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 "util/mmap.h"
14 #include "thread_map.h"
20 #include <internal/lib.h> // page_size
24 #include "bpf-event.h"
25 #include "util/string2.h"
26 #include "util/perf_api_probe.h"
32 #include "parse-events.h"
33 #include <subcmd/parse-options.h>
36 #include <sys/ioctl.h>
39 #include <linux/bitops.h>
40 #include <linux/hash.h>
41 #include <linux/log2.h>
42 #include <linux/err.h>
43 #include <linux/string.h>
44 #include <linux/zalloc.h>
45 #include <perf/evlist.h>
46 #include <perf/evsel.h>
47 #include <perf/cpumap.h>
48 #include <perf/mmap.h>
50 #include <internal/xyarray.h>
52 #ifdef LACKS_SIGQUEUE_PROTOTYPE
53 int sigqueue(pid_t pid, int sig, const union sigval value);
56 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
57 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
59 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
60 struct perf_thread_map *threads)
62 perf_evlist__init(&evlist->core);
63 perf_evlist__set_maps(&evlist->core, cpus, threads);
64 evlist->workload.pid = -1;
65 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
66 evlist->ctl_fd.fd = -1;
67 evlist->ctl_fd.ack = -1;
68 evlist->ctl_fd.pos = -1;
71 struct evlist *evlist__new(void)
73 struct evlist *evlist = zalloc(sizeof(*evlist));
76 evlist__init(evlist, NULL, NULL);
81 struct evlist *perf_evlist__new_default(void)
83 struct evlist *evlist = evlist__new();
85 if (evlist && evlist__add_default(evlist)) {
86 evlist__delete(evlist);
93 struct evlist *perf_evlist__new_dummy(void)
95 struct evlist *evlist = evlist__new();
97 if (evlist && evlist__add_dummy(evlist)) {
98 evlist__delete(evlist);
106 * perf_evlist__set_id_pos - set the positions of event ids.
107 * @evlist: selected event list
109 * Events with compatible sample types all have the same id_pos
110 * and is_pos. For convenience, put a copy on evlist.
112 void perf_evlist__set_id_pos(struct evlist *evlist)
114 struct evsel *first = evlist__first(evlist);
116 evlist->id_pos = first->id_pos;
117 evlist->is_pos = first->is_pos;
120 static void perf_evlist__update_id_pos(struct evlist *evlist)
124 evlist__for_each_entry(evlist, evsel)
125 evsel__calc_id_pos(evsel);
127 perf_evlist__set_id_pos(evlist);
130 static void evlist__purge(struct evlist *evlist)
132 struct evsel *pos, *n;
134 evlist__for_each_entry_safe(evlist, n, pos) {
135 list_del_init(&pos->core.node);
140 evlist->core.nr_entries = 0;
143 void evlist__exit(struct evlist *evlist)
145 zfree(&evlist->mmap);
146 zfree(&evlist->overwrite_mmap);
147 perf_evlist__exit(&evlist->core);
150 void evlist__delete(struct evlist *evlist)
155 evlist__munmap(evlist);
156 evlist__close(evlist);
157 evlist__purge(evlist);
158 evlist__exit(evlist);
162 void evlist__add(struct evlist *evlist, struct evsel *entry)
164 entry->evlist = evlist;
165 entry->idx = evlist->core.nr_entries;
166 entry->tracking = !entry->idx;
168 perf_evlist__add(&evlist->core, &entry->core);
170 if (evlist->core.nr_entries == 1)
171 perf_evlist__set_id_pos(evlist);
174 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
176 evsel->evlist = NULL;
177 perf_evlist__remove(&evlist->core, &evsel->core);
180 void perf_evlist__splice_list_tail(struct evlist *evlist,
181 struct list_head *list)
183 struct evsel *evsel, *temp;
185 __evlist__for_each_entry_safe(list, temp, evsel) {
186 list_del_init(&evsel->core.node);
187 evlist__add(evlist, evsel);
191 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
192 const struct evsel_str_handler *assocs, size_t nr_assocs)
198 for (i = 0; i < nr_assocs; i++) {
199 // Adding a handler for an event not in this evlist, just ignore it.
