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>
142 /* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
144 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \
145 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \
146 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \
147 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
148 #include <linux/cgroup_subsys.h>
151 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
152 static struct static_key_true *cgroup_subsys_enabled_key[] = {
153 #include <linux/cgroup_subsys.h>
157 #define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
158 static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
159 #include <linux/cgroup_subsys.h>
164 * The default hierarchy, reserved for the subsystems that are otherwise
165 * unattached - it never has more than a single cgroup, and all tasks are
166 * part of that cgroup.
168 struct cgroup_root cgrp_dfl_root;
169 EXPORT_SYMBOL_GPL(cgrp_dfl_root);
172 * The default hierarchy always exists but is hidden until mounted for the
173 * first time. This is for backward compatibility.
175 static bool cgrp_dfl_root_visible;
178 * Set by the boot param of the same name and makes subsystems with NULL
179 * ->dfl_files to use ->legacy_files on the default hierarchy.
181 static bool cgroup_legacy_files_on_dfl;
183 /* some controllers are not supported in the default hierarchy */
184 static unsigned long cgrp_dfl_root_inhibit_ss_mask;
186 /* The list of hierarchy roots */
188 static LIST_HEAD(cgroup_roots);
189 static int cgroup_root_count;
191 /* hierarchy ID allocation and mapping, protected by cgroup_mutex */
192 static DEFINE_IDR(cgroup_hierarchy_idr);
195 * Assign a monotonically increasing serial number to csses. It guarantees
196 * cgroups with bigger numbers are newer than those with smaller numbers.
197 * Also, as csses are always appended to the parent's ->children list, it
198 * guarantees that sibling csses are always sorted in the ascending serial
199 * number order on the list. Protected by cgroup_mutex.
201 static u64 css_serial_nr_next = 1;
204 * These bitmask flags indicate whether tasks in the fork and exit paths have
205 * fork/exit handlers to call. This avoids us having to do extra work in the
206 * fork/exit path to check which subsystems have fork/exit callbacks.
208 static unsigned long have_fork_callback __read_mostly;
209 static unsigned long have_exit_callback __read_mostly;
211 /* Ditto for the can_fork callback. */
212 static unsigned long have_canfork_callback __read_mostly;
214 static struct cftype cgroup_dfl_base_files[];
215 static struct cftype cgroup_legacy_base_files[];
217 static int rebind_subsystems(struct cgroup_root *dst_root,
218 unsigned long ss_mask);
219 static int cgroup_destroy_locked(struct cgroup *cgrp);
220 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
222 static void css_release(struct percpu_ref *ref);
223 static void kill_css(struct cgroup_subsys_state *css);
224 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
225 struct cgroup *cgrp, struct cftype cfts[],
229 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
230 * @ssid: subsys ID of interest
232 * cgroup_subsys_enabled() can only be used with literal subsys names which
233 * is fine for individual subsystems but unsuitable for cgroup core. This
234 * is slower static_key_enabled() based test indexed by @ssid.
236 static bool cgroup_ssid_enabled(int ssid)
238 return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
242 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
243 * @cgrp: the cgroup of interest
245 * The default hierarchy is the v2 interface of cgroup and this function
246 * can be used to test whether a cgroup is on the default hierarchy for
247 * cases where a subsystem should behave differnetly depending on the
250 * The set of behaviors which change on the default hierarchy are still
251 * being determined and the mount option is prefixed with __DEVEL__.
253 * List of changed behaviors:
255 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
256 * and "name" are disallowed.
258 * - When mounting an existing superblock, mount options should match.
260 * - Remount is disallowed.
262 * - rename(2) is disallowed.
264 * - "tasks" is removed. Everything should be at process granularity. Use
265 * "cgroup.procs" instead.
267 * - "cgroup.procs" is not sorted. pids will be unique unless they got
268 * recycled inbetween reads.
270 * - "release_agent" and "notify_on_release" are removed. Replacement
271 * notification mechanism will be implemented.
273 * - "cgroup.clone_children" is removed.
275 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup
276 * and its descendants contain no task; otherwise, 1. The file also
277 * generates kernfs notification which can be monitored through poll and
278 * [di]notify when the value of the file changes.
280 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
281 * take masks of ancestors with non-empty cpus/mems, instead of being
282 * moved to an ancestor.
284 * - cpuset: a task can be moved into an empty cpuset, and again it takes
285 * masks of ancestors.
287 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
290 * - blkcg: blk-throttle becomes properly hierarchical.
292 * - debug: disallowed on the default hierarchy.
294 static bool cgroup_on_dfl(const struct cgroup *cgrp)
296 return cgrp->root == &cgrp_dfl_root;
299 /* IDR wrappers which synchronize using cgroup_idr_lock */
300 static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
305 idr_preload(gfp_mask);
306 spin_lock_bh(&cgroup_idr_lock);
307 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_WAIT);
308 spin_unlock_bh(&cgroup_idr_lock);
313 static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
317 spin_lock_bh(&cgroup_idr_lock);
318 ret = idr_replace(idr, ptr, id);
319 spin_unlock_bh(&cgroup_idr_lock);
323 static void cgroup_idr_remove(struct idr *idr, int id)
325 spin_lock_bh(&cgroup_idr_lock);
327 spin_unlock_bh(&cgroup_idr_lock);
330 static struct cgroup *cgroup_parent(struct cgroup *cgrp)
332 struct cgroup_subsys_state *parent_css = cgrp->self.parent;
335 return container_of(parent_css, struct cgroup, self);
340 * cgroup_css - obtain a cgroup's css for the specified subsystem
341 * @cgrp: the cgroup of interest
342 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
344 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
345 * function must be called either under cgroup_mutex or rcu_read_lock() and
346 * the caller is responsible for pinning the returned css if it wants to
347 * keep accessing it outside the said locks. This function may return
348 * %NULL if @cgrp doesn't have @subsys_id enabled.
350 static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
351 struct cgroup_subsys *ss)
354 return rcu_dereference_check(cgrp->subsys[ss->id],
355 lockdep_is_held(&cgroup_mutex));
361 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
362 * @cgrp: the cgroup of interest
363 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
365 * Similar to cgroup_css() but returns the effective css, which is defined
366 * as the matching css of the nearest ancestor including self which has @ss
367 * enabled. If @ss is associated with the hierarchy @cgrp is on, this
368 * function is guaranteed to return non-NULL css.
370 static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
371 struct cgroup_subsys *ss)
373 lockdep_assert_held(&cgroup_mutex);
378 if (!(cgrp->root->subsys_mask & (1 << ss->id)))
382 * This function is used while updating css associations and thus
383 * can't test the csses directly. Use ->child_subsys_mask.
385 while (cgroup_parent(cgrp) &&
386 !(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
387 cgrp = cgroup_parent(cgrp);
389 return cgroup_css(cgrp, ss);
393 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
394 * @cgrp: the cgroup of interest
395 * @ss: the subsystem of interest
397 * Find and get the effective css of @cgrp for @ss. The effective css is
398 * defined as the matching css of the nearest ancestor including self which
399 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
400 * the root css is returned, so this function always returns a valid css.
401 * The returned css must be put using css_put().
403 struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
404 struct cgroup_subsys *ss)
406 struct cgroup_subsys_state *css;
411 css = cgroup_css(cgrp, ss);
413 if (css && css_tryget_online(css))
415 cgrp = cgroup_parent(cgrp);
418 css = init_css_set.subsys[ss->id];
425 /* convenient tests for these bits */
426 static inline bool cgroup_is_dead(const struct cgroup *cgrp)
428 return !(cgrp->self.flags & CSS_ONLINE);
431 struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
433 struct cgroup *cgrp = of->kn->parent->priv;
434 struct cftype *cft = of_cft(of);
437 * This is open and unprotected implementation of cgroup_css().
438 * seq_css() is only called from a kernfs file operation which has
439 * an active reference on the file. Because all the subsystem
440 * files are drained before a css is disassociated with a cgroup,
441 * the matching css from the cgroup's subsys table is guaranteed to
442 * be and stay valid until the enclosing operation is complete.
445 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
449 EXPORT_SYMBOL_GPL(of_css);
452 * cgroup_is_descendant - test ancestry
453 * @cgrp: the cgroup to be tested
454 * @ancestor: possible ancestor of @cgrp
456 * Test whether @cgrp is a descendant of @ancestor. It also returns %true
457 * if @cgrp == @ancestor. This function is safe to call as long as @cgrp
458 * and @ancestor are accessible.
460 bool cgroup_is_descendant(struct cgroup *cgrp, struct cgroup *ancestor)
463 if (cgrp == ancestor)
465 cgrp = cgroup_parent(cgrp);
470 static int notify_on_release(const struct cgroup *cgrp)
472 return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
476 * for_each_css - iterate all css's of a cgroup
477 * @css: the iteration cursor
478 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
479 * @cgrp: the target cgroup to iterate css's of
481 * Should be called under cgroup_[tree_]mutex.
483 #define for_each_css(css, ssid, cgrp) \
484 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
485 if (!((css) = rcu_dereference_check( \
486 (cgrp)->subsys[(ssid)], \
487 lockdep_is_held(&cgroup_mutex)))) { } \
491 * for_each_e_css - iterate all effective css's of a cgroup
492 * @css: the iteration cursor
493 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
494 * @cgrp: the target cgroup to iterate css's of
496 * Should be called under cgroup_[tree_]mutex.
498 #define for_each_e_css(css, ssid, cgrp) \
499 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
500 if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
505 * for_each_subsys - iterate all enabled cgroup subsystems
506 * @ss: the iteration cursor
507 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
509 #define for_each_subsys(ss, ssid) \
510 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
511 (((ss) = cgroup_subsys[ssid]) || true); (ssid)++)
514 * for_each_subsys_which - filter for_each_subsys with a bitmask
515 * @ss: the iteration cursor
516 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
517 * @ss_maskp: a pointer to the bitmask
519 * The block will only run for cases where the ssid-th bit (1 << ssid) of
522 #define for_each_subsys_which(ss, ssid, ss_maskp) \
523 if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
526 for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
527 if (((ss) = cgroup_subsys[ssid]) && false) \
531 /* iterate across the hierarchies */
532 #define for_each_root(root) \
533 list_for_each_entry((root), &cgroup_roots, root_list)
535 /* iterate over child cgrps, lock should be held throughout iteration */
536 #define cgroup_for_each_live_child(child, cgrp) \
537 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
538 if (({ lockdep_assert_held(&cgroup_mutex); \
539 cgroup_is_dead(child); })) \
543 static void cgroup_release_agent(struct work_struct *work);
544 static void check_for_release(struct cgroup *cgrp);
547 * A cgroup can be associated with multiple css_sets as different tasks may
548 * belong to different cgroups on different hierarchies. In the other
549 * direction, a css_set is naturally associated with multiple cgroups.
550 * This M:N relationship is represented by the following link structure
551 * which exists for each association and allows traversing the associations
554 struct cgrp_cset_link {
555 /* the cgroup and css_set this link associates */
557 struct css_set *cset;
559 /* list of cgrp_cset_links anchored at cgrp->cset_links */
560 struct list_head cset_link;
562 /* list of cgrp_cset_links anchored at css_set->cgrp_links */
563 struct list_head cgrp_link;
567 * The default css_set - used by init and its children prior to any
568 * hierarchies being mounted. It contains a pointer to the root state
569 * for each subsystem. Also used to anchor the list of css_sets. Not
570 * reference-counted, to improve performance when child cgroups
571 * haven't been created.
573 struct css_set init_css_set = {
574 .refcount = ATOMIC_INIT(1),
575 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
576 .tasks = LIST_HEAD_INIT(init_css_set.tasks),
577 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
578 .mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
579 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
582 static int css_set_count = 1; /* 1 for init_css_set */
585 * css_set_populated - does a css_set contain any tasks?
586 * @cset: target css_set
588 static bool css_set_populated(struct css_set *cset)
590 lockdep_assert_held(&css_set_rwsem);
592 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
596 * cgroup_update_populated - updated populated count of a cgroup
597 * @cgrp: the target cgroup
598 * @populated: inc or dec populated count
600 * One of the css_sets associated with @cgrp is either getting its first
601 * task or losing the last. Update @cgrp->populated_cnt accordingly. The
602 * count is propagated towards root so that a given cgroup's populated_cnt
603 * is zero iff the cgroup and all its descendants don't contain any tasks.
605 * @cgrp's interface file "cgroup.populated" is zero if
606 * @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
607 * changes from or to zero, userland is notified that the content of the
608 * interface file has changed. This can be used to detect when @cgrp and
609 * its descendants become populated or empty.
611 static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
613 lockdep_assert_held(&css_set_rwsem);
619 trigger = !cgrp->populated_cnt++;
621 trigger = !--cgrp->populated_cnt;
626 check_for_release(cgrp);
627 cgroup_file_notify(&cgrp->events_file);
629 cgrp = cgroup_parent(cgrp);
634 * css_set_update_populated - update populated state of a css_set
635 * @cset: target css_set
636 * @populated: whether @cset is populated or depopulated
638 * @cset is either getting the first task or losing the last. Update the
639 * ->populated_cnt of all associated cgroups accordingly.
641 static void css_set_update_populated(struct css_set *cset, bool populated)
643 struct cgrp_cset_link *link;
645 lockdep_assert_held(&css_set_rwsem);
647 list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
648 cgroup_update_populated(link->cgrp, populated);
652 * hash table for cgroup groups. This improves the performance to find
653 * an existing css_set. This hash doesn't (currently) take into
654 * account cgroups in empty hierarchies.
656 #define CSS_SET_HASH_BITS 7
657 static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
659 static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
661 unsigned long key = 0UL;
662 struct cgroup_subsys *ss;
665 for_each_subsys(ss, i)
666 key += (unsigned long)css[i];
667 key = (key >> 16) ^ key;
672 static void put_css_set_locked(struct css_set *cset)
674 struct cgrp_cset_link *link, *tmp_link;
675 struct cgroup_subsys *ss;
678 lockdep_assert_held(&css_set_rwsem);
680 if (!atomic_dec_and_test(&cset->refcount))
683 /* This css_set is dead. unlink it and release cgroup refcounts */
684 for_each_subsys(ss, ssid)
685 list_del(&cset->e_cset_node[ssid]);
686 hash_del(&cset->hlist);
689 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
690 list_del(&link->cset_link);
691 list_del(&link->cgrp_link);
695 kfree_rcu(cset, rcu_head);
698 static void put_css_set(struct css_set *cset)
701 * Ensure that the refcount doesn't hit zero while any readers
702 * can see it. Similar to atomic_dec_and_lock(), but for an
705 if (atomic_add_unless(&cset->refcount, -1, 1))
708 down_write(&css_set_rwsem);
709 put_css_set_locked(cset);
710 up_write(&css_set_rwsem);
714 * refcounted get/put for css_set objects
716 static inline void get_css_set(struct css_set *cset)
718 atomic_inc(&cset->refcount);
722 * compare_css_sets - helper function for find_existing_css_set().
723 * @cset: candidate css_set being tested
724 * @old_cset: existing css_set for a task
725 * @new_cgrp: cgroup that's being entered by the task
726 * @template: desired set of css pointers in css_set (pre-calculated)
728 * Returns true if "cset" matches "old_cset" except for the hierarchy
729 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
731 static bool compare_css_sets(struct css_set *cset,
732 struct css_set *old_cset,
733 struct cgroup *new_cgrp,
734 struct cgroup_subsys_state *template[])
736 struct list_head *l1, *l2;
739 * On the default hierarchy, there can be csets which are
740 * associated with the same set of cgroups but different csses.
741 * Let's first ensure that csses match.
743 if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
747 * Compare cgroup pointers in order to distinguish between
748 * different cgroups in hierarchies. As different cgroups may
749 * share the same effective css, this comparison is always
752 l1 = &cset->cgrp_links;
753 l2 = &old_cset->cgrp_links;
755 struct cgrp_cset_link *link1, *link2;
756 struct cgroup *cgrp1, *cgrp2;
760 /* See if we reached the end - both lists are equal length. */
761 if (l1 == &cset->cgrp_links) {
762 BUG_ON(l2 != &old_cset->cgrp_links);
765 BUG_ON(l2 == &old_cset->cgrp_links);
767 /* Locate the cgroups associated with these links. */
768 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
769 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
772 /* Hierarchies should be linked in the same order. */
773 BUG_ON(cgrp1->root != cgrp2->root);
776 * If this hierarchy is the hierarchy of the cgroup
777 * that's changing, then we need to check that this
778 * css_set points to the new cgroup; if it's any other
779 * hierarchy, then this css_set should point to the
780 * same cgroup as the old css_set.
782 if (cgrp1->root == new_cgrp->root) {
783 if (cgrp1 != new_cgrp)
794 * find_existing_css_set - init css array and find the matching css_set
795 * @old_cset: the css_set that we're using before the cgroup transition
796 * @cgrp: the cgroup that we're moving into
797 * @template: out param for the new set of csses, should be clear on entry
799 static struct css_set *find_existing_css_set(struct css_set *old_cset,
801 struct cgroup_subsys_state *template[])
803 struct cgroup_root *root = cgrp->root;
804 struct cgroup_subsys *ss;
805 struct css_set *cset;
810 * Build the set of subsystem state objects that we want to see in the
811 * new css_set. while subsystems can change globally, the entries here
812 * won't change, so no need for locking.
814 for_each_subsys(ss, i) {
815 if (root->subsys_mask & (1UL << i)) {
817 * @ss is in this hierarchy, so we want the
818 * effective css from @cgrp.
820 template[i] = cgroup_e_css(cgrp, ss);
823 * @ss is not in this hierarchy, so we don't want
826 template[i] = old_cset->subsys[i];
830 key = css_set_hash(template);
831 hash_for_each_possible(css_set_table, cset, hlist, key) {
832 if (!compare_css_sets(cset, old_cset, cgrp, template))
835 /* This css_set matches what we need */
839 /* No existing cgroup group matched */
843 static void free_cgrp_cset_links(struct list_head *links_to_free)
845 struct cgrp_cset_link *link, *tmp_link;
847 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
848 list_del(&link->cset_link);
854 * allocate_cgrp_cset_links - allocate cgrp_cset_links
855 * @count: the number of links to allocate
856 * @tmp_links: list_head the allocated links are put on
858 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
859 * through ->cset_link. Returns 0 on success or -errno.
