2 * Generic process-grouping system.
4 * Based originally on the cpuset system, extracted by Paul Menage
5 * Copyright (C) 2006 Google, Inc
7 * Notifications support
8 * Copyright (C) 2009 Nokia Corporation
9 * Author: Kirill A. Shutemov
11 * Copyright notices from the original cpuset code:
12 * --------------------------------------------------
13 * Copyright (C) 2003 BULL SA.
14 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
16 * Portions derived from Patrick Mochel's sysfs code.
17 * sysfs is Copyright (c) 2001-3 Patrick Mochel
19 * 2003-10-10 Written by Simon Derr.
20 * 2003-10-22 Updates by Stephen Hemminger.
21 * 2004 May-July Rework by Paul Jackson.
22 * ---------------------------------------------------
24 * This file is subject to the terms and conditions of the GNU General Public
25 * License. See the file COPYING in the main directory of the Linux
26 * distribution for more details.
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
31 #include <linux/cgroup.h>
32 #include <linux/cred.h>
33 #include <linux/ctype.h>
34 #include <linux/errno.h>
35 #include <linux/init_task.h>
36 #include <linux/kernel.h>
37 #include <linux/list.h>
38 #include <linux/magic.h>
40 #include <linux/mutex.h>
41 #include <linux/mount.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/rwsem.h>
49 #include <linux/percpu-rwsem.h>
50 #include <linux/string.h>
51 #include <linux/sort.h>
52 #include <linux/kmod.h>
53 #include <linux/delayacct.h>
54 #include <linux/cgroupstats.h>
55 #include <linux/hashtable.h>
56 #include <linux/pid_namespace.h>
57 #include <linux/idr.h>
58 #include <linux/vmalloc.h> /* TODO: replace with more sophisticated array */
59 #include <linux/kthread.h>
60 #include <linux/delay.h>
62 #include <linux/atomic.h>
65 * pidlists linger the following amount before being destroyed. The goal
66 * is avoiding frequent destruction in the middle of consecutive read calls
67 * Expiring in the middle is a performance problem not a correctness one.
68 * 1 sec should be enough.
70 #define CGROUP_PIDLIST_DESTROY_DELAY HZ
72 #define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
76 * cgroup_mutex is the master lock. Any modification to cgroup or its
77 * hierarchy must be performed while holding it.
79 * css_set_rwsem protects task->cgroups pointer, the list of css_set
80 * objects, and the chain of tasks off each css_set.
82 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
83 * cgroup.h can use them for lockdep annotations.
85 #ifdef CONFIG_PROVE_RCU
86 DEFINE_MUTEX(cgroup_mutex);
87 DECLARE_RWSEM(css_set_rwsem);
88 EXPORT_SYMBOL_GPL(cgroup_mutex);
89 EXPORT_SYMBOL_GPL(css_set_rwsem);
91 static DEFINE_MUTEX(cgroup_mutex);
92 static DECLARE_RWSEM(css_set_rwsem);
96 * Protects cgroup_idr and css_idr so that IDs can be released without
97 * grabbing cgroup_mutex.
99 static DEFINE_SPINLOCK(cgroup_idr_lock);
102 * Protects cgroup_subsys->release_agent_path. Modifying it also requires
103 * cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
105 static DEFINE_SPINLOCK(release_agent_path_lock);
107 struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
109 #define cgroup_assert_mutex_or_rcu_locked() \
110 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
111 !lockdep_is_held(&cgroup_mutex), \
112 "cgroup_mutex or RCU read lock required");
115 * cgroup destruction makes heavy use of work items and there can be a lot
116 * of concurrent destructions. Use a separate workqueue so that cgroup
117 * destruction work items don't end up filling up max_active of system_wq
118 * which may lead to deadlock.
120 static struct workqueue_struct *cgroup_destroy_wq;
123 * pidlist destructions need to be flushed on cgroup destruction. Use a
124 * separate workqueue as flush domain.
126 static struct workqueue_struct *cgroup_pidlist_destroy_wq;
128 /* generate an array of cgroup subsystem pointers */
129 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
130 static struct cgroup_subsys *cgroup_subsys[] = {
131 #include <linux/cgroup_subsys.h>
135 /* array of cgroup subsystem names */
136 #define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
137 static const char *cgroup_subsys_name[] = {
138 #include <linux/cgroup_subsys.h>
143 * The default hierarchy, reserved for the subsystems that are otherwise
144 * unattached - it never has more than a single cgroup, and all tasks are
145 * part of that cgroup.
147 struct cgroup_root cgrp_dfl_root;
148 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
151 * The default hierarchy always exists but is hidden until mounted for the
152 * first time. This is for backward compatibility.
154 static bool cgrp_dfl_root_visible;
157 * Set by the boot param of the same name and makes subsystems with NULL
158 * ->dfl_files to use ->legacy_files on the default hierarchy.
160 static bool cgroup_legacy_files_on_dfl;
162 /* some controllers are not supported in the default hierarchy */
163 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
165 /* The list of hierarchy roots */
167 static LIST_HEAD(cgroup_roots);
168 static int cgroup_root_count;
170 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
171 static DEFINE_IDR(cgroup_hierarchy_idr);
174 * Assign a monotonically increasing serial number to csses. It guarantees
175 * cgroups with bigger numbers are newer than those with smaller numbers.
176 * Also, as csses are always appended to the parent's ->children list, it
177 * guarantees that sibling csses are always sorted in the ascending serial
178 * number order on the list. Protected by cgroup_mutex.
180 static u64 css_serial_nr_next = 1;
183 * These bitmask flags indicate whether tasks in the fork and exit paths have
184 * fork/exit handlers to call. This avoids us having to do extra work in the
185 * fork/exit path to check which subsystems have fork/exit callbacks.
187 static unsigned long have_fork_callback __read_mostly;
188 static unsigned long have_exit_callback __read_mostly;
190 /* Ditto for the can_fork callback. */
191 static unsigned long have_canfork_callback __read_mostly;
193 static struct cftype cgroup_dfl_base_files[];
194 static struct cftype cgroup_legacy_base_files[];
196 static int rebind_subsystems(struct cgroup_root *dst_root,
197 unsigned long ss_mask);
198 static int cgroup_destroy_locked(struct cgroup *cgrp);
199 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
201 static void css_release(struct percpu_ref *ref);
202 static void kill_css(struct cgroup_subsys_state *css);
203 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
206 /* IDR wrappers which synchronize using cgroup_idr_lock */
207 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
212 idr_preload(gfp_mask);
213 spin_lock_bh(&cgroup_idr_lock);
214 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_WAIT);
215 spin_unlock_bh(&cgroup_idr_lock);
220 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
224 spin_lock_bh(&cgroup_idr_lock);
225 ret = idr_replace(idr, ptr, id);
226 spin_unlock_bh(&cgroup_idr_lock);
230 static void cgroup_idr_remove(struct idr *idr, int id)
232 spin_lock_bh(&cgroup_idr_lock);
234 spin_unlock_bh(&cgroup_idr_lock);
237 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
239 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
242 return container_of(parent_css, struct cgroup, self);
247 * cgroup_css - obtain a cgroup's css for the specified subsystem
248 * @cgrp: the cgroup of interest
249 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
251 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
252 * function must be called either under cgroup_mutex or rcu_read_lock() and
253 * the caller is responsible for pinning the returned css if it wants to
254 * keep accessing it outside the said locks. This function may return
255 * %NULL if @cgrp doesn't have @subsys_id enabled.
257 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
258 struct cgroup_subsys *ss)
261 return rcu_dereference_check(cgrp->subsys[ss->id],
262 lockdep_is_held(&cgroup_mutex));
268 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
269 * @cgrp: the cgroup of interest
270 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
272 * Similar to cgroup_css() but returns the effective css, which is defined
273 * as the matching css of the nearest ancestor including self which has @ss
274 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
275 * function is guaranteed to return non-NULL css.
277 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
278 struct cgroup_subsys *ss)
280 lockdep_assert_held(&cgroup_mutex);
285 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
289 * This function is used while updating css associations and thus
290 * can't test the csses directly. Use ->child_subsys_mask.
292 while (cgroup_parent(cgrp) &&
293 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
294 cgrp = cgroup_parent(cgrp);
296 return cgroup_css(cgrp, ss);
300 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
301 * @cgrp: the cgroup of interest
302 * @ss: the subsystem of interest
304 * Find and get the effective css of @cgrp for @ss. The effective css is
305 * defined as the matching css of the nearest ancestor including self which
306 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
307 * the root css is returned, so this function always returns a valid css.
308 * The returned css must be put using css_put().
310 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
311 struct cgroup_subsys *ss)
313 struct cgroup_subsys_state *css;
318 css = cgroup_css(cgrp, ss);
320 if (css && css_tryget_online(css))
322 cgrp = cgroup_parent(cgrp);
325 css = init_css_set.subsys[ss->id];
332 /* convenient tests for these bits */
333 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
335 return !(cgrp->self.flags & CSS_ONLINE);
338 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
340 struct cgroup *cgrp = of->kn->parent->priv;
341 struct cftype *cft = of_cft(of);
344 * This is open and unprotected implementation of cgroup_css().
345 * seq_css() is only called from a kernfs file operation which has
346 * an active reference on the file. Because all the subsystem
347 * files are drained before a css is disassociated with a cgroup,
348 * the matching css from the cgroup's subsys table is guaranteed to
349 * be and stay valid until the enclosing operation is complete.
352 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
356 EXPORT_SYMBOL_GPL(of_css);
359 * cgroup_is_descendant - test ancestry
360 * @cgrp: the cgroup to be tested
361 * @ancestor: possible ancestor of @cgrp
363 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
364 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
365 * and @ancestor are accessible.
367 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
370 if (cgrp == ancestor)
372 cgrp = cgroup_parent(cgrp);
377 static int notify_on_release(const struct cgroup *cgrp)
379 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
383 * for_each_css - iterate all css's of a cgroup
384 * @css: the iteration cursor
385 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
386 * @cgrp: the target cgroup to iterate css's of
388 * Should be called under cgroup_[tree_]mutex.
390 #define for_each_css(css, ssid, cgrp) \
391 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
392 if (!((css) = rcu_dereference_check( \
393 (cgrp)->subsys[(ssid)], \
394 lockdep_is_held(&cgroup_mutex)))) { } \
398 * for_each_e_css - iterate all effective css's of a cgroup
399 * @css: the iteration cursor
400 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
401 * @cgrp: the target cgroup to iterate css's of
403 * Should be called under cgroup_[tree_]mutex.
405 #define for_each_e_css(css, ssid, cgrp) \
406 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
407 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
412 * for_each_subsys - iterate all enabled cgroup subsystems
413 * @ss: the iteration cursor
414 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
416 #define for_each_subsys(ss, ssid) \
417 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
418 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
421 * for_each_subsys_which - filter for_each_subsys with a bitmask
422 * @ss: the iteration cursor
423 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
424 * @ss_maskp: a pointer to the bitmask
426 * The block will only run for cases where the ssid-th bit (1 << ssid) of
429 #define for_each_subsys_which(ss, ssid, ss_maskp) \
430 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
433 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
434 if (((ss) = cgroup_subsys[ssid]) && false) \
438 /* iterate across the hierarchies */
439 #define for_each_root(root) \
440 list_for_each_entry((root), &cgroup_roots, root_list)
442 /* iterate over child cgrps, lock should be held throughout iteration */
443 #define cgroup_for_each_live_child(child, cgrp) \
444 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
445 if (({ lockdep_assert_held(&cgroup_mutex); \
446 cgroup_is_dead(child); })) \
450 static void cgroup_release_agent(struct work_struct *work);
451 static void check_for_release(struct cgroup *cgrp);
454 * A cgroup can be associated with multiple css_sets as different tasks may
455 * belong to different cgroups on different hierarchies. In the other
456 * direction, a css_set is naturally associated with multiple cgroups.
457 * This M:N relationship is represented by the following link structure
458 * which exists for each association and allows traversing the associations
461 struct cgrp_cset_link {
462 /* the cgroup and css_set this link associates */
464 struct css_set *cset;
466 /* list of cgrp_cset_links anchored at cgrp->cset_links */
467 struct list_head cset_link;
469 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
470 struct list_head cgrp_link;
474 * The default css_set - used by init and its children prior to any
475 * hierarchies being mounted. It contains a pointer to the root state
476 * for each subsystem. Also used to anchor the list of css_sets. Not
477 * reference-counted, to improve performance when child cgroups
478 * haven't been created.
480 struct css_set init_css_set = {
481 .refcount = ATOMIC_INIT(1),
482 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
483 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
484 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
485 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
486 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
489 static int css_set_count = 1; /* 1 for init_css_set */
492 * cgroup_update_populated - updated populated count of a cgroup
493 * @cgrp: the target cgroup
494 * @populated: inc or dec populated count
496 * @cgrp is either getting the first task (css_set) or losing the last.
497 * Update @cgrp->populated_cnt accordingly. The count is propagated
498 * towards root so that a given cgroup's populated_cnt is zero iff the
499 * cgroup and all its descendants are empty.
501 * @cgrp's interface file "cgroup.populated" is zero if
502 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
503 * changes from or to zero, userland is notified that the content of the
504 * interface file has changed. This can be used to detect when @cgrp and
505 * its descendants become populated or empty.
507 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
509 lockdep_assert_held(&css_set_rwsem);
515 trigger = !cgrp->populated_cnt++;
517 trigger = !--cgrp->populated_cnt;
522 if (cgrp->populated_kn)
523 kernfs_notify(cgrp->populated_kn);
524 cgrp = cgroup_parent(cgrp);
529 * hash table for cgroup groups. This improves the performance to find
530 * an existing css_set. This hash doesn't (currently) take into
531 * account cgroups in empty hierarchies.
533 #define CSS_SET_HASH_BITS 7
534 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
536 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
538 unsigned long key = 0UL;
539 struct cgroup_subsys *ss;
542 for_each_subsys(ss, i)
543 key += (unsigned long)css[i];
544 key = (key >> 16) ^ key;
549 static void put_css_set_locked(struct css_set *cset)
551 struct cgrp_cset_link *link, *tmp_link;
552 struct cgroup_subsys *ss;
555 lockdep_assert_held(&css_set_rwsem);
557 if (!atomic_dec_and_test(&cset->refcount))
560 /* This css_set is dead. unlink it and release cgroup refcounts */
561 for_each_subsys(ss, ssid)
562 list_del(&cset->e_cset_node[ssid]);
563 hash_del(&cset->hlist);
566 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
567 struct cgroup *cgrp = link->cgrp;
569 list_del(&link->cset_link);
570 list_del(&link->cgrp_link);
572 /* @cgrp can't go away while we're holding css_set_rwsem */
573 if (list_empty(&cgrp->cset_links)) {
574 cgroup_update_populated(cgrp, false);
575 check_for_release(cgrp);
581 kfree_rcu(cset, rcu_head);
584 static void put_css_set(struct css_set *cset)
587 * Ensure that the refcount doesn't hit zero while any readers
588 * can see it. Similar to atomic_dec_and_lock(), but for an
591 if (atomic_add_unless(&cset->refcount, -1, 1))
594 down_write(&css_set_rwsem);
595 put_css_set_locked(cset);
596 up_write(&css_set_rwsem);
600 * refcounted get/put for css_set objects
602 static inline void get_css_set(struct css_set *cset)
604 atomic_inc(&cset->refcount);
608 * compare_css_sets - helper function for find_existing_css_set().
609 * @cset: candidate css_set being tested
610 * @old_cset: existing css_set for a task
611 * @new_cgrp: cgroup that's being entered by the task
612 * @template: desired set of css pointers in css_set (pre-calculated)
614 * Returns true if "cset" matches "old_cset" except for the hierarchy
615 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
617 static bool compare_css_sets(struct css_set *cset,
618 struct css_set *old_cset,
619 struct cgroup *new_cgrp,
620 struct cgroup_subsys_state *template[])
622 struct list_head *l1, *l2;
625 * On the default hierarchy, there can be csets which are
626 * associated with the same set of cgroups but different csses.
627 * Let's first ensure that csses match.
629 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
633 * Compare cgroup pointers in order to distinguish between
634 * different cgroups in hierarchies. As different cgroups may
635 * share the same effective css, this comparison is always
638 l1 = &cset->cgrp_links;
639 l2 = &old_cset->cgrp_links;
641 struct cgrp_cset_link *link1, *link2;
642 struct cgroup *cgrp1, *cgrp2;
646 /* See if we reached the end - both lists are equal length. */
647 if (l1 == &cset->cgrp_links) {
648 BUG_ON(l2 != &old_cset->cgrp_links);
651 BUG_ON(l2 == &old_cset->cgrp_links);
653 /* Locate the cgroups associated with these links. */
654 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
655 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
658 /* Hierarchies should be linked in the same order. */
659 BUG_ON(cgrp1->root != cgrp2->root);
662 * If this hierarchy is the hierarchy of the cgroup
663 * that's changing, then we need to check that this
664 * css_set points to the new cgroup; if it's any other
665 * hierarchy, then this css_set should point to the
666 * same cgroup as the old css_set.
