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
23 #include "bpf-event.h"
29 #include "parse-events.h"
30 #include <subcmd/parse-options.h>
33 #include <sys/ioctl.h>
36 #include <linux/bitops.h>
37 #include <linux/hash.h>
38 #include <linux/log2.h>
39 #include <linux/err.h>
40 #include <linux/string.h>
41 #include <linux/zalloc.h>
42 #include <perf/evlist.h>
43 #include <perf/evsel.h>
44 #include <perf/cpumap.h>
45 #include <perf/mmap.h>
47 #include <internal/xyarray.h>
49 #ifdef LACKS_SIGQUEUE_PROTOTYPE
50 int sigqueue(pid_t pid, int sig, const union sigval value);
53 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
54 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
56 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
57 struct perf_thread_map *threads)
59 perf_evlist__init(&evlist->core);
60 perf_evlist__set_maps(&evlist->core, cpus, threads);
61 evlist->workload.pid = -1;
62 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
65 struct evlist *evlist__new(void)
67 struct evlist *evlist = zalloc(sizeof(*evlist));
70 evlist__init(evlist, NULL, NULL);
75 struct evlist *perf_evlist__new_default(void)
77 struct evlist *evlist = evlist__new();
79 if (evlist && perf_evlist__add_default(evlist)) {
80 evlist__delete(evlist);
87 struct evlist *perf_evlist__new_dummy(void)
89 struct evlist *evlist = evlist__new();
91 if (evlist && perf_evlist__add_dummy(evlist)) {
92 evlist__delete(evlist);
100 * perf_evlist__set_id_pos - set the positions of event ids.
101 * @evlist: selected event list
103 * Events with compatible sample types all have the same id_pos
104 * and is_pos. For convenience, put a copy on evlist.
106 void perf_evlist__set_id_pos(struct evlist *evlist)
108 struct evsel *first = evlist__first(evlist);
110 evlist->id_pos = first->id_pos;
111 evlist->is_pos = first->is_pos;
114 static void perf_evlist__update_id_pos(struct evlist *evlist)
118 evlist__for_each_entry(evlist, evsel)
119 perf_evsel__calc_id_pos(evsel);
121 perf_evlist__set_id_pos(evlist);
124 static void evlist__purge(struct evlist *evlist)
126 struct evsel *pos, *n;
128 evlist__for_each_entry_safe(evlist, n, pos) {
129 list_del_init(&pos->core.node);
134 evlist->core.nr_entries = 0;
137 void evlist__exit(struct evlist *evlist)
139 zfree(&evlist->mmap);
140 zfree(&evlist->overwrite_mmap);
141 perf_evlist__exit(&evlist->core);
144 void evlist__delete(struct evlist *evlist)
149 evlist__munmap(evlist);
150 evlist__close(evlist);
151 evlist__purge(evlist);
152 evlist__exit(evlist);
156 void evlist__add(struct evlist *evlist, struct evsel *entry)
158 entry->evlist = evlist;
159 entry->idx = evlist->core.nr_entries;
160 entry->tracking = !entry->idx;
162 perf_evlist__add(&evlist->core, &entry->core);
164 if (evlist->core.nr_entries == 1)
165 perf_evlist__set_id_pos(evlist);
168 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
170 evsel->evlist = NULL;
171 perf_evlist__remove(&evlist->core, &evsel->core);
174 void perf_evlist__splice_list_tail(struct evlist *evlist,
175 struct list_head *list)
177 struct evsel *evsel, *temp;
179 __evlist__for_each_entry_safe(list, temp, evsel) {
180 list_del_init(&evsel->core.node);
181 evlist__add(evlist, evsel);
185 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
186 const struct evsel_str_handler *assocs, size_t nr_assocs)
192 for (i = 0; i < nr_assocs; i++) {
193 // Adding a handler for an event not in this evlist, just ignore it.
