Commit | Line | Data |
---|---|---|
0a268dbd TH |
1 | #include "cgroup-internal.h" |
2 | ||
1592c9b2 | 3 | #include <linux/ctype.h> |
0a268dbd TH |
4 | #include <linux/kmod.h> |
5 | #include <linux/sort.h> | |
1592c9b2 | 6 | #include <linux/delay.h> |
0a268dbd | 7 | #include <linux/mm.h> |
c3edc401 | 8 | #include <linux/sched/signal.h> |
56cd6973 | 9 | #include <linux/sched/task.h> |
50ff9d13 | 10 | #include <linux/magic.h> |
0a268dbd TH |
11 | #include <linux/slab.h> |
12 | #include <linux/vmalloc.h> | |
13 | #include <linux/delayacct.h> | |
14 | #include <linux/pid_namespace.h> | |
15 | #include <linux/cgroupstats.h> | |
16 | ||
17 | #include <trace/events/cgroup.h> | |
18 | ||
19 | /* | |
20 | * pidlists linger the following amount before being destroyed. The goal | |
21 | * is avoiding frequent destruction in the middle of consecutive read calls | |
22 | * Expiring in the middle is a performance problem not a correctness one. | |
23 | * 1 sec should be enough. | |
24 | */ | |
25 | #define CGROUP_PIDLIST_DESTROY_DELAY HZ | |
26 | ||
27 | /* Controllers blocked by the commandline in v1 */ | |
28 | static u16 cgroup_no_v1_mask; | |
29 | ||
30 | /* | |
31 | * pidlist destructions need to be flushed on cgroup destruction. Use a | |
32 | * separate workqueue as flush domain. | |
33 | */ | |
34 | static struct workqueue_struct *cgroup_pidlist_destroy_wq; | |
35 | ||
36 | /* | |
37 | * Protects cgroup_subsys->release_agent_path. Modifying it also requires | |
38 | * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock. | |
39 | */ | |
1592c9b2 | 40 | static DEFINE_SPINLOCK(release_agent_path_lock); |
0a268dbd | 41 | |
d62beb7f | 42 | bool cgroup1_ssid_disabled(int ssid) |
0a268dbd TH |
43 | { |
44 | return cgroup_no_v1_mask & (1 << ssid); | |
45 | } | |
46 | ||
47 | /** | |
48 | * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from' | |
49 | * @from: attach to all cgroups of a given task | |
50 | * @tsk: the task to be attached | |
51 | */ | |
52 | int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk) | |
53 | { | |
54 | struct cgroup_root *root; | |
55 | int retval = 0; | |
56 | ||
57 | mutex_lock(&cgroup_mutex); | |
58 | percpu_down_write(&cgroup_threadgroup_rwsem); | |
59 | for_each_root(root) { | |
60 | struct cgroup *from_cgrp; | |
61 | ||
62 | if (root == &cgrp_dfl_root) | |
63 | continue; | |
64 | ||
65 | spin_lock_irq(&css_set_lock); | |
66 | from_cgrp = task_cgroup_from_root(from, root); | |
67 | spin_unlock_irq(&css_set_lock); | |
68 | ||
69 | retval = cgroup_attach_task(from_cgrp, tsk, false); | |
70 | if (retval) | |
71 | break; | |
72 | } | |
73 | percpu_up_write(&cgroup_threadgroup_rwsem); | |
74 | mutex_unlock(&cgroup_mutex); | |
75 | ||
76 | return retval; | |
77 | } | |
78 | EXPORT_SYMBOL_GPL(cgroup_attach_task_all); | |
79 | ||
80 | /** | |
81 | * cgroup_trasnsfer_tasks - move tasks from one cgroup to another | |
82 | * @to: cgroup to which the tasks will be moved | |
83 | * @from: cgroup in which the tasks currently reside | |
84 | * | |
85 | * Locking rules between cgroup_post_fork() and the migration path | |
86 | * guarantee that, if a task is forking while being migrated, the new child | |
87 | * is guaranteed to be either visible in the source cgroup after the | |
88 | * parent's migration is complete or put into the target cgroup. No task | |
89 | * can slip out of migration through forking. | |
90 | */ | |
91 | int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from) | |
92 | { | |
e595cd70 | 93 | DEFINE_CGROUP_MGCTX(mgctx); |
0a268dbd TH |
94 | struct cgrp_cset_link *link; |
95 | struct css_task_iter it; | |
96 | struct task_struct *task; | |
97 | int ret; | |
98 | ||
99 | if (cgroup_on_dfl(to)) | |
100 | return -EINVAL; | |
101 | ||
102 | if (!cgroup_may_migrate_to(to)) | |
103 | return -EBUSY; | |
104 | ||
105 | mutex_lock(&cgroup_mutex); | |
106 | ||
107 | percpu_down_write(&cgroup_threadgroup_rwsem); | |
108 | ||
109 | /* all tasks in @from are being moved, all csets are source */ | |
110 | spin_lock_irq(&css_set_lock); | |
111 | list_for_each_entry(link, &from->cset_links, cset_link) | |
e595cd70 | 112 | cgroup_migrate_add_src(link->cset, to, &mgctx); |
0a268dbd TH |
113 | spin_unlock_irq(&css_set_lock); |
114 | ||
e595cd70 | 115 | ret = cgroup_migrate_prepare_dst(&mgctx); |
0a268dbd TH |
116 | if (ret) |
117 | goto out_err; | |
118 | ||
119 | /* | |
120 | * Migrate tasks one-by-one until @from is empty. This fails iff | |
121 | * ->can_attach() fails. | |
122 | */ | |
123 | do { | |
124 | css_task_iter_start(&from->self, &it); | |
125 | task = css_task_iter_next(&it); | |
126 | if (task) | |
127 | get_task_struct(task); | |
128 | css_task_iter_end(&it); | |
129 | ||
130 | if (task) { | |
bfc2cf6f | 131 | ret = cgroup_migrate(task, false, &mgctx); |
0a268dbd TH |
132 | if (!ret) |
133 | trace_cgroup_transfer_tasks(to, task, false); | |
134 | put_task_struct(task); | |
135 | } | |
136 | } while (task && !ret); | |
137 | out_err: | |
e595cd70 | 138 | cgroup_migrate_finish(&mgctx); |
0a268dbd TH |
139 | percpu_up_write(&cgroup_threadgroup_rwsem); |
140 | mutex_unlock(&cgroup_mutex); | |
141 | return ret; | |
142 | } | |
143 | ||
144 | /* | |
145 | * Stuff for reading the 'tasks'/'procs' files. | |
146 | * | |
147 | * Reading this file can return large amounts of data if a cgroup has | |
148 | * *lots* of attached tasks. So it may need several calls to read(), | |
149 | * but we cannot guarantee that the information we produce is correct | |
150 | * unless we produce it entirely atomically. | |
151 | * | |
152 | */ | |
153 | ||
154 | /* which pidlist file are we talking about? */ | |
155 | enum cgroup_filetype { | |
156 | CGROUP_FILE_PROCS, | |
157 | CGROUP_FILE_TASKS, | |
158 | }; | |
159 | ||
160 | /* | |
161 | * A pidlist is a list of pids that virtually represents the contents of one | |
162 | * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists, | |
163 | * a pair (one each for procs, tasks) for each pid namespace that's relevant | |
164 | * to the cgroup. | |
165 | */ | |
166 | struct cgroup_pidlist { | |
167 | /* | |
168 | * used to find which pidlist is wanted. doesn't change as long as | |
