xref: /openbmc/linux/tools/perf/util/cgroup.c (revision 69e874db)
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 <fcntl.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include <api/fs/fs.h>
16 #include <ftw.h>
17 #include <regex.h>
18 
19 int nr_cgroups;
20 
21 /* used to match cgroup name with patterns */
22 struct cgroup_name {
23 	struct list_head list;
24 	bool used;
25 	char name[];
26 };
27 static LIST_HEAD(cgroup_list);
28 
29 static int open_cgroup(const char *name)
30 {
31 	char path[PATH_MAX + 1];
32 	char mnt[PATH_MAX + 1];
33 	int fd;
34 
35 
36 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
37 		return -1;
38 
39 	scnprintf(path, PATH_MAX, "%s/%s", mnt, name);
40 
41 	fd = open(path, O_RDONLY);
42 	if (fd == -1)
43 		fprintf(stderr, "no access to cgroup %s\n", path);
44 
45 	return fd;
46 }
47 
48 #ifdef HAVE_FILE_HANDLE
49 int read_cgroup_id(struct cgroup *cgrp)
50 {
51 	char path[PATH_MAX + 1];
52 	char mnt[PATH_MAX + 1];
53 	struct {
54 		struct file_handle fh;
55 		uint64_t cgroup_id;
56 	} handle;
57 	int mount_id;
58 
59 	if (cgroupfs_find_mountpoint(mnt, PATH_MAX + 1, "perf_event"))
60 		return -1;
61 
62 	scnprintf(path, PATH_MAX, "%s/%s", mnt, cgrp->name);
63 
64 	handle.fh.handle_bytes = sizeof(handle.cgroup_id);
65 	if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0)
66 		return -1;
67 
68 	cgrp->id = handle.cgroup_id;
69 	return 0;
70 }
71 #endif  /* HAVE_FILE_HANDLE */
72 
73 static struct cgroup *evlist__find_cgroup(struct evlist *evlist, const char *str)
74 {
75 	struct evsel *counter;
76 	/*
77 	 * check if cgrp is already defined, if so we reuse it
78 	 */
79 	evlist__for_each_entry(evlist, counter) {
80 		if (!counter->cgrp)
81 			continue;
82 		if (!strcmp(counter->cgrp->name, str))
83 			return cgroup__get(counter->cgrp);
84 	}
85 
86 	return NULL;
87 }
88 
89 static struct cgroup *cgroup__new(const char *name, bool do_open)
90 {
91 	struct cgroup *cgroup = zalloc(sizeof(*cgroup));
92 
93 	if (cgroup != NULL) {
94 		refcount_set(&cgroup->refcnt, 1);
95 
96 		cgroup->name = strdup(name);
97 		if (!cgroup->name)
98 			goto out_err;
99 
100 		if (do_open) {
101 			cgroup->fd = open_cgroup(name);
102 			if (cgroup->fd == -1)
103 				goto out_free_name;
104 		} else {
105 			cgroup->fd = -1;
106 		}
107 	}
108 
109 	return cgroup;
110 
111 out_free_name:
112 	zfree(&cgroup->name);
113 out_err:
114 	free(cgroup);
115 	return NULL;
116 }
117 
118 struct cgroup *evlist__findnew_cgroup(struct evlist *evlist, const char *name)
119 {
120 	struct cgroup *cgroup = evlist__find_cgroup(evlist, name);
121 
122 	return cgroup ?: cgroup__new(name, true);
123 }
124 
125 static int add_cgroup(struct evlist *evlist, const char *str)
126 {
127 	struct evsel *counter;
128 	struct cgroup *cgrp = evlist__findnew_cgroup(evlist, str);
129 	int n;
130 
131 	if (!cgrp)
132 		return -1;
133 	/*
134 	 * find corresponding event
135 	 * if add cgroup N, then need to find event N
136 	 */
137 	n = 0;
138 	evlist__for_each_entry(evlist, counter) {
139 		if (n == nr_cgroups)
140 			goto found;
141 		n++;
142 	}
143 
144 	cgroup__put(cgrp);
145 	return -1;
146 found:
147 	counter->cgrp = cgrp;
148 	return 0;
149 }
150 
151 static void cgroup__delete(struct cgroup *cgroup)
152 {
153 	if (cgroup->fd >= 0)
154 		close(cgroup->fd);
