1// SPDX-License-Identifier: GPL-2.0-only
2#include "cgroup-internal.h"
3
4#include <linux/ctype.h>
5#include <linux/kmod.h>
6#include <linux/sort.h>
7#include <linux/delay.h>
8#include <linux/mm.h>
9#include <linux/sched/signal.h>
10#include <linux/sched/task.h>
11#include <linux/magic.h>
12#include <linux/slab.h>
13#include <linux/vmalloc.h>
14#include <linux/delayacct.h>
15#include <linux/pid_namespace.h>
16#include <linux/cgroupstats.h>
17#include <linux/fs_parser.h>
18
19#include <trace/events/cgroup.h>
20
21/*
22 * pidlists linger the following amount before being destroyed. The goal
23 * is avoiding frequent destruction in the middle of consecutive read calls
24 * Expiring in the middle is a performance problem not a correctness one.
25 * 1 sec should be enough.
26 */
27#define CGROUP_PIDLIST_DESTROY_DELAY HZ
28
29/* Controllers blocked by the commandline in v1 */
30static u16 cgroup_no_v1_mask;
31
32/* disable named v1 mounts */
33static bool cgroup_no_v1_named;
34
35/*
36 * pidlist destructions need to be flushed on cgroup destruction. Use a
37 * separate workqueue as flush domain.
38 */
39static struct workqueue_struct *cgroup_pidlist_destroy_wq;
40
41/* protects cgroup_subsys->release_agent_path */
42static DEFINE_SPINLOCK(release_agent_path_lock);
43
44bool cgroup1_ssid_disabled(int ssid)
45{
46 return cgroup_no_v1_mask & (1 << ssid);
47}
48
49/**
50 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
51 * @from: attach to all cgroups of a given task
52 * @tsk: the task to be attached
53 *
54 * Return: %0 on success or a negative errno code on failure
55 */
56int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
57{
58 struct cgroup_root *root;
59 int retval = 0;
60
61 cgroup_lock();
62 cgroup_attach_lock(lock_threadgroup: true);
63 for_each_root(root) {
64 struct cgroup *from_cgrp;
65
66 spin_lock_irq(lock: &css_set_lock);
67 from_cgrp = task_cgroup_from_root(task: from, root);
68 spin_unlock_irq(lock: &css_set_lock);
69
70 retval = cgroup_attach_task(dst_cgrp: from_cgrp, leader: tsk, threadgroup: false);
71 if (retval)
72 break;
73 }
74 cgroup_attach_unlock(lock_threadgroup: true);
75 cgroup_unlock();
76
77 return retval;
78}
79EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
80
81/**
82 * cgroup_transfer_tasks - move tasks from one cgroup to another
83 * @to: cgroup to which the tasks will be moved
84 * @from: cgroup in which the tasks currently reside
85 *
86 * Locking rules between cgroup_post_fork() and the migration path
87 * guarantee that, if a task is forking while being migrated, the new child
88 * is guaranteed to be either visible in the source cgroup after the
89 * parent's migration is complete or put into the target cgroup. No task
90 * can slip out of migration through forking.
91 *
92 * Return: %0 on success or a negative errno code on failure
93 */
94int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
95{
96 DEFINE_CGROUP_MGCTX(mgctx);
97 struct cgrp_cset_link *link;
98 struct css_task_iter it;
99 struct task_struct *task;
100 int ret;
101
102 if (cgroup_on_dfl(cgrp: to))
103 return -EINVAL;
104
105 ret = cgroup_migrate_vet_dst(dst_cgrp: to);
106 if (ret)
107 return ret;
108
109 cgroup_lock();
110
111 cgroup_attach_lock(lock_threadgroup: true);
112
113 /* all tasks in @from are being moved, all csets are source */
114 spin_lock_irq(lock: &css_set_lock);
115 list_for_each_entry(link, &from->cset_links, cset_link)
116 cgroup_migrate_add_src(src_cset: link->cset, dst_cgrp: to, mgctx: &mgctx);
117 spin_unlock_irq(lock: &css_set_lock);
118
119 ret = cgroup_migrate_prepare_dst(mgctx: &mgctx);
120 if (ret)
121 goto out_err;
122
123 /*
124 * Migrate tasks one-by-one until @from is empty. This fails iff
125 * ->can_attach() fails.
126 */
127 do {
128 css_task_iter_start(css: &from->self, flags: 0, it: &it);
129
130 do {
131 task = css_task_iter_next(it: &it);
132 } while (task && (task->flags & PF_EXITING));
133
134 if (task)
135 get_task_struct(t: task);
136 css_task_iter_end(it: &it);
137
138 if (task) {
139 ret = cgroup_migrate(leader: task, threadgroup: false, mgctx: &mgctx);
140 if (!ret)
141 TRACE_CGROUP_PATH(transfer_tasks, to, task, false);
142 put_task_struct(t: task);
143 }
144 } while (task && !ret);
145out_err:
146 cgroup_migrate_finish(mgctx: &mgctx);
147 cgroup_attach_unlock(lock_threadgroup: true);
148 cgroup_unlock();
149 return ret;
150}
151
152/*
153 * Stuff for reading the 'tasks'/'procs' files.
154 *
155 * Reading this file can return large amounts of data if a cgroup has
156 * *lots* of attached tasks. So it may need several calls to read(),
157 * but we cannot guarantee that the information we produce is correct
158 * unless we produce it entirely atomically.
159 *
160 */
161
162/* which pidlist file are we talking about? */
163enum cgroup_filetype {
164 CGROUP_FILE_PROCS,
165 CGROUP_FILE_TASKS,
166};
167
168/*
169 * A pidlist is a list of pids that virtually represents the contents of one
170 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
171 * a pair (one each for procs, tasks) for each pid namespace that's relevant
172 * to the cgroup.
