projects/linux·blob master · refs

1372 lines · 35.9 KiB

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/string.h>
14#include <linux/vmalloc.h>
15#include <linux/delayacct.h>
16#include <linux/pid_namespace.h>
17#include <linux/cgroupstats.h>
18#include <linux/fs_parser.h>
19
20#include <trace/events/cgroup.h>
21
22/*
23 * pidlists linger the following amount before being destroyed. The goal
24 * is avoiding frequent destruction in the middle of consecutive read calls
25 * Expiring in the middle is a performance problem not a correctness one.
26 * 1 sec should be enough.
27 */
28#define CGROUP_PIDLIST_DESTROY_DELAY HZ
29
30/* Controllers blocked by the commandline in v1 */
31static u32 cgroup_no_v1_mask;
32
33/* disable named v1 mounts */
34static bool cgroup_no_v1_named;
35
36/* Show unavailable controllers in /proc/cgroups */
37static bool proc_show_all;
38
39/*
40 * pidlist destructions need to be flushed on cgroup destruction. Use a
41 * separate workqueue as flush domain.
42 */
43static struct workqueue_struct *cgroup_pidlist_destroy_wq;
44
45/* protects cgroup_subsys->release_agent_path */
46static DEFINE_SPINLOCK(release_agent_path_lock);
47
48bool cgroup1_ssid_disabled(int ssid)
49{
50 return cgroup_no_v1_mask & (1 << ssid);
51}
52
53static bool cgroup1_subsys_absent(struct cgroup_subsys *ss)
54{
55 /* Check also dfl_cftypes for file-less controllers, i.e. perf_event */
56 return ss->legacy_cftypes == NULL && ss->dfl_cftypes;
57}
58
59/**
60 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
61 * @from: attach to all cgroups of a given task
62 * @tsk: the task to be attached
63 *
64 * Return: %0 on success or a negative errno code on failure
65 */
66int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
67{
68 struct cgroup_root *root;
69 int retval = 0;
70
71 cgroup_lock();
72 cgroup_attach_lock(CGRP_ATTACH_LOCK_GLOBAL, NULL);
73 for_each_root(root) {
74 struct cgroup *from_cgrp;
75
76 spin_lock_irq(&css_set_lock);
77 from_cgrp = task_cgroup_from_root(from, root);
78 spin_unlock_irq(&css_set_lock);
79
80 retval = cgroup_attach_task(from_cgrp, tsk, false);
81 if (retval)
82 break;
83 }
84 cgroup_attach_unlock(CGRP_ATTACH_LOCK_GLOBAL, NULL);
85 cgroup_unlock();
86
87 return retval;
88}
89EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
90
91/**
92 * cgroup_transfer_tasks - move tasks from one cgroup to another
93 * @to: cgroup to which the tasks will be moved
94 * @from: cgroup in which the tasks currently reside
95 *
96 * Locking rules between cgroup_post_fork() and the migration path
97 * guarantee that, if a task is forking while being migrated, the new child
98 * is guaranteed to be either visible in the source cgroup after the
99 * parent's migration is complete or put into the target cgroup. No task
100 * can slip out of migration through forking.
101 *
102 * Return: %0 on success or a negative errno code on failure
103 */
104int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
105{
106 DEFINE_CGROUP_MGCTX(mgctx);
107 struct cgrp_cset_link *link;
108 struct css_task_iter it;
109 struct task_struct *task;
110 int ret;
111
112 if (cgroup_on_dfl(to))
113 return -EINVAL;
114
115 ret = cgroup_migrate_vet_dst(to);
116 if (ret)
117 return ret;
118
119 cgroup_lock();
120
121 cgroup_attach_lock(CGRP_ATTACH_LOCK_GLOBAL, NULL);
122
123 /* all tasks in @from are being moved, all csets are source */
124 spin_lock_irq(&css_set_lock);
125 list_for_each_entry(link, &from->cset_links, cset_link)
126 cgroup_migrate_add_src(link->cset, to, &mgctx);
127 spin_unlock_irq(&css_set_lock);
128
129 ret = cgroup_migrate_prepare_dst(&mgctx);
130 if (ret)
131 goto out_err;
132
133 /*
134 * Migrate tasks one-by-one until @from is empty. This fails iff
135 * ->can_attach() fails.
136 */
137 do {
138 css_task_iter_start(&from->self, 0, &it);
139
140 do {
141 task = css_task_iter_next(&it);
142 } while (task && (task->flags & PF_EXITING));
143
144 if (task)
145 get_task_struct(task);
146 css_task_iter_end(&it);
147
148 if (task) {
149 ret = cgroup_migrate(task, false, &mgctx);
150 if (!ret)
151 TRACE_CGROUP_PATH(transfer_tasks, to, task, false);
152 put_task_struct(task);
153 }
154 } while (task && !ret);
155out_err:
156 cgroup_migrate_finish(&mgctx);
157 cgroup_attach_unlock(CGRP_ATTACH_LOCK_GLOBAL, NULL);
158 cgroup_unlock();
159 return ret;
160}
161
162/*
163 * Stuff for reading the 'tasks'/'procs' files.
164 *
165 * Reading this file can return large amounts of data if a cgroup has
166 * *lots* of attached tasks. So it may need several calls to read(),
167 * but we cannot guarantee that the information we produce is correct
168 * unless we produce it entirely atomically.
169 *
170 */
171
172/* which pidlist file are we talking about? */
173enum cgroup_filetype {
174 CGROUP_FILE_PROCS,
175 CGROUP_FILE_TASKS,
176};
177
178/*
179 * A pidlist is a list of pids that virtually represents the contents of one
180 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
181 * a pair (one each for procs, tasks) for each pid namespace that's relevant
182 * to the cgroup.
