1 // SPDX-License-Identifier: GPL-2.0
4 #include <sys/syscall.h>
5 #include <perf/evsel.h>
6 #include <perf/cpumap.h>
7 #include <perf/threadmap.h>
8 #include <linux/hash.h>
9 #include <linux/list.h>
10 #include <internal/evsel.h>
11 #include <linux/zalloc.h>
13 #include <internal/xyarray.h>
14 #include <internal/cpumap.h>
15 #include <internal/mmap.h>
16 #include <internal/threadmap.h>
17 #include <internal/lib.h>
18 #include <linux/string.h>
19 #include <sys/ioctl.h>
23 void perf_evsel__init(struct perf_evsel *evsel, struct perf_event_attr *attr,
26 INIT_LIST_HEAD(&evsel->node);
27 INIT_LIST_HEAD(&evsel->per_stream_periods);
30 evsel->leader = evsel;
33 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr)
35 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
38 perf_evsel__init(evsel, attr, 0);
43 void perf_evsel__delete(struct perf_evsel *evsel)
48 #define FD(_evsel, _cpu_map_idx, _thread) \
49 ((int *)xyarray__entry(_evsel->fd, _cpu_map_idx, _thread))
50 #define MMAP(_evsel, _cpu_map_idx, _thread) \
51 (_evsel->mmap ? ((struct perf_mmap *) xyarray__entry(_evsel->mmap, _cpu_map_idx, _thread)) \
54 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
56 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
61 for (idx = 0; idx < ncpus; idx++) {
62 for (thread = 0; thread < nthreads; thread++) {
63 int *fd = FD(evsel, idx, thread);
71 return evsel->fd != NULL ? 0 : -ENOMEM;
74 static int perf_evsel__alloc_mmap(struct perf_evsel *evsel, int ncpus, int nthreads)
76 evsel->mmap = xyarray__new(ncpus, nthreads, sizeof(struct perf_mmap));
78 return evsel->mmap != NULL ? 0 : -ENOMEM;
82 sys_perf_event_open(struct perf_event_attr *attr,
83 pid_t pid, struct perf_cpu cpu, int group_fd,
86 return syscall(__NR_perf_event_open, attr, pid, cpu.cpu, group_fd, flags);
89 static int get_group_fd(struct perf_evsel *evsel, int cpu_map_idx, int thread, int *group_fd)
91 struct perf_evsel *leader = evsel->leader;
94 if (evsel == leader) {
100 * Leader must be already processed/open,
106 fd = FD(leader, cpu_map_idx, thread);
107 if (fd == NULL || *fd == -1)
115 int perf_evsel__open(struct perf_evsel *evsel, struct perf_cpu_map *cpus,
116 struct perf_thread_map *threads)
119 int idx, thread, err = 0;
122 static struct perf_cpu_map *empty_cpu_map;
124 if (empty_cpu_map == NULL) {
125 empty_cpu_map = perf_cpu_map__new_any_cpu();
126 if (empty_cpu_map == NULL)
130 cpus = empty_cpu_map;
133 if (threads == NULL) {
134 static struct perf_thread_map *empty_thread_map;
136 if (empty_thread_map == NULL) {
137 empty_thread_map = perf_thread_map__new_dummy();
138 if (empty_thread_map == NULL)
142 threads = empty_thread_map;
145 if (evsel->fd == NULL &&
146 perf_evsel__alloc_fd(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
149 perf_cpu_map__for_each_cpu(cpu, idx, cpus) {
150 for (thread = 0; thread < threads->nr; thread++) {
151 int fd, group_fd, *evsel_fd;
153 evsel_fd = FD(evsel, idx, thread);
154 if (evsel_fd == NULL) {
159 err = get_group_fd(evsel, idx, thread, &group_fd);
163 fd = sys_perf_event_open(&evsel->attr,
164 threads->map[thread].pid,
177 perf_evsel__close(evsel);
182 static void perf_evsel__close_fd_cpu(struct perf_evsel *evsel, int cpu_map_idx)
186 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
187 int *fd = FD(evsel, cpu_map_idx, thread);
189 if (fd && *fd >= 0) {
196 void perf_evsel__close_fd(struct perf_evsel *evsel)