169 | * this particular list stays in the list. | |
170 | */ | |
171 | struct { enum cgroup_filetype type; struct pid_namespace *ns; } key; | |
172 | /* array of xids */ | |
173 | pid_t *list; | |
174 | /* how many elements the above list has */ | |
175 | int length; | |
176 | /* each of these stored in a list by its cgroup */ | |
177 | struct list_head links; | |
178 | /* pointer to the cgroup we belong to, for list removal purposes */ | |
179 | struct cgroup *owner; | |
180 | /* for delayed destruction */ | |
181 | struct delayed_work destroy_dwork; | |
182 | }; | |
183 | ||
184 | /* | |
185 | * The following two functions "fix" the issue where there are more pids | |
186 | * than kmalloc will give memory for; in such cases, we use vmalloc/vfree. | |
187 | * TODO: replace with a kernel-wide solution to this problem | |
188 | */ | |
189 | #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2)) | |
190 | static void *pidlist_allocate(int count) | |
191 | { | |
192 | if (PIDLIST_TOO_LARGE(count)) | |
193 | return vmalloc(count * sizeof(pid_t)); | |
194 | else | |
195 | return kmalloc(count * sizeof(pid_t), GFP_KERNEL); | |
196 | } | |
197 | ||
198 | static void pidlist_free(void *p) | |
199 | { | |
200 | kvfree(p); | |
201 | } | |
202 | ||
203 | /* | |
204 | * Used to destroy all pidlists lingering waiting for destroy timer. None | |
205 | * should be left afterwards. | |
206 | */ | |
d62beb7f | 207 | void cgroup1_pidlist_destroy_all(struct cgroup *cgrp) |
0a268dbd TH |
208 | { |
209 | struct cgroup_pidlist *l, *tmp_l; | |
210 | ||
211 | mutex_lock(&cgrp->pidlist_mutex); | |
212 | list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links) | |
213 | mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0); | |
214 | mutex_unlock(&cgrp->pidlist_mutex); | |
215 | ||
216 | flush_workqueue(cgroup_pidlist_destroy_wq); | |
217 | BUG_ON(!list_empty(&cgrp->pidlists)); | |
218 | } | |
219 | ||
220 | static void cgroup_pidlist_destroy_work_fn(struct work_struct *work) | |
221 | { | |
222 | struct delayed_work *dwork = to_delayed_work(work); | |
223 | struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist, | |
224 | destroy_dwork); | |
225 | struct cgroup_pidlist *tofree = NULL; | |
226 | ||
227 | mutex_lock(&l->owner->pidlist_mutex); | |
228 | ||
229 | /* | |
230 | * Destroy iff we didn't get queued again. The state won't change | |
231 | * as destroy_dwork can only be queued while locked. | |
232 | */ | |
233 | if (!delayed_work_pending(dwork)) { | |
234 | list_del(&l->links); | |
235 | pidlist_free(l->list); | |
236 | put_pid_ns(l->key.ns); | |
237 | tofree = l; | |
238 | } | |
239 | ||
240 | mutex_unlock(&l->owner->pidlist_mutex); | |
241 | kfree(tofree); | |
242 | } | |
243 | ||
244 | /* | |
245 | * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries | |
246 | * Returns the number of unique elements. | |
247 | */ | |
248 | static int pidlist_uniq(pid_t *list, int length) | |
249 | { | |
250 | int src, dest = 1; | |
251 | ||
252 | /* | |
253 | * we presume the 0th element is unique, so i starts at 1. trivial | |
254 | * edge cases first; no work needs to be done for either | |
255 | */ | |
256 | if (length == 0 || length == 1) | |
257 | return length; | |
258 | /* src and dest walk down the list; dest counts unique elements */ | |
259 | for (src = 1; src < length; src++) { | |
260 | /* find next unique element */ | |
261 | while (list[src] == list[src-1]) { | |
262 | src++; | |
263 | if (src == length) | |
264 | goto after; | |
265 | } | |
266 | /* dest always points to where the next unique element goes */ | |
267 | list[dest] = list[src]; | |
268 | dest++; | |
269 | } | |
270 | after: | |
271 | return dest; | |
272 | } | |
273 | ||
274 | /* | |
275 | * The two pid files - task and cgroup.procs - guaranteed that the result | |
276 | * is sorted, which forced this whole pidlist fiasco. As pid order is | |
277 | * different per namespace, each namespace needs differently sorted list, | |
278 | * making it impossible to use, for example, single rbtree of member tasks | |
279 | * sorted by task pointer. As pidlists can be fairly large, allocating one | |
280 | * per open file is dangerous, so cgroup had to implement shared pool of | |
281 | * pidlists keyed by cgroup and namespace. | |
282 | */ | |
283 | static int cmppid(const void *a, const void *b) | |
284 | { | |
285 | return *(pid_t *)a - *(pid_t *)b; | |
286 | } | |
287 | ||
288 | static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp, | |
289 | enum cgroup_filetype type) | |
290 | { | |
291 | struct cgroup_pidlist *l; | |
292 | /* don't need task_nsproxy() if we're looking at ourself */ | |
293 | struct pid_namespace *ns = task_active_pid_ns(current); | |
294 | ||
295 | lockdep_assert_held(&cgrp->pidlist_mutex); | |
296 | ||
297 | list_for_each_entry(l, &cgrp->pidlists, links) | |
298 | if (l->key.type == type && l->key.ns == ns) | |
299 | return l; | |
300 | return NULL; | |
301 | } | |
302 | ||
303 | /* | |
304 | * find the appropriate pidlist for our purpose (given procs vs tasks) | |
305 | * returns with the lock on that pidlist already held, and takes care | |
306 | * of the use count, or returns NULL with no locks held if we're out of | |
307 | * memory. | |
308 | */ | |
309 | static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp, | |
310 | enum cgroup_filetype type) | |
311 | { | |
312 | struct cgroup_pidlist *l; | |
313 | ||
314 | lockdep_assert_held(&cgrp->pidlist_mutex); | |
315 | ||
316 | l = cgroup_pidlist_find(cgrp, type); | |
317 | if (l) | |
318 | return l; | |
319 | ||
320 | /* entry not found; create a new one */ | |
321 | l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL); | |
322 | if (!l) | |
323 | return l; | |
324 | ||
325 | INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn); | |
326 | l->key.type = type; | |
327 | /* don't need task_nsproxy() if we're looking at ourself */ | |
328 | l->key.ns = get_pid_ns(task_active_pid_ns(current)); | |
329 | l->owner = cgrp; | |
330 | list_add(&l->links, &cgrp->pidlists); | |
331 | return l; | |
332 | } | |
333 | ||
334 | /** | |
335 | * cgroup_task_count - count the number of tasks in a cgroup. | |