155 	zfree(&cgroup->name);
156 	free(cgroup);
157 }
158 
159 void cgroup__put(struct cgroup *cgrp)
160 {
161 	if (cgrp && refcount_dec_and_test(&cgrp->refcnt)) {
162 		cgroup__delete(cgrp);
163 	}
164 }
165 
166 struct cgroup *cgroup__get(struct cgroup *cgroup)
167 {
168        if (cgroup)
169 		refcount_inc(&cgroup->refcnt);
170        return cgroup;
171 }
172 
173 static void evsel__set_default_cgroup(struct evsel *evsel, struct cgroup *cgroup)
174 {
175 	if (evsel->cgrp == NULL)
176 		evsel->cgrp = cgroup__get(cgroup);
177 }
178 
179 void evlist__set_default_cgroup(struct evlist *evlist, struct cgroup *cgroup)
180 {
181 	struct evsel *evsel;
182 
183 	evlist__for_each_entry(evlist, evsel)
184 		evsel__set_default_cgroup(evsel, cgroup);
185 }
186 
187 /* helper function for ftw() in match_cgroups and list_cgroups */
188 static int add_cgroup_name(const char *fpath, const struct stat *sb __maybe_unused,
189 			   int typeflag, struct FTW *ftwbuf __maybe_unused)
190 {
191 	struct cgroup_name *cn;
192 
193 	if (typeflag != FTW_D)
194 		return 0;
195 
196 	cn = malloc(sizeof(*cn) + strlen(fpath) + 1);
197 	if (cn == NULL)
198 		return -1;
199 
200 	cn->used = false;
201 	strcpy(cn->name, fpath);
202 
203 	list_add_tail(&cn->list, &cgroup_list);
204 	return 0;
205 }
206 
207 static void release_cgroup_list(void)
208 {
209 	struct cgroup_name *cn;
210 
211 	while (!list_empty(&cgroup_list)) {
212 		cn = list_first_entry(&cgroup_list, struct cgroup_name, list);
213 		list_del(&cn->list);
214 		free(cn);
215 	}
216 }
217 
218 /* collect given cgroups only */
219 static int list_cgroups(const char *str)
220 {
221 	const char *p, *e, *eos = str + strlen(str);
222 	struct cgroup_name *cn;
223 	char *s;
224 
225 	/* use given name as is - for testing purpose */
226 	for (;;) {
227 		p = strchr(str, ',');
228 		e = p ? p : eos;
229 
230 		if (e - str) {
231 			int ret;
232 
233 			s = strndup(str, e - str);
234 			if (!s)
235 				return -1;
236 			/* pretend if it's added by ftw() */
237 			ret = add_cgroup_name(s, NULL, FTW_D, NULL);
238 			free(s);
239 			if (ret)
240 				return -1;
241 		} else {
242 			if (add_cgroup_name("", NULL, FTW_D, NULL) < 0)
243 				return -1;
244 		}
245 
246 		if (!p)
247 			break;
248 		str = p+1;
249 	}
250 
251 	/* these groups will be used */
252 	list_for_each_entry(cn, &cgroup_list, list)
253 		cn->used = true;
254 
255 	return 0;
256 }
257 
258 /* collect all cgroups first and then match with the pattern */
259 static int match_cgroups(const char *str)
260 {
261 	char mnt[PATH_MAX];
262 	const char *p, *e, *eos = str + strlen(str);
263 	struct cgroup_name *cn;
264 	regex_t reg;
265 	int prefix_len;
266 	char *s;
267 
268 	if (cgroupfs_find_mountpoint(mnt, sizeof(mnt), "perf_event"))
269 		return -1;
270 
271 	/* cgroup_name will have a full path, skip the root directory */
272 	prefix_len = strlen(mnt);
273 
274 	/* collect all cgroups in the cgroup_list */
275 	if (nftw(mnt, add_cgroup_name, 20, 0) < 0)
276 		return -1;
277 
278 	for (;;) {
279 		p = strchr(str, ',');
280 		e = p ? p : eos;
281 
282 		/* allow empty cgroups, i.e., skip */
283 		if (e - str) {
284 			/* termination added */
285 			s = strndup(str, e - str);
286 			if (!s)
287 				return -1;
288 			if (regcomp(&reg, s, REG_NOSUB)) {
289 				free(s);
290 				return -1;