861 static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
863 struct cgrp_cset_link *link;
866 INIT_LIST_HEAD(tmp_links);
868 for (i = 0; i < count; i++) {
869 link = kzalloc(sizeof(*link), GFP_KERNEL);
871 free_cgrp_cset_links(tmp_links);
874 list_add(&link->cset_link, tmp_links);
880 * link_css_set - a helper function to link a css_set to a cgroup
881 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
882 * @cset: the css_set to be linked
883 * @cgrp: the destination cgroup
885 static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
888 struct cgrp_cset_link *link;
890 BUG_ON(list_empty(tmp_links));
892 if (cgroup_on_dfl(cgrp))
893 cset->dfl_cgrp = cgrp;
895 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
899 list_move(&link->cset_link, &cgrp->cset_links);
902 * Always add links to the tail of the list so that the list
903 * is sorted by order of hierarchy creation
905 list_add_tail(&link->cgrp_link, &cset->cgrp_links);
909 * find_css_set - return a new css_set with one cgroup updated
910 * @old_cset: the baseline css_set
911 * @cgrp: the cgroup to be updated
913 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
914 * substituted into the appropriate hierarchy.
916 static struct css_set *find_css_set(struct css_set *old_cset,
919 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
920 struct css_set *cset;
921 struct list_head tmp_links;
922 struct cgrp_cset_link *link;
923 struct cgroup_subsys *ss;
927 lockdep_assert_held(&cgroup_mutex);
929 /* First see if we already have a cgroup group that matches
931 down_read(&css_set_rwsem);
932 cset = find_existing_css_set(old_cset, cgrp, template);
935 up_read(&css_set_rwsem);
940 cset = kzalloc(sizeof(*cset), GFP_KERNEL);
944 /* Allocate all the cgrp_cset_link objects that we'll need */
945 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
950 atomic_set(&cset->refcount, 1);
951 INIT_LIST_HEAD(&cset->cgrp_links);
952 INIT_LIST_HEAD(&cset->tasks);
953 INIT_LIST_HEAD(&cset->mg_tasks);
954 INIT_LIST_HEAD(&cset->mg_preload_node);
955 INIT_LIST_HEAD(&cset->mg_node);
956 INIT_HLIST_NODE(&cset->hlist);
958 /* Copy the set of subsystem state objects generated in
959 * find_existing_css_set() */
960 memcpy(cset->subsys, template, sizeof(cset->subsys));
962 down_write(&css_set_rwsem);
963 /* Add reference counts and links from the new css_set. */
964 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
965 struct cgroup *c = link->cgrp;
967 if (c->root == cgrp->root)
969 link_css_set(&tmp_links, cset, c);
972 BUG_ON(!list_empty(&tmp_links));
976 /* Add @cset to the hash table */
977 key = css_set_hash(cset->subsys);
978 hash_add(css_set_table, &cset->hlist, key);
980 for_each_subsys(ss, ssid)
981 list_add_tail(&cset->e_cset_node[ssid],
982 &cset->subsys[ssid]->cgroup->e_csets[ssid]);
984 up_write(&css_set_rwsem);
989 static struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
991 struct cgroup *root_cgrp = kf_root->kn->priv;
993 return root_cgrp->root;
996 static int cgroup_init_root_id(struct cgroup_root *root)
1000 lockdep_assert_held(&cgroup_mutex);
1002 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1006 root->hierarchy_id = id;
1010 static void cgroup_exit_root_id(struct cgroup_root *root)
1012 lockdep_assert_held(&cgroup_mutex);
1014 if (root->hierarchy_id) {
1015 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1016 root->hierarchy_id = 0;
1020 static void cgroup_free_root(struct cgroup_root *root)
1023 /* hierarchy ID should already have been released */
1024 WARN_ON_ONCE(root->hierarchy_id);
1026 idr_destroy(&root->cgroup_idr);
1031 static void cgroup_destroy_root(struct cgroup_root *root)
1033 struct cgroup *cgrp = &root->cgrp;
1034 struct cgrp_cset_link *link, *tmp_link;
1036 mutex_lock(&cgroup_mutex);
1038 BUG_ON(atomic_read(&root->nr_cgrps));
1039 BUG_ON(!list_empty(&cgrp->self.children));
1041 /* Rebind all subsystems back to the default hierarchy */
1042 rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
1045 * Release all the links from cset_links to this hierarchy's
1048 down_write(&css_set_rwsem);
1050 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1051 list_del(&link->cset_link);
1052 list_del(&link->cgrp_link);
1055 up_write(&css_set_rwsem);
1057 if (!list_empty(&root->root_list)) {
1058 list_del(&root->root_list);
1059 cgroup_root_count--;
1062 cgroup_exit_root_id(root);
1064 mutex_unlock(&cgroup_mutex);
1066 kernfs_destroy_root(root->kf_root);
1067 cgroup_free_root(root);
1070 /* look up cgroup associated with given css_set on the specified hierarchy */
1071 static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1072 struct cgroup_root *root)
1074 struct cgroup *res = NULL;
1076 lockdep_assert_held(&cgroup_mutex);
1077 lockdep_assert_held(&css_set_rwsem);
1079 if (cset == &init_css_set) {
1082 struct cgrp_cset_link *link;
1084 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1085 struct cgroup *c = link->cgrp;
1087 if (c->root == root) {
1099 * Return the cgroup for "task" from the given hierarchy. Must be
1100 * called with cgroup_mutex and css_set_rwsem held.
1102 static struct cgroup *task_cgroup_from_root(struct task_struct *task,
1103 struct cgroup_root *root)
1106 * No need to lock the task - since we hold cgroup_mutex the
1107 * task can't change groups, so the only thing that can happen
1108 * is that it exits and its css is set back to init_css_set.
1110 return cset_cgroup_from_root(task_css_set(task), root);
1114 * A task must hold cgroup_mutex to modify cgroups.
1116 * Any task can increment and decrement the count field without lock.
1117 * So in general, code holding cgroup_mutex can't rely on the count
1118 * field not changing. However, if the count goes to zero, then only
1119 * cgroup_attach_task() can increment it again. Because a count of zero
1120 * means that no tasks are currently attached, therefore there is no
1121 * way a task attached to that cgroup can fork (the other way to
1122 * increment the count). So code holding cgroup_mutex can safely
1123 * assume that if the count is zero, it will stay zero. Similarly, if
1124 * a task holds cgroup_mutex on a cgroup with zero count, it
1125 * knows that the cgroup won't be removed, as cgroup_rmdir()
1128 * A cgroup can only be deleted if both its 'count' of using tasks
1129 * is zero, and its list of 'children' cgroups is empty. Since all
1130 * tasks in the system use _some_ cgroup, and since there is always at
1131 * least one task in the system (init, pid == 1), therefore, root cgroup
1132 * always has either children cgroups and/or using tasks. So we don't
1133 * need a special hack to ensure that root cgroup cannot be deleted.
1135 * P.S. One more locking exception. RCU is used to guard the
1136 * update of a tasks cgroup pointer by cgroup_attach_task()
1139 static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1140 static const struct file_operations proc_cgroupstats_operations;
1142 static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1145 struct cgroup_subsys *ss = cft->ss;
1147 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1148 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1149 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1150 cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1153 strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1158 * cgroup_file_mode - deduce file mode of a control file
1159 * @cft: the control file in question
1161 * S_IRUGO for read, S_IWUSR for write.
1163 static umode_t cgroup_file_mode(const struct cftype *cft)
1167 if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1170 if (cft->write_u64 || cft->write_s64 || cft->write) {
1171 if (cft->flags & CFTYPE_WORLD_WRITABLE)
1180 static void cgroup_get(struct cgroup *cgrp)
1182 WARN_ON_ONCE(cgroup_is_dead(cgrp));
1183 css_get(&cgrp->self);
1186 static bool cgroup_tryget(struct cgroup *cgrp)
1188 return css_tryget(&cgrp->self);
1191 static void cgroup_put(struct cgroup *cgrp)
1193 css_put(&cgrp->self);
1197 * cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1198 * @cgrp: the target cgroup
1199 * @subtree_control: the new subtree_control mask to consider
1201 * On the default hierarchy, a subsystem may request other subsystems to be
1202 * enabled together through its ->depends_on mask. In such cases, more
1203 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1205 * This function calculates which subsystems need to be enabled if
1206 * @subtree_control is to be applied to @cgrp. The returned mask is always
1207 * a superset of @subtree_control and follows the usual hierarchy rules.
1209 static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1210 unsigned long subtree_control)
1212 struct cgroup *parent = cgroup_parent(cgrp);
1213 unsigned long cur_ss_mask = subtree_control;
1214 struct cgroup_subsys *ss;
1217 lockdep_assert_held(&cgroup_mutex);
1219 if (!cgroup_on_dfl(cgrp))
1223 unsigned long new_ss_mask = cur_ss_mask;
1225 for_each_subsys_which(ss, ssid, &cur_ss_mask)
1226 new_ss_mask |= ss->depends_on;
1229 * Mask out subsystems which aren't available. This can
1230 * happen only if some depended-upon subsystems were bound
1231 * to non-default hierarchies.
1234 new_ss_mask &= parent->child_subsys_mask;
1236 new_ss_mask &= cgrp->root->subsys_mask;
1238 if (new_ss_mask == cur_ss_mask)
1240 cur_ss_mask = new_ss_mask;
1247 * cgroup_refresh_child_subsys_mask - update child_subsys_mask
1248 * @cgrp: the target cgroup
1250 * Update @cgrp->child_subsys_mask according to the current
1251 * @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1253 static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1255 cgrp->child_subsys_mask =
1256 cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1260 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1261 * @kn: the kernfs_node being serviced
1263 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1264 * the method finishes if locking succeeded. Note that once this function
1265 * returns the cgroup returned by cgroup_kn_lock_live() may become
1266 * inaccessible any time. If the caller intends to continue to access the
1267 * cgroup, it should pin it before invoking this function.
1269 static void cgroup_kn_unlock(struct kernfs_node *kn)
1271 struct cgroup *cgrp;
1273 if (kernfs_type(kn) == KERNFS_DIR)
1276 cgrp = kn->parent->priv;
1278 mutex_unlock(&cgroup_mutex);
1280 kernfs_unbreak_active_protection(kn);
1285 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1286 * @kn: the kernfs_node being serviced
1288 * This helper is to be used by a cgroup kernfs method currently servicing
1289 * @kn. It breaks the active protection, performs cgroup locking and
1290 * verifies that the associated cgroup is alive. Returns the cgroup if
1291 * alive; otherwise, %NULL. A successful return should be undone by a
1292 * matching cgroup_kn_unlock() invocation.
1294 * Any cgroup kernfs method implementation which requires locking the
1295 * associated cgroup should use this helper. It avoids nesting cgroup
1296 * locking under kernfs active protection and allows all kernfs operations
1297 * including self-removal.
1299 static struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn)
1301 struct cgroup *cgrp;
1303 if (kernfs_type(kn) == KERNFS_DIR)
1306 cgrp = kn->parent->priv;
1309 * We're gonna grab cgroup_mutex which nests outside kernfs
1310 * active_ref. cgroup liveliness check alone provides enough
1311 * protection against removal. Ensure @cgrp stays accessible and
1312 * break the active_ref protection.
1314 if (!cgroup_tryget(cgrp))
1316 kernfs_break_active_protection(kn);
1318 mutex_lock(&cgroup_mutex);
1320 if (!cgroup_is_dead(cgrp))
1323 cgroup_kn_unlock(kn);
1327 static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1329 char name[CGROUP_FILE_NAME_MAX];
1331 lockdep_assert_held(&cgroup_mutex);
1332 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1336 * css_clear_dir - remove subsys files in a cgroup directory
1338 * @cgrp_override: specify if target cgroup is different from css->cgroup
1340 static void css_clear_dir(struct cgroup_subsys_state *css,
1341 struct cgroup *cgrp_override)
1343 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1344 struct cftype *cfts;
1346 list_for_each_entry(cfts, &css->ss->cfts, node)
1347 cgroup_addrm_files(css, cgrp, cfts, false);
1351 * css_populate_dir - create subsys files in a cgroup directory
1353 * @cgrp_overried: specify if target cgroup is different from css->cgroup
1355 * On failure, no file is added.
1357 static int css_populate_dir(struct cgroup_subsys_state *css,
1358 struct cgroup *cgrp_override)
1360 struct cgroup *cgrp = cgrp_override ?: css->cgroup;
1361 struct cftype *cfts, *failed_cfts;
1365 if (cgroup_on_dfl(cgrp))
1366 cfts = cgroup_dfl_base_files;
1368 cfts = cgroup_legacy_base_files;
1370 return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1373 list_for_each_entry(cfts, &css->ss->cfts, node) {
1374 ret = cgroup_addrm_files(css, cgrp, cfts, true);
1382 list_for_each_entry(cfts, &css->ss->cfts, node) {
1383 if (cfts == failed_cfts)
1385 cgroup_addrm_files(css, cgrp, cfts, false);
1390 static int rebind_subsystems(struct cgroup_root *dst_root,
1391 unsigned long ss_mask)
1393 struct cgroup *dcgrp = &dst_root->cgrp;
1394 struct cgroup_subsys *ss;
1395 unsigned long tmp_ss_mask;
1398 lockdep_assert_held(&cgroup_mutex);
1400 for_each_subsys_which(ss, ssid, &ss_mask) {
1401 /* if @ss has non-root csses attached to it, can't move */
1402 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1405 /* can't move between two non-dummy roots either */
1406 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1410 /* skip creating root files on dfl_root for inhibited subsystems */
1411 tmp_ss_mask = ss_mask;
1412 if (dst_root == &cgrp_dfl_root)
1413 tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1415 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
1416 struct cgroup *scgrp = &ss->root->cgrp;
1419 ret = css_populate_dir(cgroup_css(scgrp, ss), dcgrp);
1424 * Rebinding back to the default root is not allowed to
1425 * fail. Using both default and non-default roots should
1426 * be rare. Moving subsystems back and forth even more so.
1427 * Just warn about it and continue.
1429 if (dst_root == &cgrp_dfl_root) {
1430 if (cgrp_dfl_root_visible) {
1431 pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1433 pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1438 for_each_subsys_which(ss, tssid, &tmp_ss_mask) {
1441 css_clear_dir(cgroup_css(scgrp, ss), dcgrp);
1447 * Nothing can fail from this point on. Remove files for the
1448 * removed subsystems and rebind each subsystem.
1450 for_each_subsys_which(ss, ssid, &ss_mask) {
1451 struct cgroup_root *src_root = ss->root;
1452 struct cgroup *scgrp = &src_root->cgrp;
1453 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1454 struct css_set *cset;
1456 WARN_ON(!css || cgroup_css(dcgrp, ss));
1458 css_clear_dir(css, NULL);
1460 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1461 rcu_assign_pointer(dcgrp->subsys[ssid], css);
1462 ss->root = dst_root;
1463 css->cgroup = dcgrp;
1465 down_write(&css_set_rwsem);
1466 hash_for_each(css_set_table, i, cset, hlist)
1467 list_move_tail(&cset->e_cset_node[ss->id],
1468 &dcgrp->e_csets[ss->id]);
1469 up_write(&css_set_rwsem);
1471 src_root->subsys_mask &= ~(1 << ssid);
1472 scgrp->subtree_control &= ~(1 << ssid);
1473 cgroup_refresh_child_subsys_mask(scgrp);
1475 /* default hierarchy doesn't enable controllers by default */
1476 dst_root->subsys_mask |= 1 << ssid;
1477 if (dst_root == &cgrp_dfl_root) {
1478 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1480 dcgrp->subtree_control |= 1 << ssid;
1481 cgroup_refresh_child_subsys_mask(dcgrp);
1482 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1489 kernfs_activate(dcgrp->kn);
1493 static int cgroup_show_options(struct seq_file *seq,
1494 struct kernfs_root *kf_root)
1496 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1497 struct cgroup_subsys *ss;
1500 if (root != &cgrp_dfl_root)
1501 for_each_subsys(ss, ssid)
1502 if (root->subsys_mask & (1 << ssid))
1503 seq_show_option(seq, ss->legacy_name, NULL);
1504 if (root->flags & CGRP_ROOT_NOPREFIX)
1505 seq_puts(seq, ",noprefix");
1506 if (root->flags & CGRP_ROOT_XATTR)
1507 seq_puts(seq, ",xattr");
1509 spin_lock(&release_agent_path_lock);
1510 if (strlen(root->release_agent_path))
1511 seq_show_option(seq, "release_agent",
1512 root->release_agent_path);
1513 spin_unlock(&release_agent_path_lock);
1515 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1516 seq_puts(seq, ",clone_children");
1517 if (strlen(root->name))
1518 seq_show_option(seq, "name", root->name);
1522 struct cgroup_sb_opts {
1523 unsigned long subsys_mask;
1525 char *release_agent;
1526 bool cpuset_clone_children;
1528 /* User explicitly requested empty subsystem */
1532 static int parse_cgroupfs_options(char *data, struct cgroup_sb_opts *opts)
1534 char *token, *o = data;
1535 bool all_ss = false, one_ss = false;
1536 unsigned long mask = -1UL;
1537 struct cgroup_subsys *ss;
1541 #ifdef CONFIG_CPUSETS
1542 mask = ~(1U << cpuset_cgrp_id);
1545 memset(opts, 0, sizeof(*opts));
1547 while ((token = strsep(&o, ",")) != NULL) {
1552 if (!strcmp(token, "none")) {
1553 /* Explicitly have no subsystems */
1557 if (!strcmp(token, "all")) {
1558 /* Mutually exclusive option 'all' + subsystem name */
1564 if (!strcmp(token, "__DEVEL__sane_behavior")) {
1565 opts->flags |= CGRP_ROOT_SANE_BEHAVIOR;
1568 if (!strcmp(token, "noprefix")) {
1569 opts->flags |= CGRP_ROOT_NOPREFIX;
1572 if (!strcmp(token, "clone_children")) {
1573 opts->cpuset_clone_children = true;
1576 if (!strcmp(token, "xattr")) {
1577 opts->flags |= CGRP_ROOT_XATTR;
1580 if (!strncmp(token, "release_agent=", 14)) {
1581 /* Specifying two release agents is forbidden */
1582 if (opts->release_agent)
1584 opts->release_agent =
1585 kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1586 if (!opts->release_agent)
1590 if (!strncmp(token, "name=", 5)) {
1591 const char *name = token + 5;
1592 /* Can't specify an empty name */
1595 /* Must match [\w.-]+ */
1596 for (i = 0; i < strlen(name); i++) {
1600 if ((c == '.') || (c == '-') || (c == '_'))
1604 /* Specifying two names is forbidden */
1607 opts->name = kstrndup(name,
1608 MAX_CGROUP_ROOT_NAMELEN - 1,
1616 for_each_subsys(ss, i) {
1617 if (strcmp(token, ss->legacy_name))
1619 if (!cgroup_ssid_enabled(i))
1622 /* Mutually exclusive option 'all' + subsystem name */
1625 opts->subsys_mask |= (1 << i);
1630 if (i == CGROUP_SUBSYS_COUNT)
1634 if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1635 pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1637 pr_err("sane_behavior: no other mount options allowed\n");
1644 * If the 'all' option was specified select all the subsystems,
1645 * otherwise if 'none', 'name=' and a subsystem name options were
1646 * not specified, let's default to 'all'
1648 if (all_ss || (!one_ss && !opts->none && !opts->name))
1649 for_each_subsys(ss, i)
1650 if (cgroup_ssid_enabled(i))
1651 opts->subsys_mask |= (1 << i);
1654 * We either have to specify by name or by subsystems. (So all
1655 * empty hierarchies must have a name).