198 for (int idx = 0; idx < xyarray__max_x(evsel->fd); idx++)
199 perf_evsel__close_fd_cpu(evsel, idx);
202 void perf_evsel__free_fd(struct perf_evsel *evsel)
204 xyarray__delete(evsel->fd);
208 void perf_evsel__close(struct perf_evsel *evsel)
210 if (evsel->fd == NULL)
213 perf_evsel__close_fd(evsel);
214 perf_evsel__free_fd(evsel);
217 void perf_evsel__close_cpu(struct perf_evsel *evsel, int cpu_map_idx)
219 if (evsel->fd == NULL)
222 perf_evsel__close_fd_cpu(evsel, cpu_map_idx);
225 void perf_evsel__munmap(struct perf_evsel *evsel)
229 if (evsel->fd == NULL || evsel->mmap == NULL)
232 for (idx = 0; idx < xyarray__max_x(evsel->fd); idx++) {
233 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
234 int *fd = FD(evsel, idx, thread);
236 if (fd == NULL || *fd < 0)
239 perf_mmap__munmap(MMAP(evsel, idx, thread));
243 xyarray__delete(evsel->mmap);
247 int perf_evsel__mmap(struct perf_evsel *evsel, int pages)
249 int ret, idx, thread;
250 struct perf_mmap_param mp = {
251 .prot = PROT_READ | PROT_WRITE,
252 .mask = (pages * page_size) - 1,
255 if (evsel->fd == NULL || evsel->mmap)
258 if (perf_evsel__alloc_mmap(evsel, xyarray__max_x(evsel->fd), xyarray__max_y(evsel->fd)) < 0)
261 for (idx = 0; idx < xyarray__max_x(evsel->fd); idx++) {
262 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
263 int *fd = FD(evsel, idx, thread);
264 struct perf_mmap *map;
265 struct perf_cpu cpu = perf_cpu_map__cpu(evsel->cpus, idx);
267 if (fd == NULL || *fd < 0)
270 map = MMAP(evsel, idx, thread);
271 perf_mmap__init(map, NULL, false, NULL);
273 ret = perf_mmap__mmap(map, &mp, *fd, cpu);
275 perf_evsel__munmap(evsel);
284 void *perf_evsel__mmap_base(struct perf_evsel *evsel, int cpu_map_idx, int thread)
286 int *fd = FD(evsel, cpu_map_idx, thread);
288 if (fd == NULL || *fd < 0 || MMAP(evsel, cpu_map_idx, thread) == NULL)
291 return MMAP(evsel, cpu_map_idx, thread)->base;
294 int perf_evsel__read_size(struct perf_evsel *evsel)
296 u64 read_format = evsel->attr.read_format;
297 int entry = sizeof(u64); /* value */
301 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
304 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
307 if (read_format & PERF_FORMAT_ID)
308 entry += sizeof(u64);
310 if (read_format & PERF_FORMAT_LOST)
311 entry += sizeof(u64);
313 if (read_format & PERF_FORMAT_GROUP) {
314 nr = evsel->nr_members;
322 /* This only reads values for the leader */
323 static int perf_evsel__read_group(struct perf_evsel *evsel, int cpu_map_idx,
324 int thread, struct perf_counts_values *count)
326 size_t size = perf_evsel__read_size(evsel);
327 int *fd = FD(evsel, cpu_map_idx, thread);
328 u64 read_format = evsel->attr.read_format;
332 if (fd == NULL || *fd < 0)
335 data = calloc(1, size);
339 if (readn(*fd, data, size) <= 0) {
345 * This reads only the leader event intentionally since we don't have
346 * perf counts values for sibling events.
348 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
349 count->ena = data[idx++];
350 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
351 count->run = data[idx++];
353 /* value is always available */
354 count->val = data[idx++];
355 if (read_format & PERF_FORMAT_ID)
356 count->id = data[idx++];
357 if (read_format & PERF_FORMAT_LOST)
358 count->lost = data[idx++];
365 * The perf read format is very flexible. It needs to set the proper
366 * values according to the read format.