173 */
174struct cgroup_pidlist {
175 /*
176 * used to find which pidlist is wanted. doesn't change as long as
177 * this particular list stays in the list.
178 */
179 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
180 /* array of xids */
181 pid_t *list;
182 /* how many elements the above list has */
183 int length;
184 /* each of these stored in a list by its cgroup */
185 struct list_head links;
186 /* pointer to the cgroup we belong to, for list removal purposes */
187 struct cgroup *owner;
188 /* for delayed destruction */
189 struct delayed_work destroy_dwork;
190};
191
192/*
193 * Used to destroy all pidlists lingering waiting for destroy timer. None
194 * should be left afterwards.
195 */
196void cgroup1_pidlist_destroy_all(struct cgroup *cgrp)
197{
198 struct cgroup_pidlist *l, *tmp_l;
199
200 mutex_lock(&cgrp->pidlist_mutex);
201 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
202 mod_delayed_work(wq: cgroup_pidlist_destroy_wq, dwork: &l->destroy_dwork, delay: 0);
203 mutex_unlock(lock: &cgrp->pidlist_mutex);
204
205 flush_workqueue(cgroup_pidlist_destroy_wq);
206 BUG_ON(!list_empty(&cgrp->pidlists));
207}
208
209static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
210{
211 struct delayed_work *dwork = to_delayed_work(work);
212 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
213 destroy_dwork);
214 struct cgroup_pidlist *tofree = NULL;
215
216 mutex_lock(&l->owner->pidlist_mutex);
217
218 /*
219 * Destroy iff we didn't get queued again. The state won't change
220 * as destroy_dwork can only be queued while locked.
221 */
222 if (!delayed_work_pending(dwork)) {
223 list_del(entry: &l->links);
224 kvfree(addr: l->list);
225 put_pid_ns(ns: l->key.ns);
226 tofree = l;
227 }
228
229 mutex_unlock(lock: &l->owner->pidlist_mutex);
230 kfree(objp: tofree);
231}
232
233/*
234 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
235 * Returns the number of unique elements.
236 */
237static int pidlist_uniq(pid_t *list, int length)
238{
239 int src, dest = 1;
240
241 /*
242 * we presume the 0th element is unique, so i starts at 1. trivial
243 * edge cases first; no work needs to be done for either
244 */
245 if (length == 0 || length == 1)
246 return length;
247 /* src and dest walk down the list; dest counts unique elements */
248 for (src = 1; src < length; src++) {
249 /* find next unique element */
250 while (list[src] == list[src-1]) {
251 src++;
252 if (src == length)
253 goto after;
254 }
255 /* dest always points to where the next unique element goes */
256 list[dest] = list[src];
257 dest++;
258 }
259after:
260 return dest;
261}
262
263/*
264 * The two pid files - task and cgroup.procs - guaranteed that the result
265 * is sorted, which forced this whole pidlist fiasco. As pid order is
266 * different per namespace, each namespace needs differently sorted list,
267 * making it impossible to use, for example, single rbtree of member tasks
268 * sorted by task pointer. As pidlists can be fairly large, allocating one
269 * per open file is dangerous, so cgroup had to implement shared pool of
270 * pidlists keyed by cgroup and namespace.
271 */
272static int cmppid(const void *a, const void *b)
273{
274 return *(pid_t *)a - *(pid_t *)b;
275}
276
277static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
278 enum cgroup_filetype type)
279{
280 struct cgroup_pidlist *l;
281 /* don't need task_nsproxy() if we're looking at ourself */
282 struct pid_namespace *ns = task_active_pid_ns(current);
283
284 lockdep_assert_held(&cgrp->pidlist_mutex);
285
286 list_for_each_entry(l, &cgrp->pidlists, links)
287 if (l->key.type == type && l->key.ns == ns)
288 return l;
289 return NULL;
290}
291
292/*
293 * find the appropriate pidlist for our purpose (given procs vs tasks)
294 * returns with the lock on that pidlist already held, and takes care
295 * of the use count, or returns NULL with no locks held if we're out of
296 * memory.
297 */
298static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
299 enum cgroup_filetype type)
300{
301 struct cgroup_pidlist *l;
302
303 lockdep_assert_held(&cgrp->pidlist_mutex);
304
305 l = cgroup_pidlist_find(cgrp, type);
306 if (l)
307 return l;
308
309 /* entry not found; create a new one */
310 l = kzalloc(size: sizeof(struct cgroup_pidlist), GFP_KERNEL);
311 if (!l)
312 return l;
313
314 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
315 l->key.type = type;
316 /* don't need task_nsproxy() if we're looking at ourself */
317 l->key.ns = get_pid_ns(ns: task_active_pid_ns(current));
318 l->owner = cgrp;
319 list_add(new: &l->links, head: &cgrp->pidlists);
320 return l;
321}
322
323/*
324 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
325 */
326static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
327 struct cgroup_pidlist **lp)
328{
329 pid_t *array;
330 int length;
331 int pid, n = 0; /* used for populating the array */
332 struct css_task_iter it;
333 struct task_struct *tsk;
334 struct cgroup_pidlist *l;
335
336 lockdep_assert_held(&cgrp->pidlist_mutex);
337
338 /*
339 * If cgroup gets more users after we read count, we won't have
340 * enough space - tough. This race is indistinguishable to the
341 * caller from the case that the additional cgroup users didn't
342 * show up until sometime later on.