183 */
184struct cgroup_pidlist {
185 /*
186 * used to find which pidlist is wanted. doesn't change as long as
187 * this particular list stays in the list.
188 */
189 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
190 /* array of xids */
191 pid_t *list;
192 /* how many elements the above list has */
193 int length;
194 /* each of these stored in a list by its cgroup */
195 struct list_head links;
196 /* pointer to the cgroup we belong to, for list removal purposes */
197 struct cgroup *owner;
198 /* for delayed destruction */
199 struct delayed_work destroy_dwork;
200};
201
202/*
203 * Used to destroy all pidlists lingering waiting for destroy timer. None
204 * should be left afterwards.
205 */
206void cgroup1_pidlist_destroy_all(struct cgroup *cgrp)
207{
208 struct cgroup_pidlist *l, *tmp_l;
209
210 mutex_lock(&cgrp->pidlist_mutex);
211 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
212 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
213 mutex_unlock(&cgrp->pidlist_mutex);
214
215 flush_workqueue(cgroup_pidlist_destroy_wq);
216 BUG_ON(!list_empty(&cgrp->pidlists));
217}
218
219static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
220{
221 struct delayed_work *dwork = to_delayed_work(work);
222 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
223 destroy_dwork);
224 struct cgroup_pidlist *tofree = NULL;
225
226 mutex_lock(&l->owner->pidlist_mutex);
227
228 /*
229 * Destroy iff we didn't get queued again. The state won't change
230 * as destroy_dwork can only be queued while locked.
231 */
232 if (!delayed_work_pending(dwork)) {
233 list_del(&l->links);
234 kvfree(l->list);
235 put_pid_ns(l->key.ns);
236 tofree = l;
237 }
238
239 mutex_unlock(&l->owner->pidlist_mutex);
240 kfree(tofree);
241}
242
243/*
244 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
245 * Returns the number of unique elements.
246 */
247static int pidlist_uniq(pid_t *list, int length)
248{
249 int src, dest = 1;
250
251 /*
252 * we presume the 0th element is unique, so i starts at 1. trivial
253 * edge cases first; no work needs to be done for either
254 */
255 if (length == 0 || length == 1)
256 return length;
257 /* src and dest walk down the list; dest counts unique elements */
258 for (src = 1; src < length; src++) {
259 /* find next unique element */
260 while (list[src] == list[src-1]) {
261 src++;
262 if (src == length)
263 goto after;
264 }
265 /* dest always points to where the next unique element goes */
266 list[dest] = list[src];
267 dest++;
268 }
269after:
270 return dest;
271}
272
273/*
274 * The two pid files - task and cgroup.procs - guaranteed that the result
275 * is sorted, which forced this whole pidlist fiasco. As pid order is
276 * different per namespace, each namespace needs differently sorted list,
277 * making it impossible to use, for example, single rbtree of member tasks
278 * sorted by task pointer. As pidlists can be fairly large, allocating one
279 * per open file is dangerous, so cgroup had to implement shared pool of
280 * pidlists keyed by cgroup and namespace.
281 */
282static int cmppid(const void *a, const void *b)
283{
284 return *(pid_t *)a - *(pid_t *)b;
285}
286
287static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
288 enum cgroup_filetype type)
289{
290 struct cgroup_pidlist *l;
291 /* don't need task_nsproxy() if we're looking at ourself */
292 struct pid_namespace *ns = task_active_pid_ns(current);
293
294 lockdep_assert_held(&cgrp->pidlist_mutex);
295
296 list_for_each_entry(l, &cgrp->pidlists, links)
297 if (l->key.type == type && l->key.ns == ns)
298 return l;
299 return NULL;
300}
301
302/*
303 * find the appropriate pidlist for our purpose (given procs vs tasks)
304 * returns with the lock on that pidlist already held, and takes care
305 * of the use count, or returns NULL with no locks held if we're out of
306 * memory.
307 */
308static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
309 enum cgroup_filetype type)
310{
311 struct cgroup_pidlist *l;
312
313 lockdep_assert_held(&cgrp->pidlist_mutex);
314
315 l = cgroup_pidlist_find(cgrp, type);
316 if (l)
317 return l;
318
319 /* entry not found; create a new one */
320 l = kzalloc_obj(struct cgroup_pidlist);
321 if (!l)
322 return l;
323
324 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
325 l->key.type = type;
326 /* don't need task_nsproxy() if we're looking at ourself */
327 l->key.ns = get_pid_ns(task_active_pid_ns(current));
328 l->owner = cgrp;
329 list_add(&l->links, &cgrp->pidlists);
330 return l;
331}
332
333/*
334 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
335 */
336static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
337 struct cgroup_pidlist **lp)
338{
339 pid_t *array;
340 int length;
341 int pid, n = 0; /* used for populating the array */
342 struct css_task_iter it;
343 struct task_struct *tsk;
344 struct cgroup_pidlist *l;
345
346 lockdep_assert_held(&cgrp->pidlist_mutex);
347
348 /*
349 * If cgroup gets more users after we read count, we won't have
350 * enough space - tough. This race is indistinguishable to the
351 * caller from the case that the additional cgroup users didn't
352 * show up until sometime later on.
