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)
16 #include "thread_map.h"
18 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
19 #define GROUP_FD(group_fd, cpu) (*(int *)xyarray__entry(group_fd, cpu, 0))
21 int __perf_evsel__sample_size(u64 sample_type)
23 u64 mask = sample_type & PERF_SAMPLE_MASK;
27 for (i = 0; i < 64; i++) {
28 if (mask & (1ULL << i))
37 static void hists__init(struct hists *hists)
39 memset(hists, 0, sizeof(*hists));
40 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
41 hists->entries_in = &hists->entries_in_array[0];
42 hists->entries_collapsed = RB_ROOT;
43 hists->entries = RB_ROOT;
44 pthread_mutex_init(&hists->lock, NULL);
47 void perf_evsel__init(struct perf_evsel *evsel,
48 struct perf_event_attr *attr, int idx)
52 INIT_LIST_HEAD(&evsel->node);
53 hists__init(&evsel->hists);
56 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
58 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
61 perf_evsel__init(evsel, attr, idx);
66 void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts)
68 struct perf_event_attr *attr = &evsel->attr;
69 int track = !evsel->idx; /* only the first counter needs these */
71 attr->sample_id_all = opts->sample_id_all_avail ? 1 : 0;
72 attr->inherit = !opts->no_inherit;
73 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
74 PERF_FORMAT_TOTAL_TIME_RUNNING |
77 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
80 * We default some events to a 1 default interval. But keep
81 * it a weak assumption overridable by the user.
83 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
84 opts->user_interval != ULLONG_MAX)) {
86 attr->sample_type |= PERF_SAMPLE_PERIOD;
88 attr->sample_freq = opts->freq;
90 attr->sample_period = opts->default_interval;
95 attr->sample_freq = 0;
97 if (opts->inherit_stat)
98 attr->inherit_stat = 1;
100 if (opts->sample_address) {
101 attr->sample_type |= PERF_SAMPLE_ADDR;
102 attr->mmap_data = track;
105 if (opts->call_graph)
106 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
108 if (opts->system_wide)
109 attr->sample_type |= PERF_SAMPLE_CPU;
112 attr->sample_type |= PERF_SAMPLE_PERIOD;
114 if (opts->sample_id_all_avail &&
115 (opts->sample_time || opts->system_wide ||
116 !opts->no_inherit || opts->cpu_list))
117 attr->sample_type |= PERF_SAMPLE_TIME;
119 if (opts->raw_samples) {
120 attr->sample_type |= PERF_SAMPLE_TIME;
121 attr->sample_type |= PERF_SAMPLE_RAW;
122 attr->sample_type |= PERF_SAMPLE_CPU;
125 if (opts->no_delay) {
127 attr->wakeup_events = 1;
133 if (opts->target_pid == -1 && opts->target_tid == -1 && !opts->system_wide) {
135 attr->enable_on_exec = 1;
139 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
142 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
145 for (cpu = 0; cpu < ncpus; cpu++) {
146 for (thread = 0; thread < nthreads; thread++) {
147 FD(evsel, cpu, thread) = -1;
152 return evsel->fd != NULL ? 0 : -ENOMEM;
155 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
157 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
158 if (evsel->sample_id == NULL)
161 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
162 if (evsel->id == NULL) {
163 xyarray__delete(evsel->sample_id);
164 evsel->sample_id = NULL;
171 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
173 evsel->counts = zalloc((sizeof(*evsel->counts) +
174 (ncpus * sizeof(struct perf_counts_values))));
175 return evsel->counts != NULL ? 0 : -ENOMEM;
178 void perf_evsel__free_fd(struct perf_evsel *evsel)
180 xyarray__delete(evsel->fd);
184 void perf_evsel__free_id(struct perf_evsel *evsel)
186 xyarray__delete(evsel->sample_id);
187 evsel->sample_id = NULL;
192 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
196 for (cpu = 0; cpu < ncpus; cpu++)
197 for (thread = 0; thread < nthreads; ++thread) {
198 close(FD(evsel, cpu, thread));
199 FD(evsel, cpu, thread) = -1;
203 void perf_evsel__exit(struct perf_evsel *evsel)
205 assert(list_empty(&evsel->node));
206 xyarray__delete(evsel->fd);
207 xyarray__delete(evsel->sample_id);
