2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
10 #include <lk/debugfs.h>
13 #include "thread_map.h"
19 #include "parse-events.h"
23 #include <linux/bitops.h>
24 #include <linux/hash.h>
26 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
27 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
29 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
30 struct thread_map *threads)
34 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
35 INIT_HLIST_HEAD(&evlist->heads[i]);
36 INIT_LIST_HEAD(&evlist->entries);
37 perf_evlist__set_maps(evlist, cpus, threads);
38 evlist->workload.pid = -1;
41 struct perf_evlist *perf_evlist__new(void)
43 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
46 perf_evlist__init(evlist, NULL, NULL);
51 void perf_evlist__config(struct perf_evlist *evlist,
52 struct perf_record_opts *opts)
54 struct perf_evsel *evsel;
56 * Set the evsel leader links before we configure attributes,
57 * since some might depend on this info.
60 perf_evlist__set_leader(evlist);
62 if (evlist->cpus->map[0] < 0)
63 opts->no_inherit = true;
65 list_for_each_entry(evsel, &evlist->entries, node) {
66 perf_evsel__config(evsel, opts);
68 if (evlist->nr_entries > 1)
69 perf_evsel__set_sample_id(evsel);
73 static void perf_evlist__purge(struct perf_evlist *evlist)
75 struct perf_evsel *pos, *n;
77 list_for_each_entry_safe(pos, n, &evlist->entries, node) {
78 list_del_init(&pos->node);
79 perf_evsel__delete(pos);
82 evlist->nr_entries = 0;
85 void perf_evlist__exit(struct perf_evlist *evlist)
90 evlist->pollfd = NULL;
93 void perf_evlist__delete(struct perf_evlist *evlist)
95 perf_evlist__purge(evlist);
96 perf_evlist__exit(evlist);
100 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
102 list_add_tail(&entry->node, &evlist->entries);
103 ++evlist->nr_entries;
106 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
107 struct list_head *list,
110 list_splice_tail(list, &evlist->entries);
111 evlist->nr_entries += nr_entries;
114 void __perf_evlist__set_leader(struct list_head *list)
116 struct perf_evsel *evsel, *leader;
118 leader = list_entry(list->next, struct perf_evsel, node);
119 evsel = list_entry(list->prev, struct perf_evsel, node);
121 leader->nr_members = evsel->idx - leader->idx + 1;
123 list_for_each_entry(evsel, list, node) {
124 evsel->leader = leader;
128 void perf_evlist__set_leader(struct perf_evlist *evlist)
130 if (evlist->nr_entries) {
131 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
132 __perf_evlist__set_leader(&evlist->entries);
136 int perf_evlist__add_default(struct perf_evlist *evlist)
138 struct perf_event_attr attr = {
139 .type = PERF_TYPE_HARDWARE,
140 .config = PERF_COUNT_HW_CPU_CYCLES,
142 struct perf_evsel *evsel;
144 event_attr_init(&attr);
146 evsel = perf_evsel__new(&attr, 0);
150 /* use strdup() because free(evsel) assumes name is allocated */
151 evsel->name = strdup("cycles");
155 perf_evlist__add(evlist, evsel);
158 perf_evsel__delete(evsel);
163 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
164 struct perf_event_attr *attrs, size_t nr_attrs)
166 struct perf_evsel *evsel, *n;
170 for (i = 0; i < nr_attrs; i++) {
171 evsel = perf_evsel__new(attrs + i, evlist->nr_entries + i);
173 goto out_delete_partial_list;
174 list_add_tail(&evsel->node, &head);
177 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
181 out_delete_partial_list:
182 list_for_each_entry_safe(evsel, n, &head, node)
183 perf_evsel__delete(evsel);
187 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
188 struct perf_event_attr *attrs, size_t nr_attrs)
192 for (i = 0; i < nr_attrs; i++)
193 event_attr_init(attrs + i);
195 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
199 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
201 struct perf_evsel *evsel;
203 list_for_each_entry(evsel, &evlist->entries, node) {
204 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
205 (int)evsel->attr.config == id)
212 int perf_evlist__add_newtp(struct perf_evlist *evlist,
