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"
19 #include "util.h" // page_size
22 #include "bpf-event.h"
28 #include "parse-events.h"
29 #include <subcmd/parse-options.h>
32 #include <sys/ioctl.h>
35 #include <linux/bitops.h>
36 #include <linux/hash.h>
37 #include <linux/log2.h>
38 #include <linux/err.h>
39 #include <linux/string.h>
40 #include <linux/zalloc.h>
41 #include <perf/evlist.h>
42 #include <perf/evsel.h>
43 #include <perf/cpumap.h>
45 #include <internal/xyarray.h>
47 #ifdef LACKS_SIGQUEUE_PROTOTYPE
48 int sigqueue(pid_t pid, int sig, const union sigval value);
51 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
52 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
54 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
55 struct perf_thread_map *threads)
59 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
60 INIT_HLIST_HEAD(&evlist->heads[i]);
61 perf_evlist__init(&evlist->core);
62 perf_evlist__set_maps(&evlist->core, cpus, threads);
63 fdarray__init(&evlist->pollfd, 64);
64 evlist->workload.pid = -1;
65 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
68 struct evlist *evlist__new(void)
70 struct evlist *evlist = zalloc(sizeof(*evlist));
73 evlist__init(evlist, NULL, NULL);
78 struct evlist *perf_evlist__new_default(void)
80 struct evlist *evlist = evlist__new();
82 if (evlist && perf_evlist__add_default(evlist)) {
83 evlist__delete(evlist);
90 struct evlist *perf_evlist__new_dummy(void)
92 struct evlist *evlist = evlist__new();
94 if (evlist && perf_evlist__add_dummy(evlist)) {
95 evlist__delete(evlist);
103 * perf_evlist__set_id_pos - set the positions of event ids.
104 * @evlist: selected event list
106 * Events with compatible sample types all have the same id_pos
107 * and is_pos. For convenience, put a copy on evlist.
109 void perf_evlist__set_id_pos(struct evlist *evlist)
111 struct evsel *first = perf_evlist__first(evlist);
113 evlist->id_pos = first->id_pos;
114 evlist->is_pos = first->is_pos;
117 static void perf_evlist__update_id_pos(struct evlist *evlist)
121 evlist__for_each_entry(evlist, evsel)
122 perf_evsel__calc_id_pos(evsel);
124 perf_evlist__set_id_pos(evlist);
127 static void perf_evlist__purge(struct evlist *evlist)
129 struct evsel *pos, *n;
131 evlist__for_each_entry_safe(evlist, n, pos) {
132 list_del_init(&pos->core.node);
137 evlist->core.nr_entries = 0;
140 void perf_evlist__exit(struct evlist *evlist)
142 zfree(&evlist->mmap);
143 zfree(&evlist->overwrite_mmap);
144 fdarray__exit(&evlist->pollfd);
147 void evlist__delete(struct evlist *evlist)
152 perf_evlist__munmap(evlist);
153 evlist__close(evlist);
154 perf_cpu_map__put(evlist->core.cpus);
155 perf_thread_map__put(evlist->core.threads);
156 evlist->core.cpus = NULL;
157 evlist->core.threads = NULL;
158 perf_evlist__purge(evlist);
159 perf_evlist__exit(evlist);
163 void evlist__add(struct evlist *evlist, struct evsel *entry)
165 entry->evlist = evlist;
166 entry->idx = evlist->core.nr_entries;
167 entry->tracking = !entry->idx;
169 perf_evlist__add(&evlist->core, &entry->core);
171 if (evlist->core.nr_entries == 1)
172 perf_evlist__set_id_pos(evlist);
175 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
177 evsel->evlist = NULL;
178 perf_evlist__remove(&evlist->core, &evsel->core);
181 void perf_evlist__splice_list_tail(struct evlist *evlist,
182 struct list_head *list)
184 struct evsel *evsel, *temp;
186 __evlist__for_each_entry_safe(list, temp, evsel) {
187 list_del_init(&evsel->core.node);
188 evlist__add(evlist, evsel);
192 void __perf_evlist__set_leader(struct list_head *list)
194 struct evsel *evsel, *leader;
196 leader = list_entry(list->next, struct evsel, core.node);
197 evsel = list_entry(list->prev, struct evsel, core.node);
199 leader->core.nr_members = evsel->idx - leader->idx + 1;
201 __evlist__for_each_entry(list, evsel) {
202 evsel->leader = leader;
206 void perf_evlist__set_leader(struct evlist *evlist)
208 if (evlist->core.nr_entries) {
