xref: /linux/tools/perf/util/cgroup.c (revision 046fd8206d820b71e7870f7b894b46f8a15ae974)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <subcmd/parse-options.h>
3 #include "evsel.h"
4 #include "cgroup.h"
5 #include "evlist.h"
6 #include "rblist.h"
7 #include "metricgroup.h"
8 #include "stat.h"
9 #include <linux/zalloc.h>
10 #include <sys/types.h>
11 #include <sys/stat.h>
12 #include <sys/statfs.h>
13 #include <errno.h>
14 #include <fcntl.h>
15 #include <stdlib.h>
16 #include <string.h>
17 #include <api/fs/fs.h>
18 #include <ftw.h>
19 #include <regex.h>
20 
21 int nr_cgroups;
22 bool cgrp_event_expanded;
23 
24 /* used to match cgroup name with patterns */
25 struct cgroup_name {
26 	struct list_head list;
27 	bool used;
28 	char name[];
29 };
30 static LIST_HEAD(cgroup_list);
31 
32 static int open_cgroup(const char *name)
33 {
34 	char path[PATH_MAX + 1];
35 	char mnt[PATH_MAX + 1];
36 	int fd;
37 
38 
39 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
40 		return -1;
41 
42 	scnprintf(path, PATH_MAX, "%s/%s", mnt, name);
43 
44 	fd = open(path, O_RDONLY);
45 	if (fd == -1)
46 		fprintf(stderr, "no access to cgroup %s\n", path);
47 
48 	return fd;
49 }
50 
51 #ifdef HAVE_FILE_HANDLE
52 static u64 __read_cgroup_id(const char *path)
53 {
54 	struct {
55 		struct file_handle fh;
56 		uint64_t cgroup_id;
57 	} handle;
58 	int mount_id;
59 
60 	handle.fh.handle_bytes = sizeof(handle.cgroup_id);
61 	if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0)
62 		return -1ULL;
63 
64 	return handle.cgroup_id;
65 }
66 
67 int read_cgroup_id(struct cgroup *cgrp)
68 {
69 	char path[PATH_MAX + 1];
70 	char mnt[PATH_MAX + 1];
71 
72 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
73 		return -1;
74 
75 	scnprintf(path, PATH_MAX, "%s/%s", mnt, cgrp->name);
76 
77 	cgrp->id = __read_cgroup_id(path);
78 	return 0;
79 }
80 #else
81 static inline u64 __read_cgroup_id(const char *path __maybe_unused) { return -1ULL; }
82 #endif  /* HAVE_FILE_HANDLE */
83 
84 #ifndef CGROUP2_SUPER_MAGIC
85 #define CGROUP2_SUPER_MAGIC  0x63677270
86 #endif
87 
88 int cgroup_is_v2(const char *subsys)
89 {
90 	char mnt[PATH_MAX + 1];
91 	struct statfs stbuf;
92 
93 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, subsys))
94 		return -1;
95 
96 	if (statfs(mnt, &stbuf) < 0)
97 		return -1;
98 
99 	return (stbuf.f_type == CGROUP2_SUPER_MAGIC);
100 }
101 
102 static struct cgroup *evlist__find_cgroup(struct evlist *evlist, const char *str)
103 {
104 	struct evsel *counter;
105 	/*
106 	 * check if cgrp is already defined, if so we reuse it
107 	 */
108 	evlist__for_each_entry(evlist, counter) {
109 		if (!counter->cgrp)
110 			continue;
111 		if (!strcmp(counter->cgrp->name, str))
112 			return cgroup__get(counter->cgrp);
113 	}
114 
115 	return NULL;
116 }
117 
118 struct cgroup *cgroup__new(const char *name, bool do_open)
119 {
120 	struct cgroup *cgroup = zalloc(sizeof(*cgroup));
121 
122 	if (cgroup != NULL) {
123 		refcount_set(&cgroup->refcnt, 1);
124 
125 		cgroup->name = strdup(name);
126 		if (!cgroup->name)
127 			goto out_err;
128 
129 		if (do_open) {
130 			cgroup->fd = open_cgroup(name);
131 			if (cgroup->fd == -1)
132 				goto out_free_name;
133 		} else {