1657 if (!opts->subsys_mask && !opts->name)
1661 * Option noprefix was introduced just for backward compatibility
1662 * with the old cpuset, so we allow noprefix only if mounting just
1663 * the cpuset subsystem.
1665 if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1668 /* Can't specify "none" and some subsystems */
1669 if (opts->subsys_mask && opts->none)
1675 static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1678 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1679 struct cgroup_sb_opts opts;
1680 unsigned long added_mask, removed_mask;
1682 if (root == &cgrp_dfl_root) {
1683 pr_err("remount is not allowed\n");
1687 mutex_lock(&cgroup_mutex);
1689 /* See what subsystems are wanted */
1690 ret = parse_cgroupfs_options(data, &opts);
1694 if (opts.subsys_mask != root->subsys_mask || opts.release_agent)
1695 pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1696 task_tgid_nr(current), current->comm);
1698 added_mask = opts.subsys_mask & ~root->subsys_mask;
1699 removed_mask = root->subsys_mask & ~opts.subsys_mask;
1701 /* Don't allow flags or name to change at remount */
1702 if ((opts.flags ^ root->flags) ||
1703 (opts.name && strcmp(opts.name, root->name))) {
1704 pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1705 opts.flags, opts.name ?: "", root->flags, root->name);
1710 /* remounting is not allowed for populated hierarchies */
1711 if (!list_empty(&root->cgrp.self.children)) {
1716 ret = rebind_subsystems(root, added_mask);
1720 rebind_subsystems(&cgrp_dfl_root, removed_mask);
1722 if (opts.release_agent) {
1723 spin_lock(&release_agent_path_lock);
1724 strcpy(root->release_agent_path, opts.release_agent);
1725 spin_unlock(&release_agent_path_lock);
1728 kfree(opts.release_agent);
1730 mutex_unlock(&cgroup_mutex);
1735 * To reduce the fork() overhead for systems that are not actually using
1736 * their cgroups capability, we don't maintain the lists running through
1737 * each css_set to its tasks until we see the list actually used - in other
1738 * words after the first mount.
1740 static bool use_task_css_set_links __read_mostly;
1742 static void cgroup_enable_task_cg_lists(void)
1744 struct task_struct *p, *g;
1746 down_write(&css_set_rwsem);
1748 if (use_task_css_set_links)
1751 use_task_css_set_links = true;
1754 * We need tasklist_lock because RCU is not safe against
1755 * while_each_thread(). Besides, a forking task that has passed
1756 * cgroup_post_fork() without seeing use_task_css_set_links = 1
1757 * is not guaranteed to have its child immediately visible in the
1758 * tasklist if we walk through it with RCU.
1760 read_lock(&tasklist_lock);
1761 do_each_thread(g, p) {
1762 WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1763 task_css_set(p) != &init_css_set);
1766 * We should check if the process is exiting, otherwise
1767 * it will race with cgroup_exit() in that the list
1768 * entry won't be deleted though the process has exited.
1769 * Do it while holding siglock so that we don't end up
1770 * racing against cgroup_exit().
1772 spin_lock_irq(&p->sighand->siglock);
1773 if (!(p->flags & PF_EXITING)) {
1774 struct css_set *cset = task_css_set(p);
1776 if (!css_set_populated(cset))
1777 css_set_update_populated(cset, true);
1778 list_add(&p->cg_list, &cset->tasks);
1781 spin_unlock_irq(&p->sighand->siglock);
1782 } while_each_thread(g, p);
1783 read_unlock(&tasklist_lock);
1785 up_write(&css_set_rwsem);
1788 static void init_cgroup_housekeeping(struct cgroup *cgrp)
1790 struct cgroup_subsys *ss;
1793 INIT_LIST_HEAD(&cgrp->self.sibling);
1794 INIT_LIST_HEAD(&cgrp->self.children);
1795 INIT_LIST_HEAD(&cgrp->self.files);
1796 INIT_LIST_HEAD(&cgrp->cset_links);
1797 INIT_LIST_HEAD(&cgrp->pidlists);
1798 mutex_init(&cgrp->pidlist_mutex);
1799 cgrp->self.cgroup = cgrp;
1800 cgrp->self.flags |= CSS_ONLINE;
1802 for_each_subsys(ss, ssid)
1803 INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1805 init_waitqueue_head(&cgrp->offline_waitq);
1806 INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1809 static void init_cgroup_root(struct cgroup_root *root,
1810 struct cgroup_sb_opts *opts)
1812 struct cgroup *cgrp = &root->cgrp;
1814 INIT_LIST_HEAD(&root->root_list);
1815 atomic_set(&root->nr_cgrps, 1);
1817 init_cgroup_housekeeping(cgrp);
1818 idr_init(&root->cgroup_idr);
1820 root->flags = opts->flags;
1821 if (opts->release_agent)
1822 strcpy(root->release_agent_path, opts->release_agent);
1824 strcpy(root->name, opts->name);
1825 if (opts->cpuset_clone_children)
1826 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1829 static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1831 LIST_HEAD(tmp_links);
1832 struct cgroup *root_cgrp = &root->cgrp;
1833 struct css_set *cset;
1836 lockdep_assert_held(&cgroup_mutex);
1838 ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1841 root_cgrp->id = ret;
1843 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1849 * We're accessing css_set_count without locking css_set_rwsem here,
1850 * but that's OK - it can only be increased by someone holding
1851 * cgroup_lock, and that's us. The worst that can happen is that we
1852 * have some link structures left over
1854 ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1858 ret = cgroup_init_root_id(root);
1862 root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1863 KERNFS_ROOT_CREATE_DEACTIVATED,
1865 if (IS_ERR(root->kf_root)) {
1866 ret = PTR_ERR(root->kf_root);
1869 root_cgrp->kn = root->kf_root->kn;
1871 ret = css_populate_dir(&root_cgrp->self, NULL);
1875 ret = rebind_subsystems(root, ss_mask);
1880 * There must be no failure case after here, since rebinding takes
1881 * care of subsystems' refcounts, which are explicitly dropped in
1882 * the failure exit path.
1884 list_add(&root->root_list, &cgroup_roots);
1885 cgroup_root_count++;
1888 * Link the root cgroup in this hierarchy into all the css_set
1891 down_write(&css_set_rwsem);
1892 hash_for_each(css_set_table, i, cset, hlist) {
1893 link_css_set(&tmp_links, cset, root_cgrp);
1894 if (css_set_populated(cset))
1895 cgroup_update_populated(root_cgrp, true);
1897 up_write(&css_set_rwsem);
1899 BUG_ON(!list_empty(&root_cgrp->self.children));
1900 BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1902 kernfs_activate(root_cgrp->kn);
1907 kernfs_destroy_root(root->kf_root);
1908 root->kf_root = NULL;
1910 cgroup_exit_root_id(root);
1912 percpu_ref_exit(&root_cgrp->self.refcnt);
1914 free_cgrp_cset_links(&tmp_links);
1918 static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1919 int flags, const char *unused_dev_name,
1922 struct super_block *pinned_sb = NULL;
1923 struct cgroup_subsys *ss;
1924 struct cgroup_root *root;
1925 struct cgroup_sb_opts opts;
1926 struct dentry *dentry;
1932 * The first time anyone tries to mount a cgroup, enable the list
1933 * linking each css_set to its tasks and fix up all existing tasks.
1935 if (!use_task_css_set_links)
1936 cgroup_enable_task_cg_lists();
1938 mutex_lock(&cgroup_mutex);
1940 /* First find the desired set of subsystems */
1941 ret = parse_cgroupfs_options(data, &opts);
1945 /* look for a matching existing root */
1946 if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1947 cgrp_dfl_root_visible = true;
1948 root = &cgrp_dfl_root;
1949 cgroup_get(&root->cgrp);
1955 * Destruction of cgroup root is asynchronous, so subsystems may
1956 * still be dying after the previous unmount. Let's drain the
1957 * dying subsystems. We just need to ensure that the ones
1958 * unmounted previously finish dying and don't care about new ones
1959 * starting. Testing ref liveliness is good enough.
1961 for_each_subsys(ss, i) {
1962 if (!(opts.subsys_mask & (1 << i)) ||
1963 ss->root == &cgrp_dfl_root)
1966 if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1967 mutex_unlock(&cgroup_mutex);
1969 ret = restart_syscall();
1972 cgroup_put(&ss->root->cgrp);
1975 for_each_root(root) {
1976 bool name_match = false;
1978 if (root == &cgrp_dfl_root)
1982 * If we asked for a name then it must match. Also, if
1983 * name matches but sybsys_mask doesn't, we should fail.
1984 * Remember whether name matched.
1987 if (strcmp(opts.name, root->name))
1993 * If we asked for subsystems (or explicitly for no
1994 * subsystems) then they must match.
1996 if ((opts.subsys_mask || opts.none) &&
1997 (opts.subsys_mask != root->subsys_mask)) {
2004 if (root->flags ^ opts.flags)
2005 pr_warn("new mount options do not match the existing superblock, will be ignored\n");
2008 * We want to reuse @root whose lifetime is governed by its
2009 * ->cgrp. Let's check whether @root is alive and keep it
2010 * that way. As cgroup_kill_sb() can happen anytime, we
2011 * want to block it by pinning the sb so that @root doesn't
2012 * get killed before mount is complete.
2014 * With the sb pinned, tryget_live can reliably indicate
2015 * whether @root can be reused. If it's being killed,
2016 * drain it. We can use wait_queue for the wait but this
2017 * path is super cold. Let's just sleep a bit and retry.
2019 pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
2020 if (IS_ERR(pinned_sb) ||
2021 !percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
2022 mutex_unlock(&cgroup_mutex);
2023 if (!IS_ERR_OR_NULL(pinned_sb))
2024 deactivate_super(pinned_sb);
2026 ret = restart_syscall();
2035 * No such thing, create a new one. name= matching without subsys
2036 * specification is allowed for already existing hierarchies but we
2037 * can't create new one without subsys specification.
2039 if (!opts.subsys_mask && !opts.none) {
2044 root = kzalloc(sizeof(*root), GFP_KERNEL);
2050 init_cgroup_root(root, &opts);
2052 ret = cgroup_setup_root(root, opts.subsys_mask);
2054 cgroup_free_root(root);
2057 mutex_unlock(&cgroup_mutex);
2059 kfree(opts.release_agent);
2063 return ERR_PTR(ret);
2065 dentry = kernfs_mount(fs_type, flags, root->kf_root,
2066 CGROUP_SUPER_MAGIC, &new_sb);
2067 if (IS_ERR(dentry) || !new_sb)
2068 cgroup_put(&root->cgrp);
2071 * If @pinned_sb, we're reusing an existing root and holding an
2072 * extra ref on its sb. Mount is complete. Put the extra ref.
2076 deactivate_super(pinned_sb);
2082 static void cgroup_kill_sb(struct super_block *sb)
2084 struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2085 struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2088 * If @root doesn't have any mounts or children, start killing it.
2089 * This prevents new mounts by disabling percpu_ref_tryget_live().
2090 * cgroup_mount() may wait for @root's release.
2092 * And don't kill the default root.
2094 if (!list_empty(&root->cgrp.self.children) ||
2095 root == &cgrp_dfl_root)
2096 cgroup_put(&root->cgrp);
2098 percpu_ref_kill(&root->cgrp.self.refcnt);
2103 static struct file_system_type cgroup_fs_type = {
2105 .mount = cgroup_mount,
2106 .kill_sb = cgroup_kill_sb,
2110 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2111 * @task: target task
2112 * @buf: the buffer to write the path into
2113 * @buflen: the length of the buffer
2115 * Determine @task's cgroup on the first (the one with the lowest non-zero
2116 * hierarchy_id) cgroup hierarchy and copy its path into @buf. This
2117 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2118 * cgroup controller callbacks.
2120 * Return value is the same as kernfs_path().
2122 char *task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2124 struct cgroup_root *root;
2125 struct cgroup *cgrp;
2126 int hierarchy_id = 1;
2129 mutex_lock(&cgroup_mutex);
2130 down_read(&css_set_rwsem);
2132 root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2135 cgrp = task_cgroup_from_root(task, root);
2136 path = cgroup_path(cgrp, buf, buflen);
2138 /* if no hierarchy exists, everyone is in "/" */
2139 if (strlcpy(buf, "/", buflen) < buflen)
2143 up_read(&css_set_rwsem);
2144 mutex_unlock(&cgroup_mutex);
2147 EXPORT_SYMBOL_GPL(task_cgroup_path);
2149 /* used to track tasks and other necessary states during migration */
2150 struct cgroup_taskset {
2151 /* the src and dst cset list running through cset->mg_node */
2152 struct list_head src_csets;
2153 struct list_head dst_csets;
2156 * Fields for cgroup_taskset_*() iteration.
2158 * Before migration is committed, the target migration tasks are on
2159 * ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
2160 * the csets on ->dst_csets. ->csets point to either ->src_csets
2161 * or ->dst_csets depending on whether migration is committed.
2163 * ->cur_csets and ->cur_task point to the current task position
2166 struct list_head *csets;
2167 struct css_set *cur_cset;
2168 struct task_struct *cur_task;
2171 #define CGROUP_TASKSET_INIT(tset) (struct cgroup_taskset){ \
2172 .src_csets = LIST_HEAD_INIT(tset.src_csets), \
2173 .dst_csets = LIST_HEAD_INIT(tset.dst_csets), \
2174 .csets = &tset.src_csets, \
2178 * cgroup_taskset_add - try to add a migration target task to a taskset
2179 * @task: target task
2180 * @tset: target taskset
2182 * Add @task, which is a migration target, to @tset. This function becomes
2183 * noop if @task doesn't need to be migrated. @task's css_set should have
2184 * been added as a migration source and @task->cg_list will be moved from
2185 * the css_set's tasks list to mg_tasks one.
2187 static void cgroup_taskset_add(struct task_struct *task,
2188 struct cgroup_taskset *tset)
2190 struct css_set *cset;
2192 lockdep_assert_held(&css_set_rwsem);
2194 /* @task either already exited or can't exit until the end */
2195 if (task->flags & PF_EXITING)
2198 /* leave @task alone if post_fork() hasn't linked it yet */
2199 if (list_empty(&task->cg_list))
2202 cset = task_css_set(task);
2203 if (!cset->mg_src_cgrp)
2206 list_move_tail(&task->cg_list, &cset->mg_tasks);
2207 if (list_empty(&cset->mg_node))
2208 list_add_tail(&cset->mg_node, &tset->src_csets);
2209 if (list_empty(&cset->mg_dst_cset->mg_node))
2210 list_move_tail(&cset->mg_dst_cset->mg_node,
2215 * cgroup_taskset_first - reset taskset and return the first task
2216 * @tset: taskset of interest
2218 * @tset iteration is initialized and the first task is returned.
2220 struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset)
2222 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2223 tset->cur_task = NULL;
2225 return cgroup_taskset_next(tset);
2229 * cgroup_taskset_next - iterate to the next task in taskset
2230 * @tset: taskset of interest
2232 * Return the next task in @tset. Iteration must have been initialized
2233 * with cgroup_taskset_first().
2235 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset)
2237 struct css_set *cset = tset->cur_cset;
2238 struct task_struct *task = tset->cur_task;
2240 while (&cset->mg_node != tset->csets) {
2242 task = list_first_entry(&cset->mg_tasks,
2243 struct task_struct, cg_list);
2245 task = list_next_entry(task, cg_list);
2247 if (&task->cg_list != &cset->mg_tasks) {
2248 tset->cur_cset = cset;
2249 tset->cur_task = task;
2253 cset = list_next_entry(cset, mg_node);
2261 * cgroup_task_migrate - move a task from one cgroup to another.