343 */
344 length = cgroup_task_count(cgrp);
345 array = kvmalloc_array(n: length, size: sizeof(pid_t), GFP_KERNEL);
346 if (!array)
347 return -ENOMEM;
348 /* now, populate the array */
349 css_task_iter_start(css: &cgrp->self, flags: 0, it: &it);
350 while ((tsk = css_task_iter_next(it: &it))) {
351 if (unlikely(n == length))
352 break;
353 /* get tgid or pid for procs or tasks file respectively */
354 if (type == CGROUP_FILE_PROCS)
355 pid = task_tgid_vnr(tsk);
356 else
357 pid = task_pid_vnr(tsk);
358 if (pid > 0) /* make sure to only use valid results */
359 array[n++] = pid;
360 }
361 css_task_iter_end(it: &it);
362 length = n;
363 /* now sort & strip out duplicates (tgids or recycled thread PIDs) */
364 sort(base: array, num: length, size: sizeof(pid_t), cmp_func: cmppid, NULL);
365 length = pidlist_uniq(list: array, length);
366
367 l = cgroup_pidlist_find_create(cgrp, type);
368 if (!l) {
369 kvfree(addr: array);
370 return -ENOMEM;
371 }
372
373 /* store array, freeing old if necessary */
374 kvfree(addr: l->list);
375 l->list = array;
376 l->length = length;
377 *lp = l;
378 return 0;
379}
380
381/*
382 * seq_file methods for the tasks/procs files. The seq_file position is the
383 * next pid to display; the seq_file iterator is a pointer to the pid
384 * in the cgroup->l->list array.
385 */
386
387static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
388{
389 /*
390 * Initially we receive a position value that corresponds to
391 * one more than the last pid shown (or 0 on the first call or
392 * after a seek to the start). Use a binary-search to find the
393 * next pid to display, if any
394 */
395 struct kernfs_open_file *of = s->private;
396 struct cgroup_file_ctx *ctx = of->priv;
397 struct cgroup *cgrp = seq_css(seq: s)->cgroup;
398 struct cgroup_pidlist *l;
399 enum cgroup_filetype type = seq_cft(seq: s)->private;
400 int index = 0, pid = *pos;
401 int *iter, ret;
402
403 mutex_lock(&cgrp->pidlist_mutex);
404
405 /*
406 * !NULL @ctx->procs1.pidlist indicates that this isn't the first
407 * start() after open. If the matching pidlist is around, we can use
408 * that. Look for it. Note that @ctx->procs1.pidlist can't be used
409 * directly. It could already have been destroyed.
410 */
411 if (ctx->procs1.pidlist)
412 ctx->procs1.pidlist = cgroup_pidlist_find(cgrp, type);
413
414 /*
415 * Either this is the first start() after open or the matching
416 * pidlist has been destroyed inbetween. Create a new one.
417 */
418 if (!ctx->procs1.pidlist) {
419 ret = pidlist_array_load(cgrp, type, lp: &ctx->procs1.pidlist);
420 if (ret)
421 return ERR_PTR(error: ret);
422 }
423 l = ctx->procs1.pidlist;
424
425 if (pid) {
426 int end = l->length;
427
428 while (index < end) {
429 int mid = (index + end) / 2;
430 if (l->list[mid] == pid) {
431 index = mid;
432 break;
433 } else if (l->list[mid] < pid)
434 index = mid + 1;
435 else
436 end = mid;
437 }
438 }
439 /* If we're off the end of the array, we're done */
440 if (index >= l->length)
441 return NULL;
442 /* Update the abstract position to be the actual pid that we found */
443 iter = l->list + index;
444 *pos = *iter;
445 return iter;
446}
447
448static void cgroup_pidlist_stop(struct seq_file *s, void *v)
449{
450 struct kernfs_open_file *of = s->private;
451 struct cgroup_file_ctx *ctx = of->priv;
452 struct cgroup_pidlist *l = ctx->procs1.pidlist;
453
454 if (l)
455 mod_delayed_work(wq: cgroup_pidlist_destroy_wq, dwork: &l->destroy_dwork,
456 CGROUP_PIDLIST_DESTROY_DELAY);
457 mutex_unlock(lock: &seq_css(seq: s)->cgroup->pidlist_mutex);
458}
459
460static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
461{
462 struct kernfs_open_file *of = s->private;
463 struct cgroup_file_ctx *ctx = of->priv;
464 struct cgroup_pidlist *l = ctx->procs1.pidlist;
465 pid_t *p = v;
466 pid_t *end = l->list + l->length;
467 /*
468 * Advance to the next pid in the array. If this goes off the
469 * end, we're done
470 */
471 p++;
472 if (p >= end) {
473 (*pos)++;
474 return NULL;
475 } else {
476 *pos = *p;
477 return p;
478 }
479}
480
481static int cgroup_pidlist_show(struct seq_file *s, void *v)
482{
483 seq_printf(m: s, fmt: "%d\n", *(int *)v);
484
485 return 0;
486}
487
488static ssize_t __cgroup1_procs_write(struct kernfs_open_file *of,
489 char *buf, size_t nbytes, loff_t off,
490 bool threadgroup)
491{
492 struct cgroup *cgrp;
493 struct task_struct *task;
494 const struct cred *cred, *tcred;
495 ssize_t ret;
496 bool locked;
497
498 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
499 if (!cgrp)
500 return -ENODEV;
501
502 task = cgroup_procs_write_start(buf, threadgroup, locked: &locked);
503 ret = PTR_ERR_OR_ZERO(ptr: task);
504 if (ret)
505 goto out_unlock;
506
507 /*
508 * Even if we're attaching all tasks in the thread group, we only need
509 * to check permissions on one of them. Check permissions using the
510 * credentials from file open to protect against inherited fd attacks.