194 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
199 if (evsel->handler != NULL)
201 evsel->handler = assocs[i].handler;
209 void __perf_evlist__set_leader(struct list_head *list)
211 struct evsel *evsel, *leader;
213 leader = list_entry(list->next, struct evsel, core.node);
214 evsel = list_entry(list->prev, struct evsel, core.node);
216 leader->core.nr_members = evsel->idx - leader->idx + 1;
218 __evlist__for_each_entry(list, evsel) {
219 evsel->leader = leader;
223 void perf_evlist__set_leader(struct evlist *evlist)
225 if (evlist->core.nr_entries) {
226 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
227 __perf_evlist__set_leader(&evlist->core.entries);
231 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
233 struct evsel *evsel = perf_evsel__new_cycles(precise);
238 evlist__add(evlist, evsel);
242 int perf_evlist__add_dummy(struct evlist *evlist)
244 struct perf_event_attr attr = {
245 .type = PERF_TYPE_SOFTWARE,
246 .config = PERF_COUNT_SW_DUMMY,
247 .size = sizeof(attr), /* to capture ABI version */
249 struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
254 evlist__add(evlist, evsel);
258 static int evlist__add_attrs(struct evlist *evlist,
259 struct perf_event_attr *attrs, size_t nr_attrs)
261 struct evsel *evsel, *n;
265 for (i = 0; i < nr_attrs; i++) {
266 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
268 goto out_delete_partial_list;
269 list_add_tail(&evsel->core.node, &head);
272 perf_evlist__splice_list_tail(evlist, &head);
276 out_delete_partial_list:
277 __evlist__for_each_entry_safe(&head, n, evsel)
278 evsel__delete(evsel);
282 int __perf_evlist__add_default_attrs(struct evlist *evlist,
283 struct perf_event_attr *attrs, size_t nr_attrs)
287 for (i = 0; i < nr_attrs; i++)
288 event_attr_init(attrs + i);
290 return evlist__add_attrs(evlist, attrs, nr_attrs);
294 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
298 evlist__for_each_entry(evlist, evsel) {
299 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
300 (int)evsel->core.attr.config == id)
308 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
313 evlist__for_each_entry(evlist, evsel) {
314 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
315 (strcmp(evsel->name, name) == 0))
322 int perf_evlist__add_newtp(struct evlist *evlist,
323 const char *sys, const char *name, void *handler)
325 struct evsel *evsel = perf_evsel__newtp(sys, name);
330 evsel->handler = handler;
331 evlist__add(evlist, evsel);
335 static int perf_evlist__nr_threads(struct evlist *evlist,
338 if (evsel->core.system_wide)
341 return perf_thread_map__nr(evlist->core.threads);
344 void evlist__disable(struct evlist *evlist)
348 evlist__for_each_entry(evlist, pos) {
349 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
354 evlist->enabled = false;
357 void evlist__enable(struct evlist *evlist)
361 evlist__for_each_entry(evlist, pos) {
362 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
367 evlist->enabled = true;
370 void perf_evlist__toggle_enable(struct evlist *evlist)
372 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
375 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
376 struct evsel *evsel, int cpu)
379 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
384 for (thread = 0; thread < nr_threads; thread++) {
385 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
392 static int perf_evlist__enable_event_thread(struct evlist *evlist,
397 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
402 for (cpu = 0; cpu < nr_cpus; cpu++) {
403 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
410 int perf_evlist__enable_event_idx(struct evlist *evlist,
411 struct evsel *evsel, int idx)
413 bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
416 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
