xref: /linux/tools/perf/util/cgroup.c (revision 6beeaf48db6c548fcfc2ad32739d33af2fef3a5b)
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 int read_cgroup_id(struct cgroup *cgrp)
52 {
53 	char path[PATH_MAX + 1];
54 	char mnt[PATH_MAX + 1];
55 	struct {
56 		struct file_handle fh;
57 		uint64_t cgroup_id;
58 	} handle;
59 	int mount_id;
60 
61 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
62 		return -1;
63 
64 	scnprintf(path, PATH_MAX, "%s/%s", mnt, cgrp->name);
65 
66 	handle.fh.handle_bytes = sizeof(handle.cgroup_id);
67 	if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0)
68 		return -1;
69 
70 	cgrp->id = handle.cgroup_id;
71 	return 0;
72 }
73 #endif  /* HAVE_FILE_HANDLE */
74 
75 #ifndef CGROUP2_SUPER_MAGIC
76 #define CGROUP2_SUPER_MAGIC  0x63677270
77 #endif
78 
79 int cgroup_is_v2(const char *subsys)
80 {
81 	char mnt[PATH_MAX + 1];
82 	struct statfs stbuf;
83 
84 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, subsys))
85 		return -1;
86 
87 	if (statfs(mnt, &stbuf) < 0)
88 		return -1;
89 
90 	return (stbuf.f_type == CGROUP2_SUPER_MAGIC);
91 }
92 
93 static struct cgroup *evlist__find_cgroup(struct evlist *evlist, const char *str)
94 {
95 	struct evsel *counter;
96 	/*
97 	 * check if cgrp is already defined, if so we reuse it
98 	 */
99 	evlist__for_each_entry(evlist, counter) {
100 		if (!counter->cgrp)
101 			continue;
102 		if (!strcmp(counter->cgrp->name, str))
103 			return cgroup__get(counter->cgrp);
104 	}
105 
106 	return NULL;
107 }
108 
109 static struct cgroup *cgroup__new(const char *name, bool do_open)
110 {
111 	struct cgroup *cgroup = zalloc(sizeof(*cgroup));
112 
113 	if (cgroup != NULL) {
114 		refcount_set(&cgroup->refcnt, 1);
115 
116 		cgroup->name = strdup(name);
117 		if (!cgroup->name)
118 			goto out_err;
119 
120 		if (do_open) {
121 			cgroup->fd = open_cgroup(name);
122 			if (cgroup->fd == -1)
123 				goto out_free_name;
124 		} else {
125 			cgroup->fd = -1;
126 		}
127 	}
128 
129 	return cgroup;
130 
131 out_free_name:
132 	zfree(&cgroup->name);
133 out_err:
134 	free(cgroup);
135 	return NULL;
136 }
137 
138 struct cgroup *evlist__findnew_cgroup(struct evlist *evlist, const char *name)
139 {
140 	struct cgroup *cgroup = evlist__find_cgroup(evlist, name);
141 
142 	return cgroup ?: cgroup__new(name, true);
143 }
144 
145 static int add_cgroup(struct evlist *evlist, const char *str)
146 {
147 	struct evsel *counter;
148 	struct cgroup *cgrp = evlist__findnew_cgroup(evlist, str);
149 	int n;
150 
151 	if (!cgrp)
152 		return -1;
153 	/*
154 	 * find corresponding event
155 	 * if add cgroup N, then need to find event N
156 	 */
157 	n = 0;
158 	evlist__for_each_entry(evlist, counter) {
159 		if (n == nr_cgroups)
160 			goto found;
161 		n++;
162 	}
163 
164 	cgroup__put(cgrp);
165 	return -1;
166 found:
167 	counter->cgrp = cgrp;
168 	return 0;
169 }
170 
171 static void cgroup__delete(struct cgroup *cgroup)
172 {
173 	if (cgroup->fd >= 0)
174 		close(cgroup->fd);
175 	zfree(&cgroup->name);
176 	free(cgroup);
177 }
178 
179 void cgroup__put(struct cgroup *cgrp)
180 {
181 	if (cgrp && refcount_dec_and_test(&cgrp->refcnt)) {
182 		cgroup__delete(cgrp);
183 	}
184 }
185 
186 struct cgroup *cgroup__get(struct cgroup *cgroup)