209 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
210 __perf_evlist__set_leader(&evlist->core.entries);
214 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
216 struct evsel *evsel = perf_evsel__new_cycles(precise);
221 evlist__add(evlist, evsel);
225 int perf_evlist__add_dummy(struct evlist *evlist)
227 struct perf_event_attr attr = {
228 .type = PERF_TYPE_SOFTWARE,
229 .config = PERF_COUNT_SW_DUMMY,
230 .size = sizeof(attr), /* to capture ABI version */
232 struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
237 evlist__add(evlist, evsel);
241 static int evlist__add_attrs(struct evlist *evlist,
242 struct perf_event_attr *attrs, size_t nr_attrs)
244 struct evsel *evsel, *n;
248 for (i = 0; i < nr_attrs; i++) {
249 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
251 goto out_delete_partial_list;
252 list_add_tail(&evsel->core.node, &head);
255 perf_evlist__splice_list_tail(evlist, &head);
259 out_delete_partial_list:
260 __evlist__for_each_entry_safe(&head, n, evsel)
261 evsel__delete(evsel);
265 int __perf_evlist__add_default_attrs(struct evlist *evlist,
266 struct perf_event_attr *attrs, size_t nr_attrs)
270 for (i = 0; i < nr_attrs; i++)
271 event_attr_init(attrs + i);
273 return evlist__add_attrs(evlist, attrs, nr_attrs);
277 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
281 evlist__for_each_entry(evlist, evsel) {
282 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
283 (int)evsel->core.attr.config == id)
291 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
296 evlist__for_each_entry(evlist, evsel) {
297 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
298 (strcmp(evsel->name, name) == 0))
305 int perf_evlist__add_newtp(struct evlist *evlist,
306 const char *sys, const char *name, void *handler)
308 struct evsel *evsel = perf_evsel__newtp(sys, name);
313 evsel->handler = handler;
314 evlist__add(evlist, evsel);
318 static int perf_evlist__nr_threads(struct evlist *evlist,
321 if (evsel->system_wide)
324 return perf_thread_map__nr(evlist->core.threads);
327 void evlist__disable(struct evlist *evlist)
331 evlist__for_each_entry(evlist, pos) {
332 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
337 evlist->enabled = false;
340 void evlist__enable(struct evlist *evlist)
344 evlist__for_each_entry(evlist, pos) {
345 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
350 evlist->enabled = true;
353 void perf_evlist__toggle_enable(struct evlist *evlist)
355 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
358 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
359 struct evsel *evsel, int cpu)
362 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
367 for (thread = 0; thread < nr_threads; thread++) {
368 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
375 static int perf_evlist__enable_event_thread(struct evlist *evlist,
380 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
385 for (cpu = 0; cpu < nr_cpus; cpu++) {
386 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
393 int perf_evlist__enable_event_idx(struct evlist *evlist,
394 struct evsel *evsel, int idx)
396 bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
399 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
401 return perf_evlist__enable_event_thread(evlist, evsel, idx);
404 int perf_evlist__alloc_pollfd(struct evlist *evlist)
406 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
407 int nr_threads = perf_thread_map__nr(evlist->core.threads);
411 evlist__for_each_entry(evlist, evsel) {
412 if (evsel->system_wide)
415 nfds += nr_cpus * nr_threads;
418 if (fdarray__available_entries(&evlist->pollfd) < nfds &&
419 fdarray__grow(&evlist->pollfd, nfds) < 0)
425 static int __perf_evlist__add_pollfd(struct evlist *evlist, int fd,
426 struct mmap *map, short revent)
428 int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
430 * Save the idx so that when we filter out fds POLLHUP'ed we can
431 * close the associated evlist->mmap[] entry.