291 			}
292 
293 			/* check cgroup name with the pattern */
294 			list_for_each_entry(cn, &cgroup_list, list) {
295 				char *name = cn->name + prefix_len;
296 
297 				if (name[0] == '/' && name[1])
298 					name++;
299 				if (!regexec(&reg, name, 0, NULL, 0))
300 					cn->used = true;
301 			}
302 			regfree(&reg);
303 			free(s);
304 		} else {
305 			/* first entry to root cgroup */
306 			cn = list_first_entry(&cgroup_list, struct cgroup_name,
307 					      list);
308 			cn->used = true;
309 		}
310 
311 		if (!p)
312 			break;
313 		str = p+1;
314 	}
315 	return prefix_len;
316 }
317 
318 int parse_cgroups(const struct option *opt, const char *str,
319 		  int unset __maybe_unused)
320 {
321 	struct evlist *evlist = *(struct evlist **)opt->value;
322 	struct evsel *counter;
323 	struct cgroup *cgrp = NULL;
324 	const char *p, *e, *eos = str + strlen(str);
325 	char *s;
326 	int ret, i;
327 
328 	if (list_empty(&evlist->core.entries)) {
329 		fprintf(stderr, "must define events before cgroups\n");
330 		return -1;
331 	}
332 
333 	for (;;) {
334 		p = strchr(str, ',');
335 		e = p ? p : eos;
336 
337 		/* allow empty cgroups, i.e., skip */
338 		if (e - str) {
339 			/* termination added */
340 			s = strndup(str, e - str);
341 			if (!s)
342 				return -1;
343 			ret = add_cgroup(evlist, s);
344 			free(s);
345 			if (ret)
346 				return -1;
347 		}
348 		/* nr_cgroups is increased een for empty cgroups */
349 		nr_cgroups++;
350 		if (!p)
351 			break;
352 		str = p+1;
353 	}
354 	/* for the case one cgroup combine to multiple events */
355 	i = 0;
356 	if (nr_cgroups == 1) {
357 		evlist__for_each_entry(evlist, counter) {
358 			if (i == 0)
359 				cgrp = counter->cgrp;
360 			else {
361 				counter->cgrp = cgrp;
362 				refcount_inc(&cgrp->refcnt);
363 			}
364 			i++;
365 		}
366 	}
367 	return 0;
368 }
369 
370 static bool has_pattern_string(const char *str)
371 {
372 	return !!strpbrk(str, "{}[]()|*+?^$");
373 }
374 
375 int evlist__expand_cgroup(struct evlist *evlist, const char *str,
376 			  struct rblist *metric_events, bool open_cgroup)
377 {
378 	struct evlist *orig_list, *tmp_list;
379 	struct evsel *pos, *evsel, *leader;
380 	struct rblist orig_metric_events;
381 	struct cgroup *cgrp = NULL;
382 	struct cgroup_name *cn;
383 	int ret = -1;
384 	int prefix_len;
385 
386 	if (evlist->core.nr_entries == 0) {
387 		fprintf(stderr, "must define events before cgroups\n");
388 		return -EINVAL;
389 	}
390 
391 	orig_list = evlist__new();
392 	tmp_list = evlist__new();
393 	if (orig_list == NULL || tmp_list == NULL) {
394 		fprintf(stderr, "memory allocation failed\n");
395 		return -ENOMEM;
396 	}
397 
398 	/* save original events and init evlist */
399 	evlist__splice_list_tail(orig_list, &evlist->core.entries);
400 	evlist->core.nr_entries = 0;
401 
402 	if (metric_events) {
403 		orig_metric_events = *metric_events;
404 		rblist__init(metric_events);
405 	} else {
406 		rblist__init(&orig_metric_events);
407 	}
408 
409 	if (has_pattern_string(str))
410 		prefix_len = match_cgroups(str);
411 	else
412 		prefix_len = list_cgroups(str);
413 
414 	if (prefix_len < 0)
415 		goto out_err;
416 
417 	list_for_each_entry(cn, &cgroup_list, list) {
418 		char *name;
419 
420 		if (!cn->used)
421 			continue;
422 
423 		/* cgroup_name might have a full path, skip the prefix */