187 {
188        if (cgroup)
189 		refcount_inc(&cgroup->refcnt);
190        return cgroup;
191 }
192 
193 static void evsel__set_default_cgroup(struct evsel *evsel, struct cgroup *cgroup)
194 {
195 	if (evsel->cgrp == NULL)
196 		evsel->cgrp = cgroup__get(cgroup);
197 }
198 
199 void evlist__set_default_cgroup(struct evlist *evlist, struct cgroup *cgroup)
200 {
201 	struct evsel *evsel;
202 
203 	evlist__for_each_entry(evlist, evsel)
204 		evsel__set_default_cgroup(evsel, cgroup);
205 }
206 
207 /* helper function for ftw() in match_cgroups and list_cgroups */
208 static int add_cgroup_name(const char *fpath, const struct stat *sb __maybe_unused,
209 			   int typeflag, struct FTW *ftwbuf __maybe_unused)
210 {
211 	struct cgroup_name *cn;
212 
213 	if (typeflag != FTW_D)
214 		return 0;
215 
216 	cn = malloc(sizeof(*cn) + strlen(fpath) + 1);
217 	if (cn == NULL)
218 		return -1;
219 
220 	cn->used = false;
221 	strcpy(cn->name, fpath);
222 
223 	list_add_tail(&cn->list, &cgroup_list);
224 	return 0;
225 }
226 
227 static void release_cgroup_list(void)
228 {
229 	struct cgroup_name *cn;
230 
231 	while (!list_empty(&cgroup_list)) {
232 		cn = list_first_entry(&cgroup_list, struct cgroup_name, list);
233 		list_del(&cn->list);
234 		free(cn);
235 	}
236 }
237 
238 /* collect given cgroups only */
239 static int list_cgroups(const char *str)
240 {
241 	const char *p, *e, *eos = str + strlen(str);
242 	struct cgroup_name *cn;
243 	char *s;
244 
245 	/* use given name as is - for testing purpose */
246 	for (;;) {
247 		p = strchr(str, ',');
248 		e = p ? p : eos;
249 
250 		if (e - str) {
251 			int ret;
252 
253 			s = strndup(str, e - str);
254 			if (!s)
255 				return -1;
256 			/* pretend if it's added by ftw() */
257 			ret = add_cgroup_name(s, NULL, FTW_D, NULL);
258 			free(s);
259 			if (ret)
260 				return -1;
261 		} else {
262 			if (add_cgroup_name("", NULL, FTW_D, NULL) < 0)
263 				return -1;
264 		}
265 
266 		if (!p)
267 			break;
268 		str = p+1;
269 	}
270 
271 	/* these groups will be used */
272 	list_for_each_entry(cn, &cgroup_list, list)
273 		cn->used = true;
274 
275 	return 0;
276 }
277 
278 /* collect all cgroups first and then match with the pattern */
279 static int match_cgroups(const char *str)
280 {
281 	char mnt[PATH_MAX];
282 	const char *p, *e, *eos = str + strlen(str);
283 	struct cgroup_name *cn;
284 	regex_t reg;
285 	int prefix_len;
286 	char *s;
287 
288 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
289 		return -1;
290 
291 	/* cgroup_name will have a full path, skip the root directory */
292 	prefix_len = strlen(mnt);
293 
294 	/* collect all cgroups in the cgroup_list */
295 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
296 		return -1;
297 
298 	for (;;) {
299 		p = strchr(str, ',');
300 		e = p ? p : eos;
301 
302 		/* allow empty cgroups, i.e., skip */
303 		if (e - str) {
304 			/* termination added */
305 			s = strndup(str, e - str);
306 			if (!s)
307 				return -1;
308 			if (regcomp(&reg, s, REG_NOSUB)) {
309 				free(s);
310 				return -1;
311 			}
312 
313 			/* check cgroup name with the pattern */
314 			list_for_each_entry(cn, &cgroup_list, list) {
315 				char *name = cn->name + prefix_len;
316 
317 				if (name[0] == '/' && name[1])