668 if (cgrp1->root == new_cgrp->root) {
669 if (cgrp1 != new_cgrp)
680 * find_existing_css_set - init css array and find the matching css_set
681 * @old_cset: the css_set that we're using before the cgroup transition
682 * @cgrp: the cgroup that we're moving into
683 * @template: out param for the new set of csses, should be clear on entry
685 static struct css_set *find_existing_css_set(struct css_set *old_cset,
687 struct cgroup_subsys_state *template[])
689 struct cgroup_root *root = cgrp->root;
690 struct cgroup_subsys *ss;
691 struct css_set *cset;
696 * Build the set of subsystem state objects that we want to see in the
697 * new css_set. while subsystems can change globally, the entries here
698 * won't change, so no need for locking.
700 for_each_subsys(ss, i) {
701 if (root->subsys_mask & (1UL << i)) {
703 * @ss is in this hierarchy, so we want the
704 * effective css from @cgrp.
706 template[i] = cgroup_e_css(cgrp, ss);
709 * @ss is not in this hierarchy, so we don't want
712 template[i] = old_cset->subsys[i];
716 key = css_set_hash(template);
717 hash_for_each_possible(css_set_table, cset, hlist, key) {
718 if (!compare_css_sets(cset, old_cset, cgrp, template))
721 /* This css_set matches what we need */
725 /* No existing cgroup group matched */
729 static void free_cgrp_cset_links(struct list_head *links_to_free)
731 struct cgrp_cset_link *link, *tmp_link;
733 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
734 list_del(&link->cset_link);
740 * allocate_cgrp_cset_links - allocate cgrp_cset_links
741 * @count: the number of links to allocate
742 * @tmp_links: list_head the allocated links are put on
744 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
745 * through ->cset_link. Returns 0 on success or -errno.
747 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
749 struct cgrp_cset_link *link;
752 INIT_LIST_HEAD(tmp_links);
754 for (i = 0; i < count; i++) {
755 link = kzalloc(sizeof(*link), GFP_KERNEL);
757 free_cgrp_cset_links(tmp_links);
760 list_add(&link->cset_link, tmp_links);
766 * link_css_set - a helper function to link a css_set to a cgroup
767 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
768 * @cset: the css_set to be linked
769 * @cgrp: the destination cgroup
771 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
774 struct cgrp_cset_link *link;
776 BUG_ON(list_empty(tmp_links));
778 if (cgroup_on_dfl(cgrp))
779 cset->dfl_cgrp = cgrp;
781 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
785 if (list_empty(&cgrp->cset_links))
786 cgroup_update_populated(cgrp, true);
787 list_move(&link->cset_link, &cgrp->cset_links);
790 * Always add links to the tail of the list so that the list
791 * is sorted by order of hierarchy creation
793 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
797 * find_css_set - return a new css_set with one cgroup updated
798 * @old_cset: the baseline css_set
799 * @cgrp: the cgroup to be updated
801 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
802 * substituted into the appropriate hierarchy.
804 static struct css_set *find_css_set(struct css_set *old_cset,
807 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
808 struct css_set *cset;
809 struct list_head tmp_links;
810 struct cgrp_cset_link *link;
811 struct cgroup_subsys *ss;
815 lockdep_assert_held(&cgroup_mutex);
817 /* First see if we already have a cgroup group that matches
819 down_read(&css_set_rwsem);
820 cset = find_existing_css_set(old_cset, cgrp, template);
823 up_read(&css_set_rwsem);
828 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
832 /* Allocate all the cgrp_cset_link objects that we'll need */
833 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
838 atomic_set(&cset->refcount, 1);
839 INIT_LIST_HEAD(&cset->cgrp_links);
840 INIT_LIST_HEAD(&cset->tasks);
841 INIT_LIST_HEAD(&cset->mg_tasks);
842 INIT_LIST_HEAD(&cset->mg_preload_node);
843 INIT_LIST_HEAD(&cset->mg_node);
844 INIT_HLIST_NODE(&cset->hlist);
846 /* Copy the set of subsystem state objects generated in
847 * find_existing_css_set() */
848 memcpy(cset->subsys, template, sizeof(cset->subsys));
850 down_write(&css_set_rwsem);
851 /* Add reference counts and links from the new css_set. */
852 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
853 struct cgroup *c = link->cgrp;
855 if (c->root == cgrp->root)
857 link_css_set(&tmp_links, cset, c);
860 BUG_ON(!list_empty(&tmp_links));
864 /* Add @cset to the hash table */
865 key = css_set_hash(cset->subsys);
866 hash_add(css_set_table, &cset->hlist, key);
868 for_each_subsys(ss, ssid)
869 list_add_tail(&cset->e_cset_node[ssid],
870 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
872 up_write(&css_set_rwsem);
877 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
879 struct cgroup *root_cgrp = kf_root->kn->priv;
881 return root_cgrp->root;
884 static int cgroup_init_root_id(struct cgroup_root *root)
888 lockdep_assert_held(&cgroup_mutex);
890 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
894 root->hierarchy_id = id;
898 static void cgroup_exit_root_id(struct cgroup_root *root)
900 lockdep_assert_held(&cgroup_mutex);
902 if (root->hierarchy_id) {
903 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
904 root->hierarchy_id = 0;
908 static void cgroup_free_root(struct cgroup_root *root)
911 /* hierarchy ID should already have been released */
912 WARN_ON_ONCE(root->hierarchy_id);
914 idr_destroy(&root->cgroup_idr);
919 static void cgroup_destroy_root(struct cgroup_root *root)
921 struct cgroup *cgrp = &root->cgrp;
922 struct cgrp_cset_link *link, *tmp_link;
924 mutex_lock(&cgroup_mutex);
926 BUG_ON(atomic_read(&root->nr_cgrps));
927 BUG_ON(!list_empty(&cgrp->self.children));
929 /* Rebind all subsystems back to the default hierarchy */
930 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
933 * Release all the links from cset_links to this hierarchy's
936 down_write(&css_set_rwsem);
938 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
939 list_del(&link->cset_link);
940 list_del(&link->cgrp_link);
943 up_write(&css_set_rwsem);
945 if (!list_empty(&root->root_list)) {
946 list_del(&root->root_list);
950 cgroup_exit_root_id(root);
952 mutex_unlock(&cgroup_mutex);
954 kernfs_destroy_root(root->kf_root);
955 cgroup_free_root(root);
958 /* look up cgroup associated with given css_set on the specified hierarchy */
959 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
960 struct cgroup_root *root)
962 struct cgroup *res = NULL;
964 lockdep_assert_held(&cgroup_mutex);
965 lockdep_assert_held(&css_set_rwsem);
967 if (cset == &init_css_set) {
970 struct cgrp_cset_link *link;
972 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
973 struct cgroup *c = link->cgrp;
975 if (c->root == root) {
987 * Return the cgroup for "task" from the given hierarchy. Must be
988 * called with cgroup_mutex and css_set_rwsem held.
990 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
991 struct cgroup_root *root)
994 * No need to lock the task - since we hold cgroup_mutex the
995 * task can't change groups, so the only thing that can happen
996 * is that it exits and its css is set back to init_css_set.
998 return cset_cgroup_from_root(task_css_set(task), root);
1002 * A task must hold cgroup_mutex to modify cgroups.
1004 * Any task can increment and decrement the count field without lock.
1005 * So in general, code holding cgroup_mutex can't rely on the count
1006 * field not changing. However, if the count goes to zero, then only
1007 * cgroup_attach_task() can increment it again. Because a count of zero
1008 * means that no tasks are currently attached, therefore there is no
1009 * way a task attached to that cgroup can fork (the other way to
1010 * increment the count). So code holding cgroup_mutex can safely
1011 * assume that if the count is zero, it will stay zero. Similarly, if
1012 * a task holds cgroup_mutex on a cgroup with zero count, it
1013 * knows that the cgroup won't be removed, as cgroup_rmdir()
1016 * A cgroup can only be deleted if both its 'count' of using tasks
1017 * is zero, and its list of 'children' cgroups is empty. Since all
1018 * tasks in the system use _some_ cgroup, and since there is always at
1019 * least one task in the system (init, pid == 1), therefore, root cgroup
1020 * always has either children cgroups and/or using tasks. So we don't
1021 * need a special hack to ensure that root cgroup cannot be deleted.
1023 * P.S. One more locking exception. RCU is used to guard the
1024 * update of a tasks cgroup pointer by cgroup_attach_task()
1027 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
1028 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1029 static const struct file_operations proc_cgroupstats_operations;
1031 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1034 struct cgroup_subsys *ss = cft->ss;
1036 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1037 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1038 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1039 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1042 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1047 * cgroup_file_mode - deduce file mode of a control file
1048 * @cft: the control file in question
1050 * returns cft->mode if ->mode is not 0
1051 * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
1052 * returns S_IRUGO if it has only a read handler
1053 * returns S_IWUSR if it has only a write hander
1055 static umode_t cgroup_file_mode(const struct cftype *cft)
1062 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1065 if (cft->write_u64 || cft->write_s64 || cft->write)
1071 static void cgroup_get(struct cgroup *cgrp)
1073 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1074 css_get(&cgrp->self);
1077 static bool cgroup_tryget(struct cgroup *cgrp)
1079 return css_tryget(&cgrp->self);
1082 static void cgroup_put(struct cgroup *cgrp)
1084 css_put(&cgrp->self);
1088 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1089 * @cgrp: the target cgroup
1090 * @subtree_control: the new subtree_control mask to consider
1092 * On the default hierarchy, a subsystem may request other subsystems to be
1093 * enabled together through its ->depends_on mask. In such cases, more
1094 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1096 * This function calculates which subsystems need to be enabled if
1097 * @subtree_control is to be applied to @cgrp. The returned mask is always
1098 * a superset of @subtree_control and follows the usual hierarchy rules.
1100 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1101 unsigned long subtree_control)
1103 struct cgroup *parent = cgroup_parent(cgrp);
1104 unsigned long cur_ss_mask = subtree_control;
1105 struct cgroup_subsys *ss;
1108 lockdep_assert_held(&cgroup_mutex);
1110 if (!cgroup_on_dfl(cgrp))
1114 unsigned long new_ss_mask = cur_ss_mask;
1116 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1117 new_ss_mask |= ss->depends_on;
1120 * Mask out subsystems which aren't available. This can
1121 * happen only if some depended-upon subsystems were bound
1122 * to non-default hierarchies.
1125 new_ss_mask &= parent->child_subsys_mask;
1127 new_ss_mask &= cgrp->root->subsys_mask;
1129 if (new_ss_mask == cur_ss_mask)
1131 cur_ss_mask = new_ss_mask;
1138 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1139 * @cgrp: the target cgroup
1141 * Update @cgrp->child_subsys_mask according to the current
1142 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1144 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1146 cgrp->child_subsys_mask =
1147 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1151 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1152 * @kn: the kernfs_node being serviced
1154 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1155 * the method finishes if locking succeeded. Note that once this function
1156 * returns the cgroup returned by cgroup_kn_lock_live() may become
1157 * inaccessible any time. If the caller intends to continue to access the
1158 * cgroup, it should pin it before invoking this function.
1160 static void cgroup_kn_unlock(struct kernfs_node *kn)
1162 struct cgroup *cgrp;
1164 if (kernfs_type(kn) == KERNFS_DIR)
1167 cgrp = kn->parent->priv;
1169 mutex_unlock(&cgroup_mutex);
1171 kernfs_unbreak_active_protection(kn);
1176 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1177 * @kn: the kernfs_node being serviced
1179 * This helper is to be used by a cgroup kernfs method currently servicing
1180 * @kn. It breaks the active protection, performs cgroup locking and
1181 * verifies that the associated cgroup is alive. Returns the cgroup if
1182 * alive; otherwise, %NULL. A successful return should be undone by a
1183 * matching cgroup_kn_unlock() invocation.
1185 * Any cgroup kernfs method implementation which requires locking the
1186 * associated cgroup should use this helper. It avoids nesting cgroup
1187 * locking under kernfs active protection and allows all kernfs operations
1188 * including self-removal.
1190 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1192 struct cgroup *cgrp;
1194 if (kernfs_type(kn) == KERNFS_DIR)
1197 cgrp = kn->parent->priv;
1200 * We're gonna grab cgroup_mutex which nests outside kernfs
1201 * active_ref. cgroup liveliness check alone provides enough
1202 * protection against removal. Ensure @cgrp stays accessible and
1203 * break the active_ref protection.
1205 if (!cgroup_tryget(cgrp))
1207 kernfs_break_active_protection(kn);
1209 mutex_lock(&cgroup_mutex);
1211 if (!cgroup_is_dead(cgrp))
1214 cgroup_kn_unlock(kn);
1218 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1220 char name[CGROUP_FILE_NAME_MAX];
1222 lockdep_assert_held(&cgroup_mutex);
1223 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1227 * cgroup_clear_dir - remove subsys files in a cgroup directory
1228 * @cgrp: target cgroup
1229 * @subsys_mask: mask of the subsystem ids whose files should be removed
1231 static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1233 struct cgroup_subsys *ss;
1236 for_each_subsys(ss, i) {
1237 struct cftype *cfts;
1239 if (!(subsys_mask & (1 << i)))
1241 list_for_each_entry(cfts, &ss->cfts, node)
1242 cgroup_addrm_files(cgrp, cfts, false);
1246 static int rebind_subsystems(struct cgroup_root *dst_root,
1247 unsigned long ss_mask)
1249 struct cgroup_subsys *ss;
1250 unsigned long tmp_ss_mask;
1253 lockdep_assert_held(&cgroup_mutex);
1255 for_each_subsys_which(ss, ssid, &ss_mask) {
1256 /* if @ss has non-root csses attached to it, can't move */
1257 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1260 /* can't move between two non-dummy roots either */
1261 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1265 /* skip creating root files on dfl_root for inhibited subsystems */
1266 tmp_ss_mask = ss_mask;
1267 if (dst_root == &cgrp_dfl_root)
1268 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1270 ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1272 if (dst_root != &cgrp_dfl_root)
1276 * Rebinding back to the default root is not allowed to
1277 * fail. Using both default and non-default roots should
1278 * be rare. Moving subsystems back and forth even more so.
1279 * Just warn about it and continue.
1281 if (cgrp_dfl_root_visible) {
1282 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1284 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1289 * Nothing can fail from this point on. Remove files for the
1290 * removed subsystems and rebind each subsystem.
1292 for_each_subsys_which(ss, ssid, &ss_mask)
1293 cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1295 for_each_subsys_which(ss, ssid, &ss_mask) {
1296 struct cgroup_root *src_root;
1297 struct cgroup_subsys_state *css;
1298 struct css_set *cset;
1300 src_root = ss->root;
1301 css = cgroup_css(&src_root->cgrp, ss);
1303 WARN_ON(!css || cgroup_css(&dst_root->cgrp, ss));
1305 RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1306 rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1307 ss->root = dst_root;
1308 css->cgroup = &dst_root->cgrp;
1310 down_write(&css_set_rwsem);
1311 hash_for_each(css_set_table, i, cset, hlist)
1312 list_move_tail(&cset->e_cset_node[ss->id],
1313 &dst_root->cgrp.e_csets[ss->id]);
1314 up_write(&css_set_rwsem);
1316 src_root->subsys_mask &= ~(1 << ssid);
1317 src_root->cgrp.subtree_control &= ~(1 << ssid);
1318 cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1320 /* default hierarchy doesn't enable controllers by default */
1321 dst_root->subsys_mask |= 1 << ssid;
1322 if (dst_root != &cgrp_dfl_root) {
1323 dst_root->cgrp.subtree_control |= 1 << ssid;
1324 cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1331 kernfs_activate(dst_root->cgrp.kn);
1335 static int cgroup_show_options(struct seq_file *seq,
1336 struct kernfs_root *kf_root)
1338 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1339 struct cgroup_subsys *ss;
1342 if (root != &cgrp_dfl_root)
1343 for_each_subsys(ss, ssid)
1344 if (root->subsys_mask & (1 << ssid))
1345 seq_show_option(seq, ss->legacy_name, NULL);
1346 if (root->flags & CGRP_ROOT_NOPREFIX)
1347 seq_puts(seq, ",noprefix");
1348 if (root->flags & CGRP_ROOT_XATTR)
1349 seq_puts(seq, ",xattr");
1351 spin_lock(&release_agent_path_lock);
1352 if (strlen(root->release_agent_path))
1353 seq_show_option(seq, "release_agent",
1354 root->release_agent_path);
1355 spin_unlock(&release_agent_path_lock);
1357 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1358 seq_puts(seq, ",clone_children");
1359 if (strlen(root->name))
1360 seq_show_option(seq, "name", root->name);
1364 struct cgroup_sb_opts {
1365 unsigned long subsys_mask;
1367 char *release_agent;
1368 bool cpuset_clone_children;
1370 /* User explicitly requested empty subsystem */
1374 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1376 char *token, *o = data;
1377 bool all_ss = false, one_ss = false;
1378 unsigned long mask = -1UL;
1379 struct cgroup_subsys *ss;
1383 #ifdef CONFIG_CPUSETS
1384 mask = ~(1U << cpuset_cgrp_id);
1387 memset(opts, 0, sizeof(*opts));
1389 while ((token = strsep(&o, ",")) != NULL) {
1394 if (!strcmp(token, "none")) {
1395 /* Explicitly have no subsystems */
1399 if (!strcmp(token, "all")) {
1400 /* Mutually exclusive option 'all' + subsystem name */
1406 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1407 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1410 if (!strcmp(token, "noprefix")) {
1411 opts->flags |= CGRP_ROOT_NOPREFIX;
1414 if (!strcmp(token, "clone_children")) {
1415 opts->cpuset_clone_children = true;
1418 if (!strcmp(token, "xattr")) {
1419 opts->flags |= CGRP_ROOT_XATTR;
1422 if (!strncmp(token, "release_agent=", 14)) {
1423 /* Specifying two release agents is forbidden */
1424 if (opts->release_agent)
1426 opts->release_agent =
1427 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1428 if (!opts->release_agent)
1432 if (!strncmp(token, "name=", 5)) {
1433 const char *name = token + 5;
1434 /* Can't specify an empty name */
1437 /* Must match [\w.-]+ */
1438 for (i = 0; i < strlen(name); i++) {
1442 if ((c == '.') || (c == '-') || (c == '_'))
1446 /* Specifying two names is forbidden */
1449 opts->name = kstrndup(name,
1450 MAX_CGROUP_ROOT_NAMELEN - 1,
1458 for_each_subsys(ss, i) {
1459 if (strcmp(token, ss->legacy_name))
1464 /* Mutually exclusive option 'all' + subsystem name */
1467 opts->subsys_mask |= (1 << i);
1472 if (i == CGROUP_SUBSYS_COUNT)
1476 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1477 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1479 pr_err("sane_behavior: no other mount options allowed\n");
1486 * If the 'all' option was specified select all the subsystems,
1487 * otherwise if 'none', 'name=' and a subsystem name options were
1488 * not specified, let's default to 'all'
1490 if (all_ss || (!one_ss && !opts->none && !opts->name))
1491 for_each_subsys(ss, i)
1493 opts->subsys_mask |= (1 << i);
1496 * We either have to specify by name or by subsystems. (So all
1497 * empty hierarchies must have a name).