418 return perf_evlist__enable_event_thread(evlist, evsel, idx);
421 int evlist__add_pollfd(struct evlist *evlist, int fd)
423 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
426 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
428 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
431 int evlist__poll(struct evlist *evlist, int timeout)
433 return perf_evlist__poll(&evlist->core, timeout);
436 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
438 struct hlist_head *head;
439 struct perf_sample_id *sid;
442 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
443 head = &evlist->core.heads[hash];
445 hlist_for_each_entry(sid, head, node)
452 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
454 struct perf_sample_id *sid;
456 if (evlist->core.nr_entries == 1 || !id)
457 return evlist__first(evlist);
459 sid = perf_evlist__id2sid(evlist, id);
461 return container_of(sid->evsel, struct evsel, core);
463 if (!perf_evlist__sample_id_all(evlist))
464 return evlist__first(evlist);
469 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
472 struct perf_sample_id *sid;
477 sid = perf_evlist__id2sid(evlist, id);
479 return container_of(sid->evsel, struct evsel, core);
484 static int perf_evlist__event2id(struct evlist *evlist,
485 union perf_event *event, u64 *id)
487 const __u64 *array = event->sample.array;
490 n = (event->header.size - sizeof(event->header)) >> 3;
492 if (event->header.type == PERF_RECORD_SAMPLE) {
493 if (evlist->id_pos >= n)
495 *id = array[evlist->id_pos];
497 if (evlist->is_pos > n)
505 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
506 union perf_event *event)
508 struct evsel *first = evlist__first(evlist);
509 struct hlist_head *head;
510 struct perf_sample_id *sid;
514 if (evlist->core.nr_entries == 1)
517 if (!first->core.attr.sample_id_all &&
518 event->header.type != PERF_RECORD_SAMPLE)
521 if (perf_evlist__event2id(evlist, event, &id))
524 /* Synthesized events have an id of zero */
528 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
529 head = &evlist->core.heads[hash];
531 hlist_for_each_entry(sid, head, node) {
533 return container_of(sid->evsel, struct evsel, core);
538 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
542 if (!evlist->overwrite_mmap)
545 for (i = 0; i < evlist->core.nr_mmaps; i++) {
546 int fd = evlist->overwrite_mmap[i].core.fd;
551 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
558 static int perf_evlist__pause(struct evlist *evlist)
560 return perf_evlist__set_paused(evlist, true);
563 static int perf_evlist__resume(struct evlist *evlist)
565 return perf_evlist__set_paused(evlist, false);
568 static void evlist__munmap_nofree(struct evlist *evlist)
573 for (i = 0; i < evlist->core.nr_mmaps; i++)
574 perf_mmap__munmap(&evlist->mmap[i].core);
576 if (evlist->overwrite_mmap)
577 for (i = 0; i < evlist->core.nr_mmaps; i++)
578 perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
581 void evlist__munmap(struct evlist *evlist)
583 evlist__munmap_nofree(evlist);
584 zfree(&evlist->mmap);
585 zfree(&evlist->overwrite_mmap);
588 static void perf_mmap__unmap_cb(struct perf_mmap *map)
590 struct mmap *m = container_of(map, struct mmap, core);
595 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
601 evlist->core.nr_mmaps = perf_cpu_map__nr(evlist->core.cpus);
602 if (perf_cpu_map__empty(evlist->core.cpus))
603 evlist->core.nr_mmaps = perf_thread_map__nr(evlist->core.threads);
604 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
608 for (i = 0; i < evlist->core.nr_mmaps; i++) {
610 * When the perf_mmap() call is made we grab one refcount, plus
611 * one extra to let perf_mmap__consume() get the last
612 * events after all real references (perf_mmap__get()) are
615 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
616 * thus does perf_mmap__get() on it.
