xref: /linux/tools/perf/util/cgroup.c (revision af8e51644a70f612974a6e767fa7d896d3c23f88)
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 <fcntl.h>
14 #include <stdlib.h>
15 #include <string.h>
16 #include <api/fs/fs.h>
17 #include <ftw.h>
18 #include <regex.h>
19 
20 int nr_cgroups;
21 bool cgrp_event_expanded;
22 
23 /* used to match cgroup name with patterns */
24 struct cgroup_name {
25 	struct list_head list;
26 	bool used;
27 	char name[];
28 };
29 static LIST_HEAD(cgroup_list);
30 
31 static int open_cgroup(const char *name)
32 {
33 	char path[PATH_MAX + 1];
34 	char mnt[PATH_MAX + 1];
35 	int fd;
36 
37 
38 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
39 		return -1;
40 
41 	scnprintf(path, PATH_MAX, "%s/%s", mnt, name);
42 
43 	fd = open(path, O_RDONLY);
44 	if (fd == -1)
45 		fprintf(stderr, "no access to cgroup %s\n", path);
46 
47 	return fd;
48 }
49 
50 #ifdef HAVE_FILE_HANDLE
51 static u64 __read_cgroup_id(const char *path)
52 {
53 	struct {
54 		struct file_handle fh;
55 		uint64_t cgroup_id;
56 	} handle;
57 	int mount_id;
58 
59 	handle.fh.handle_bytes = sizeof(handle.cgroup_id);
60 	if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0)
61 		return -1ULL;
62 
63 	return handle.cgroup_id;
64 }
65 
66 int read_cgroup_id(struct cgroup *cgrp)
67 {
68 	char path[PATH_MAX + 1];
69 	char mnt[PATH_MAX + 1];
70 
71 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
72 		return -1;
73 
74 	scnprintf(path, PATH_MAX, "%s/%s", mnt, cgrp->name);
75 
76 	cgrp->id = __read_cgroup_id(path);
77 	return 0;
78 }
79 #else
80 static inline u64 __read_cgroup_id(const char *path __maybe_unused) { return -1ULL; }
81 #endif  /* HAVE_FILE_HANDLE */
82 
83 #ifndef CGROUP2_SUPER_MAGIC
84 #define CGROUP2_SUPER_MAGIC  0x63677270
85 #endif
86 
87 int cgroup_is_v2(const char *subsys)
88 {
89 	char mnt[PATH_MAX + 1];
90 	struct statfs stbuf;
91 
92 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, subsys))
93 		return -1;
94 
95 	if (statfs(mnt, &stbuf) < 0)
96 		return -1;
97 
98 	return (stbuf.f_type == CGROUP2_SUPER_MAGIC);
99 }
100 
101 static struct cgroup *evlist__find_cgroup(struct evlist *evlist, const char *str)
102 {
103 	struct evsel *counter;
104 	/*
105 	 * check if cgrp is already defined, if so we reuse it
106 	 */
107 	evlist__for_each_entry(evlist, counter) {
108 		if (!counter->cgrp)
109 			continue;
110 		if (!strcmp(counter->cgrp->name, str))
111 			return cgroup__get(counter->cgrp);
112 	}
113 
114 	return NULL;
115 }
116 
117 struct cgroup *cgroup__new(const char *name, bool do_open)
118 {
119 	struct cgroup *cgroup = zalloc(sizeof(*cgroup));
120 
121 	if (cgroup != NULL) {
122 		refcount_set(&cgroup->refcnt, 1);
123 
124 		cgroup->name = strdup(name);
125 		if (!cgroup->name)
126 			goto out_err;
127 
128 		if (do_open) {
129 			cgroup->fd = open_cgroup(name);
130 			if (cgroup->fd == -1)
131 				goto out_free_name;
132 		} else {
133 			cgroup->fd = -1;
134 		}
135 	}
136 
137 	return cgroup;
138 
139 out_free_name:
140 	zfree(&cgroup->name);
141 out_err:
142 	free(cgroup);
143 	return NULL;
144 }
145 
146 struct cgroup *evlist__findnew_cgroup(struct evlist *evlist, const char *name)
147 {
148 	struct cgroup *cgroup = evlist__find_cgroup(evlist, name);
149 
150 	return cgroup ?: cgroup__new(name, true);
151 }
152 
153 static int add_cgroup(struct evlist *evlist, const char *str)
154 {
155 	struct evsel *counter;
156 	struct cgroup *cgrp = evlist__findnew_cgroup(evlist, str);
157 	int n;
158 
159 	if (!cgrp)
160 		return -1;
161 	/*
162 	 * find corresponding event
163 	 * if add cgroup N, then need to find event N
164 	 */
165 	n = 0;