1499 if (!opts->subsys_mask && !opts->name)
1503 * Option noprefix was introduced just for backward compatibility
1504 * with the old cpuset, so we allow noprefix only if mounting just
1505 * the cpuset subsystem.
1507 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1510 /* Can't specify "none" and some subsystems */
1511 if (opts->subsys_mask && opts->none)
1517 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1520 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1521 struct cgroup_sb_opts opts;
1522 unsigned long added_mask, removed_mask;
1524 if (root == &cgrp_dfl_root) {
1525 pr_err("remount is not allowed\n");
1529 mutex_lock(&cgroup_mutex);
1531 /* See what subsystems are wanted */
1532 ret = parse_cgroupfs_options(data, &opts);
1536 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1537 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1538 task_tgid_nr(current), current->comm);
1540 added_mask = opts.subsys_mask & ~root->subsys_mask;
1541 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1543 /* Don't allow flags or name to change at remount */
1544 if ((opts.flags ^ root->flags) ||
1545 (opts.name && strcmp(opts.name, root->name))) {
1546 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1547 opts.flags, opts.name ?: "", root->flags, root->name);
1552 /* remounting is not allowed for populated hierarchies */
1553 if (!list_empty(&root->cgrp.self.children)) {
1558 ret = rebind_subsystems(root, added_mask);
1562 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1564 if (opts.release_agent) {
1565 spin_lock(&release_agent_path_lock);
1566 strcpy(root->release_agent_path, opts.release_agent);
1567 spin_unlock(&release_agent_path_lock);
1570 kfree(opts.release_agent);
1572 mutex_unlock(&cgroup_mutex);
1577 * To reduce the fork() overhead for systems that are not actually using
1578 * their cgroups capability, we don't maintain the lists running through
1579 * each css_set to its tasks until we see the list actually used - in other
1580 * words after the first mount.
1582 static bool use_task_css_set_links __read_mostly;
1584 static void cgroup_enable_task_cg_lists(void)
1586 struct task_struct *p, *g;
1588 down_write(&css_set_rwsem);
1590 if (use_task_css_set_links)
1593 use_task_css_set_links = true;
1596 * We need tasklist_lock because RCU is not safe against
1597 * while_each_thread(). Besides, a forking task that has passed
1598 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1599 * is not guaranteed to have its child immediately visible in the
1600 * tasklist if we walk through it with RCU.
1602 read_lock(&tasklist_lock);
1603 do_each_thread(g, p) {
1604 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1605 task_css_set(p) != &init_css_set);
1608 * We should check if the process is exiting, otherwise
1609 * it will race with cgroup_exit() in that the list
1610 * entry won't be deleted though the process has exited.
1611 * Do it while holding siglock so that we don't end up
1612 * racing against cgroup_exit().
1614 spin_lock_irq(&p->sighand->siglock);
1615 if (!(p->flags & PF_EXITING)) {
1616 struct css_set *cset = task_css_set(p);
1618 list_add(&p->cg_list, &cset->tasks);
1621 spin_unlock_irq(&p->sighand->siglock);
1622 } while_each_thread(g, p);
1623 read_unlock(&tasklist_lock);
1625 up_write(&css_set_rwsem);
1628 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1630 struct cgroup_subsys *ss;
1633 INIT_LIST_HEAD(&cgrp->self.sibling);
1634 INIT_LIST_HEAD(&cgrp->self.children);
1635 INIT_LIST_HEAD(&cgrp->cset_links);
1636 INIT_LIST_HEAD(&cgrp->pidlists);
1637 mutex_init(&cgrp->pidlist_mutex);
1638 cgrp->self.cgroup = cgrp;
1639 cgrp->self.flags |= CSS_ONLINE;
1641 for_each_subsys(ss, ssid)
1642 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1644 init_waitqueue_head(&cgrp->offline_waitq);
1645 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1648 static void init_cgroup_root(struct cgroup_root *root,
1649 struct cgroup_sb_opts *opts)
1651 struct cgroup *cgrp = &root->cgrp;
1653 INIT_LIST_HEAD(&root->root_list);
1654 atomic_set(&root->nr_cgrps, 1);
1656 init_cgroup_housekeeping(cgrp);
1657 idr_init(&root->cgroup_idr);
1659 root->flags = opts->flags;
1660 if (opts->release_agent)
1661 strcpy(root->release_agent_path, opts->release_agent);
1663 strcpy(root->name, opts->name);
1664 if (opts->cpuset_clone_children)
1665 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1668 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1670 LIST_HEAD(tmp_links);
1671 struct cgroup *root_cgrp = &root->cgrp;
1672 struct cftype *base_files;
1673 struct css_set *cset;
1676 lockdep_assert_held(&cgroup_mutex);
1678 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1681 root_cgrp->id = ret;
1683 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1689 * We're accessing css_set_count without locking css_set_rwsem here,
1690 * but that's OK - it can only be increased by someone holding
1691 * cgroup_lock, and that's us. The worst that can happen is that we
1692 * have some link structures left over
1694 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1698 ret = cgroup_init_root_id(root);
1702 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1703 KERNFS_ROOT_CREATE_DEACTIVATED,
1705 if (IS_ERR(root->kf_root)) {
1706 ret = PTR_ERR(root->kf_root);
1709 root_cgrp->kn = root->kf_root->kn;
1711 if (root == &cgrp_dfl_root)
1712 base_files = cgroup_dfl_base_files;
1714 base_files = cgroup_legacy_base_files;
1716 ret = cgroup_addrm_files(root_cgrp, base_files, true);
1720 ret = rebind_subsystems(root, ss_mask);
1725 * There must be no failure case after here, since rebinding takes
1726 * care of subsystems' refcounts, which are explicitly dropped in
1727 * the failure exit path.
1729 list_add(&root->root_list, &cgroup_roots);
1730 cgroup_root_count++;
1733 * Link the root cgroup in this hierarchy into all the css_set
1736 down_write(&css_set_rwsem);
1737 hash_for_each(css_set_table, i, cset, hlist)
1738 link_css_set(&tmp_links, cset, root_cgrp);
1739 up_write(&css_set_rwsem);
1741 BUG_ON(!list_empty(&root_cgrp->self.children));
1742 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1744 kernfs_activate(root_cgrp->kn);
1749 kernfs_destroy_root(root->kf_root);
1750 root->kf_root = NULL;
1752 cgroup_exit_root_id(root);
1754 percpu_ref_exit(&root_cgrp->self.refcnt);
1756 free_cgrp_cset_links(&tmp_links);
1760 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1761 int flags, const char *unused_dev_name,
1764 struct super_block *pinned_sb = NULL;
1765 struct cgroup_subsys *ss;
1766 struct cgroup_root *root;
1767 struct cgroup_sb_opts opts;
1768 struct dentry *dentry;
1774 * The first time anyone tries to mount a cgroup, enable the list
1775 * linking each css_set to its tasks and fix up all existing tasks.
1777 if (!use_task_css_set_links)
1778 cgroup_enable_task_cg_lists();
1780 mutex_lock(&cgroup_mutex);
1782 /* First find the desired set of subsystems */
1783 ret = parse_cgroupfs_options(data, &opts);
1787 /* look for a matching existing root */
1788 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1789 cgrp_dfl_root_visible = true;
1790 root = &cgrp_dfl_root;
1791 cgroup_get(&root->cgrp);
1797 * Destruction of cgroup root is asynchronous, so subsystems may
1798 * still be dying after the previous unmount. Let's drain the
1799 * dying subsystems. We just need to ensure that the ones
1800 * unmounted previously finish dying and don't care about new ones
1801 * starting. Testing ref liveliness is good enough.
1803 for_each_subsys(ss, i) {
1804 if (!(opts.subsys_mask & (1 << i)) ||
1805 ss->root == &cgrp_dfl_root)
1808 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1809 mutex_unlock(&cgroup_mutex);
1811 ret = restart_syscall();
1814 cgroup_put(&ss->root->cgrp);
1817 for_each_root(root) {
1818 bool name_match = false;
1820 if (root == &cgrp_dfl_root)
1824 * If we asked for a name then it must match. Also, if
1825 * name matches but sybsys_mask doesn't, we should fail.
1826 * Remember whether name matched.
1829 if (strcmp(opts.name, root->name))
1835 * If we asked for subsystems (or explicitly for no
1836 * subsystems) then they must match.
1838 if ((opts.subsys_mask || opts.none) &&
1839 (opts.subsys_mask != root->subsys_mask)) {
1846 if (root->flags ^ opts.flags)
1847 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1850 * We want to reuse @root whose lifetime is governed by its
1851 * ->cgrp. Let's check whether @root is alive and keep it
1852 * that way. As cgroup_kill_sb() can happen anytime, we
1853 * want to block it by pinning the sb so that @root doesn't
1854 * get killed before mount is complete.
1856 * With the sb pinned, tryget_live can reliably indicate
1857 * whether @root can be reused. If it's being killed,
1858 * drain it. We can use wait_queue for the wait but this
1859 * path is super cold. Let's just sleep a bit and retry.
1861 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1862 if (IS_ERR(pinned_sb) ||
1863 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1864 mutex_unlock(&cgroup_mutex);
1865 if (!IS_ERR_OR_NULL(pinned_sb))
1866 deactivate_super(pinned_sb);
1868 ret = restart_syscall();
1877 * No such thing, create a new one. name= matching without subsys
1878 * specification is allowed for already existing hierarchies but we
1879 * can't create new one without subsys specification.
1881 if (!opts.subsys_mask && !opts.none) {
1886 root = kzalloc(sizeof(*root), GFP_KERNEL);
1892 init_cgroup_root(root, &opts);
1894 ret = cgroup_setup_root(root, opts.subsys_mask);
1896 cgroup_free_root(root);
1899 mutex_unlock(&cgroup_mutex);
1901 kfree(opts.release_agent);
1905 return ERR_PTR(ret);
1907 dentry = kernfs_mount(fs_type, flags, root->kf_root,
1908 CGROUP_SUPER_MAGIC, &new_sb);
1909 if (IS_ERR(dentry) || !new_sb)
1910 cgroup_put(&root->cgrp);
1913 * If @pinned_sb, we're reusing an existing root and holding an
1914 * extra ref on its sb. Mount is complete. Put the extra ref.
1918 deactivate_super(pinned_sb);
1924 static void cgroup_kill_sb(struct super_block *sb)
1926 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1927 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1930 * If @root doesn't have any mounts or children, start killing it.
1931 * This prevents new mounts by disabling percpu_ref_tryget_live().
1932 * cgroup_mount() may wait for @root's release.
1934 * And don't kill the default root.
1936 if (!list_empty(&root->cgrp.self.children) ||
1937 root == &cgrp_dfl_root)
1938 cgroup_put(&root->cgrp);
1940 percpu_ref_kill(&root->cgrp.self.refcnt);
1945 static struct file_system_type cgroup_fs_type = {
1947 .mount = cgroup_mount,
1948 .kill_sb = cgroup_kill_sb,
1952 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1953 * @task: target task
1954 * @buf: the buffer to write the path into
1955 * @buflen: the length of the buffer
1957 * Determine @task's cgroup on the first (the one with the lowest non-zero
1958 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1959 * function grabs cgroup_mutex and shouldn't be used inside locks used by
1960 * cgroup controller callbacks.
1962 * Return value is the same as kernfs_path().
1964 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
1966 struct cgroup_root *root;
1967 struct cgroup *cgrp;
1968 int hierarchy_id = 1;
1971 mutex_lock(&cgroup_mutex);
1972 down_read(&css_set_rwsem);
1974 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1977 cgrp = task_cgroup_from_root(task, root);
1978 path = cgroup_path(cgrp, buf, buflen);
1980 /* if no hierarchy exists, everyone is in "/" */
1981 if (strlcpy(buf, "/", buflen) < buflen)
1985 up_read(&css_set_rwsem);
1986 mutex_unlock(&cgroup_mutex);
1989 EXPORT_SYMBOL_GPL(task_cgroup_path);
1991 /* used to track tasks and other necessary states during migration */
1992 struct cgroup_taskset {
1993 /* the src and dst cset list running through cset->mg_node */
1994 struct list_head src_csets;
1995 struct list_head dst_csets;
1998 * Fields for cgroup_taskset_*() iteration.
2000 * Before migration is committed, the target migration tasks are on
2001 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2002 * the csets on ->dst_csets. ->csets point to either ->src_csets
2003 * or ->dst_csets depending on whether migration is committed.
2005 * ->cur_csets and ->cur_task point to the current task position
2008 struct list_head *csets;
2009 struct css_set *cur_cset;
2010 struct task_struct *cur_task;
2014 * cgroup_taskset_first - reset taskset and return the first task
2015 * @tset: taskset of interest
2017 * @tset iteration is initialized and the first task is returned.
2019 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2021 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2022 tset->cur_task = NULL;
2024 return cgroup_taskset_next(tset);
2028 * cgroup_taskset_next - iterate to the next task in taskset
2029 * @tset: taskset of interest
2031 * Return the next task in @tset. Iteration must have been initialized
2032 * with cgroup_taskset_first().
2034 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2036 struct css_set *cset = tset->cur_cset;
2037 struct task_struct *task = tset->cur_task;
2039 while (&cset->mg_node != tset->csets) {
2041 task = list_first_entry(&cset->mg_tasks,
2042 struct task_struct, cg_list);
2044 task = list_next_entry(task, cg_list);
2046 if (&task->cg_list != &cset->mg_tasks) {
2047 tset->cur_cset = cset;
2048 tset->cur_task = task;
2052 cset = list_next_entry(cset, mg_node);
2060 * cgroup_task_migrate - move a task from one cgroup to another.
2061 * @old_cgrp: the cgroup @tsk is being migrated from
2062 * @tsk: the task being migrated
2063 * @new_cset: the new css_set @tsk is being attached to
2065 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2067 static void cgroup_task_migrate(struct cgroup *old_cgrp,
2068 struct task_struct *tsk,
2069 struct css_set *new_cset)
2071 struct css_set *old_cset;
2073 lockdep_assert_held(&cgroup_mutex);
2074 lockdep_assert_held(&css_set_rwsem);
2077 * We are synchronized through cgroup_threadgroup_rwsem against
2078 * PF_EXITING setting such that we can't race against cgroup_exit()
2079 * changing the css_set to init_css_set and dropping the old one.
2081 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2082 old_cset = task_css_set(tsk);
2084 get_css_set(new_cset);
2085 rcu_assign_pointer(tsk->cgroups, new_cset);
2088 * Use move_tail so that cgroup_taskset_first() still returns the
2089 * leader after migration. This works because cgroup_migrate()
2090 * ensures that the dst_cset of the leader is the first on the
2091 * tset's dst_csets list.