213 const char *sys, const char *name, void *handler)
215 struct perf_evsel *evsel;
217 evsel = perf_evsel__newtp(sys, name, evlist->nr_entries);
221 evsel->handler.func = handler;
222 perf_evlist__add(evlist, evsel);
226 void perf_evlist__disable(struct perf_evlist *evlist)
229 struct perf_evsel *pos;
230 int nr_cpus = cpu_map__nr(evlist->cpus);
231 int nr_threads = thread_map__nr(evlist->threads);
233 for (cpu = 0; cpu < nr_cpus; cpu++) {
234 list_for_each_entry(pos, &evlist->entries, node) {
235 if (!perf_evsel__is_group_leader(pos))
237 for (thread = 0; thread < nr_threads; thread++)
238 ioctl(FD(pos, cpu, thread),
239 PERF_EVENT_IOC_DISABLE, 0);
244 void perf_evlist__enable(struct perf_evlist *evlist)
247 struct perf_evsel *pos;
248 int nr_cpus = cpu_map__nr(evlist->cpus);
249 int nr_threads = thread_map__nr(evlist->threads);
251 for (cpu = 0; cpu < nr_cpus; cpu++) {
252 list_for_each_entry(pos, &evlist->entries, node) {
253 if (!perf_evsel__is_group_leader(pos))
255 for (thread = 0; thread < nr_threads; thread++)
256 ioctl(FD(pos, cpu, thread),
257 PERF_EVENT_IOC_ENABLE, 0);
262 static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
264 int nr_cpus = cpu_map__nr(evlist->cpus);
265 int nr_threads = thread_map__nr(evlist->threads);
266 int nfds = nr_cpus * nr_threads * evlist->nr_entries;
267 evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
268 return evlist->pollfd != NULL ? 0 : -ENOMEM;
271 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
273 fcntl(fd, F_SETFL, O_NONBLOCK);
274 evlist->pollfd[evlist->nr_fds].fd = fd;
275 evlist->pollfd[evlist->nr_fds].events = POLLIN;
279 static void perf_evlist__id_hash(struct perf_evlist *evlist,
280 struct perf_evsel *evsel,
281 int cpu, int thread, u64 id)
284 struct perf_sample_id *sid = SID(evsel, cpu, thread);
288 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
289 hlist_add_head(&sid->node, &evlist->heads[hash]);
292 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
293 int cpu, int thread, u64 id)
295 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
296 evsel->id[evsel->ids++] = id;
299 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
300 struct perf_evsel *evsel,
301 int cpu, int thread, int fd)
303 u64 read_data[4] = { 0, };
304 int id_idx = 1; /* The first entry is the counter value */
306 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
307 read(fd, &read_data, sizeof(read_data)) == -1)
310 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
312 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
315 perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]);
319 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
321 struct hlist_head *head;
322 struct perf_sample_id *sid;
325 if (evlist->nr_entries == 1)
326 return perf_evlist__first(evlist);
328 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
329 head = &evlist->heads[hash];
331 hlist_for_each_entry(sid, head, node)
335 if (!perf_evlist__sample_id_all(evlist))
336 return perf_evlist__first(evlist);
341 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
343 struct perf_mmap *md = &evlist->mmap[idx];
344 unsigned int head = perf_mmap__read_head(md);
345 unsigned int old = md->prev;
346 unsigned char *data = md->base + page_size;
347 union perf_event *event = NULL;
349 if (evlist->overwrite) {
351 * If we're further behind than half the buffer, there's a chance
352 * the writer will bite our tail and mess up the samples under us.
354 * If we somehow ended up ahead of the head, we got messed up.
356 * In either case, truncate and restart at head.
358 int diff = head - old;
359 if (diff > md->mask / 2 || diff < 0) {
360 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
363 * head points to a known good entry, start there.
372 event = (union perf_event *)&data[old & md->mask];
373 size = event->header.size;
376 * Event straddles the mmap boundary -- header should always
377 * be inside due to u64 alignment of output.