2262 * @tsk: the task being migrated
2263 * @new_cset: the new css_set @tsk is being attached to
2265 * Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2267 static void cgroup_task_migrate(struct task_struct *tsk,
2268 struct css_set *new_cset)
2270 struct css_set *old_cset;
2272 lockdep_assert_held(&cgroup_mutex);
2273 lockdep_assert_held(&css_set_rwsem);
2276 * We are synchronized through cgroup_threadgroup_rwsem against
2277 * PF_EXITING setting such that we can't race against cgroup_exit()
2278 * changing the css_set to init_css_set and dropping the old one.
2280 WARN_ON_ONCE(tsk->flags & PF_EXITING);
2281 old_cset = task_css_set(tsk);
2283 if (!css_set_populated(new_cset))
2284 css_set_update_populated(new_cset, true);
2286 get_css_set(new_cset);
2287 rcu_assign_pointer(tsk->cgroups, new_cset);
2288 list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2290 if (!css_set_populated(old_cset))
2291 css_set_update_populated(old_cset, false);
2294 * We just gained a reference on old_cset by taking it from the
2295 * task. As trading it for new_cset is protected by cgroup_mutex,
2296 * we're safe to drop it here; it will be freed under RCU.
2298 put_css_set_locked(old_cset);
2302 * cgroup_taskset_migrate - migrate a taskset to a cgroup
2303 * @tset: taget taskset
2304 * @dst_cgrp: destination cgroup
2306 * Migrate tasks in @tset to @dst_cgrp. This function fails iff one of the
2307 * ->can_attach callbacks fails and guarantees that either all or none of
2308 * the tasks in @tset are migrated. @tset is consumed regardless of
2311 static int cgroup_taskset_migrate(struct cgroup_taskset *tset,
2312 struct cgroup *dst_cgrp)
2314 struct cgroup_subsys_state *css, *failed_css = NULL;
2315 struct task_struct *task, *tmp_task;
2316 struct css_set *cset, *tmp_cset;
2319 /* methods shouldn't be called if no task is actually migrating */
2320 if (list_empty(&tset->src_csets))
2323 /* check that we can legitimately attach to the cgroup */
2324 for_each_e_css(css, i, dst_cgrp) {
2325 if (css->ss->can_attach) {
2326 ret = css->ss->can_attach(css, tset);
2329 goto out_cancel_attach;
2335 * Now that we're guaranteed success, proceed to move all tasks to
2336 * the new cgroup. There are no failure cases after here, so this
2337 * is the commit point.
2339 down_write(&css_set_rwsem);
2340 list_for_each_entry(cset, &tset->src_csets, mg_node) {
2341 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2342 cgroup_task_migrate(task, cset->mg_dst_cset);
2344 up_write(&css_set_rwsem);
2347 * Migration is committed, all target tasks are now on dst_csets.
2348 * Nothing is sensitive to fork() after this point. Notify
2349 * controllers that migration is complete.
2351 tset->csets = &tset->dst_csets;
2353 for_each_e_css(css, i, dst_cgrp)
2354 if (css->ss->attach)
2355 css->ss->attach(css, tset);
2358 goto out_release_tset;
2361 for_each_e_css(css, i, dst_cgrp) {
2362 if (css == failed_css)
2364 if (css->ss->cancel_attach)
2365 css->ss->cancel_attach(css, tset);
2368 down_write(&css_set_rwsem);
2369 list_splice_init(&tset->dst_csets, &tset->src_csets);
2370 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2371 list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2372 list_del_init(&cset->mg_node);
2374 up_write(&css_set_rwsem);
2379 * cgroup_migrate_finish - cleanup after attach
2380 * @preloaded_csets: list of preloaded css_sets
2382 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2383 * those functions for details.
2385 static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2387 struct css_set *cset, *tmp_cset;
2389 lockdep_assert_held(&cgroup_mutex);
2391 down_write(&css_set_rwsem);
2392 list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2393 cset->mg_src_cgrp = NULL;
2394 cset->mg_dst_cset = NULL;
2395 list_del_init(&cset->mg_preload_node);
2396 put_css_set_locked(cset);
2398 up_write(&css_set_rwsem);
2402 * cgroup_migrate_add_src - add a migration source css_set
2403 * @src_cset: the source css_set to add
2404 * @dst_cgrp: the destination cgroup
2405 * @preloaded_csets: list of preloaded css_sets
2407 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2408 * @src_cset and add it to @preloaded_csets, which should later be cleaned
2409 * up by cgroup_migrate_finish().
2411 * This function may be called without holding cgroup_threadgroup_rwsem
2412 * even if the target is a process. Threads may be created and destroyed
2413 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2414 * into play and the preloaded css_sets are guaranteed to cover all
2417 static void cgroup_migrate_add_src(struct css_set *src_cset,
2418 struct cgroup *dst_cgrp,
2419 struct list_head *preloaded_csets)
2421 struct cgroup *src_cgrp;
2423 lockdep_assert_held(&cgroup_mutex);
2424 lockdep_assert_held(&css_set_rwsem);
2426 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2428 if (!list_empty(&src_cset->mg_preload_node))
2431 WARN_ON(src_cset->mg_src_cgrp);
2432 WARN_ON(!list_empty(&src_cset->mg_tasks));
2433 WARN_ON(!list_empty(&src_cset->mg_node));
2435 src_cset->mg_src_cgrp = src_cgrp;
2436 get_css_set(src_cset);
2437 list_add(&src_cset->mg_preload_node, preloaded_csets);
2441 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2442 * @dst_cgrp: the destination cgroup (may be %NULL)
2443 * @preloaded_csets: list of preloaded source css_sets
2445 * Tasks are about to be moved to @dst_cgrp and all the source css_sets
2446 * have been preloaded to @preloaded_csets. This function looks up and
2447 * pins all destination css_sets, links each to its source, and append them
2448 * to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2449 * source css_set is assumed to be its cgroup on the default hierarchy.
2451 * This function must be called after cgroup_migrate_add_src() has been
2452 * called on each migration source css_set. After migration is performed
2453 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2456 static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2457 struct list_head *preloaded_csets)
2460 struct css_set *src_cset, *tmp_cset;
2462 lockdep_assert_held(&cgroup_mutex);
2465 * Except for the root, child_subsys_mask must be zero for a cgroup
2466 * with tasks so that child cgroups don't compete against tasks.
2468 if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2469 dst_cgrp->child_subsys_mask)
2472 /* look up the dst cset for each src cset and link it to src */
2473 list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2474 struct css_set *dst_cset;
2476 dst_cset = find_css_set(src_cset,
2477 dst_cgrp ?: src_cset->dfl_cgrp);
2481 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2484 * If src cset equals dst, it's noop. Drop the src.
2485 * cgroup_migrate() will skip the cset too. Note that we
2486 * can't handle src == dst as some nodes are used by both.
2488 if (src_cset == dst_cset) {
2489 src_cset->mg_src_cgrp = NULL;
2490 list_del_init(&src_cset->mg_preload_node);
2491 put_css_set(src_cset);
2492 put_css_set(dst_cset);
2496 src_cset->mg_dst_cset = dst_cset;
2498 if (list_empty(&dst_cset->mg_preload_node))
2499 list_add(&dst_cset->mg_preload_node, &csets);
2501 put_css_set(dst_cset);
2504 list_splice_tail(&csets, preloaded_csets);
2507 cgroup_migrate_finish(&csets);
2512 * cgroup_migrate - migrate a process or task to a cgroup
2513 * @leader: the leader of the process or the task to migrate
2514 * @threadgroup: whether @leader points to the whole process or a single task
2515 * @cgrp: the destination cgroup
2517 * Migrate a process or task denoted by @leader to @cgrp. If migrating a
2518 * process, the caller must be holding cgroup_threadgroup_rwsem. The
2519 * caller is also responsible for invoking cgroup_migrate_add_src() and
2520 * cgroup_migrate_prepare_dst() on the targets before invoking this
2521 * function and following up with cgroup_migrate_finish().
2523 * As long as a controller's ->can_attach() doesn't fail, this function is
2524 * guaranteed to succeed. This means that, excluding ->can_attach()
2525 * failure, when migrating multiple targets, the success or failure can be
2526 * decided for all targets by invoking group_migrate_prepare_dst() before
2527 * actually starting migrating.
2529 static int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2530 struct cgroup *cgrp)
2532 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2533 struct task_struct *task;
2536 * Prevent freeing of tasks while we take a snapshot. Tasks that are
2537 * already PF_EXITING could be freed from underneath us unless we
2538 * take an rcu_read_lock.
2540 down_write(&css_set_rwsem);
2544 cgroup_taskset_add(task, &tset);
2547 } while_each_thread(leader, task);
2549 up_write(&css_set_rwsem);
2551 return cgroup_taskset_migrate(&tset, cgrp);
2555 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2556 * @dst_cgrp: the cgroup to attach to
2557 * @leader: the task or the leader of the threadgroup to be attached
2558 * @threadgroup: attach the whole threadgroup?
2560 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2562 static int cgroup_attach_task(struct cgroup *dst_cgrp,
2563 struct task_struct *leader, bool threadgroup)
2565 LIST_HEAD(preloaded_csets);
2566 struct task_struct *task;
2569 /* look up all src csets */
2570 down_read(&css_set_rwsem);
2574 cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2578 } while_each_thread(leader, task);
2580 up_read(&css_set_rwsem);
2582 /* prepare dst csets and commit */
2583 ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2585 ret = cgroup_migrate(leader, threadgroup, dst_cgrp);
2587 cgroup_migrate_finish(&preloaded_csets);
2591 static int cgroup_procs_write_permission(struct task_struct *task,
2592 struct cgroup *dst_cgrp,
2593 struct kernfs_open_file *of)
2595 const struct cred *cred = current_cred();
2596 const struct cred *tcred = get_task_cred(task);
2600 * even if we're attaching all tasks in the thread group, we only
2601 * need to check permissions on one of them.
2603 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2604 !uid_eq(cred->euid, tcred->uid) &&
2605 !uid_eq(cred->euid, tcred->suid))
2608 if (!ret && cgroup_on_dfl(dst_cgrp)) {
2609 struct super_block *sb = of->file->f_path.dentry->d_sb;
2610 struct cgroup *cgrp;
2611 struct inode *inode;
2613 down_read(&css_set_rwsem);
2614 cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2615 up_read(&css_set_rwsem);
2617 while (!cgroup_is_descendant(dst_cgrp, cgrp))
2618 cgrp = cgroup_parent(cgrp);
2621 inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
2623 ret = inode_permission(inode, MAY_WRITE);
2633 * Find the task_struct of the task to attach by vpid and pass it along to the
2634 * function to attach either it or all tasks in its threadgroup. Will lock
2635 * cgroup_mutex and threadgroup.
2637 static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf,
2638 size_t nbytes, loff_t off, bool threadgroup)
2640 struct task_struct *tsk;
2641 struct cgroup *cgrp;
2645 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2648 cgrp = cgroup_kn_lock_live(of->kn);
2652 percpu_down_write(&cgroup_threadgroup_rwsem);
2655 tsk = find_task_by_vpid(pid);
2658 goto out_unlock_rcu;
2665 tsk = tsk->group_leader;
2668 * Workqueue threads may acquire PF_NO_SETAFFINITY and become
2669 * trapped in a cpuset, or RT worker may be born in a cgroup
2670 * with no rt_runtime allocated. Just say no.
2672 if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
2674 goto out_unlock_rcu;
2677 get_task_struct(tsk);
2680 ret = cgroup_procs_write_permission(tsk, cgrp, of);
2682 ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2684 put_task_struct(tsk);
2685 goto out_unlock_threadgroup;
2689 out_unlock_threadgroup:
2690 percpu_up_write(&cgroup_threadgroup_rwsem);
2691 cgroup_kn_unlock(of->kn);
2692 return ret ?: nbytes;
2696 * cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2697 * @from: attach to all cgroups of a given task
2698 * @tsk: the task to be attached
2700 int cgroup_attach_task_all(struct task_struct *from, struct task_struct *tsk)
2702 struct cgroup_root *root;
2705 mutex_lock(&cgroup_mutex);
2706 for_each_root(root) {
2707 struct cgroup *from_cgrp;
2709 if (root == &cgrp_dfl_root)
2712 down_read(&css_set_rwsem);
2713 from_cgrp = task_cgroup_from_root(from, root);
2714 up_read(&css_set_rwsem);
2716 retval = cgroup_attach_task(from_cgrp, tsk, false);
2720 mutex_unlock(&cgroup_mutex);
2724 EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2726 static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2727 char *buf, size_t nbytes, loff_t off)
2729 return __cgroup_procs_write(of, buf, nbytes, off, false);
2732 static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2733 char *buf, size_t nbytes, loff_t off)
2735 return __cgroup_procs_write(of, buf, nbytes, off, true);
2738 static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2739 char *buf, size_t nbytes, loff_t off)
2741 struct cgroup *cgrp;
2743 BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2745 cgrp = cgroup_kn_lock_live(of->kn);
2748 spin_lock(&release_agent_path_lock);
2749 strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2750 sizeof(cgrp->root->release_agent_path));
2751 spin_unlock(&release_agent_path_lock);
2752 cgroup_kn_unlock(of->kn);
2756 static int cgroup_release_agent_show(struct seq_file *seq, void *v)
2758 struct cgroup *cgrp = seq_css(seq)->cgroup;
2760 spin_lock(&release_agent_path_lock);
2761 seq_puts(seq, cgrp->root->release_agent_path);
2762 spin_unlock(&release_agent_path_lock);
2763 seq_putc(seq, '\n');
2767 static int cgroup_sane_behavior_show(struct seq_file *seq, void *v)
2769 seq_puts(seq, "0\n");
2773 static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2775 struct cgroup_subsys *ss;
2776 bool printed = false;
2779 for_each_subsys_which(ss, ssid, &ss_mask) {
2782 seq_printf(seq, "%s", ss->name);
2786 seq_putc(seq, '\n');
2789 /* show controllers which are currently attached to the default hierarchy */
2790 static int cgroup_root_controllers_show(struct seq_file *seq, void *v)
2792 struct cgroup *cgrp = seq_css(seq)->cgroup;
2794 cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2795 ~cgrp_dfl_root_inhibit_ss_mask);
2799 /* show controllers which are enabled from the parent */
2800 static int cgroup_controllers_show(struct seq_file *seq, void *v)
2802 struct cgroup *cgrp = seq_css(seq)->cgroup;
2804 cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2808 /* show controllers which are enabled for a given cgroup's children */
2809 static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2811 struct cgroup *cgrp = seq_css(seq)->cgroup;
2813 cgroup_print_ss_mask(seq, cgrp->subtree_control);
2818 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2819 * @cgrp: root of the subtree to update csses for
2821 * @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2822 * css associations need to be updated accordingly. This function looks up
2823 * all css_sets which are attached to the subtree, creates the matching
2824 * updated css_sets and migrates the tasks to the new ones.
2826 static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2828 LIST_HEAD(preloaded_csets);
2829 struct cgroup_taskset tset = CGROUP_TASKSET_INIT(tset);
2830 struct cgroup_subsys_state *css;
2831 struct css_set *src_cset;
2834 lockdep_assert_held(&cgroup_mutex);
2836 percpu_down_write(&cgroup_threadgroup_rwsem);
2838 /* look up all csses currently attached to @cgrp's subtree */
2839 down_read(&css_set_rwsem);
2840 css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2841 struct cgrp_cset_link *link;
2843 /* self is not affected by child_subsys_mask change */
2844 if (css->cgroup == cgrp)
2847 list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2848 cgroup_migrate_add_src(link->cset, cgrp,
2851 up_read(&css_set_rwsem);
2853 /* NULL dst indicates self on default hierarchy */
2854 ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2858 down_write(&css_set_rwsem);
2859 list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2860 struct task_struct *task, *ntask;
2862 /* src_csets precede dst_csets, break on the first dst_cset */
2863 if (!src_cset->mg_src_cgrp)
2866 /* all tasks in src_csets need to be migrated */
2867 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2868 cgroup_taskset_add(task, &tset);
2870 up_write(&css_set_rwsem);
2872 ret = cgroup_taskset_migrate(&tset, cgrp);
2874 cgroup_migrate_finish(&preloaded_csets);
2875 percpu_up_write(&cgroup_threadgroup_rwsem);
2879 /* change the enabled child controllers for a cgroup in the default hierarchy */
2880 static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2881 char *buf, size_t nbytes,
2884 unsigned long enable = 0, disable = 0;
2885 unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2886 struct cgroup *cgrp, *child;
2887 struct cgroup_subsys *ss;
2892 * Parse input - space separated list of subsystem names prefixed
2893 * with either + or -.
2895 buf = strstrip(buf);
2896 while ((tok = strsep(&buf, " "))) {
2897 unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2901 for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2902 if (!cgroup_ssid_enabled(ssid) ||
2903 strcmp(tok + 1, ss->name))
2907 enable |= 1 << ssid;
2908 disable &= ~(1 << ssid);
2909 } else if (*tok == '-') {
2910 disable |= 1 << ssid;
2911 enable &= ~(1 << ssid);
2917 if (ssid == CGROUP_SUBSYS_COUNT)
2921 cgrp = cgroup_kn_lock_live(of->kn);
2925 for_each_subsys(ss, ssid) {
2926 if (enable & (1 << ssid)) {
2927 if (cgrp->subtree_control & (1 << ssid)) {
2928 enable &= ~(1 << ssid);
2932 /* unavailable or not enabled on the parent? */
2933 if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) ||
2934 (cgroup_parent(cgrp) &&
2935 !(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2939 } else if (disable & (1 << ssid)) {
2940 if (!(cgrp->subtree_control & (1 << ssid))) {
2941 disable &= ~(1 << ssid);
2945 /* a child has it enabled? */
2946 cgroup_for_each_live_child(child, cgrp) {
2947 if (child->subtree_control & (1 << ssid)) {
2955 if (!enable && !disable) {
2961 * Except for the root, subtree_control must be zero for a cgroup
2962 * with tasks so that child cgroups don't compete against tasks.
2964 if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2970 * Update subsys masks and calculate what needs to be done. More
2971 * subsystems than specified may need to be enabled or disabled
2972 * depending on subsystem dependencies.