2093 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2096 * We just gained a reference on old_cset by taking it from the
2097 * task. As trading it for new_cset is protected by cgroup_mutex,
2098 * we're safe to drop it here; it will be freed under RCU.
2100 put_css_set_locked(old_cset);
2104 * cgroup_migrate_finish - cleanup after attach
2105 * @preloaded_csets: list of preloaded css_sets
2107 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2108 * those functions for details.
2110 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2112 struct css_set *cset, *tmp_cset;
2114 lockdep_assert_held(&cgroup_mutex);
2116 down_write(&css_set_rwsem);
2117 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2118 cset->mg_src_cgrp = NULL;
2119 cset->mg_dst_cset = NULL;
2120 list_del_init(&cset->mg_preload_node);
2121 put_css_set_locked(cset);
2123 up_write(&css_set_rwsem);
2127 * cgroup_migrate_add_src - add a migration source css_set
2128 * @src_cset: the source css_set to add
2129 * @dst_cgrp: the destination cgroup
2130 * @preloaded_csets: list of preloaded css_sets
2132 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2133 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2134 * up by cgroup_migrate_finish().
2136 * This function may be called without holding cgroup_threadgroup_rwsem
2137 * even if the target is a process. Threads may be created and destroyed
2138 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2139 * into play and the preloaded css_sets are guaranteed to cover all
2142 static void cgroup_migrate_add_src(struct css_set *src_cset,
2143 struct cgroup *dst_cgrp,
2144 struct list_head *preloaded_csets)
2146 struct cgroup *src_cgrp;
2148 lockdep_assert_held(&cgroup_mutex);
2149 lockdep_assert_held(&css_set_rwsem);
2151 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2153 if (!list_empty(&src_cset->mg_preload_node))
2156 WARN_ON(src_cset->mg_src_cgrp);
2157 WARN_ON(!list_empty(&src_cset->mg_tasks));
2158 WARN_ON(!list_empty(&src_cset->mg_node));
2160 src_cset->mg_src_cgrp = src_cgrp;
2161 get_css_set(src_cset);
2162 list_add(&src_cset->mg_preload_node, preloaded_csets);
2166 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2167 * @dst_cgrp: the destination cgroup (may be %NULL)
2168 * @preloaded_csets: list of preloaded source css_sets
2170 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2171 * have been preloaded to @preloaded_csets. This function looks up and
2172 * pins all destination css_sets, links each to its source, and append them
2173 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2174 * source css_set is assumed to be its cgroup on the default hierarchy.
2176 * This function must be called after cgroup_migrate_add_src() has been
2177 * called on each migration source css_set. After migration is performed
2178 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2181 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2182 struct list_head *preloaded_csets)
2185 struct css_set *src_cset, *tmp_cset;
2187 lockdep_assert_held(&cgroup_mutex);
2190 * Except for the root, child_subsys_mask must be zero for a cgroup
2191 * with tasks so that child cgroups don't compete against tasks.
2193 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2194 dst_cgrp->child_subsys_mask)
2197 /* look up the dst cset for each src cset and link it to src */
2198 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2199 struct css_set *dst_cset;
2201 dst_cset = find_css_set(src_cset,
2202 dst_cgrp ?: src_cset->dfl_cgrp);
2206 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2209 * If src cset equals dst, it's noop. Drop the src.
2210 * cgroup_migrate() will skip the cset too. Note that we
2211 * can't handle src == dst as some nodes are used by both.
2213 if (src_cset == dst_cset) {
2214 src_cset->mg_src_cgrp = NULL;
2215 list_del_init(&src_cset->mg_preload_node);
2216 put_css_set(src_cset);
2217 put_css_set(dst_cset);
2221 src_cset->mg_dst_cset = dst_cset;
2223 if (list_empty(&dst_cset->mg_preload_node))
2224 list_add(&dst_cset->mg_preload_node, &csets);
2226 put_css_set(dst_cset);
2229 list_splice_tail(&csets, preloaded_csets);
2232 cgroup_migrate_finish(&csets);
2237 * cgroup_migrate - migrate a process or task to a cgroup
2238 * @cgrp: the destination cgroup
2239 * @leader: the leader of the process or the task to migrate
2240 * @threadgroup: whether @leader points to the whole process or a single task
2242 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2243 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2244 * caller is also responsible for invoking cgroup_migrate_add_src() and
2245 * cgroup_migrate_prepare_dst() on the targets before invoking this
2246 * function and following up with cgroup_migrate_finish().
2248 * As long as a controller's ->can_attach() doesn't fail, this function is
2249 * guaranteed to succeed. This means that, excluding ->can_attach()
2250 * failure, when migrating multiple targets, the success or failure can be
2251 * decided for all targets by invoking group_migrate_prepare_dst() before
2252 * actually starting migrating.
2254 static int cgroup_migrate(struct cgroup *cgrp, struct task_struct *leader,
2257 struct cgroup_taskset tset = {
2258 .src_csets = LIST_HEAD_INIT(tset.src_csets),
2259 .dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2260 .csets = &tset.src_csets,
2262 struct cgroup_subsys_state *css, *failed_css = NULL;
2263 struct css_set *cset, *tmp_cset;
2264 struct task_struct *task, *tmp_task;
2268 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2269 * already PF_EXITING could be freed from underneath us unless we
2270 * take an rcu_read_lock.
2272 down_write(&css_set_rwsem);
2276 /* @task either already exited or can't exit until the end */
2277 if (task->flags & PF_EXITING)
2280 /* leave @task alone if post_fork() hasn't linked it yet */
2281 if (list_empty(&task->cg_list))
2284 cset = task_css_set(task);
2285 if (!cset->mg_src_cgrp)
2289 * cgroup_taskset_first() must always return the leader.
2290 * Take care to avoid disturbing the ordering.
2292 list_move_tail(&task->cg_list, &cset->mg_tasks);
2293 if (list_empty(&cset->mg_node))
2294 list_add_tail(&cset->mg_node, &tset.src_csets);
2295 if (list_empty(&cset->mg_dst_cset->mg_node))
2296 list_move_tail(&cset->mg_dst_cset->mg_node,
2301 } while_each_thread(leader, task);
2303 up_write(&css_set_rwsem);
2305 /* methods shouldn't be called if no task is actually migrating */
2306 if (list_empty(&tset.src_csets))
2309 /* check that we can legitimately attach to the cgroup */
2310 for_each_e_css(css, i, cgrp) {
2311 if (css->ss->can_attach) {
2312 ret = css->ss->can_attach(css, &tset);
2315 goto out_cancel_attach;
2321 * Now that we're guaranteed success, proceed to move all tasks to
2322 * the new cgroup. There are no failure cases after here, so this
2323 * is the commit point.
2325 down_write(&css_set_rwsem);
2326 list_for_each_entry(cset, &tset.src_csets, mg_node) {
2327 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2328 cgroup_task_migrate(cset->mg_src_cgrp, task,
2331 up_write(&css_set_rwsem);
2334 * Migration is committed, all target tasks are now on dst_csets.
2335 * Nothing is sensitive to fork() after this point. Notify
2336 * controllers that migration is complete.
2338 tset.csets = &tset.dst_csets;
2340 for_each_e_css(css, i, cgrp)
2341 if (css->ss->attach)
2342 css->ss->attach(css, &tset);
2345 goto out_release_tset;
2348 for_each_e_css(css, i, cgrp) {
2349 if (css == failed_css)
2351 if (css->ss->cancel_attach)
2352 css->ss->cancel_attach(css, &tset);
2355 down_write(&css_set_rwsem);
2356 list_splice_init(&tset.dst_csets, &tset.src_csets);
2357 list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2358 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2359 list_del_init(&cset->mg_node);
2361 up_write(&css_set_rwsem);
2366 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2367 * @dst_cgrp: the cgroup to attach to
2368 * @leader: the task or the leader of the threadgroup to be attached
2369 * @threadgroup: attach the whole threadgroup?
2371 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2373 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2374 struct task_struct *leader, bool threadgroup)
2376 LIST_HEAD(preloaded_csets);
2377 struct task_struct *task;
2380 /* look up all src csets */
2381 down_read(&css_set_rwsem);
2385 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2389 } while_each_thread(leader, task);
2391 up_read(&css_set_rwsem);
2393 /* prepare dst csets and commit */
2394 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2396 ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2398 cgroup_migrate_finish(&preloaded_csets);
2402 static int cgroup_procs_write_permission(struct task_struct *task,
2403 struct cgroup *dst_cgrp,
2404 struct kernfs_open_file *of)
2406 const struct cred *cred = current_cred();
2407 const struct cred *tcred = get_task_cred(task);
2411 * even if we're attaching all tasks in the thread group, we only
2412 * need to check permissions on one of them.
2414 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2415 !uid_eq(cred->euid, tcred->uid) &&
2416 !uid_eq(cred->euid, tcred->suid))
2419 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2420 struct super_block *sb = of->file->f_path.dentry->d_sb;
2421 struct cgroup *cgrp;
2422 struct inode *inode;
2424 down_read(&css_set_rwsem);
2425 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2426 up_read(&css_set_rwsem);
2428 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2429 cgrp = cgroup_parent(cgrp);
2432 inode = kernfs_get_inode(sb, cgrp->procs_kn);
2434 ret = inode_permission(inode, MAY_WRITE);
2444 * Find the task_struct of the task to attach by vpid and pass it along to the
2445 * function to attach either it or all tasks in its threadgroup. Will lock
2446 * cgroup_mutex and threadgroup.
2448 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2449 size_t nbytes, loff_t off, bool threadgroup)
2451 struct task_struct *tsk;
2452 struct cgroup *cgrp;
2456 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2459 cgrp = cgroup_kn_lock_live(of->kn);
2466 tsk = find_task_by_vpid(pid);
2470 goto out_unlock_cgroup;
2477 tsk = tsk->group_leader;
2480 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2481 * trapped in a cpuset, or RT worker may be born in a cgroup
2482 * with no rt_runtime allocated. Just say no.
2484 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2487 goto out_unlock_cgroup;
2490 get_task_struct(tsk);
2493 percpu_down_write(&cgroup_threadgroup_rwsem);
2495 if (!thread_group_leader(tsk)) {
2497 * a race with de_thread from another thread's exec()
2498 * may strip us of our leadership, if this happens,
2499 * there is no choice but to throw this task away and
2500 * try again; this is
2501 * "double-double-toil-and-trouble-check locking".
2503 percpu_up_write(&cgroup_threadgroup_rwsem);
2504 put_task_struct(tsk);
2505 goto retry_find_task;
2509 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2511 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2513 percpu_up_write(&cgroup_threadgroup_rwsem);
2515 put_task_struct(tsk);
2517 cgroup_kn_unlock(of->kn);
2518 return ret ?: nbytes;
2522 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2523 * @from: attach to all cgroups of a given task
2524 * @tsk: the task to be attached
2526 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2528 struct cgroup_root *root;
2531 mutex_lock(&cgroup_mutex);
2532 for_each_root(root) {
2533 struct cgroup *from_cgrp;
2535 if (root == &cgrp_dfl_root)
2538 down_read(&css_set_rwsem);
2539 from_cgrp = task_cgroup_from_root(from, root);
2540 up_read(&css_set_rwsem);
2542 retval = cgroup_attach_task(from_cgrp, tsk, false);
2546 mutex_unlock(&cgroup_mutex);
2550 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2552 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2553 char *buf, size_t nbytes, loff_t off)
2555 return __cgroup_procs_write(of, buf, nbytes, off, false);
2558 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2559 char *buf, size_t nbytes, loff_t off)
2561 return __cgroup_procs_write(of, buf, nbytes, off, true);
2564 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2565 char *buf, size_t nbytes, loff_t off)
2567 struct cgroup *cgrp;
2569 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2571 cgrp = cgroup_kn_lock_live(of->kn);
2574 spin_lock(&release_agent_path_lock);
2575 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2576 sizeof(cgrp->root->release_agent_path));
2577 spin_unlock(&release_agent_path_lock);
2578 cgroup_kn_unlock(of->kn);
2582 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2584 struct cgroup *cgrp = seq_css(seq)->cgroup;
2586 spin_lock(&release_agent_path_lock);
2587 seq_puts(seq, cgrp->root->release_agent_path);
2588 spin_unlock(&release_agent_path_lock);
2589 seq_putc(seq, '\n');
2593 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2595 seq_puts(seq, "0\n");
2599 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2601 struct cgroup_subsys *ss;
2602 bool printed = false;
2605 for_each_subsys_which(ss, ssid, &ss_mask) {
2608 seq_printf(seq, "%s", ss->name);
2612 seq_putc(seq, '\n');
2615 /* show controllers which are currently attached to the default hierarchy */
2616 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2618 struct cgroup *cgrp = seq_css(seq)->cgroup;
2620 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2621 ~cgrp_dfl_root_inhibit_ss_mask);
2625 /* show controllers which are enabled from the parent */
2626 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2628 struct cgroup *cgrp = seq_css(seq)->cgroup;
2630 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2634 /* show controllers which are enabled for a given cgroup's children */
2635 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2637 struct cgroup *cgrp = seq_css(seq)->cgroup;
2639 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2644 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2645 * @cgrp: root of the subtree to update csses for
2647 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2648 * css associations need to be updated accordingly. This function looks up
2649 * all css_sets which are attached to the subtree, creates the matching
2650 * updated css_sets and migrates the tasks to the new ones.
2652 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2654 LIST_HEAD(preloaded_csets);
2655 struct cgroup_subsys_state *css;
2656 struct css_set *src_cset;
2659 lockdep_assert_held(&cgroup_mutex);
2661 /* look up all csses currently attached to @cgrp's subtree */
2662 down_read(&css_set_rwsem);
2663 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2664 struct cgrp_cset_link *link;
2666 /* self is not affected by child_subsys_mask change */
2667 if (css->cgroup == cgrp)
2670 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2671 cgroup_migrate_add_src(link->cset, cgrp,
2674 up_read(&css_set_rwsem);
2676 /* NULL dst indicates self on default hierarchy */
2677 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2681 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2682 struct task_struct *last_task = NULL, *task;
2684 /* src_csets precede dst_csets, break on the first dst_cset */
2685 if (!src_cset->mg_src_cgrp)
2689 * All tasks in src_cset need to be migrated to the
2690 * matching dst_cset. Empty it process by process. We
2691 * walk tasks but migrate processes. The leader might even
2692 * belong to a different cset but such src_cset would also
2693 * be among the target src_csets because the default
2694 * hierarchy enforces per-process membership.
2697 down_read(&css_set_rwsem);
2698 task = list_first_entry_or_null(&src_cset->tasks,
2699 struct task_struct, cg_list);
2701 task = task->group_leader;
2702 WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2703 get_task_struct(task);
2705 up_read(&css_set_rwsem);
2710 /* guard against possible infinite loop */
2711 if (WARN(last_task == task,
2712 "cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2716 percpu_down_write(&cgroup_threadgroup_rwsem);
2717 /* raced against de_thread() from another thread? */
2718 if (!thread_group_leader(task)) {
2719 percpu_up_write(&cgroup_threadgroup_rwsem);
2720 put_task_struct(task);
2724 ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2726 percpu_up_write(&cgroup_threadgroup_rwsem);
2727 put_task_struct(task);
2729 if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2735 cgroup_migrate_finish(&preloaded_csets);
2739 /* change the enabled child controllers for a cgroup in the default hierarchy */
2740 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2741 char *buf, size_t nbytes,
2744 unsigned long enable = 0, disable = 0;
2745 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2746 struct cgroup *cgrp, *child;
2747 struct cgroup_subsys *ss;
2752 * Parse input - space separated list of subsystem names prefixed
2753 * with either + or -.
2755 buf = strstrip(buf);
2756 while ((tok = strsep(&buf, " "))) {
2757 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2761 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2762 if (ss->disabled || strcmp(tok + 1, ss->name))
2766 enable |= 1 << ssid;
2767 disable &= ~(1 << ssid);
2768 } else if (*tok == '-') {
2769 disable |= 1 << ssid;
2770 enable &= ~(1 << ssid);
2776 if (ssid == CGROUP_SUBSYS_COUNT)
2780 cgrp = cgroup_kn_lock_live(of->kn);
2784 for_each_subsys(ss, ssid) {
2785 if (enable & (1 << ssid)) {
2786 if (cgrp->subtree_control & (1 << ssid)) {
2787 enable &= ~(1 << ssid);
2791 /* unavailable or not enabled on the parent? */
2792 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2793 (cgroup_parent(cgrp) &&
2794 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2798 } else if (disable & (1 << ssid)) {
2799 if (!(cgrp->subtree_control & (1 << ssid))) {
2800 disable &= ~(1 << ssid);
2804 /* a child has it enabled? */
2805 cgroup_for_each_live_child(child, cgrp) {
2806 if (child->subtree_control & (1 << ssid)) {
2814 if (!enable && !disable) {
2820 * Except for the root, subtree_control must be zero for a cgroup
2821 * with tasks so that child cgroups don't compete against tasks.