200 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
205 if (evsel->handler != NULL)
207 evsel->handler = assocs[i].handler;
215 void __perf_evlist__set_leader(struct list_head *list)
217 struct evsel *evsel, *leader;
219 leader = list_entry(list->next, struct evsel, core.node);
220 evsel = list_entry(list->prev, struct evsel, core.node);
222 leader->core.nr_members = evsel->idx - leader->idx + 1;
224 __evlist__for_each_entry(list, evsel) {
225 evsel->leader = leader;
229 void perf_evlist__set_leader(struct evlist *evlist)
231 if (evlist->core.nr_entries) {
232 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
233 __perf_evlist__set_leader(&evlist->core.entries);
237 int __evlist__add_default(struct evlist *evlist, bool precise)
239 struct evsel *evsel = evsel__new_cycles(precise);
244 evlist__add(evlist, evsel);
248 int evlist__add_dummy(struct evlist *evlist)
250 struct perf_event_attr attr = {
251 .type = PERF_TYPE_SOFTWARE,
252 .config = PERF_COUNT_SW_DUMMY,
253 .size = sizeof(attr), /* to capture ABI version */
255 struct evsel *evsel = evsel__new_idx(&attr, evlist->core.nr_entries);
260 evlist__add(evlist, evsel);
264 static int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
266 struct evsel *evsel, *n;
270 for (i = 0; i < nr_attrs; i++) {
271 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
273 goto out_delete_partial_list;
274 list_add_tail(&evsel->core.node, &head);
277 perf_evlist__splice_list_tail(evlist, &head);
281 out_delete_partial_list:
282 __evlist__for_each_entry_safe(&head, n, evsel)
283 evsel__delete(evsel);
287 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
291 for (i = 0; i < nr_attrs; i++)
292 event_attr_init(attrs + i);
294 return evlist__add_attrs(evlist, attrs, nr_attrs);
298 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
302 evlist__for_each_entry(evlist, evsel) {
303 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
304 (int)evsel->core.attr.config == id)
312 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
317 evlist__for_each_entry(evlist, evsel) {
318 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
319 (strcmp(evsel->name, name) == 0))
326 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
328 struct evsel *evsel = evsel__newtp(sys, name);
333 evsel->handler = handler;
334 evlist__add(evlist, evsel);
338 static int perf_evlist__nr_threads(struct evlist *evlist,
341 if (evsel->core.system_wide)
344 return perf_thread_map__nr(evlist->core.threads);
347 void evlist__cpu_iter_start(struct evlist *evlist)
352 * Reset the per evsel cpu_iter. This is needed because
353 * each evsel's cpumap may have a different index space,
354 * and some operations need the index to modify
355 * the FD xyarray (e.g. open, close)
357 evlist__for_each_entry(evlist, pos)
361 bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
363 if (ev->cpu_iter >= ev->core.cpus->nr)
365 if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
370 bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
372 if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
379 void evlist__disable(struct evlist *evlist)
382 struct affinity affinity;
384 bool has_imm = false;
386 if (affinity__setup(&affinity) < 0)
389 /* Disable 'immediate' events last */
390 for (imm = 0; imm <= 1; imm++) {
391 evlist__for_each_cpu(evlist, i, cpu) {
392 affinity__set(&affinity, cpu);
394 evlist__for_each_entry(evlist, pos) {
395 if (evsel__cpu_iter_skip(pos, cpu))
397 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
401 if (pos->immediate != imm)
403 evsel__disable_cpu(pos, pos->cpu_iter - 1);
410 affinity__cleanup(&affinity);
411 evlist__for_each_entry(evlist, pos) {
412 if (!evsel__is_group_leader(pos) || !pos->core.fd)
414 pos->disabled = true;
417 evlist->enabled = false;
420 void evlist__enable(struct evlist *evlist)
423 struct affinity affinity;
426 if (affinity__setup(&affinity) < 0)
429 evlist__for_each_cpu(evlist, i, cpu) {
430 affinity__set(&affinity, cpu);
432 evlist__for_each_entry(evlist, pos) {
433 if (evsel__cpu_iter_skip(pos, cpu))
435 if (!evsel__is_group_leader(pos) || !pos->core.fd)
437 evsel__enable_cpu(pos, pos->cpu_iter - 1);
440 affinity__cleanup(&affinity);
441 evlist__for_each_entry(evlist, pos) {
442 if (!evsel__is_group_leader(pos) || !pos->core.fd)
444 pos->disabled = false;
447 evlist->enabled = true;
450 void perf_evlist__toggle_enable(struct evlist *evlist)
452 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
