xref: /openbmc/linux/tools/lib/perf/evlist.c (revision 5086ea4b)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <perf/evlist.h>
3 #include <perf/evsel.h>
4 #include <linux/bitops.h>
5 #include <linux/list.h>
6 #include <linux/hash.h>
7 #include <sys/ioctl.h>
8 #include <internal/evlist.h>
9 #include <internal/evsel.h>
10 #include <internal/xyarray.h>
11 #include <internal/mmap.h>
12 #include <internal/cpumap.h>
13 #include <internal/threadmap.h>
14 #include <internal/lib.h>
15 #include <linux/zalloc.h>
16 #include <stdlib.h>
17 #include <errno.h>
18 #include <unistd.h>
19 #include <fcntl.h>
20 #include <signal.h>
21 #include <poll.h>
22 #include <sys/mman.h>
23 #include <perf/cpumap.h>
24 #include <perf/threadmap.h>
25 #include <api/fd/array.h>
26 
27 void perf_evlist__init(struct perf_evlist *evlist)
28 {
29 	int i;
30 
31 	for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
32 		INIT_HLIST_HEAD(&evlist->heads[i]);
33 	INIT_LIST_HEAD(&evlist->entries);
34 	evlist->nr_entries = 0;
35 	fdarray__init(&evlist->pollfd, 64);
36 }
37 
38 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
39 					  struct perf_evsel *evsel)
40 {
41 	/*
42 	 * We already have cpus for evsel (via PMU sysfs) so
43 	 * keep it, if there's no target cpu list defined.
44 	 */
45 	if (!evsel->own_cpus || evlist->has_user_cpus) {
46 		perf_cpu_map__put(evsel->cpus);
47 		evsel->cpus = perf_cpu_map__get(evlist->cpus);
48 	} else if (!evsel->system_wide && perf_cpu_map__empty(evlist->cpus)) {
49 		perf_cpu_map__put(evsel->cpus);
50 		evsel->cpus = perf_cpu_map__get(evlist->cpus);
51 	} else if (evsel->cpus != evsel->own_cpus) {
52 		perf_cpu_map__put(evsel->cpus);
53 		evsel->cpus = perf_cpu_map__get(evsel->own_cpus);
54 	}
55 
56 	perf_thread_map__put(evsel->threads);
57 	evsel->threads = perf_thread_map__get(evlist->threads);
58 	evlist->all_cpus = perf_cpu_map__merge(evlist->all_cpus, evsel->cpus);
59 }
60 
61 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
62 {
63 	struct perf_evsel *evsel;
64 
65 	perf_evlist__for_each_evsel(evlist, evsel)
66 		__perf_evlist__propagate_maps(evlist, evsel);
67 }
68 
69 void perf_evlist__add(struct perf_evlist *evlist,
70 		      struct perf_evsel *evsel)
71 {
72 	list_add_tail(&evsel->node, &evlist->entries);
73 	evlist->nr_entries += 1;
74 	__perf_evlist__propagate_maps(evlist, evsel);
75 }
76 
77 void perf_evlist__remove(struct perf_evlist *evlist,
78 			 struct perf_evsel *evsel)
79 {
80 	list_del_init(&evsel->node);
81 	evlist->nr_entries -= 1;
82 }
83 
84 struct perf_evlist *perf_evlist__new(void)
85 {
86 	struct perf_evlist *evlist = zalloc(sizeof(*evlist));
87 
88 	if (evlist != NULL)
89 		perf_evlist__init(evlist);
90 
91 	return evlist;
92 }
93 
94 struct perf_evsel *
95 perf_evlist__next(struct perf_evlist *evlist, struct perf_evsel *prev)
96 {
97 	struct perf_evsel *next;
98 
99 	if (!prev) {
100 		next = list_first_entry(&evlist->entries,
101 					struct perf_evsel,
102 					node);
103 	} else {
104 		next = list_next_entry(prev, node);
105 	}
106 
107 	/* Empty list is noticed here so don't need checking on entry. */
108 	if (&next->node == &evlist->entries)
109 		return NULL;
110 
111 	return next;
112 }
113 
114 static void perf_evlist__purge(struct perf_evlist *evlist)
115 {
116 	struct perf_evsel *pos, *n;
117 
118 	perf_evlist__for_each_entry_safe(evlist, n, pos) {
119 		list_del_init(&pos->node);
120 		perf_evsel__delete(pos);
121 	}
122 
123 	evlist->nr_entries = 0;
124 }
125 
126 void perf_evlist__exit(struct perf_evlist *evlist)
127 {
128 	perf_cpu_map__put(evlist->cpus);
129 	perf_cpu_map__put(evlist->all_cpus);
130 	perf_thread_map__put(evlist->threads);
131 	evlist->cpus = NULL;