2823 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2829 * Update subsys masks and calculate what needs to be done. More
2830 * subsystems than specified may need to be enabled or disabled
2831 * depending on subsystem dependencies.
2833 old_sc = cgrp->subtree_control;
2834 old_ss = cgrp->child_subsys_mask;
2835 new_sc = (old_sc | enable) & ~disable;
2836 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2838 css_enable = ~old_ss & new_ss;
2839 css_disable = old_ss & ~new_ss;
2840 enable |= css_enable;
2841 disable |= css_disable;
2844 * Because css offlining is asynchronous, userland might try to
2845 * re-enable the same controller while the previous instance is
2846 * still around. In such cases, wait till it's gone using
2849 for_each_subsys_which(ss, ssid, &css_enable) {
2850 cgroup_for_each_live_child(child, cgrp) {
2853 if (!cgroup_css(child, ss))
2857 prepare_to_wait(&child->offline_waitq, &wait,
2858 TASK_UNINTERRUPTIBLE);
2859 cgroup_kn_unlock(of->kn);
2861 finish_wait(&child->offline_waitq, &wait);
2864 return restart_syscall();
2868 cgrp->subtree_control = new_sc;
2869 cgrp->child_subsys_mask = new_ss;
2872 * Create new csses or make the existing ones visible. A css is
2873 * created invisible if it's being implicitly enabled through
2874 * dependency. An invisible css is made visible when the userland
2875 * explicitly enables it.
2877 for_each_subsys(ss, ssid) {
2878 if (!(enable & (1 << ssid)))
2881 cgroup_for_each_live_child(child, cgrp) {
2882 if (css_enable & (1 << ssid))
2883 ret = create_css(child, ss,
2884 cgrp->subtree_control & (1 << ssid));
2886 ret = cgroup_populate_dir(child, 1 << ssid);
2893 * At this point, cgroup_e_css() results reflect the new csses
2894 * making the following cgroup_update_dfl_csses() properly update
2895 * css associations of all tasks in the subtree.
2897 ret = cgroup_update_dfl_csses(cgrp);
2902 * All tasks are migrated out of disabled csses. Kill or hide
2903 * them. A css is hidden when the userland requests it to be
2904 * disabled while other subsystems are still depending on it. The
2905 * css must not actively control resources and be in the vanilla
2906 * state if it's made visible again later. Controllers which may
2907 * be depended upon should provide ->css_reset() for this purpose.
2909 for_each_subsys(ss, ssid) {
2910 if (!(disable & (1 << ssid)))
2913 cgroup_for_each_live_child(child, cgrp) {
2914 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2916 if (css_disable & (1 << ssid)) {
2919 cgroup_clear_dir(child, 1 << ssid);
2927 * The effective csses of all the descendants (excluding @cgrp) may
2928 * have changed. Subsystems can optionally subscribe to this event
2929 * by implementing ->css_e_css_changed() which is invoked if any of
2930 * the effective csses seen from the css's cgroup may have changed.
2932 for_each_subsys(ss, ssid) {
2933 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2934 struct cgroup_subsys_state *css;
2936 if (!ss->css_e_css_changed || !this_css)
2939 css_for_each_descendant_pre(css, this_css)
2940 if (css != this_css)
2941 ss->css_e_css_changed(css);
2944 kernfs_activate(cgrp->kn);
2947 cgroup_kn_unlock(of->kn);
2948 return ret ?: nbytes;
2951 cgrp->subtree_control = old_sc;
2952 cgrp->child_subsys_mask = old_ss;
2954 for_each_subsys(ss, ssid) {
2955 if (!(enable & (1 << ssid)))
2958 cgroup_for_each_live_child(child, cgrp) {
2959 struct cgroup_subsys_state *css = cgroup_css(child, ss);
2964 if (css_enable & (1 << ssid))
2967 cgroup_clear_dir(child, 1 << ssid);
2973 static int cgroup_populated_show(struct seq_file *seq, void *v)
2975 seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2979 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
2980 size_t nbytes, loff_t off)
2982 struct cgroup *cgrp = of->kn->parent->priv;
2983 struct cftype *cft = of->kn->priv;
2984 struct cgroup_subsys_state *css;
2988 return cft->write(of, buf, nbytes, off);
2991 * kernfs guarantees that a file isn't deleted with operations in
2992 * flight, which means that the matching css is and stays alive and
2993 * doesn't need to be pinned. The RCU locking is not necessary
2994 * either. It's just for the convenience of using cgroup_css().
2997 css = cgroup_css(cgrp, cft->ss);
3000 if (cft->write_u64) {
3001 unsigned long long v;
3002 ret = kstrtoull(buf, 0, &v);
3004 ret = cft->write_u64(css, cft, v);
3005 } else if (cft->write_s64) {
3007 ret = kstrtoll(buf, 0, &v);
3009 ret = cft->write_s64(css, cft, v);
3014 return ret ?: nbytes;
3017 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3019 return seq_cft(seq)->seq_start(seq, ppos);
3022 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3024 return seq_cft(seq)->seq_next(seq, v, ppos);
3027 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3029 seq_cft(seq)->seq_stop(seq, v);
3032 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3034 struct cftype *cft = seq_cft(m);
3035 struct cgroup_subsys_state *css = seq_css(m);
3038 return cft->seq_show(m, arg);
3041 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3042 else if (cft->read_s64)
3043 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3049 static struct kernfs_ops cgroup_kf_single_ops = {
3050 .atomic_write_len = PAGE_SIZE,
3051 .write = cgroup_file_write,
3052 .seq_show = cgroup_seqfile_show,
3055 static struct kernfs_ops cgroup_kf_ops = {
3056 .atomic_write_len = PAGE_SIZE,
3057 .write = cgroup_file_write,
3058 .seq_start = cgroup_seqfile_start,
3059 .seq_next = cgroup_seqfile_next,
3060 .seq_stop = cgroup_seqfile_stop,
3061 .seq_show = cgroup_seqfile_show,
3065 * cgroup_rename - Only allow simple rename of directories in place.
3067 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3068 const char *new_name_str)
3070 struct cgroup *cgrp = kn->priv;
3073 if (kernfs_type(kn) != KERNFS_DIR)
3075 if (kn->parent != new_parent)
3079 * This isn't a proper migration and its usefulness is very
3080 * limited. Disallow on the default hierarchy.
3082 if (cgroup_on_dfl(cgrp))
3086 * We're gonna grab cgroup_mutex which nests outside kernfs
3087 * active_ref. kernfs_rename() doesn't require active_ref
3088 * protection. Break them before grabbing cgroup_mutex.
3090 kernfs_break_active_protection(new_parent);
3091 kernfs_break_active_protection(kn);
3093 mutex_lock(&cgroup_mutex);
3095 ret = kernfs_rename(kn, new_parent, new_name_str);
3097 mutex_unlock(&cgroup_mutex);
3099 kernfs_unbreak_active_protection(kn);
3100 kernfs_unbreak_active_protection(new_parent);
3104 /* set uid and gid of cgroup dirs and files to that of the creator */
3105 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3107 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3108 .ia_uid = current_fsuid(),
3109 .ia_gid = current_fsgid(), };
3111 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3112 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3115 return kernfs_setattr(kn, &iattr);
3118 static int cgroup_add_file(struct cgroup *cgrp, struct cftype *cft)
3120 char name[CGROUP_FILE_NAME_MAX];
3121 struct kernfs_node *kn;
3122 struct lock_class_key *key = NULL;
3125 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3126 key = &cft->lockdep_key;
3128 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3129 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3134 ret = cgroup_kn_set_ugid(kn);
3140 if (cft->write == cgroup_procs_write)
3141 cgrp->procs_kn = kn;
3142 else if (cft->seq_show == cgroup_populated_show)
3143 cgrp->populated_kn = kn;
3148 * cgroup_addrm_files - add or remove files to a cgroup directory
3149 * @cgrp: the target cgroup
3150 * @cfts: array of cftypes to be added
3151 * @is_add: whether to add or remove
3153 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3154 * For removals, this function never fails. If addition fails, this
3155 * function doesn't remove files already added. The caller is responsible
3158 static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3164 lockdep_assert_held(&cgroup_mutex);
3166 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3167 /* does cft->flags tell us to skip this file on @cgrp? */
3168 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3170 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3172 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3174 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3178 ret = cgroup_add_file(cgrp, cft);
3180 pr_warn("%s: failed to add %s, err=%d\n",
3181 __func__, cft->name, ret);
3185 cgroup_rm_file(cgrp, cft);
3191 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3194 struct cgroup_subsys *ss = cfts[0].ss;
3195 struct cgroup *root = &ss->root->cgrp;
3196 struct cgroup_subsys_state *css;
3199 lockdep_assert_held(&cgroup_mutex);
3201 /* add/rm files for all cgroups created before */
3202 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3203 struct cgroup *cgrp = css->cgroup;
3205 if (cgroup_is_dead(cgrp))
3208 ret = cgroup_addrm_files(cgrp, cfts, is_add);
3214 kernfs_activate(root->kn);
3218 static void cgroup_exit_cftypes(struct cftype *cfts)
3222 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3223 /* free copy for custom atomic_write_len, see init_cftypes() */
3224 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3229 /* revert flags set by cgroup core while adding @cfts */
3230 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3234 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3238 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3239 struct kernfs_ops *kf_ops;
3241 WARN_ON(cft->ss || cft->kf_ops);
3244 kf_ops = &cgroup_kf_ops;
3246 kf_ops = &cgroup_kf_single_ops;
3249 * Ugh... if @cft wants a custom max_write_len, we need to
3250 * make a copy of kf_ops to set its atomic_write_len.
3252 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3253 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3255 cgroup_exit_cftypes(cfts);
3258 kf_ops->atomic_write_len = cft->max_write_len;
3261 cft->kf_ops = kf_ops;
3268 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3270 lockdep_assert_held(&cgroup_mutex);
3272 if (!cfts || !cfts[0].ss)
3275 list_del(&cfts->node);
3276 cgroup_apply_cftypes(cfts, false);
3277 cgroup_exit_cftypes(cfts);
3282 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3283 * @cfts: zero-length name terminated array of cftypes
3285 * Unregister @cfts. Files described by @cfts are removed from all
3286 * existing cgroups and all future cgroups won't have them either. This
3287 * function can be called anytime whether @cfts' subsys is attached or not.
3289 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3292 int cgroup_rm_cftypes(struct cftype *cfts)
3296 mutex_lock(&cgroup_mutex);
3297 ret = cgroup_rm_cftypes_locked(cfts);
3298 mutex_unlock(&cgroup_mutex);
3303 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3304 * @ss: target cgroup subsystem
3305 * @cfts: zero-length name terminated array of cftypes
3307 * Register @cfts to @ss. Files described by @cfts are created for all
3308 * existing cgroups to which @ss is attached and all future cgroups will
3309 * have them too. This function can be called anytime whether @ss is
3312 * Returns 0 on successful registration, -errno on failure. Note that this
3313 * function currently returns 0 as long as @cfts registration is successful
3314 * even if some file creation attempts on existing cgroups fail.
3316 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3323 if (!cfts || cfts[0].name[0] == '\0')
3326 ret = cgroup_init_cftypes(ss, cfts);
3330 mutex_lock(&cgroup_mutex);
3332 list_add_tail(&cfts->node, &ss->cfts);
3333 ret = cgroup_apply_cftypes(cfts, true);
3335 cgroup_rm_cftypes_locked(cfts);
3337 mutex_unlock(&cgroup_mutex);
3342 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3343 * @ss: target cgroup subsystem
3344 * @cfts: zero-length name terminated array of cftypes
3346 * Similar to cgroup_add_cftypes() but the added files are only used for
3347 * the default hierarchy.
3349 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3353 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3354 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3355 return cgroup_add_cftypes(ss, cfts);
3359 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3360 * @ss: target cgroup subsystem
3361 * @cfts: zero-length name terminated array of cftypes
3363 * Similar to cgroup_add_cftypes() but the added files are only used for
3364 * the legacy hierarchies.
3366 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3371 * If legacy_flies_on_dfl, we want to show the legacy files on the
3372 * dfl hierarchy but iff the target subsystem hasn't been updated
3373 * for the dfl hierarchy yet.
3375 if (!cgroup_legacy_files_on_dfl ||
3376 ss->dfl_cftypes != ss->legacy_cftypes) {
3377 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3378 cft->flags |= __CFTYPE_NOT_ON_DFL;
3381 return cgroup_add_cftypes(ss, cfts);
3385 * cgroup_task_count - count the number of tasks in a cgroup.
3386 * @cgrp: the cgroup in question
3388 * Return the number of tasks in the cgroup.
3390 static int cgroup_task_count(const struct cgroup *cgrp)
3393 struct cgrp_cset_link *link;
3395 down_read(&css_set_rwsem);
3396 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3397 count += atomic_read(&link->cset->refcount);
3398 up_read(&css_set_rwsem);
3403 * css_next_child - find the next child of a given css
3404 * @pos: the current position (%NULL to initiate traversal)
3405 * @parent: css whose children to walk
3407 * This function returns the next child of @parent and should be called
3408 * under either cgroup_mutex or RCU read lock. The only requirement is
3409 * that @parent and @pos are accessible. The next sibling is guaranteed to
3410 * be returned regardless of their states.
3412 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3413 * css which finished ->css_online() is guaranteed to be visible in the
3414 * future iterations and will stay visible until the last reference is put.
3415 * A css which hasn't finished ->css_online() or already finished
3416 * ->css_offline() may show up during traversal. It's each subsystem's
3417 * responsibility to synchronize against on/offlining.
3419 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3420 struct cgroup_subsys_state *parent)
3422 struct cgroup_subsys_state *next;
3424 cgroup_assert_mutex_or_rcu_locked();
3427 * @pos could already have been unlinked from the sibling list.
3428 * Once a cgroup is removed, its ->sibling.next is no longer
3429 * updated when its next sibling changes. CSS_RELEASED is set when
3430 * @pos is taken off list, at which time its next pointer is valid,
3431 * and, as releases are serialized, the one pointed to by the next
3432 * pointer is guaranteed to not have started release yet. This
3433 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3434 * critical section, the one pointed to by its next pointer is
3435 * guaranteed to not have finished its RCU grace period even if we
3436 * have dropped rcu_read_lock() inbetween iterations.
3438 * If @pos has CSS_RELEASED set, its next pointer can't be
3439 * dereferenced; however, as each css is given a monotonically
3440 * increasing unique serial number and always appended to the
3441 * sibling list, the next one can be found by walking the parent's
3442 * children until the first css with higher serial number than
3443 * @pos's. While this path can be slower, it happens iff iteration
3444 * races against release and the race window is very small.
3447 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3448 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3449 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3451 list_for_each_entry_rcu(next, &parent->children, sibling)
3452 if (next->serial_nr > pos->serial_nr)
3457 * @next, if not pointing to the head, can be dereferenced and is
3460 if (&next->sibling != &parent->children)
3466 * css_next_descendant_pre - find the next descendant for pre-order walk
3467 * @pos: the current position (%NULL to initiate traversal)
3468 * @root: css whose descendants to walk
3470 * To be used by css_for_each_descendant_pre(). Find the next descendant
3471 * to visit for pre-order traversal of @root's descendants. @root is
3472 * included in the iteration and the first node to be visited.
3474 * While this function requires cgroup_mutex or RCU read locking, it
3475 * doesn't require the whole traversal to be contained in a single critical
3476 * section. This function will return the correct next descendant as long
3477 * as both @pos and @root are accessible and @pos is a descendant of @root.
3479 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3480 * css which finished ->css_online() is guaranteed to be visible in the
3481 * future iterations and will stay visible until the last reference is put.
3482 * A css which hasn't finished ->css_online() or already finished
3483 * ->css_offline() may show up during traversal. It's each subsystem's
3484 * responsibility to synchronize against on/offlining.
3486 struct cgroup_subsys_state *
3487 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3488 struct cgroup_subsys_state *root)
3490 struct cgroup_subsys_state *next;
3492 cgroup_assert_mutex_or_rcu_locked();
3494 /* if first iteration, visit @root */
3498 /* visit the first child if exists */
3499 next = css_next_child(NULL, pos);
3503 /* no child, visit my or the closest ancestor's next sibling */
3504 while (pos != root) {
3505 next = css_next_child(pos, pos->parent);
3515 * css_rightmost_descendant - return the rightmost descendant of a css
3516 * @pos: css of interest
3518 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3519 * is returned. This can be used during pre-order traversal to skip
3522 * While this function requires cgroup_mutex or RCU read locking, it
3523 * doesn't require the whole traversal to be contained in a single critical
3524 * section. This function will return the correct rightmost descendant as
3525 * long as @pos is accessible.