211 void perf_evsel__delete(struct perf_evsel *evsel)
213 perf_evsel__exit(evsel);
214 close_cgroup(evsel->cgrp);
219 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
220 int cpu, int thread, bool scale)
222 struct perf_counts_values count;
223 size_t nv = scale ? 3 : 1;
225 if (FD(evsel, cpu, thread) < 0)
228 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
231 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
237 else if (count.run < count.ena)
238 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
240 count.ena = count.run = 0;
242 evsel->counts->cpu[cpu] = count;
246 int __perf_evsel__read(struct perf_evsel *evsel,
247 int ncpus, int nthreads, bool scale)
249 size_t nv = scale ? 3 : 1;
251 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
253 aggr->val = aggr->ena = aggr->run = 0;
255 for (cpu = 0; cpu < ncpus; cpu++) {
256 for (thread = 0; thread < nthreads; thread++) {
257 if (FD(evsel, cpu, thread) < 0)
260 if (readn(FD(evsel, cpu, thread),
261 &count, nv * sizeof(u64)) < 0)
264 aggr->val += count.val;
266 aggr->ena += count.ena;
267 aggr->run += count.run;
272 evsel->counts->scaled = 0;
274 if (aggr->run == 0) {
275 evsel->counts->scaled = -1;
280 if (aggr->run < aggr->ena) {
281 evsel->counts->scaled = 1;
282 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
285 aggr->ena = aggr->run = 0;
290 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
291 struct thread_map *threads, bool group,
292 struct xyarray *group_fds)
295 unsigned long flags = 0;
298 if (evsel->fd == NULL &&
299 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
303 flags = PERF_FLAG_PID_CGROUP;
304 pid = evsel->cgrp->fd;
307 for (cpu = 0; cpu < cpus->nr; cpu++) {
308 int group_fd = group_fds ? GROUP_FD(group_fds, cpu) : -1;
310 for (thread = 0; thread < threads->nr; thread++) {
313 pid = threads->map[thread];
315 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
319 if (FD(evsel, cpu, thread) < 0) {
324 if (group && group_fd == -1)
325 group_fd = FD(evsel, cpu, thread);
333 while (--thread >= 0) {
334 close(FD(evsel, cpu, thread));
335 FD(evsel, cpu, thread) = -1;
337 thread = threads->nr;
338 } while (--cpu >= 0);
342 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
344 if (evsel->fd == NULL)
347 perf_evsel__close_fd(evsel, ncpus, nthreads);
348 perf_evsel__free_fd(evsel);
361 struct thread_map map;
363 } empty_thread_map = {
368 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
369 struct thread_map *threads, bool group,
370 struct xyarray *group_fd)
373 /* Work around old compiler warnings about strict aliasing */
374 cpus = &empty_cpu_map.map;
378 threads = &empty_thread_map.map;
380 return __perf_evsel__open(evsel, cpus, threads, group, group_fd);
383 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
384 struct cpu_map *cpus, bool group,
385 struct xyarray *group_fd)
387 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group,
391 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
392 struct thread_map *threads, bool group,
393 struct xyarray *group_fd)
395 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group,
399 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
400 struct perf_sample *sample)
402 const u64 *array = event->sample.array;
404 array += ((event->header.size -
405 sizeof(event->header)) / sizeof(u64)) - 1;
407 if (type & PERF_SAMPLE_CPU) {
408 u32 *p = (u32 *)array;
413 if (type & PERF_SAMPLE_STREAM_ID) {
414 sample->stream_id = *array;
418 if (type & PERF_SAMPLE_ID) {
423 if (type & PERF_SAMPLE_TIME) {
424 sample->time = *array;
428 if (type & PERF_SAMPLE_TID) {
429 u32 *p = (u32 *)array;
437 static bool sample_overlap(const union perf_event *event,
438 const void *offset, u64 size)
440 const void *base = event;
442 if (offset + size > base + event->header.size)
448 int perf_event__parse_sample(const union perf_event *event, u64 type,
449 int sample_size, bool sample_id_all,
450 struct perf_sample *data, bool swapped)
455 * used for cross-endian analysis. See git commit 65014ab3
456 * for why this goofiness is needed.