336 | * @cgrp: the cgroup in question | |
0a268dbd TH |
337 | */ |
338 | static int cgroup_task_count(const struct cgroup *cgrp) | |
339 | { | |
340 | int count = 0; | |
341 | struct cgrp_cset_link *link; | |
342 | ||
343 | spin_lock_irq(&css_set_lock); | |
344 | list_for_each_entry(link, &cgrp->cset_links, cset_link) | |
73a7242a | 345 | count += link->cset->nr_tasks; |
0a268dbd TH |
346 | spin_unlock_irq(&css_set_lock); |
347 | return count; | |
348 | } | |
349 | ||
350 | /* | |
351 | * Load a cgroup's pidarray with either procs' tgids or tasks' pids | |
352 | */ | |
353 | static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type, | |
354 | struct cgroup_pidlist **lp) | |
355 | { | |
356 | pid_t *array; | |
357 | int length; | |
358 | int pid, n = 0; /* used for populating the array */ | |
359 | struct css_task_iter it; | |
360 | struct task_struct *tsk; | |
361 | struct cgroup_pidlist *l; | |
362 | ||
363 | lockdep_assert_held(&cgrp->pidlist_mutex); | |
364 | ||
365 | /* | |
366 | * If cgroup gets more users after we read count, we won't have | |
367 | * enough space - tough. This race is indistinguishable to the | |
368 | * caller from the case that the additional cgroup users didn't | |
369 | * show up until sometime later on. | |
370 | */ | |
371 | length = cgroup_task_count(cgrp); | |
372 | array = pidlist_allocate(length); | |
373 | if (!array) | |
374 | return -ENOMEM; | |
375 | /* now, populate the array */ | |
376 | css_task_iter_start(&cgrp->self, &it); | |
377 | while ((tsk = css_task_iter_next(&it))) { | |
378 | if (unlikely(n == length)) | |
379 | break; | |
380 | /* get tgid or pid for procs or tasks file respectively */ | |
381 | if (type == CGROUP_FILE_PROCS) | |
382 | pid = task_tgid_vnr(tsk); | |
383 | else | |
384 | pid = task_pid_vnr(tsk); | |
385 | if (pid > 0) /* make sure to only use valid results */ | |
386 | array[n++] = pid; | |
387 | } | |
388 | css_task_iter_end(&it); | |
389 | length = n; | |
390 | /* now sort & (if procs) strip out duplicates */ | |
391 | sort(array, length, sizeof(pid_t), cmppid, NULL); | |
392 | if (type == CGROUP_FILE_PROCS) | |
393 | length = pidlist_uniq(array, length); | |
394 | ||
395 | l = cgroup_pidlist_find_create(cgrp, type); | |
396 | if (!l) { | |
397 | pidlist_free(array); | |
398 | return -ENOMEM; | |
399 | } | |
400 | ||
401 | /* store array, freeing old if necessary */ | |
402 | pidlist_free(l->list); | |
403 | l->list = array; | |
404 | l->length = length; | |
405 | *lp = l; | |
406 | return 0; | |
407 | } | |
408 | ||
409 | /* | |
410 | * seq_file methods for the tasks/procs files. The seq_file position is the | |
411 | * next pid to display; the seq_file iterator is a pointer to the pid | |
412 | * in the cgroup->l->list array. | |
413 | */ | |
414 | ||
415 | static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos) | |
416 | { | |
417 | /* | |
418 | * Initially we receive a position value that corresponds to | |
419 | * one more than the last pid shown (or 0 on the first call or | |
420 | * after a seek to the start). Use a binary-search to find the | |
421 | * next pid to display, if any | |
422 | */ | |
423 | struct kernfs_open_file *of = s->private; | |
424 | struct cgroup *cgrp = seq_css(s)->cgroup; | |
425 | struct cgroup_pidlist *l; | |
426 | enum cgroup_filetype type = seq_cft(s)->private; | |
427 | int index = 0, pid = *pos; | |
428 | int *iter, ret; | |
429 | ||
430 | mutex_lock(&cgrp->pidlist_mutex); | |
431 | ||
432 | /* | |
433 | * !NULL @of->priv indicates that this isn't the first start() | |
434 | * after open. If the matching pidlist is around, we can use that. | |
435 | * Look for it. Note that @of->priv can't be used directly. It | |
436 | * could already have been destroyed. | |
437 | */ | |
438 | if (of->priv) | |
439 | of->priv = cgroup_pidlist_find(cgrp, type); | |
440 | ||
441 | /* | |
442 | * Either this is the first start() after open or the matching | |
443 | * pidlist has been destroyed inbetween. Create a new one. | |
444 | */ | |
445 | if (!of->priv) { | |
446 | ret = pidlist_array_load(cgrp, type, | |
447 | (struct cgroup_pidlist **)&of->priv); | |
448 | if (ret) | |
449 | return ERR_PTR(ret); | |
450 | } | |
451 | l = of->priv; | |
452 | ||
453 | if (pid) { | |
454 | int end = l->length; | |
455 | ||
456 | while (index < end) { | |
457 | int mid = (index + end) / 2; | |
458 | if (l->list[mid] == pid) { | |
459 | index = mid; | |
460 | break; | |
461 | } else if (l->list[mid] <= pid) | |
462 | index = mid + 1; | |
463 | else | |
464 | end = mid; | |
465 | } | |
466 | } | |
467 | /* If we're off the end of the array, we're done */ | |
468 | if (index >= l->length) | |
469 | return NULL; | |
470 | /* Update the abstract position to be the actual pid that we found */ | |
471 | iter = l->list + index; | |
472 | *pos = *iter; | |
473 | return iter; | |
474 | } | |
475 | ||
476 | static void cgroup_pidlist_stop(struct seq_file *s, void *v) | |
477 | { | |
478 | struct kernfs_open_file *of = s->private; | |
479 | struct cgroup_pidlist *l = of->priv; | |
480 | ||
481 | if (l) | |
482 | mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, | |
483 | CGROUP_PIDLIST_DESTROY_DELAY); | |
484 | mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex); | |
485 | } | |
486 | ||
487 | static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos) | |
488 | { | |
489 | struct kernfs_open_file *of = s->private; | |
490 | struct cgroup_pidlist *l = of->priv; | |
491 | pid_t *p = v; | |
492 | pid_t *end = l->list + l->length; | |
493 | /* | |
494 | * Advance to the next pid in the array. If this goes off the | |
495 | * end, we're done | |
496 | */ | |
497 | p++; | |
498 | if (p >= end) { | |
499 | return NULL; | |
500 | } else { | |
501 | *pos = *p; | |
502 | return p; | |
503 | } | |
504 | } | |
505 | ||
506 | static int cgroup_pidlist_show(struct seq_file *s, void *v) | |
507 | { | |
508 | seq_printf(s, "%d\n", *(int *)v); | |
509 | ||
510 | return 0; | |
511 | } | |
512 | ||
513 | static ssize_t cgroup_tasks_write(struct kernfs_open_file *of, | |
514 | char *buf, size_t nbytes, loff_t off) | |
515 | { | |
