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)
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(struct cpu_map *cpus,
42 struct thread_map *threads)
44 struct perf_evlist *evlist = zalloc(sizeof(*evlist));
47 perf_evlist__init(evlist, cpus, threads);
52 void perf_evlist__config_attrs(struct perf_evlist *evlist,
53 struct perf_record_opts *opts)
55 struct perf_evsel *evsel, *first;
57 if (evlist->cpus->map[0] < 0)
58 opts->no_inherit = true;
60 first = perf_evlist__first(evlist);
62 list_for_each_entry(evsel, &evlist->entries, node) {
63 perf_evsel__config(evsel, opts, first);
65 if (evlist->nr_entries > 1)
66 evsel->attr.sample_type |= PERF_SAMPLE_ID;
70 static void perf_evlist__purge(struct perf_evlist *evlist)
72 struct perf_evsel *pos, *n;
74 list_for_each_entry_safe(pos, n, &evlist->entries, node) {
75 list_del_init(&pos->node);
76 perf_evsel__delete(pos);
79 evlist->nr_entries = 0;
82 void perf_evlist__exit(struct perf_evlist *evlist)
87 evlist->pollfd = NULL;
90 void perf_evlist__delete(struct perf_evlist *evlist)
92 perf_evlist__purge(evlist);
93 perf_evlist__exit(evlist);
97 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
99 list_add_tail(&entry->node, &evlist->entries);
100 ++evlist->nr_entries;
103 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
104 struct list_head *list,
107 list_splice_tail(list, &evlist->entries);
108 evlist->nr_entries += nr_entries;
111 void __perf_evlist__set_leader(struct list_head *list)
113 struct perf_evsel *evsel, *leader;
115 leader = list_entry(list->next, struct perf_evsel, node);
116 leader->leader = NULL;
118 list_for_each_entry(evsel, list, node) {
120 evsel->leader = leader;
124 void perf_evlist__set_leader(struct perf_evlist *evlist)
126 if (evlist->nr_entries)
127 __perf_evlist__set_leader(&evlist->entries);
130 int perf_evlist__add_default(struct perf_evlist *evlist)
132 struct perf_event_attr attr = {
133 .type = PERF_TYPE_HARDWARE,
134 .config = PERF_COUNT_HW_CPU_CYCLES,
136 struct perf_evsel *evsel;
138 event_attr_init(&attr);
140 evsel = perf_evsel__new(&attr, 0);
144 /* use strdup() because free(evsel) assumes name is allocated */
145 evsel->name = strdup("cycles");
149 perf_evlist__add(evlist, evsel);
152 perf_evsel__delete(evsel);
157 int perf_evlist__add_attrs(struct perf_evlist *evlist,
158 struct perf_event_attr *attrs, size_t nr_attrs)
160 struct perf_evsel *evsel, *n;
164 for (i = 0; i < nr_attrs; i++) {
165 evsel = perf_evsel__new(attrs + i, evlist->nr_entries + i);
167 goto out_delete_partial_list;
168 list_add_tail(&evsel->node, &head);
171 perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
175 out_delete_partial_list:
176 list_for_each_entry_safe(evsel, n, &head, node)
177 perf_evsel__delete(evsel);
181 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
182 struct perf_event_attr *attrs, size_t nr_attrs)
186 for (i = 0; i < nr_attrs; i++)
187 event_attr_init(attrs + i);
189 return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
192 static int trace_event__id(const char *evname)
194 char *filename, *colon;
197 if (asprintf(&filename, "%s/%s/id", tracing_events_path, evname) < 0)
200 colon = strrchr(filename, ':');
204 fd = open(filename, O_RDONLY);
207 if (read(fd, id, sizeof(id)) > 0)
216 int perf_evlist__add_tracepoints(struct perf_evlist *evlist,
217 const char *tracepoints[],
218 size_t nr_tracepoints)
222 struct perf_event_attr *attrs = zalloc(nr_tracepoints * sizeof(*attrs));
227 for (i = 0; i < nr_tracepoints; i++) {
228 err = trace_event__id(tracepoints[i]);
233 attrs[i].type = PERF_TYPE_TRACEPOINT;
234 attrs[i].config = err;
235 attrs[i].sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
236 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD);
237 attrs[i].sample_period = 1;
240 err = perf_evlist__add_attrs(evlist, attrs, nr_tracepoints);
247 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
249 struct perf_evsel *evsel;
251 list_for_each_entry(evsel, &evlist->entries, node) {
252 if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
