel) {
1651 if (evsel != tracking_evsel)
1652 evsel->tracking = false;
1655 tracking_evsel->tracking = true;
1659 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1662 struct evsel *evsel;
1664 evlist__for_each_entry(evlist, evsel) {
1667 if (strcmp(str, evsel->name) == 0)
1674 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1675 enum bkw_mmap_state state)
1677 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1684 if (!evlist->overwrite_mmap)
1687 switch (old_state) {
1688 case BKW_MMAP_NOTREADY: {
1689 if (state != BKW_MMAP_RUNNING)
1693 case BKW_MMAP_RUNNING: {
1694 if (state != BKW_MMAP_DATA_PENDING)
1699 case BKW_MMAP_DATA_PENDING: {
1700 if (state != BKW_MMAP_EMPTY)
1704 case BKW_MMAP_EMPTY: {
1705 if (state != BKW_MMAP_RUNNING)
1711 WARN_ONCE(1, "Shouldn't get there\n");
1714 evlist->bkw_mmap_state = state;
1718 perf_evlist__pause(evlist);
1721 perf_evlist__resume(evlist);
1732 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1734 struct evsel *evsel;
1736 evlist__for_each_entry(evlist, evsel) {
1737 if (!evsel->core.attr.exclude_kernel)
1745 * Events in data file are not collect in groups, but we still want
1746 * the group display. Set the artificial group and set the leader's
1747 * forced_leader flag to notify the display code.
1749 void perf_evlist__force_leader(struct evlist *evlist)
1751 if (!evlist->nr_groups) {
1752 struct evsel *leader = perf_evlist__first(evlist);
1754 perf_evlist__set_leader(evlist);
1755 leader->forced_leader = true;
1759 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1760 struct evsel *evsel)
1762 struct evsel *c2, *leader;
1763 bool is_open = true;
1765 leader = evsel->leader;
1766 pr_debug("Weak group for %s/%d failed\n",
1767 leader->name, leader->core.nr_members);
1770 * for_each_group_member doesn't work here because it doesn't
1771 * include the first entry.
1773 evlist__for_each_entry(evsel_list, c2) {
1776 if (c2->leader == leader) {
1780 c2->core.nr_members = 0;
1786 int perf_evlist__add_sb_event(struct evlist **evlist,
1787 struct perf_event_attr *attr,
1788 perf_evsel__sb_cb_t cb,
1791 struct evsel *evsel;
1792 bool new_evlist = (*evlist) == NULL;
1794 if (*evlist == NULL)
1795 *evlist = evlist__new();
1796 if (*evlist == NULL)
1799 if (!attr->sample_id_all) {
1800 pr_warning("enabling sample_id_all for all side band events\n");
1801 attr->sample_id_all = 1;
1804 evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1808 evsel->side_band.cb = cb;
1809 evsel->side_band.data = data;
1810 evlist__add(*evlist, evsel);
1815 evlist__delete(*evlist);
1821 static void *perf_evlist__poll_thread(void *arg)
1823 struct evlist *evlist = arg;
1824 bool draining = false;
1828 bool got_data = false;
1830 if (evlist->thread.done)
1834 perf_evlist__poll(evlist, 1000);
1836 for (i = 0; i < evlist->nr_mmaps; i++) {
1837 struct perf_mmap *map = &evlist->mmap[i];
1838 union perf_event *event;
1840 if (perf_mmap__read_init(map))
1842 while ((event = perf_mmap__read_event(map)) != NULL) {
1843 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1845 if (evsel && evsel->side_band.cb)
1846 evsel->side_band.cb(event, evsel->side_band.data);
1848 pr_warning("cannot locate proper evsel for the side band event\n");
1850 perf_mmap__consume(map);
1853 perf_mmap__read_done(map);
1856 if (draining && !got_data)
1862 int perf_evlist__start_sb_thread(struct evlist *evlist,
1863 struct target *target)
1865 struct evsel *counter;
1870 if (perf_evlist__create_maps(evlist, target))
1871 goto out_delete_evlist;
1873 evlist__for_each_entry(evlist, counter) {
1874 if (evsel__open(counter, evlist->core.cpus,
1875 evlist->core.threads) < 0)
1876 goto out_delete_evlist;
1879 if (perf_evlist__mmap(evlist, UINT_MAX))
1880 goto out_delete_evlist;
1882 evlist__for_each_entry(evlist, counter) {
1883 if (evsel__enable(counter))
1884 goto out_delete_evlist;
1887 evlist->thread.done = 0;
1888 if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1889 goto out_delete_evlist;
1894 evlist__delete(evlist);
1899 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1903 evlist->thread.done = 1;
1904 pthread_join(evlist->thread.th, NULL);
1905 evlist__delete(evlist);