353 */
354 length = cgroup_task_count(cgrp);
355 array = kvmalloc_objs(pid_t, length);
356 if (!array)
357 return -ENOMEM;
358 /* now, populate the array */
359 css_task_iter_start(&cgrp->self, 0, &it);
360 while ((tsk = css_task_iter_next(&it))) {
361 if (unlikely(n == length))
362 break;
363 /* get tgid or pid for procs or tasks file respectively */
364 if (type == CGROUP_FILE_PROCS)
365 pid = task_tgid_vnr(tsk);
366 else
367 pid = task_pid_vnr(tsk);
368 if (pid > 0) /* make sure to only use valid results */
369 array[n++] = pid;
370 }
371 css_task_iter_end(&it);
372 length = n;
373 /* now sort & strip out duplicates (tgids or recycled thread PIDs) */
374 sort(array, length, sizeof(pid_t), cmppid, NULL);
375 length = pidlist_uniq(array, length);
376
377 l = cgroup_pidlist_find_create(cgrp, type);
378 if (!l) {
379 kvfree(array);
380 return -ENOMEM;
381 }
382
383 /* store array, freeing old if necessary */
384 kvfree(l->list);
385 l->list = array;
386 l->length = length;
387 *lp = l;
388 return 0;
389}
390
391/*
392 * seq_file methods for the tasks/procs files. The seq_file position is the
393 * next pid to display; the seq_file iterator is a pointer to the pid
394 * in the cgroup->l->list array.
395 */
396
397static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
398{
399 /*
400 * Initially we receive a position value that corresponds to
401 * one more than the last pid shown (or 0 on the first call or
402 * after a seek to the start). Use a binary-search to find the
403 * next pid to display, if any
404 */
405 struct kernfs_open_file *of = s->private;
406 struct cgroup_file_ctx *ctx = of->priv;
407 struct cgroup *cgrp = seq_css(s)->cgroup;
408 struct cgroup_pidlist *l;
409 enum cgroup_filetype type = seq_cft(s)->private;
410 int index = 0, pid = *pos;
411 int *iter, ret;
412
413 mutex_lock(&cgrp->pidlist_mutex);
414
415 /*
416 * !NULL @ctx->procs1.pidlist indicates that this isn't the first
417 * start() after open. If the matching pidlist is around, we can use
418 * that. Look for it. Note that @ctx->procs1.pidlist can't be used
419 * directly. It could already have been destroyed.
420 */
421 if (ctx->procs1.pidlist)
422 ctx->procs1.pidlist = cgroup_pidlist_find(cgrp, type);
423
424 /*
425 * Either this is the first start() after open or the matching
426 * pidlist has been destroyed inbetween. Create a new one.
427 */
428 if (!ctx->procs1.pidlist) {
429 ret = pidlist_array_load(cgrp, type, &ctx->procs1.pidlist);
430 if (ret)
431 return ERR_PTR(ret);
432 }
433 l = ctx->procs1.pidlist;
434
435 if (pid) {
436 int end = l->length;
437
438 while (index < end) {
439 int mid = (index + end) / 2;
440 if (l->list[mid] == pid) {
441 index = mid;
442 break;
443 } else if (l->list[mid] < pid)
444 index = mid + 1;
445 else
446 end = mid;
447 }
448 }
449 /* If we're off the end of the array, we're done */
450 if (index >= l->length)
451 return NULL;
452 /* Update the abstract position to be the actual pid that we found */
453 iter = l->list + index;
454 *pos = *iter;
455 return iter;
456}
457
458static void cgroup_pidlist_stop(struct seq_file *s, void *v)
459{
460 struct kernfs_open_file *of = s->private;
461 struct cgroup_file_ctx *ctx = of->priv;
462 struct cgroup_pidlist *l = ctx->procs1.pidlist;
463
464 if (l)
465 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
466 CGROUP_PIDLIST_DESTROY_DELAY);
467 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
468}
469
470static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
471{
472 struct kernfs_open_file *of = s->private;
473 struct cgroup_file_ctx *ctx = of->priv;
474 struct cgroup_pidlist *l = ctx->procs1.pidlist;
475 pid_t *p = v;
476 pid_t *end = l->list + l->length;
477 /*
478 * Advance to the next pid in the array. If this goes off the
479 * end, we're done
480 */
481 p++;
482 if (p >= end) {
483 (*pos)++;
484 return NULL;
485 } else {
486 *pos = *p;
487 return p;
488 }
489}
490
491static int cgroup_pidlist_show(struct seq_file *s, void *v)
492{
493 seq_printf(s, "%d\n", *(int *)v);
494
495 return 0;
496}
497
498static ssize_t __cgroup1_procs_write(struct kernfs_open_file *of,
499 char *buf, size_t nbytes, loff_t off,
500 bool threadgroup)
501{
502 struct cgroup *cgrp;
503 struct task_struct *task;
504 const struct cred *cred, *tcred;
505 ssize_t ret;
506 enum cgroup_attach_lock_mode lock_mode;
507
508 cgrp = cgroup_kn_lock_live(of->kn, false);
509 if (!cgrp)
510 return -ENODEV;
511
512 task = cgroup_procs_write_start(buf, threadgroup, &lock_mode);
513 ret = PTR_ERR_OR_ZERO(task);
514 if (ret)
515 goto out_unlock;
516
517 /*
518 * Even if we're attaching all tasks in the thread group, we only need
519 * to check permissions on one of them. Check permissions using the
520 * credentials from file open to protect against inherited fd attacks.