455 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
456 struct evsel *evsel, int cpu)
459 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
464 for (thread = 0; thread < nr_threads; thread++) {
465 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
472 static int perf_evlist__enable_event_thread(struct evlist *evlist,
477 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
482 for (cpu = 0; cpu < nr_cpus; cpu++) {
483 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
490 int perf_evlist__enable_event_idx(struct evlist *evlist,
491 struct evsel *evsel, int idx)
493 bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
496 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
498 return perf_evlist__enable_event_thread(evlist, evsel, idx);
501 int evlist__add_pollfd(struct evlist *evlist, int fd)
503 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
506 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
508 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
511 int evlist__poll(struct evlist *evlist, int timeout)
513 return perf_evlist__poll(&evlist->core, timeout);
516 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
518 struct hlist_head *head;
519 struct perf_sample_id *sid;
522 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
523 head = &evlist->core.heads[hash];
525 hlist_for_each_entry(sid, head, node)
532 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
534 struct perf_sample_id *sid;
536 if (evlist->core.nr_entries == 1 || !id)
537 return evlist__first(evlist);
539 sid = perf_evlist__id2sid(evlist, id);
541 return container_of(sid->evsel, struct evsel, core);
543 if (!evlist__sample_id_all(evlist))
544 return evlist__first(evlist);
549 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
552 struct perf_sample_id *sid;
557 sid = perf_evlist__id2sid(evlist, id);
559 return container_of(sid->evsel, struct evsel, core);
564 static int perf_evlist__event2id(struct evlist *evlist,
565 union perf_event *event, u64 *id)
567 const __u64 *array = event->sample.array;
570 n = (event->header.size - sizeof(event->header)) >> 3;
572 if (event->header.type == PERF_RECORD_SAMPLE) {
573 if (evlist->id_pos >= n)
575 *id = array[evlist->id_pos];
577 if (evlist->is_pos > n)
585 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
586 union perf_event *event)
588 struct evsel *first = evlist__first(evlist);
589 struct hlist_head *head;
590 struct perf_sample_id *sid;
594 if (evlist->core.nr_entries == 1)
597 if (!first->core.attr.sample_id_all &&
598 event->header.type != PERF_RECORD_SAMPLE)
601 if (perf_evlist__event2id(evlist, event, &id))
604 /* Synthesized events have an id of zero */
608 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
609 head = &evlist->core.heads[hash];
611 hlist_for_each_entry(sid, head, node) {
613 return container_of(sid->evsel, struct evsel, core);
618 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
622 if (!evlist->overwrite_mmap)
625 for (i = 0; i < evlist->core.nr_mmaps; i++) {
626 int fd = evlist->overwrite_mmap[i].core.fd;
631 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
638 static int perf_evlist__pause(struct evlist *evlist)
640 return perf_evlist__set_paused(evlist, true);
643 static int perf_evlist__resume(struct evlist *evlist)
645 return perf_evlist__set_paused(evlist, false);
648 static void evlist__munmap_nofree(struct evlist *evlist)
653 for (i = 0; i < evlist->core.nr_mmaps; i++)
654 perf_mmap__munmap(&evlist->mmap[i].core);
656 if (evlist->overwrite_mmap)
657 for (i = 0; i < evlist->core.nr_mmaps; i++)
658 perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
661 void evlist__munmap(struct evlist *evlist)
663 evlist__munmap_nofree(evlist);
664 zfree(&evlist->mmap);
665 zfree(&evlist->overwrite_mmap);
668 static void perf_mmap__unmap_cb(struct perf_mmap *map)
670 struct mmap *m = container_of(map, struct mmap, core);
675 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
681 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
685 for (i = 0; i < evlist->core.nr_mmaps; i++) {
686 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
689 * When the perf_mmap() call is made we grab one refcount, plus
690 * one extra to let perf_mmap__consume() get the last
691 * events after all real references (perf_mmap__get()) are
694 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
695 * thus does perf_mmap__get() on it.