132 	evlist->all_cpus = NULL;
133 	evlist->threads = NULL;
134 	fdarray__exit(&evlist->pollfd);
135 }
136 
137 void perf_evlist__delete(struct perf_evlist *evlist)
138 {
139 	if (evlist == NULL)
140 		return;
141 
142 	perf_evlist__munmap(evlist);
143 	perf_evlist__close(evlist);
144 	perf_evlist__purge(evlist);
145 	perf_evlist__exit(evlist);
146 	free(evlist);
147 }
148 
149 void perf_evlist__set_maps(struct perf_evlist *evlist,
150 			   struct perf_cpu_map *cpus,
151 			   struct perf_thread_map *threads)
152 {
153 	/*
154 	 * Allow for the possibility that one or another of the maps isn't being
155 	 * changed i.e. don't put it.  Note we are assuming the maps that are
156 	 * being applied are brand new and evlist is taking ownership of the
157 	 * original reference count of 1.  If that is not the case it is up to
158 	 * the caller to increase the reference count.
159 	 */
160 	if (cpus != evlist->cpus) {
161 		perf_cpu_map__put(evlist->cpus);
162 		evlist->cpus = perf_cpu_map__get(cpus);
163 	}
164 
165 	if (threads != evlist->threads) {
166 		perf_thread_map__put(evlist->threads);
167 		evlist->threads = perf_thread_map__get(threads);
168 	}
169 
170 	if (!evlist->all_cpus && cpus)
171 		evlist->all_cpus = perf_cpu_map__get(cpus);
172 
173 	perf_evlist__propagate_maps(evlist);
174 }
175 
176 int perf_evlist__open(struct perf_evlist *evlist)
177 {
178 	struct perf_evsel *evsel;
179 	int err;
180 
181 	perf_evlist__for_each_entry(evlist, evsel) {
182 		err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
183 		if (err < 0)
184 			goto out_err;
185 	}
186 
187 	return 0;
188 
189 out_err:
190 	perf_evlist__close(evlist);
191 	return err;
192 }
193 
194 void perf_evlist__close(struct perf_evlist *evlist)
195 {
196 	struct perf_evsel *evsel;
197 
198 	perf_evlist__for_each_entry_reverse(evlist, evsel)
199 		perf_evsel__close(evsel);
200 }
201 
202 void perf_evlist__enable(struct perf_evlist *evlist)
203 {
204 	struct perf_evsel *evsel;
205 
206 	perf_evlist__for_each_entry(evlist, evsel)
207 		perf_evsel__enable(evsel);
208 }
209 
210 void perf_evlist__disable(struct perf_evlist *evlist)
211 {
212 	struct perf_evsel *evsel;
213 
214 	perf_evlist__for_each_entry(evlist, evsel)
215 		perf_evsel__disable(evsel);
216 }
217 
218 u64 perf_evlist__read_format(struct perf_evlist *evlist)
219 {
220 	struct perf_evsel *first = perf_evlist__first(evlist);
221 
222 	return first->attr.read_format;
223 }
224 
225 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
226 
227 static void perf_evlist__id_hash(struct perf_evlist *evlist,
228 				 struct perf_evsel *evsel,
229 				 int cpu, int thread, u64 id)
230 {
231 	int hash;
232 	struct perf_sample_id *sid = SID(evsel, cpu, thread);
233 
234 	sid->id = id;
235 	sid->evsel = evsel;
236 	hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
237 	hlist_add_head(&sid->node, &evlist->heads[hash]);
238 }
239 
240 void perf_evlist__id_add(struct perf_evlist *evlist,
241 			 struct perf_evsel *evsel,
242 			 int cpu, int thread, u64 id)
243 {
244 	perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
245 	evsel->id[evsel->ids++] = id;
246 }
247 
248 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
249 			   struct perf_evsel *evsel,
250 			   int cpu, int thread, int fd)
251 {
252 	u64 read_data[4] = { 0, };
253 	int id_idx = 1; /* The first entry is the counter value */
254 	u64 id;
255 	int ret;
256 
257 	ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
258 	if (!ret)
259 		goto add;
260 
261 	if (errno != ENOTTY)
262 		return -1;
263 
264 	/* Legacy way to get event id.. All hail to old kernels! */
265 
266 	/*
267 	 * This way does not work with group format read, so bail
268 	 * out in that case.