3527 struct cgroup_subsys_state *
3528 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3530 struct cgroup_subsys_state *last, *tmp;
3532 cgroup_assert_mutex_or_rcu_locked();
3536 /* ->prev isn't RCU safe, walk ->next till the end */
3538 css_for_each_child(tmp, last)
3545 static struct cgroup_subsys_state *
3546 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3548 struct cgroup_subsys_state *last;
3552 pos = css_next_child(NULL, pos);
3559 * css_next_descendant_post - find the next descendant for post-order walk
3560 * @pos: the current position (%NULL to initiate traversal)
3561 * @root: css whose descendants to walk
3563 * To be used by css_for_each_descendant_post(). Find the next descendant
3564 * to visit for post-order traversal of @root's descendants. @root is
3565 * included in the iteration and the last node to be visited.
3567 * While this function requires cgroup_mutex or RCU read locking, it
3568 * doesn't require the whole traversal to be contained in a single critical
3569 * section. This function will return the correct next descendant as long
3570 * as both @pos and @cgroup are accessible and @pos is a descendant of
3573 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3574 * css which finished ->css_online() is guaranteed to be visible in the
3575 * future iterations and will stay visible until the last reference is put.
3576 * A css which hasn't finished ->css_online() or already finished
3577 * ->css_offline() may show up during traversal. It's each subsystem's
3578 * responsibility to synchronize against on/offlining.
3580 struct cgroup_subsys_state *
3581 css_next_descendant_post(struct cgroup_subsys_state *pos,
3582 struct cgroup_subsys_state *root)
3584 struct cgroup_subsys_state *next;
3586 cgroup_assert_mutex_or_rcu_locked();
3588 /* if first iteration, visit leftmost descendant which may be @root */
3590 return css_leftmost_descendant(root);
3592 /* if we visited @root, we're done */
3596 /* if there's an unvisited sibling, visit its leftmost descendant */
3597 next = css_next_child(pos, pos->parent);
3599 return css_leftmost_descendant(next);
3601 /* no sibling left, visit parent */
3606 * css_has_online_children - does a css have online children
3607 * @css: the target css
3609 * Returns %true if @css has any online children; otherwise, %false. This
3610 * function can be called from any context but the caller is responsible
3611 * for synchronizing against on/offlining as necessary.
3613 bool css_has_online_children(struct cgroup_subsys_state *css)
3615 struct cgroup_subsys_state *child;
3619 css_for_each_child(child, css) {
3620 if (child->flags & CSS_ONLINE) {
3630 * css_advance_task_iter - advance a task itererator to the next css_set
3631 * @it: the iterator to advance
3633 * Advance @it to the next css_set to walk.
3635 static void css_advance_task_iter(struct css_task_iter *it)
3637 struct list_head *l = it->cset_pos;
3638 struct cgrp_cset_link *link;
3639 struct css_set *cset;
3641 /* Advance to the next non-empty css_set */
3644 if (l == it->cset_head) {
3645 it->cset_pos = NULL;
3650 cset = container_of(l, struct css_set,
3651 e_cset_node[it->ss->id]);
3653 link = list_entry(l, struct cgrp_cset_link, cset_link);
3656 } while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3660 if (!list_empty(&cset->tasks))
3661 it->task_pos = cset->tasks.next;
3663 it->task_pos = cset->mg_tasks.next;
3665 it->tasks_head = &cset->tasks;
3666 it->mg_tasks_head = &cset->mg_tasks;
3670 * css_task_iter_start - initiate task iteration
3671 * @css: the css to walk tasks of
3672 * @it: the task iterator to use
3674 * Initiate iteration through the tasks of @css. The caller can call
3675 * css_task_iter_next() to walk through the tasks until the function
3676 * returns NULL. On completion of iteration, css_task_iter_end() must be
3679 * Note that this function acquires a lock which is released when the
3680 * iteration finishes. The caller can't sleep while iteration is in
3683 void css_task_iter_start(struct cgroup_subsys_state *css,
3684 struct css_task_iter *it)
3685 __acquires(css_set_rwsem)
3687 /* no one should try to iterate before mounting cgroups */
3688 WARN_ON_ONCE(!use_task_css_set_links);
3690 down_read(&css_set_rwsem);
3695 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3697 it->cset_pos = &css->cgroup->cset_links;
3699 it->cset_head = it->cset_pos;
3701 css_advance_task_iter(it);
3705 * css_task_iter_next - return the next task for the iterator
3706 * @it: the task iterator being iterated
3708 * The "next" function for task iteration. @it should have been
3709 * initialized via css_task_iter_start(). Returns NULL when the iteration
3712 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3714 struct task_struct *res;
3715 struct list_head *l = it->task_pos;
3717 /* If the iterator cg is NULL, we have no tasks */
3720 res = list_entry(l, struct task_struct, cg_list);
3723 * Advance iterator to find next entry. cset->tasks is consumed
3724 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3729 if (l == it->tasks_head)
3730 l = it->mg_tasks_head->next;
3732 if (l == it->mg_tasks_head)
3733 css_advance_task_iter(it);
3741 * css_task_iter_end - finish task iteration
3742 * @it: the task iterator to finish
3744 * Finish task iteration started by css_task_iter_start().
3746 void css_task_iter_end(struct css_task_iter *it)
3747 __releases(css_set_rwsem)
3749 up_read(&css_set_rwsem);
3753 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3754 * @to: cgroup to which the tasks will be moved
3755 * @from: cgroup in which the tasks currently reside
3757 * Locking rules between cgroup_post_fork() and the migration path
3758 * guarantee that, if a task is forking while being migrated, the new child
3759 * is guaranteed to be either visible in the source cgroup after the
3760 * parent's migration is complete or put into the target cgroup. No task
3761 * can slip out of migration through forking.
3763 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3765 LIST_HEAD(preloaded_csets);
3766 struct cgrp_cset_link *link;
3767 struct css_task_iter it;
3768 struct task_struct *task;
3771 mutex_lock(&cgroup_mutex);
3773 /* all tasks in @from are being moved, all csets are source */
3774 down_read(&css_set_rwsem);
3775 list_for_each_entry(link, &from->cset_links, cset_link)
3776 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3777 up_read(&css_set_rwsem);
3779 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3784 * Migrate tasks one-by-one until @form is empty. This fails iff
3785 * ->can_attach() fails.
3788 css_task_iter_start(&from->self, &it);
3789 task = css_task_iter_next(&it);
3791 get_task_struct(task);
3792 css_task_iter_end(&it);
3795 ret = cgroup_migrate(to, task, false);
3796 put_task_struct(task);
3798 } while (task && !ret);
3800 cgroup_migrate_finish(&preloaded_csets);
3801 mutex_unlock(&cgroup_mutex);
3806 * Stuff for reading the 'tasks'/'procs' files.
3808 * Reading this file can return large amounts of data if a cgroup has
3809 * *lots* of attached tasks. So it may need several calls to read(),
3810 * but we cannot guarantee that the information we produce is correct
3811 * unless we produce it entirely atomically.
3815 /* which pidlist file are we talking about? */
3816 enum cgroup_filetype {
3822 * A pidlist is a list of pids that virtually represents the contents of one
3823 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3824 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3827 struct cgroup_pidlist {
3829 * used to find which pidlist is wanted. doesn't change as long as
3830 * this particular list stays in the list.
3832 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3835 /* how many elements the above list has */
3837 /* each of these stored in a list by its cgroup */
3838 struct list_head links;
3839 /* pointer to the cgroup we belong to, for list removal purposes */
3840 struct cgroup *owner;
3841 /* for delayed destruction */
3842 struct delayed_work destroy_dwork;
3846 * The following two functions "fix" the issue where there are more pids
3847 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3848 * TODO: replace with a kernel-wide solution to this problem
3850 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3851 static void *pidlist_allocate(int count)
3853 if (PIDLIST_TOO_LARGE(count))
3854 return vmalloc(count * sizeof(pid_t));
3856 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3859 static void pidlist_free(void *p)
3865 * Used to destroy all pidlists lingering waiting for destroy timer. None
3866 * should be left afterwards.
3868 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3870 struct cgroup_pidlist *l, *tmp_l;
3872 mutex_lock(&cgrp->pidlist_mutex);
3873 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3874 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3875 mutex_unlock(&cgrp->pidlist_mutex);
3877 flush_workqueue(cgroup_pidlist_destroy_wq);
3878 BUG_ON(!list_empty(&cgrp->pidlists));
3881 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3883 struct delayed_work *dwork = to_delayed_work(work);
3884 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3886 struct cgroup_pidlist *tofree = NULL;
3888 mutex_lock(&l->owner->pidlist_mutex);
3891 * Destroy iff we didn't get queued again. The state won't change
3892 * as destroy_dwork can only be queued while locked.
3894 if (!delayed_work_pending(dwork)) {
3895 list_del(&l->links);
3896 pidlist_free(l->list);
3897 put_pid_ns(l->key.ns);
3901 mutex_unlock(&l->owner->pidlist_mutex);
3906 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3907 * Returns the number of unique elements.
3909 static int pidlist_uniq(pid_t *list, int length)
3914 * we presume the 0th element is unique, so i starts at 1. trivial
3915 * edge cases first; no work needs to be done for either
3917 if (length == 0 || length == 1)
3919 /* src and dest walk down the list; dest counts unique elements */
3920 for (src = 1; src < length; src++) {
3921 /* find next unique element */
3922 while (list[src] == list[src-1]) {
3927 /* dest always points to where the next unique element goes */
3928 list[dest] = list[src];
3936 * The two pid files - task and cgroup.procs - guaranteed that the result
3937 * is sorted, which forced this whole pidlist fiasco. As pid order is
3938 * different per namespace, each namespace needs differently sorted list,
3939 * making it impossible to use, for example, single rbtree of member tasks
3940 * sorted by task pointer. As pidlists can be fairly large, allocating one
3941 * per open file is dangerous, so cgroup had to implement shared pool of
3942 * pidlists keyed by cgroup and namespace.
3944 * All this extra complexity was caused by the original implementation
3945 * committing to an entirely unnecessary property. In the long term, we
3946 * want to do away with it. Explicitly scramble sort order if on the
3947 * default hierarchy so that no such expectation exists in the new
3950 * Scrambling is done by swapping every two consecutive bits, which is
3951 * non-identity one-to-one mapping which disturbs sort order sufficiently.
3953 static pid_t pid_fry(pid_t pid)
3955 unsigned a = pid & 0x55555555;
3956 unsigned b = pid & 0xAAAAAAAA;
3958 return (a << 1) | (b >> 1);
3961 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3963 if (cgroup_on_dfl(cgrp))
3964 return pid_fry(pid);
3969 static int cmppid(const void *a, const void *b)
3971 return *(pid_t *)a - *(pid_t *)b;
3974 static int fried_cmppid(const void *a, const void *b)
3976 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
3979 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
3980 enum cgroup_filetype type)
3982 struct cgroup_pidlist *l;
3983 /* don't need task_nsproxy() if we're looking at ourself */
3984 struct pid_namespace *ns = task_active_pid_ns(current);
3986 lockdep_assert_held(&cgrp->pidlist_mutex);
3988 list_for_each_entry(l, &cgrp->pidlists, links)
3989 if (l->key.type == type && l->key.ns == ns)
3995 * find the appropriate pidlist for our purpose (given procs vs tasks)
3996 * returns with the lock on that pidlist already held, and takes care
3997 * of the use count, or returns NULL with no locks held if we're out of
4000 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4001 enum cgroup_filetype type)
4003 struct cgroup_pidlist *l;
4005 lockdep_assert_held(&cgrp->pidlist_mutex);
4007 l = cgroup_pidlist_find(cgrp, type);
4011 /* entry not found; create a new one */
4012 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4016 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4018 /* don't need task_nsproxy() if we're looking at ourself */
4019 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4021 list_add(&l->links, &cgrp->pidlists);
4026 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4028 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4029 struct cgroup_pidlist **lp)
4033 int pid, n = 0; /* used for populating the array */
4034 struct css_task_iter it;
4035 struct task_struct *tsk;
4036 struct cgroup_pidlist *l;
4038 lockdep_assert_held(&cgrp->pidlist_mutex);
4041 * If cgroup gets more users after we read count, we won't have
4042 * enough space - tough. This race is indistinguishable to the
4043 * caller from the case that the additional cgroup users didn't
4044 * show up until sometime later on.
4046 length = cgroup_task_count(cgrp);
4047 array = pidlist_allocate(length);
4050 /* now, populate the array */
4051 css_task_iter_start(&cgrp->self, &it);
4052 while ((tsk = css_task_iter_next(&it))) {
4053 if (unlikely(n == length))
4055 /* get tgid or pid for procs or tasks file respectively */
4056 if (type == CGROUP_FILE_PROCS)
4057 pid = task_tgid_vnr(tsk);
4059 pid = task_pid_vnr(tsk);
4060 if (pid > 0) /* make sure to only use valid results */
4063 css_task_iter_end(&it);
4065 /* now sort & (if procs) strip out duplicates */
4066 if (cgroup_on_dfl(cgrp))
4067 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4069 sort(array, length, sizeof(pid_t), cmppid, NULL);
4070 if (type == CGROUP_FILE_PROCS)
4071 length = pidlist_uniq(array, length);
4073 l = cgroup_pidlist_find_create(cgrp, type);
4075 pidlist_free(array);
4079 /* store array, freeing old if necessary */
4080 pidlist_free(l->list);
4088 * cgroupstats_build - build and fill cgroupstats
4089 * @stats: cgroupstats to fill information into
4090 * @dentry: A dentry entry belonging to the cgroup for which stats have
4093 * Build and fill cgroupstats so that taskstats can export it to user
4096 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4098 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4099 struct cgroup *cgrp;
4100 struct css_task_iter it;
4101 struct task_struct *tsk;
4103 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4104 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4105 kernfs_type(kn) != KERNFS_DIR)
4108 mutex_lock(&cgroup_mutex);
4111 * We aren't being called from kernfs and there's no guarantee on
4112 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4113 * @kn->priv is RCU safe. Let's do the RCU dancing.
4116 cgrp = rcu_dereference(kn->priv);
4117 if (!cgrp || cgroup_is_dead(cgrp)) {
4119 mutex_unlock(&cgroup_mutex);
4124 css_task_iter_start(&cgrp->self, &it);
4125 while ((tsk = css_task_iter_next(&it))) {
4126 switch (tsk->state) {
4128 stats->nr_running++;
4130 case TASK_INTERRUPTIBLE:
4131 stats->nr_sleeping++;
4133 case TASK_UNINTERRUPTIBLE:
4134 stats->nr_uninterruptible++;
4137 stats->nr_stopped++;
4140 if (delayacct_is_task_waiting_on_io(tsk))
4141 stats->nr_io_wait++;
4145 css_task_iter_end(&it);
4147 mutex_unlock(&cgroup_mutex);
4153 * seq_file methods for the tasks/procs files. The seq_file position is the
4154 * next pid to display; the seq_file iterator is a pointer to the pid
4155 * in the cgroup->l->list array.
4158 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4161 * Initially we receive a position value that corresponds to
4162 * one more than the last pid shown (or 0 on the first call or
4163 * after a seek to the start). Use a binary-search to find the
4164 * next pid to display, if any
4166 struct kernfs_open_file *of = s->private;
4167 struct cgroup *cgrp = seq_css(s)->cgroup;
4168 struct cgroup_pidlist *l;
4169 enum cgroup_filetype type = seq_cft(s)->private;
4170 int index = 0, pid = *pos;
4173 mutex_lock(&cgrp->pidlist_mutex);
4176 * !NULL @of->priv indicates that this isn't the first start()
4177 * after open. If the matching pidlist is around, we can use that.
4178 * Look for it. Note that @of->priv can't be used directly. It
4179 * could already have been destroyed.
4182 of->priv = cgroup_pidlist_find(cgrp, type);
4185 * Either this is the first start() after open or the matching
4186 * pidlist has been destroyed inbetween. Create a new one.