318 					name++;
319 				if (!regexec(&reg, name, 0, NULL, 0))
320 					cn->used = true;
321 			}
322 			regfree(&reg);
323 			free(s);
324 		} else {
325 			/* first entry to root cgroup */
326 			cn = list_first_entry(&cgroup_list, struct cgroup_name,
327 					      list);
328 			cn->used = true;
329 		}
330 
331 		if (!p)
332 			break;
333 		str = p+1;
334 	}
335 	return prefix_len;
336 }
337 
338 int parse_cgroups(const struct option *opt, const char *str,
339 		  int unset __maybe_unused)
340 {
341 	struct evlist *evlist = *(struct evlist **)opt->value;
342 	struct evsel *counter;
343 	struct cgroup *cgrp = NULL;
344 	const char *p, *e, *eos = str + strlen(str);
345 	char *s;
346 	int ret, i;
347 
348 	if (list_empty(&evlist->core.entries)) {
349 		fprintf(stderr, "must define events before cgroups\n");
350 		return -1;
351 	}
352 
353 	for (;;) {
354 		p = strchr(str, ',');
355 		e = p ? p : eos;
356 
357 		/* allow empty cgroups, i.e., skip */
358 		if (e - str) {
359 			/* termination added */
360 			s = strndup(str, e - str);
361 			if (!s)
362 				return -1;
363 			ret = add_cgroup(evlist, s);
364 			free(s);
365 			if (ret)
366 				return -1;
367 		}
368 		/* nr_cgroups is increased een for empty cgroups */
369 		nr_cgroups++;
370 		if (!p)
371 			break;
372 		str = p+1;
373 	}
374 	/* for the case one cgroup combine to multiple events */
375 	i = 0;
376 	if (nr_cgroups == 1) {
377 		evlist__for_each_entry(evlist, counter) {
378 			if (i == 0)
379 				cgrp = counter->cgrp;
380 			else {
381 				counter->cgrp = cgrp;
382 				refcount_inc(&cgrp->refcnt);
383 			}
384 			i++;
385 		}
386 	}
387 	return 0;
388 }
389 
390 static bool has_pattern_string(const char *str)
391 {
392 	return !!strpbrk(str, "{}[]()|*+?^$");
393 }
394 
395 int evlist__expand_cgroup(struct evlist *evlist, const char *str,
396 			  struct rblist *metric_events, bool open_cgroup)
397 {
398 	struct evlist *orig_list, *tmp_list;
399 	struct evsel *pos, *evsel, *leader;
400 	struct rblist orig_metric_events;
401 	struct cgroup *cgrp = NULL;
402 	struct cgroup_name *cn;
403 	int ret = -1;
404 	int prefix_len;
405 
406 	if (evlist->core.nr_entries == 0) {
407 		fprintf(stderr, "must define events before cgroups\n");
408 		return -EINVAL;
409 	}
410 
411 	orig_list = evlist__new();
412 	tmp_list = evlist__new();
413 	if (orig_list == NULL || tmp_list == NULL) {
414 		fprintf(stderr, "memory allocation failed\n");
415 		return -ENOMEM;
416 	}
417 
418 	/* save original events and init evlist */
419 	evlist__splice_list_tail(orig_list, &evlist->core.entries);
420 	evlist->core.nr_entries = 0;
421 
422 	if (metric_events) {
423 		orig_metric_events = *metric_events;
424 		rblist__init(metric_events);
425 	} else {
426 		rblist__init(&orig_metric_events);
427 	}
428 
429 	if (has_pattern_string(str))
430 		prefix_len = match_cgroups(str);
431 	else
432 		prefix_len = list_cgroups(str);
433 
434 	if (prefix_len < 0)
435 		goto out_err;
436 
437 	list_for_each_entry(cn, &cgroup_list, list) {
438 		char *name;
439 
440 		if (!cn->used)
441 			continue;
442 
443 		/* cgroup_name might have a full path, skip the prefix */
444 		name = cn->name + prefix_len;
445 		if (name[0] == '/' && name[1])
446 			name++;