253 (int)evsel->attr.config == id)
260 int perf_evlist__set_tracepoints_handlers(struct perf_evlist *evlist,
261 const struct perf_evsel_str_handler *assocs,
264 struct perf_evsel *evsel;
268 for (i = 0; i < nr_assocs; i++) {
269 err = trace_event__id(assocs[i].name);
273 evsel = perf_evlist__find_tracepoint_by_id(evlist, err);
278 if (evsel->handler.func != NULL)
280 evsel->handler.func = assocs[i].handler;
288 void perf_evlist__disable(struct perf_evlist *evlist)
291 struct perf_evsel *pos;
293 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
294 list_for_each_entry(pos, &evlist->entries, node) {
295 for (thread = 0; thread < evlist->threads->nr; thread++)
296 ioctl(FD(pos, cpu, thread),
297 PERF_EVENT_IOC_DISABLE, 0);
302 void perf_evlist__enable(struct perf_evlist *evlist)
305 struct perf_evsel *pos;
307 for (cpu = 0; cpu < cpu_map__nr(evlist->cpus); cpu++) {
308 list_for_each_entry(pos, &evlist->entries, node) {
309 for (thread = 0; thread < evlist->threads->nr; thread++)
310 ioctl(FD(pos, cpu, thread),
311 PERF_EVENT_IOC_ENABLE, 0);
316 static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
318 int nfds = cpu_map__nr(evlist->cpus) * evlist->threads->nr * evlist->nr_entries;
319 evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
320 return evlist->pollfd != NULL ? 0 : -ENOMEM;
323 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
325 fcntl(fd, F_SETFL, O_NONBLOCK);
326 evlist->pollfd[evlist->nr_fds].fd = fd;
327 evlist->pollfd[evlist->nr_fds].events = POLLIN;
331 static void perf_evlist__id_hash(struct perf_evlist *evlist,
332 struct perf_evsel *evsel,
333 int cpu, int thread, u64 id)
336 struct perf_sample_id *sid = SID(evsel, cpu, thread);
340 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
341 hlist_add_head(&sid->node, &evlist->heads[hash]);
344 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
345 int cpu, int thread, u64 id)
347 perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
348 evsel->id[evsel->ids++] = id;
351 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
352 struct perf_evsel *evsel,
353 int cpu, int thread, int fd)
355 u64 read_data[4] = { 0, };
356 int id_idx = 1; /* The first entry is the counter value */
358 if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
359 read(fd, &read_data, sizeof(read_data)) == -1)
362 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
364 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
367 perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]);
371 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
373 struct hlist_head *head;
374 struct hlist_node *pos;
375 struct perf_sample_id *sid;
378 if (evlist->nr_entries == 1)
379 return perf_evlist__first(evlist);
381 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
382 head = &evlist->heads[hash];
384 hlist_for_each_entry(sid, pos, head, node)
388 if (!perf_evlist__sample_id_all(evlist))
389 return perf_evlist__first(evlist);
394 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
396 /* XXX Move this to perf.c, making it generally available */
397 unsigned int page_size = sysconf(_SC_PAGE_SIZE);
398 struct perf_mmap *md = &evlist->mmap[idx];
399 unsigned int head = perf_mmap__read_head(md);
400 unsigned int old = md->prev;
401 unsigned char *data = md->base + page_size;
402 union perf_event *event = NULL;
404 if (evlist->overwrite) {
406 * If we're further behind than half the buffer, there's a chance
407 * the writer will bite our tail and mess up the samples under us.
409 * If we somehow ended up ahead of the head, we got messed up.
411 * In either case, truncate and restart at head.
413 int diff = head - old;
414 if (diff > md->mask / 2 || diff < 0) {
415 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
418 * head points to a known good entry, start there.
427 event = (union perf_event *)&data[old & md->mask];
428 size = event->header.size;
431 * Event straddles the mmap boundary -- header should always
432 * be inside due to u64 alignment of output.