511 */
512 cred = of->file->f_cred;
513 tcred = get_task_cred(task);
514 if (!uid_eq(left: cred->euid, GLOBAL_ROOT_UID) &&
515 !uid_eq(cred->euid, tcred->uid) &&
516 !uid_eq(cred->euid, tcred->suid))
517 ret = -EACCES;
518 put_cred(tcred);
519 if (ret)
520 goto out_finish;
521
522 ret = cgroup_attach_task(cgrp, task, threadgroup);
523
524out_finish:
525 cgroup_procs_write_finish(task, locked);
526out_unlock:
527 cgroup_kn_unlock(of->kn);
528
529 return ret ?: nbytes;
530}
531
532static ssize_t cgroup1_procs_write(struct kernfs_open_file *of,
533 char *buf, size_t nbytes, loff_t off)
534{
535 return __cgroup1_procs_write(of, buf, nbytes, off, threadgroup: true);
536}
537
538static ssize_t cgroup1_tasks_write(struct kernfs_open_file *of,
539 char *buf, size_t nbytes, loff_t off)
540{
541 return __cgroup1_procs_write(of, buf, nbytes, off, threadgroup: false);
542}
543
544static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
545 char *buf, size_t nbytes, loff_t off)
546{
547 struct cgroup *cgrp;
548 struct cgroup_file_ctx *ctx;
549
550 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
551
552 /*
553 * Release agent gets called with all capabilities,
554 * require capabilities to set release agent.
555 */
556 ctx = of->priv;
557 if ((ctx->ns->user_ns != &init_user_ns) ||
558 !file_ns_capable(file: of->file, ns: &init_user_ns, CAP_SYS_ADMIN))
559 return -EPERM;
560
561 cgrp = cgroup_kn_lock_live(kn: of->kn, drain_offline: false);
562 if (!cgrp)
563 return -ENODEV;
564 spin_lock(lock: &release_agent_path_lock);
565 strscpy(p: cgrp->root->release_agent_path, q: strstrip(str: buf),
566 size: sizeof(cgrp->root->release_agent_path));
567 spin_unlock(lock: &release_agent_path_lock);
568 cgroup_kn_unlock(kn: of->kn);
569 return nbytes;
570}
571
572static int cgroup_release_agent_show(struct seq_file *seq, void *v)
573{
574 struct cgroup *cgrp = seq_css(seq)->cgroup;
575
576 spin_lock(lock: &release_agent_path_lock);
577 seq_puts(m: seq, s: cgrp->root->release_agent_path);
578 spin_unlock(lock: &release_agent_path_lock);
579 seq_putc(m: seq, c: '\n');
580 return 0;
581}
582
583static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
584{
585 seq_puts(m: seq, s: "0\n");
586 return 0;
587}
588
589static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
590 struct cftype *cft)
591{
592 return notify_on_release(cgrp: css->cgroup);
593}
594
595static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
596 struct cftype *cft, u64 val)
597{
598 if (val)
599 set_bit(nr: CGRP_NOTIFY_ON_RELEASE, addr: &css->cgroup->flags);
600 else
601 clear_bit(nr: CGRP_NOTIFY_ON_RELEASE, addr: &css->cgroup->flags);
602 return 0;
603}
604
605static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
606 struct cftype *cft)
607{
608 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
609}
610
611static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
612 struct cftype *cft, u64 val)
613{
614 if (val)
615 set_bit(nr: CGRP_CPUSET_CLONE_CHILDREN, addr: &css->cgroup->flags);
616 else
617 clear_bit(nr: CGRP_CPUSET_CLONE_CHILDREN, addr: &css->cgroup->flags);
618 return 0;
619}
620
621/* cgroup core interface files for the legacy hierarchies */
622struct cftype cgroup1_base_files[] = {
623 {
624 .name = "cgroup.procs",
625 .seq_start = cgroup_pidlist_start,
626 .seq_next = cgroup_pidlist_next,
627 .seq_stop = cgroup_pidlist_stop,
628 .seq_show = cgroup_pidlist_show,
629 .private = CGROUP_FILE_PROCS,
630 .write = cgroup1_procs_write,
631 },
632 {
633 .name = "cgroup.clone_children",
634 .read_u64 = cgroup_clone_children_read,
635 .write_u64 = cgroup_clone_children_write,
636 },
637 {
638 .name = "cgroup.sane_behavior",
639 .flags = CFTYPE_ONLY_ON_ROOT,
640 .seq_show = cgroup_sane_behavior_show,
641 },
642 {
643 .name = "tasks",
644 .seq_start = cgroup_pidlist_start,
645 .seq_next = cgroup_pidlist_next,
646 .seq_stop = cgroup_pidlist_stop,
647 .seq_show = cgroup_pidlist_show,
648 .private = CGROUP_FILE_TASKS,
649 .write = cgroup1_tasks_write,
650 },
651 {
652 .name = "notify_on_release",
653 .read_u64 = cgroup_read_notify_on_release,
654 .write_u64 = cgroup_write_notify_on_release,
655 },
656 {
657 .name = "release_agent",
658 .flags = CFTYPE_ONLY_ON_ROOT,
659 .seq_show = cgroup_release_agent_show,
660 .write = cgroup_release_agent_write,
661 .max_write_len = PATH_MAX - 1,
662 },
663 { } /* terminate */
664};
665
666/* Display information about each subsystem and each hierarchy */
667int proc_cgroupstats_show(struct seq_file *m, void *v)
668{
669 struct cgroup_subsys *ss;
670 int i;
671
672 seq_puts(m, s: "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
673 /*
674 * Grab the subsystems state racily. No need to add avenue to
675 * cgroup_mutex contention.