368 static void perf_evsel__adjust_values(struct perf_evsel *evsel, u64 *buf,
369 struct perf_counts_values *count)
371 u64 read_format = evsel->attr.read_format;
374 count->val = buf[n++];
376 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
377 count->ena = buf[n++];
379 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
380 count->run = buf[n++];
382 if (read_format & PERF_FORMAT_ID)
383 count->id = buf[n++];
385 if (read_format & PERF_FORMAT_LOST)
386 count->lost = buf[n++];
389 int perf_evsel__read(struct perf_evsel *evsel, int cpu_map_idx, int thread,
390 struct perf_counts_values *count)
392 size_t size = perf_evsel__read_size(evsel);
393 int *fd = FD(evsel, cpu_map_idx, thread);
394 u64 read_format = evsel->attr.read_format;
395 struct perf_counts_values buf;
397 memset(count, 0, sizeof(*count));
399 if (fd == NULL || *fd < 0)
402 if (read_format & PERF_FORMAT_GROUP)
403 return perf_evsel__read_group(evsel, cpu_map_idx, thread, count);
405 if (MMAP(evsel, cpu_map_idx, thread) &&
406 !(read_format & (PERF_FORMAT_ID | PERF_FORMAT_LOST)) &&
407 !perf_mmap__read_self(MMAP(evsel, cpu_map_idx, thread), count))
410 if (readn(*fd, buf.values, size) <= 0)
413 perf_evsel__adjust_values(evsel, buf.values, count);
417 static int perf_evsel__ioctl(struct perf_evsel *evsel, int ioc, void *arg,
418 int cpu_map_idx, int thread)
420 int *fd = FD(evsel, cpu_map_idx, thread);
422 if (fd == NULL || *fd < 0)
425 return ioctl(*fd, ioc, arg);
428 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
434 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
435 int err = perf_evsel__ioctl(evsel, ioc, arg, cpu_map_idx, thread);
444 int perf_evsel__enable_cpu(struct perf_evsel *evsel, int cpu_map_idx)
446 return perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, NULL, cpu_map_idx);
449 int perf_evsel__enable_thread(struct perf_evsel *evsel, int thread)
451 struct perf_cpu cpu __maybe_unused;
455 perf_cpu_map__for_each_cpu(cpu, idx, evsel->cpus) {
456 err = perf_evsel__ioctl(evsel, PERF_EVENT_IOC_ENABLE, NULL, idx, thread);
464 int perf_evsel__enable(struct perf_evsel *evsel)
469 for (i = 0; i < xyarray__max_x(evsel->fd) && !err; i++)
470 err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, NULL, i);
474 int perf_evsel__disable_cpu(struct perf_evsel *evsel, int cpu_map_idx)
476 return perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, NULL, cpu_map_idx);
479 int perf_evsel__disable(struct perf_evsel *evsel)
484 for (i = 0; i < xyarray__max_x(evsel->fd) && !err; i++)
485 err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, NULL, i);
489 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
493 for (i = 0; i < perf_cpu_map__nr(evsel->cpus) && !err; i++)
494 err = perf_evsel__run_ioctl(evsel,
495 PERF_EVENT_IOC_SET_FILTER,
500 struct perf_cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
505 struct perf_thread_map *perf_evsel__threads(struct perf_evsel *evsel)
507 return evsel->threads;
510 struct perf_event_attr *perf_evsel__attr(struct perf_evsel *evsel)
515 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
517 if (ncpus == 0 || nthreads == 0)
520 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
521 if (evsel->sample_id == NULL)
524 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
525 if (evsel->id == NULL) {
526 xyarray__delete(evsel->sample_id);
527 evsel->sample_id = NULL;
534 void perf_evsel__free_id(struct perf_evsel *evsel)
536 struct perf_sample_id_period *pos, *n;
538 xyarray__delete(evsel->sample_id);
539 evsel->sample_id = NULL;
543 perf_evsel_for_each_per_thread_period_safe(evsel, n, pos) {
544 list_del_init(&pos->node);
549 bool perf_evsel__attr_has_per_thread_sample_period(struct perf_evsel *evsel)
551 return (evsel->attr.sample_type & PERF_SAMPLE_READ) &&
552 (evsel->attr.sample_type & PERF_SAMPLE_TID) &&
556 u64 *perf_sample_id__get_period_storage(struct perf_sample_id *sid, u32 tid, bool per_thread)
558 struct hlist_head *head;
559 struct perf_sample_id_period *res;
565 hash = hash_32(tid, PERF_SAMPLE_ID__HLIST_BITS);
566 head = &sid->periods[hash];
568 hlist_for_each_entry(res, head, hnode)
572 if (sid->evsel == NULL)
575 res = zalloc(sizeof(struct perf_sample_id_period));
579 INIT_LIST_HEAD(&res->node);
582 list_add_tail(&res->node, &sid->evsel->per_stream_periods);
583 hlist_add_head(&res->hnode, &sid->periods[hash]);
588 void perf_counts_values__scale(struct perf_counts_values *count,
589 bool scale, __s8 *pscaled)
594 if (count->run == 0) {
597 } else if (count->run < count->ena) {
599 count->val = (u64)((double)count->val * count->ena / count->run);