379 if ((old & md->mask) + size != ((old + size) & md->mask)) {
380 unsigned int offset = old;
381 unsigned int len = min(sizeof(*event), size), cpy;
382 void *dst = &md->event_copy;
385 cpy = min(md->mask + 1 - (offset & md->mask), len);
386 memcpy(dst, &data[offset & md->mask], cpy);
392 event = &md->event_copy;
400 if (!evlist->overwrite)
401 perf_mmap__write_tail(md, old);
406 void perf_evlist__munmap(struct perf_evlist *evlist)
410 for (i = 0; i < evlist->nr_mmaps; i++) {
411 if (evlist->mmap[i].base != NULL) {
412 munmap(evlist->mmap[i].base, evlist->mmap_len);
413 evlist->mmap[i].base = NULL;
421 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
423 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
424 if (cpu_map__all(evlist->cpus))
425 evlist->nr_mmaps = thread_map__nr(evlist->threads);
426 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
427 return evlist->mmap != NULL ? 0 : -ENOMEM;
430 static int __perf_evlist__mmap(struct perf_evlist *evlist,
431 int idx, int prot, int mask, int fd)
433 evlist->mmap[idx].prev = 0;
434 evlist->mmap[idx].mask = mask;
435 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
437 if (evlist->mmap[idx].base == MAP_FAILED) {
438 evlist->mmap[idx].base = NULL;
442 perf_evlist__add_pollfd(evlist, fd);
446 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
448 struct perf_evsel *evsel;
450 int nr_cpus = cpu_map__nr(evlist->cpus);
451 int nr_threads = thread_map__nr(evlist->threads);
453 for (cpu = 0; cpu < nr_cpus; cpu++) {
456 for (thread = 0; thread < nr_threads; thread++) {
457 list_for_each_entry(evsel, &evlist->entries, node) {
458 int fd = FD(evsel, cpu, thread);
462 if (__perf_evlist__mmap(evlist, cpu,
463 prot, mask, output) < 0)
466 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
470 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
471 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
480 for (cpu = 0; cpu < nr_cpus; cpu++) {
481 if (evlist->mmap[cpu].base != NULL) {
482 munmap(evlist->mmap[cpu].base, evlist->mmap_len);
483 evlist->mmap[cpu].base = NULL;
489 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
491 struct perf_evsel *evsel;
493 int nr_threads = thread_map__nr(evlist->threads);
495 for (thread = 0; thread < nr_threads; thread++) {
498 list_for_each_entry(evsel, &evlist->entries, node) {
499 int fd = FD(evsel, 0, thread);
503 if (__perf_evlist__mmap(evlist, thread,
504 prot, mask, output) < 0)
507 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
511 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
512 perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
520 for (thread = 0; thread < nr_threads; thread++) {
521 if (evlist->mmap[thread].base != NULL) {
522 munmap(evlist->mmap[thread].base, evlist->mmap_len);
523 evlist->mmap[thread].base = NULL;
529 /** perf_evlist__mmap - Create per cpu maps to receive events
531 * @evlist - list of events
532 * @pages - map length in pages
533 * @overwrite - overwrite older events?
535 * If overwrite is false the user needs to signal event consuption using:
537 * struct perf_mmap *m = &evlist->mmap[cpu];
538 * unsigned int head = perf_mmap__read_head(m);
540 * perf_mmap__write_tail(m, head)
542 * Using perf_evlist__read_on_cpu does this automatically.
544 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
547 struct perf_evsel *evsel;
548 const struct cpu_map *cpus = evlist->cpus;
549 const struct thread_map *threads = evlist->threads;
550 int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
552 /* 512 kiB: default amount of unprivileged mlocked memory */
553 if (pages == UINT_MAX)
554 pages = (512 * 1024) / page_size;
555 else if (!is_power_of_2(pages))
558 mask = pages * page_size - 1;
560 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
563 if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
566 evlist->overwrite = overwrite;
567 evlist->mmap_len = (pages + 1) * page_size;
569 list_for_each_entry(evsel, &evlist->entries, node) {
570 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
571 evsel->sample_id == NULL &&
572 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
576 if (cpu_map__all(cpus))
577 return perf_evlist__mmap_per_thread(evlist, prot, mask);
579 return perf_evlist__mmap_per_cpu(evlist, prot, mask);
582 int perf_evlist__create_maps(struct perf_evlist *evlist,
583 struct perf_target *target)
585 evlist->threads = thread_map__new_str(target->pid, target->tid,
588 if (evlist->threads == NULL)
591 if (perf_target__has_task(target))
592 evlist->cpus = cpu_map__dummy_new();
593 else if (!perf_target__has_cpu(target) && !target->uses_mmap)
594 evlist->cpus = cpu_map__dummy_new();
596 evlist->cpus = cpu_map__new(target->cpu_list);
598 if (evlist->cpus == NULL)
599 goto out_delete_threads;
604 thread_map__delete(evlist->threads);
608 void perf_evlist__delete_maps(struct perf_evlist *evlist)