676 */
677
678 for_each_subsys(ss, i)
679 seq_printf(m, fmt: "%s\t%d\t%d\t%d\n",
680 ss->legacy_name, ss->root->hierarchy_id,
681 atomic_read(v: &ss->root->nr_cgrps),
682 cgroup_ssid_enabled(ssid: i));
683
684 return 0;
685}
686
687/**
688 * cgroupstats_build - build and fill cgroupstats
689 * @stats: cgroupstats to fill information into
690 * @dentry: A dentry entry belonging to the cgroup for which stats have
691 * been requested.
692 *
693 * Build and fill cgroupstats so that taskstats can export it to user
694 * space.
695 *
696 * Return: %0 on success or a negative errno code on failure
697 */
698int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
699{
700 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
701 struct cgroup *cgrp;
702 struct css_task_iter it;
703 struct task_struct *tsk;
704
705 /* it should be kernfs_node belonging to cgroupfs and is a directory */
706 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
707 kernfs_type(kn) != KERNFS_DIR)
708 return -EINVAL;
709
710 /*
711 * We aren't being called from kernfs and there's no guarantee on
712 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
713 * @kn->priv is RCU safe. Let's do the RCU dancing.
714 */
715 rcu_read_lock();
716 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
717 if (!cgrp || !cgroup_tryget(cgrp)) {
718 rcu_read_unlock();
719 return -ENOENT;
720 }
721 rcu_read_unlock();
722
723 css_task_iter_start(css: &cgrp->self, flags: 0, it: &it);
724 while ((tsk = css_task_iter_next(it: &it))) {
725 switch (READ_ONCE(tsk->__state)) {
726 case TASK_RUNNING:
727 stats->nr_running++;
728 break;
729 case TASK_INTERRUPTIBLE:
730 stats->nr_sleeping++;
731 break;
732 case TASK_UNINTERRUPTIBLE:
733 stats->nr_uninterruptible++;
734 break;
735 case TASK_STOPPED:
736 stats->nr_stopped++;
737 break;
738 default:
739 if (tsk->in_iowait)
740 stats->nr_io_wait++;
741 break;
742 }
743 }
744 css_task_iter_end(it: &it);
745
746 cgroup_put(cgrp);
747 return 0;
748}
749
750void cgroup1_check_for_release(struct cgroup *cgrp)
751{
752 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
753 !css_has_online_children(css: &cgrp->self) && !cgroup_is_dead(cgrp))
754 schedule_work(work: &cgrp->release_agent_work);
755}
756
757/*
758 * Notify userspace when a cgroup is released, by running the
759 * configured release agent with the name of the cgroup (path
760 * relative to the root of cgroup file system) as the argument.
761 *
762 * Most likely, this user command will try to rmdir this cgroup.
763 *
764 * This races with the possibility that some other task will be
765 * attached to this cgroup before it is removed, or that some other
766 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
767 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
768 * unused, and this cgroup will be reprieved from its death sentence,
769 * to continue to serve a useful existence. Next time it's released,
770 * we will get notified again, if it still has 'notify_on_release' set.
771 *
772 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
773 * means only wait until the task is successfully execve()'d. The
774 * separate release agent task is forked by call_usermodehelper(),
775 * then control in this thread returns here, without waiting for the
776 * release agent task. We don't bother to wait because the caller of
777 * this routine has no use for the exit status of the release agent
778 * task, so no sense holding our caller up for that.
779 */
780void cgroup1_release_agent(struct work_struct *work)
781{
782 struct cgroup *cgrp =
783 container_of(work, struct cgroup, release_agent_work);
784 char *pathbuf, *agentbuf;
785 char *argv[3], *envp[3];
786 int ret;
787
788 /* snoop agent path and exit early if empty */
789 if (!cgrp->root->release_agent_path[0])
790 return;
791
792 /* prepare argument buffers */
793 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
794 agentbuf = kmalloc(PATH_MAX, GFP_KERNEL);
795 if (!pathbuf || !agentbuf)
796 goto out_free;
797
798 spin_lock(lock: &release_agent_path_lock);
799 strscpy(p: agentbuf, q: cgrp->root->release_agent_path, PATH_MAX);
800 spin_unlock(lock: &release_agent_path_lock);
801 if (!agentbuf[0])
802 goto out_free;
803
804 ret = cgroup_path_ns(cgrp, buf: pathbuf, PATH_MAX, ns: &init_cgroup_ns);
805 if (ret < 0 || ret >= PATH_MAX)
806 goto out_free;
807
808 argv[0] = agentbuf;
809 argv[1] = pathbuf;
810 argv[2] = NULL;
811
812 /* minimal command environment */
813 envp[0] = "HOME=/";
814 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
815 envp[2] = NULL;
816
817 call_usermodehelper(path: argv[0], argv, envp, UMH_WAIT_EXEC);
818out_free:
819 kfree(objp: agentbuf);
820 kfree(objp: pathbuf);
821}
822
823/*
824 * cgroup_rename - Only allow simple rename of directories in place.
825 */
826static int cgroup1_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
827 const char *new_name_str)
828{
829 struct cgroup *cgrp = kn->priv;
830 int ret;
831
832 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
833 if (strchr(new_name_str, '\n'))
834 return -EINVAL;
835
836 if (kernfs_type(kn) != KERNFS_DIR)
837 return -ENOTDIR;
838 if (kn->parent != new_parent)
839 return -EIO;
840
841 /*
842 * We're gonna grab cgroup_mutex which nests outside kernfs
843 * active_ref. kernfs_rename() doesn't require active_ref
844 * protection. Break them before grabbing cgroup_mutex.