521 */
522 cred = of->file->f_cred;
523 tcred = get_task_cred(task);
524 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
525 !uid_eq(cred->euid, tcred->uid) &&
526 !uid_eq(cred->euid, tcred->suid))
527 ret = -EACCES;
528 put_cred(tcred);
529 if (ret)
530 goto out_finish;
531
532 ret = cgroup_attach_task(cgrp, task, threadgroup);
533
534out_finish:
535 cgroup_procs_write_finish(task, lock_mode);
536out_unlock:
537 cgroup_kn_unlock(of->kn);
538
539 return ret ?: nbytes;
540}
541
542static ssize_t cgroup1_procs_write(struct kernfs_open_file *of,
543 char *buf, size_t nbytes, loff_t off)
544{
545 return __cgroup1_procs_write(of, buf, nbytes, off, true);
546}
547
548static ssize_t cgroup1_tasks_write(struct kernfs_open_file *of,
549 char *buf, size_t nbytes, loff_t off)
550{
551 return __cgroup1_procs_write(of, buf, nbytes, off, false);
552}
553
554static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
555 char *buf, size_t nbytes, loff_t off)
556{
557 struct cgroup *cgrp;
558 struct cgroup_file_ctx *ctx;
559
560 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
561
562 /*
563 * Release agent gets called with all capabilities,
564 * require capabilities to set release agent.
565 */
566 ctx = of->priv;
567 if ((ctx->ns->user_ns != &init_user_ns) ||
568 !file_ns_capable(of->file, &init_user_ns, CAP_SYS_ADMIN))
569 return -EPERM;
570
571 cgrp = cgroup_kn_lock_live(of->kn, false);
572 if (!cgrp)
573 return -ENODEV;
574 spin_lock(&release_agent_path_lock);
575 strscpy(cgrp->root->release_agent_path, strstrip(buf),
576 sizeof(cgrp->root->release_agent_path));
577 spin_unlock(&release_agent_path_lock);
578 cgroup_kn_unlock(of->kn);
579 return nbytes;
580}
581
582static int cgroup_release_agent_show(struct seq_file *seq, void *v)
583{
584 struct cgroup *cgrp = seq_css(seq)->cgroup;
585
586 spin_lock(&release_agent_path_lock);
587 seq_puts(seq, cgrp->root->release_agent_path);
588 spin_unlock(&release_agent_path_lock);
589 seq_putc(seq, '\n');
590 return 0;
591}
592
593static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
594{
595 seq_puts(seq, "0\n");
596 return 0;
597}
598
599static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
600 struct cftype *cft)
601{
602 return notify_on_release(css->cgroup);
603}
604
605static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
606 struct cftype *cft, u64 val)
607{
608 if (val)
609 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
610 else
611 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
612 return 0;
613}
614
615static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
616 struct cftype *cft)
617{
618 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
619}
620
621static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
622 struct cftype *cft, u64 val)
623{
624 if (val)
625 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
626 else
627 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
628 return 0;
629}
630
631/* cgroup core interface files for the legacy hierarchies */
632struct cftype cgroup1_base_files[] = {
633 {
634 .name = "cgroup.procs",
635 .seq_start = cgroup_pidlist_start,
636 .seq_next = cgroup_pidlist_next,
637 .seq_stop = cgroup_pidlist_stop,
638 .seq_show = cgroup_pidlist_show,
639 .private = CGROUP_FILE_PROCS,
640 .write = cgroup1_procs_write,
641 },
642 {
643 .name = "cgroup.clone_children",
644 .read_u64 = cgroup_clone_children_read,
645 .write_u64 = cgroup_clone_children_write,
646 },
647 {
648 .name = "cgroup.sane_behavior",
649 .flags = CFTYPE_ONLY_ON_ROOT,
650 .seq_show = cgroup_sane_behavior_show,
651 },
652 {
653 .name = "tasks",
654 .seq_start = cgroup_pidlist_start,
655 .seq_next = cgroup_pidlist_next,
656 .seq_stop = cgroup_pidlist_stop,
657 .seq_show = cgroup_pidlist_show,
658 .private = CGROUP_FILE_TASKS,
659 .write = cgroup1_tasks_write,
660 },
661 {
662 .name = "notify_on_release",
663 .read_u64 = cgroup_read_notify_on_release,
664 .write_u64 = cgroup_write_notify_on_release,
665 },
666 {
667 .name = "release_agent",
668 .flags = CFTYPE_ONLY_ON_ROOT,
669 .seq_show = cgroup_release_agent_show,
670 .write = cgroup_release_agent_write,
671 .max_write_len = PATH_MAX - 1,
672 },
673 { } /* terminate */
674};
675
676/* Display information about each subsystem and each hierarchy */
677int proc_cgroupstats_show(struct seq_file *m, void *v)
678{
679 struct cgroup_subsys *ss;
680 bool cgrp_v1_visible = false;
681 int i;
682
683 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
684 /*
685 * Grab the subsystems state racily. No need to add avenue to
686 * cgroup_mutex contention.
687 */
688
689 for_each_subsys(ss, i) {
690 cgrp_v1_visible |= ss->root != &cgrp_dfl_root;
691
692 if (!proc_show_all && cgroup1_subsys_absent(ss))
693 continue;
694
695 seq_printf(m, "%s\t%d\t%d\t%d\n",
696 ss->legacy_name, ss->root->hierarchy_id,
697 atomic_read(&ss->root->nr_cgrps),
698 cgroup_ssid_enabled(i));
699 }
700
701 if (cgrp_dfl_visible && !cgrp_v1_visible)
702 pr_info_once("/proc/cgroups lists only v1 controllers, use cgroup.controllers of root cgroup for v2 info\n");
703
704
705 return 0;
706}
707
708/**
709 * cgroupstats_build - build and fill cgroupstats
710 * @stats: cgroupstats to fill information into
711 * @dentry: A dentry entry belonging to the cgroup for which stats have
712 * been requested.
713 *
714 * Build and fill cgroupstats so that taskstats can export it to user
715 * space.