424 		name = cn->name + prefix_len;
425 		if (name[0] == '/' && name[1])
426 			name++;
427 		cgrp = cgroup__new(name, open_cgroup);
428 		if (cgrp == NULL)
429 			goto out_err;
430 
431 		leader = NULL;
432 		evlist__for_each_entry(orig_list, pos) {
433 			evsel = evsel__clone(pos);
434 			if (evsel == NULL)
435 				goto out_err;
436 
437 			cgroup__put(evsel->cgrp);
438 			evsel->cgrp = cgroup__get(cgrp);
439 
440 			if (evsel__is_group_leader(pos))
441 				leader = evsel;
442 			evsel->leader = leader;
443 
444 			evlist__add(tmp_list, evsel);
445 		}
446 		/* cgroup__new() has a refcount, release it here */
447 		cgroup__put(cgrp);
448 		nr_cgroups++;
449 
450 		if (metric_events) {
451 			perf_stat__collect_metric_expr(tmp_list);
452 			if (metricgroup__copy_metric_events(tmp_list, cgrp,
453 							    metric_events,
454 							    &orig_metric_events) < 0)
455 				goto out_err;
456 		}
457 
458 		evlist__splice_list_tail(evlist, &tmp_list->core.entries);
459 		tmp_list->core.nr_entries = 0;
460 	}
461 
462 	if (list_empty(&evlist->core.entries)) {
463 		fprintf(stderr, "no cgroup matched: %s\n", str);
464 		goto out_err;
465 	}
466 
467 	ret = 0;
468 
469 out_err:
470 	evlist__delete(orig_list);
471 	evlist__delete(tmp_list);
472 	rblist__exit(&orig_metric_events);
473 	release_cgroup_list();
474 
475 	return ret;
476 }
477 
478 static struct cgroup *__cgroup__findnew(struct rb_root *root, uint64_t id,
479 					bool create, const char *path)
480 {
481 	struct rb_node **p = &root->rb_node;
482 	struct rb_node *parent = NULL;
483 	struct cgroup *cgrp;
484 
485 	while (*p != NULL) {
486 		parent = *p;
487 		cgrp = rb_entry(parent, struct cgroup, node);
488 
489 		if (cgrp->id == id)
490 			return cgrp;
491 
492 		if (cgrp->id < id)
493 			p = &(*p)->rb_left;
494 		else
495 			p = &(*p)->rb_right;
496 	}
497 
498 	if (!create)
499 		return NULL;
500 
501 	cgrp = malloc(sizeof(*cgrp));
502 	if (cgrp == NULL)
503 		return NULL;
504 
505 	cgrp->name = strdup(path);
506 	if (cgrp->name == NULL) {
507 		free(cgrp);
508 		return NULL;
509 	}
510 
511 	cgrp->fd = -1;
512 	cgrp->id = id;
513 	refcount_set(&cgrp->refcnt, 1);
514 
515 	rb_link_node(&cgrp->node, parent, p);
516 	rb_insert_color(&cgrp->node, root);
517 
518 	return cgrp;
519 }
520 
521 struct cgroup *cgroup__findnew(struct perf_env *env, uint64_t id,
522 			       const char *path)
523 {
524 	struct cgroup *cgrp;
525 
526 	down_write(&env->cgroups.lock);
527 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, true, path);
528 	up_write(&env->cgroups.lock);
529 	return cgrp;
530 }
531 
532 struct cgroup *cgroup__find(struct perf_env *env, uint64_t id)
533 {
534 	struct cgroup *cgrp;
535 
536 	down_read(&env->cgroups.lock);
537 	cgrp = __cgroup__findnew(&env->cgroups.tree, id, false, NULL);
538 	up_read(&env->cgroups.lock);
539 	return cgrp;
540 }
541 
542 void perf_env__purge_cgroups(struct perf_env *env)
543 {
544 	struct rb_node *node;
545 	struct cgroup *cgrp;
546 
547 	down_write(&env->cgroups.lock);
548 	while (!RB_EMPTY_ROOT(&env->cgroups.tree)) {
549 		node = rb_first(&env->cgroups.tree);
550 		cgrp = rb_entry(node, struct cgroup, node);
551 
552 		rb_erase(node, &env->cgroups.tree);
553 		cgroup__put(cgrp);
554 	}
555 	up_write(&env->cgroups.lock);
556 }
557