2974 old_sc = cgrp->subtree_control;
2975 old_ss = cgrp->child_subsys_mask;
2976 new_sc = (old_sc | enable) & ~disable;
2977 new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2979 css_enable = ~old_ss & new_ss;
2980 css_disable = old_ss & ~new_ss;
2981 enable |= css_enable;
2982 disable |= css_disable;
2985 * Because css offlining is asynchronous, userland might try to
2986 * re-enable the same controller while the previous instance is
2987 * still around. In such cases, wait till it's gone using
2990 for_each_subsys_which(ss, ssid, &css_enable) {
2991 cgroup_for_each_live_child(child, cgrp) {
2994 if (!cgroup_css(child, ss))
2998 prepare_to_wait(&child->offline_waitq, &wait,
2999 TASK_UNINTERRUPTIBLE);
3000 cgroup_kn_unlock(of->kn);
3002 finish_wait(&child->offline_waitq, &wait);
3005 return restart_syscall();
3009 cgrp->subtree_control = new_sc;
3010 cgrp->child_subsys_mask = new_ss;
3013 * Create new csses or make the existing ones visible. A css is
3014 * created invisible if it's being implicitly enabled through
3015 * dependency. An invisible css is made visible when the userland
3016 * explicitly enables it.
3018 for_each_subsys(ss, ssid) {
3019 if (!(enable & (1 << ssid)))
3022 cgroup_for_each_live_child(child, cgrp) {
3023 if (css_enable & (1 << ssid))
3024 ret = create_css(child, ss,
3025 cgrp->subtree_control & (1 << ssid));
3027 ret = css_populate_dir(cgroup_css(child, ss),
3035 * At this point, cgroup_e_css() results reflect the new csses
3036 * making the following cgroup_update_dfl_csses() properly update
3037 * css associations of all tasks in the subtree.
3039 ret = cgroup_update_dfl_csses(cgrp);
3044 * All tasks are migrated out of disabled csses. Kill or hide
3045 * them. A css is hidden when the userland requests it to be
3046 * disabled while other subsystems are still depending on it. The
3047 * css must not actively control resources and be in the vanilla
3048 * state if it's made visible again later. Controllers which may
3049 * be depended upon should provide ->css_reset() for this purpose.
3051 for_each_subsys(ss, ssid) {
3052 if (!(disable & (1 << ssid)))
3055 cgroup_for_each_live_child(child, cgrp) {
3056 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3058 if (css_disable & (1 << ssid)) {
3061 css_clear_dir(css, NULL);
3069 * The effective csses of all the descendants (excluding @cgrp) may
3070 * have changed. Subsystems can optionally subscribe to this event
3071 * by implementing ->css_e_css_changed() which is invoked if any of
3072 * the effective csses seen from the css's cgroup may have changed.
3074 for_each_subsys(ss, ssid) {
3075 struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
3076 struct cgroup_subsys_state *css;
3078 if (!ss->css_e_css_changed || !this_css)
3081 css_for_each_descendant_pre(css, this_css)
3082 if (css != this_css)
3083 ss->css_e_css_changed(css);
3086 kernfs_activate(cgrp->kn);
3089 cgroup_kn_unlock(of->kn);
3090 return ret ?: nbytes;
3093 cgrp->subtree_control = old_sc;
3094 cgrp->child_subsys_mask = old_ss;
3096 for_each_subsys(ss, ssid) {
3097 if (!(enable & (1 << ssid)))
3100 cgroup_for_each_live_child(child, cgrp) {
3101 struct cgroup_subsys_state *css = cgroup_css(child, ss);
3106 if (css_enable & (1 << ssid))
3109 css_clear_dir(css, NULL);
3115 static int cgroup_events_show(struct seq_file *seq, void *v)
3117 seq_printf(seq, "populated %d\n",
3118 cgroup_is_populated(seq_css(seq)->cgroup));
3122 static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3123 size_t nbytes, loff_t off)
3125 struct cgroup *cgrp = of->kn->parent->priv;
3126 struct cftype *cft = of->kn->priv;
3127 struct cgroup_subsys_state *css;
3131 return cft->write(of, buf, nbytes, off);
3134 * kernfs guarantees that a file isn't deleted with operations in
3135 * flight, which means that the matching css is and stays alive and
3136 * doesn't need to be pinned. The RCU locking is not necessary
3137 * either. It's just for the convenience of using cgroup_css().
3140 css = cgroup_css(cgrp, cft->ss);
3143 if (cft->write_u64) {
3144 unsigned long long v;
3145 ret = kstrtoull(buf, 0, &v);
3147 ret = cft->write_u64(css, cft, v);
3148 } else if (cft->write_s64) {
3150 ret = kstrtoll(buf, 0, &v);
3152 ret = cft->write_s64(css, cft, v);
3157 return ret ?: nbytes;
3160 static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3162 return seq_cft(seq)->seq_start(seq, ppos);
3165 static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3167 return seq_cft(seq)->seq_next(seq, v, ppos);
3170 static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3172 seq_cft(seq)->seq_stop(seq, v);
3175 static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3177 struct cftype *cft = seq_cft(m);
3178 struct cgroup_subsys_state *css = seq_css(m);
3181 return cft->seq_show(m, arg);
3184 seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3185 else if (cft->read_s64)
3186 seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3192 static struct kernfs_ops cgroup_kf_single_ops = {
3193 .atomic_write_len = PAGE_SIZE,
3194 .write = cgroup_file_write,
3195 .seq_show = cgroup_seqfile_show,
3198 static struct kernfs_ops cgroup_kf_ops = {
3199 .atomic_write_len = PAGE_SIZE,
3200 .write = cgroup_file_write,
3201 .seq_start = cgroup_seqfile_start,
3202 .seq_next = cgroup_seqfile_next,
3203 .seq_stop = cgroup_seqfile_stop,
3204 .seq_show = cgroup_seqfile_show,
3208 * cgroup_rename - Only allow simple rename of directories in place.
3210 static int cgroup_rename(struct kernfs_node *kn, struct kernfs_node *new_parent,
3211 const char *new_name_str)
3213 struct cgroup *cgrp = kn->priv;
3216 if (kernfs_type(kn) != KERNFS_DIR)
3218 if (kn->parent != new_parent)
3222 * This isn't a proper migration and its usefulness is very
3223 * limited. Disallow on the default hierarchy.
3225 if (cgroup_on_dfl(cgrp))
3229 * We're gonna grab cgroup_mutex which nests outside kernfs
3230 * active_ref. kernfs_rename() doesn't require active_ref
3231 * protection. Break them before grabbing cgroup_mutex.
3233 kernfs_break_active_protection(new_parent);
3234 kernfs_break_active_protection(kn);
3236 mutex_lock(&cgroup_mutex);
3238 ret = kernfs_rename(kn, new_parent, new_name_str);
3240 mutex_unlock(&cgroup_mutex);
3242 kernfs_unbreak_active_protection(kn);
3243 kernfs_unbreak_active_protection(new_parent);
3247 /* set uid and gid of cgroup dirs and files to that of the creator */
3248 static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3250 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3251 .ia_uid = current_fsuid(),
3252 .ia_gid = current_fsgid(), };
3254 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3255 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3258 return kernfs_setattr(kn, &iattr);
3261 static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3264 char name[CGROUP_FILE_NAME_MAX];
3265 struct kernfs_node *kn;
3266 struct lock_class_key *key = NULL;
3269 #ifdef CONFIG_DEBUG_LOCK_ALLOC
3270 key = &cft->lockdep_key;
3272 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3273 cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3278 ret = cgroup_kn_set_ugid(kn);
3284 if (cft->file_offset) {
3285 struct cgroup_file *cfile = (void *)css + cft->file_offset;
3289 list_add(&cfile->node, &css->files);
3296 * cgroup_addrm_files - add or remove files to a cgroup directory
3297 * @css: the target css
3298 * @cgrp: the target cgroup (usually css->cgroup)
3299 * @cfts: array of cftypes to be added
3300 * @is_add: whether to add or remove
3302 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3303 * For removals, this function never fails.
3305 static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3306 struct cgroup *cgrp, struct cftype cfts[],
3309 struct cftype *cft, *cft_end = NULL;
3312 lockdep_assert_held(&cgroup_mutex);
3315 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3316 /* does cft->flags tell us to skip this file on @cgrp? */
3317 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3319 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3321 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3323 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3327 ret = cgroup_add_file(css, cgrp, cft);
3329 pr_warn("%s: failed to add %s, err=%d\n",
3330 __func__, cft->name, ret);
3336 cgroup_rm_file(cgrp, cft);
3342 static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3345 struct cgroup_subsys *ss = cfts[0].ss;
3346 struct cgroup *root = &ss->root->cgrp;
3347 struct cgroup_subsys_state *css;
3350 lockdep_assert_held(&cgroup_mutex);
3352 /* add/rm files for all cgroups created before */
3353 css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3354 struct cgroup *cgrp = css->cgroup;
3356 if (cgroup_is_dead(cgrp))
3359 ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3365 kernfs_activate(root->kn);
3369 static void cgroup_exit_cftypes(struct cftype *cfts)
3373 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3374 /* free copy for custom atomic_write_len, see init_cftypes() */
3375 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3380 /* revert flags set by cgroup core while adding @cfts */
3381 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3385 static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3389 for (cft = cfts; cft->name[0] != '\0'; cft++) {
3390 struct kernfs_ops *kf_ops;
3392 WARN_ON(cft->ss || cft->kf_ops);
3395 kf_ops = &cgroup_kf_ops;
3397 kf_ops = &cgroup_kf_single_ops;
3400 * Ugh... if @cft wants a custom max_write_len, we need to
3401 * make a copy of kf_ops to set its atomic_write_len.
3403 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3404 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3406 cgroup_exit_cftypes(cfts);
3409 kf_ops->atomic_write_len = cft->max_write_len;
3412 cft->kf_ops = kf_ops;
3419 static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3421 lockdep_assert_held(&cgroup_mutex);
3423 if (!cfts || !cfts[0].ss)
3426 list_del(&cfts->node);
3427 cgroup_apply_cftypes(cfts, false);
3428 cgroup_exit_cftypes(cfts);
3433 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3434 * @cfts: zero-length name terminated array of cftypes
3436 * Unregister @cfts. Files described by @cfts are removed from all
3437 * existing cgroups and all future cgroups won't have them either. This
3438 * function can be called anytime whether @cfts' subsys is attached or not.
3440 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3443 int cgroup_rm_cftypes(struct cftype *cfts)
3447 mutex_lock(&cgroup_mutex);
3448 ret = cgroup_rm_cftypes_locked(cfts);
3449 mutex_unlock(&cgroup_mutex);
3454 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3455 * @ss: target cgroup subsystem
3456 * @cfts: zero-length name terminated array of cftypes
3458 * Register @cfts to @ss. Files described by @cfts are created for all
3459 * existing cgroups to which @ss is attached and all future cgroups will
3460 * have them too. This function can be called anytime whether @ss is
3463 * Returns 0 on successful registration, -errno on failure. Note that this
3464 * function currently returns 0 as long as @cfts registration is successful
3465 * even if some file creation attempts on existing cgroups fail.
3467 static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3471 if (!cgroup_ssid_enabled(ss->id))
3474 if (!cfts || cfts[0].name[0] == '\0')
3477 ret = cgroup_init_cftypes(ss, cfts);
3481 mutex_lock(&cgroup_mutex);
3483 list_add_tail(&cfts->node, &ss->cfts);
3484 ret = cgroup_apply_cftypes(cfts, true);
3486 cgroup_rm_cftypes_locked(cfts);
3488 mutex_unlock(&cgroup_mutex);
3493 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3494 * @ss: target cgroup subsystem
3495 * @cfts: zero-length name terminated array of cftypes
3497 * Similar to cgroup_add_cftypes() but the added files are only used for
3498 * the default hierarchy.
3500 int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3504 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3505 cft->flags |= __CFTYPE_ONLY_ON_DFL;
3506 return cgroup_add_cftypes(ss, cfts);
3510 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3511 * @ss: target cgroup subsystem
3512 * @cfts: zero-length name terminated array of cftypes
3514 * Similar to cgroup_add_cftypes() but the added files are only used for
3515 * the legacy hierarchies.
3517 int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3522 * If legacy_flies_on_dfl, we want to show the legacy files on the
3523 * dfl hierarchy but iff the target subsystem hasn't been updated
3524 * for the dfl hierarchy yet.
3526 if (!cgroup_legacy_files_on_dfl ||
3527 ss->dfl_cftypes != ss->legacy_cftypes) {
3528 for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3529 cft->flags |= __CFTYPE_NOT_ON_DFL;
3532 return cgroup_add_cftypes(ss, cfts);
3536 * cgroup_task_count - count the number of tasks in a cgroup.
3537 * @cgrp: the cgroup in question
3539 * Return the number of tasks in the cgroup.
3541 static int cgroup_task_count(const struct cgroup *cgrp)
3544 struct cgrp_cset_link *link;
3546 down_read(&css_set_rwsem);
3547 list_for_each_entry(link, &cgrp->cset_links, cset_link)
3548 count += atomic_read(&link->cset->refcount);
3549 up_read(&css_set_rwsem);
3554 * css_next_child - find the next child of a given css
3555 * @pos: the current position (%NULL to initiate traversal)
3556 * @parent: css whose children to walk
3558 * This function returns the next child of @parent and should be called
3559 * under either cgroup_mutex or RCU read lock. The only requirement is
3560 * that @parent and @pos are accessible. The next sibling is guaranteed to
3561 * be returned regardless of their states.
3563 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3564 * css which finished ->css_online() is guaranteed to be visible in the
3565 * future iterations and will stay visible until the last reference is put.
3566 * A css which hasn't finished ->css_online() or already finished
3567 * ->css_offline() may show up during traversal. It's each subsystem's
3568 * responsibility to synchronize against on/offlining.
3570 struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3571 struct cgroup_subsys_state *parent)
3573 struct cgroup_subsys_state *next;
3575 cgroup_assert_mutex_or_rcu_locked();
3578 * @pos could already have been unlinked from the sibling list.
3579 * Once a cgroup is removed, its ->sibling.next is no longer
3580 * updated when its next sibling changes. CSS_RELEASED is set when
3581 * @pos is taken off list, at which time its next pointer is valid,
3582 * and, as releases are serialized, the one pointed to by the next
3583 * pointer is guaranteed to not have started release yet. This
3584 * implies that if we observe !CSS_RELEASED on @pos in this RCU
3585 * critical section, the one pointed to by its next pointer is
3586 * guaranteed to not have finished its RCU grace period even if we
3587 * have dropped rcu_read_lock() inbetween iterations.
3589 * If @pos has CSS_RELEASED set, its next pointer can't be
3590 * dereferenced; however, as each css is given a monotonically
3591 * increasing unique serial number and always appended to the
3592 * sibling list, the next one can be found by walking the parent's
3593 * children until the first css with higher serial number than
3594 * @pos's. While this path can be slower, it happens iff iteration
3595 * races against release and the race window is very small.
3598 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3599 } else if (likely(!(pos->flags & CSS_RELEASED))) {
3600 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3602 list_for_each_entry_rcu(next, &parent->children, sibling)
3603 if (next->serial_nr > pos->serial_nr)
3608 * @next, if not pointing to the head, can be dereferenced and is
3611 if (&next->sibling != &parent->children)
3617 * css_next_descendant_pre - find the next descendant for pre-order walk
3618 * @pos: the current position (%NULL to initiate traversal)
3619 * @root: css whose descendants to walk
3621 * To be used by css_for_each_descendant_pre(). Find the next descendant
3622 * to visit for pre-order traversal of @root's descendants. @root is
3623 * included in the iteration and the first node to be visited.
3625 * While this function requires cgroup_mutex or RCU read locking, it
3626 * doesn't require the whole traversal to be contained in a single critical
3627 * section. This function will return the correct next descendant as long
3628 * as both @pos and @root are accessible and @pos is a descendant of @root.
3630 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3631 * css which finished ->css_online() is guaranteed to be visible in the
3632 * future iterations and will stay visible until the last reference is put.
3633 * A css which hasn't finished ->css_online() or already finished
3634 * ->css_offline() may show up during traversal. It's each subsystem's
3635 * responsibility to synchronize against on/offlining.
3637 struct cgroup_subsys_state *
3638 css_next_descendant_pre(struct cgroup_subsys_state *pos,
3639 struct cgroup_subsys_state *root)
3641 struct cgroup_subsys_state *next;
3643 cgroup_assert_mutex_or_rcu_locked();
3645 /* if first iteration, visit @root */
3649 /* visit the first child if exists */
3650 next = css_next_child(NULL, pos);
3654 /* no child, visit my or the closest ancestor's next sibling */
3655 while (pos != root) {
3656 next = css_next_child(pos, pos->parent);
3666 * css_rightmost_descendant - return the rightmost descendant of a css
3667 * @pos: css of interest
3669 * Return the rightmost descendant of @pos. If there's no descendant, @pos
3670 * is returned. This can be used during pre-order traversal to skip
3673 * While this function requires cgroup_mutex or RCU read locking, it
3674 * doesn't require the whole traversal to be contained in a single critical
3675 * section. This function will return the correct rightmost descendant as
3676 * long as @pos is accessible.
3678 struct cgroup_subsys_state *
3679 css_rightmost_descendant(struct cgroup_subsys_state *pos)
3681 struct cgroup_subsys_state *last, *tmp;
3683 cgroup_assert_mutex_or_rcu_locked();
3687 /* ->prev isn't RCU safe, walk ->next till the end */
3689 css_for_each_child(tmp, last)
3696 static struct cgroup_subsys_state *
3697 css_leftmost_descendant(struct cgroup_subsys_state *pos)
3699 struct cgroup_subsys_state *last;
3703 pos = css_next_child(NULL, pos);
3710 * css_next_descendant_post - find the next descendant for post-order walk
3711 * @pos: the current position (%NULL to initiate traversal)
3712 * @root: css whose descendants to walk
3714 * To be used by css_for_each_descendant_post(). Find the next descendant
3715 * to visit for post-order traversal of @root's descendants. @root is
3716 * included in the iteration and the last node to be visited.