845 */
846 kernfs_break_active_protection(kn: new_parent);
847 kernfs_break_active_protection(kn);
848
849 cgroup_lock();
850
851 ret = kernfs_rename(kn, new_parent, new_name: new_name_str);
852 if (!ret)
853 TRACE_CGROUP_PATH(rename, cgrp);
854
855 cgroup_unlock();
856
857 kernfs_unbreak_active_protection(kn);
858 kernfs_unbreak_active_protection(kn: new_parent);
859 return ret;
860}
861
862static int cgroup1_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
863{
864 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
865 struct cgroup_subsys *ss;
866 int ssid;
867
868 for_each_subsys(ss, ssid)
869 if (root->subsys_mask & (1 << ssid))
870 seq_show_option(m: seq, name: ss->legacy_name, NULL);
871 if (root->flags & CGRP_ROOT_NOPREFIX)
872 seq_puts(m: seq, s: ",noprefix");
873 if (root->flags & CGRP_ROOT_XATTR)
874 seq_puts(m: seq, s: ",xattr");
875 if (root->flags & CGRP_ROOT_CPUSET_V2_MODE)
876 seq_puts(m: seq, s: ",cpuset_v2_mode");
877 if (root->flags & CGRP_ROOT_FAVOR_DYNMODS)
878 seq_puts(m: seq, s: ",favordynmods");
879
880 spin_lock(lock: &release_agent_path_lock);
881 if (strlen(root->release_agent_path))
882 seq_show_option(m: seq, name: "release_agent",
883 value: root->release_agent_path);
884 spin_unlock(lock: &release_agent_path_lock);
885
886 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
887 seq_puts(m: seq, s: ",clone_children");
888 if (strlen(root->name))
889 seq_show_option(m: seq, name: "name", value: root->name);
890 return 0;
891}
892
893enum cgroup1_param {
894 Opt_all,
895 Opt_clone_children,
896 Opt_cpuset_v2_mode,
897 Opt_name,
898 Opt_none,
899 Opt_noprefix,
900 Opt_release_agent,
901 Opt_xattr,
902 Opt_favordynmods,
903 Opt_nofavordynmods,
904};
905
906const struct fs_parameter_spec cgroup1_fs_parameters[] = {
907 fsparam_flag ("all", Opt_all),
908 fsparam_flag ("clone_children", Opt_clone_children),
909 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode),
910 fsparam_string("name", Opt_name),
911 fsparam_flag ("none", Opt_none),
912 fsparam_flag ("noprefix", Opt_noprefix),
913 fsparam_string("release_agent", Opt_release_agent),
914 fsparam_flag ("xattr", Opt_xattr),
915 fsparam_flag ("favordynmods", Opt_favordynmods),
916 fsparam_flag ("nofavordynmods", Opt_nofavordynmods),
917 {}
918};
919
920int cgroup1_parse_param(struct fs_context *fc, struct fs_parameter *param)
921{
922 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
923 struct cgroup_subsys *ss;
924 struct fs_parse_result result;
925 int opt, i;
926
927 opt = fs_parse(fc, desc: cgroup1_fs_parameters, param, result: &result);
928 if (opt == -ENOPARAM) {
929 int ret;
930
931 ret = vfs_parse_fs_param_source(fc, param);
932 if (ret != -ENOPARAM)
933 return ret;
934 for_each_subsys(ss, i) {
935 if (strcmp(param->key, ss->legacy_name))
936 continue;
937 if (!cgroup_ssid_enabled(ssid: i) || cgroup1_ssid_disabled(ssid: i))
938 return invalfc(fc, "Disabled controller '%s'",
939 param->key);
940 ctx->subsys_mask |= (1 << i);
941 return 0;
942 }
943 return invalfc(fc, "Unknown subsys name '%s'", param->key);
944 }
945 if (opt < 0)
946 return opt;
947
948 switch (opt) {
949 case Opt_none:
950 /* Explicitly have no subsystems */
951 ctx->none = true;
952 break;
953 case Opt_all:
954 ctx->all_ss = true;
955 break;
956 case Opt_noprefix:
957 ctx->flags |= CGRP_ROOT_NOPREFIX;
958 break;
959 case Opt_clone_children:
960 ctx->cpuset_clone_children = true;
961 break;
962 case Opt_cpuset_v2_mode:
963 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE;
964 break;
965 case Opt_xattr:
966 ctx->flags |= CGRP_ROOT_XATTR;
967 break;
968 case Opt_favordynmods:
969 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
970 break;
971 case Opt_nofavordynmods:
972 ctx->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
973 break;
974 case Opt_release_agent:
975 /* Specifying two release agents is forbidden */
976 if (ctx->release_agent)
977 return invalfc(fc, "release_agent respecified");
978 /*
979 * Release agent gets called with all capabilities,
980 * require capabilities to set release agent.