269 	 */
270 	if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
271 		return -1;
272 
273 	if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
274 	    read(fd, &read_data, sizeof(read_data)) == -1)
275 		return -1;
276 
277 	if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
278 		++id_idx;
279 	if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
280 		++id_idx;
281 
282 	id = read_data[id_idx];
283 
284 add:
285 	perf_evlist__id_add(evlist, evsel, cpu, thread, id);
286 	return 0;
287 }
288 
289 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
290 {
291 	int nr_cpus = perf_cpu_map__nr(evlist->cpus);
292 	int nr_threads = perf_thread_map__nr(evlist->threads);
293 	int nfds = 0;
294 	struct perf_evsel *evsel;
295 
296 	perf_evlist__for_each_entry(evlist, evsel) {
297 		if (evsel->system_wide)
298 			nfds += nr_cpus;
299 		else
300 			nfds += nr_cpus * nr_threads;
301 	}
302 
303 	if (fdarray__available_entries(&evlist->pollfd) < nfds &&
304 	    fdarray__grow(&evlist->pollfd, nfds) < 0)
305 		return -ENOMEM;
306 
307 	return 0;
308 }
309 
310 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
311 			    void *ptr, short revent, enum fdarray_flags flags)
312 {
313 	int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP, flags);
314 
315 	if (pos >= 0) {
316 		evlist->pollfd.priv[pos].ptr = ptr;
317 		fcntl(fd, F_SETFL, O_NONBLOCK);
318 	}
319 
320 	return pos;
321 }
322 
323 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
324 					 void *arg __maybe_unused)
325 {
326 	struct perf_mmap *map = fda->priv[fd].ptr;
327 
328 	if (map)
329 		perf_mmap__put(map);
330 }
331 
332 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
333 {
334 	return fdarray__filter(&evlist->pollfd, revents_and_mask,
335 			       perf_evlist__munmap_filtered, NULL);
336 }
337 
338 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
339 {
340 	return fdarray__poll(&evlist->pollfd, timeout);
341 }
342 
343 static struct perf_mmap* perf_evlist__alloc_mmap(struct perf_evlist *evlist, bool overwrite)
344 {
345 	int i;
346 	struct perf_mmap *map;
347 
348 	map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
349 	if (!map)
350 		return NULL;
351 
352 	for (i = 0; i < evlist->nr_mmaps; i++) {
353 		struct perf_mmap *prev = i ? &map[i - 1] : NULL;
354 
355 		/*
356 		 * When the perf_mmap() call is made we grab one refcount, plus
357 		 * one extra to let perf_mmap__consume() get the last
358 		 * events after all real references (perf_mmap__get()) are
359 		 * dropped.
360 		 *
361 		 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
362 		 * thus does perf_mmap__get() on it.