434 if ((old & md->mask) + size != ((old + size) & md->mask)) {
435 unsigned int offset = old;
436 unsigned int len = min(sizeof(*event), size), cpy;
437 void *dst = &evlist->event_copy;
440 cpy = min(md->mask + 1 - (offset & md->mask), len);
441 memcpy(dst, &data[offset & md->mask], cpy);
447 event = &evlist->event_copy;
455 if (!evlist->overwrite)
456 perf_mmap__write_tail(md, old);
461 void perf_evlist__munmap(struct perf_evlist *evlist)
465 for (i = 0; i < evlist->nr_mmaps; i++) {
466 if (evlist->mmap[i].base != NULL) {
467 munmap(evlist->mmap[i].base, evlist->mmap_len);
468 evlist->mmap[i].base = NULL;
476 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
478 evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
479 if (cpu_map__all(evlist->cpus))
480 evlist->nr_mmaps = evlist->threads->nr;
481 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
482 return evlist->mmap != NULL ? 0 : -ENOMEM;
485 static int __perf_evlist__mmap(struct perf_evlist *evlist,
486 int idx, int prot, int mask, int fd)
488 evlist->mmap[idx].prev = 0;
489 evlist->mmap[idx].mask = mask;
490 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
492 if (evlist->mmap[idx].base == MAP_FAILED) {
493 evlist->mmap[idx].base = NULL;
497 perf_evlist__add_pollfd(evlist, fd);
501 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
503 struct perf_evsel *evsel;
506 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
509 for (thread = 0; thread < evlist->threads->nr; thread++) {
510 list_for_each_entry(evsel, &evlist->entries, node) {
511 int fd = FD(evsel, cpu, thread);
515 if (__perf_evlist__mmap(evlist, cpu,
516 prot, mask, output) < 0)
519 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
523 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
524 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
533 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
534 if (evlist->mmap[cpu].base != NULL) {
535 munmap(evlist->mmap[cpu].base, evlist->mmap_len);
536 evlist->mmap[cpu].base = NULL;
542 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
544 struct perf_evsel *evsel;
547 for (thread = 0; thread < evlist->threads->nr; thread++) {
550 list_for_each_entry(evsel, &evlist->entries, node) {
551 int fd = FD(evsel, 0, thread);
555 if (__perf_evlist__mmap(evlist, thread,
556 prot, mask, output) < 0)
559 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
563 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
564 perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
572 for (thread = 0; thread < evlist->threads->nr; thread++) {
573 if (evlist->mmap[thread].base != NULL) {
574 munmap(evlist->mmap[thread].base, evlist->mmap_len);
575 evlist->mmap[thread].base = NULL;
581 /** perf_evlist__mmap - Create per cpu maps to receive events
583 * @evlist - list of events
584 * @pages - map length in pages
585 * @overwrite - overwrite older events?
587 * If overwrite is false the user needs to signal event consuption using:
589 * struct perf_mmap *m = &evlist->mmap[cpu];
590 * unsigned int head = perf_mmap__read_head(m);
592 * perf_mmap__write_tail(m, head)
594 * Using perf_evlist__read_on_cpu does this automatically.
596 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
599 unsigned int page_size = sysconf(_SC_PAGE_SIZE);
600 struct perf_evsel *evsel;
601 const struct cpu_map *cpus = evlist->cpus;
602 const struct thread_map *threads = evlist->threads;
603 int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
605 /* 512 kiB: default amount of unprivileged mlocked memory */
606 if (pages == UINT_MAX)
607 pages = (512 * 1024) / page_size;
608 else if (!is_power_of_2(pages))
611 mask = pages * page_size - 1;
613 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
616 if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
619 evlist->overwrite = overwrite;
620 evlist->mmap_len = (pages + 1) * page_size;
622 list_for_each_entry(evsel, &evlist->entries, node) {
623 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
624 evsel->sample_id == NULL &&
625 perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
629 if (cpu_map__all(cpus))
630 return perf_evlist__mmap_per_thread(evlist, prot, mask);
632 return perf_evlist__mmap_per_cpu(evlist, prot, mask);
635 int perf_evlist__create_maps(struct perf_evlist *evlist,
636 struct perf_target *target)
638 evlist->threads = thread_map__new_str(target->pid, target->tid,
641 if (evlist->threads == NULL)
644 if (perf_target__has_task(target))
645 evlist->cpus = cpu_map__dummy_new();
646 else if (!perf_target__has_cpu(target) && !target->uses_mmap)
647 evlist->cpus = cpu_map__dummy_new();
649 evlist->cpus = cpu_map__new(target->cpu_list);
651 if (evlist->cpus == NULL)
652 goto out_delete_threads;
657 thread_map__delete(evlist->threads);