716 *
717 * Return: %0 on success or a negative errno code on failure
718 */
719int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
720{
721 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
722 struct cgroup *cgrp;
723 struct css_task_iter it;
724 struct task_struct *tsk;
725
726 /* it should be kernfs_node belonging to cgroupfs and is a directory */
727 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
728 kernfs_type(kn) != KERNFS_DIR)
729 return -EINVAL;
730
731 /*
732 * We aren't being called from kernfs and there's no guarantee on
733 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
734 * @kn->priv is RCU safe. Let's do the RCU dancing.
735 */
736 rcu_read_lock();
737 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
738 if (!cgrp || !cgroup_tryget(cgrp)) {
739 rcu_read_unlock();
740 return -ENOENT;
741 }
742 rcu_read_unlock();
743
744 css_task_iter_start(&cgrp->self, 0, &it);
745 while ((tsk = css_task_iter_next(&it))) {
746 switch (READ_ONCE(tsk->__state)) {
747 case TASK_RUNNING:
748 stats->nr_running++;
749 break;
750 case TASK_INTERRUPTIBLE:
751 stats->nr_sleeping++;
752 break;
753 case TASK_UNINTERRUPTIBLE:
754 stats->nr_uninterruptible++;
755 break;
756 case TASK_STOPPED:
757 stats->nr_stopped++;
758 break;
759 default:
760 if (tsk->in_iowait)
761 stats->nr_io_wait++;
762 break;
763 }
764 }
765 css_task_iter_end(&it);
766
767 cgroup_put(cgrp);
768 return 0;
769}
770
771void cgroup1_check_for_release(struct cgroup *cgrp)
772{
773 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
774 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
775 schedule_work(&cgrp->release_agent_work);
776}
777
778/*
779 * Notify userspace when a cgroup is released, by running the
780 * configured release agent with the name of the cgroup (path
781 * relative to the root of cgroup file system) as the argument.
782 *
783 * Most likely, this user command will try to rmdir this cgroup.
784 *
785 * This races with the possibility that some other task will be
786 * attached to this cgroup before it is removed, or that some other
787 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
788 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
789 * unused, and this cgroup will be reprieved from its death sentence,
790 * to continue to serve a useful existence. Next time it's released,
791 * we will get notified again, if it still has 'notify_on_release' set.
792 *
793 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
794 * means only wait until the task is successfully execve()'d. The
795 * separate release agent task is forked by call_usermodehelper(),
796 * then control in this thread returns here, without waiting for the
797 * release agent task. We don't bother to wait because the caller of
798 * this routine has no use for the exit status of the release agent
799 * task, so no sense holding our caller up for that.
800 */
801void cgroup1_release_agent(struct work_struct *work)
802{
803 struct cgroup *cgrp =
804 container_of(work, struct cgroup, release_agent_work);
805 char *pathbuf, *agentbuf;
806 char *argv[3], *envp[3];
807 int ret;
808
809 /* snoop agent path and exit early if empty */
810 if (!cgrp->root->release_agent_path[0])
811 return;
812
813 /* prepare argument buffers */
814 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
815 agentbuf = kmalloc(PATH_MAX, GFP_KERNEL);
816 if (!pathbuf || !agentbuf)
817 goto out_free;
818
819 spin_lock(&release_agent_path_lock);
820 strscpy(agentbuf, cgrp->root->release_agent_path, PATH_MAX);
821 spin_unlock(&release_agent_path_lock);
822 if (!agentbuf[0])
823 goto out_free;
824
825 ret = cgroup_path_ns(cgrp, pathbuf, PATH_MAX, &init_cgroup_ns);
826 if (ret < 0)
827 goto out_free;
828
829 argv[0] = agentbuf;
830 argv[1] = pathbuf;
831 argv[2] = NULL;
832
833 /* minimal command environment */
834 envp[0] = "HOME=/";
835 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
836 envp[2] = NULL;
837
838 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
839out_free:
840 kfree(agentbuf);
841 kfree(pathbuf);
842}
843
844/*
845 * cgroup_rename - Only allow simple rename of directories in place.
846 */
847static int cgroup1_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
848 const char *new_name_str)
849{
850 struct cgroup *cgrp = kn->priv;
851 int ret;
852
853 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
854 if (strchr(new_name_str, '\n'))
855 return -EINVAL;
856
857 if (kernfs_type(kn) != KERNFS_DIR)
858 return -ENOTDIR;
859 if (rcu_access_pointer(kn->__parent) != new_parent)
860 return -EIO;
861
862 /*
863 * We're gonna grab cgroup_mutex which nests outside kernfs
864 * active_ref. kernfs_rename() doesn't require active_ref
865 * protection. Break them before grabbing cgroup_mutex.