618 perf_mmap__init(&map[i].core, overwrite, perf_mmap__unmap_cb);
625 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
626 struct perf_mmap_param *_mp,
627 int idx, bool per_cpu)
629 struct evlist *evlist = container_of(_evlist, struct evlist, core);
630 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
632 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
635 static struct perf_mmap*
636 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
638 struct evlist *evlist = container_of(_evlist, struct evlist, core);
639 struct mmap *maps = evlist->mmap;
642 maps = evlist->overwrite_mmap;
645 maps = evlist__alloc_mmap(evlist, true);
649 evlist->overwrite_mmap = maps;
650 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
651 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
655 return &maps[idx].core;
659 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
662 struct mmap *map = container_of(_map, struct mmap, core);
663 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
665 return mmap__mmap(map, mp, output, cpu);
668 unsigned long perf_event_mlock_kb_in_pages(void)
673 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
675 * Pick a once upon a time good value, i.e. things look
676 * strange since we can't read a sysctl value, but lets not
681 max -= (page_size / 1024);
684 pages = (max * 1024) / page_size;
685 if (!is_power_of_2(pages))
686 pages = rounddown_pow_of_two(pages);
691 size_t evlist__mmap_size(unsigned long pages)
693 if (pages == UINT_MAX)
694 pages = perf_event_mlock_kb_in_pages();
695 else if (!is_power_of_2(pages))
698 return (pages + 1) * page_size;
701 static long parse_pages_arg(const char *str, unsigned long min,
704 unsigned long pages, val;
705 static struct parse_tag tags[] = {
706 { .tag = 'B', .mult = 1 },
707 { .tag = 'K', .mult = 1 << 10 },
708 { .tag = 'M', .mult = 1 << 20 },
709 { .tag = 'G', .mult = 1 << 30 },
716 val = parse_tag_value(str, tags);
717 if (val != (unsigned long) -1) {
718 /* we got file size value */
719 pages = PERF_ALIGN(val, page_size) / page_size;
721 /* we got pages count value */
723 pages = strtoul(str, &eptr, 10);
728 if (pages == 0 && min == 0) {
729 /* leave number of pages at 0 */
730 } else if (!is_power_of_2(pages)) {
733 /* round pages up to next power of 2 */
734 pages = roundup_pow_of_two(pages);
738 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
739 pr_info("rounding mmap pages size to %s (%lu pages)\n",
749 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
751 unsigned long max = UINT_MAX;
754 if (max > SIZE_MAX / page_size)
755 max = SIZE_MAX / page_size;
757 pages = parse_pages_arg(str, 1, max);
759 pr_err("Invalid argument for --mmap_pages/-m\n");
767 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
768 int unset __maybe_unused)
770 return __perf_evlist__parse_mmap_pages(opt->value, str);
774 * evlist__mmap_ex - Create mmaps to receive events.
775 * @evlist: list of events
776 * @pages: map length in pages
777 * @overwrite: overwrite older events?
778 * @auxtrace_pages - auxtrace map length in pages
779 * @auxtrace_overwrite - overwrite older auxtrace data?
781 * If @overwrite is %false the user needs to signal event consumption using
782 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
785 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
786 * consumption using auxtrace_mmap__write_tail().
788 * Return: %0 on success, negative error code otherwise.
790 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
791 unsigned int auxtrace_pages,
792 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
796 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
797 * Its value is decided by evsel's write_backward.
798 * So &mp should not be passed through const pointer.
800 struct mmap_params mp = {
801 .nr_cblocks = nr_cblocks,
802 .affinity = affinity,
804 .comp_level = comp_level
806 struct perf_evlist_mmap_ops ops = {
807 .idx = perf_evlist__mmap_cb_idx,
808 .get = perf_evlist__mmap_cb_get,
809 .mmap = perf_evlist__mmap_cb_mmap,
813 evlist->mmap = evlist__alloc_mmap(evlist, false);
817 evlist->core.mmap_len = evlist__mmap_size(pages);
818 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
819 mp.core.mask = evlist->core.mmap_len - page_size - 1;
821 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
822 auxtrace_pages, auxtrace_overwrite);
824 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
827 int evlist__mmap(struct evlist *evlist, unsigned int pages)
829 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
832 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
834 bool all_threads = (target->per_thread && target->system_wide);
835 struct perf_cpu_map *cpus;
836 struct perf_thread_map *threads;
839 * If specify '-a' and '--per-thread' to perf record, perf record
840 * will override '--per-thread'. target->per_thread = false and
841 * target->system_wide = true.
843 * If specify '--per-thread' only to perf record,
844 * target->per_thread = true and target->system_wide = false.
846 * So target->per_thread && target->system_wide is false.
