linux - kernel/cgroup.c

5620 lines · 150.9 KiB

1	/*
2	* Generic process-grouping system.
3	*
4	* Based originally on the cpuset system, extracted by Paul Menage
5	* Copyright (C) 2006 Google, Inc
6	*
7	* Notifications support
8	* Copyright (C) 2009 Nokia Corporation
9	* Author: Kirill A. Shutemov
10	*
11	* Copyright notices from the original cpuset code:
12	* --------------------------------------------------
13	* Copyright (C) 2003 BULL SA.
14	* Copyright (C) 2004-2006 Silicon Graphics, Inc.
15	*
16	* Portions derived from Patrick Mochel's sysfs code.
17	* sysfs is Copyright (c) 2001-3 Patrick Mochel
18	*
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	* ---------------------------------------------------
23	*
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.
27	*/
28
29	#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
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>
39	#include <linux/mm.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>
61
62	#include <linux/atomic.h>
63
64	/*
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.
69	*/
70	#define CGROUP_PIDLIST_DESTROY_DELAY HZ
71
72	#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \
73	MAX_CFTYPE_NAME + 2)
74
75	/*
76	* cgroup_mutex is the master lock. Any modification to cgroup or its
77	* hierarchy must be performed while holding it.
78	*
79	* css_set_rwsem protects task->cgroups pointer, the list of css_set
80	* objects, and the chain of tasks off each css_set.
81	*
82	* These locks are exported if CONFIG_PROVE_RCU so that accessors in
83	* cgroup.h can use them for lockdep annotations.
84	*/
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);
90	#else
91	static DEFINE_MUTEX(cgroup_mutex);
92	static DECLARE_RWSEM(css_set_rwsem);
93	#endif
94
95	/*
96	* Protects cgroup_idr and css_idr so that IDs can be released without
97	* grabbing cgroup_mutex.
98	*/
99	static DEFINE_SPINLOCK(cgroup_idr_lock);
100
101	/*
102	* Protects cgroup_subsys->release_agent_path. Modifying it also requires
103	* cgroup_mutex. Reading requires either cgroup_mutex or this spinlock.
104	*/
105	static DEFINE_SPINLOCK(release_agent_path_lock);
106
107	struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
108
109	#define cgroup_assert_mutex_or_rcu_locked() \
110	rcu_lockdep_assert(rcu_read_lock_held() \|\| \
111	lockdep_is_held(&cgroup_mutex), \
112	"cgroup_mutex or RCU read lock required");
113
114	/*
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.
119	*/
120	static struct workqueue_struct *cgroup_destroy_wq;
121
122	/*
123	* pidlist destructions need to be flushed on cgroup destruction. Use a
124	* separate workqueue as flush domain.
125	*/
126	static struct workqueue_struct *cgroup_pidlist_destroy_wq;
127
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>
132	};
133	#undef SUBSYS
134
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>
139	};
140	#undef SUBSYS
141
142	/*
143	* The default hierarchy, reserved for the subsystems that are otherwise
144	* unattached - it never has more than a single cgroup, and all tasks are
145	* part of that cgroup.
146	*/
147	struct cgroup_root cgrp_dfl_root;
148
149	/*
150	* The default hierarchy always exists but is hidden until mounted for the
151	* first time. This is for backward compatibility.
152	*/
153	static bool cgrp_dfl_root_visible;
154
155	/*
156	* Set by the boot param of the same name and makes subsystems with NULL
157	* ->dfl_files to use ->legacy_files on the default hierarchy.
158	*/
159	static bool cgroup_legacy_files_on_dfl;
160
161	/* some controllers are not supported in the default hierarchy */
162	static unsigned long cgrp_dfl_root_inhibit_ss_mask;
163
164	/* The list of hierarchy roots */
165
166	static LIST_HEAD(cgroup_roots);
167	static int cgroup_root_count;
168
169	/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
170	static DEFINE_IDR(cgroup_hierarchy_idr);
171
172	/*
173	* Assign a monotonically increasing serial number to csses. It guarantees
174	* cgroups with bigger numbers are newer than those with smaller numbers.
175	* Also, as csses are always appended to the parent's ->children list, it
176	* guarantees that sibling csses are always sorted in the ascending serial
177	* number order on the list. Protected by cgroup_mutex.
178	*/
179	static u64 css_serial_nr_next = 1;
180
181	/*
182	* These bitmask flags indicate whether tasks in the fork and exit paths have
183	* fork/exit handlers to call. This avoids us having to do extra work in the
184	* fork/exit path to check which subsystems have fork/exit callbacks.