434 evlist->pollfd.priv[pos].ptr = map;
436 fcntl(fd, F_SETFL, O_NONBLOCK);
442 int perf_evlist__add_pollfd(struct evlist *evlist, int fd)
444 return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
447 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
448 void *arg __maybe_unused)
450 struct mmap *map = fda->priv[fd].ptr;
456 int perf_evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
458 return fdarray__filter(&evlist->pollfd, revents_and_mask,
459 perf_evlist__munmap_filtered, NULL);
462 int perf_evlist__poll(struct evlist *evlist, int timeout)
464 return fdarray__poll(&evlist->pollfd, timeout);
467 static void perf_evlist__id_hash(struct evlist *evlist,
469 int cpu, int thread, u64 id)
472 struct perf_sample_id *sid = SID(evsel, cpu, thread);
476 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
477 hlist_add_head(&sid->node, &evlist->heads[hash]);
480 void perf_evlist__id_add(struct evlist *evlist, struct evsel *evsel,
481 int cpu, int thread, u64 id)
483 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
484 evsel->id[evsel->ids++] = id;
487 int perf_evlist__id_add_fd(struct evlist *evlist,
489 int cpu, int thread, int fd)
491 u64 read_data[4] = { 0, };
492 int id_idx = 1; /* The first entry is the counter value */
496 ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
503 /* Legacy way to get event id.. All hail to old kernels! */
506 * This way does not work with group format read, so bail
509 if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
512 if (!(evsel->core.attr.read_format & PERF_FORMAT_ID) ||
513 read(fd, &read_data, sizeof(read_data)) == -1)
516 if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
518 if (evsel->core.attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
521 id = read_data[id_idx];
524 perf_evlist__id_add(evlist, evsel, cpu, thread, id);
528 static void perf_evlist__set_sid_idx(struct evlist *evlist,
529 struct evsel *evsel, int idx, int cpu,
532 struct perf_sample_id *sid = SID(evsel, cpu, thread);
534 if (evlist->core.cpus && cpu >= 0)
535 sid->cpu = evlist->core.cpus->map[cpu];
538 if (!evsel->system_wide && evlist->core.threads && thread >= 0)
539 sid->tid = perf_thread_map__pid(evlist->core.threads, thread);
544 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
546 struct hlist_head *head;
547 struct perf_sample_id *sid;
550 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
551 head = &evlist->heads[hash];
553 hlist_for_each_entry(sid, head, node)
560 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
562 struct perf_sample_id *sid;
564 if (evlist->core.nr_entries == 1 || !id)
565 return perf_evlist__first(evlist);
567 sid = perf_evlist__id2sid(evlist, id);
571 if (!perf_evlist__sample_id_all(evlist))
572 return perf_evlist__first(evlist);
577 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
580 struct perf_sample_id *sid;
585 sid = perf_evlist__id2sid(evlist, id);
592 static int perf_evlist__event2id(struct evlist *evlist,
593 union perf_event *event, u64 *id)
595 const __u64 *array = event->sample.array;
598 n = (event->header.size - sizeof(event->header)) >> 3;
600 if (event->header.type == PERF_RECORD_SAMPLE) {
601 if (evlist->id_pos >= n)
603 *id = array[evlist->id_pos];
605 if (evlist->is_pos > n)
613 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
614 union perf_event *event)
616 struct evsel *first = perf_evlist__first(evlist);
617 struct hlist_head *head;
618 struct perf_sample_id *sid;
622 if (evlist->core.nr_entries == 1)
625 if (!first->core.attr.sample_id_all &&
626 event->header.type != PERF_RECORD_SAMPLE)
629 if (perf_evlist__event2id(evlist, event, &id))
632 /* Synthesized events have an id of zero */
636 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
637 head = &evlist->heads[hash];
639 hlist_for_each_entry(sid, head, node) {
646 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
650 if (!evlist->overwrite_mmap)
653 for (i = 0; i < evlist->nr_mmaps; i++) {
654 int fd = evlist->overwrite_mmap[i].fd;
659 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
666 static int perf_evlist__pause(struct evlist *evlist)
668 return perf_evlist__set_paused(evlist, true);
671 static int perf_evlist__resume(struct evlist *evlist)
673 return perf_evlist__set_paused(evlist, false);
676 static void perf_evlist__munmap_nofree(struct evlist *evlist)
681 for (i = 0; i < evlist->nr_mmaps; i++)
682 perf_mmap__munmap(&evlist->mmap[i]);
684 if (evlist->overwrite_mmap)
685 for (i = 0; i < evlist->nr_mmaps; i++)
686 perf_mmap__munmap(&evlist->overwrite_mmap[i]);
689 void perf_evlist__munmap(struct evlist *evlist)
691 perf_evlist__munmap_nofree(evlist);
692 zfree(&evlist->mmap);
693 zfree(&evlist->overwrite_mmap);
696 static struct mmap *perf_evlist__alloc_mmap(struct evlist *evlist,
702 evlist->nr_mmaps = perf_cpu_map__nr(evlist->core.cpus);
703 if (perf_cpu_map__empty(evlist->core.cpus))
704 evlist->nr_mmaps = perf_thread_map__nr(evlist->core.threads);
705 map = zalloc(evlist->nr_mmaps * sizeof(struct mmap));
709 for (i = 0; i < evlist->nr_mmaps; i++) {
711 map[i].overwrite = overwrite;
713 * When the perf_mmap() call is made we grab one refcount, plus
714 * one extra to let perf_mmap__consume() get the last
715 * events after all real references (perf_mmap__get()) are
718 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
719 * thus does perf_mmap__get() on it.
