xref: /linux/tools/perf/util/cgroup.c (revision 9e906a9dead17d81d6c2687f65e159231d0e3286)
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 
open_cgroup(const char * name)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
__read_cgroup_id(const char * path)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 
read_cgroup_id(struct cgroup * cgrp)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
__read_cgroup_id(const char * path __maybe_unused)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 
cgroup_is_v2(const char * subsys)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 
evlist__find_cgroup(struct evlist * evlist,const char * str)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 
cgroup__new(const char * name,bool do_open)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 
evlist__findnew_cgroup(struct evlist * evlist,const char * name)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 
add_cgroup(struct evlist * evlist,const char * str)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 
cgroup__delete(struct cgroup * cgroup)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 
cgroup__put(struct cgroup * cgrp)188 void cgroup__put(struct cgroup *cgrp)
189 {
190 	if (cgrp && refcount_dec_and_test(&cgrp->refcnt)) {
191 		cgroup__delete(cgrp);
192 	}
193 }
194 
cgroup__get(struct cgroup * cgroup)195 struct cgroup *cgroup__get(struct cgroup *cgroup)
196 {
197        if (cgroup)
198 		refcount_inc(&cgroup->refcnt);
199        return cgroup;
200 }
201 
evsel__set_default_cgroup(struct evsel * evsel,struct cgroup * cgroup)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 
evlist__set_default_cgroup(struct evlist * evlist,struct cgroup * cgroup)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 */
add_cgroup_name(const char * fpath,const struct stat * sb __maybe_unused,int typeflag,struct FTW * ftwbuf __maybe_unused)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 
check_and_add_cgroup_name(const char * fpath)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 
release_cgroup_list(void)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 */
list_cgroups(const char * str)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 */
match_cgroups(const char * str)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 
parse_cgroups(const struct option * opt,const char * str,int unset __maybe_unused)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 
has_pattern_string(const char * str)412 static bool has_pattern_string(const char *str)
413 {
414 	return !!strpbrk(str, "{}[]()|*+?^$");
415 }
416 
evlist__expand_cgroup(struct evlist * evlist,const char * str,bool open_cgroup)417 int evlist__expand_cgroup(struct evlist *evlist, const char *str, bool open_cgroup)
418 {
419 	struct evlist *orig_list, *tmp_list;
420 	struct evsel *pos, *evsel, *leader;
421 	struct rblist orig_metric_events;
422 	struct cgroup *cgrp = NULL;
423 	struct cgroup_name *cn;
424 	int ret = -1;
425 	int prefix_len;
426 
427 	if (evlist->core.nr_entries == 0) {
428 		fprintf(stderr, "must define events before cgroups\n");
429 		return -EINVAL;
430 	}
431 
432 	orig_list = evlist__new();
433 	tmp_list = evlist__new();
434 	if (orig_list == NULL || tmp_list == NULL) {
435 		fprintf(stderr, "memory allocation failed\n");
436 		return -ENOMEM;
437 	}
438 
439 	/* save original events and init evlist */
440 	evlist__splice_list_tail(orig_list, &evlist->core.entries);
441 	evlist->core.nr_entries = 0;
442 
443 	orig_metric_events = evlist->metric_events;
444 	metricgroup__rblist_init(&evlist->metric_events);
445 
446 	if (has_pattern_string(str))
447 		prefix_len = match_cgroups(str);
448 	else
449 		prefix_len = list_cgroups(str);
450 
451 	if (prefix_len < 0)
452 		goto out_err;
453 
454 	list_for_each_entry(cn, &cgroup_list, list) {
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 		leader = NULL;
471 		evlist__for_each_entry(orig_list, pos) {
472 			evsel = evsel__clone(/*dest=*/NULL, pos);
473 			if (evsel == NULL)
474 				goto out_err;
475 
476 			cgroup__put(evsel->cgrp);
477 			evsel->cgrp = cgroup__get(cgrp);
478 
479 			if (evsel__is_group_leader(pos))