166 	evlist__for_each_entry(evlist, counter) {
167 		if (n == nr_cgroups)
168 			goto found;
169 		n++;
170 	}
171 
172 	cgroup__put(cgrp);
173 	return -1;
174 found:
175 	counter->cgrp = cgrp;
176 	return 0;
177 }
178 
179 static void cgroup__delete(struct cgroup *cgroup)
180 {
181 	if (cgroup->fd >= 0)
182 		close(cgroup->fd);
183 	zfree(&cgroup->name);
184 	free(cgroup);
185 }
186 
187 void cgroup__put(struct cgroup *cgrp)
188 {
189 	if (cgrp && refcount_dec_and_test(&cgrp->refcnt)) {
190 		cgroup__delete(cgrp);
191 	}
192 }
193 
194 struct cgroup *cgroup__get(struct cgroup *cgroup)
195 {
196        if (cgroup)
197 		refcount_inc(&cgroup->refcnt);
198        return cgroup;
199 }
200 
201 static void evsel__set_default_cgroup(struct evsel *evsel, struct cgroup *cgroup)
202 {
203 	if (evsel->cgrp == NULL)
204 		evsel->cgrp = cgroup__get(cgroup);
205 }
206 
207 void evlist__set_default_cgroup(struct evlist *evlist, struct cgroup *cgroup)
208 {
209 	struct evsel *evsel;
210 
211 	evlist__for_each_entry(evlist, evsel)
212 		evsel__set_default_cgroup(evsel, cgroup);
213 }
214 
215 /* helper function for ftw() in match_cgroups and list_cgroups */
216 static int add_cgroup_name(const char *fpath, const struct stat *sb __maybe_unused,
217 			   int typeflag, struct FTW *ftwbuf __maybe_unused)
218 {
219 	struct cgroup_name *cn;
220 
221 	if (typeflag != FTW_D)
222 		return 0;
223 
224 	cn = malloc(sizeof(*cn) + strlen(fpath) + 1);
225 	if (cn == NULL)
226 		return -1;
227 
228 	cn->used = false;
229 	strcpy(cn->name, fpath);
230 
231 	list_add_tail(&cn->list, &cgroup_list);
232 	return 0;
233 }
234 
235 static int check_and_add_cgroup_name(const char *fpath)
236 {
237 	struct cgroup_name *cn;
238 
239 	list_for_each_entry(cn, &cgroup_list, list) {
240 		if (!strcmp(cn->name, fpath))
241 			return 0;
242 	}
243 
244 	/* pretend if it's added by ftw() */
245 	return add_cgroup_name(fpath, NULL, FTW_D, NULL);
246 }
247 
248 static void release_cgroup_list(void)
249 {
250 	struct cgroup_name *cn;
251 
252 	while (!list_empty(&cgroup_list)) {
253 		cn = list_first_entry(&cgroup_list, struct cgroup_name, list);
254 		list_del(&cn->list);
255 		free(cn);
256 	}
257 }
258 
259 /* collect given cgroups only */
260 static int list_cgroups(const char *str)
261 {
262 	const char *p, *e, *eos = str + strlen(str);
263 	struct cgroup_name *cn;
264 	char *s;
265 
266 	/* use given name as is when no regex is given */
267 	for (;;) {
268 		p = strchr(str, ',');
269 		e = p ? p : eos;
270 
271 		if (e - str) {
272 			int ret;
273 
274 			s = strndup(str, e - str);
275 			if (!s)
276 				return -1;
277 
278 			ret = check_and_add_cgroup_name(s);
279 			free(s);
280 			if (ret < 0)
281 				return -1;
282 		} else {
283 			if (check_and_add_cgroup_name("/") < 0)
284 				return -1;
285 		}
286 
287 		if (!p)
288 			break;
289 		str = p+1;
290 	}
291 
292 	/* these groups will be used */
293 	list_for_each_entry(cn, &cgroup_list, list)
294 		cn->used = true;
295 
296 	return 0;
297 }
298 
299 /* collect all cgroups first and then match with the pattern */
300 static int match_cgroups(const char *str)
301 {
302 	char mnt[PATH_MAX];
303 	const char *p, *e, *eos = str + strlen(str);
304 	struct cgroup_name *cn;
305 	regex_t reg;
306 	int prefix_len;
307 	char *s;
308 
309 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
310 		return -1;
311 
312 	/* cgroup_name will have a full path, skip the root directory */
313 	prefix_len = strlen(mnt);
314 
315 	/* collect all cgroups in the cgroup_list */
316 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
317 		return -1;
318 
319 	for (;;) {
320 		p = strchr(str, ',');