721 refcount_set(&map[i].refcnt, 0);
727 perf_evlist__should_poll(struct evlist *evlist __maybe_unused,
730 if (evsel->core.attr.write_backward)
735 static int perf_evlist__mmap_per_evsel(struct evlist *evlist, int idx,
736 struct mmap_params *mp, int cpu_idx,
737 int thread, int *_output, int *_output_overwrite)
741 int evlist_cpu = cpu_map__cpu(evlist->core.cpus, cpu_idx);
743 evlist__for_each_entry(evlist, evsel) {
744 struct mmap *maps = evlist->mmap;
745 int *output = _output;
749 mp->prot = PROT_READ | PROT_WRITE;
750 if (evsel->core.attr.write_backward) {
751 output = _output_overwrite;
752 maps = evlist->overwrite_mmap;
755 maps = perf_evlist__alloc_mmap(evlist, true);
758 evlist->overwrite_mmap = maps;
759 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
760 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
762 mp->prot &= ~PROT_WRITE;
765 if (evsel->system_wide && thread)
768 cpu = perf_cpu_map__idx(evsel->core.cpus, evlist_cpu);
772 fd = FD(evsel, cpu, thread);
777 if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
780 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
783 perf_mmap__get(&maps[idx]);
786 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
789 * The system_wide flag causes a selected event to be opened
790 * always without a pid. Consequently it will never get a
791 * POLLHUP, but it is used for tracking in combination with
792 * other events, so it should not need to be polled anyway.
793 * Therefore don't add it for polling.
795 if (!evsel->system_wide &&
796 __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
797 perf_mmap__put(&maps[idx]);
801 if (evsel->core.attr.read_format & PERF_FORMAT_ID) {
802 if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
805 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
813 static int perf_evlist__mmap_per_cpu(struct evlist *evlist,
814 struct mmap_params *mp)
817 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
818 int nr_threads = perf_thread_map__nr(evlist->core.threads);
820 pr_debug2("perf event ring buffer mmapped per cpu\n");
821 for (cpu = 0; cpu < nr_cpus; cpu++) {
823 int output_overwrite = -1;
825 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
828 for (thread = 0; thread < nr_threads; thread++) {
829 if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
830 thread, &output, &output_overwrite))
838 perf_evlist__munmap_nofree(evlist);
842 static int perf_evlist__mmap_per_thread(struct evlist *evlist,
843 struct mmap_params *mp)
846 int nr_threads = perf_thread_map__nr(evlist->core.threads);
848 pr_debug2("perf event ring buffer mmapped per thread\n");
849 for (thread = 0; thread < nr_threads; thread++) {
851 int output_overwrite = -1;
853 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
856 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
857 &output, &output_overwrite))
864 perf_evlist__munmap_nofree(evlist);
868 unsigned long perf_event_mlock_kb_in_pages(void)
873 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
875 * Pick a once upon a time good value, i.e. things look
876 * strange since we can't read a sysctl value, but lets not
881 max -= (page_size / 1024);
884 pages = (max * 1024) / page_size;
885 if (!is_power_of_2(pages))
886 pages = rounddown_pow_of_two(pages);
891 size_t perf_evlist__mmap_size(unsigned long pages)
893 if (pages == UINT_MAX)
894 pages = perf_event_mlock_kb_in_pages();
895 else if (!is_power_of_2(pages))
898 return (pages + 1) * page_size;
901 static long parse_pages_arg(const char *str, unsigned long min,
904 unsigned long pages, val;
905 static struct parse_tag tags[] = {
906 { .tag = 'B', .mult = 1 },
907 { .tag = 'K', .mult = 1 << 10 },
908 { .tag = 'M', .mult = 1 << 20 },
909 { .tag = 'G', .mult = 1 << 30 },
916 val = parse_tag_value(str, tags);
917 if (val != (unsigned long) -1) {
918 /* we got file size value */
919 pages = PERF_ALIGN(val, page_size) / page_size;
921 /* we got pages count value */
923 pages = strtoul(str, &eptr, 10);
928 if (pages == 0 && min == 0) {
929 /* leave number of pages at 0 */
930 } else if (!is_power_of_2(pages)) {
933 /* round pages up to next power of 2 */
934 pages = roundup_pow_of_two(pages);
938 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
939 pr_info("rounding mmap pages size to %s (%lu pages)\n",
949 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
951 unsigned long max = UINT_MAX;
954 if (max > SIZE_MAX / page_size)
955 max = SIZE_MAX / page_size;
957 pages = parse_pages_arg(str, 1, max);
959 pr_err("Invalid argument for --mmap_pages/-m\n");
967 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
968 int unset __maybe_unused)
970 return __perf_evlist__parse_mmap_pages(opt->value, str);
974 * perf_evlist__mmap_ex - Create mmaps to receive events.