866 */
867 kernfs_break_active_protection(new_parent);
868 kernfs_break_active_protection(kn);
869
870 cgroup_lock();
871
872 ret = kernfs_rename(kn, new_parent, new_name_str);
873 if (!ret)
874 TRACE_CGROUP_PATH(rename, cgrp);
875
876 cgroup_unlock();
877
878 kernfs_unbreak_active_protection(kn);
879 kernfs_unbreak_active_protection(new_parent);
880 return ret;
881}
882
883static int cgroup1_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
884{
885 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
886 struct cgroup_subsys *ss;
887 int ssid;
888
889 for_each_subsys(ss, ssid)
890 if (root->subsys_mask & (1 << ssid))
891 seq_show_option(seq, ss->legacy_name, NULL);
892 if (root->flags & CGRP_ROOT_NOPREFIX)
893 seq_puts(seq, ",noprefix");
894 if (root->flags & CGRP_ROOT_XATTR)
895 seq_puts(seq, ",xattr");
896 if (root->flags & CGRP_ROOT_CPUSET_V2_MODE)
897 seq_puts(seq, ",cpuset_v2_mode");
898 if (root->flags & CGRP_ROOT_FAVOR_DYNMODS)
899 seq_puts(seq, ",favordynmods");
900
901 spin_lock(&release_agent_path_lock);
902 if (strlen(root->release_agent_path))
903 seq_show_option(seq, "release_agent",
904 root->release_agent_path);
905 spin_unlock(&release_agent_path_lock);
906
907 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
908 seq_puts(seq, ",clone_children");
909 if (strlen(root->name))
910 seq_show_option(seq, "name", root->name);
911 return 0;
912}
913
914enum cgroup1_param {
915 Opt_all,
916 Opt_clone_children,
917 Opt_cpuset_v2_mode,
918 Opt_name,
919 Opt_none,
920 Opt_noprefix,
921 Opt_release_agent,
922 Opt_xattr,
923 Opt_favordynmods,
924 Opt_nofavordynmods,
925};
926
927const struct fs_parameter_spec cgroup1_fs_parameters[] = {
928 fsparam_flag ("all", Opt_all),
929 fsparam_flag ("clone_children", Opt_clone_children),
930 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode),
931 fsparam_string("name", Opt_name),
932 fsparam_flag ("none", Opt_none),
933 fsparam_flag ("noprefix", Opt_noprefix),
934 fsparam_string("release_agent", Opt_release_agent),
935 fsparam_flag ("xattr", Opt_xattr),
936 fsparam_flag ("favordynmods", Opt_favordynmods),
937 fsparam_flag ("nofavordynmods", Opt_nofavordynmods),
938 {}
939};
940
941int cgroup1_parse_param(struct fs_context *fc, struct fs_parameter *param)
942{
943 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
944 struct cgroup_subsys *ss;
945 struct fs_parse_result result;
946 int opt, i;
947
948 opt = fs_parse(fc, cgroup1_fs_parameters, param, &result);
949 if (opt == -ENOPARAM) {
950 int ret;
951
952 ret = vfs_parse_fs_param_source(fc, param);
953 if (ret != -ENOPARAM)
954 return ret;
955 for_each_subsys(ss, i) {
956 if (strcmp(param->key, ss->legacy_name) ||
957 cgroup1_subsys_absent(ss))
958 continue;
959 if (!cgroup_ssid_enabled(i) || cgroup1_ssid_disabled(i))
960 return invalfc(fc, "Disabled controller '%s'",
961 param->key);
962 ctx->subsys_mask |= (1 << i);
963 return 0;
964 }
965 return invalfc(fc, "Unknown subsys name '%s'", param->key);
966 }
967 if (opt < 0)
968 return opt;
969
970 switch (opt) {
971 case Opt_none:
972 /* Explicitly have no subsystems */
973 ctx->none = true;
974 break;
975 case Opt_all:
976 ctx->all_ss = true;
977 break;
978 case Opt_noprefix:
979 ctx->flags |= CGRP_ROOT_NOPREFIX;
980 break;
981 case Opt_clone_children:
982 ctx->cpuset_clone_children = true;
983 break;
984 case Opt_cpuset_v2_mode:
985 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE;
986 break;
987 case Opt_xattr:
988 ctx->flags |= CGRP_ROOT_XATTR;
989 break;
990 case Opt_favordynmods:
991 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
992 break;
993 case Opt_nofavordynmods:
994 ctx->flags &= ~CGRP_ROOT_FAVOR_DYNMODS;
995 break;
996 case Opt_release_agent:
997 /* Specifying two release agents is forbidden */
998 if (ctx->release_agent)
999 return invalfc(fc, "release_agent respecified");
1000 /*
1001 * Release agent gets called with all capabilities,
1002 * require capabilities to set release agent.
1003 */
1004 if ((fc->user_ns != &init_user_ns) || !capable(CAP_SYS_ADMIN))
1005 return invalfc(fc, "Setting release_agent not allowed");
1006 ctx->release_agent = param->string;
1007 param->string = NULL;
1008 break;
1009 case Opt_name:
1010 /* blocked by boot param? */
1011 if (cgroup_no_v1_named)
1012 return -ENOENT;
1013 /* Can't specify an empty name */
1014 if (!param->size)
1015 return invalfc(fc, "Empty name");
1016 if (param->size > MAX_CGROUP_ROOT_NAMELEN - 1)
1017 return invalfc(fc, "Name too long");
1018 /* Must match [\w.-]+ */
1019 for (i = 0; i < param->size; i++) {
1020 char c = param->string[i];
1021 if (isalnum(c))
1022 continue;
1023 if ((c == '.') || (c == '-') || (c == '_'))
1024 continue;
1025 return invalfc(fc, "Invalid name");
1026 }
1027 /* Specifying two names is forbidden */
1028 if (ctx->name)
1029 return invalfc(fc, "name respecified");
1030 ctx->name = param->string;
1031 param->string = NULL;
1032 break;
1033 }
1034 return 0;
1035}
1036
1037static int check_cgroupfs_options(struct fs_context *fc)
1038{
1039 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1040 u32 mask = U32_MAX;
1041 u32 enabled = 0;
1042 struct cgroup_subsys *ss;
1043 int i;
1044
1045#ifdef CONFIG_CPUSETS
1046 mask = ~((u32)1 << cpuset_cgrp_id);
1047#endif
1048 for_each_subsys(ss, i)
1049 if (cgroup_ssid_enabled(i) && !cgroup1_ssid_disabled(i) &&
1050 !cgroup1_subsys_absent(ss))
1051 enabled |= 1 << i;
1052
1053 ctx->subsys_mask &= enabled;
1054
1055 /*
1056 * In absence of 'none', 'name=' and subsystem name options,
1057 * let's default to 'all'.