134 			cgroup->fd = -1;
135 		}
136 	}
137 
138 	return cgroup;
139 
140 out_free_name:
141 	zfree(&cgroup->name);
142 out_err:
143 	free(cgroup);
144 	return NULL;
145 }
146 
147 struct cgroup *evlist__findnew_cgroup(struct evlist *evlist, const char *name)
148 {
149 	struct cgroup *cgroup = evlist__find_cgroup(evlist, name);
150 
151 	return cgroup ?: cgroup__new(name, true);
152 }
153 
154 static int add_cgroup(struct evlist *evlist, const char *str)
155 {
156 	struct evsel *counter;
157 	struct cgroup *cgrp = evlist__findnew_cgroup(evlist, str);
158 	int n;
159 
160 	if (!cgrp)
161 		return -1;
162 	/*
163 	 * find corresponding event
164 	 * if add cgroup N, then need to find event N
165 	 */
166 	n = 0;
167 	evlist__for_each_entry(evlist, counter) {
168 		if (n == nr_cgroups)
169 			goto found;
170 		n++;
171 	}
172 
173 	cgroup__put(cgrp);
174 	return -1;
175 found:
176 	counter->cgrp = cgrp;
177 	return 0;
178 }
179 
180 static void cgroup__delete(struct cgroup *cgroup)
181 {
182 	if (cgroup->fd >= 0)
183 		close(cgroup->fd);
184 	zfree(&cgroup->name);
185 	free(cgroup);
186 }
187 
188 void cgroup__put(struct cgroup *cgrp)
189 {
190 	if (cgrp && refcount_dec_and_test(&cgrp->refcnt)) {
191 		cgroup__delete(cgrp);
192 	}
193 }
194 
195 struct cgroup *cgroup__get(struct cgroup *cgroup)
196 {
197        if (cgroup)
198 		refcount_inc(&cgroup->refcnt);
199        return cgroup;
200 }
201 
202 static void evsel__set_default_cgroup(struct evsel *evsel, struct cgroup *cgroup)
203 {
204 	if (evsel->cgrp == NULL)
205 		evsel->cgrp = cgroup__get(cgroup);
206 }
207 
208 void evlist__set_default_cgroup(struct evlist *evlist, struct cgroup *cgroup)
209 {
210 	struct evsel *evsel;
211 
212 	evlist__for_each_entry(evlist, evsel)
213 		evsel__set_default_cgroup(evsel, cgroup);
214 }
215 
216 /* helper function for ftw() in match_cgroups and list_cgroups */
217 static int add_cgroup_name(const char *fpath, const struct stat *sb __maybe_unused,
218 			   int typeflag, struct FTW *ftwbuf __maybe_unused)
219 {
220 	struct cgroup_name *cn;
221 
222 	if (typeflag != FTW_D)
223 		return 0;
224 
225 	cn = malloc(sizeof(*cn) + strlen(fpath) + 1);
226 	if (cn == NULL)
227 		return -1;
228 
229 	cn->used = false;
230 	strcpy(cn->name, fpath);
231 
232 	list_add_tail(&cn->list, &cgroup_list);
233 	return 0;
234 }
235 
236 static int check_and_add_cgroup_name(const char *fpath)
237 {
238 	struct cgroup_name *cn;
239 
240 	list_for_each_entry(cn, &cgroup_list, list) {
241 		if (!strcmp(cn->name, fpath))
242 			return 0;
243 	}
244 
245 	/* pretend if it's added by ftw() */
246 	return add_cgroup_name(fpath, NULL, FTW_D, NULL);
247 }
248 
249 static void release_cgroup_list(void)
250 {
251 	struct cgroup_name *cn;
252 
253 	while (!list_empty(&cgroup_list)) {
254 		cn = list_first_entry(&cgroup_list, struct cgroup_name, list);
255 		list_del(&cn->list);
256 		free(cn);
257 	}
258 }
259 
260 /* collect given cgroups only */
261 static int list_cgroups(const char *str)
262 {
263 	const char *p, *e, *eos = str + strlen(str);
264 	struct cgroup_name *cn;
265 	char *s;
266 
267 	/* use given name as is when no regex is given */
268 	for (;;) {
269 		p = strchr(str, ',');