697 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
704 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
705 struct perf_mmap_param *_mp,
706 int idx, bool per_cpu)
708 struct evlist *evlist = container_of(_evlist, struct evlist, core);
709 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
711 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
714 static struct perf_mmap*
715 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
717 struct evlist *evlist = container_of(_evlist, struct evlist, core);
720 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
723 maps = evlist__alloc_mmap(evlist, overwrite);
728 evlist->overwrite_mmap = maps;
729 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
730 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
736 return &maps[idx].core;
740 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
743 struct mmap *map = container_of(_map, struct mmap, core);
744 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
746 return mmap__mmap(map, mp, output, cpu);
749 unsigned long perf_event_mlock_kb_in_pages(void)
754 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
756 * Pick a once upon a time good value, i.e. things look
757 * strange since we can't read a sysctl value, but lets not
762 max -= (page_size / 1024);
765 pages = (max * 1024) / page_size;
766 if (!is_power_of_2(pages))
767 pages = rounddown_pow_of_two(pages);
772 size_t evlist__mmap_size(unsigned long pages)
774 if (pages == UINT_MAX)
775 pages = perf_event_mlock_kb_in_pages();
776 else if (!is_power_of_2(pages))
779 return (pages + 1) * page_size;
782 static long parse_pages_arg(const char *str, unsigned long min,
785 unsigned long pages, val;
786 static struct parse_tag tags[] = {
787 { .tag = 'B', .mult = 1 },
788 { .tag = 'K', .mult = 1 << 10 },
789 { .tag = 'M', .mult = 1 << 20 },
790 { .tag = 'G', .mult = 1 << 30 },
797 val = parse_tag_value(str, tags);
798 if (val != (unsigned long) -1) {
799 /* we got file size value */
800 pages = PERF_ALIGN(val, page_size) / page_size;
802 /* we got pages count value */
804 pages = strtoul(str, &eptr, 10);
809 if (pages == 0 && min == 0) {
810 /* leave number of pages at 0 */
811 } else if (!is_power_of_2(pages)) {
814 /* round pages up to next power of 2 */
815 pages = roundup_pow_of_two(pages);
819 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
820 pr_info("rounding mmap pages size to %s (%lu pages)\n",
830 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
832 unsigned long max = UINT_MAX;
835 if (max > SIZE_MAX / page_size)
836 max = SIZE_MAX / page_size;
838 pages = parse_pages_arg(str, 1, max);
840 pr_err("Invalid argument for --mmap_pages/-m\n");
848 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
849 int unset __maybe_unused)
851 return __perf_evlist__parse_mmap_pages(opt->value, str);
855 * evlist__mmap_ex - Create mmaps to receive events.
856 * @evlist: list of events
857 * @pages: map length in pages
858 * @overwrite: overwrite older events?
859 * @auxtrace_pages - auxtrace map length in pages
860 * @auxtrace_overwrite - overwrite older auxtrace data?
862 * If @overwrite is %false the user needs to signal event consumption using
863 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
866 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
867 * consumption using auxtrace_mmap__write_tail().
869 * Return: %0 on success, negative error code otherwise.
871 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
872 unsigned int auxtrace_pages,
873 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
877 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
878 * Its value is decided by evsel's write_backward.
879 * So &mp should not be passed through const pointer.
881 struct mmap_params mp = {
882 .nr_cblocks = nr_cblocks,
883 .affinity = affinity,
885 .comp_level = comp_level
887 struct perf_evlist_mmap_ops ops = {
888 .idx = perf_evlist__mmap_cb_idx,
889 .get = perf_evlist__mmap_cb_get,
890 .mmap = perf_evlist__mmap_cb_mmap,
893 evlist->core.mmap_len = evlist__mmap_size(pages);
894 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
896 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
897 auxtrace_pages, auxtrace_overwrite);
899 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
902 int evlist__mmap(struct evlist *evlist, unsigned int pages)
904 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
907 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
909 bool all_threads = (target->per_thread && target->system_wide);
910 struct perf_cpu_map *cpus;
911 struct perf_thread_map *threads;
914 * If specify '-a' and '--per-thread' to perf record, perf record
915 * will override '--per-thread'. target->per_thread = false and
916 * target->system_wide = true.
918 * If specify '--per-thread' only to perf record,
919 * target->per_thread = true and target->system_wide = false.
921 * So target->per_thread && target->system_wide is false.
922 * For perf record, thread_map__new_str doesn't call
923 * thread_map__new_all_cpus. That will keep perf record's
926 * For perf stat, it allows the case that target->per_thread and
927 * target->system_wide are all true. It means to collect system-wide
928 * per-thread data. thread_map__new_str will call
929 * thread_map__new_all_cpus to enumerate all threads.
931 threads = thread_map__new_str(target->pid, target->tid, target->uid,
937 if (target__uses_dummy_map(target))
938 cpus = perf_cpu_map__dummy_new();
940 cpus = perf_cpu_map__new(target->cpu_list);
943 goto out_delete_threads;
945 evlist->core.has_user_cpus = !!target->cpu_list;
947 perf_evlist__set_maps(&evlist->core, cpus, threads);
949 /* as evlist now has references, put count here */
950 perf_cpu_map__put(cpus);
951 perf_thread_map__put(threads);
956 perf_thread_map__put(threads);
960 void __perf_evlist__set_sample_bit(struct evlist *evlist,
961 enum perf_event_sample_format bit)
965 evlist__for_each_entry(evlist, evsel)
966 __evsel__set_sample_bit(evsel, bit);
969 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
970 enum perf_event_sample_format bit)
974 evlist__for_each_entry(evlist, evsel)
975 __evsel__reset_sample_bit(evsel, bit);
978 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
983 evlist__for_each_entry(evlist, evsel) {
984 if (evsel->filter == NULL)
988 * filters only work for tracepoint event, which doesn't have cpu limit.