1058 */
1059 if (!ctx->subsys_mask && !ctx->none && !ctx->name)
1060 ctx->all_ss = true;
1061
1062 if (ctx->all_ss) {
1063 /* Mutually exclusive option 'all' + subsystem name */
1064 if (ctx->subsys_mask)
1065 return invalfc(fc, "subsys name conflicts with all");
1066 /* 'all' => select all the subsystems */
1067 ctx->subsys_mask = enabled;
1068 }
1069
1070 /*
1071 * We either have to specify by name or by subsystems. (So all
1072 * empty hierarchies must have a name).
1073 */
1074 if (!ctx->subsys_mask && !ctx->name)
1075 return invalfc(fc, "Need name or subsystem set");
1076
1077 /*
1078 * Option noprefix was introduced just for backward compatibility
1079 * with the old cpuset, so we allow noprefix only if mounting just
1080 * the cpuset subsystem.
1081 */
1082 if ((ctx->flags & CGRP_ROOT_NOPREFIX) && (ctx->subsys_mask & mask))
1083 return invalfc(fc, "noprefix used incorrectly");
1084
1085 /* Can't specify "none" and some subsystems */
1086 if (ctx->subsys_mask && ctx->none)
1087 return invalfc(fc, "none used incorrectly");
1088
1089 return 0;
1090}
1091
1092int cgroup1_reconfigure(struct fs_context *fc)
1093{
1094 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1095 struct kernfs_root *kf_root = kernfs_root_from_sb(fc->root->d_sb);
1096 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1097 int ret = 0;
1098 u32 added_mask, removed_mask;
1099
1100 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1101
1102 /* See what subsystems are wanted */
1103 ret = check_cgroupfs_options(fc);
1104 if (ret)
1105 goto out_unlock;
1106
1107 if (ctx->subsys_mask != root->subsys_mask || ctx->release_agent)
1108 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1109 task_tgid_nr(current), current->comm);
1110
1111 added_mask = ctx->subsys_mask & ~root->subsys_mask;
1112 removed_mask = root->subsys_mask & ~ctx->subsys_mask;
1113
1114 /* Don't allow flags or name to change at remount */
1115 if ((ctx->flags ^ root->flags) ||
1116 (ctx->name && strcmp(ctx->name, root->name))) {
1117 errorfc(fc, "option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"",
1118 ctx->flags, ctx->name ?: "", root->flags, root->name);
1119 ret = -EINVAL;
1120 goto out_unlock;
1121 }
1122
1123 /* remounting is not allowed for populated hierarchies */
1124 if (!list_empty(&root->cgrp.self.children)) {
1125 ret = -EBUSY;
1126 goto out_unlock;
1127 }
1128
1129 ret = rebind_subsystems(root, added_mask);
1130 if (ret)
1131 goto out_unlock;
1132
1133 WARN_ON(rebind_subsystems(&cgrp_dfl_root, removed_mask));
1134
1135 if (ctx->release_agent) {
1136 spin_lock(&release_agent_path_lock);
1137 strscpy(root->release_agent_path, ctx->release_agent);
1138 spin_unlock(&release_agent_path_lock);
1139 }
1140
1141 trace_cgroup_remount(root);
1142
1143 out_unlock:
1144 cgroup_unlock();
1145 return ret;
1146}
1147
1148struct kernfs_syscall_ops cgroup1_kf_syscall_ops = {
1149 .rename = cgroup1_rename,
1150 .show_options = cgroup1_show_options,
1151 .mkdir = cgroup_mkdir,
1152 .rmdir = cgroup_rmdir,
1153 .show_path = cgroup_show_path,
1154};
1155
1156/*
1157 * The guts of cgroup1 mount - find or create cgroup_root to use.
1158 * Called with cgroup_mutex held; returns 0 on success, -E... on
1159 * error and positive - in case when the candidate is busy dying.
1160 * On success it stashes a reference to cgroup_root into given
1161 * cgroup_fs_context; that reference is *NOT* counting towards the
1162 * cgroup_root refcount.
1163 */
1164static int cgroup1_root_to_use(struct fs_context *fc)
1165{
1166 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1167 struct cgroup_root *root;
1168 struct cgroup_subsys *ss;
1169 int i, ret;
1170
1171 /* First find the desired set of subsystems */
1172 ret = check_cgroupfs_options(fc);
1173 if (ret)
1174 return ret;
1175
1176 /*
1177 * Destruction of cgroup root is asynchronous, so subsystems may
1178 * still be dying after the previous unmount. Let's drain the
1179 * dying subsystems. We just need to ensure that the ones
1180 * unmounted previously finish dying and don't care about new ones
1181 * starting. Testing ref liveliness is good enough.
1182 */
1183 for_each_subsys(ss, i) {
1184 if (!(ctx->subsys_mask & (1 << i)) ||
1185 ss->root == &cgrp_dfl_root)
1186 continue;
1187
1188 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt))
1189 return 1; /* restart */
1190 cgroup_put(&ss->root->cgrp);
1191 }
1192
1193 for_each_root(root) {
1194 bool name_match = false;
1195
1196 if (root == &cgrp_dfl_root)
1197 continue;
1198
1199 /*
1200 * If we asked for a name then it must match. Also, if
1201 * name matches but sybsys_mask doesn't, we should fail.
1202 * Remember whether name matched.