463 memset(data, 0, sizeof(*data));
464 data->cpu = data->pid = data->tid = -1;
465 data->stream_id = data->id = data->time = -1ULL;
467 if (event->header.type != PERF_RECORD_SAMPLE) {
470 return perf_event__parse_id_sample(event, type, data);
473 array = event->sample.array;
475 if (sample_size + sizeof(event->header) > event->header.size)
478 if (type & PERF_SAMPLE_IP) {
479 data->ip = event->ip.ip;
483 if (type & PERF_SAMPLE_TID) {
486 /* undo swap of u64, then swap on individual u32s */
487 u.val64 = bswap_64(u.val64);
488 u.val32[0] = bswap_32(u.val32[0]);
489 u.val32[1] = bswap_32(u.val32[1]);
492 data->pid = u.val32[0];
493 data->tid = u.val32[1];
497 if (type & PERF_SAMPLE_TIME) {
503 if (type & PERF_SAMPLE_ADDR) {
509 if (type & PERF_SAMPLE_ID) {
514 if (type & PERF_SAMPLE_STREAM_ID) {
515 data->stream_id = *array;
519 if (type & PERF_SAMPLE_CPU) {
523 /* undo swap of u64, then swap on individual u32s */
524 u.val64 = bswap_64(u.val64);
525 u.val32[0] = bswap_32(u.val32[0]);
528 data->cpu = u.val32[0];
532 if (type & PERF_SAMPLE_PERIOD) {
533 data->period = *array;
537 if (type & PERF_SAMPLE_READ) {
538 fprintf(stderr, "PERF_SAMPLE_READ is unsuported for now\n");
542 if (type & PERF_SAMPLE_CALLCHAIN) {
543 if (sample_overlap(event, array, sizeof(data->callchain->nr)))
546 data->callchain = (struct ip_callchain *)array;
548 if (sample_overlap(event, array, data->callchain->nr))
551 array += 1 + data->callchain->nr;
554 if (type & PERF_SAMPLE_RAW) {
558 if (WARN_ONCE(swapped,
559 "Endianness of raw data not corrected!\n")) {
560 /* undo swap of u64, then swap on individual u32s */
561 u.val64 = bswap_64(u.val64);
562 u.val32[0] = bswap_32(u.val32[0]);
563 u.val32[1] = bswap_32(u.val32[1]);
566 if (sample_overlap(event, array, sizeof(u32)))
569 data->raw_size = u.val32[0];
570 pdata = (void *) array + sizeof(u32);
572 if (sample_overlap(event, pdata, data->raw_size))
575 data->raw_data = (void *) pdata;
581 int perf_event__synthesize_sample(union perf_event *event, u64 type,
582 const struct perf_sample *sample,
588 * used for cross-endian analysis. See git commit 65014ab3
589 * for why this goofiness is needed.
596 array = event->sample.array;
598 if (type & PERF_SAMPLE_IP) {
599 event->ip.ip = sample->ip;
603 if (type & PERF_SAMPLE_TID) {
604 u.val32[0] = sample->pid;
605 u.val32[1] = sample->tid;
608 * Inverse of what is done in perf_event__parse_sample
610 u.val32[0] = bswap_32(u.val32[0]);
611 u.val32[1] = bswap_32(u.val32[1]);
612 u.val64 = bswap_64(u.val64);
619 if (type & PERF_SAMPLE_TIME) {
620 *array = sample->time;
624 if (type & PERF_SAMPLE_ADDR) {
625 *array = sample->addr;
629 if (type & PERF_SAMPLE_ID) {
634 if (type & PERF_SAMPLE_STREAM_ID) {
635 *array = sample->stream_id;
639 if (type & PERF_SAMPLE_CPU) {
640 u.val32[0] = sample->cpu;
643 * Inverse of what is done in perf_event__parse_sample
645 u.val32[0] = bswap_32(u.val32[0]);
646 u.val64 = bswap_64(u.val64);
652 if (type & PERF_SAMPLE_PERIOD) {
653 *array = sample->period;