661 void perf_evlist__delete_maps(struct perf_evlist *evlist)
663 cpu_map__delete(evlist->cpus);
664 thread_map__delete(evlist->threads);
666 evlist->threads = NULL;
669 int perf_evlist__set_filters(struct perf_evlist *evlist)
671 const struct thread_map *threads = evlist->threads;
672 const struct cpu_map *cpus = evlist->cpus;
673 struct perf_evsel *evsel;
680 list_for_each_entry(evsel, &evlist->entries, node) {
681 filter = evsel->filter;
684 for (cpu = 0; cpu < cpus->nr; cpu++) {
685 for (thread = 0; thread < threads->nr; thread++) {
686 fd = FD(evsel, cpu, thread);
687 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
697 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
699 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
701 list_for_each_entry_continue(pos, &evlist->entries, node) {
702 if (first->attr.sample_type != pos->attr.sample_type)
709 u64 perf_evlist__sample_type(struct perf_evlist *evlist)
711 struct perf_evsel *first = perf_evlist__first(evlist);
712 return first->attr.sample_type;
715 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
717 struct perf_evsel *first = perf_evlist__first(evlist);
718 struct perf_sample *data;
722 if (!first->attr.sample_id_all)
725 sample_type = first->attr.sample_type;
727 if (sample_type & PERF_SAMPLE_TID)
728 size += sizeof(data->tid) * 2;
730 if (sample_type & PERF_SAMPLE_TIME)
731 size += sizeof(data->time);
733 if (sample_type & PERF_SAMPLE_ID)
734 size += sizeof(data->id);
736 if (sample_type & PERF_SAMPLE_STREAM_ID)
737 size += sizeof(data->stream_id);
739 if (sample_type & PERF_SAMPLE_CPU)
740 size += sizeof(data->cpu) * 2;
745 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
747 struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
749 list_for_each_entry_continue(pos, &evlist->entries, node) {
750 if (first->attr.sample_id_all != pos->attr.sample_id_all)
757 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
759 struct perf_evsel *first = perf_evlist__first(evlist);
760 return first->attr.sample_id_all;
763 void perf_evlist__set_selected(struct perf_evlist *evlist,
764 struct perf_evsel *evsel)
766 evlist->selected = evsel;
769 int perf_evlist__open(struct perf_evlist *evlist)
771 struct perf_evsel *evsel;
772 int err, ncpus, nthreads;
774 list_for_each_entry(evsel, &evlist->entries, node) {
775 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
782 ncpus = evlist->cpus ? evlist->cpus->nr : 1;
783 nthreads = evlist->threads ? evlist->threads->nr : 1;
785 list_for_each_entry_reverse(evsel, &evlist->entries, node)
786 perf_evsel__close(evsel, ncpus, nthreads);
792 int perf_evlist__prepare_workload(struct perf_evlist *evlist,
793 struct perf_record_opts *opts,
796 int child_ready_pipe[2], go_pipe[2];
799 if (pipe(child_ready_pipe) < 0) {
800 perror("failed to create 'ready' pipe");
804 if (pipe(go_pipe) < 0) {
805 perror("failed to create 'go' pipe");
806 goto out_close_ready_pipe;
809 evlist->workload.pid = fork();
810 if (evlist->workload.pid < 0) {
811 perror("failed to fork");
812 goto out_close_pipes;
815 if (!evlist->workload.pid) {
816 if (opts->pipe_output)
819 close(child_ready_pipe[0]);
821 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
824 * Do a dummy execvp to get the PLT entry resolved,
825 * so we avoid the resolver overhead on the real
828 execvp("", (char **)argv);
831 * Tell the parent we're ready to go
833 close(child_ready_pipe[1]);
836 * Wait until the parent tells us to go.
838 if (read(go_pipe[0], &bf, 1) == -1)
839 perror("unable to read pipe");
841 execvp(argv[0], (char **)argv);
844 kill(getppid(), SIGUSR1);
848 if (perf_target__none(&opts->target))
849 evlist->threads->map[0] = evlist->workload.pid;
851 close(child_ready_pipe[1]);
854 * wait for child to settle
856 if (read(child_ready_pipe[0], &bf, 1) == -1) {
857 perror("unable to read pipe");
858 goto out_close_pipes;
861 evlist->workload.cork_fd = go_pipe[1];
862 close(child_ready_pipe[0]);
868 out_close_ready_pipe:
869 close(child_ready_pipe[0]);
870 close(child_ready_pipe[1]);
874 int perf_evlist__start_workload(struct perf_evlist *evlist)
876 if (evlist->workload.cork_fd > 0) {
878 * Remove the cork, let it rip!
880 return close(evlist->workload.cork_fd);
886 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
887 struct perf_sample *sample, bool swapped)
889 struct perf_evsel *evsel = perf_evlist__first(evlist);
890 return perf_evsel__parse_sample(evsel, event, sample, swapped);
893 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
895 struct perf_evsel *evsel;
898 list_for_each_entry(evsel, &evlist->entries, node) {
899 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
900 perf_evsel__name(evsel));
903 return printed + fprintf(fp, "\n");;