270 		e = p ? p : eos;
271 
272 		if (e - str) {
273 			int ret;
274 
275 			s = strndup(str, e - str);
276 			if (!s)
277 				return -1;
278 
279 			ret = check_and_add_cgroup_name(s);
280 			free(s);
281 			if (ret < 0)
282 				return -1;
283 		} else {
284 			if (check_and_add_cgroup_name("/") < 0)
285 				return -1;
286 		}
287 
288 		if (!p)
289 			break;
290 		str = p+1;
291 	}
292 
293 	/* these groups will be used */
294 	list_for_each_entry(cn, &cgroup_list, list)
295 		cn->used = true;
296 
297 	return 0;
298 }
299 
300 /* collect all cgroups first and then match with the pattern */
301 static int match_cgroups(const char *str)
302 {
303 	char mnt[PATH_MAX];
304 	const char *p, *e, *eos = str + strlen(str);
305 	struct cgroup_name *cn;
306 	regex_t reg;
307 	int prefix_len;
308 	char *s;
309 
310 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
311 		return -1;
312 
313 	/* cgroup_name will have a full path, skip the root directory */
314 	prefix_len = strlen(mnt);
315 
316 	/* collect all cgroups in the cgroup_list */
317 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
318 		return -1;
319 
320 	for (;;) {
321 		p = strchr(str, ',');
322 		e = p ? p : eos;
323 
324 		/* allow empty cgroups, i.e., skip */
325 		if (e - str) {
326 			/* termination added */
327 			s = strndup(str, e - str);
328 			if (!s)
329 				return -1;
330 			if (regcomp(&reg, s, REG_NOSUB)) {
331 				free(s);
332 				return -1;
333 			}
334 
335 			/* check cgroup name with the pattern */
336 			list_for_each_entry(cn, &cgroup_list, list) {
337 				char *name = cn->name + prefix_len;
338 
339 				if (name[0] == '/' && name[1])
340 					name++;
341 				if (!regexec(&reg, name, 0, NULL, 0))
342 					cn->used = true;
343 			}
344 			regfree(&reg);
345 			free(s);
346 		} else {
347 			/* first entry to root cgroup */
348 			cn = list_first_entry(&cgroup_list, struct cgroup_name,
349 					      list);
350 			cn->used = true;
351 		}
352 
353 		if (!p)
354 			break;
355 		str = p+1;
356 	}
357 	return prefix_len;
358 }
359 
360 int parse_cgroups(const struct option *opt, const char *str,
361 		  int unset __maybe_unused)
362 {
363 	struct evlist *evlist = *(struct evlist **)opt->value;
364 	struct evsel *counter;
365 	struct cgroup *cgrp = NULL;
366 	const char *p, *e, *eos = str + strlen(str);
367 	char *s;
368 	int ret, i;
369 
370 	if (list_empty(&evlist->core.entries)) {
371 		fprintf(stderr, "must define events before cgroups\n");
372 		return -1;
373 	}
374 
375 	for (;;) {
376 		p = strchr(str, ',');
377 		e = p ? p : eos;
378 
379 		/* allow empty cgroups, i.e., skip */
380 		if (e - str) {
381 			/* termination added */
382 			s = strndup(str, e - str);
383 			if (!s)
384 				return -1;
385 			ret = add_cgroup(evlist, s);
386 			free(s);
387 			if (ret)
388 				return -1;
389 		}
390 		/* nr_cgroups is increased een for empty cgroups */
391 		nr_cgroups++;
392 		if (!p)
393 			break;
394 		str = p+1;
395 	}
396 	/* for the case one cgroup combine to multiple events */
397 	i = 0;
398 	if (nr_cgroups == 1) {
399 		evlist__for_each_entry(evlist, counter) {
400 			if (i == 0)
401 				cgrp = counter->cgrp;
402 			else {
403 				counter->cgrp = cgrp;