185	*/
186	static unsigned long have_fork_callback __read_mostly;
187	static unsigned long have_exit_callback __read_mostly;
188
189	static struct cftype cgroup_dfl_base_files[];
190	static struct cftype cgroup_legacy_base_files[];
191
192	static int rebind_subsystems(struct cgroup_root *dst_root,
193	unsigned long ss_mask);
194	static int cgroup_destroy_locked(struct cgroup *cgrp);
195	static int create_css(struct cgroup cgrp, struct cgroup_subsys ss,
196	bool visible);
197	static void css_release(struct percpu_ref *ref);
198	static void kill_css(struct cgroup_subsys_state *css);
199	static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
200	bool is_add);
201
202	/* IDR wrappers which synchronize using cgroup_idr_lock */
203	static int cgroup_idr_alloc(struct idr idr, void ptr, int start, int end,
204	gfp_t gfp_mask)
205	{
206	int ret;
207
208	idr_preload(gfp_mask);
209	spin_lock_bh(&cgroup_idr_lock);
210	ret = idr_alloc(idr, ptr, start, end, gfp_mask);
211	spin_unlock_bh(&cgroup_idr_lock);
212	idr_preload_end();
213	return ret;
214	}
215
216	static void cgroup_idr_replace(struct idr idr, void *ptr, int id)
217	{
218	void *ret;
219
220	spin_lock_bh(&cgroup_idr_lock);
221	ret = idr_replace(idr, ptr, id);
222	spin_unlock_bh(&cgroup_idr_lock);
223	return ret;
224	}
225
226	static void cgroup_idr_remove(struct idr *idr, int id)
227	{
228	spin_lock_bh(&cgroup_idr_lock);
229	idr_remove(idr, id);
230	spin_unlock_bh(&cgroup_idr_lock);
231	}
232
233	static struct cgroup cgroup_parent(struct cgroup cgrp)
234	{
235	struct cgroup_subsys_state *parent_css = cgrp->self.parent;
236
237	if (parent_css)
238	return container_of(parent_css, struct cgroup, self);
239	return NULL;
240	}
241
242	/**
243	* cgroup_css - obtain a cgroup's css for the specified subsystem
244	* @cgrp: the cgroup of interest
245	* @ss: the subsystem of interest (%NULL returns @cgrp->self)
246	*
247	* Return @cgrp's css (cgroup_subsys_state) associated with @ss. This
248	* function must be called either under cgroup_mutex or rcu_read_lock() and
249	* the caller is responsible for pinning the returned css if it wants to
250	* keep accessing it outside the said locks. This function may return
251	* %NULL if @cgrp doesn't have @subsys_id enabled.
252	*/
253	static struct cgroup_subsys_state cgroup_css(struct cgroup cgrp,
254	struct cgroup_subsys *ss)
255	{
256	if (ss)
257	return rcu_dereference_check(cgrp->subsys[ss->id],
258	lockdep_is_held(&cgroup_mutex));
259	else
260	return &cgrp->self;
261	}
262
263	/**
264	* cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
265	* @cgrp: the cgroup of interest
266	* @ss: the subsystem of interest (%NULL returns @cgrp->self)
267	*
268	* Similar to cgroup_css() but returns the effective css, which is defined
269	* as the matching css of the nearest ancestor including self which has @ss
270	* enabled. If @ss is associated with the hierarchy @cgrp is on, this
271	* function is guaranteed to return non-NULL css.
272	*/
273	static struct cgroup_subsys_state cgroup_e_css(struct cgroup cgrp,
274	struct cgroup_subsys *ss)
275	{
276	lockdep_assert_held(&cgroup_mutex);
277
278	if (!ss)
279	return &cgrp->self;
280
281	if (!(cgrp->root->subsys_mask & (1 << ss->id)))
282	return NULL;
283
284	/*
285	* This function is used while updating css associations and thus
286	* can't test the csses directly. Use ->child_subsys_mask.
287	*/
288	while (cgroup_parent(cgrp) &&
289	!(cgroup_parent(cgrp)->child_subsys_mask & (1 << ss->id)))
290	cgrp = cgroup_parent(cgrp);
291
292	return cgroup_css(cgrp, ss);
293	}
294
295	/**
296	* cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
297	* @cgrp: the cgroup of interest
298	* @ss: the subsystem of interest
299	*
300	* Find and get the effective css of @cgrp for @ss. The effective css is
301	* defined as the matching css of the nearest ancestor including self which
302	* has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on,
303	* the root css is returned, so this function always returns a valid css.