516 | return __cgroup_procs_write(of, buf, nbytes, off, false); | |
517 | } | |
518 | ||
519 | static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of, | |
520 | char *buf, size_t nbytes, loff_t off) | |
521 | { | |
522 | struct cgroup *cgrp; | |
523 | ||
524 | BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX); | |
525 | ||
526 | cgrp = cgroup_kn_lock_live(of->kn, false); | |
527 | if (!cgrp) | |
528 | return -ENODEV; | |
529 | spin_lock(&release_agent_path_lock); | |
530 | strlcpy(cgrp->root->release_agent_path, strstrip(buf), | |
531 | sizeof(cgrp->root->release_agent_path)); | |
532 | spin_unlock(&release_agent_path_lock); | |
533 | cgroup_kn_unlock(of->kn); | |
534 | return nbytes; | |
535 | } | |
536 | ||
537 | static int cgroup_release_agent_show(struct seq_file *seq, void *v) | |
538 | { | |
539 | struct cgroup *cgrp = seq_css(seq)->cgroup; | |
540 | ||
541 | spin_lock(&release_agent_path_lock); | |
542 | seq_puts(seq, cgrp->root->release_agent_path); | |
543 | spin_unlock(&release_agent_path_lock); | |
544 | seq_putc(seq, '\n'); | |
545 | return 0; | |
546 | } | |
547 | ||
548 | static int cgroup_sane_behavior_show(struct seq_file *seq, void *v) | |
549 | { | |
550 | seq_puts(seq, "0\n"); | |
551 | return 0; | |
552 | } | |
553 | ||
554 | static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css, | |
555 | struct cftype *cft) | |
556 | { | |
557 | return notify_on_release(css->cgroup); | |
558 | } | |
559 | ||
560 | static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css, | |
561 | struct cftype *cft, u64 val) | |
562 | { | |
563 | if (val) | |
564 | set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags); | |
565 | else | |
566 | clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags); | |
567 | return 0; | |
568 | } | |
569 | ||
570 | static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css, | |
571 | struct cftype *cft) | |
572 | { | |
573 | return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); | |
574 | } | |
575 | ||
576 | static int cgroup_clone_children_write(struct cgroup_subsys_state *css, | |
577 | struct cftype *cft, u64 val) | |
578 | { | |
579 | if (val) | |
580 | set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); | |
581 | else | |
582 | clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags); | |
583 | return 0; | |
584 | } | |
585 | ||
586 | /* cgroup core interface files for the legacy hierarchies */ | |
d62beb7f | 587 | struct cftype cgroup1_base_files[] = { |
0a268dbd TH |
588 | { |
589 | .name = "cgroup.procs", | |
590 | .seq_start = cgroup_pidlist_start, | |
591 | .seq_next = cgroup_pidlist_next, | |
592 | .seq_stop = cgroup_pidlist_stop, | |
593 | .seq_show = cgroup_pidlist_show, | |
594 | .private = CGROUP_FILE_PROCS, | |
595 | .write = cgroup_procs_write, | |
596 | }, | |
597 | { | |
598 | .name = "cgroup.clone_children", | |
599 | .read_u64 = cgroup_clone_children_read, | |
600 | .write_u64 = cgroup_clone_children_write, | |
601 | }, | |
602 | { | |
603 | .name = "cgroup.sane_behavior", | |
604 | .flags = CFTYPE_ONLY_ON_ROOT, | |
605 | .seq_show = cgroup_sane_behavior_show, | |
606 | }, | |
607 | { | |
608 | .name = "tasks", | |
609 | .seq_start = cgroup_pidlist_start, | |
610 | .seq_next = cgroup_pidlist_next, | |
611 | .seq_stop = cgroup_pidlist_stop, | |
612 | .seq_show = cgroup_pidlist_show, | |
613 | .private = CGROUP_FILE_TASKS, | |
614 | .write = cgroup_tasks_write, | |
615 | }, | |
616 | { | |
617 | .name = "notify_on_release", | |
618 | .read_u64 = cgroup_read_notify_on_release, | |
619 | .write_u64 = cgroup_write_notify_on_release, | |
620 | }, | |
621 | { | |
622 | .name = "release_agent", | |
623 | .flags = CFTYPE_ONLY_ON_ROOT, | |
624 | .seq_show = cgroup_release_agent_show, | |
625 | .write = cgroup_release_agent_write, | |
626 | .max_write_len = PATH_MAX - 1, | |
627 | }, | |
628 | { } /* terminate */ | |
629 | }; | |
630 | ||
631 | /* Display information about each subsystem and each hierarchy */ | |
632 | static int proc_cgroupstats_show(struct seq_file *m, void *v) | |
633 | { | |
634 | struct cgroup_subsys *ss; | |
635 | int i; | |
636 | ||
637 | seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n"); | |
638 | /* | |
639 | * ideally we don't want subsystems moving around while we do this. | |
640 | * cgroup_mutex is also necessary to guarantee an atomic snapshot of | |
641 | * subsys/hierarchy state. | |
642 | */ | |
643 | mutex_lock(&cgroup_mutex); | |
644 | ||
645 | for_each_subsys(ss, i) | |
646 | seq_printf(m, "%s\t%d\t%d\t%d\n", | |
647 | ss->legacy_name, ss->root->hierarchy_id, | |
648 | atomic_read(&ss->root->nr_cgrps), | |
649 | cgroup_ssid_enabled(i)); | |
650 | ||
651 | mutex_unlock(&cgroup_mutex); | |
652 | return 0; | |
653 | } | |
654 | ||
655 | static int cgroupstats_open(struct inode *inode, struct file *file) | |
656 | { | |
657 | return single_open(file, proc_cgroupstats_show, NULL); | |
658 | } | |
659 | ||
660 | const struct file_operations proc_cgroupstats_operations = { | |
661 | .open = cgroupstats_open, | |
662 | .read = seq_read, | |
663 | .llseek = seq_lseek, | |
664 | .release = single_release, | |
665 | }; | |
666 | ||
667 | /** | |
668 | * cgroupstats_build - build and fill cgroupstats | |
669 | * @stats: cgroupstats to fill information into | |
670 | * @dentry: A dentry entry belonging to the cgroup for which stats have | |
671 | * been requested. | |
672 | * | |
673 | * Build and fill cgroupstats so that taskstats can export it to user | |
674 | * space. | |
675 | */ | |
676 | int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry) | |
677 | { | |
678 | struct kernfs_node *kn = kernfs_node_from_dentry(dentry); | |
679 | struct cgroup *cgrp; | |
680 | struct css_task_iter it; | |
681 | struct task_struct *tsk; | |
682 | ||
683 | /* it should be kernfs_node belonging to cgroupfs and is a directory */ | |
684 | if (dentry->d_sb->s_type != &cgroup_fs_type || !kn || | |
685 | kernfs_type(kn) != KERNFS_DIR) | |
686 | return -EINVAL; | |
687 | ||
688 | mutex_lock(&cgroup_mutex); | |
689 | ||
690 | /* | |
691 | * We aren't being called from kernfs and there's no guarantee on | |