404 				refcount_inc(&cgrp->refcnt);
405 			}
406 			i++;
407 		}
408 	}
409 	return 0;
410 }
411 
412 static bool has_pattern_string(const char *str)
413 {
414 	return !!strpbrk(str, "{}[]()|*+?^$");
415 }
416 
417 int evlist__expand_cgroup(struct evlist *evlist, const char *str, bool open_cgroup)
418 {
419 	struct evlist *orig_list, *tmp_list;
420 	struct rblist orig_metric_events;
421 	struct cgroup *cgrp = NULL;
422 	struct cgroup_name *cn;
423 	int ret = -1;
424 	int prefix_len;
425 
426 	if (evlist->core.nr_entries == 0) {
427 		fprintf(stderr, "must define events before cgroups\n");
428 		return -EINVAL;
429 	}
430 
431 	orig_list = evlist__new();
432 	tmp_list = evlist__new();
433 	if (orig_list == NULL || tmp_list == NULL) {
434 		fprintf(stderr, "memory allocation failed\n");
435 		return -ENOMEM;
436 	}
437 
438 	/* save original events and init evlist */
439 	evlist__splice_list_tail(orig_list, &evlist->core.entries);
440 	evlist->core.nr_entries = 0;
441 
442 	orig_metric_events = evlist->metric_events;
443 	metricgroup__rblist_init(&evlist->metric_events);
444 
445 	if (has_pattern_string(str))
446 		prefix_len = match_cgroups(str);
447 	else
448 		prefix_len = list_cgroups(str);
449 
450 	if (prefix_len < 0)
451 		goto out_err;
452 
453 	list_for_each_entry(cn, &cgroup_list, list) {
454 		struct evsel *pos;
455 		char *name;
456 
457 		if (!cn->used)
458 			continue;
459 
460 		/* cgroup_name might have a full path, skip the prefix */
461 		name = cn->name + prefix_len;
462 		if (name[0] == '/' && name[1])
463 			name++;
464 
465 		/* the cgroup can go away in the meantime */
466 		cgrp = cgroup__new(name, open_cgroup);
467 		if (cgrp == NULL)
468 			continue;
469 
470 		/* copy the list and set to the new cgroup. */
471 		evlist__for_each_entry(orig_list, pos) {
472 			struct evsel *evsel = evsel__clone(/*dest=*/NULL, pos);
473 
474 			if (evsel == NULL)
475 				goto out_err;
476 
477 			/* stash the copy during the copying. */
478 			pos->priv = evsel;
479 			cgroup__put(evsel->cgrp);
480 			evsel->cgrp = cgroup__get(cgrp);
481 
482 			evlist__add(tmp_list, evsel);
483 		}
484 		/* update leader information using stashed pointer to copy. */
485 		evlist__for_each_entry(orig_list, pos) {
486 			struct evsel *evsel = pos->priv;
487 
488 			if (evsel__leader(pos))
489 				evsel__set_leader(evsel, evsel__leader(pos)->priv);
490 
491 			if (pos->metric_leader)
492 				evsel->metric_leader = pos->metric_leader->priv;
493 
494 			if (pos->first_wildcard_match)
495 				evsel->first_wildcard_match = pos->first_wildcard_match->priv;
496 		}
497 		/* the stashed copy is no longer used. */
498 		evlist__for_each_entry(orig_list, pos)
499 			pos->priv = NULL;
500 
501 		/* cgroup__new() has a refcount, release it here */
502 		cgroup__put(cgrp);
503 		nr_cgroups++;
504 
505 		if (metricgroup__copy_metric_events(tmp_list, cgrp,
506 						    &evlist->metric_events,
507 						    &orig_metric_events) < 0)
508 			goto out_err;
509 
510 		evlist__splice_list_tail(evlist, &tmp_list->core.entries);
511 		tmp_list->core.nr_entries = 0;
512 	}
513 
514 	if (list_empty(&evlist->core.entries)) {
515 		fprintf(stderr, "no cgroup matched: %s\n", str);
516 		goto out_err;