304	* The returned css must be put using css_put().
305	*/
306	struct cgroup_subsys_state cgroup_get_e_css(struct cgroup cgrp,
307	struct cgroup_subsys *ss)
308	{
309	struct cgroup_subsys_state *css;
310
311	rcu_read_lock();
312
313	do {
314	css = cgroup_css(cgrp, ss);
315
316	if (css && css_tryget_online(css))
317	goto out_unlock;
318	cgrp = cgroup_parent(cgrp);
319	} while (cgrp);
320
321	css = init_css_set.subsys[ss->id];
322	css_get(css);
323	out_unlock:
324	rcu_read_unlock();
325	return css;
326	}
327
328	/* convenient tests for these bits */
329	static inline bool cgroup_is_dead(const struct cgroup *cgrp)
330	{
331	return !(cgrp->self.flags & CSS_ONLINE);
332	}
333
334	struct cgroup_subsys_state of_css(struct kernfs_open_file of)
335	{
336	struct cgroup *cgrp = of->kn->parent->priv;
337	struct cftype *cft = of_cft(of);
338
339	/*
340	* This is open and unprotected implementation of cgroup_css().
341	* seq_css() is only called from a kernfs file operation which has
342	* an active reference on the file. Because all the subsystem
343	* files are drained before a css is disassociated with a cgroup,
344	* the matching css from the cgroup's subsys table is guaranteed to
345	* be and stay valid until the enclosing operation is complete.
346	*/
347	if (cft->ss)
348	return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
349	else
350	return &cgrp->self;
351	}
352	EXPORT_SYMBOL_GPL(of_css);
353
354	/**
355	* cgroup_is_descendant - test ancestry
356	* @cgrp: the cgroup to be tested
357	* @ancestor: possible ancestor of @cgrp
358	*
359	* Test whether @cgrp is a descendant of @ancestor. It also returns %true
360	* if @cgrp == @ancestor. This function is safe to call as long as @cgrp
361	* and @ancestor are accessible.
362	*/
363	bool cgroup_is_descendant(struct cgroup cgrp, struct cgroup ancestor)
364	{
365	while (cgrp) {
366	if (cgrp == ancestor)
367	return true;
368	cgrp = cgroup_parent(cgrp);
369	}
370	return false;
371	}
372
373	static int notify_on_release(const struct cgroup *cgrp)
374	{
375	return test_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
376	}
377
378	/**
379	* for_each_css - iterate all css's of a cgroup
380	* @css: the iteration cursor
381	* @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
382	* @cgrp: the target cgroup to iterate css's of
383	*
384	* Should be called under cgroup_[tree_]mutex.
385	*/
386	#define for_each_css(css, ssid, cgrp) \
387	for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
388	if (!((css) = rcu_dereference_check( \
389	(cgrp)->subsys[(ssid)], \
390	lockdep_is_held(&cgroup_mutex)))) { } \
391	else
392
393	/**
394	* for_each_e_css - iterate all effective css's of a cgroup
395	* @css: the iteration cursor
396	* @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
397	* @cgrp: the target cgroup to iterate css's of
398	*
399	* Should be called under cgroup_[tree_]mutex.
400	*/
401	#define for_each_e_css(css, ssid, cgrp) \
402	for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \
403	if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
404	; \
405	else
406
407	/**
408	* for_each_subsys - iterate all enabled cgroup subsystems
409	* @ss: the iteration cursor
410	* @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
411	*/
412	#define for_each_subsys(ss, ssid) \
413	for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT && \
414	(((ss) = cgroup_subsys[ssid]) \|\| true); (ssid)++)
415
416	/**
417	* for_each_subsys_which - filter for_each_subsys with a bitmask
418	* @ss: the iteration cursor
419	* @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
420	* @ss_maskp: a pointer to the bitmask
421	*
422	* The block will only run for cases where the ssid-th bit (1 << ssid) of
423	* mask is set to 1.