447 		cgrp = cgroup__new(name, open_cgroup);
448 		if (cgrp == NULL)
449 			goto out_err;
450 
451 		leader = NULL;
452 		evlist__for_each_entry(orig_list, pos) {
453 			evsel = evsel__clone(pos);
454 			if (evsel == NULL)
455 				goto out_err;
456 
457 			cgroup__put(evsel->cgrp);
458 			evsel->cgrp = cgroup__get(cgrp);
459 
460 			if (evsel__is_group_leader(pos))
461 				leader = evsel;
462 			evsel__set_leader(evsel, leader);
463 
464 			evlist__add(tmp_list, evsel);
465 		}
466 		/* cgroup__new() has a refcount, release it here */
467 		cgroup__put(cgrp);
468 		nr_cgroups++;
469 
470 		if (metric_events) {
471 			perf_stat__collect_metric_expr(tmp_list);
472 			if (metricgroup__copy_metric_events(tmp_list, cgrp,
473 							    metric_events,
474 							    &orig_metric_events) < 0)
475 				goto out_err;
476 		}
477 
478 		evlist__splice_list_tail(evlist, &tmp_list->core.entries);
479 		tmp_list->core.nr_entries = 0;
480 	}
481 
482 	if (list_empty(&evlist->core.entries)) {
483 		fprintf(stderr, "no cgroup matched: %s\n", str);
484 		goto out_err;
485 	}
486 
487 	ret = 0;
488 	cgrp_event_expanded = true;
489 
490 out_err:
491 	evlist__delete(orig_list);
492 	evlist__delete(tmp_list);
493 	rblist__exit(&orig_metric_events);
494 	release_cgroup_list();
495 
496 	return ret;
497 }
498 
499 static struct cgroup *__cgroup__findnew(struct rb_root *root, uint64_t id,
500 					bool create, const char *path)
501 {
502 	struct rb_node **p = &root->rb_node;
503 	struct rb_node *parent = NULL;
504 	struct cgroup *cgrp;
505 
506 	while (*p != NULL) {
507 		parent = *p;
508 		cgrp = rb_entry(parent, struct cgroup, node);
509 
510 		if (cgrp->id == id)
511 			return cgrp;
512 
513 		if (cgrp->id < id)
514 			p = &(*p)->rb_left;
515 		else
516 			p = &(*p)->rb_right;
517 	}
518 
519 	if (!create)
520 		return NULL;
521 
522 	cgrp = malloc(sizeof(*cgrp));
523 	if (cgrp == NULL)
524 		return NULL;
525 
526 	cgrp->name = strdup(path);
527 	if (cgrp->name == NULL) {
528 		free(cgrp);
529 		return NULL;
530 	}
531 
532 	cgrp->fd = -1;
533 	cgrp->id = id;
534 	refcount_set(&cgrp->refcnt, 1);
535 
536 	rb_link_node(&cgrp->node, parent, p);
537 	rb_insert_color(&cgrp->node, root);
538 
539 	return cgrp;
540 }
541 
542 struct cgroup *cgroup__findnew(struct perf_env *env, uint64_t id,
543 			       const char *path)
544 {
545 	struct cgroup *cgrp;
546 
547 	down_write(&env->cgroups.lock);
548 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, true, path);
549 	up_write(&env->cgroups.lock);
550 	return cgrp;
551 }
552 
553 struct cgroup *cgroup__find(struct perf_env *env, uint64_t id)
554 {
555 	struct cgroup *cgrp;
556 
557 	down_read(&env->cgroups.lock);
558 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, false, NULL);
559 	up_read(&env->cgroups.lock);
560 	return cgrp;
561 }
562 
563 void perf_env__purge_cgroups(struct perf_env *env)
564 {
565 	struct rb_node *node;
566 	struct cgroup *cgrp;
567 
568 	down_write(&env->cgroups.lock);
569 	while (!RB_EMPTY_ROOT(&env->cgroups.tree)) {
570 		node = rb_first(&env->cgroups.tree);
571 		cgrp = rb_entry(node, struct cgroup, node);
572 
573 		rb_erase(node, &env->cgroups.tree);
574 		cgroup__put(cgrp);
575 	}
576 	up_write(&env->cgroups.lock);
577 }
578