480 				leader = evsel;
481 			evsel__set_leader(evsel, leader);
482 
483 			evlist__add(tmp_list, evsel);
484 		}
485 		/* cgroup__new() has a refcount, release it here */
486 		cgroup__put(cgrp);
487 		nr_cgroups++;
488 
489 		if (metricgroup__copy_metric_events(tmp_list, cgrp,
490 						    &evlist->metric_events,
491 						    &orig_metric_events) < 0)
492 			goto out_err;
493 
494 		evlist__splice_list_tail(evlist, &tmp_list->core.entries);
495 		tmp_list->core.nr_entries = 0;
496 	}
497 
498 	if (list_empty(&evlist->core.entries)) {
499 		fprintf(stderr, "no cgroup matched: %s\n", str);
500 		goto out_err;
501 	}
502 
503 	ret = 0;
504 	cgrp_event_expanded = true;
505 
506 out_err:
507 	evlist__delete(orig_list);
508 	evlist__delete(tmp_list);
509 	metricgroup__rblist_exit(&orig_metric_events);
510 	release_cgroup_list();
511 
512 	return ret;
513 }
514 
__cgroup__findnew(struct rb_root * root,uint64_t id,bool create,const char * path)515 static struct cgroup *__cgroup__findnew(struct rb_root *root, uint64_t id,
516 					bool create, const char *path)
517 {
518 	struct rb_node **p = &root->rb_node;
519 	struct rb_node *parent = NULL;
520 	struct cgroup *cgrp;
521 
522 	while (*p != NULL) {
523 		parent = *p;
524 		cgrp = rb_entry(parent, struct cgroup, node);
525 
526 		if (cgrp->id == id)
527 			return cgrp;
528 
529 		if (cgrp->id < id)
530 			p = &(*p)->rb_left;
531 		else
532 			p = &(*p)->rb_right;
533 	}
534 
535 	if (!create)
536 		return NULL;
537 
538 	cgrp = malloc(sizeof(*cgrp));
539 	if (cgrp == NULL)
540 		return NULL;
541 
542 	cgrp->name = strdup(path);
543 	if (cgrp->name == NULL) {
544 		free(cgrp);
545 		return NULL;
546 	}
547 
548 	cgrp->fd = -1;
549 	cgrp->id = id;
550 	refcount_set(&cgrp->refcnt, 1);
551 
552 	rb_link_node(&cgrp->node, parent, p);
553 	rb_insert_color(&cgrp->node, root);
554 
555 	return cgrp;
556 }
557 
cgroup__findnew(struct perf_env * env,uint64_t id,const char * path)558 struct cgroup *cgroup__findnew(struct perf_env *env, uint64_t id,
559 			       const char *path)
560 {
561 	struct cgroup *cgrp;
562 
563 	down_write(&env->cgroups.lock);
564 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, true, path);
565 	up_write(&env->cgroups.lock);
566 	return cgrp;
567 }
568 
__cgroup__find(struct rb_root * root,uint64_t id)569 struct cgroup *__cgroup__find(struct rb_root *root, uint64_t id)
570 {
571 	return __cgroup__findnew(root, id, /*create=*/false, /*path=*/NULL);
572 }
573 
cgroup__find(struct perf_env * env,uint64_t id)574 struct cgroup *cgroup__find(struct perf_env *env, uint64_t id)
575 {
576 	struct cgroup *cgrp;
577 
578 	down_read(&env->cgroups.lock);
579 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, false, NULL);
580 	up_read(&env->cgroups.lock);
581 	return cgrp;
582 }
583 
perf_env__purge_cgroups(struct perf_env * env)584 void perf_env__purge_cgroups(struct perf_env *env)
585 {
586 	struct rb_node *node;
587 	struct cgroup *cgrp;
588 
589 	down_write(&env->cgroups.lock);
590 	while (!RB_EMPTY_ROOT(&env->cgroups.tree)) {
591 		node = rb_first(&env->cgroups.tree);
592 		cgrp = rb_entry(node, struct cgroup, node);
593 
594 		rb_erase(node, &env->cgroups.tree);
595 		cgroup__put(cgrp);
596 	}
597 	up_write(&env->cgroups.lock);
598 }
599 
read_all_cgroups(struct rb_root * root)600 void read_all_cgroups(struct rb_root *root)
601 {
602 	char mnt[PATH_MAX];
603 	struct cgroup_name *cn;
604 	int prefix_len;
605 
606 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
607 		return;
608 
609 	/* cgroup_name will have a full path, skip the root directory */
610 	prefix_len = strlen(mnt);
611 
612 	/* collect all cgroups in the cgroup_list */
613 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
614 		return;
615 
616 	list_for_each_entry(cn, &cgroup_list, list) {
617 		const char *name;
618 		u64 cgrp_id;
619 
620 		/* cgroup_name might have a full path, skip the prefix */
621 		name = cn->name + prefix_len;
622 		if (name[0] == '\0')
623 			name = "/";
624 
625 		cgrp_id = __read_cgroup_id(cn->name);
626 		__cgroup__findnew(root, cgrp_id, /*create=*/true, name);
627 	}
628 
629 	release_cgroup_list();
630 }
631