692 | * @kn->priv's validity. For this and css_tryget_online_from_dir(), | |
693 | * @kn->priv is RCU safe. Let's do the RCU dancing. | |
694 | */ | |
695 | rcu_read_lock(); | |
e0aed7c7 | 696 | cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); |
0a268dbd TH |
697 | if (!cgrp || cgroup_is_dead(cgrp)) { |
698 | rcu_read_unlock(); | |
699 | mutex_unlock(&cgroup_mutex); | |
700 | return -ENOENT; | |
701 | } | |
702 | rcu_read_unlock(); | |
703 | ||
704 | css_task_iter_start(&cgrp->self, &it); | |
705 | while ((tsk = css_task_iter_next(&it))) { | |
706 | switch (tsk->state) { | |
707 | case TASK_RUNNING: | |
708 | stats->nr_running++; | |
709 | break; | |
710 | case TASK_INTERRUPTIBLE: | |
711 | stats->nr_sleeping++; | |
712 | break; | |
713 | case TASK_UNINTERRUPTIBLE: | |
714 | stats->nr_uninterruptible++; | |
715 | break; | |
716 | case TASK_STOPPED: | |
717 | stats->nr_stopped++; | |
718 | break; | |
719 | default: | |
720 | if (delayacct_is_task_waiting_on_io(tsk)) | |
721 | stats->nr_io_wait++; | |
722 | break; | |
723 | } | |
724 | } | |
725 | css_task_iter_end(&it); | |
726 | ||
727 | mutex_unlock(&cgroup_mutex); | |
728 | return 0; | |
729 | } | |
730 | ||
d62beb7f | 731 | void cgroup1_check_for_release(struct cgroup *cgrp) |
0a268dbd TH |
732 | { |
733 | if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) && | |
734 | !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp)) | |
735 | schedule_work(&cgrp->release_agent_work); | |
736 | } | |
737 | ||
738 | /* | |
739 | * Notify userspace when a cgroup is released, by running the | |
740 | * configured release agent with the name of the cgroup (path | |
741 | * relative to the root of cgroup file system) as the argument. | |
742 | * | |
743 | * Most likely, this user command will try to rmdir this cgroup. | |
744 | * | |
745 | * This races with the possibility that some other task will be | |
746 | * attached to this cgroup before it is removed, or that some other | |
747 | * user task will 'mkdir' a child cgroup of this cgroup. That's ok. | |
748 | * The presumed 'rmdir' will fail quietly if this cgroup is no longer | |
749 | * unused, and this cgroup will be reprieved from its death sentence, | |
750 | * to continue to serve a useful existence. Next time it's released, | |
751 | * we will get notified again, if it still has 'notify_on_release' set. | |
752 | * | |
753 | * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which | |
754 | * means only wait until the task is successfully execve()'d. The | |
755 | * separate release agent task is forked by call_usermodehelper(), | |
756 | * then control in this thread returns here, without waiting for the | |
757 | * release agent task. We don't bother to wait because the caller of | |
758 | * this routine has no use for the exit status of the release agent | |
759 | * task, so no sense holding our caller up for that. | |
760 | */ | |
d62beb7f | 761 | void cgroup1_release_agent(struct work_struct *work) |
0a268dbd TH |
762 | { |
763 | struct cgroup *cgrp = | |
764 | container_of(work, struct cgroup, release_agent_work); | |
765 | char *pathbuf = NULL, *agentbuf = NULL; | |
766 | char *argv[3], *envp[3]; | |
767 | int ret; | |
768 | ||
769 | mutex_lock(&cgroup_mutex); | |
770 | ||
771 | pathbuf = kmalloc(PATH_MAX, GFP_KERNEL); | |
772 | agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL); | |
773 | if (!pathbuf || !agentbuf) | |
774 | goto out; | |
775 | ||
776 | spin_lock_irq(&css_set_lock); | |
777 | ret = cgroup_path_ns_locked(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns); | |
778 | spin_unlock_irq(&css_set_lock); | |
779 | if (ret < 0 || ret >= PATH_MAX) | |
780 | goto out; | |
781 | ||
782 | argv[0] = agentbuf; | |
783 | argv[1] = pathbuf; | |
784 | argv[2] = NULL; | |
785 | ||
786 | /* minimal command environment */ | |
787 | envp[0] = "HOME=/"; | |
788 | envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; | |
789 | envp[2] = NULL; | |
790 | ||
791 | mutex_unlock(&cgroup_mutex); | |
792 | call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC); | |
793 | goto out_free; | |
794 | out: | |
795 | mutex_unlock(&cgroup_mutex); | |
796 | out_free: | |
797 | kfree(agentbuf); | |
798 | kfree(pathbuf); | |
799 | } | |
800 | ||
801 | /* | |
802 | * cgroup_rename - Only allow simple rename of directories in place. | |
803 | */ | |
1592c9b2 TH |
804 | static int cgroup1_rename(struct kernfs_node *kn, struct kernfs_node *new_parent, |
805 | const char *new_name_str) | |
0a268dbd TH |
806 | { |
807 | struct cgroup *cgrp = kn->priv; | |
808 | int ret; | |
809 | ||
810 | if (kernfs_type(kn) != KERNFS_DIR) | |
811 | return -ENOTDIR; | |
812 | if (kn->parent != new_parent) | |
813 | return -EIO; | |
814 | ||
0a268dbd TH |
815 | /* |
816 | * We're gonna grab cgroup_mutex which nests outside kernfs | |
817 | * active_ref. kernfs_rename() doesn't require active_ref | |
818 | * protection. Break them before grabbing cgroup_mutex. | |
819 | */ | |
820 | kernfs_break_active_protection(new_parent); | |
821 | kernfs_break_active_protection(kn); | |
822 | ||
823 | mutex_lock(&cgroup_mutex); | |
824 | ||
825 | ret = kernfs_rename(kn, new_parent, new_name_str); | |
826 | if (!ret) | |
827 | trace_cgroup_rename(cgrp); | |
828 | ||
829 | mutex_unlock(&cgroup_mutex); | |
830 | ||
831 | kernfs_unbreak_active_protection(kn); | |
832 | kernfs_unbreak_active_protection(new_parent); | |
833 | return ret; | |
834 | } | |
835 | ||
1592c9b2 TH |
836 | static int cgroup1_show_options(struct seq_file *seq, struct kernfs_root *kf_root) |
837 | { | |
838 | struct cgroup_root *root = cgroup_root_from_kf(kf_root); | |
839 | struct cgroup_subsys *ss; | |
840 | int ssid; | |
841 | ||
842 | for_each_subsys(ss, ssid) | |
843 | if (root->subsys_mask & (1 << ssid)) | |
844 | seq_show_option(seq, ss->legacy_name, NULL); | |
845 | if (root->flags & CGRP_ROOT_NOPREFIX) | |
846 | seq_puts(seq, ",noprefix"); | |
847 | if (root->flags & CGRP_ROOT_XATTR) | |
848 | seq_puts(seq, ",xattr"); | |
849 | ||
850 | spin_lock(&release_agent_path_lock); | |
851 | if (strlen(root->release_agent_path)) | |
852 | seq_show_option(seq, "release_agent", | |