3718 * While this function requires cgroup_mutex or RCU read locking, it
3719 * doesn't require the whole traversal to be contained in a single critical
3720 * section. This function will return the correct next descendant as long
3721 * as both @pos and @cgroup are accessible and @pos is a descendant of
3724 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3725 * css which finished ->css_online() is guaranteed to be visible in the
3726 * future iterations and will stay visible until the last reference is put.
3727 * A css which hasn't finished ->css_online() or already finished
3728 * ->css_offline() may show up during traversal. It's each subsystem's
3729 * responsibility to synchronize against on/offlining.
3731 struct cgroup_subsys_state *
3732 css_next_descendant_post(struct cgroup_subsys_state *pos,
3733 struct cgroup_subsys_state *root)
3735 struct cgroup_subsys_state *next;
3737 cgroup_assert_mutex_or_rcu_locked();
3739 /* if first iteration, visit leftmost descendant which may be @root */
3741 return css_leftmost_descendant(root);
3743 /* if we visited @root, we're done */
3747 /* if there's an unvisited sibling, visit its leftmost descendant */
3748 next = css_next_child(pos, pos->parent);
3750 return css_leftmost_descendant(next);
3752 /* no sibling left, visit parent */
3757 * css_has_online_children - does a css have online children
3758 * @css: the target css
3760 * Returns %true if @css has any online children; otherwise, %false. This
3761 * function can be called from any context but the caller is responsible
3762 * for synchronizing against on/offlining as necessary.
3764 bool css_has_online_children(struct cgroup_subsys_state *css)
3766 struct cgroup_subsys_state *child;
3770 css_for_each_child(child, css) {
3771 if (child->flags & CSS_ONLINE) {
3781 * css_advance_task_iter - advance a task itererator to the next css_set
3782 * @it: the iterator to advance
3784 * Advance @it to the next css_set to walk.
3786 static void css_advance_task_iter(struct css_task_iter *it)
3788 struct list_head *l = it->cset_pos;
3789 struct cgrp_cset_link *link;
3790 struct css_set *cset;
3792 /* Advance to the next non-empty css_set */
3795 if (l == it->cset_head) {
3796 it->cset_pos = NULL;
3801 cset = container_of(l, struct css_set,
3802 e_cset_node[it->ss->id]);
3804 link = list_entry(l, struct cgrp_cset_link, cset_link);
3807 } while (!css_set_populated(cset));
3811 if (!list_empty(&cset->tasks))
3812 it->task_pos = cset->tasks.next;
3814 it->task_pos = cset->mg_tasks.next;
3816 it->tasks_head = &cset->tasks;
3817 it->mg_tasks_head = &cset->mg_tasks;
3821 * css_task_iter_start - initiate task iteration
3822 * @css: the css to walk tasks of
3823 * @it: the task iterator to use
3825 * Initiate iteration through the tasks of @css. The caller can call
3826 * css_task_iter_next() to walk through the tasks until the function
3827 * returns NULL. On completion of iteration, css_task_iter_end() must be
3830 * Note that this function acquires a lock which is released when the
3831 * iteration finishes. The caller can't sleep while iteration is in
3834 void css_task_iter_start(struct cgroup_subsys_state *css,
3835 struct css_task_iter *it)
3836 __acquires(css_set_rwsem)
3838 /* no one should try to iterate before mounting cgroups */
3839 WARN_ON_ONCE(!use_task_css_set_links);
3841 down_read(&css_set_rwsem);
3846 it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3848 it->cset_pos = &css->cgroup->cset_links;
3850 it->cset_head = it->cset_pos;
3852 css_advance_task_iter(it);
3856 * css_task_iter_next - return the next task for the iterator
3857 * @it: the task iterator being iterated
3859 * The "next" function for task iteration. @it should have been
3860 * initialized via css_task_iter_start(). Returns NULL when the iteration
3863 struct task_struct *css_task_iter_next(struct css_task_iter *it)
3865 struct task_struct *res;
3866 struct list_head *l = it->task_pos;
3868 /* If the iterator cg is NULL, we have no tasks */
3871 res = list_entry(l, struct task_struct, cg_list);
3874 * Advance iterator to find next entry. cset->tasks is consumed
3875 * first and then ->mg_tasks. After ->mg_tasks, we move onto the
3880 if (l == it->tasks_head)
3881 l = it->mg_tasks_head->next;
3883 if (l == it->mg_tasks_head)
3884 css_advance_task_iter(it);
3892 * css_task_iter_end - finish task iteration
3893 * @it: the task iterator to finish
3895 * Finish task iteration started by css_task_iter_start().
3897 void css_task_iter_end(struct css_task_iter *it)
3898 __releases(css_set_rwsem)
3900 up_read(&css_set_rwsem);
3904 * cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3905 * @to: cgroup to which the tasks will be moved
3906 * @from: cgroup in which the tasks currently reside
3908 * Locking rules between cgroup_post_fork() and the migration path
3909 * guarantee that, if a task is forking while being migrated, the new child
3910 * is guaranteed to be either visible in the source cgroup after the
3911 * parent's migration is complete or put into the target cgroup. No task
3912 * can slip out of migration through forking.
3914 int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
3916 LIST_HEAD(preloaded_csets);
3917 struct cgrp_cset_link *link;
3918 struct css_task_iter it;
3919 struct task_struct *task;
3922 mutex_lock(&cgroup_mutex);
3924 /* all tasks in @from are being moved, all csets are source */
3925 down_read(&css_set_rwsem);
3926 list_for_each_entry(link, &from->cset_links, cset_link)
3927 cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3928 up_read(&css_set_rwsem);
3930 ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3935 * Migrate tasks one-by-one until @form is empty. This fails iff
3936 * ->can_attach() fails.
3939 css_task_iter_start(&from->self, &it);
3940 task = css_task_iter_next(&it);
3942 get_task_struct(task);
3943 css_task_iter_end(&it);
3946 ret = cgroup_migrate(task, false, to);
3947 put_task_struct(task);
3949 } while (task && !ret);
3951 cgroup_migrate_finish(&preloaded_csets);
3952 mutex_unlock(&cgroup_mutex);
3957 * Stuff for reading the 'tasks'/'procs' files.
3959 * Reading this file can return large amounts of data if a cgroup has
3960 * *lots* of attached tasks. So it may need several calls to read(),
3961 * but we cannot guarantee that the information we produce is correct
3962 * unless we produce it entirely atomically.
3966 /* which pidlist file are we talking about? */
3967 enum cgroup_filetype {
3973 * A pidlist is a list of pids that virtually represents the contents of one
3974 * of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3975 * a pair (one each for procs, tasks) for each pid namespace that's relevant
3978 struct cgroup_pidlist {
3980 * used to find which pidlist is wanted. doesn't change as long as
3981 * this particular list stays in the list.
3983 struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3986 /* how many elements the above list has */
3988 /* each of these stored in a list by its cgroup */
3989 struct list_head links;
3990 /* pointer to the cgroup we belong to, for list removal purposes */
3991 struct cgroup *owner;
3992 /* for delayed destruction */
3993 struct delayed_work destroy_dwork;
3997 * The following two functions "fix" the issue where there are more pids
3998 * than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3999 * TODO: replace with a kernel-wide solution to this problem
4001 #define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
4002 static void *pidlist_allocate(int count)
4004 if (PIDLIST_TOO_LARGE(count))
4005 return vmalloc(count * sizeof(pid_t));
4007 return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
4010 static void pidlist_free(void *p)
4016 * Used to destroy all pidlists lingering waiting for destroy timer. None
4017 * should be left afterwards.
4019 static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
4021 struct cgroup_pidlist *l, *tmp_l;
4023 mutex_lock(&cgrp->pidlist_mutex);
4024 list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
4025 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
4026 mutex_unlock(&cgrp->pidlist_mutex);
4028 flush_workqueue(cgroup_pidlist_destroy_wq);
4029 BUG_ON(!list_empty(&cgrp->pidlists));
4032 static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
4034 struct delayed_work *dwork = to_delayed_work(work);
4035 struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
4037 struct cgroup_pidlist *tofree = NULL;
4039 mutex_lock(&l->owner->pidlist_mutex);
4042 * Destroy iff we didn't get queued again. The state won't change
4043 * as destroy_dwork can only be queued while locked.
4045 if (!delayed_work_pending(dwork)) {
4046 list_del(&l->links);
4047 pidlist_free(l->list);
4048 put_pid_ns(l->key.ns);
4052 mutex_unlock(&l->owner->pidlist_mutex);
4057 * pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
4058 * Returns the number of unique elements.
4060 static int pidlist_uniq(pid_t *list, int length)
4065 * we presume the 0th element is unique, so i starts at 1. trivial
4066 * edge cases first; no work needs to be done for either
4068 if (length == 0 || length == 1)
4070 /* src and dest walk down the list; dest counts unique elements */
4071 for (src = 1; src < length; src++) {
4072 /* find next unique element */
4073 while (list[src] == list[src-1]) {
4078 /* dest always points to where the next unique element goes */
4079 list[dest] = list[src];
4087 * The two pid files - task and cgroup.procs - guaranteed that the result
4088 * is sorted, which forced this whole pidlist fiasco. As pid order is
4089 * different per namespace, each namespace needs differently sorted list,
4090 * making it impossible to use, for example, single rbtree of member tasks
4091 * sorted by task pointer. As pidlists can be fairly large, allocating one
4092 * per open file is dangerous, so cgroup had to implement shared pool of
4093 * pidlists keyed by cgroup and namespace.
4095 * All this extra complexity was caused by the original implementation
4096 * committing to an entirely unnecessary property. In the long term, we
4097 * want to do away with it. Explicitly scramble sort order if on the
4098 * default hierarchy so that no such expectation exists in the new
4101 * Scrambling is done by swapping every two consecutive bits, which is
4102 * non-identity one-to-one mapping which disturbs sort order sufficiently.
4104 static pid_t pid_fry(pid_t pid)
4106 unsigned a = pid & 0x55555555;
4107 unsigned b = pid & 0xAAAAAAAA;
4109 return (a << 1) | (b >> 1);
4112 static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
4114 if (cgroup_on_dfl(cgrp))
4115 return pid_fry(pid);
4120 static int cmppid(const void *a, const void *b)
4122 return *(pid_t *)a - *(pid_t *)b;
4125 static int fried_cmppid(const void *a, const void *b)
4127 return pid_fry(*(pid_t *)a) - pid_fry(*(pid_t *)b);
4130 static struct cgroup_pidlist *cgroup_pidlist_find(struct cgroup *cgrp,
4131 enum cgroup_filetype type)
4133 struct cgroup_pidlist *l;
4134 /* don't need task_nsproxy() if we're looking at ourself */
4135 struct pid_namespace *ns = task_active_pid_ns(current);
4137 lockdep_assert_held(&cgrp->pidlist_mutex);
4139 list_for_each_entry(l, &cgrp->pidlists, links)
4140 if (l->key.type == type && l->key.ns == ns)
4146 * find the appropriate pidlist for our purpose (given procs vs tasks)
4147 * returns with the lock on that pidlist already held, and takes care
4148 * of the use count, or returns NULL with no locks held if we're out of
4151 static struct cgroup_pidlist *cgroup_pidlist_find_create(struct cgroup *cgrp,
4152 enum cgroup_filetype type)
4154 struct cgroup_pidlist *l;
4156 lockdep_assert_held(&cgrp->pidlist_mutex);
4158 l = cgroup_pidlist_find(cgrp, type);
4162 /* entry not found; create a new one */
4163 l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
4167 INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
4169 /* don't need task_nsproxy() if we're looking at ourself */
4170 l->key.ns = get_pid_ns(task_active_pid_ns(current));
4172 list_add(&l->links, &cgrp->pidlists);
4177 * Load a cgroup's pidarray with either procs' tgids or tasks' pids
4179 static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
4180 struct cgroup_pidlist **lp)
4184 int pid, n = 0; /* used for populating the array */
4185 struct css_task_iter it;
4186 struct task_struct *tsk;
4187 struct cgroup_pidlist *l;
4189 lockdep_assert_held(&cgrp->pidlist_mutex);
4192 * If cgroup gets more users after we read count, we won't have
4193 * enough space - tough. This race is indistinguishable to the
4194 * caller from the case that the additional cgroup users didn't
4195 * show up until sometime later on.
4197 length = cgroup_task_count(cgrp);
4198 array = pidlist_allocate(length);
4201 /* now, populate the array */
4202 css_task_iter_start(&cgrp->self, &it);
4203 while ((tsk = css_task_iter_next(&it))) {
4204 if (unlikely(n == length))
4206 /* get tgid or pid for procs or tasks file respectively */
4207 if (type == CGROUP_FILE_PROCS)
4208 pid = task_tgid_vnr(tsk);
4210 pid = task_pid_vnr(tsk);
4211 if (pid > 0) /* make sure to only use valid results */
4214 css_task_iter_end(&it);
4216 /* now sort & (if procs) strip out duplicates */
4217 if (cgroup_on_dfl(cgrp))
4218 sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4220 sort(array, length, sizeof(pid_t), cmppid, NULL);
4221 if (type == CGROUP_FILE_PROCS)
4222 length = pidlist_uniq(array, length);
4224 l = cgroup_pidlist_find_create(cgrp, type);
4226 pidlist_free(array);
4230 /* store array, freeing old if necessary */
4231 pidlist_free(l->list);
4239 * cgroupstats_build - build and fill cgroupstats
4240 * @stats: cgroupstats to fill information into
4241 * @dentry: A dentry entry belonging to the cgroup for which stats have
4244 * Build and fill cgroupstats so that taskstats can export it to user
4247 int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
4249 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4250 struct cgroup *cgrp;
4251 struct css_task_iter it;
4252 struct task_struct *tsk;
4254 /* it should be kernfs_node belonging to cgroupfs and is a directory */
4255 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
4256 kernfs_type(kn) != KERNFS_DIR)
4259 mutex_lock(&cgroup_mutex);
4262 * We aren't being called from kernfs and there's no guarantee on
4263 * @kn->priv's validity. For this and css_tryget_online_from_dir(),
4264 * @kn->priv is RCU safe. Let's do the RCU dancing.
4267 cgrp = rcu_dereference(kn->priv);
4268 if (!cgrp || cgroup_is_dead(cgrp)) {
4270 mutex_unlock(&cgroup_mutex);
4275 css_task_iter_start(&cgrp->self, &it);
4276 while ((tsk = css_task_iter_next(&it))) {
4277 switch (tsk->state) {
4279 stats->nr_running++;
4281 case TASK_INTERRUPTIBLE:
4282 stats->nr_sleeping++;
4284 case TASK_UNINTERRUPTIBLE:
4285 stats->nr_uninterruptible++;
4288 stats->nr_stopped++;
4291 if (delayacct_is_task_waiting_on_io(tsk))
4292 stats->nr_io_wait++;
4296 css_task_iter_end(&it);
4298 mutex_unlock(&cgroup_mutex);
4304 * seq_file methods for the tasks/procs files. The seq_file position is the
4305 * next pid to display; the seq_file iterator is a pointer to the pid
4306 * in the cgroup->l->list array.
4309 static void *cgroup_pidlist_start(struct seq_file *s, loff_t *pos)
4312 * Initially we receive a position value that corresponds to
4313 * one more than the last pid shown (or 0 on the first call or
4314 * after a seek to the start). Use a binary-search to find the
4315 * next pid to display, if any
4317 struct kernfs_open_file *of = s->private;
4318 struct cgroup *cgrp = seq_css(s)->cgroup;
4319 struct cgroup_pidlist *l;
4320 enum cgroup_filetype type = seq_cft(s)->private;
4321 int index = 0, pid = *pos;
4324 mutex_lock(&cgrp->pidlist_mutex);
4327 * !NULL @of->priv indicates that this isn't the first start()
4328 * after open. If the matching pidlist is around, we can use that.
4329 * Look for it. Note that @of->priv can't be used directly. It
4330 * could already have been destroyed.
4333 of->priv = cgroup_pidlist_find(cgrp, type);
4336 * Either this is the first start() after open or the matching
4337 * pidlist has been destroyed inbetween. Create a new one.