989 * So evlist and evsel should always be same.
991 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1001 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1003 struct evsel *evsel;
1009 evlist__for_each_entry(evlist, evsel) {
1010 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1013 err = evsel__set_filter(evsel, filter);
1021 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1023 struct evsel *evsel;
1029 evlist__for_each_entry(evlist, evsel) {
1030 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1033 err = evsel__append_tp_filter(evsel, filter);
1041 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1046 for (i = 0; i < npids; ++i) {
1048 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1053 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1067 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1069 char *filter = asprintf__tp_filter_pids(npids, pids);
1070 int ret = perf_evlist__set_tp_filter(evlist, filter);
1076 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1078 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1081 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1083 char *filter = asprintf__tp_filter_pids(npids, pids);
1084 int ret = perf_evlist__append_tp_filter(evlist, filter);
1090 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1092 return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1095 bool evlist__valid_sample_type(struct evlist *evlist)
1099 if (evlist->core.nr_entries == 1)
1102 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1105 evlist__for_each_entry(evlist, pos) {
1106 if (pos->id_pos != evlist->id_pos ||
1107 pos->is_pos != evlist->is_pos)
1114 u64 __evlist__combined_sample_type(struct evlist *evlist)
1116 struct evsel *evsel;
1118 if (evlist->combined_sample_type)
1119 return evlist->combined_sample_type;
1121 evlist__for_each_entry(evlist, evsel)
1122 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1124 return evlist->combined_sample_type;
1127 u64 evlist__combined_sample_type(struct evlist *evlist)
1129 evlist->combined_sample_type = 0;
1130 return __evlist__combined_sample_type(evlist);
1133 u64 evlist__combined_branch_type(struct evlist *evlist)
1135 struct evsel *evsel;
1136 u64 branch_type = 0;
1138 evlist__for_each_entry(evlist, evsel)
1139 branch_type |= evsel->core.attr.branch_sample_type;
1143 bool perf_evlist__valid_read_format(struct evlist *evlist)
1145 struct evsel *first = evlist__first(evlist), *pos = first;
1146 u64 read_format = first->core.attr.read_format;
1147 u64 sample_type = first->core.attr.sample_type;
1149 evlist__for_each_entry(evlist, pos) {
1150 if (read_format != pos->core.attr.read_format) {
1151 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1152 read_format, (u64)pos->core.attr.read_format);
1156 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1157 if ((sample_type & PERF_SAMPLE_READ) &&
1158 !(read_format & PERF_FORMAT_ID)) {
1165 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1167 struct evsel *first = evlist__first(evlist);
1168 struct perf_sample *data;
1172 if (!first->core.attr.sample_id_all)
1175 sample_type = first->core.attr.sample_type;
1177 if (sample_type & PERF_SAMPLE_TID)
1178 size += sizeof(data->tid) * 2;
1180 if (sample_type & PERF_SAMPLE_TIME)
1181 size += sizeof(data->time);
1183 if (sample_type & PERF_SAMPLE_ID)
1184 size += sizeof(data->id);
1186 if (sample_type & PERF_SAMPLE_STREAM_ID)
1187 size += sizeof(data->stream_id);
1189 if (sample_type & PERF_SAMPLE_CPU)
1190 size += sizeof(data->cpu) * 2;
1192 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1193 size += sizeof(data->id);
1198 bool evlist__valid_sample_id_all(struct evlist *evlist)
1200 struct evsel *first = evlist__first(evlist), *pos = first;
1202 evlist__for_each_entry_continue(evlist, pos) {
1203 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1210 bool evlist__sample_id_all(struct evlist *evlist)
1212 struct evsel *first = evlist__first(evlist);
1213 return first->core.attr.sample_id_all;
1216 void perf_evlist__set_selected(struct evlist *evlist,
1217 struct evsel *evsel)
1219 evlist->selected = evsel;
1222 void evlist__close(struct evlist *evlist)
1224 struct evsel *evsel;
1225 struct affinity affinity;
1229 * With perf record core.cpus is usually NULL.
1230 * Use the old method to handle this for now.