847 * For perf record, thread_map__new_str doesn't call
848 * thread_map__new_all_cpus. That will keep perf record's
851 * For perf stat, it allows the case that target->per_thread and
852 * target->system_wide are all true. It means to collect system-wide
853 * per-thread data. thread_map__new_str will call
854 * thread_map__new_all_cpus to enumerate all threads.
856 threads = thread_map__new_str(target->pid, target->tid, target->uid,
862 if (target__uses_dummy_map(target))
863 cpus = perf_cpu_map__dummy_new();
865 cpus = perf_cpu_map__new(target->cpu_list);
868 goto out_delete_threads;
870 evlist->core.has_user_cpus = !!target->cpu_list;
872 perf_evlist__set_maps(&evlist->core, cpus, threads);
877 perf_thread_map__put(threads);
881 void __perf_evlist__set_sample_bit(struct evlist *evlist,
882 enum perf_event_sample_format bit)
886 evlist__for_each_entry(evlist, evsel)
887 __perf_evsel__set_sample_bit(evsel, bit);
890 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
891 enum perf_event_sample_format bit)
895 evlist__for_each_entry(evlist, evsel)
896 __perf_evsel__reset_sample_bit(evsel, bit);
899 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
904 evlist__for_each_entry(evlist, evsel) {
905 if (evsel->filter == NULL)
909 * filters only work for tracepoint event, which doesn't have cpu limit.
910 * So evlist and evsel should always be same.
912 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
922 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
930 evlist__for_each_entry(evlist, evsel) {
931 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
934 err = perf_evsel__set_filter(evsel, filter);
942 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
950 evlist__for_each_entry(evlist, evsel) {
951 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
954 err = perf_evsel__append_tp_filter(evsel, filter);
962 static char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
967 for (i = 0; i < npids; ++i) {
969 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
974 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
988 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
990 char *filter = asprintf__tp_filter_pids(npids, pids);
991 int ret = perf_evlist__set_tp_filter(evlist, filter);
997 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
999 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1002 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1004 char *filter = asprintf__tp_filter_pids(npids, pids);
1005 int ret = perf_evlist__append_tp_filter(evlist, filter);
1011 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1013 return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1016 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1020 if (evlist->core.nr_entries == 1)
1023 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1026 evlist__for_each_entry(evlist, pos) {
1027 if (pos->id_pos != evlist->id_pos ||
1028 pos->is_pos != evlist->is_pos)
1035 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1037 struct evsel *evsel;
1039 if (evlist->combined_sample_type)
1040 return evlist->combined_sample_type;
1042 evlist__for_each_entry(evlist, evsel)
1043 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1045 return evlist->combined_sample_type;
1048 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1050 evlist->combined_sample_type = 0;
1051 return __perf_evlist__combined_sample_type(evlist);
1054 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1056 struct evsel *evsel;
1057 u64 branch_type = 0;
1059 evlist__for_each_entry(evlist, evsel)
1060 branch_type |= evsel->core.attr.branch_sample_type;
1064 bool perf_evlist__valid_read_format(struct evlist *evlist)
1066 struct evsel *first = evlist__first(evlist), *pos = first;
1067 u64 read_format = first->core.attr.read_format;
1068 u64 sample_type = first->core.attr.sample_type;
1070 evlist__for_each_entry(evlist, pos) {
1071 if (read_format != pos->core.attr.read_format)
1075 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1076 if ((sample_type & PERF_SAMPLE_READ) &&
1077 !(read_format & PERF_FORMAT_ID)) {
1084 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1086 struct evsel *first = evlist__first(evlist);
1087 struct perf_sample *data;
1091 if (!first->core.attr.sample_id_all)
1094 sample_type = first->core.attr.sample_type;
1096 if (sample_type & PERF_SAMPLE_TID)