1203 */
1204 if (ctx->name) {
1205 if (strcmp(ctx->name, root->name))
1206 continue;
1207 name_match = true;
1208 }
1209
1210 /*
1211 * If we asked for subsystems (or explicitly for no
1212 * subsystems) then they must match.
1213 */
1214 if ((ctx->subsys_mask || ctx->none) &&
1215 (ctx->subsys_mask != root->subsys_mask)) {
1216 if (!name_match)
1217 continue;
1218 return -EBUSY;
1219 }
1220
1221 if (root->flags ^ ctx->flags)
1222 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1223
1224 ctx->root = root;
1225 return 0;
1226 }
1227
1228 /*
1229 * No such thing, create a new one. name= matching without subsys
1230 * specification is allowed for already existing hierarchies but we
1231 * can't create new one without subsys specification.
1232 */
1233 if (!ctx->subsys_mask && !ctx->none)
1234 return invalfc(fc, "No subsys list or none specified");
1235
1236 /* Hierarchies may only be created in the initial cgroup namespace. */
1237 if (ctx->ns != &init_cgroup_ns)
1238 return -EPERM;
1239
1240 root = kzalloc_obj(*root);
1241 if (!root)
1242 return -ENOMEM;
1243
1244 ctx->root = root;
1245 init_cgroup_root(ctx);
1246
1247 ret = cgroup_setup_root(root, ctx->subsys_mask);
1248 if (!ret)
1249 cgroup_favor_dynmods(root, ctx->flags & CGRP_ROOT_FAVOR_DYNMODS);
1250 else
1251 cgroup_free_root(root);
1252
1253 return ret;
1254}
1255
1256int cgroup1_get_tree(struct fs_context *fc)
1257{
1258 struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1259 int ret;
1260
1261 /* Check if the caller has permission to mount. */
1262 if (!ns_capable(ctx->ns->user_ns, CAP_SYS_ADMIN))
1263 return -EPERM;
1264
1265 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1266
1267 ret = cgroup1_root_to_use(fc);
1268 if (!ret && !percpu_ref_tryget_live(&ctx->root->cgrp.self.refcnt))
1269 ret = 1; /* restart */
1270
1271 cgroup_unlock();
1272
1273 if (!ret)
1274 ret = cgroup_do_get_tree(fc);
1275
1276 if (!ret && percpu_ref_is_dying(&ctx->root->cgrp.self.refcnt)) {
1277 fc_drop_locked(fc);
1278 ret = 1;
1279 }
1280
1281 if (unlikely(ret > 0)) {
1282 msleep(10);
1283 return restart_syscall();
1284 }
1285 return ret;
1286}
1287
1288/**
1289 * task_get_cgroup1 - Acquires the associated cgroup of a task within a
1290 * specific cgroup1 hierarchy. The cgroup1 hierarchy is identified by its
1291 * hierarchy ID.
1292 * @tsk: The target task
1293 * @hierarchy_id: The ID of a cgroup1 hierarchy
1294 *
1295 * On success, the cgroup is returned. On failure, ERR_PTR is returned.
1296 * We limit it to cgroup1 only.
1297 */
1298struct cgroup *task_get_cgroup1(struct task_struct *tsk, int hierarchy_id)
1299{
1300 struct cgroup *cgrp = ERR_PTR(-ENOENT);
1301 struct cgroup_root *root;
1302 unsigned long flags;
1303
1304 rcu_read_lock();
1305 for_each_root(root) {
1306 /* cgroup1 only*/
1307 if (root == &cgrp_dfl_root)
1308 continue;
1309 if (root->hierarchy_id != hierarchy_id)
1310 continue;
1311 spin_lock_irqsave(&css_set_lock, flags);
1312 cgrp = task_cgroup_from_root(tsk, root);
1313 if (!cgrp || !cgroup_tryget(cgrp))
1314 cgrp = ERR_PTR(-ENOENT);
1315 spin_unlock_irqrestore(&css_set_lock, flags);
1316 break;
1317 }
1318 rcu_read_unlock();
1319 return cgrp;
1320}
1321
1322static int __init cgroup1_wq_init(void)
1323{
1324 /*
1325 * Used to destroy pidlists and separate to serve as flush domain.
1326 * Cap @max_active to 1 too.
1327 */
1328 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
1329 WQ_PERCPU, 1);
1330 BUG_ON(!cgroup_pidlist_destroy_wq);
1331 return 0;
1332}
1333core_initcall(cgroup1_wq_init);
1334
1335static int __init cgroup_no_v1(char *str)
1336{
1337 struct cgroup_subsys *ss;
1338 char *token;
1339 int i;
1340
1341 while ((token = strsep(&str, ",")) != NULL) {
1342 if (!*token)
1343 continue;
1344
1345 if (!strcmp(token, "all")) {
1346 cgroup_no_v1_mask = U32_MAX;
1347 continue;
1348 }
1349
1350 if (!strcmp(token, "named")) {
1351 cgroup_no_v1_named = true;
1352 continue;
1353 }
1354
1355 for_each_subsys(ss, i) {
1356 if (strcmp(token, ss->name) &&
1357 strcmp(token, ss->legacy_name))
1358 continue;
1359
1360 cgroup_no_v1_mask |= 1 << i;
1361 break;
1362 }
1363 }
1364 return 1;
1365}
1366__setup("cgroup_no_v1=", cgroup_no_v1);
1367
1368static int __init cgroup_v1_proc(char *str)
1369{
1370 return (kstrtobool(str, &proc_show_all) == 0);
1371}
1372__setup("cgroup_v1_proc=", cgroup_v1_proc);