4189 ret = pidlist_array_load(cgrp, type,
4190 (struct cgroup_pidlist **)&of->priv);
4192 return ERR_PTR(ret);
4197 int end = l->length;
4199 while (index < end) {
4200 int mid = (index + end) / 2;
4201 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4204 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4210 /* If we're off the end of the array, we're done */
4211 if (index >= l->length)
4213 /* Update the abstract position to be the actual pid that we found */
4214 iter = l->list + index;
4215 *pos = cgroup_pid_fry(cgrp, *iter);
4219 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4221 struct kernfs_open_file *of = s->private;
4222 struct cgroup_pidlist *l = of->priv;
4225 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4226 CGROUP_PIDLIST_DESTROY_DELAY);
4227 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4230 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4232 struct kernfs_open_file *of = s->private;
4233 struct cgroup_pidlist *l = of->priv;
4235 pid_t *end = l->list + l->length;
4237 * Advance to the next pid in the array. If this goes off the
4244 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4249 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4251 seq_printf(s, "%d\n", *(int *)v);
4256 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4259 return notify_on_release(css->cgroup);
4262 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4263 struct cftype *cft, u64 val)
4266 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4268 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4272 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4275 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4278 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4279 struct cftype *cft, u64 val)
4282 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4284 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4288 /* cgroup core interface files for the default hierarchy */
4289 static struct cftype cgroup_dfl_base_files[] = {
4291 .name = "cgroup.procs",
4292 .seq_start = cgroup_pidlist_start,
4293 .seq_next = cgroup_pidlist_next,
4294 .seq_stop = cgroup_pidlist_stop,
4295 .seq_show = cgroup_pidlist_show,
4296 .private = CGROUP_FILE_PROCS,
4297 .write = cgroup_procs_write,
4298 .mode = S_IRUGO | S_IWUSR,
4301 .name = "cgroup.controllers",
4302 .flags = CFTYPE_ONLY_ON_ROOT,
4303 .seq_show = cgroup_root_controllers_show,
4306 .name = "cgroup.controllers",
4307 .flags = CFTYPE_NOT_ON_ROOT,
4308 .seq_show = cgroup_controllers_show,
4311 .name = "cgroup.subtree_control",
4312 .seq_show = cgroup_subtree_control_show,
4313 .write = cgroup_subtree_control_write,
4316 .name = "cgroup.populated",
4317 .flags = CFTYPE_NOT_ON_ROOT,
4318 .seq_show = cgroup_populated_show,
4323 /* cgroup core interface files for the legacy hierarchies */
4324 static struct cftype cgroup_legacy_base_files[] = {
4326 .name = "cgroup.procs",
4327 .seq_start = cgroup_pidlist_start,
4328 .seq_next = cgroup_pidlist_next,
4329 .seq_stop = cgroup_pidlist_stop,
4330 .seq_show = cgroup_pidlist_show,
4331 .private = CGROUP_FILE_PROCS,
4332 .write = cgroup_procs_write,
4333 .mode = S_IRUGO | S_IWUSR,
4336 .name = "cgroup.clone_children",
4337 .read_u64 = cgroup_clone_children_read,
4338 .write_u64 = cgroup_clone_children_write,
4341 .name = "cgroup.sane_behavior",
4342 .flags = CFTYPE_ONLY_ON_ROOT,
4343 .seq_show = cgroup_sane_behavior_show,
4347 .seq_start = cgroup_pidlist_start,
4348 .seq_next = cgroup_pidlist_next,
4349 .seq_stop = cgroup_pidlist_stop,
4350 .seq_show = cgroup_pidlist_show,
4351 .private = CGROUP_FILE_TASKS,
4352 .write = cgroup_tasks_write,
4353 .mode = S_IRUGO | S_IWUSR,
4356 .name = "notify_on_release",
4357 .read_u64 = cgroup_read_notify_on_release,
4358 .write_u64 = cgroup_write_notify_on_release,
4361 .name = "release_agent",
4362 .flags = CFTYPE_ONLY_ON_ROOT,
4363 .seq_show = cgroup_release_agent_show,
4364 .write = cgroup_release_agent_write,
4365 .max_write_len = PATH_MAX - 1,
4371 * cgroup_populate_dir - create subsys files in a cgroup directory
4372 * @cgrp: target cgroup
4373 * @subsys_mask: mask of the subsystem ids whose files should be added
4375 * On failure, no file is added.
4377 static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4379 struct cgroup_subsys *ss;
4382 /* process cftsets of each subsystem */
4383 for_each_subsys(ss, i) {
4384 struct cftype *cfts;
4386 if (!(subsys_mask & (1 << i)))
4389 list_for_each_entry(cfts, &ss->cfts, node) {
4390 ret = cgroup_addrm_files(cgrp, cfts, true);
4397 cgroup_clear_dir(cgrp, subsys_mask);
4402 * css destruction is four-stage process.
4404 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4405 * Implemented in kill_css().
4407 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4408 * and thus css_tryget_online() is guaranteed to fail, the css can be
4409 * offlined by invoking offline_css(). After offlining, the base ref is
4410 * put. Implemented in css_killed_work_fn().
4412 * 3. When the percpu_ref reaches zero, the only possible remaining
4413 * accessors are inside RCU read sections. css_release() schedules the
4416 * 4. After the grace period, the css can be freed. Implemented in
4417 * css_free_work_fn().
4419 * It is actually hairier because both step 2 and 4 require process context
4420 * and thus involve punting to css->destroy_work adding two additional
4421 * steps to the already complex sequence.
4423 static void css_free_work_fn(struct work_struct *work)
4425 struct cgroup_subsys_state *css =
4426 container_of(work, struct cgroup_subsys_state, destroy_work);
4427 struct cgroup_subsys *ss = css->ss;
4428 struct cgroup *cgrp = css->cgroup;
4430 percpu_ref_exit(&css->refcnt);
4437 css_put(css->parent);
4440 cgroup_idr_remove(&ss->css_idr, id);
4443 /* cgroup free path */
4444 atomic_dec(&cgrp->root->nr_cgrps);
4445 cgroup_pidlist_destroy_all(cgrp);
4446 cancel_work_sync(&cgrp->release_agent_work);
4448 if (cgroup_parent(cgrp)) {
4450 * We get a ref to the parent, and put the ref when
4451 * this cgroup is being freed, so it's guaranteed
4452 * that the parent won't be destroyed before its
4455 cgroup_put(cgroup_parent(cgrp));
4456 kernfs_put(cgrp->kn);
4460 * This is root cgroup's refcnt reaching zero,
4461 * which indicates that the root should be
4464 cgroup_destroy_root(cgrp->root);
4469 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4471 struct cgroup_subsys_state *css =
4472 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4474 INIT_WORK(&css->destroy_work, css_free_work_fn);
4475 queue_work(cgroup_destroy_wq, &css->destroy_work);
4478 static void css_release_work_fn(struct work_struct *work)
4480 struct cgroup_subsys_state *css =
4481 container_of(work, struct cgroup_subsys_state, destroy_work);
4482 struct cgroup_subsys *ss = css->ss;
4483 struct cgroup *cgrp = css->cgroup;
4485 mutex_lock(&cgroup_mutex);
4487 css->flags |= CSS_RELEASED;
4488 list_del_rcu(&css->sibling);
4491 /* css release path */
4492 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4493 if (ss->css_released)
4494 ss->css_released(css);
4496 /* cgroup release path */
4497 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4501 * There are two control paths which try to determine
4502 * cgroup from dentry without going through kernfs -
4503 * cgroupstats_build() and css_tryget_online_from_dir().
4504 * Those are supported by RCU protecting clearing of
4505 * cgrp->kn->priv backpointer.
4507 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4510 mutex_unlock(&cgroup_mutex);
4512 call_rcu(&css->rcu_head, css_free_rcu_fn);
4515 static void css_release(struct percpu_ref *ref)
4517 struct cgroup_subsys_state *css =
4518 container_of(ref, struct cgroup_subsys_state, refcnt);
4520 INIT_WORK(&css->destroy_work, css_release_work_fn);
4521 queue_work(cgroup_destroy_wq, &css->destroy_work);
4524 static void init_and_link_css(struct cgroup_subsys_state *css,
4525 struct cgroup_subsys *ss, struct cgroup *cgrp)
4527 lockdep_assert_held(&cgroup_mutex);
4531 memset(css, 0, sizeof(*css));
4534 INIT_LIST_HEAD(&css->sibling);
4535 INIT_LIST_HEAD(&css->children);
4536 css->serial_nr = css_serial_nr_next++;
4538 if (cgroup_parent(cgrp)) {
4539 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4540 css_get(css->parent);
4543 BUG_ON(cgroup_css(cgrp, ss));
4546 /* invoke ->css_online() on a new CSS and mark it online if successful */
4547 static int online_css(struct cgroup_subsys_state *css)
4549 struct cgroup_subsys *ss = css->ss;
4552 lockdep_assert_held(&cgroup_mutex);
4555 ret = ss->css_online(css);
4557 css->flags |= CSS_ONLINE;
4558 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4563 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4564 static void offline_css(struct cgroup_subsys_state *css)
4566 struct cgroup_subsys *ss = css->ss;
4568 lockdep_assert_held(&cgroup_mutex);
4570 if (!(css->flags & CSS_ONLINE))
4573 if (ss->css_offline)
4574 ss->css_offline(css);
4576 css->flags &= ~CSS_ONLINE;
4577 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4579 wake_up_all(&css->cgroup->offline_waitq);
4583 * create_css - create a cgroup_subsys_state
4584 * @cgrp: the cgroup new css will be associated with
4585 * @ss: the subsys of new css
4586 * @visible: whether to create control knobs for the new css or not
4588 * Create a new css associated with @cgrp - @ss pair. On success, the new
4589 * css is online and installed in @cgrp with all interface files created if
4590 * @visible. Returns 0 on success, -errno on failure.
4592 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4595 struct cgroup *parent = cgroup_parent(cgrp);
4596 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4597 struct cgroup_subsys_state *css;
4600 lockdep_assert_held(&cgroup_mutex);
4602 css = ss->css_alloc(parent_css);
4604 return PTR_ERR(css);
4606 init_and_link_css(css, ss, cgrp);
4608 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4612 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4614 goto err_free_percpu_ref;
4618 err = cgroup_populate_dir(cgrp, 1 << ss->id);
4623 /* @css is ready to be brought online now, make it visible */
4624 list_add_tail_rcu(&css->sibling, &parent_css->children);
4625 cgroup_idr_replace(&ss->css_idr, css, css->id);
4627 err = online_css(css);
4631 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4632 cgroup_parent(parent)) {
4633 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4634 current->comm, current->pid, ss->name);
4635 if (!strcmp(ss->name, "memory"))
4636 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4637 ss->warned_broken_hierarchy = true;
4643 list_del_rcu(&css->sibling);
4644 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4646 cgroup_idr_remove(&ss->css_idr, css->id);
4647 err_free_percpu_ref:
4648 percpu_ref_exit(&css->refcnt);
4650 call_rcu(&css->rcu_head, css_free_rcu_fn);
4654 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4657 struct cgroup *parent, *cgrp;
4658 struct cgroup_root *root;
4659 struct cgroup_subsys *ss;
4660 struct kernfs_node *kn;
4661 struct cftype *base_files;
4664 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4666 if (strchr(name, '\n'))
4669 parent = cgroup_kn_lock_live(parent_kn);
4672 root = parent->root;
4674 /* allocate the cgroup and its ID, 0 is reserved for the root */
4675 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4681 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4686 * Temporarily set the pointer to NULL, so idr_find() won't return
4687 * a half-baked cgroup.
4689 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4692 goto out_cancel_ref;
4695 init_cgroup_housekeeping(cgrp);
4697 cgrp->self.parent = &parent->self;
4700 if (notify_on_release(parent))
4701 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4703 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4704 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4706 /* create the directory */
4707 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4715 * This extra ref will be put in cgroup_free_fn() and guarantees
4716 * that @cgrp->kn is always accessible.
4720 cgrp->self.serial_nr = css_serial_nr_next++;
4722 /* allocation complete, commit to creation */
4723 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4724 atomic_inc(&root->nr_cgrps);
4728 * @cgrp is now fully operational. If something fails after this
4729 * point, it'll be released via the normal destruction path.
4731 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4733 ret = cgroup_kn_set_ugid(kn);
4737 if (cgroup_on_dfl(cgrp))
4738 base_files = cgroup_dfl_base_files;
4740 base_files = cgroup_legacy_base_files;
4742 ret = cgroup_addrm_files(cgrp, base_files, true);
4746 /* let's create and online css's */
4747 for_each_subsys(ss, ssid) {
4748 if (parent->child_subsys_mask & (1 << ssid)) {
4749 ret = create_css(cgrp, ss,
4750 parent->subtree_control & (1 << ssid));
4757 * On the default hierarchy, a child doesn't automatically inherit
4758 * subtree_control from the parent. Each is configured manually.
4760 if (!cgroup_on_dfl(cgrp)) {
4761 cgrp->subtree_control = parent->subtree_control;
4762 cgroup_refresh_child_subsys_mask(cgrp);
4765 kernfs_activate(kn);
4771 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4773 percpu_ref_exit(&cgrp->self.refcnt);
4777 cgroup_kn_unlock(parent_kn);
4781 cgroup_destroy_locked(cgrp);
4786 * This is called when the refcnt of a css is confirmed to be killed.
4787 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4788 * initate destruction and put the css ref from kill_css().
4790 static void css_killed_work_fn(struct work_struct *work)
4792 struct cgroup_subsys_state *css =
4793 container_of(work, struct cgroup_subsys_state, destroy_work);
4795 mutex_lock(&cgroup_mutex);
4797 mutex_unlock(&cgroup_mutex);
4802 /* css kill confirmation processing requires process context, bounce */
4803 static void css_killed_ref_fn(struct percpu_ref *ref)
4805 struct cgroup_subsys_state *css =
4806 container_of(ref, struct cgroup_subsys_state, refcnt);
4808 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4809 queue_work(cgroup_destroy_wq, &css->destroy_work);
4813 * kill_css - destroy a css
4814 * @css: css to destroy
4816 * This function initiates destruction of @css by removing cgroup interface
4817 * files and putting its base reference. ->css_offline() will be invoked
4818 * asynchronously once css_tryget_online() is guaranteed to fail and when
4819 * the reference count reaches zero, @css will be released.
4821 static void kill_css(struct cgroup_subsys_state *css)
4823 lockdep_assert_held(&cgroup_mutex);
4826 * This must happen before css is disassociated with its cgroup.
4827 * See seq_css() for details.
4829 cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4832 * Killing would put the base ref, but we need to keep it alive
4833 * until after ->css_offline().
4838 * cgroup core guarantees that, by the time ->css_offline() is
4839 * invoked, no new css reference will be given out via
4840 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4841 * proceed to offlining css's because percpu_ref_kill() doesn't
4842 * guarantee that the ref is seen as killed on all CPUs on return.
4844 * Use percpu_ref_kill_and_confirm() to get notifications as each
4845 * css is confirmed to be seen as killed on all CPUs.
4847 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4851 * cgroup_destroy_locked - the first stage of cgroup destruction
4852 * @cgrp: cgroup to be destroyed
4854 * css's make use of percpu refcnts whose killing latency shouldn't be
4855 * exposed to userland and are RCU protected. Also, cgroup core needs to
4856 * guarantee that css_tryget_online() won't succeed by the time
4857 * ->css_offline() is invoked. To satisfy all the requirements,
4858 * destruction is implemented in the following two steps.
4860 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4861 * userland visible parts and start killing the percpu refcnts of
4862 * css's. Set up so that the next stage will be kicked off once all
4863 * the percpu refcnts are confirmed to be killed.
4865 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4866 * rest of destruction. Once all cgroup references are gone, the
4867 * cgroup is RCU-freed.
4869 * This function implements s1. After this step, @cgrp is gone as far as
4870 * the userland is concerned and a new cgroup with the same name may be
4871 * created. As cgroup doesn't care about the names internally, this
4872 * doesn't cause any problem.
4874 static int cgroup_destroy_locked(struct cgroup *cgrp)
4875 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4877 struct cgroup_subsys_state *css;
4881 lockdep_assert_held(&cgroup_mutex);
4884 * css_set_rwsem synchronizes access to ->cset_links and prevents
4885 * @cgrp from being removed while put_css_set() is in progress.
4887 down_read(&css_set_rwsem);
4888 empty = list_empty(&cgrp->cset_links);
4889 up_read(&css_set_rwsem);
4894 * Make sure there's no live children. We can't test emptiness of
4895 * ->self.children as dead children linger on it while being
4896 * drained; otherwise, "rmdir parent/child parent" may fail.
4898 if (css_has_online_children(&cgrp->self))
4902 * Mark @cgrp dead. This prevents further task migration and child
4903 * creation by disabling cgroup_lock_live_group().
4905 cgrp->self.flags &= ~CSS_ONLINE;
4907 /* initiate massacre of all css's */
4908 for_each_css(css, ssid, cgrp)
4912 * Remove @cgrp directory along with the base files. @cgrp has an
4913 * extra ref on its kn.
4915 kernfs_remove(cgrp->kn);
4917 check_for_release(cgroup_parent(cgrp));
4919 /* put the base reference */
4920 percpu_ref_kill(&cgrp->self.refcnt);
4925 static int cgroup_rmdir(struct kernfs_node *kn)
4927 struct cgroup *cgrp;
4930 cgrp = cgroup_kn_lock_live(kn);
4934 ret = cgroup_destroy_locked(cgrp);
4936 cgroup_kn_unlock(kn);
4940 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4941 .remount_fs = cgroup_remount,
4942 .show_options = cgroup_show_options,
4943 .mkdir = cgroup_mkdir,
4944 .rmdir = cgroup_rmdir,
4945 .rename = cgroup_rename,
4948 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4950 struct cgroup_subsys_state *css;
4952 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4954 mutex_lock(&cgroup_mutex);
4956 idr_init(&ss->css_idr);
4957 INIT_LIST_HEAD(&ss->cfts);
4959 /* Create the root cgroup state for this subsystem */
4960 ss->root = &cgrp_dfl_root;
4961 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4962 /* We don't handle early failures gracefully */
4963 BUG_ON(IS_ERR(css));
4964 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4967 * Root csses are never destroyed and we can't initialize
4968 * percpu_ref during early init. Disable refcnting.