1232 if (!evlist->core.cpus) {
1233 evlist__for_each_entry_reverse(evlist, evsel)
1234 evsel__close(evsel);
1238 if (affinity__setup(&affinity) < 0)
1240 evlist__for_each_cpu(evlist, i, cpu) {
1241 affinity__set(&affinity, cpu);
1243 evlist__for_each_entry_reverse(evlist, evsel) {
1244 if (evsel__cpu_iter_skip(evsel, cpu))
1246 perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1249 affinity__cleanup(&affinity);
1250 evlist__for_each_entry_reverse(evlist, evsel) {
1251 perf_evsel__free_fd(&evsel->core);
1252 perf_evsel__free_id(&evsel->core);
1256 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1258 struct perf_cpu_map *cpus;
1259 struct perf_thread_map *threads;
1263 * Try reading /sys/devices/system/cpu/online to get
1266 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1267 * code needs an overhaul to properly forward the
1268 * error, and we may not want to do that fallback to a
1269 * default cpu identity map :-\
1271 cpus = perf_cpu_map__new(NULL);
1275 threads = perf_thread_map__new_dummy();
1279 perf_evlist__set_maps(&evlist->core, cpus, threads);
1281 perf_thread_map__put(threads);
1283 perf_cpu_map__put(cpus);
1288 int evlist__open(struct evlist *evlist)
1290 struct evsel *evsel;
1294 * Default: one fd per CPU, all threads, aka systemwide
1295 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1297 if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1298 err = perf_evlist__create_syswide_maps(evlist);
1303 perf_evlist__update_id_pos(evlist);
1305 evlist__for_each_entry(evlist, evsel) {
1306 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1313 evlist__close(evlist);
1318 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1319 const char *argv[], bool pipe_output,
1320 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1322 int child_ready_pipe[2], go_pipe[2];
1325 if (pipe(child_ready_pipe) < 0) {
1326 perror("failed to create 'ready' pipe");
1330 if (pipe(go_pipe) < 0) {
1331 perror("failed to create 'go' pipe");
1332 goto out_close_ready_pipe;
1335 evlist->workload.pid = fork();
1336 if (evlist->workload.pid < 0) {
1337 perror("failed to fork");
1338 goto out_close_pipes;
1341 if (!evlist->workload.pid) {
1347 signal(SIGTERM, SIG_DFL);
1349 close(child_ready_pipe[0]);
1351 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1354 * Tell the parent we're ready to go
1356 close(child_ready_pipe[1]);
1359 * Wait until the parent tells us to go.
1361 ret = read(go_pipe[0], &bf, 1);
1363 * The parent will ask for the execvp() to be performed by
1364 * writing exactly one byte, in workload.cork_fd, usually via
1365 * perf_evlist__start_workload().
1367 * For cancelling the workload without actually running it,
1368 * the parent will just close workload.cork_fd, without writing
1369 * anything, i.e. read will return zero and we just exit()
1374 perror("unable to read pipe");
1378 execvp(argv[0], (char **)argv);
1383 val.sival_int = errno;
1384 if (sigqueue(getppid(), SIGUSR1, val))
1392 struct sigaction act = {
1393 .sa_flags = SA_SIGINFO,
1394 .sa_sigaction = exec_error,
1396 sigaction(SIGUSR1, &act, NULL);
1399 if (target__none(target)) {
1400 if (evlist->core.threads == NULL) {
1401 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1402 __func__, __LINE__);
1403 goto out_close_pipes;
1405 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1408 close(child_ready_pipe[1]);
1411 * wait for child to settle
1413 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1414 perror("unable to read pipe");
1415 goto out_close_pipes;
1418 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1419 evlist->workload.cork_fd = go_pipe[1];
1420 close(child_ready_pipe[0]);
1426 out_close_ready_pipe:
1427 close(child_ready_pipe[0]);
1428 close(child_ready_pipe[1]);
1432 int perf_evlist__start_workload(struct evlist *evlist)
1434 if (evlist->workload.cork_fd > 0) {
1438 * Remove the cork, let it rip!