424	*/
425	#define for_each_subsys_which(ss, ssid, ss_maskp) \
426	if (!CGROUP_SUBSYS_COUNT) /* to avoid spurious gcc warning */ \
427	(ssid) = 0; \
428	else \
429	for_each_set_bit(ssid, ss_maskp, CGROUP_SUBSYS_COUNT) \
430	if (((ss) = cgroup_subsys[ssid]) && false) \
431	break; \
432	else
433
434	/* iterate across the hierarchies */
435	#define for_each_root(root) \
436	list_for_each_entry((root), &cgroup_roots, root_list)
437
438	/* iterate over child cgrps, lock should be held throughout iteration */
439	#define cgroup_for_each_live_child(child, cgrp) \
440	list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
441	if (({ lockdep_assert_held(&cgroup_mutex); \
442	cgroup_is_dead(child); })) \
443	; \
444	else
445
446	static void cgroup_release_agent(struct work_struct *work);
447	static void check_for_release(struct cgroup *cgrp);
448
449	/*
450	* A cgroup can be associated with multiple css_sets as different tasks may
451	* belong to different cgroups on different hierarchies. In the other
452	* direction, a css_set is naturally associated with multiple cgroups.
453	* This M:N relationship is represented by the following link structure
454	* which exists for each association and allows traversing the associations
455	* from both sides.
456	*/
457	struct cgrp_cset_link {
458	/* the cgroup and css_set this link associates */
459	struct cgroup *cgrp;
460	struct css_set *cset;
461
462	/* list of cgrp_cset_links anchored at cgrp->cset_links */
463	struct list_head cset_link;
464
465	/* list of cgrp_cset_links anchored at css_set->cgrp_links */
466	struct list_head cgrp_link;
467	};
468
469	/*
470	* The default css_set - used by init and its children prior to any
471	* hierarchies being mounted. It contains a pointer to the root state
472	* for each subsystem. Also used to anchor the list of css_sets. Not
473	* reference-counted, to improve performance when child cgroups
474	* haven't been created.
475	*/
476	struct css_set init_css_set = {
477	.refcount = ATOMIC_INIT(1),
478	.cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links),
479	.tasks = LIST_HEAD_INIT(init_css_set.tasks),
480	.mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks),
481	.mg_preload_node = LIST_HEAD_INIT(init_css_set.mg_preload_node),
482	.mg_node = LIST_HEAD_INIT(init_css_set.mg_node),
483	};
484
485	static int css_set_count = 1; /* 1 for init_css_set */
486
487	/**
488	* cgroup_update_populated - updated populated count of a cgroup
489	* @cgrp: the target cgroup
490	* @populated: inc or dec populated count
491	*
492	* @cgrp is either getting the first task (css_set) or losing the last.
493	* Update @cgrp->populated_cnt accordingly. The count is propagated
494	* towards root so that a given cgroup's populated_cnt is zero iff the
495	* cgroup and all its descendants are empty.
496	*
497	* @cgrp's interface file "cgroup.populated" is zero if
498	* @cgrp->populated_cnt is zero and 1 otherwise. When @cgrp->populated_cnt
499	* changes from or to zero, userland is notified that the content of the
500	* interface file has changed. This can be used to detect when @cgrp and
501	* its descendants become populated or empty.
502	*/
503	static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
504	{
505	lockdep_assert_held(&css_set_rwsem);
506
507	do {
508	bool trigger;
509
510	if (populated)
511	trigger = !cgrp->populated_cnt++;
512	else
513	trigger = !--cgrp->populated_cnt;
514
515	if (!trigger)
516	break;
517
518	if (cgrp->populated_kn)
519	kernfs_notify(cgrp->populated_kn);
520	cgrp = cgroup_parent(cgrp);
521	} while (cgrp);
522	}
523
524	/*
525	* hash table for cgroup groups. This improves the performance to find
526	* an existing css_set. This hash doesn't (currently) take into
527	* account cgroups in empty hierarchies.