975 * @evlist: list of events
976 * @pages: map length in pages
977 * @overwrite: overwrite older events?
978 * @auxtrace_pages - auxtrace map length in pages
979 * @auxtrace_overwrite - overwrite older auxtrace data?
981 * If @overwrite is %false the user needs to signal event consumption using
982 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
985 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
986 * consumption using auxtrace_mmap__write_tail().
988 * Return: %0 on success, negative error code otherwise.
990 int perf_evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
991 unsigned int auxtrace_pages,
992 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
996 const struct perf_cpu_map *cpus = evlist->core.cpus;
997 const struct perf_thread_map *threads = evlist->core.threads;
999 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1000 * Its value is decided by evsel's write_backward.
1001 * So &mp should not be passed through const pointer.
1003 struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
1004 .comp_level = comp_level };
1007 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1011 if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1014 evlist->mmap_len = perf_evlist__mmap_size(pages);
1015 pr_debug("mmap size %zuB\n", evlist->mmap_len);
1016 mp.mask = evlist->mmap_len - page_size - 1;
1018 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1019 auxtrace_pages, auxtrace_overwrite);
1021 evlist__for_each_entry(evlist, evsel) {
1022 if ((evsel->core.attr.read_format & PERF_FORMAT_ID) &&
1023 evsel->sample_id == NULL &&
1024 perf_evsel__alloc_id(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
1028 if (perf_cpu_map__empty(cpus))
1029 return perf_evlist__mmap_per_thread(evlist, &mp);
1031 return perf_evlist__mmap_per_cpu(evlist, &mp);
1034 int perf_evlist__mmap(struct evlist *evlist, unsigned int pages)
1036 return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1039 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
1041 bool all_threads = (target->per_thread && target->system_wide);
1042 struct perf_cpu_map *cpus;
1043 struct perf_thread_map *threads;
1046 * If specify '-a' and '--per-thread' to perf record, perf record
1047 * will override '--per-thread'. target->per_thread = false and
1048 * target->system_wide = true.
1050 * If specify '--per-thread' only to perf record,
1051 * target->per_thread = true and target->system_wide = false.
1053 * So target->per_thread && target->system_wide is false.
1054 * For perf record, thread_map__new_str doesn't call
1055 * thread_map__new_all_cpus. That will keep perf record's
1058 * For perf stat, it allows the case that target->per_thread and
1059 * target->system_wide are all true. It means to collect system-wide
1060 * per-thread data. thread_map__new_str will call
1061 * thread_map__new_all_cpus to enumerate all threads.
1063 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1069 if (target__uses_dummy_map(target))
1070 cpus = perf_cpu_map__dummy_new();
1072 cpus = perf_cpu_map__new(target->cpu_list);
1075 goto out_delete_threads;
1077 evlist->core.has_user_cpus = !!target->cpu_list;
1079 perf_evlist__set_maps(&evlist->core, cpus, threads);
1084 perf_thread_map__put(threads);
1088 void __perf_evlist__set_sample_bit(struct evlist *evlist,
1089 enum perf_event_sample_format bit)
1091 struct evsel *evsel;
1093 evlist__for_each_entry(evlist, evsel)
1094 __perf_evsel__set_sample_bit(evsel, bit);
1097 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
1098 enum perf_event_sample_format bit)
1100 struct evsel *evsel;
1102 evlist__for_each_entry(evlist, evsel)
1103 __perf_evsel__reset_sample_bit(evsel, bit);
1106 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1108 struct evsel *evsel;
1111 evlist__for_each_entry(evlist, evsel) {
1112 if (evsel->filter == NULL)
1116 * filters only work for tracepoint event, which doesn't have cpu limit.
1117 * So evlist and evsel should always be same.