610 cpu_map__delete(evlist->cpus);
611 thread_map__delete(evlist->threads);
613 evlist->threads = NULL;
616 int perf_evlist__apply_filters(struct perf_evlist *evlist)
618 struct perf_evsel *evsel;
620 const int ncpus = cpu_map__nr(evlist->cpus),
621 nthreads = thread_map__nr(evlist->threads);
623 list_for_each_entry(evsel, &evlist->entries, node) {
624 if (evsel->filter == NULL)
627 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
635 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
637 struct perf_evsel *evsel;
639 const int ncpus = cpu_map__nr(evlist->cpus),
640 nthreads = thread_map__nr(evlist->threads);
642 list_for_each_entry(evsel, &evlist->entries, node) {
643 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
651 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
653 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
655 list_for_each_entry_continue(pos, &evlist->entries, node) {
656 if (first->attr.sample_type != pos->attr.sample_type)
663 u64 perf_evlist__sample_type(struct perf_evlist *evlist)
665 struct perf_evsel *first = perf_evlist__first(evlist);
666 return first->attr.sample_type;
669 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
671 struct perf_evsel *first = perf_evlist__first(evlist);
672 struct perf_sample *data;
676 if (!first->attr.sample_id_all)
679 sample_type = first->attr.sample_type;
681 if (sample_type & PERF_SAMPLE_TID)
682 size += sizeof(data->tid) * 2;
684 if (sample_type & PERF_SAMPLE_TIME)
685 size += sizeof(data->time);
687 if (sample_type & PERF_SAMPLE_ID)
688 size += sizeof(data->id);
690 if (sample_type & PERF_SAMPLE_STREAM_ID)
691 size += sizeof(data->stream_id);
693 if (sample_type & PERF_SAMPLE_CPU)
694 size += sizeof(data->cpu) * 2;
699 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
701 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
703 list_for_each_entry_continue(pos, &evlist->entries, node) {
704 if (first->attr.sample_id_all != pos->attr.sample_id_all)
711 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
713 struct perf_evsel *first = perf_evlist__first(evlist);
714 return first->attr.sample_id_all;
717 void perf_evlist__set_selected(struct perf_evlist *evlist,
718 struct perf_evsel *evsel)
720 evlist->selected = evsel;
723 void perf_evlist__close(struct perf_evlist *evlist)
725 struct perf_evsel *evsel;
726 int ncpus = cpu_map__nr(evlist->cpus);
727 int nthreads = thread_map__nr(evlist->threads);
729 list_for_each_entry_reverse(evsel, &evlist->entries, node)
730 perf_evsel__close(evsel, ncpus, nthreads);
733 int perf_evlist__open(struct perf_evlist *evlist)
735 struct perf_evsel *evsel;
738 list_for_each_entry(evsel, &evlist->entries, node) {
739 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
746 perf_evlist__close(evlist);
751 int perf_evlist__prepare_workload(struct perf_evlist *evlist,
752 struct perf_target *target,
753 const char *argv[], bool pipe_output,
756 int child_ready_pipe[2], go_pipe[2];
759 if (pipe(child_ready_pipe) < 0) {
760 perror("failed to create 'ready' pipe");
764 if (pipe(go_pipe) < 0) {
765 perror("failed to create 'go' pipe");
766 goto out_close_ready_pipe;
769 evlist->workload.pid = fork();
770 if (evlist->workload.pid < 0) {
771 perror("failed to fork");
772 goto out_close_pipes;
775 if (!evlist->workload.pid) {
779 close(child_ready_pipe[0]);
781 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
784 * Do a dummy execvp to get the PLT entry resolved,
785 * so we avoid the resolver overhead on the real
788 execvp("", (char **)argv);
791 * Tell the parent we're ready to go
793 close(child_ready_pipe[1]);
796 * Wait until the parent tells us to go.
798 if (read(go_pipe[0], &bf, 1) == -1)
799 perror("unable to read pipe");
801 execvp(argv[0], (char **)argv);
805 kill(getppid(), SIGUSR1);
809 if (perf_target__none(target))
810 evlist->threads->map[0] = evlist->workload.pid;
812 close(child_ready_pipe[1]);
815 * wait for child to settle
817 if (read(child_ready_pipe[0], &bf, 1) == -1) {
818 perror("unable to read pipe");
819 goto out_close_pipes;
822 evlist->workload.cork_fd = go_pipe[1];
823 close(child_ready_pipe[0]);
829 out_close_ready_pipe:
830 close(child_ready_pipe[0]);
831 close(child_ready_pipe[1]);
835 int perf_evlist__start_workload(struct perf_evlist *evlist)
837 if (evlist->workload.cork_fd > 0) {
839 * Remove the cork, let it rip!
841 return close(evlist->workload.cork_fd);
847 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
848 struct perf_sample *sample)
850 struct perf_evsel *evsel = perf_evlist__first(evlist);
851 return perf_evsel__parse_sample(evsel, event, sample);
854 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
856 struct perf_evsel *evsel;
859 list_for_each_entry(evsel, &evlist->entries, node) {
860 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
861 perf_evsel__name(evsel));
864 return printed + fprintf(fp, "\n");;