4970 css->flags |= CSS_NO_REF;
4973 /* allocation can't be done safely during early init */
4976 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4977 BUG_ON(css->id < 0);
4980 /* Update the init_css_set to contain a subsys
4981 * pointer to this state - since the subsystem is
4982 * newly registered, all tasks and hence the
4983 * init_css_set is in the subsystem's root cgroup. */
4984 init_css_set.subsys[ss->id] = css;
4986 have_fork_callback |= (bool)ss->fork << ss->id;
4987 have_exit_callback |= (bool)ss->exit << ss->id;
4988 have_canfork_callback |= (bool)ss->can_fork << ss->id;
4990 /* At system boot, before all subsystems have been
4991 * registered, no tasks have been forked, so we don't
4992 * need to invoke fork callbacks here. */
4993 BUG_ON(!list_empty(&init_task.tasks));
4995 BUG_ON(online_css(css));
4997 mutex_unlock(&cgroup_mutex);
5001 * cgroup_init_early - cgroup initialization at system boot
5003 * Initialize cgroups at system boot, and initialize any
5004 * subsystems that request early init.
5006 int __init cgroup_init_early(void)
5008 static struct cgroup_sb_opts __initdata opts;
5009 struct cgroup_subsys *ss;
5012 init_cgroup_root(&cgrp_dfl_root, &opts);
5013 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5015 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5017 for_each_subsys(ss, i) {
5018 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5019 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5020 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5022 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5023 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5026 ss->name = cgroup_subsys_name[i];
5027 if (!ss->legacy_name)
5028 ss->legacy_name = cgroup_subsys_name[i];
5031 cgroup_init_subsys(ss, true);
5037 * cgroup_init - cgroup initialization
5039 * Register cgroup filesystem and /proc file, and initialize
5040 * any subsystems that didn't request early init.
5042 int __init cgroup_init(void)
5044 struct cgroup_subsys *ss;
5048 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5049 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5050 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5052 mutex_lock(&cgroup_mutex);
5054 /* Add init_css_set to the hash table */
5055 key = css_set_hash(init_css_set.subsys);
5056 hash_add(css_set_table, &init_css_set.hlist, key);
5058 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5060 mutex_unlock(&cgroup_mutex);
5062 for_each_subsys(ss, ssid) {
5063 if (ss->early_init) {
5064 struct cgroup_subsys_state *css =
5065 init_css_set.subsys[ss->id];
5067 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5069 BUG_ON(css->id < 0);
5071 cgroup_init_subsys(ss, false);
5074 list_add_tail(&init_css_set.e_cset_node[ssid],
5075 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5078 * Setting dfl_root subsys_mask needs to consider the
5079 * disabled flag and cftype registration needs kmalloc,
5080 * both of which aren't available during early_init.
5085 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5087 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5088 ss->dfl_cftypes = ss->legacy_cftypes;
5090 if (!ss->dfl_cftypes)
5091 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5093 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5094 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5096 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5097 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5101 ss->bind(init_css_set.subsys[ssid]);
5104 err = sysfs_create_mount_point(fs_kobj, "cgroup");
5108 err = register_filesystem(&cgroup_fs_type);
5110 sysfs_remove_mount_point(fs_kobj, "cgroup");
5114 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5118 static int __init cgroup_wq_init(void)
5121 * There isn't much point in executing destruction path in
5122 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5123 * Use 1 for @max_active.
5125 * We would prefer to do this in cgroup_init() above, but that
5126 * is called before init_workqueues(): so leave this until after.
5128 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5129 BUG_ON(!cgroup_destroy_wq);
5132 * Used to destroy pidlists and separate to serve as flush domain.
5133 * Cap @max_active to 1 too.
5135 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5137 BUG_ON(!cgroup_pidlist_destroy_wq);
5141 core_initcall(cgroup_wq_init);
5144 * proc_cgroup_show()
5145 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5146 * - Used for /proc/<pid>/cgroup.
5148 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5149 struct pid *pid, struct task_struct *tsk)
5153 struct cgroup_root *root;
5156 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5160 mutex_lock(&cgroup_mutex);
5161 down_read(&css_set_rwsem);
5163 for_each_root(root) {
5164 struct cgroup_subsys *ss;
5165 struct cgroup *cgrp;
5166 int ssid, count = 0;
5168 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5171 seq_printf(m, "%d:", root->hierarchy_id);
5172 if (root != &cgrp_dfl_root)
5173 for_each_subsys(ss, ssid)
5174 if (root->subsys_mask & (1 << ssid))
5175 seq_printf(m, "%s%s", count++ ? "," : "",
5177 if (strlen(root->name))
5178 seq_printf(m, "%sname=%s", count ? "," : "",
5181 cgrp = task_cgroup_from_root(tsk, root);
5182 path = cgroup_path(cgrp, buf, PATH_MAX);
5184 retval = -ENAMETOOLONG;
5193 up_read(&css_set_rwsem);
5194 mutex_unlock(&cgroup_mutex);
5200 /* Display information about each subsystem and each hierarchy */
5201 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5203 struct cgroup_subsys *ss;
5206 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5208 * ideally we don't want subsystems moving around while we do this.
5209 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5210 * subsys/hierarchy state.
5212 mutex_lock(&cgroup_mutex);
5214 for_each_subsys(ss, i)
5215 seq_printf(m, "%s\t%d\t%d\t%d\n",
5216 ss->legacy_name, ss->root->hierarchy_id,
5217 atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5219 mutex_unlock(&cgroup_mutex);
5223 static int cgroupstats_open(struct inode *inode, struct file *file)
5225 return single_open(file, proc_cgroupstats_show, NULL);
5228 static const struct file_operations proc_cgroupstats_operations = {
5229 .open = cgroupstats_open,
5231 .llseek = seq_lseek,
5232 .release = single_release,
5235 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5237 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5238 return &ss_priv[i - CGROUP_CANFORK_START];
5242 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5244 void **private = subsys_canfork_priv_p(ss_priv, i);
5245 return private ? *private : NULL;
5249 * cgroup_fork - initialize cgroup related fields during copy_process()
5250 * @child: pointer to task_struct of forking parent process.
5252 * A task is associated with the init_css_set until cgroup_post_fork()
5253 * attaches it to the parent's css_set. Empty cg_list indicates that
5254 * @child isn't holding reference to its css_set.
5256 void cgroup_fork(struct task_struct *child)
5258 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5259 INIT_LIST_HEAD(&child->cg_list);
5263 * cgroup_can_fork - called on a new task before the process is exposed
5264 * @child: the task in question.
5266 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5267 * returns an error, the fork aborts with that error code. This allows for
5268 * a cgroup subsystem to conditionally allow or deny new forks.
5270 int cgroup_can_fork(struct task_struct *child,
5271 void *ss_priv[CGROUP_CANFORK_COUNT])
5273 struct cgroup_subsys *ss;
5276 for_each_subsys_which(ss, i, &have_canfork_callback) {
5277 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5285 for_each_subsys(ss, j) {
5288 if (ss->cancel_fork)
5289 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5296 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5297 * @child: the task in question
5299 * This calls the cancel_fork() callbacks if a fork failed *after*
5300 * cgroup_can_fork() succeded.
5302 void cgroup_cancel_fork(struct task_struct *child,
5303 void *ss_priv[CGROUP_CANFORK_COUNT])
5305 struct cgroup_subsys *ss;
5308 for_each_subsys(ss, i)
5309 if (ss->cancel_fork)
5310 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5314 * cgroup_post_fork - called on a new task after adding it to the task list
5315 * @child: the task in question
5317 * Adds the task to the list running through its css_set if necessary and
5318 * call the subsystem fork() callbacks. Has to be after the task is
5319 * visible on the task list in case we race with the first call to
5320 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5323 void cgroup_post_fork(struct task_struct *child,
5324 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5326 struct cgroup_subsys *ss;
5330 * This may race against cgroup_enable_task_cg_lists(). As that
5331 * function sets use_task_css_set_links before grabbing
5332 * tasklist_lock and we just went through tasklist_lock to add
5333 * @child, it's guaranteed that either we see the set
5334 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5335 * @child during its iteration.
5337 * If we won the race, @child is associated with %current's
5338 * css_set. Grabbing css_set_rwsem guarantees both that the
5339 * association is stable, and, on completion of the parent's
5340 * migration, @child is visible in the source of migration or
5341 * already in the destination cgroup. This guarantee is necessary
5342 * when implementing operations which need to migrate all tasks of
5343 * a cgroup to another.
5345 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5346 * will remain in init_css_set. This is safe because all tasks are
5347 * in the init_css_set before cg_links is enabled and there's no
5348 * operation which transfers all tasks out of init_css_set.
5350 if (use_task_css_set_links) {
5351 struct css_set *cset;
5353 down_write(&css_set_rwsem);
5354 cset = task_css_set(current);
5355 if (list_empty(&child->cg_list)) {
5356 rcu_assign_pointer(child->cgroups, cset);
5357 list_add(&child->cg_list, &cset->tasks);
5360 up_write(&css_set_rwsem);
5364 * Call ss->fork(). This must happen after @child is linked on
5365 * css_set; otherwise, @child might change state between ->fork()
5366 * and addition to css_set.
5368 for_each_subsys_which(ss, i, &have_fork_callback)
5369 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5373 * cgroup_exit - detach cgroup from exiting task
5374 * @tsk: pointer to task_struct of exiting process
5376 * Description: Detach cgroup from @tsk and release it.
5378 * Note that cgroups marked notify_on_release force every task in
5379 * them to take the global cgroup_mutex mutex when exiting.
5380 * This could impact scaling on very large systems. Be reluctant to
5381 * use notify_on_release cgroups where very high task exit scaling
5382 * is required on large systems.
5384 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5385 * call cgroup_exit() while the task is still competent to handle
5386 * notify_on_release(), then leave the task attached to the root cgroup in
5387 * each hierarchy for the remainder of its exit. No need to bother with
5388 * init_css_set refcnting. init_css_set never goes away and we can't race
5389 * with migration path - PF_EXITING is visible to migration path.
5391 void cgroup_exit(struct task_struct *tsk)
5393 struct cgroup_subsys *ss;
5394 struct css_set *cset;
5395 bool put_cset = false;
5399 * Unlink from @tsk from its css_set. As migration path can't race
5400 * with us, we can check cg_list without grabbing css_set_rwsem.
5402 if (!list_empty(&tsk->cg_list)) {
5403 down_write(&css_set_rwsem);
5404 list_del_init(&tsk->cg_list);
5405 up_write(&css_set_rwsem);
5409 /* Reassign the task to the init_css_set. */
5410 cset = task_css_set(tsk);
5411 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5413 /* see cgroup_post_fork() for details */
5414 for_each_subsys_which(ss, i, &have_exit_callback) {
5415 struct cgroup_subsys_state *old_css = cset->subsys[i];
5416 struct cgroup_subsys_state *css = task_css(tsk, i);
5418 ss->exit(css, old_css, tsk);
5425 static void check_for_release(struct cgroup *cgrp)
5427 if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5428 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5429 schedule_work(&cgrp->release_agent_work);
5433 * Notify userspace when a cgroup is released, by running the
5434 * configured release agent with the name of the cgroup (path
5435 * relative to the root of cgroup file system) as the argument.
5437 * Most likely, this user command will try to rmdir this cgroup.
5439 * This races with the possibility that some other task will be
5440 * attached to this cgroup before it is removed, or that some other
5441 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5442 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5443 * unused, and this cgroup will be reprieved from its death sentence,
5444 * to continue to serve a useful existence. Next time it's released,
5445 * we will get notified again, if it still has 'notify_on_release' set.
5447 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5448 * means only wait until the task is successfully execve()'d. The
5449 * separate release agent task is forked by call_usermodehelper(),
5450 * then control in this thread returns here, without waiting for the
5451 * release agent task. We don't bother to wait because the caller of
5452 * this routine has no use for the exit status of the release agent
5453 * task, so no sense holding our caller up for that.
5455 static void cgroup_release_agent(struct work_struct *work)
5457 struct cgroup *cgrp =
5458 container_of(work, struct cgroup, release_agent_work);
5459 char *pathbuf = NULL, *agentbuf = NULL, *path;
5460 char *argv[3], *envp[3];
5462 mutex_lock(&cgroup_mutex);
5464 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5465 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5466 if (!pathbuf || !agentbuf)
5469 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5477 /* minimal command environment */
5479 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5482 mutex_unlock(&cgroup_mutex);
5483 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5486 mutex_unlock(&cgroup_mutex);
5492 static int __init cgroup_disable(char *str)
5494 struct cgroup_subsys *ss;
5498 while ((token = strsep(&str, ",")) != NULL) {
5502 for_each_subsys(ss, i) {
5503 if (strcmp(token, ss->name) &&
5504 strcmp(token, ss->legacy_name))
5508 printk(KERN_INFO "Disabling %s control group subsystem\n",
5515 __setup("cgroup_disable=", cgroup_disable);
5517 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5519 printk("cgroup: using legacy files on the default hierarchy\n");
5520 cgroup_legacy_files_on_dfl = true;
5523 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5526 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5527 * @dentry: directory dentry of interest
5528 * @ss: subsystem of interest
5530 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5531 * to get the corresponding css and return it. If such css doesn't exist
5532 * or can't be pinned, an ERR_PTR value is returned.
5534 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5535 struct cgroup_subsys *ss)
5537 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5538 struct cgroup_subsys_state *css = NULL;
5539 struct cgroup *cgrp;
5541 /* is @dentry a cgroup dir? */
5542 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5543 kernfs_type(kn) != KERNFS_DIR)
5544 return ERR_PTR(-EBADF);
5549 * This path doesn't originate from kernfs and @kn could already
5550 * have been or be removed at any point. @kn->priv is RCU
5551 * protected for this access. See css_release_work_fn() for details.
5553 cgrp = rcu_dereference(kn->priv);
5555 css = cgroup_css(cgrp, ss);
5557 if (!css || !css_tryget_online(css))
5558 css = ERR_PTR(-ENOENT);
5565 * css_from_id - lookup css by id
5566 * @id: the cgroup id
5567 * @ss: cgroup subsys to be looked into
5569 * Returns the css if there's valid one with @id, otherwise returns NULL.
5570 * Should be called under rcu_read_lock().
5572 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5574 WARN_ON_ONCE(!rcu_read_lock_held());
5575 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5578 #ifdef CONFIG_CGROUP_DEBUG
5579 static struct cgroup_subsys_state *
5580 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5582 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5585 return ERR_PTR(-ENOMEM);
5590 static void debug_css_free(struct cgroup_subsys_state *css)
5595 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5598 return cgroup_task_count(css->cgroup);
5601 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5604 return (u64)(unsigned long)current->cgroups;
5607 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5613 count = atomic_read(&task_css_set(current)->refcount);
5618 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5620 struct cgrp_cset_link *link;
5621 struct css_set *cset;
5624 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5628 down_read(&css_set_rwsem);
5630 cset = rcu_dereference(current->cgroups);
5631 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5632 struct cgroup *c = link->cgrp;
5634 cgroup_name(c, name_buf, NAME_MAX + 1);
5635 seq_printf(seq, "Root %d group %s\n",
5636 c->root->hierarchy_id, name_buf);
5639 up_read(&css_set_rwsem);
5644 #define MAX_TASKS_SHOWN_PER_CSS 25
5645 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5647 struct cgroup_subsys_state *css = seq_css(seq);
5648 struct cgrp_cset_link *link;
5650 down_read(&css_set_rwsem);
5651 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5652 struct css_set *cset = link->cset;
5653 struct task_struct *task;
5656 seq_printf(seq, "css_set %p\n", cset);
5658 list_for_each_entry(task, &cset->tasks, cg_list) {
5659 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5661 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5664 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5665 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5667 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5671 seq_puts(seq, " ...\n");
5673 up_read(&css_set_rwsem);
5677 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5679 return (!cgroup_has_tasks(css->cgroup) &&
5680 !css_has_online_children(&css->cgroup->self));
5683 static struct cftype debug_files[] = {
5685 .name = "taskcount",
5686 .read_u64 = debug_taskcount_read,
5690 .name = "current_css_set",
5691 .read_u64 = current_css_set_read,
5695 .name = "current_css_set_refcount",
5696 .read_u64 = current_css_set_refcount_read,
5700 .name = "current_css_set_cg_links",
5701 .seq_show = current_css_set_cg_links_read,
5705 .name = "cgroup_css_links",
5706 .seq_show = cgroup_css_links_read,
5710 .name = "releasable",
5711 .read_u64 = releasable_read,
5717 struct cgroup_subsys debug_cgrp_subsys = {
5718 .css_alloc = debug_css_alloc,
5719 .css_free = debug_css_free,
5720 .legacy_cftypes = debug_files,
5722 #endif /* CONFIG_CGROUP_DEBUG */