1119 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1129 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1131 struct evsel *evsel;
1134 evlist__for_each_entry(evlist, evsel) {
1135 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1138 err = perf_evsel__set_filter(evsel, filter);
1146 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1152 for (i = 0; i < npids; ++i) {
1154 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1159 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1167 ret = perf_evlist__set_tp_filter(evlist, filter);
1173 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1175 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1178 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1182 if (evlist->core.nr_entries == 1)
1185 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1188 evlist__for_each_entry(evlist, pos) {
1189 if (pos->id_pos != evlist->id_pos ||
1190 pos->is_pos != evlist->is_pos)
1197 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1199 struct evsel *evsel;
1201 if (evlist->combined_sample_type)
1202 return evlist->combined_sample_type;
1204 evlist__for_each_entry(evlist, evsel)
1205 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1207 return evlist->combined_sample_type;
1210 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1212 evlist->combined_sample_type = 0;
1213 return __perf_evlist__combined_sample_type(evlist);
1216 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1218 struct evsel *evsel;
1219 u64 branch_type = 0;
1221 evlist__for_each_entry(evlist, evsel)
1222 branch_type |= evsel->core.attr.branch_sample_type;
1226 bool perf_evlist__valid_read_format(struct evlist *evlist)
1228 struct evsel *first = perf_evlist__first(evlist), *pos = first;
1229 u64 read_format = first->core.attr.read_format;
1230 u64 sample_type = first->core.attr.sample_type;
1232 evlist__for_each_entry(evlist, pos) {
1233 if (read_format != pos->core.attr.read_format)
1237 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1238 if ((sample_type & PERF_SAMPLE_READ) &&
1239 !(read_format & PERF_FORMAT_ID)) {
1246 u64 perf_evlist__read_format(struct evlist *evlist)
1248 struct evsel *first = perf_evlist__first(evlist);
1249 return first->core.attr.read_format;
1252 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1254 struct evsel *first = perf_evlist__first(evlist);
1255 struct perf_sample *data;
1259 if (!first->core.attr.sample_id_all)
1262 sample_type = first->core.attr.sample_type;
1264 if (sample_type & PERF_SAMPLE_TID)
1265 size += sizeof(data->tid) * 2;
1267 if (sample_type & PERF_SAMPLE_TIME)
1268 size += sizeof(data->time);
1270 if (sample_type & PERF_SAMPLE_ID)
1271 size += sizeof(data->id);
1273 if (sample_type & PERF_SAMPLE_STREAM_ID)
1274 size += sizeof(data->stream_id);
1276 if (sample_type & PERF_SAMPLE_CPU)
1277 size += sizeof(data->cpu) * 2;
1279 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1280 size += sizeof(data->id);
1285 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1287 struct evsel *first = perf_evlist__first(evlist), *pos = first;
1289 evlist__for_each_entry_continue(evlist, pos) {
1290 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1297 bool perf_evlist__sample_id_all(struct evlist *evlist)
1299 struct evsel *first = perf_evlist__first(evlist);
1300 return first->core.attr.sample_id_all;
1303 void perf_evlist__set_selected(struct evlist *evlist,
1304 struct evsel *evsel)
1306 evlist->selected = evsel;
1309 void evlist__close(struct evlist *evlist)
1311 struct evsel *evsel;
1313 evlist__for_each_entry_reverse(evlist, evsel)
1314 evsel__close(evsel);
1317 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1319 struct perf_cpu_map *cpus;
1320 struct perf_thread_map *threads;
1324 * Try reading /sys/devices/system/cpu/online to get
1327 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1328 * code needs an overhaul to properly forward the
1329 * error, and we may not want to do that fallback to a
1330 * default cpu identity map :-\
1332 cpus = perf_cpu_map__new(NULL);
1336 threads = perf_thread_map__new_dummy();
1340 perf_evlist__set_maps(&evlist->core, cpus, threads);
1344 perf_cpu_map__put(cpus);
1348 int evlist__open(struct evlist *evlist)
1350 struct evsel *evsel;
1354 * Default: one fd per CPU, all threads, aka systemwide
1355 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1357 if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1358 err = perf_evlist__create_syswide_maps(evlist);
1363 perf_evlist__update_id_pos(evlist);
1365 evlist__for_each_entry(evlist, evsel) {
1366 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1373 evlist__close(evlist);
1378 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1379 const char *argv[], bool pipe_output,
1380 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1382 int child_ready_pipe[2], go_pipe[2];
1385 if (pipe(child_ready_pipe) < 0) {
1386 perror("failed to create 'ready' pipe");
1390 if (pipe(go_pipe) < 0) {
1391 perror("failed to create 'go' pipe");
1392 goto out_close_ready_pipe;
1395 evlist->workload.pid = fork();
1396 if (evlist->workload.pid < 0) {
1397 perror("failed to fork");
1398 goto out_close_pipes;
1401 if (!evlist->workload.pid) {
1407 signal(SIGTERM, SIG_DFL);
1409 close(child_ready_pipe[0]);
1411 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1414 * Tell the parent we're ready to go
1416 close(child_ready_pipe[1]);
1419 * Wait until the parent tells us to go.
1421 ret = read(go_pipe[0], &bf, 1);
1423 * The parent will ask for the execvp() to be performed by
1424 * writing exactly one byte, in workload.cork_fd, usually via
1425 * perf_evlist__start_workload().