528	*/
529	#define CSS_SET_HASH_BITS 7
530	static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
531
532	static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
533	{
534	unsigned long key = 0UL;
535	struct cgroup_subsys *ss;
536	int i;
537
538	for_each_subsys(ss, i)
539	key += (unsigned long)css[i];
540	key = (key >> 16) ^ key;
541
542	return key;
543	}
544
545	static void put_css_set_locked(struct css_set *cset)
546	{
547	struct cgrp_cset_link link, tmp_link;
548	struct cgroup_subsys *ss;
549	int ssid;
550
551	lockdep_assert_held(&css_set_rwsem);
552
553	if (!atomic_dec_and_test(&cset->refcount))
554	return;
555
556	/* This css_set is dead. unlink it and release cgroup refcounts */
557	for_each_subsys(ss, ssid)
558	list_del(&cset->e_cset_node[ssid]);
559	hash_del(&cset->hlist);
560	css_set_count--;
561
562	list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
563	struct cgroup *cgrp = link->cgrp;
564
565	list_del(&link->cset_link);
566	list_del(&link->cgrp_link);
567
568	/* @cgrp can't go away while we're holding css_set_rwsem */
569	if (list_empty(&cgrp->cset_links)) {
570	cgroup_update_populated(cgrp, false);
571	check_for_release(cgrp);
572	}
573
574	kfree(link);
575	}
576
577	kfree_rcu(cset, rcu_head);
578	}
579
580	static void put_css_set(struct css_set *cset)
581	{
582	/*
583	* Ensure that the refcount doesn't hit zero while any readers
584	* can see it. Similar to atomic_dec_and_lock(), but for an
585	* rwlock
586	*/
587	if (atomic_add_unless(&cset->refcount, -1, 1))
588	return;
589
590	down_write(&css_set_rwsem);
591	put_css_set_locked(cset);
592	up_write(&css_set_rwsem);
593	}
594
595	/*
596	* refcounted get/put for css_set objects
597	*/
598	static inline void get_css_set(struct css_set *cset)
599	{
600	atomic_inc(&cset->refcount);
601	}
602
603	/**
604	* compare_css_sets - helper function for find_existing_css_set().
605	* @cset: candidate css_set being tested
606	* @old_cset: existing css_set for a task
607	* @new_cgrp: cgroup that's being entered by the task
608	* @template: desired set of css pointers in css_set (pre-calculated)
609	*
610	* Returns true if "cset" matches "old_cset" except for the hierarchy
611	* which "new_cgrp" belongs to, for which it should match "new_cgrp".
612	*/
613	static bool compare_css_sets(struct css_set *cset,
614	struct css_set *old_cset,
615	struct cgroup *new_cgrp,
616	struct cgroup_subsys_state *template[])
617	{
618	struct list_head l1, l2;
619
620	/*
621	* On the default hierarchy, there can be csets which are
622	* associated with the same set of cgroups but different csses.
623	* Let's first ensure that csses match.
624	*/
625	if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
626	return false;
627
628	/*
629	* Compare cgroup pointers in order to distinguish between
630	* different cgroups in hierarchies. As different cgroups may
631	* share the same effective css, this comparison is always
632	* necessary.
633	*/
634	l1 = &cset->cgrp_links;
635	l2 = &old_cset->cgrp_links;
636	while (1) {
637	struct cgrp_cset_link link1, link2;
638	struct cgroup cgrp1, cgrp2;
639
640	l1 = l1->next;
641	l2 = l2->next;
642	/* See if we reached the end - both lists are equal length. */
643	if (l1 == &cset->cgrp_links) {
644	BUG_ON(l2 != &old_cset->cgrp_links);
645	break;
646	} else {
647	BUG_ON(l2 == &old_cset->cgrp_links);
648	}
649	/* Locate the cgroups associated with these links. */
650	link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
651	link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
652	cgrp1 = link1->cgrp;
653	cgrp2 = link2->cgrp;
654	/* Hierarchies should be linked in the same order. */
655	BUG_ON(cgrp1->root != cgrp2->root);
656
657	/*
658	* If this hierarchy is the hierarchy of the cgroup
659	* that's changing, then we need to check that this
660	* css_set points to the new cgroup; if it's any other
661	* hierarchy, then this css_set should point to the
662	* same cgroup as the old css_set.
663	*/
664	if (cgrp1->root == new_cgrp->root) {
665	if (cgrp1 != new_cgrp)
666	return false;
667	} else {
668	if (cgrp1 != cgrp2)
669	return false;
670	}
671	}
672	return true;
673	}
674
675	/**
676	* find_existing_css_set - init css array and find the matching css_set
677	* @old_cset: the css_set that we're using before the cgroup transition
678	* @cgrp: the cgroup that we're moving into
679	* @template: out param for the new set of csses, should be clear on entry
680	*/
681	static struct css_set find_existing_css_set(struct css_set old_cset,
682	struct cgroup *cgrp,
683	struct cgroup_subsys_state *template[])
684	{
685	struct cgroup_root *root = cgrp->root;
686	struct cgroup_subsys *ss;
687	struct css_set *cset;
688	unsigned long key;
689	int i;
690
691	/*
692	* Build the set of subsystem state objects that we want to see in the
693	* new css_set. while subsystems can change globally, the entries here
694	* won't change, so no need for locking.