363 		 */
364 		perf_mmap__init(&map[i], prev, overwrite, NULL);
365 	}
366 
367 	return map;
368 }
369 
370 static void perf_evsel__set_sid_idx(struct perf_evsel *evsel, int idx, int cpu, int thread)
371 {
372 	struct perf_sample_id *sid = SID(evsel, cpu, thread);
373 
374 	sid->idx = idx;
375 	sid->cpu = perf_cpu_map__cpu(evsel->cpus, cpu);
376 	sid->tid = perf_thread_map__pid(evsel->threads, thread);
377 }
378 
379 static struct perf_mmap*
380 perf_evlist__mmap_cb_get(struct perf_evlist *evlist, bool overwrite, int idx)
381 {
382 	struct perf_mmap *maps;
383 
384 	maps = overwrite ? evlist->mmap_ovw : evlist->mmap;
385 
386 	if (!maps) {
387 		maps = perf_evlist__alloc_mmap(evlist, overwrite);
388 		if (!maps)
389 			return NULL;
390 
391 		if (overwrite)
392 			evlist->mmap_ovw = maps;
393 		else
394 			evlist->mmap = maps;
395 	}
396 
397 	return &maps[idx];
398 }
399 
400 #define FD(e, x, y) (*(int *) xyarray__entry(e->fd, x, y))
401 
402 static int
403 perf_evlist__mmap_cb_mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
404 			  int output, int cpu)
405 {
406 	return perf_mmap__mmap(map, mp, output, cpu);
407 }
408 
409 static void perf_evlist__set_mmap_first(struct perf_evlist *evlist, struct perf_mmap *map,
410 					bool overwrite)
411 {
412 	if (overwrite)
413 		evlist->mmap_ovw_first = map;
414 	else
415 		evlist->mmap_first = map;
416 }
417 
418 static int
419 mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
420 	       int idx, struct perf_mmap_param *mp, int cpu_idx,
421 	       int thread, int *_output, int *_output_overwrite)
422 {
423 	int evlist_cpu = perf_cpu_map__cpu(evlist->cpus, cpu_idx);
424 	struct perf_evsel *evsel;
425 	int revent;
426 
427 	perf_evlist__for_each_entry(evlist, evsel) {
428 		bool overwrite = evsel->attr.write_backward;
429 		struct perf_mmap *map;
430 		int *output, fd, cpu;
431 
432 		if (evsel->system_wide && thread)
433 			continue;
434 
435 		cpu = perf_cpu_map__idx(evsel->cpus, evlist_cpu);
436 		if (cpu == -1)
437 			continue;
438 
439 		map = ops->get(evlist, overwrite, idx);
440 		if (map == NULL)
441 			return -ENOMEM;
442 
443 		if (overwrite) {
444 			mp->prot = PROT_READ;
445 			output   = _output_overwrite;
446 		} else {
447 			mp->prot = PROT_READ | PROT_WRITE;
448 			output   = _output;
449 		}
450 
451 		fd = FD(evsel, cpu, thread);
452 
453 		if (*output == -1) {
454 			*output = fd;
455 
456 			/*
457 			 * The last one will be done at perf_mmap__consume(), so that we
458 			 * make sure we don't prevent tools from consuming every last event in
459 			 * the ring buffer.
460 			 *
461 			 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
462 			 * anymore, but the last events for it are still in the ring buffer,
463 			 * waiting to be consumed.
464 			 *
465 			 * Tools can chose to ignore this at their own discretion, but the
466 			 * evlist layer can't just drop it when filtering events in
467 			 * perf_evlist__filter_pollfd().
468 			 */
469 			refcount_set(&map->refcnt, 2);
470 
471 			if (ops->mmap(map, mp, *output, evlist_cpu) < 0)
472 				return -1;
473 
474 			if (!idx)
475 				perf_evlist__set_mmap_first(evlist, map, overwrite);
476 		} else {
477 			if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
478 				return -1;
479 
480 			perf_mmap__get(map);
481 		}
482 
483 		revent = !overwrite ? POLLIN : 0;
484 
485 		if (!evsel->system_wide &&
486 		    perf_evlist__add_pollfd(evlist, fd, map, revent, fdarray_flag__default) < 0) {
487 			perf_mmap__put(map);
488 			return -1;
489 		}
490 
491 		if (evsel->attr.read_format & PERF_FORMAT_ID) {
492 			if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
493 						   fd) < 0)
494 				return -1;