1427 * For cancelling the workload without actually running it,
1428 * the parent will just close workload.cork_fd, without writing
1429 * anything, i.e. read will return zero and we just exit()
1434 perror("unable to read pipe");
1438 execvp(argv[0], (char **)argv);
1443 val.sival_int = errno;
1444 if (sigqueue(getppid(), SIGUSR1, val))
1452 struct sigaction act = {
1453 .sa_flags = SA_SIGINFO,
1454 .sa_sigaction = exec_error,
1456 sigaction(SIGUSR1, &act, NULL);
1459 if (target__none(target)) {
1460 if (evlist->core.threads == NULL) {
1461 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1462 __func__, __LINE__);
1463 goto out_close_pipes;
1465 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1468 close(child_ready_pipe[1]);
1471 * wait for child to settle
1473 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1474 perror("unable to read pipe");
1475 goto out_close_pipes;
1478 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1479 evlist->workload.cork_fd = go_pipe[1];
1480 close(child_ready_pipe[0]);
1486 out_close_ready_pipe:
1487 close(child_ready_pipe[0]);
1488 close(child_ready_pipe[1]);
1492 int perf_evlist__start_workload(struct evlist *evlist)
1494 if (evlist->workload.cork_fd > 0) {
1498 * Remove the cork, let it rip!
1500 ret = write(evlist->workload.cork_fd, &bf, 1);
1502 perror("unable to write to pipe");
1504 close(evlist->workload.cork_fd);
1511 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1512 struct perf_sample *sample)
1514 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1518 return perf_evsel__parse_sample(evsel, event, sample);
1521 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1522 union perf_event *event,
1525 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1529 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1532 size_t perf_evlist__fprintf(struct evlist *evlist, FILE *fp)
1534 struct evsel *evsel;
1537 evlist__for_each_entry(evlist, evsel) {
1538 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1539 perf_evsel__name(evsel));
1542 return printed + fprintf(fp, "\n");
1545 int perf_evlist__strerror_open(struct evlist *evlist,
1546 int err, char *buf, size_t size)
1549 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1554 printed = scnprintf(buf, size,
1556 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1558 value = perf_event_paranoid();
1560 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1563 printed += scnprintf(buf + printed, size - printed,
1564 "For your workloads it needs to be <= 1\nHint:\t");
1566 printed += scnprintf(buf + printed, size - printed,
1567 "For system wide tracing it needs to be set to -1.\n");
1569 printed += scnprintf(buf + printed, size - printed,
1570 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1571 "Hint:\tThe current value is %d.", value);
1574 struct evsel *first = perf_evlist__first(evlist);
1577 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1580 if (first->core.attr.sample_freq < (u64)max_freq)
1583 printed = scnprintf(buf, size,
1585 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1586 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1587 emsg, max_freq, first->core.attr.sample_freq);
1592 scnprintf(buf, size, "%s", emsg);
1599 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1601 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1602 int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1606 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1607 printed += scnprintf(buf + printed, size - printed,
1609 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1610 "Hint:\tTried using %zd kB.\n",
1611 emsg, pages_max_per_user, pages_attempted);
1613 if (pages_attempted >= pages_max_per_user) {
1614 printed += scnprintf(buf + printed, size - printed,
1615 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1616 pages_max_per_user + pages_attempted);
1619 printed += scnprintf(buf + printed, size - printed,
1620 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1623 scnprintf(buf, size, "%s", emsg);
1630 void perf_evlist__to_front(struct evlist *evlist,
1631 struct evsel *move_evsel)
1633 struct evsel *evsel, *n;
1636 if (move_evsel == perf_evlist__first(evlist))
1639 evlist__for_each_entry_safe(evlist, n, evsel) {
1640 if (evsel->leader == move_evsel->leader)
1641 list_move_tail(&evsel->core.node, &move);
1644 list_splice(&move, &evlist->core.entries);
1647 void perf_evlist__set_tracking_event(struct evlist *evlist,
1648 struct evsel *tracking_evsel)
1650 struct evsel *evsel;
1652 if (tracking_evsel->tracking)
1655 evlist__for_each_entry(evlist, evsel) {
1656 if (evsel != tracking_evsel)
1657 evsel->tracking = false;
1660 tracking_evsel->tracking = true;
1664 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1667 struct evsel *evsel;
1669 evlist__for_each_entry(evlist, evsel) {
1672 if (strcmp(str, evsel->name) == 0)
1679 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1680 enum bkw_mmap_state state)
1682 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1689 if (!evlist->overwrite_mmap)
1692 switch (old_state) {
1693 case BKW_MMAP_NOTREADY: {
1694 if (state != BKW_MMAP_RUNNING)
1698 case BKW_MMAP_RUNNING: {
1699 if (state != BKW_MMAP_DATA_PENDING)
1704 case BKW_MMAP_DATA_PENDING: {
1705 if (state != BKW_MMAP_EMPTY)
1709 case BKW_MMAP_EMPTY: {
1710 if (state != BKW_MMAP_RUNNING)
1716 WARN_ONCE(1, "Shouldn't get there\n");
1719 evlist->bkw_mmap_state = state;
1723 perf_evlist__pause(evlist);
1726 perf_evlist__resume(evlist);
1737 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1739 struct evsel *evsel;
1741 evlist__for_each_entry(evlist, evsel) {
1742 if (!evsel->core.attr.exclude_kernel)
1750 * Events in data file are not collect in groups, but we still want
1751 * the group display. Set the artificial group and set the leader's
1752 * forced_leader flag to notify the display code.