1097 size += sizeof(data->tid) * 2;
1099 if (sample_type & PERF_SAMPLE_TIME)
1100 size += sizeof(data->time);
1102 if (sample_type & PERF_SAMPLE_ID)
1103 size += sizeof(data->id);
1105 if (sample_type & PERF_SAMPLE_STREAM_ID)
1106 size += sizeof(data->stream_id);
1108 if (sample_type & PERF_SAMPLE_CPU)
1109 size += sizeof(data->cpu) * 2;
1111 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1112 size += sizeof(data->id);
1117 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1119 struct evsel *first = evlist__first(evlist), *pos = first;
1121 evlist__for_each_entry_continue(evlist, pos) {
1122 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1129 bool perf_evlist__sample_id_all(struct evlist *evlist)
1131 struct evsel *first = evlist__first(evlist);
1132 return first->core.attr.sample_id_all;
1135 void perf_evlist__set_selected(struct evlist *evlist,
1136 struct evsel *evsel)
1138 evlist->selected = evsel;
1141 void evlist__close(struct evlist *evlist)
1143 struct evsel *evsel;
1145 evlist__for_each_entry_reverse(evlist, evsel)
1146 evsel__close(evsel);
1149 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1151 struct perf_cpu_map *cpus;
1152 struct perf_thread_map *threads;
1156 * Try reading /sys/devices/system/cpu/online to get
1159 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1160 * code needs an overhaul to properly forward the
1161 * error, and we may not want to do that fallback to a
1162 * default cpu identity map :-\
1164 cpus = perf_cpu_map__new(NULL);
1168 threads = perf_thread_map__new_dummy();
1172 perf_evlist__set_maps(&evlist->core, cpus, threads);
1176 perf_cpu_map__put(cpus);
1180 int evlist__open(struct evlist *evlist)
1182 struct evsel *evsel;
1186 * Default: one fd per CPU, all threads, aka systemwide
1187 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1189 if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1190 err = perf_evlist__create_syswide_maps(evlist);
1195 perf_evlist__update_id_pos(evlist);
1197 evlist__for_each_entry(evlist, evsel) {
1198 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1205 evlist__close(evlist);
1210 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1211 const char *argv[], bool pipe_output,
1212 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1214 int child_ready_pipe[2], go_pipe[2];
1217 if (pipe(child_ready_pipe) < 0) {
1218 perror("failed to create 'ready' pipe");
1222 if (pipe(go_pipe) < 0) {
1223 perror("failed to create 'go' pipe");
1224 goto out_close_ready_pipe;
1227 evlist->workload.pid = fork();
1228 if (evlist->workload.pid < 0) {
1229 perror("failed to fork");
1230 goto out_close_pipes;
1233 if (!evlist->workload.pid) {
1239 signal(SIGTERM, SIG_DFL);
1241 close(child_ready_pipe[0]);
1243 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1246 * Tell the parent we're ready to go
1248 close(child_ready_pipe[1]);
1251 * Wait until the parent tells us to go.
1253 ret = read(go_pipe[0], &bf, 1);
1255 * The parent will ask for the execvp() to be performed by
1256 * writing exactly one byte, in workload.cork_fd, usually via
1257 * perf_evlist__start_workload().
1259 * For cancelling the workload without actually running it,
1260 * the parent will just close workload.cork_fd, without writing
1261 * anything, i.e. read will return zero and we just exit()
1266 perror("unable to read pipe");
1270 execvp(argv[0], (char **)argv);
1275 val.sival_int = errno;
1276 if (sigqueue(getppid(), SIGUSR1, val))
1284 struct sigaction act = {
1285 .sa_flags = SA_SIGINFO,
1286 .sa_sigaction = exec_error,
1288 sigaction(SIGUSR1, &act, NULL);
1291 if (target__none(target)) {
1292 if (evlist->core.threads == NULL) {
1293 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1294 __func__, __LINE__);
1295 goto out_close_pipes;
1297 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1300 close(child_ready_pipe[1]);
1303 * wait for child to settle
1305 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1306 perror("unable to read pipe");
1307 goto out_close_pipes;
1310 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1311 evlist->workload.cork_fd = go_pipe[1];
1312 close(child_ready_pipe[0]);
1318 out_close_ready_pipe:
1319 close(child_ready_pipe[0]);
1320 close(child_ready_pipe[1]);
1324 int perf_evlist__start_workload(struct evlist *evlist)
1326 if (evlist->workload.cork_fd > 0) {
1330 * Remove the cork, let it rip!