517 	}
518 
519 	ret = 0;
520 	cgrp_event_expanded = true;
521 
522 out_err:
523 	evlist__delete(orig_list);
524 	evlist__delete(tmp_list);
525 	metricgroup__rblist_exit(&orig_metric_events);
526 	release_cgroup_list();
527 
528 	return ret;
529 }
530 
531 static struct cgroup *__cgroup__findnew(struct rb_root *root, uint64_t id,
532 					bool create, const char *path)
533 {
534 	struct rb_node **p = &root->rb_node;
535 	struct rb_node *parent = NULL;
536 	struct cgroup *cgrp;
537 
538 	while (*p != NULL) {
539 		parent = *p;
540 		cgrp = rb_entry(parent, struct cgroup, node);
541 
542 		if (cgrp->id == id)
543 			return cgrp;
544 
545 		if (cgrp->id < id)
546 			p = &(*p)->rb_left;
547 		else
548 			p = &(*p)->rb_right;
549 	}
550 
551 	if (!create)
552 		return NULL;
553 
554 	cgrp = malloc(sizeof(*cgrp));
555 	if (cgrp == NULL)
556 		return NULL;
557 
558 	cgrp->name = strdup(path);
559 	if (cgrp->name == NULL) {
560 		free(cgrp);
561 		return NULL;
562 	}
563 
564 	cgrp->fd = -1;
565 	cgrp->id = id;
566 	refcount_set(&cgrp->refcnt, 1);
567 
568 	rb_link_node(&cgrp->node, parent, p);
569 	rb_insert_color(&cgrp->node, root);
570 
571 	return cgrp;
572 }
573 
574 struct cgroup *cgroup__findnew(struct perf_env *env, uint64_t id,
575 			       const char *path)
576 {
577 	struct cgroup *cgrp;
578 
579 	down_write(&env->cgroups.lock);
580 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, true, path);
581 	up_write(&env->cgroups.lock);
582 	return cgrp;
583 }
584 
585 struct cgroup *__cgroup__find(struct rb_root *root, uint64_t id)
586 {
587 	return __cgroup__findnew(root, id, /*create=*/false, /*path=*/NULL);
588 }
589 
590 struct cgroup *cgroup__find(struct perf_env *env, uint64_t id)
591 {
592 	struct cgroup *cgrp;
593 
594 	down_read(&env->cgroups.lock);
595 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, false, NULL);
596 	up_read(&env->cgroups.lock);
597 	return cgrp;
598 }
599 
600 void perf_env__purge_cgroups(struct perf_env *env)
601 {
602 	struct rb_node *node;
603 	struct cgroup *cgrp;
604 
605 	down_write(&env->cgroups.lock);
606 	while (!RB_EMPTY_ROOT(&env->cgroups.tree)) {
607 		node = rb_first(&env->cgroups.tree);
608 		cgrp = rb_entry(node, struct cgroup, node);
609 
610 		rb_erase(node, &env->cgroups.tree);
611 		cgroup__put(cgrp);
612 	}
613 	up_write(&env->cgroups.lock);
614 }
615 
616 void read_all_cgroups(struct rb_root *root)
617 {
618 	char mnt[PATH_MAX];
619 	struct cgroup_name *cn;
620 	int prefix_len;
621 
622 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
623 		return;
624 
625 	/* cgroup_name will have a full path, skip the root directory */
626 	prefix_len = strlen(mnt);
627 
628 	/* collect all cgroups in the cgroup_list */
629 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
630 		return;
631 
632 	list_for_each_entry(cn, &cgroup_list, list) {
633 		const char *name;
634 		u64 cgrp_id;
635 
636 		/* cgroup_name might have a full path, skip the prefix */
637 		name = cn->name + prefix_len;
638 		if (name[0] == '\0')
639 			name = "/";
640 
641 		cgrp_id = __read_cgroup_id(cn->name);
642 		__cgroup__findnew(root, cgrp_id, /*create=*/true, name);
643 	}
644 
645 	release_cgroup_list();
646 }
647