853 | root->release_agent_path); | |
854 | spin_unlock(&release_agent_path_lock); | |
855 | ||
856 | if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags)) | |
857 | seq_puts(seq, ",clone_children"); | |
858 | if (strlen(root->name)) | |
859 | seq_show_option(seq, "name", root->name); | |
860 | return 0; | |
861 | } | |
862 | ||
863 | static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts) | |
864 | { | |
865 | char *token, *o = data; | |
866 | bool all_ss = false, one_ss = false; | |
867 | u16 mask = U16_MAX; | |
868 | struct cgroup_subsys *ss; | |
869 | int nr_opts = 0; | |
870 | int i; | |
871 | ||
872 | #ifdef CONFIG_CPUSETS | |
873 | mask = ~((u16)1 << cpuset_cgrp_id); | |
874 | #endif | |
875 | ||
876 | memset(opts, 0, sizeof(*opts)); | |
877 | ||
878 | while ((token = strsep(&o, ",")) != NULL) { | |
879 | nr_opts++; | |
880 | ||
881 | if (!*token) | |
882 | return -EINVAL; | |
883 | if (!strcmp(token, "none")) { | |
884 | /* Explicitly have no subsystems */ | |
885 | opts->none = true; | |
886 | continue; | |
887 | } | |
888 | if (!strcmp(token, "all")) { | |
889 | /* Mutually exclusive option 'all' + subsystem name */ | |
890 | if (one_ss) | |
891 | return -EINVAL; | |
892 | all_ss = true; | |
893 | continue; | |
894 | } | |
895 | if (!strcmp(token, "noprefix")) { | |
896 | opts->flags |= CGRP_ROOT_NOPREFIX; | |
897 | continue; | |
898 | } | |
899 | if (!strcmp(token, "clone_children")) { | |
900 | opts->cpuset_clone_children = true; | |
901 | continue; | |
902 | } | |
903 | if (!strcmp(token, "xattr")) { | |
904 | opts->flags |= CGRP_ROOT_XATTR; | |
905 | continue; | |
906 | } | |
907 | if (!strncmp(token, "release_agent=", 14)) { | |
908 | /* Specifying two release agents is forbidden */ | |
909 | if (opts->release_agent) | |
910 | return -EINVAL; | |
911 | opts->release_agent = | |
912 | kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL); | |
913 | if (!opts->release_agent) | |
914 | return -ENOMEM; | |
915 | continue; | |
916 | } | |
917 | if (!strncmp(token, "name=", 5)) { | |
918 | const char *name = token + 5; | |
919 | /* Can't specify an empty name */ | |
920 | if (!strlen(name)) | |
921 | return -EINVAL; | |
922 | /* Must match [\w.-]+ */ | |
923 | for (i = 0; i < strlen(name); i++) { | |
924 | char c = name[i]; | |
925 | if (isalnum(c)) | |
926 | continue; | |
927 | if ((c == '.') || (c == '-') || (c == '_')) | |
928 | continue; | |
929 | return -EINVAL; | |
930 | } | |
931 | /* Specifying two names is forbidden */ | |
932 | if (opts->name) | |
933 | return -EINVAL; | |
934 | opts->name = kstrndup(name, | |
935 | MAX_CGROUP_ROOT_NAMELEN - 1, | |
936 | GFP_KERNEL); | |
937 | if (!opts->name) | |
938 | return -ENOMEM; | |
939 | ||
940 | continue; | |
941 | } | |
942 | ||
943 | for_each_subsys(ss, i) { | |
944 | if (strcmp(token, ss->legacy_name)) | |
945 | continue; | |
946 | if (!cgroup_ssid_enabled(i)) | |
947 | continue; | |
d62beb7f | 948 | if (cgroup1_ssid_disabled(i)) |
1592c9b2 TH |
949 | continue; |
950 | ||
951 | /* Mutually exclusive option 'all' + subsystem name */ | |
952 | if (all_ss) | |
953 | return -EINVAL; | |
954 | opts->subsys_mask |= (1 << i); | |
955 | one_ss = true; | |
956 | ||
957 | break; | |
958 | } | |
959 | if (i == CGROUP_SUBSYS_COUNT) | |
960 | return -ENOENT; | |
961 | } | |
962 | ||
963 | /* | |
964 | * If the 'all' option was specified select all the subsystems, | |
965 | * otherwise if 'none', 'name=' and a subsystem name options were | |
966 | * not specified, let's default to 'all' | |
967 | */ | |
968 | if (all_ss || (!one_ss && !opts->none && !opts->name)) | |
969 | for_each_subsys(ss, i) | |
d62beb7f | 970 | if (cgroup_ssid_enabled(i) && !cgroup1_ssid_disabled(i)) |
1592c9b2 TH |
971 | opts->subsys_mask |= (1 << i); |
972 | ||
973 | /* | |
974 | * We either have to specify by name or by subsystems. (So all | |
975 | * empty hierarchies must have a name). | |
976 | */ | |
977 | if (!opts->subsys_mask && !opts->name) | |
978 | return -EINVAL; | |
979 | ||
980 | /* | |
981 | * Option noprefix was introduced just for backward compatibility | |
982 | * with the old cpuset, so we allow noprefix only if mounting just | |
983 | * the cpuset subsystem. | |
984 | */ | |
985 | if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask)) | |
986 | return -EINVAL; | |
987 | ||
988 | /* Can't specify "none" and some subsystems */ | |
989 | if (opts->subsys_mask && opts->none) | |
990 | return -EINVAL; | |
991 | ||
992 | return 0; | |
993 | } | |
994 | ||
995 | static int cgroup1_remount(struct kernfs_root *kf_root, int *flags, char *data) | |
996 | { | |
997 | int ret = 0; | |
998 | struct cgroup_root *root = cgroup_root_from_kf(kf_root); | |
999 | struct cgroup_sb_opts opts; | |
1000 | u16 added_mask, removed_mask; | |
1001 | ||
1002 | cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); | |
1003 | ||
1004 | /* See what subsystems are wanted */ | |
1005 | ret = parse_cgroupfs_options(data, &opts); | |
1006 | if (ret) | |
1007 | goto out_unlock; | |
1008 | ||
1009 | if (opts.subsys_mask != root->subsys_mask || opts.release_agent) | |
1010 | pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n", | |
1011 | task_tgid_nr(current), current->comm); | |
1012 | ||
1013 | added_mask = opts.subsys_mask & ~root->subsys_mask; | |
1014 | removed_mask = root->subsys_mask & ~opts.subsys_mask; | |
1015 | ||
1016 | /* Don't allow flags or name to change at remount */ | |
1017 | if ((opts.flags ^ root->flags) || | |
1018 | (opts.name && strcmp(opts.name, root->name))) { | |
1019 | pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n", | |
1020 | opts.flags, opts.name ?: "", root->flags, root->name); | |
1021 | ret = -EINVAL; | |
1022 | goto out_unlock; | |
1023 | } | |
1024 | ||
1025 | /* remounting is not allowed for populated hierarchies */ | |
1026 | if (!list_empty(&root->cgrp.self.children)) { | |
1027 | ret = -EBUSY; | |
1028 | goto out_unlock; | |
1029 | } | |
1030 | ||
1031 | ret = rebind_subsystems(root, added_mask); | |
1032 | if (ret) | |
1033 | goto out_unlock; | |
1034 | ||
1035 | WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask)); | |
1036 | ||
1037 | if (opts.release_agent) { | |