981 */
982 if ((fc->user_ns != &init_user_ns) || !capable(CAP_SYS_ADMIN))
983 return invalfc(fc, "Setting release_agent not allowed");
984 ctx->release_agent = param->string;
985 param->string = NULL;
986 break;
987 case Opt_name:
988 /* blocked by boot param? */
989 if (cgroup_no_v1_named)
990 return -ENOENT;
991 /* Can't specify an empty name */
992 if (!param->size)
993 return invalfc(fc, "Empty name");
994 if (param->size > MAX_CGROUP_ROOT_NAMELEN - 1)
995 return invalfc(fc, "Name too long");
996 /* Must match [\w.-]+ */
997 for (i = 0; i < param->size; i++) {
998 char c = param->string[i];
999 if (isalnum(c))
1000 continue;
1001 if ((c == '.') || (c == '-') || (c == '_'))
1002 continue;
1003 return invalfc(fc, "Invalid name");
1004 }
1005 /* Specifying two names is forbidden */
1006 if (ctx->name)
1007 return invalfc(fc, "name respecified");
1008 ctx->name = param->string;
1009 param->string = NULL;
1010 break;
1011 }
1012 return 0;
1013}
1014
1015static int check_cgroupfs_options(struct fs_context *fc)
1016{
1017 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1018 u16 mask = U16_MAX;
1019 u16 enabled = 0;
1020 struct cgroup_subsys *ss;
1021 int i;
1022
1023#ifdef CONFIG_CPUSETS
1024 mask = ~((u16)1 << cpuset_cgrp_id);
1025#endif
1026 for_each_subsys(ss, i)
1027 if (cgroup_ssid_enabled(ssid: i) && !cgroup1_ssid_disabled(ssid: i))
1028 enabled |= 1 << i;
1029
1030 ctx->subsys_mask &= enabled;
1031
1032 /*
1033 * In absence of 'none', 'name=' and subsystem name options,
1034 * let's default to 'all'.
1035 */
1036 if (!ctx->subsys_mask && !ctx->none && !ctx->name)
1037 ctx->all_ss = true;
1038
1039 if (ctx->all_ss) {
1040 /* Mutually exclusive option 'all' + subsystem name */
1041 if (ctx->subsys_mask)
1042 return invalfc(fc, "subsys name conflicts with all");
1043 /* 'all' => select all the subsystems */
1044 ctx->subsys_mask = enabled;
1045 }
1046
1047 /*
1048 * We either have to specify by name or by subsystems. (So all
1049 * empty hierarchies must have a name).
1050 */
1051 if (!ctx->subsys_mask && !ctx->name)
1052 return invalfc(fc, "Need name or subsystem set");
1053
1054 /*
1055 * Option noprefix was introduced just for backward compatibility
1056 * with the old cpuset, so we allow noprefix only if mounting just
1057 * the cpuset subsystem.
1058 */
1059 if ((ctx->flags & CGRP_ROOT_NOPREFIX) && (ctx->subsys_mask & mask))
1060 return invalfc(fc, "noprefix used incorrectly");
1061
1062 /* Can't specify "none" and some subsystems */
1063 if (ctx->subsys_mask && ctx->none)
1064 return invalfc(fc, "none used incorrectly");
1065
1066 return 0;
1067}
1068
1069int cgroup1_reconfigure(struct fs_context *fc)
1070{
1071 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1072 struct kernfs_root *kf_root = kernfs_root_from_sb(sb: fc->root->d_sb);
1073 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1074 int ret = 0;
1075 u16 added_mask, removed_mask;
1076
1077 cgroup_lock_and_drain_offline(cgrp: &cgrp_dfl_root.cgrp);
1078
1079 /* See what subsystems are wanted */
1080 ret = check_cgroupfs_options(fc);
1081 if (ret)
1082 goto out_unlock;
1083
1084 if (ctx->subsys_mask != root->subsys_mask || ctx->release_agent)
1085 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1086 task_tgid_nr(current), current->comm);
1087
1088 added_mask = ctx->subsys_mask & ~root->subsys_mask;
1089 removed_mask = root->subsys_mask & ~ctx->subsys_mask;
1090
1091 /* Don't allow flags or name to change at remount */
1092 if ((ctx->flags ^ root->flags) ||
1093 (ctx->name && strcmp(ctx->name, root->name))) {
1094 errorfc(fc, "option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"",
1095 ctx->flags, ctx->name ?: "", root->flags, root->name);
1096 ret = -EINVAL;
1097 goto out_unlock;
1098 }
1099
1100 /* remounting is not allowed for populated hierarchies */
1101 if (!list_empty(head: &root->cgrp.self.children)) {
1102 ret = -EBUSY;
1103 goto out_unlock;
1104 }
1105
1106 ret = rebind_subsystems(dst_root: root, ss_mask: added_mask);
1107 if (ret)
1108 goto out_unlock;
1109
1110 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1111
1112 if (ctx->release_agent) {
1113 spin_lock(lock: &release_agent_path_lock);
1114 strcpy(p: root->release_agent_path, q: ctx->release_agent);
1115 spin_unlock(lock: &release_agent_path_lock);
1116 }
1117
1118 trace_cgroup_remount(root);
1119
1120 out_unlock:
1121 cgroup_unlock();
1122 return ret;
1123}
1124
1125struct kernfs_syscall_ops cgroup1_kf_syscall_ops = {
1126 .rename = cgroup1_rename,
1127 .show_options = cgroup1_show_options,
1128 .mkdir = cgroup_mkdir,
1129 .rmdir = cgroup_rmdir,
1130 .show_path = cgroup_show_path,
1131};
1132
1133/*
1134 * The guts of cgroup1 mount - find or create cgroup_root to use.
1135 * Called with cgroup_mutex held; returns 0 on success, -E... on
1136 * error and positive - in case when the candidate is busy dying.
1137 * On success it stashes a reference to cgroup_root into given
1138 * cgroup_fs_context; that reference is *NOT* counting towards the
1139 * cgroup_root refcount.