321 		e = p ? p : eos;
322 
323 		/* allow empty cgroups, i.e., skip */
324 		if (e - str) {
325 			/* termination added */
326 			s = strndup(str, e - str);
327 			if (!s)
328 				return -1;
329 			if (regcomp(&reg, s, REG_NOSUB)) {
330 				free(s);
331 				return -1;
332 			}
333 
334 			/* check cgroup name with the pattern */
335 			list_for_each_entry(cn, &cgroup_list, list) {
336 				char *name = cn->name + prefix_len;
337 
338 				if (name[0] == '/' && name[1])
339 					name++;
340 				if (!regexec(&reg, name, 0, NULL, 0))
341 					cn->used = true;
342 			}
343 			regfree(&reg);
344 			free(s);
345 		} else {
346 			/* first entry to root cgroup */
347 			cn = list_first_entry(&cgroup_list, struct cgroup_name,
348 					      list);
349 			cn->used = true;
350 		}
351 
352 		if (!p)
353 			break;
354 		str = p+1;
355 	}
356 	return prefix_len;
357 }
358 
359 int parse_cgroups(const struct option *opt, const char *str,
360 		  int unset __maybe_unused)
361 {
362 	struct evlist *evlist = *(struct evlist **)opt->value;
363 	struct evsel *counter;
364 	struct cgroup *cgrp = NULL;
365 	const char *p, *e, *eos = str + strlen(str);
366 	char *s;
367 	int ret, i;
368 
369 	if (list_empty(&evlist->core.entries)) {
370 		fprintf(stderr, "must define events before cgroups\n");
371 		return -1;
372 	}
373 
374 	for (;;) {
375 		p = strchr(str, ',');
376 		e = p ? p : eos;
377 
378 		/* allow empty cgroups, i.e., skip */
379 		if (e - str) {
380 			/* termination added */
381 			s = strndup(str, e - str);
382 			if (!s)
383 				return -1;
384 			ret = add_cgroup(evlist, s);
385 			free(s);
386 			if (ret)
387 				return -1;
388 		}
389 		/* nr_cgroups is increased een for empty cgroups */
390 		nr_cgroups++;
391 		if (!p)
392 			break;
393 		str = p+1;
394 	}
395 	/* for the case one cgroup combine to multiple events */
396 	i = 0;
397 	if (nr_cgroups == 1) {
398 		evlist__for_each_entry(evlist, counter) {
399 			if (i == 0)
400 				cgrp = counter->cgrp;
401 			else {
402 				counter->cgrp = cgrp;
403 				refcount_inc(&cgrp->refcnt);
404 			}
405 			i++;
406 		}
407 	}
408 	return 0;
409 }
410 
411 static bool has_pattern_string(const char *str)
412 {
413 	return !!strpbrk(str, "{}[]()|*+?^$");
414 }
415 
416 int evlist__expand_cgroup(struct evlist *evlist, const char *str,
417 			  struct rblist *metric_events, 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 	if (metric_events) {
444 		orig_metric_events = *metric_events;
445 		rblist__init(metric_events);
446 	} else {
447 		rblist__init(&orig_metric_events);
448 	}
449 
450 	if (has_pattern_string(str))
451 		prefix_len = match_cgroups(str);
452 	else
453 		prefix_len = list_cgroups(str);
454 
455 	if (prefix_len < 0)
456 		goto out_err;
457 
458 	list_for_each_entry(cn, &cgroup_list, list) {
459 		char *name;
460 
461 		if (!cn->used)
462 			continue;
463 
464 		/* cgroup_name might have a full path, skip the prefix */
465 		name = cn->name + prefix_len;
466 		if (name[0] == '/' && name[1])
467 			name++;
468 
469 		/* the cgroup can go away in the meantime */
470 		cgrp = cgroup__new(name, open_cgroup);
471 		if (cgrp == NULL)
472 			continue;
473 
474 		leader = NULL;
475 		evlist__for_each_entry(orig_list, pos) {
476 			evsel = evsel__clone(pos);
477 			if (evsel == NULL)
478 				goto out_err;
479 
480 			cgroup__put(evsel->cgrp);
481 			evsel->cgrp = cgroup__get(cgrp);
482 
483 			if (evsel__is_group_leader(pos))
484 				leader = evsel;
485 			evsel__set_leader(evsel, leader);
486 
487 			evlist__add(tmp_list, evsel);
488 		}
489 		/* cgroup__new() has a refcount, release it here */
490 		cgroup__put(cgrp);