4340 ret = pidlist_array_load(cgrp, type,
4341 (struct cgroup_pidlist **)&of->priv);
4343 return ERR_PTR(ret);
4348 int end = l->length;
4350 while (index < end) {
4351 int mid = (index + end) / 2;
4352 if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4355 } else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4361 /* If we're off the end of the array, we're done */
4362 if (index >= l->length)
4364 /* Update the abstract position to be the actual pid that we found */
4365 iter = l->list + index;
4366 *pos = cgroup_pid_fry(cgrp, *iter);
4370 static void cgroup_pidlist_stop(struct seq_file *s, void *v)
4372 struct kernfs_open_file *of = s->private;
4373 struct cgroup_pidlist *l = of->priv;
4376 mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4377 CGROUP_PIDLIST_DESTROY_DELAY);
4378 mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4381 static void *cgroup_pidlist_next(struct seq_file *s, void *v, loff_t *pos)
4383 struct kernfs_open_file *of = s->private;
4384 struct cgroup_pidlist *l = of->priv;
4386 pid_t *end = l->list + l->length;
4388 * Advance to the next pid in the array. If this goes off the
4395 *pos = cgroup_pid_fry(seq_css(s)->cgroup, *p);
4400 static int cgroup_pidlist_show(struct seq_file *s, void *v)
4402 seq_printf(s, "%d\n", *(int *)v);
4407 static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4410 return notify_on_release(css->cgroup);
4413 static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4414 struct cftype *cft, u64 val)
4417 set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4419 clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4423 static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4426 return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4429 static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4430 struct cftype *cft, u64 val)
4433 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4435 clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4439 /* cgroup core interface files for the default hierarchy */
4440 static struct cftype cgroup_dfl_base_files[] = {
4442 .name = "cgroup.procs",
4443 .file_offset = offsetof(struct cgroup, procs_file),
4444 .seq_start = cgroup_pidlist_start,
4445 .seq_next = cgroup_pidlist_next,
4446 .seq_stop = cgroup_pidlist_stop,
4447 .seq_show = cgroup_pidlist_show,
4448 .private = CGROUP_FILE_PROCS,
4449 .write = cgroup_procs_write,
4452 .name = "cgroup.controllers",
4453 .flags = CFTYPE_ONLY_ON_ROOT,
4454 .seq_show = cgroup_root_controllers_show,
4457 .name = "cgroup.controllers",
4458 .flags = CFTYPE_NOT_ON_ROOT,
4459 .seq_show = cgroup_controllers_show,
4462 .name = "cgroup.subtree_control",
4463 .seq_show = cgroup_subtree_control_show,
4464 .write = cgroup_subtree_control_write,
4467 .name = "cgroup.events",
4468 .flags = CFTYPE_NOT_ON_ROOT,
4469 .file_offset = offsetof(struct cgroup, events_file),
4470 .seq_show = cgroup_events_show,
4475 /* cgroup core interface files for the legacy hierarchies */
4476 static struct cftype cgroup_legacy_base_files[] = {
4478 .name = "cgroup.procs",
4479 .seq_start = cgroup_pidlist_start,
4480 .seq_next = cgroup_pidlist_next,
4481 .seq_stop = cgroup_pidlist_stop,
4482 .seq_show = cgroup_pidlist_show,
4483 .private = CGROUP_FILE_PROCS,
4484 .write = cgroup_procs_write,
4487 .name = "cgroup.clone_children",
4488 .read_u64 = cgroup_clone_children_read,
4489 .write_u64 = cgroup_clone_children_write,
4492 .name = "cgroup.sane_behavior",
4493 .flags = CFTYPE_ONLY_ON_ROOT,
4494 .seq_show = cgroup_sane_behavior_show,
4498 .seq_start = cgroup_pidlist_start,
4499 .seq_next = cgroup_pidlist_next,
4500 .seq_stop = cgroup_pidlist_stop,
4501 .seq_show = cgroup_pidlist_show,
4502 .private = CGROUP_FILE_TASKS,
4503 .write = cgroup_tasks_write,
4506 .name = "notify_on_release",
4507 .read_u64 = cgroup_read_notify_on_release,
4508 .write_u64 = cgroup_write_notify_on_release,
4511 .name = "release_agent",
4512 .flags = CFTYPE_ONLY_ON_ROOT,
4513 .seq_show = cgroup_release_agent_show,
4514 .write = cgroup_release_agent_write,
4515 .max_write_len = PATH_MAX - 1,
4521 * css destruction is four-stage process.
4523 * 1. Destruction starts. Killing of the percpu_ref is initiated.
4524 * Implemented in kill_css().
4526 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4527 * and thus css_tryget_online() is guaranteed to fail, the css can be
4528 * offlined by invoking offline_css(). After offlining, the base ref is
4529 * put. Implemented in css_killed_work_fn().
4531 * 3. When the percpu_ref reaches zero, the only possible remaining
4532 * accessors are inside RCU read sections. css_release() schedules the
4535 * 4. After the grace period, the css can be freed. Implemented in
4536 * css_free_work_fn().
4538 * It is actually hairier because both step 2 and 4 require process context
4539 * and thus involve punting to css->destroy_work adding two additional
4540 * steps to the already complex sequence.
4542 static void css_free_work_fn(struct work_struct *work)
4544 struct cgroup_subsys_state *css =
4545 container_of(work, struct cgroup_subsys_state, destroy_work);
4546 struct cgroup_subsys *ss = css->ss;
4547 struct cgroup *cgrp = css->cgroup;
4548 struct cgroup_file *cfile;
4550 percpu_ref_exit(&css->refcnt);
4552 list_for_each_entry(cfile, &css->files, node)
4553 kernfs_put(cfile->kn);
4560 css_put(css->parent);
4563 cgroup_idr_remove(&ss->css_idr, id);
4566 /* cgroup free path */
4567 atomic_dec(&cgrp->root->nr_cgrps);
4568 cgroup_pidlist_destroy_all(cgrp);
4569 cancel_work_sync(&cgrp->release_agent_work);
4571 if (cgroup_parent(cgrp)) {
4573 * We get a ref to the parent, and put the ref when
4574 * this cgroup is being freed, so it's guaranteed
4575 * that the parent won't be destroyed before its
4578 cgroup_put(cgroup_parent(cgrp));
4579 kernfs_put(cgrp->kn);
4583 * This is root cgroup's refcnt reaching zero,
4584 * which indicates that the root should be
4587 cgroup_destroy_root(cgrp->root);
4592 static void css_free_rcu_fn(struct rcu_head *rcu_head)
4594 struct cgroup_subsys_state *css =
4595 container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4597 INIT_WORK(&css->destroy_work, css_free_work_fn);
4598 queue_work(cgroup_destroy_wq, &css->destroy_work);
4601 static void css_release_work_fn(struct work_struct *work)
4603 struct cgroup_subsys_state *css =
4604 container_of(work, struct cgroup_subsys_state, destroy_work);
4605 struct cgroup_subsys *ss = css->ss;
4606 struct cgroup *cgrp = css->cgroup;
4608 mutex_lock(&cgroup_mutex);
4610 css->flags |= CSS_RELEASED;
4611 list_del_rcu(&css->sibling);
4614 /* css release path */
4615 cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4616 if (ss->css_released)
4617 ss->css_released(css);
4619 /* cgroup release path */
4620 cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4624 * There are two control paths which try to determine
4625 * cgroup from dentry without going through kernfs -
4626 * cgroupstats_build() and css_tryget_online_from_dir().
4627 * Those are supported by RCU protecting clearing of
4628 * cgrp->kn->priv backpointer.
4630 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, NULL);
4633 mutex_unlock(&cgroup_mutex);
4635 call_rcu(&css->rcu_head, css_free_rcu_fn);
4638 static void css_release(struct percpu_ref *ref)
4640 struct cgroup_subsys_state *css =
4641 container_of(ref, struct cgroup_subsys_state, refcnt);
4643 INIT_WORK(&css->destroy_work, css_release_work_fn);
4644 queue_work(cgroup_destroy_wq, &css->destroy_work);
4647 static void init_and_link_css(struct cgroup_subsys_state *css,
4648 struct cgroup_subsys *ss, struct cgroup *cgrp)
4650 lockdep_assert_held(&cgroup_mutex);
4654 memset(css, 0, sizeof(*css));
4657 INIT_LIST_HEAD(&css->sibling);
4658 INIT_LIST_HEAD(&css->children);
4659 INIT_LIST_HEAD(&css->files);
4660 css->serial_nr = css_serial_nr_next++;
4662 if (cgroup_parent(cgrp)) {
4663 css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4664 css_get(css->parent);
4667 BUG_ON(cgroup_css(cgrp, ss));
4670 /* invoke ->css_online() on a new CSS and mark it online if successful */
4671 static int online_css(struct cgroup_subsys_state *css)
4673 struct cgroup_subsys *ss = css->ss;
4676 lockdep_assert_held(&cgroup_mutex);
4679 ret = ss->css_online(css);
4681 css->flags |= CSS_ONLINE;
4682 rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4687 /* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4688 static void offline_css(struct cgroup_subsys_state *css)
4690 struct cgroup_subsys *ss = css->ss;
4692 lockdep_assert_held(&cgroup_mutex);
4694 if (!(css->flags & CSS_ONLINE))
4697 if (ss->css_offline)
4698 ss->css_offline(css);
4700 css->flags &= ~CSS_ONLINE;
4701 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4703 wake_up_all(&css->cgroup->offline_waitq);
4707 * create_css - create a cgroup_subsys_state
4708 * @cgrp: the cgroup new css will be associated with
4709 * @ss: the subsys of new css
4710 * @visible: whether to create control knobs for the new css or not
4712 * Create a new css associated with @cgrp - @ss pair. On success, the new
4713 * css is online and installed in @cgrp with all interface files created if
4714 * @visible. Returns 0 on success, -errno on failure.
4716 static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
4719 struct cgroup *parent = cgroup_parent(cgrp);
4720 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4721 struct cgroup_subsys_state *css;
4724 lockdep_assert_held(&cgroup_mutex);
4726 css = ss->css_alloc(parent_css);
4728 return PTR_ERR(css);
4730 init_and_link_css(css, ss, cgrp);
4732 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4736 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4738 goto err_free_percpu_ref;
4742 err = css_populate_dir(css, NULL);
4747 /* @css is ready to be brought online now, make it visible */
4748 list_add_tail_rcu(&css->sibling, &parent_css->children);
4749 cgroup_idr_replace(&ss->css_idr, css, css->id);
4751 err = online_css(css);
4755 if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4756 cgroup_parent(parent)) {
4757 pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4758 current->comm, current->pid, ss->name);
4759 if (!strcmp(ss->name, "memory"))
4760 pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4761 ss->warned_broken_hierarchy = true;
4767 list_del_rcu(&css->sibling);
4768 css_clear_dir(css, NULL);
4770 cgroup_idr_remove(&ss->css_idr, css->id);
4771 err_free_percpu_ref:
4772 percpu_ref_exit(&css->refcnt);
4774 call_rcu(&css->rcu_head, css_free_rcu_fn);
4778 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
4781 struct cgroup *parent, *cgrp;
4782 struct cgroup_root *root;
4783 struct cgroup_subsys *ss;
4784 struct kernfs_node *kn;
4787 /* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4789 if (strchr(name, '\n'))
4792 parent = cgroup_kn_lock_live(parent_kn);
4795 root = parent->root;
4797 /* allocate the cgroup and its ID, 0 is reserved for the root */
4798 cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4804 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4809 * Temporarily set the pointer to NULL, so idr_find() won't return
4810 * a half-baked cgroup.
4812 cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4815 goto out_cancel_ref;
4818 init_cgroup_housekeeping(cgrp);
4820 cgrp->self.parent = &parent->self;
4823 if (notify_on_release(parent))
4824 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4826 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4827 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4829 /* create the directory */
4830 kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4838 * This extra ref will be put in cgroup_free_fn() and guarantees
4839 * that @cgrp->kn is always accessible.
4843 cgrp->self.serial_nr = css_serial_nr_next++;
4845 /* allocation complete, commit to creation */
4846 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4847 atomic_inc(&root->nr_cgrps);
4851 * @cgrp is now fully operational. If something fails after this
4852 * point, it'll be released via the normal destruction path.
4854 cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4856 ret = cgroup_kn_set_ugid(kn);
4860 ret = css_populate_dir(&cgrp->self, NULL);
4864 /* let's create and online css's */
4865 for_each_subsys(ss, ssid) {
4866 if (parent->child_subsys_mask & (1 << ssid)) {
4867 ret = create_css(cgrp, ss,
4868 parent->subtree_control & (1 << ssid));
4875 * On the default hierarchy, a child doesn't automatically inherit
4876 * subtree_control from the parent. Each is configured manually.
4878 if (!cgroup_on_dfl(cgrp)) {
4879 cgrp->subtree_control = parent->subtree_control;
4880 cgroup_refresh_child_subsys_mask(cgrp);
4883 kernfs_activate(kn);
4889 cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4891 percpu_ref_exit(&cgrp->self.refcnt);
4895 cgroup_kn_unlock(parent_kn);
4899 cgroup_destroy_locked(cgrp);
4904 * This is called when the refcnt of a css is confirmed to be killed.
4905 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4906 * initate destruction and put the css ref from kill_css().
4908 static void css_killed_work_fn(struct work_struct *work)
4910 struct cgroup_subsys_state *css =
4911 container_of(work, struct cgroup_subsys_state, destroy_work);
4913 mutex_lock(&cgroup_mutex);
4915 mutex_unlock(&cgroup_mutex);
4920 /* css kill confirmation processing requires process context, bounce */
4921 static void css_killed_ref_fn(struct percpu_ref *ref)
4923 struct cgroup_subsys_state *css =
4924 container_of(ref, struct cgroup_subsys_state, refcnt);
4926 INIT_WORK(&css->destroy_work, css_killed_work_fn);
4927 queue_work(cgroup_destroy_wq, &css->destroy_work);
4931 * kill_css - destroy a css
4932 * @css: css to destroy
4934 * This function initiates destruction of @css by removing cgroup interface
4935 * files and putting its base reference. ->css_offline() will be invoked
4936 * asynchronously once css_tryget_online() is guaranteed to fail and when
4937 * the reference count reaches zero, @css will be released.
4939 static void kill_css(struct cgroup_subsys_state *css)
4941 lockdep_assert_held(&cgroup_mutex);
4944 * This must happen before css is disassociated with its cgroup.
4945 * See seq_css() for details.
4947 css_clear_dir(css, NULL);
4950 * Killing would put the base ref, but we need to keep it alive
4951 * until after ->css_offline().
4956 * cgroup core guarantees that, by the time ->css_offline() is
4957 * invoked, no new css reference will be given out via
4958 * css_tryget_online(). We can't simply call percpu_ref_kill() and
4959 * proceed to offlining css's because percpu_ref_kill() doesn't
4960 * guarantee that the ref is seen as killed on all CPUs on return.
4962 * Use percpu_ref_kill_and_confirm() to get notifications as each
4963 * css is confirmed to be seen as killed on all CPUs.
4965 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4969 * cgroup_destroy_locked - the first stage of cgroup destruction
4970 * @cgrp: cgroup to be destroyed
4972 * css's make use of percpu refcnts whose killing latency shouldn't be
4973 * exposed to userland and are RCU protected. Also, cgroup core needs to
4974 * guarantee that css_tryget_online() won't succeed by the time
4975 * ->css_offline() is invoked. To satisfy all the requirements,
4976 * destruction is implemented in the following two steps.
4978 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4979 * userland visible parts and start killing the percpu refcnts of
4980 * css's. Set up so that the next stage will be kicked off once all
4981 * the percpu refcnts are confirmed to be killed.
4983 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4984 * rest of destruction. Once all cgroup references are gone, the
4985 * cgroup is RCU-freed.
4987 * This function implements s1. After this step, @cgrp is gone as far as
4988 * the userland is concerned and a new cgroup with the same name may be
4989 * created. As cgroup doesn't care about the names internally, this
4990 * doesn't cause any problem.
4992 static int cgroup_destroy_locked(struct cgroup *cgrp)
4993 __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4995 struct cgroup_subsys_state *css;
4999 lockdep_assert_held(&cgroup_mutex);
5002 * css_set_rwsem synchronizes access to ->cset_links and prevents
5003 * @cgrp from being removed while put_css_set() is in progress.
5005 down_read(&css_set_rwsem);
5006 empty = list_empty(&cgrp->cset_links);
5007 up_read(&css_set_rwsem);
5012 * Make sure there's no live children. We can't test emptiness of
5013 * ->self.children as dead children linger on it while being
5014 * drained; otherwise, "rmdir parent/child parent" may fail.
5016 if (css_has_online_children(&cgrp->self))
5020 * Mark @cgrp dead. This prevents further task migration and child
5021 * creation by disabling cgroup_lock_live_group().
5023 cgrp->self.flags &= ~CSS_ONLINE;
5025 /* initiate massacre of all css's */
5026 for_each_css(css, ssid, cgrp)
5030 * Remove @cgrp directory along with the base files. @cgrp has an
5031 * extra ref on its kn.
5033 kernfs_remove(cgrp->kn);
5035 check_for_release(cgroup_parent(cgrp));
5037 /* put the base reference */
5038 percpu_ref_kill(&cgrp->self.refcnt);
5043 static int cgroup_rmdir(struct kernfs_node *kn)
5045 struct cgroup *cgrp;
5048 cgrp = cgroup_kn_lock_live(kn);
5052 ret = cgroup_destroy_locked(cgrp);
5054 cgroup_kn_unlock(kn);
5058 static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5059 .remount_fs = cgroup_remount,
5060 .show_options = cgroup_show_options,
5061 .mkdir = cgroup_mkdir,
5062 .rmdir = cgroup_rmdir,
5063 .rename = cgroup_rename,
5066 static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5068 struct cgroup_subsys_state *css;
5070 printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
5072 mutex_lock(&cgroup_mutex);
5074 idr_init(&ss->css_idr);
5075 INIT_LIST_HEAD(&ss->cfts);
5077 /* Create the root cgroup state for this subsystem */
5078 ss->root = &cgrp_dfl_root;
5079 css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5080 /* We don't handle early failures gracefully */
5081 BUG_ON(IS_ERR(css));
5082 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5085 * Root csses are never destroyed and we can't initialize
5086 * percpu_ref during early init. Disable refcnting.
5088 css->flags |= CSS_NO_REF;
5091 /* allocation can't be done safely during early init */
5094 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5095 BUG_ON(css->id < 0);
5098 /* Update the init_css_set to contain a subsys
5099 * pointer to this state - since the subsystem is
5100 * newly registered, all tasks and hence the
5101 * init_css_set is in the subsystem's root cgroup. */
5102 init_css_set.subsys[ss->id] = css;
5104 have_fork_callback |= (bool)ss->fork << ss->id;
5105 have_exit_callback |= (bool)ss->exit << ss->id;
5106 have_canfork_callback |= (bool)ss->can_fork << ss->id;
5108 /* At system boot, before all subsystems have been
5109 * registered, no tasks have been forked, so we don't
5110 * need to invoke fork callbacks here. */
5111 BUG_ON(!list_empty(&init_task.tasks));
5113 BUG_ON(online_css(css));
5115 mutex_unlock(&cgroup_mutex);
5119 * cgroup_init_early - cgroup initialization at system boot
5121 * Initialize cgroups at system boot, and initialize any
5122 * subsystems that request early init.