1440 ret = write(evlist->workload.cork_fd, &bf, 1);
1442 perror("unable to write to pipe");
1444 close(evlist->workload.cork_fd);
1451 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1452 struct perf_sample *sample)
1454 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1458 return evsel__parse_sample(evsel, event, sample);
1461 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1462 union perf_event *event,
1465 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1469 return evsel__parse_sample_timestamp(evsel, event, timestamp);
1472 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1475 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1480 printed = scnprintf(buf, size,
1482 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1484 value = perf_event_paranoid();
1486 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1489 printed += scnprintf(buf + printed, size - printed,
1490 "For your workloads it needs to be <= 1\nHint:\t");
1492 printed += scnprintf(buf + printed, size - printed,
1493 "For system wide tracing it needs to be set to -1.\n");
1495 printed += scnprintf(buf + printed, size - printed,
1496 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1497 "Hint:\tThe current value is %d.", value);
1500 struct evsel *first = evlist__first(evlist);
1503 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1506 if (first->core.attr.sample_freq < (u64)max_freq)
1509 printed = scnprintf(buf, size,
1511 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1512 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1513 emsg, max_freq, first->core.attr.sample_freq);
1518 scnprintf(buf, size, "%s", emsg);
1525 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1527 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1528 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1532 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1533 printed += scnprintf(buf + printed, size - printed,
1535 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1536 "Hint:\tTried using %zd kB.\n",
1537 emsg, pages_max_per_user, pages_attempted);
1539 if (pages_attempted >= pages_max_per_user) {
1540 printed += scnprintf(buf + printed, size - printed,
1541 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1542 pages_max_per_user + pages_attempted);
1545 printed += scnprintf(buf + printed, size - printed,
1546 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1549 scnprintf(buf, size, "%s", emsg);
1556 void perf_evlist__to_front(struct evlist *evlist,
1557 struct evsel *move_evsel)
1559 struct evsel *evsel, *n;
1562 if (move_evsel == evlist__first(evlist))
1565 evlist__for_each_entry_safe(evlist, n, evsel) {
1566 if (evsel->leader == move_evsel->leader)
1567 list_move_tail(&evsel->core.node, &move);
1570 list_splice(&move, &evlist->core.entries);
1573 struct evsel *perf_evlist__get_tracking_event(struct evlist *evlist)
1575 struct evsel *evsel;
1577 evlist__for_each_entry(evlist, evsel) {
1578 if (evsel->tracking)
1582 return evlist__first(evlist);
1585 void perf_evlist__set_tracking_event(struct evlist *evlist,
1586 struct evsel *tracking_evsel)
1588 struct evsel *evsel;
1590 if (tracking_evsel->tracking)
1593 evlist__for_each_entry(evlist, evsel) {
1594 if (evsel != tracking_evsel)
1595 evsel->tracking = false;
1598 tracking_evsel->tracking = true;
1602 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1605 struct evsel *evsel;
1607 evlist__for_each_entry(evlist, evsel) {
1610 if (strcmp(str, evsel->name) == 0)
1617 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1618 enum bkw_mmap_state state)
1620 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1627 if (!evlist->overwrite_mmap)
1630 switch (old_state) {
1631 case BKW_MMAP_NOTREADY: {
1632 if (state != BKW_MMAP_RUNNING)
1636 case BKW_MMAP_RUNNING: {
1637 if (state != BKW_MMAP_DATA_PENDING)
1642 case BKW_MMAP_DATA_PENDING: {
1643 if (state != BKW_MMAP_EMPTY)
1647 case BKW_MMAP_EMPTY: {
1648 if (state != BKW_MMAP_RUNNING)
1654 WARN_ONCE(1, "Shouldn't get there\n");
1657 evlist->bkw_mmap_state = state;
1661 perf_evlist__pause(evlist);
1664 perf_evlist__resume(evlist);
1675 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1677 struct evsel *evsel;
1679 evlist__for_each_entry(evlist, evsel) {
1680 if (!evsel->core.attr.exclude_kernel)
1688 * Events in data file are not collect in groups, but we still want
1689 * the group display. Set the artificial group and set the leader's
1690 * forced_leader flag to notify the display code.
1692 void perf_evlist__force_leader(struct evlist *evlist)
1694 if (!evlist->nr_groups) {
1695 struct evsel *leader = evlist__first(evlist);
1697 perf_evlist__set_leader(evlist);
1698 leader->forced_leader = true;
1702 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1703 struct evsel *evsel,
1706 struct evsel *c2, *leader;
1707 bool is_open = true;
1709 leader = evsel->leader;
1710 pr_debug("Weak group for %s/%d failed\n",
1711 leader->name, leader->core.nr_members);
1714 * for_each_group_member doesn't work here because it doesn't
1715 * include the first entry.
1717 evlist__for_each_entry(evsel_list, c2) {
1720 if (c2->leader == leader) {
1721 if (is_open && close)
1722 perf_evsel__close(&c2->core);
1724 c2->core.nr_members = 0;
1726 * Set this for all former members of the group
1727 * to indicate they get reopened.