491 		nr_cgroups++;
492 
493 		if (metric_events) {
494 			if (metricgroup__copy_metric_events(tmp_list, cgrp,
495 							    metric_events,
496 							    &orig_metric_events) < 0)
497 				goto out_err;
498 		}
499 
500 		evlist__splice_list_tail(evlist, &tmp_list->core.entries);
501 		tmp_list->core.nr_entries = 0;
502 	}
503 
504 	if (list_empty(&evlist->core.entries)) {
505 		fprintf(stderr, "no cgroup matched: %s\n", str);
506 		goto out_err;
507 	}
508 
509 	ret = 0;
510 	cgrp_event_expanded = true;
511 
512 out_err:
513 	evlist__delete(orig_list);
514 	evlist__delete(tmp_list);
515 	rblist__exit(&orig_metric_events);
516 	release_cgroup_list();
517 
518 	return ret;
519 }
520 
521 static struct cgroup *__cgroup__findnew(struct rb_root *root, uint64_t id,
522 					bool create, const char *path)
523 {
524 	struct rb_node **p = &root->rb_node;
525 	struct rb_node *parent = NULL;
526 	struct cgroup *cgrp;
527 
528 	while (*p != NULL) {
529 		parent = *p;
530 		cgrp = rb_entry(parent, struct cgroup, node);
531 
532 		if (cgrp->id == id)
533 			return cgrp;
534 
535 		if (cgrp->id < id)
536 			p = &(*p)->rb_left;
537 		else
538 			p = &(*p)->rb_right;
539 	}
540 
541 	if (!create)
542 		return NULL;
543 
544 	cgrp = malloc(sizeof(*cgrp));
545 	if (cgrp == NULL)
546 		return NULL;
547 
548 	cgrp->name = strdup(path);
549 	if (cgrp->name == NULL) {
550 		free(cgrp);
551 		return NULL;
552 	}
553 
554 	cgrp->fd = -1;
555 	cgrp->id = id;
556 	refcount_set(&cgrp->refcnt, 1);
557 
558 	rb_link_node(&cgrp->node, parent, p);
559 	rb_insert_color(&cgrp->node, root);
560 
561 	return cgrp;
562 }
563 
564 struct cgroup *cgroup__findnew(struct perf_env *env, uint64_t id,
565 			       const char *path)
566 {
567 	struct cgroup *cgrp;
568 
569 	down_write(&env->cgroups.lock);
570 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, true, path);
571 	up_write(&env->cgroups.lock);
572 	return cgrp;
573 }
574 
575 struct cgroup *__cgroup__find(struct rb_root *root, uint64_t id)
576 {
577 	return __cgroup__findnew(root, id, /*create=*/false, /*path=*/NULL);
578 }
579 
580 struct cgroup *cgroup__find(struct perf_env *env, uint64_t id)
581 {
582 	struct cgroup *cgrp;
583 
584 	down_read(&env->cgroups.lock);
585 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, false, NULL);
586 	up_read(&env->cgroups.lock);
587 	return cgrp;
588 }
589 
590 void perf_env__purge_cgroups(struct perf_env *env)
591 {
592 	struct rb_node *node;
593 	struct cgroup *cgrp;
594 
595 	down_write(&env->cgroups.lock);
596 	while (!RB_EMPTY_ROOT(&env->cgroups.tree)) {
597 		node = rb_first(&env->cgroups.tree);
598 		cgrp = rb_entry(node, struct cgroup, node);
599 
600 		rb_erase(node, &env->cgroups.tree);
601 		cgroup__put(cgrp);
602 	}
603 	up_write(&env->cgroups.lock);
604 }
605 
606 void read_all_cgroups(struct rb_root *root)
607 {
608 	char mnt[PATH_MAX];
609 	struct cgroup_name *cn;
610 	int prefix_len;
611 
612 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
613 		return;
614 
615 	/* cgroup_name will have a full path, skip the root directory */
616 	prefix_len = strlen(mnt);
617 
618 	/* collect all cgroups in the cgroup_list */
619 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
620 		return;
621 
622 	list_for_each_entry(cn, &cgroup_list, list) {
623 		const char *name;
624 		u64 cgrp_id;
625 
626 		/* cgroup_name might have a full path, skip the prefix */
627 		name = cn->name + prefix_len;
628 		if (name[0] == '\0')
629 			name = "/";
630 
631 		cgrp_id = __read_cgroup_id(cn->name);
632 		__cgroup__findnew(root, cgrp_id, /*create=*/true, name);
633 	}
634 
635 	release_cgroup_list();
636 }
637