1038 | spin_lock(&release_agent_path_lock); | |
1039 | strcpy(root->release_agent_path, opts.release_agent); | |
1040 | spin_unlock(&release_agent_path_lock); | |
1041 | } | |
1042 | ||
1043 | trace_cgroup_remount(root); | |
1044 | ||
1045 | out_unlock: | |
1046 | kfree(opts.release_agent); | |
1047 | kfree(opts.name); | |
1048 | mutex_unlock(&cgroup_mutex); | |
1049 | return ret; | |
1050 | } | |
1051 | ||
1052 | struct kernfs_syscall_ops cgroup1_kf_syscall_ops = { | |
1053 | .rename = cgroup1_rename, | |
1054 | .show_options = cgroup1_show_options, | |
1055 | .remount_fs = cgroup1_remount, | |
1056 | .mkdir = cgroup_mkdir, | |
1057 | .rmdir = cgroup_rmdir, | |
1058 | .show_path = cgroup_show_path, | |
1059 | }; | |
1060 | ||
1061 | struct dentry *cgroup1_mount(struct file_system_type *fs_type, int flags, | |
1062 | void *data, unsigned long magic, | |
1063 | struct cgroup_namespace *ns) | |
1064 | { | |
1065 | struct super_block *pinned_sb = NULL; | |
1066 | struct cgroup_sb_opts opts; | |
1067 | struct cgroup_root *root; | |
1068 | struct cgroup_subsys *ss; | |
1069 | struct dentry *dentry; | |
1070 | int i, ret; | |
9732adc5 | 1071 | bool new_root = false; |
1592c9b2 TH |
1072 | |
1073 | cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); | |
1074 | ||
1075 | /* First find the desired set of subsystems */ | |
1076 | ret = parse_cgroupfs_options(data, &opts); | |
1077 | if (ret) | |
1078 | goto out_unlock; | |
1079 | ||
1080 | /* | |
1081 | * Destruction of cgroup root is asynchronous, so subsystems may | |
1082 | * still be dying after the previous unmount. Let's drain the | |
1083 | * dying subsystems. We just need to ensure that the ones | |
1084 | * unmounted previously finish dying and don't care about new ones | |
1085 | * starting. Testing ref liveliness is good enough. | |
1086 | */ | |
1087 | for_each_subsys(ss, i) { | |
1088 | if (!(opts.subsys_mask & (1 << i)) || | |
1089 | ss->root == &cgrp_dfl_root) | |
1090 | continue; | |
1091 | ||
1092 | if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) { | |
1093 | mutex_unlock(&cgroup_mutex); | |
1094 | msleep(10); | |
1095 | ret = restart_syscall(); | |
1096 | goto out_free; | |
1097 | } | |
1098 | cgroup_put(&ss->root->cgrp); | |
1099 | } | |
1100 | ||
1101 | for_each_root(root) { | |
1102 | bool name_match = false; | |
1103 | ||
1104 | if (root == &cgrp_dfl_root) | |
1105 | continue; | |
1106 | ||
1107 | /* | |
1108 | * If we asked for a name then it must match. Also, if | |
1109 | * name matches but sybsys_mask doesn't, we should fail. | |
1110 | * Remember whether name matched. | |
1111 | */ | |
1112 | if (opts.name) { | |
1113 | if (strcmp(opts.name, root->name)) | |
1114 | continue; | |
1115 | name_match = true; | |
1116 | } | |
1117 | ||
1118 | /* | |
1119 | * If we asked for subsystems (or explicitly for no | |
1120 | * subsystems) then they must match. | |
1121 | */ | |
1122 | if ((opts.subsys_mask || opts.none) && | |
1123 | (opts.subsys_mask != root->subsys_mask)) { | |
1124 | if (!name_match) | |
1125 | continue; | |
1126 | ret = -EBUSY; | |
1127 | goto out_unlock; | |
1128 | } | |
1129 | ||
1130 | if (root->flags ^ opts.flags) | |
1131 | pr_warn("new mount options do not match the existing superblock, will be ignored\n"); | |
1132 | ||
1133 | /* | |
1134 | * We want to reuse @root whose lifetime is governed by its | |
1135 | * ->cgrp. Let's check whether @root is alive and keep it | |
1136 | * that way. As cgroup_kill_sb() can happen anytime, we | |
1137 | * want to block it by pinning the sb so that @root doesn't | |
1138 | * get killed before mount is complete. | |
1139 | * | |
1140 | * With the sb pinned, tryget_live can reliably indicate | |
1141 | * whether @root can be reused. If it's being killed, | |
1142 | * drain it. We can use wait_queue for the wait but this | |
1143 | * path is super cold. Let's just sleep a bit and retry. | |
1144 | */ | |
1145 | pinned_sb = kernfs_pin_sb(root->kf_root, NULL); | |
330c4186 | 1146 | if (IS_ERR(pinned_sb) || |
1592c9b2 TH |
1147 | !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) { |
1148 | mutex_unlock(&cgroup_mutex); | |
1149 | if (!IS_ERR_OR_NULL(pinned_sb)) | |
1150 | deactivate_super(pinned_sb); | |
1151 | msleep(10); | |
1152 | ret = restart_syscall(); | |
1153 | goto out_free; | |
1154 | } | |
1155 | ||
1156 | ret = 0; | |
1157 | goto out_unlock; | |
1158 | } | |
1159 | ||
1160 | /* | |
1161 | * No such thing, create a new one. name= matching without subsys | |
1162 | * specification is allowed for already existing hierarchies but we | |
1163 | * can't create new one without subsys specification. | |
1164 | */ | |
1165 | if (!opts.subsys_mask && !opts.none) { | |
1166 | ret = -EINVAL; | |
1167 | goto out_unlock; | |
1168 | } | |
1169 | ||
1170 | /* Hierarchies may only be created in the initial cgroup namespace. */ | |
1171 | if (ns != &init_cgroup_ns) { | |
1172 | ret = -EPERM; | |
1173 | goto out_unlock; | |
1174 | } | |
1175 | ||
1176 | root = kzalloc(sizeof(*root), GFP_KERNEL); | |
1177 | if (!root) { | |
1178 | ret = -ENOMEM; | |
1179 | goto out_unlock; | |
1180 | } | |
9732adc5 | 1181 | new_root = true; |
1592c9b2 TH |
1182 | |
1183 | init_cgroup_root(root, &opts); | |
1184 | ||
9732adc5 | 1185 | ret = cgroup_setup_root(root, opts.subsys_mask, PERCPU_REF_INIT_DEAD); |
1592c9b2 TH |
1186 | if (ret) |
1187 | cgroup_free_root(root); | |
1188 | ||
1189 | out_unlock: | |
1190 | mutex_unlock(&cgroup_mutex); | |
1191 | out_free: | |
1192 | kfree(opts.release_agent); | |
1193 | kfree(opts.name); | |
1194 | ||
1195 | if (ret) | |
1196 | return ERR_PTR(ret); | |
1197 | ||
1198 | dentry = cgroup_do_mount(&cgroup_fs_type, flags, root, | |
1199 | CGROUP_SUPER_MAGIC, ns); | |
1200 | ||
9732adc5 ZL |
1201 | /* |
1202 | * There's a race window after we release cgroup_mutex and before | |
1203 | * allocating a superblock. Make sure a concurrent process won't | |
1204 | * be able to re-use the root during this window by delaying the | |
1205 | * initialization of root refcnt. | |
1206 | */ | |
1207 | if (new_root) { | |
1208 | mutex_lock(&cgroup_mutex); | |
1209 | percpu_ref_reinit(&root->cgrp.self.refcnt); | |