1754 void perf_evlist__force_leader(struct evlist *evlist)
1756 if (!evlist->nr_groups) {
1757 struct evsel *leader = perf_evlist__first(evlist);
1759 perf_evlist__set_leader(evlist);
1760 leader->forced_leader = true;
1764 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1765 struct evsel *evsel)
1767 struct evsel *c2, *leader;
1768 bool is_open = true;
1770 leader = evsel->leader;
1771 pr_debug("Weak group for %s/%d failed\n",
1772 leader->name, leader->core.nr_members);
1775 * for_each_group_member doesn't work here because it doesn't
1776 * include the first entry.
1778 evlist__for_each_entry(evsel_list, c2) {
1781 if (c2->leader == leader) {
1785 c2->core.nr_members = 0;
1791 int perf_evlist__add_sb_event(struct evlist **evlist,
1792 struct perf_event_attr *attr,
1793 perf_evsel__sb_cb_t cb,
1796 struct evsel *evsel;
1797 bool new_evlist = (*evlist) == NULL;
1799 if (*evlist == NULL)
1800 *evlist = evlist__new();
1801 if (*evlist == NULL)
1804 if (!attr->sample_id_all) {
1805 pr_warning("enabling sample_id_all for all side band events\n");
1806 attr->sample_id_all = 1;
1809 evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1813 evsel->side_band.cb = cb;
1814 evsel->side_band.data = data;
1815 evlist__add(*evlist, evsel);
1820 evlist__delete(*evlist);
1826 static void *perf_evlist__poll_thread(void *arg)
1828 struct evlist *evlist = arg;
1829 bool draining = false;
1832 * In order to read symbols from other namespaces perf to needs to call
1833 * setns(2). This isn't permitted if the struct_fs has multiple users.
1834 * unshare(2) the fs so that we may continue to setns into namespaces
1835 * that we're observing when, for instance, reading the build-ids at
1836 * the end of a 'perf record' session.
1841 bool got_data = false;
1843 if (evlist->thread.done)
1847 perf_evlist__poll(evlist, 1000);
1849 for (i = 0; i < evlist->nr_mmaps; i++) {
1850 struct mmap *map = &evlist->mmap[i];
1851 union perf_event *event;
1853 if (perf_mmap__read_init(map))
1855 while ((event = perf_mmap__read_event(map)) != NULL) {
1856 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1858 if (evsel && evsel->side_band.cb)
1859 evsel->side_band.cb(event, evsel->side_band.data);
1861 pr_warning("cannot locate proper evsel for the side band event\n");
1863 perf_mmap__consume(map);
1866 perf_mmap__read_done(map);
1869 if (draining && !got_data)
1875 int perf_evlist__start_sb_thread(struct evlist *evlist,
1876 struct target *target)
1878 struct evsel *counter;
1883 if (perf_evlist__create_maps(evlist, target))
1884 goto out_delete_evlist;
1886 evlist__for_each_entry(evlist, counter) {
1887 if (evsel__open(counter, evlist->core.cpus,
1888 evlist->core.threads) < 0)
1889 goto out_delete_evlist;
1892 if (perf_evlist__mmap(evlist, UINT_MAX))
1893 goto out_delete_evlist;
1895 evlist__for_each_entry(evlist, counter) {
1896 if (evsel__enable(counter))
1897 goto out_delete_evlist;
1900 evlist->thread.done = 0;
1901 if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1902 goto out_delete_evlist;
1907 evlist__delete(evlist);
1912 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1916 evlist->thread.done = 1;
1917 pthread_join(evlist->thread.th, NULL);
1918 evlist__delete(evlist);