495 			perf_evsel__set_sid_idx(evsel, idx, cpu, thread);
496 		}
497 	}
498 
499 	return 0;
500 }
501 
502 static int
503 mmap_per_thread(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
504 		struct perf_mmap_param *mp)
505 {
506 	int thread;
507 	int nr_threads = perf_thread_map__nr(evlist->threads);
508 
509 	for (thread = 0; thread < nr_threads; thread++) {
510 		int output = -1;
511 		int output_overwrite = -1;
512 
513 		if (ops->idx)
514 			ops->idx(evlist, mp, thread, false);
515 
516 		if (mmap_per_evsel(evlist, ops, thread, mp, 0, thread,
517 				   &output, &output_overwrite))
518 			goto out_unmap;
519 	}
520 
521 	return 0;
522 
523 out_unmap:
524 	perf_evlist__munmap(evlist);
525 	return -1;
526 }
527 
528 static int
529 mmap_per_cpu(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
530 	     struct perf_mmap_param *mp)
531 {
532 	int nr_threads = perf_thread_map__nr(evlist->threads);
533 	int nr_cpus    = perf_cpu_map__nr(evlist->cpus);
534 	int cpu, thread;
535 
536 	for (cpu = 0; cpu < nr_cpus; cpu++) {
537 		int output = -1;
538 		int output_overwrite = -1;
539 
540 		if (ops->idx)
541 			ops->idx(evlist, mp, cpu, true);
542 
543 		for (thread = 0; thread < nr_threads; thread++) {
544 			if (mmap_per_evsel(evlist, ops, cpu, mp, cpu,
545 					   thread, &output, &output_overwrite))
546 				goto out_unmap;
547 		}
548 	}
549 
550 	return 0;
551 
552 out_unmap:
553 	perf_evlist__munmap(evlist);
554 	return -1;
555 }
556 
557 static int perf_evlist__nr_mmaps(struct perf_evlist *evlist)
558 {
559 	int nr_mmaps;
560 
561 	nr_mmaps = perf_cpu_map__nr(evlist->cpus);
562 	if (perf_cpu_map__empty(evlist->cpus))
563 		nr_mmaps = perf_thread_map__nr(evlist->threads);
564 
565 	return nr_mmaps;
566 }
567 
568 int perf_evlist__mmap_ops(struct perf_evlist *evlist,
569 			  struct perf_evlist_mmap_ops *ops,
570 			  struct perf_mmap_param *mp)
571 {
572 	struct perf_evsel *evsel;
573 	const struct perf_cpu_map *cpus = evlist->cpus;
574 	const struct perf_thread_map *threads = evlist->threads;
575 
576 	if (!ops || !ops->get || !ops->mmap)
577 		return -EINVAL;
578 
579 	mp->mask = evlist->mmap_len - page_size - 1;
580 
581 	evlist->nr_mmaps = perf_evlist__nr_mmaps(evlist);
582 
583 	perf_evlist__for_each_entry(evlist, evsel) {
584 		if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
585 		    evsel->sample_id == NULL &&
586 		    perf_evsel__alloc_id(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
587 			return -ENOMEM;
588 	}
589 
590 	if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
591 		return -ENOMEM;
592 
593 	if (perf_cpu_map__empty(cpus))
594 		return mmap_per_thread(evlist, ops, mp);
595 
596 	return mmap_per_cpu(evlist, ops, mp);
597 }
598 
599 int perf_evlist__mmap(struct perf_evlist *evlist, int pages)
600 {
601 	struct perf_mmap_param mp;
602 	struct perf_evlist_mmap_ops ops = {
603 		.get  = perf_evlist__mmap_cb_get,
604 		.mmap = perf_evlist__mmap_cb_mmap,
605 	};
606 
607 	evlist->mmap_len = (pages + 1) * page_size;
608 
609 	return perf_evlist__mmap_ops(evlist, &ops, &mp);
610 }
611 
612 void perf_evlist__munmap(struct perf_evlist *evlist)
613 {
614 	int i;
615 
616 	if (evlist->mmap) {
617 		for (i = 0; i < evlist->nr_mmaps; i++)
618 			perf_mmap__munmap(&evlist->mmap[i]);
619 	}
620 
621 	if (evlist->mmap_ovw) {
622 		for (i = 0; i < evlist->nr_mmaps; i++)
623 			perf_mmap__munmap(&evlist->mmap_ovw[i]);
624 	}
625 
626 	zfree(&evlist->mmap);
627 	zfree(&evlist->mmap_ovw);
628 }
629 
630 struct perf_mmap*
631 perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,
632 		       bool overwrite)
633 {
634 	if (map)
635 		return map->next;
636 
637 	return overwrite ? evlist->mmap_ovw_first : evlist->mmap_first;
638 }
639