5124 int __init cgroup_init_early(void)
5126 static struct cgroup_sb_opts __initdata opts;
5127 struct cgroup_subsys *ss;
5130 init_cgroup_root(&cgrp_dfl_root, &opts);
5131 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5133 RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5135 for_each_subsys(ss, i) {
5136 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5137 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
5138 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5140 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5141 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5144 ss->name = cgroup_subsys_name[i];
5145 if (!ss->legacy_name)
5146 ss->legacy_name = cgroup_subsys_name[i];
5149 cgroup_init_subsys(ss, true);
5154 static unsigned long cgroup_disable_mask __initdata;
5157 * cgroup_init - cgroup initialization
5159 * Register cgroup filesystem and /proc file, and initialize
5160 * any subsystems that didn't request early init.
5162 int __init cgroup_init(void)
5164 struct cgroup_subsys *ss;
5168 BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5169 BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5170 BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5172 mutex_lock(&cgroup_mutex);
5174 /* Add init_css_set to the hash table */
5175 key = css_set_hash(init_css_set.subsys);
5176 hash_add(css_set_table, &init_css_set.hlist, key);
5178 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5180 mutex_unlock(&cgroup_mutex);
5182 for_each_subsys(ss, ssid) {
5183 if (ss->early_init) {
5184 struct cgroup_subsys_state *css =
5185 init_css_set.subsys[ss->id];
5187 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5189 BUG_ON(css->id < 0);
5191 cgroup_init_subsys(ss, false);
5194 list_add_tail(&init_css_set.e_cset_node[ssid],
5195 &cgrp_dfl_root.cgrp.e_csets[ssid]);
5198 * Setting dfl_root subsys_mask needs to consider the
5199 * disabled flag and cftype registration needs kmalloc,
5200 * both of which aren't available during early_init.
5202 if (cgroup_disable_mask & (1 << ssid)) {
5203 static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5204 printk(KERN_INFO "Disabling %s control group subsystem\n",
5209 cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5211 if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5212 ss->dfl_cftypes = ss->legacy_cftypes;
5214 if (!ss->dfl_cftypes)
5215 cgrp_dfl_root_inhibit_ss_mask |= 1 << ss->id;
5217 if (ss->dfl_cftypes == ss->legacy_cftypes) {
5218 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5220 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5221 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5225 ss->bind(init_css_set.subsys[ssid]);
5228 err = sysfs_create_mount_point(fs_kobj, "cgroup");
5232 err = register_filesystem(&cgroup_fs_type);
5234 sysfs_remove_mount_point(fs_kobj, "cgroup");
5238 proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5242 static int __init cgroup_wq_init(void)
5245 * There isn't much point in executing destruction path in
5246 * parallel. Good chunk is serialized with cgroup_mutex anyway.
5247 * Use 1 for @max_active.
5249 * We would prefer to do this in cgroup_init() above, but that
5250 * is called before init_workqueues(): so leave this until after.
5252 cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5253 BUG_ON(!cgroup_destroy_wq);
5256 * Used to destroy pidlists and separate to serve as flush domain.
5257 * Cap @max_active to 1 too.
5259 cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5261 BUG_ON(!cgroup_pidlist_destroy_wq);
5265 core_initcall(cgroup_wq_init);
5268 * proc_cgroup_show()
5269 * - Print task's cgroup paths into seq_file, one line for each hierarchy
5270 * - Used for /proc/<pid>/cgroup.
5272 int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5273 struct pid *pid, struct task_struct *tsk)
5277 struct cgroup_root *root;
5280 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5284 mutex_lock(&cgroup_mutex);
5285 down_read(&css_set_rwsem);
5287 for_each_root(root) {
5288 struct cgroup_subsys *ss;
5289 struct cgroup *cgrp;
5290 int ssid, count = 0;
5292 if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5295 seq_printf(m, "%d:", root->hierarchy_id);
5296 if (root != &cgrp_dfl_root)
5297 for_each_subsys(ss, ssid)
5298 if (root->subsys_mask & (1 << ssid))
5299 seq_printf(m, "%s%s", count++ ? "," : "",
5301 if (strlen(root->name))
5302 seq_printf(m, "%sname=%s", count ? "," : "",
5305 cgrp = task_cgroup_from_root(tsk, root);
5306 path = cgroup_path(cgrp, buf, PATH_MAX);
5308 retval = -ENAMETOOLONG;
5317 up_read(&css_set_rwsem);
5318 mutex_unlock(&cgroup_mutex);
5324 /* Display information about each subsystem and each hierarchy */
5325 static int proc_cgroupstats_show(struct seq_file *m, void *v)
5327 struct cgroup_subsys *ss;
5330 seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5332 * ideally we don't want subsystems moving around while we do this.
5333 * cgroup_mutex is also necessary to guarantee an atomic snapshot of
5334 * subsys/hierarchy state.
5336 mutex_lock(&cgroup_mutex);
5338 for_each_subsys(ss, i)
5339 seq_printf(m, "%s\t%d\t%d\t%d\n",
5340 ss->legacy_name, ss->root->hierarchy_id,
5341 atomic_read(&ss->root->nr_cgrps),
5342 cgroup_ssid_enabled(i));
5344 mutex_unlock(&cgroup_mutex);
5348 static int cgroupstats_open(struct inode *inode, struct file *file)
5350 return single_open(file, proc_cgroupstats_show, NULL);
5353 static const struct file_operations proc_cgroupstats_operations = {
5354 .open = cgroupstats_open,
5356 .llseek = seq_lseek,
5357 .release = single_release,
5360 static void **subsys_canfork_priv_p(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5362 if (CGROUP_CANFORK_START <= i && i < CGROUP_CANFORK_END)
5363 return &ss_priv[i - CGROUP_CANFORK_START];
5367 static void *subsys_canfork_priv(void *ss_priv[CGROUP_CANFORK_COUNT], int i)
5369 void **private = subsys_canfork_priv_p(ss_priv, i);
5370 return private ? *private : NULL;
5374 * cgroup_fork - initialize cgroup related fields during copy_process()
5375 * @child: pointer to task_struct of forking parent process.
5377 * A task is associated with the init_css_set until cgroup_post_fork()
5378 * attaches it to the parent's css_set. Empty cg_list indicates that
5379 * @child isn't holding reference to its css_set.
5381 void cgroup_fork(struct task_struct *child)
5383 RCU_INIT_POINTER(child->cgroups, &init_css_set);
5384 INIT_LIST_HEAD(&child->cg_list);
5388 * cgroup_can_fork - called on a new task before the process is exposed
5389 * @child: the task in question.
5391 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5392 * returns an error, the fork aborts with that error code. This allows for
5393 * a cgroup subsystem to conditionally allow or deny new forks.
5395 int cgroup_can_fork(struct task_struct *child,
5396 void *ss_priv[CGROUP_CANFORK_COUNT])
5398 struct cgroup_subsys *ss;
5401 for_each_subsys_which(ss, i, &have_canfork_callback) {
5402 ret = ss->can_fork(child, subsys_canfork_priv_p(ss_priv, i));
5410 for_each_subsys(ss, j) {
5413 if (ss->cancel_fork)
5414 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, j));
5421 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5422 * @child: the task in question
5424 * This calls the cancel_fork() callbacks if a fork failed *after*
5425 * cgroup_can_fork() succeded.
5427 void cgroup_cancel_fork(struct task_struct *child,
5428 void *ss_priv[CGROUP_CANFORK_COUNT])
5430 struct cgroup_subsys *ss;
5433 for_each_subsys(ss, i)
5434 if (ss->cancel_fork)
5435 ss->cancel_fork(child, subsys_canfork_priv(ss_priv, i));
5439 * cgroup_post_fork - called on a new task after adding it to the task list
5440 * @child: the task in question
5442 * Adds the task to the list running through its css_set if necessary and
5443 * call the subsystem fork() callbacks. Has to be after the task is
5444 * visible on the task list in case we race with the first call to
5445 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5448 void cgroup_post_fork(struct task_struct *child,
5449 void *old_ss_priv[CGROUP_CANFORK_COUNT])
5451 struct cgroup_subsys *ss;
5455 * This may race against cgroup_enable_task_cg_lists(). As that
5456 * function sets use_task_css_set_links before grabbing
5457 * tasklist_lock and we just went through tasklist_lock to add
5458 * @child, it's guaranteed that either we see the set
5459 * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5460 * @child during its iteration.
5462 * If we won the race, @child is associated with %current's
5463 * css_set. Grabbing css_set_rwsem guarantees both that the
5464 * association is stable, and, on completion of the parent's
5465 * migration, @child is visible in the source of migration or
5466 * already in the destination cgroup. This guarantee is necessary
5467 * when implementing operations which need to migrate all tasks of
5468 * a cgroup to another.
5470 * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5471 * will remain in init_css_set. This is safe because all tasks are
5472 * in the init_css_set before cg_links is enabled and there's no
5473 * operation which transfers all tasks out of init_css_set.
5475 if (use_task_css_set_links) {
5476 struct css_set *cset;
5478 down_write(&css_set_rwsem);
5479 cset = task_css_set(current);
5480 if (list_empty(&child->cg_list)) {
5481 rcu_assign_pointer(child->cgroups, cset);
5482 list_add(&child->cg_list, &cset->tasks);
5485 up_write(&css_set_rwsem);
5489 * Call ss->fork(). This must happen after @child is linked on
5490 * css_set; otherwise, @child might change state between ->fork()
5491 * and addition to css_set.
5493 for_each_subsys_which(ss, i, &have_fork_callback)
5494 ss->fork(child, subsys_canfork_priv(old_ss_priv, i));
5498 * cgroup_exit - detach cgroup from exiting task
5499 * @tsk: pointer to task_struct of exiting process
5501 * Description: Detach cgroup from @tsk and release it.
5503 * Note that cgroups marked notify_on_release force every task in
5504 * them to take the global cgroup_mutex mutex when exiting.
5505 * This could impact scaling on very large systems. Be reluctant to
5506 * use notify_on_release cgroups where very high task exit scaling
5507 * is required on large systems.
5509 * We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5510 * call cgroup_exit() while the task is still competent to handle
5511 * notify_on_release(), then leave the task attached to the root cgroup in
5512 * each hierarchy for the remainder of its exit. No need to bother with
5513 * init_css_set refcnting. init_css_set never goes away and we can't race
5514 * with migration path - PF_EXITING is visible to migration path.
5516 void cgroup_exit(struct task_struct *tsk)
5518 struct cgroup_subsys *ss;
5519 struct css_set *cset;
5520 bool put_cset = false;
5524 * Unlink from @tsk from its css_set. As migration path can't race
5525 * with us, we can check css_set and cg_list without synchronization.
5527 cset = task_css_set(tsk);
5529 if (!list_empty(&tsk->cg_list)) {
5530 down_write(&css_set_rwsem);
5531 list_del_init(&tsk->cg_list);
5532 if (!css_set_populated(cset))
5533 css_set_update_populated(cset, false);
5534 up_write(&css_set_rwsem);
5538 /* Reassign the task to the init_css_set. */
5539 RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5541 /* see cgroup_post_fork() for details */
5542 for_each_subsys_which(ss, i, &have_exit_callback) {
5543 struct cgroup_subsys_state *old_css = cset->subsys[i];
5544 struct cgroup_subsys_state *css = task_css(tsk, i);
5546 ss->exit(css, old_css, tsk);
5553 static void check_for_release(struct cgroup *cgrp)
5555 if (notify_on_release(cgrp) && !cgroup_is_populated(cgrp) &&
5556 !css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5557 schedule_work(&cgrp->release_agent_work);
5561 * Notify userspace when a cgroup is released, by running the
5562 * configured release agent with the name of the cgroup (path
5563 * relative to the root of cgroup file system) as the argument.
5565 * Most likely, this user command will try to rmdir this cgroup.
5567 * This races with the possibility that some other task will be
5568 * attached to this cgroup before it is removed, or that some other
5569 * user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5570 * The presumed 'rmdir' will fail quietly if this cgroup is no longer
5571 * unused, and this cgroup will be reprieved from its death sentence,
5572 * to continue to serve a useful existence. Next time it's released,
5573 * we will get notified again, if it still has 'notify_on_release' set.
5575 * The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5576 * means only wait until the task is successfully execve()'d. The
5577 * separate release agent task is forked by call_usermodehelper(),
5578 * then control in this thread returns here, without waiting for the
5579 * release agent task. We don't bother to wait because the caller of
5580 * this routine has no use for the exit status of the release agent
5581 * task, so no sense holding our caller up for that.
5583 static void cgroup_release_agent(struct work_struct *work)
5585 struct cgroup *cgrp =
5586 container_of(work, struct cgroup, release_agent_work);
5587 char *pathbuf = NULL, *agentbuf = NULL, *path;
5588 char *argv[3], *envp[3];
5590 mutex_lock(&cgroup_mutex);
5592 pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5593 agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5594 if (!pathbuf || !agentbuf)
5597 path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5605 /* minimal command environment */
5607 envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5610 mutex_unlock(&cgroup_mutex);
5611 call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5614 mutex_unlock(&cgroup_mutex);
5620 static int __init cgroup_disable(char *str)
5622 struct cgroup_subsys *ss;
5626 while ((token = strsep(&str, ",")) != NULL) {
5630 for_each_subsys(ss, i) {
5631 if (strcmp(token, ss->name) &&
5632 strcmp(token, ss->legacy_name))
5634 cgroup_disable_mask |= 1 << i;
5639 __setup("cgroup_disable=", cgroup_disable);
5641 static int __init cgroup_set_legacy_files_on_dfl(char *str)
5643 printk("cgroup: using legacy files on the default hierarchy\n");
5644 cgroup_legacy_files_on_dfl = true;
5647 __setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5650 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5651 * @dentry: directory dentry of interest
5652 * @ss: subsystem of interest
5654 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5655 * to get the corresponding css and return it. If such css doesn't exist
5656 * or can't be pinned, an ERR_PTR value is returned.
5658 struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5659 struct cgroup_subsys *ss)
5661 struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5662 struct cgroup_subsys_state *css = NULL;
5663 struct cgroup *cgrp;
5665 /* is @dentry a cgroup dir? */
5666 if (dentry->d_sb->s_type != &cgroup_fs_type || !kn ||
5667 kernfs_type(kn) != KERNFS_DIR)
5668 return ERR_PTR(-EBADF);
5673 * This path doesn't originate from kernfs and @kn could already
5674 * have been or be removed at any point. @kn->priv is RCU
5675 * protected for this access. See css_release_work_fn() for details.
5677 cgrp = rcu_dereference(kn->priv);
5679 css = cgroup_css(cgrp, ss);
5681 if (!css || !css_tryget_online(css))
5682 css = ERR_PTR(-ENOENT);
5689 * css_from_id - lookup css by id
5690 * @id: the cgroup id
5691 * @ss: cgroup subsys to be looked into
5693 * Returns the css if there's valid one with @id, otherwise returns NULL.
5694 * Should be called under rcu_read_lock().
5696 struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5698 WARN_ON_ONCE(!rcu_read_lock_held());
5699 return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5702 #ifdef CONFIG_CGROUP_DEBUG
5703 static struct cgroup_subsys_state *
5704 debug_css_alloc(struct cgroup_subsys_state *parent_css)
5706 struct cgroup_subsys_state *css = kzalloc(sizeof(*css), GFP_KERNEL);
5709 return ERR_PTR(-ENOMEM);
5714 static void debug_css_free(struct cgroup_subsys_state *css)
5719 static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5722 return cgroup_task_count(css->cgroup);
5725 static u64 current_css_set_read(struct cgroup_subsys_state *css,
5728 return (u64)(unsigned long)current->cgroups;
5731 static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5737 count = atomic_read(&task_css_set(current)->refcount);
5742 static int current_css_set_cg_links_read(struct seq_file *seq, void *v)
5744 struct cgrp_cset_link *link;
5745 struct css_set *cset;
5748 name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5752 down_read(&css_set_rwsem);
5754 cset = rcu_dereference(current->cgroups);
5755 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5756 struct cgroup *c = link->cgrp;
5758 cgroup_name(c, name_buf, NAME_MAX + 1);
5759 seq_printf(seq, "Root %d group %s\n",
5760 c->root->hierarchy_id, name_buf);
5763 up_read(&css_set_rwsem);
5768 #define MAX_TASKS_SHOWN_PER_CSS 25
5769 static int cgroup_css_links_read(struct seq_file *seq, void *v)
5771 struct cgroup_subsys_state *css = seq_css(seq);
5772 struct cgrp_cset_link *link;
5774 down_read(&css_set_rwsem);
5775 list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5776 struct css_set *cset = link->cset;
5777 struct task_struct *task;
5780 seq_printf(seq, "css_set %p\n", cset);
5782 list_for_each_entry(task, &cset->tasks, cg_list) {
5783 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5785 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5788 list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5789 if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5791 seq_printf(seq, " task %d\n", task_pid_vnr(task));
5795 seq_puts(seq, " ...\n");
5797 up_read(&css_set_rwsem);
5801 static u64 releasable_read(struct cgroup_subsys_state *css, struct cftype *cft)
5803 return (!cgroup_is_populated(css->cgroup) &&
5804 !css_has_online_children(&css->cgroup->self));
5807 static struct cftype debug_files[] = {
5809 .name = "taskcount",
5810 .read_u64 = debug_taskcount_read,
5814 .name = "current_css_set",
5815 .read_u64 = current_css_set_read,
5819 .name = "current_css_set_refcount",
5820 .read_u64 = current_css_set_refcount_read,
5824 .name = "current_css_set_cg_links",
5825 .seq_show = current_css_set_cg_links_read,
5829 .name = "cgroup_css_links",
5830 .seq_show = cgroup_css_links_read,
5834 .name = "releasable",
5835 .read_u64 = releasable_read,
5841 struct cgroup_subsys debug_cgrp_subsys = {
5842 .css_alloc = debug_css_alloc,
5843 .css_free = debug_css_free,
5844 .legacy_cftypes = debug_files,
5846 #endif /* CONFIG_CGROUP_DEBUG */