1332 ret = write(evlist->workload.cork_fd, &bf, 1);
1334 perror("unable to write to pipe");
1336 close(evlist->workload.cork_fd);
1343 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1344 struct perf_sample *sample)
1346 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1350 return perf_evsel__parse_sample(evsel, event, sample);
1353 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1354 union perf_event *event,
1357 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1361 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1364 int perf_evlist__strerror_open(struct evlist *evlist,
1365 int err, char *buf, size_t size)
1368 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1373 printed = scnprintf(buf, size,
1375 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1377 value = perf_event_paranoid();
1379 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1382 printed += scnprintf(buf + printed, size - printed,
1383 "For your workloads it needs to be <= 1\nHint:\t");
1385 printed += scnprintf(buf + printed, size - printed,
1386 "For system wide tracing it needs to be set to -1.\n");
1388 printed += scnprintf(buf + printed, size - printed,
1389 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1390 "Hint:\tThe current value is %d.", value);
1393 struct evsel *first = evlist__first(evlist);
1396 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1399 if (first->core.attr.sample_freq < (u64)max_freq)
1402 printed = scnprintf(buf, size,
1404 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1405 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1406 emsg, max_freq, first->core.attr.sample_freq);
1411 scnprintf(buf, size, "%s", emsg);
1418 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1420 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1421 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1425 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1426 printed += scnprintf(buf + printed, size - printed,
1428 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1429 "Hint:\tTried using %zd kB.\n",
1430 emsg, pages_max_per_user, pages_attempted);
1432 if (pages_attempted >= pages_max_per_user) {
1433 printed += scnprintf(buf + printed, size - printed,
1434 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1435 pages_max_per_user + pages_attempted);
1438 printed += scnprintf(buf + printed, size - printed,
1439 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1442 scnprintf(buf, size, "%s", emsg);
1449 void perf_evlist__to_front(struct evlist *evlist,
1450 struct evsel *move_evsel)
1452 struct evsel *evsel, *n;
1455 if (move_evsel == evlist__first(evlist))
1458 evlist__for_each_entry_safe(evlist, n, evsel) {
1459 if (evsel->leader == move_evsel->leader)
1460 list_move_tail(&evsel->core.node, &move);
1463 list_splice(&move, &evlist->core.entries);
1466 void perf_evlist__set_tracking_event(struct evlist *evlist,
1467 struct evsel *tracking_evsel)
1469 struct evsel *evsel;
1471 if (tracking_evsel->tracking)
1474 evlist__for_each_entry(evlist, evsel) {
1475 if (evsel != tracking_evsel)
1476 evsel->tracking = false;
1479 tracking_evsel->tracking = true;
1483 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1486 struct evsel *evsel;
1488 evlist__for_each_entry(evlist, evsel) {
1491 if (strcmp(str, evsel->name) == 0)
1498 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1499 enum bkw_mmap_state state)
1501 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1508 if (!evlist->overwrite_mmap)
1511 switch (old_state) {
1512 case BKW_MMAP_NOTREADY: {
1513 if (state != BKW_MMAP_RUNNING)
1517 case BKW_MMAP_RUNNING: {
1518 if (state != BKW_MMAP_DATA_PENDING)
1523 case BKW_MMAP_DATA_PENDING: {
1524 if (state != BKW_MMAP_EMPTY)
1528 case BKW_MMAP_EMPTY: {
1529 if (state != BKW_MMAP_RUNNING)
1535 WARN_ONCE(1, "Shouldn't get there\n");
1538 evlist->bkw_mmap_state = state;
1542 perf_evlist__pause(evlist);
1545 perf_evlist__resume(evlist);
1556 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1558 struct evsel *evsel;
1560 evlist__for_each_entry(evlist, evsel) {
1561 if (!evsel->core.attr.exclude_kernel)
1569 * Events in data file are not collect in groups, but we still want
1570 * the group display. Set the artificial group and set the leader's
1571 * forced_leader flag to notify the display code.