1210 | mutex_unlock(&cgroup_mutex); | |
1211 | } | |
1212 | ||
1592c9b2 TH |
1213 | /* |
1214 | * If @pinned_sb, we're reusing an existing root and holding an | |
1215 | * extra ref on its sb. Mount is complete. Put the extra ref. | |
1216 | */ | |
1217 | if (pinned_sb) | |
1218 | deactivate_super(pinned_sb); | |
1219 | ||
1220 | return dentry; | |
1221 | } | |
1222 | ||
0a268dbd TH |
1223 | static int __init cgroup1_wq_init(void) |
1224 | { | |
1225 | /* | |
1226 | * Used to destroy pidlists and separate to serve as flush domain. | |
1227 | * Cap @max_active to 1 too. | |
1228 | */ | |
1229 | cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy", | |
1230 | 0, 1); | |
1231 | BUG_ON(!cgroup_pidlist_destroy_wq); | |
1232 | return 0; | |
1233 | } | |
1234 | core_initcall(cgroup1_wq_init); | |
1235 | ||
1236 | static int __init cgroup_no_v1(char *str) | |
1237 | { | |
1238 | struct cgroup_subsys *ss; | |
1239 | char *token; | |
1240 | int i; | |
1241 | ||
1242 | while ((token = strsep(&str, ",")) != NULL) { | |
1243 | if (!*token) | |
1244 | continue; | |
1245 | ||
1246 | if (!strcmp(token, "all")) { | |
1247 | cgroup_no_v1_mask = U16_MAX; | |
1248 | break; | |
1249 | } | |
1250 | ||
1251 | for_each_subsys(ss, i) { | |
1252 | if (strcmp(token, ss->name) && | |
1253 | strcmp(token, ss->legacy_name)) | |
1254 | continue; | |
1255 | ||
1256 | cgroup_no_v1_mask |= 1 << i; | |
1257 | } | |
1258 | } | |
1259 | return 1; | |
1260 | } | |
1261 | __setup("cgroup_no_v1=", cgroup_no_v1); | |
1262 | ||
1263 | ||
1264 | #ifdef CONFIG_CGROUP_DEBUG | |
1265 | static struct cgroup_subsys_state * | |
1266 | debug_css_alloc(struct cgroup_subsys_state *parent_css) | |
1267 | { | |
1268 | struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL); | |
1269 | ||
1270 | if (!css) | |
1271 | return ERR_PTR(-ENOMEM); | |
1272 | ||
1273 | return css; | |
1274 | } | |
1275 | ||
1276 | static void debug_css_free(struct cgroup_subsys_state *css) | |
1277 | { | |
1278 | kfree(css); | |
1279 | } | |
1280 | ||
1281 | static u64 debug_taskcount_read(struct cgroup_subsys_state *css, | |
1282 | struct cftype *cft) | |
1283 | { | |
1284 | return cgroup_task_count(css->cgroup); | |
1285 | } | |
1286 | ||
1287 | static u64 current_css_set_read(struct cgroup_subsys_state *css, | |
1288 | struct cftype *cft) | |
1289 | { | |
1290 | return (u64)(unsigned long)current->cgroups; | |
1291 | } | |
1292 | ||
1293 | static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css, | |
1294 | struct cftype *cft) | |
1295 | { | |
1296 | u64 count; | |
1297 | ||
1298 | rcu_read_lock(); | |
4b9502e6 | 1299 | count = refcount_read(&task_css_set(current)->refcount); |
0a268dbd TH |
1300 | rcu_read_unlock(); |
1301 | return count; | |
1302 | } | |
1303 | ||
1304 | static int current_css_set_cg_links_read(struct seq_file *seq, void *v) | |
1305 | { | |
1306 | struct cgrp_cset_link *link; | |
1307 | struct css_set *cset; | |
1308 | char *name_buf; | |
1309 | ||
1310 | name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL); | |
1311 | if (!name_buf) | |
1312 | return -ENOMEM; | |
1313 | ||
1314 | spin_lock_irq(&css_set_lock); | |
1315 | rcu_read_lock(); | |
1316 | cset = rcu_dereference(current->cgroups); | |
1317 | list_for_each_entry(link, &cset->cgrp_links, cgrp_link) { | |
1318 | struct cgroup *c = link->cgrp; | |
1319 | ||
1320 | cgroup_name(c, name_buf, NAME_MAX + 1); | |
1321 | seq_printf(seq, "Root %d group %s\n", | |
1322 | c->root->hierarchy_id, name_buf); | |
1323 | } | |
1324 | rcu_read_unlock(); | |
1325 | spin_unlock_irq(&css_set_lock); | |
1326 | kfree(name_buf); | |
1327 | return 0; | |
1328 | } | |
1329 | ||
1330 | #define MAX_TASKS_SHOWN_PER_CSS 25 | |
1331 | static int cgroup_css_links_read(struct seq_file *seq, void *v) | |
1332 | { | |
1333 | struct cgroup_subsys_state *css = seq_css(seq); | |
1334 | struct cgrp_cset_link *link; | |
1335 | ||
1336 | spin_lock_irq(&css_set_lock); | |
1337 | list_for_each_entry(link, &css->cgroup->cset_links, cset_link) { | |
1338 | struct css_set *cset = link->cset; | |
1339 | struct task_struct *task; | |
1340 | int count = 0; | |
1341 | ||
b6a6759d | 1342 | seq_printf(seq, "css_set %pK\n", cset); |
0a268dbd TH |
1343 | |
1344 | list_for_each_entry(task, &cset->tasks, cg_list) { | |
1345 | if (count++ > MAX_TASKS_SHOWN_PER_CSS) | |
1346 | goto overflow; | |
1347 | seq_printf(seq, " task %d\n", task_pid_vnr(task)); | |
1348 | } | |
1349 | ||
1350 | list_for_each_entry(task, &cset->mg_tasks, cg_list) { | |
1351 | if (count++ > MAX_TASKS_SHOWN_PER_CSS) | |
1352 | goto overflow; | |
1353 | seq_printf(seq, " task %d\n", task_pid_vnr(task)); | |
1354 | } | |
1355 | continue; | |
1356 | overflow: | |
1357 | seq_puts(seq, " ...\n"); | |
1358 | } | |
1359 | spin_unlock_irq(&css_set_lock); | |
1360 | return 0; | |
1361 | } | |
1362 | ||
1363 | static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft) | |
1364 | { | |
1365 | return (!cgroup_is_populated(css->cgroup) && | |
1366 | !css_has_online_children(&css->cgroup->self)); | |
1367 | } | |
1368 | ||
1369 | static struct cftype debug_files[] = { | |
1370 | { | |
1371 | .name = "taskcount", | |
1372 | .read_u64 = debug_taskcount_read, | |
1373 | }, | |
1374 | ||
1375 | { | |
1376 | .name = "current_css_set", | |
1377 | .read_u64 = current_css_set_read, | |
1378 | }, | |
1379 | ||
1380 | { | |
1381 | .name = "current_css_set_refcount", | |
1382 | .read_u64 = current_css_set_refcount_read, | |
1383 | }, | |
1384 | ||
1385 | { | |
1386 | .name = "current_css_set_cg_links", | |
1387 | .seq_show = current_css_set_cg_links_read, | |
1388 | }, | |
1389 | ||
1390 | { | |
1391 | .name = "cgroup_css_links", | |
1392 | .seq_show = cgroup_css_links_read, | |
1393 | }, | |
1394 | ||
1395 | { | |
1396 | .name = "releasable", | |
1397 | .read_u64 = releasable_read, | |
1398 | }, | |
1399 | ||
1400 | { } /* terminate */ | |
1401 | }; | |
1402 | ||
1403 | struct cgroup_subsys debug_cgrp_subsys = { | |
1404 | .css_alloc = debug_css_alloc, | |
1405 | .css_free = debug_css_free, | |
1406 | .legacy_cftypes = debug_files, | |
1407 | }; | |
1408 | #endif /* CONFIG_CGROUP_DEBUG */ |