1729 c2->reset_group = true;
1735 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1740 if (strncmp(str, "fifo:", 5))
1744 if (!*str || *str == ',')
1756 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1757 * end of a FIFO to be repeatedly opened and closed.
1759 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1761 pr_err("Failed to open '%s'\n", s);
1766 *ctl_fd_close = true;
1769 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1770 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1772 pr_err("Failed to open '%s'\n", p);
1784 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1786 char *comma = NULL, *endptr = NULL;
1788 *ctl_fd_close = false;
1790 if (strncmp(str, "fd:", 3))
1791 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1793 *ctl_fd = strtoul(&str[3], &endptr, 0);
1794 if (endptr == &str[3])
1797 comma = strchr(str, ',');
1799 if (endptr != comma)
1802 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1803 if (endptr == comma + 1 || *endptr != '\0')
1810 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1812 if (*ctl_fd_close) {
1813 *ctl_fd_close = false;
1815 if (ctl_fd_ack >= 0)
1820 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1823 pr_debug("Control descriptor is not initialized\n");
1827 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1828 fdarray_flag__nonfilterable);
1829 if (evlist->ctl_fd.pos < 0) {
1830 evlist->ctl_fd.pos = -1;
1831 pr_err("Failed to add ctl fd entry: %m\n");
1835 evlist->ctl_fd.fd = fd;
1836 evlist->ctl_fd.ack = ack;
1841 bool evlist__ctlfd_initialized(struct evlist *evlist)
1843 return evlist->ctl_fd.pos >= 0;
1846 int evlist__finalize_ctlfd(struct evlist *evlist)
1848 struct pollfd *entries = evlist->core.pollfd.entries;
1850 if (!evlist__ctlfd_initialized(evlist))
1853 entries[evlist->ctl_fd.pos].fd = -1;
1854 entries[evlist->ctl_fd.pos].events = 0;
1855 entries[evlist->ctl_fd.pos].revents = 0;
1857 evlist->ctl_fd.pos = -1;
1858 evlist->ctl_fd.ack = -1;
1859 evlist->ctl_fd.fd = -1;
1864 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1865 char *cmd_data, size_t data_size)
1869 size_t bytes_read = 0;
1871 *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1872 memset(cmd_data, 0, data_size);
1876 err = read(evlist->ctl_fd.fd, &c, 1);
1878 if (c == '\n' || c == '\0')
1880 cmd_data[bytes_read++] = c;
1881 if (bytes_read == data_size)
1884 } else if (err == -1) {
1887 if (errno == EAGAIN || errno == EWOULDBLOCK)
1890 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1895 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1896 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1898 if (bytes_read > 0) {
1899 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1900 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1901 *cmd = EVLIST_CTL_CMD_ENABLE;
1902 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1903 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1904 *cmd = EVLIST_CTL_CMD_DISABLE;
1905 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
1906 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
1907 *cmd = EVLIST_CTL_CMD_SNAPSHOT;
1908 pr_debug("is snapshot\n");
1912 return bytes_read ? (int)bytes_read : err;
1915 int evlist__ctlfd_ack(struct evlist *evlist)
1919 if (evlist->ctl_fd.ack == -1)
1922 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
1923 sizeof(EVLIST_CTL_CMD_ACK_TAG));
1925 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
1930 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
1933 char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
1934 int ctlfd_pos = evlist->ctl_fd.pos;
1935 struct pollfd *entries = evlist->core.pollfd.entries;
1937 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
1940 if (entries[ctlfd_pos].revents & POLLIN) {
1941 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
1942 EVLIST_CTL_CMD_MAX_LEN);
1945 case EVLIST_CTL_CMD_ENABLE:
1946 evlist__enable(evlist);
1948 case EVLIST_CTL_CMD_DISABLE:
1949 evlist__disable(evlist);
1951 case EVLIST_CTL_CMD_SNAPSHOT:
1953 case EVLIST_CTL_CMD_ACK:
1954 case EVLIST_CTL_CMD_UNSUPPORTED:
1956 pr_debug("ctlfd: unsupported %d\n", *cmd);
1959 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
1960 *cmd == EVLIST_CTL_CMD_SNAPSHOT))
1961 evlist__ctlfd_ack(evlist);
1965 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
1966 evlist__finalize_ctlfd(evlist);
1968 entries[ctlfd_pos].revents = 0;
1973 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
1975 struct evsel *evsel;
1977 evlist__for_each_entry(evlist, evsel) {
1978 if (evsel->idx == idx)