1140 */
1141static int cgroup1_root_to_use(struct fs_context *fc)
1142{
1143 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1144 struct cgroup_root *root;
1145 struct cgroup_subsys *ss;
1146 int i, ret;
1147
1148 /* First find the desired set of subsystems */
1149 ret = check_cgroupfs_options(fc);
1150 if (ret)
1151 return ret;
1152
1153 /*
1154 * Destruction of cgroup root is asynchronous, so subsystems may
1155 * still be dying after the previous unmount. Let's drain the
1156 * dying subsystems. We just need to ensure that the ones
1157 * unmounted previously finish dying and don't care about new ones
1158 * starting. Testing ref liveliness is good enough.
1159 */
1160 for_each_subsys(ss, i) {
1161 if (!(ctx->subsys_mask & (1 << i)) ||
1162 ss->root == &cgrp_dfl_root)
1163 continue;
1164
1165 if (!percpu_ref_tryget_live(ref: &ss->root->cgrp.self.refcnt))
1166 return 1; /* restart */
1167 cgroup_put(cgrp: &ss->root->cgrp);
1168 }
1169
1170 for_each_root(root) {
1171 bool name_match = false;
1172
1173 if (root == &cgrp_dfl_root)
1174 continue;
1175
1176 /*
1177 * If we asked for a name then it must match. Also, if
1178 * name matches but sybsys_mask doesn't, we should fail.
1179 * Remember whether name matched.
1180 */
1181 if (ctx->name) {
1182 if (strcmp(ctx->name, root->name))
1183 continue;
1184 name_match = true;
1185 }
1186
1187 /*
1188 * If we asked for subsystems (or explicitly for no
1189 * subsystems) then they must match.
1190 */
1191 if ((ctx->subsys_mask || ctx->none) &&
1192 (ctx->subsys_mask != root->subsys_mask)) {
1193 if (!name_match)
1194 continue;
1195 return -EBUSY;
1196 }
1197
1198 if (root->flags ^ ctx->flags)
1199 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1200
1201 ctx->root = root;
1202 return 0;
1203 }
1204
1205 /*
1206 * No such thing, create a new one. name= matching without subsys
1207 * specification is allowed for already existing hierarchies but we
1208 * can't create new one without subsys specification.
1209 */
1210 if (!ctx->subsys_mask && !ctx->none)
1211 return invalfc(fc, "No subsys list or none specified");
1212
1213 /* Hierarchies may only be created in the initial cgroup namespace. */
1214 if (ctx->ns != &init_cgroup_ns)
1215 return -EPERM;
1216
1217 root = kzalloc(size: sizeof(*root), GFP_KERNEL);
1218 if (!root)
1219 return -ENOMEM;
1220
1221 ctx->root = root;
1222 init_cgroup_root(ctx);
1223
1224 ret = cgroup_setup_root(root, ss_mask: ctx->subsys_mask);
1225 if (!ret)
1226 cgroup_favor_dynmods(root, favor: ctx->flags & CGRP_ROOT_FAVOR_DYNMODS);
1227 else
1228 cgroup_free_root(root);
1229
1230 return ret;
1231}
1232
1233int cgroup1_get_tree(struct fs_context *fc)
1234{
1235 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1236 int ret;
1237
1238 /* Check if the caller has permission to mount. */
1239 if (!ns_capable(ns: ctx->ns->user_ns, CAP_SYS_ADMIN))
1240 return -EPERM;
1241
1242 cgroup_lock_and_drain_offline(cgrp: &cgrp_dfl_root.cgrp);
1243
1244 ret = cgroup1_root_to_use(fc);
1245 if (!ret && !percpu_ref_tryget_live(ref: &ctx->root->cgrp.self.refcnt))
1246 ret = 1; /* restart */
1247
1248 cgroup_unlock();
1249
1250 if (!ret)
1251 ret = cgroup_do_get_tree(fc);
1252
1253 if (!ret && percpu_ref_is_dying(ref: &ctx->root->cgrp.self.refcnt)) {
1254 fc_drop_locked(fc);
1255 ret = 1;
1256 }
1257
1258 if (unlikely(ret > 0)) {
1259 msleep(msecs: 10);
1260 return restart_syscall();
1261 }
1262 return ret;
1263}
1264
1265static int __init cgroup1_wq_init(void)
1266{
1267 /*
1268 * Used to destroy pidlists and separate to serve as flush domain.
1269 * Cap @max_active to 1 too.
1270 */
1271 cgroup_pidlist_destroy_wq = alloc_workqueue(fmt: "cgroup_pidlist_destroy",
1272 flags: 0, max_active: 1);
1273 BUG_ON(!cgroup_pidlist_destroy_wq);
1274 return 0;
1275}
1276core_initcall(cgroup1_wq_init);
1277
1278static int __init cgroup_no_v1(char *str)
1279{
1280 struct cgroup_subsys *ss;
1281 char *token;
1282 int i;
1283
1284 while ((token = strsep(&str, ",")) != NULL) {
1285 if (!*token)
1286 continue;
1287
1288 if (!strcmp(token, "all")) {
1289 cgroup_no_v1_mask = U16_MAX;
1290 continue;
1291 }
1292
1293 if (!strcmp(token, "named")) {
1294 cgroup_no_v1_named = true;
1295 continue;
1296 }
1297
1298 for_each_subsys(ss, i) {
1299 if (strcmp(token, ss->name) &&
1300 strcmp(token, ss->legacy_name))
1301 continue;
1302
1303 cgroup_no_v1_mask |= 1 << i;
1304 }
1305 }
1306 return 1;
1307}
1308__setup("cgroup_no_v1=", cgroup_no_v1);
1309

source code of linux/kernel/cgroup/cgroup-v1.c