1573 void perf_evlist__force_leader(struct evlist *evlist)
1575 if (!evlist->nr_groups) {
1576 struct evsel *leader = evlist__first(evlist);
1578 perf_evlist__set_leader(evlist);
1579 leader->forced_leader = true;
1583 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1584 struct evsel *evsel)
1586 struct evsel *c2, *leader;
1587 bool is_open = true;
1589 leader = evsel->leader;
1590 pr_debug("Weak group for %s/%d failed\n",
1591 leader->name, leader->core.nr_members);
1594 * for_each_group_member doesn't work here because it doesn't
1595 * include the first entry.
1597 evlist__for_each_entry(evsel_list, c2) {
1600 if (c2->leader == leader) {
1602 perf_evsel__close(&evsel->core);
1604 c2->core.nr_members = 0;
1610 int perf_evlist__add_sb_event(struct evlist **evlist,
1611 struct perf_event_attr *attr,
1612 perf_evsel__sb_cb_t cb,
1615 struct evsel *evsel;
1616 bool new_evlist = (*evlist) == NULL;
1618 if (*evlist == NULL)
1619 *evlist = evlist__new();
1620 if (*evlist == NULL)
1623 if (!attr->sample_id_all) {
1624 pr_warning("enabling sample_id_all for all side band events\n");
1625 attr->sample_id_all = 1;
1628 evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1632 evsel->side_band.cb = cb;
1633 evsel->side_band.data = data;
1634 evlist__add(*evlist, evsel);
1639 evlist__delete(*evlist);
1645 static void *perf_evlist__poll_thread(void *arg)
1647 struct evlist *evlist = arg;
1648 bool draining = false;
1651 * In order to read symbols from other namespaces perf to needs to call
1652 * setns(2). This isn't permitted if the struct_fs has multiple users.
1653 * unshare(2) the fs so that we may continue to setns into namespaces
1654 * that we're observing when, for instance, reading the build-ids at
1655 * the end of a 'perf record' session.
1660 bool got_data = false;
1662 if (evlist->thread.done)
1666 evlist__poll(evlist, 1000);
1668 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1669 struct mmap *map = &evlist->mmap[i];
1670 union perf_event *event;
1672 if (perf_mmap__read_init(&map->core))
1674 while ((event = perf_mmap__read_event(&map->core)) != NULL) {
1675 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1677 if (evsel && evsel->side_band.cb)
1678 evsel->side_band.cb(event, evsel->side_band.data);
1680 pr_warning("cannot locate proper evsel for the side band event\n");
1682 perf_mmap__consume(&map->core);
1685 perf_mmap__read_done(&map->core);
1688 if (draining && !got_data)
1694 int perf_evlist__start_sb_thread(struct evlist *evlist,
1695 struct target *target)
1697 struct evsel *counter;
1702 if (perf_evlist__create_maps(evlist, target))
1703 goto out_delete_evlist;
1705 evlist__for_each_entry(evlist, counter) {
1706 if (evsel__open(counter, evlist->core.cpus,
1707 evlist->core.threads) < 0)
1708 goto out_delete_evlist;
1711 if (evlist__mmap(evlist, UINT_MAX))
1712 goto out_delete_evlist;
1714 evlist__for_each_entry(evlist, counter) {
1715 if (evsel__enable(counter))
1716 goto out_delete_evlist;
1719 evlist->thread.done = 0;
1720 if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1721 goto out_delete_evlist;
1726 evlist__delete(evlist);
1731 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1735 evlist->thread.done = 1;
1736 pthread_join(evlist->thread.th, NULL);
1737 evlist__delete(evlist);