695	*/
696	for_each_subsys(ss, i) {
697	if (root->subsys_mask & (1UL << i)) {
698	/*
699	* @ss is in this hierarchy, so we want the
700	* effective css from @cgrp.
701	*/
702	template[i] = cgroup_e_css(cgrp, ss);
703	} else {
704	/*
705	* @ss is not in this hierarchy, so we don't want
706	* to change the css.
707	*/
708	template[i] = old_cset->subsys[i];
709	}
710	}
711
712	key = css_set_hash(template);
713	hash_for_each_possible(css_set_table, cset, hlist, key) {
714	if (!compare_css_sets(cset, old_cset, cgrp, template))
715	continue;
716
717	/* This css_set matches what we need */
718	return cset;
719	}
720
721	/* No existing cgroup group matched */
722	return NULL;
723	}
724
725	static void free_cgrp_cset_links(struct list_head *links_to_free)
726	{
727	struct cgrp_cset_link link, tmp_link;
728
729	list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
730	list_del(&link->cset_link);
731	kfree(link);
732	}
733	}
734
735	/**
736	* allocate_cgrp_cset_links - allocate cgrp_cset_links
737	* @count: the number of links to allocate
738	* @tmp_links: list_head the allocated links are put on
739	*
740	* Allocate @count cgrp_cset_link structures and chain them on @tmp_links
741	* through ->cset_link. Returns 0 on success or -errno.
742	*/
743	static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
744	{
745	struct cgrp_cset_link *link;
746	int i;
747
748	INIT_LIST_HEAD(tmp_links);
749
750	for (i = 0; i < count; i++) {
751	link = kzalloc(sizeof(*link), GFP_KERNEL);
752	if (!link) {
753	free_cgrp_cset_links(tmp_links);
754	return -ENOMEM;
755	}
756	list_add(&link->cset_link, tmp_links);
757	}
758	return 0;
759	}
760
761	/**
762	* link_css_set - a helper function to link a css_set to a cgroup
763	* @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
764	* @cset: the css_set to be linked
765	* @cgrp: the destination cgroup
766	*/
767	static void link_css_set(struct list_head tmp_links, struct css_set cset,
768	struct cgroup *cgrp)
769	{
770	struct cgrp_cset_link *link;
771
772	BUG_ON(list_empty(tmp_links));
773
774	if (cgroup_on_dfl(cgrp))
775	cset->dfl_cgrp = cgrp;
776
777	link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
778	link->cset = cset;
779	link->cgrp = cgrp;
780
781	if (list_empty(&cgrp->cset_links))
782	cgroup_update_populated(cgrp, true);
783	list_move(&link->cset_link, &cgrp->cset_links);
784
785	/*
786	* Always add links to the tail of the list so that the list
787	* is sorted by order of hierarchy creation
788	*/
789	list_add_tail(&link->cgrp_link, &cset->cgrp_links);
790	}
791
792	/**
793	* find_css_set - return a new css_set with one cgroup updated
794	* @old_cset: the baseline css_set
795	* @cgrp: the cgroup to be updated
796	*
797	* Return a new css_set that's equivalent to @old_cset, but with @cgrp
798	* substituted into the appropriate hierarchy.
799	*/
800	static struct css_set find_css_set(struct css_set old_cset,
801	struct cgroup *cgrp)
802	{
803	struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
804	struct css_set *cset;
805	struct list_head tmp_links;
806	struct cgrp_cset_link *link;
807	struct cgroup_subsys *ss;
808	unsigned long key;
809	int ssid;
810
811	lockdep_assert_held(&cgroup_mutex);
812
813	/* First see if we already have a cgroup group that matches
814	* the desired set */
815	down_read(&css_set_rwsem);
816	cset = find_existing_css_set(old_cset, cgrp, template);
817	if (cset)
818	get_css_set(cset);
819	up_read(&css_set_rwsem);
820
821	if (cset)
822	return cset;
823
824	cset = kzalloc(sizeof(*cset), GFP_KERNEL);
825	if (!cset)
826	return NULL;
827
828	/* Allocate all the cgrp_cset_link objects that we'll need */
829	if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
830	kfree(cset);
831	return NULL;
832	}
833
834	atomic_set(&cset->refcount, 1);
835	INIT_LIST_HEAD(&cset->cgrp_links);
836	INIT_LIST_HEAD(&cset->tasks);
837	INIT_LIST_HEAD(&cset->mg_tasks);
838	INIT_LIST_HEAD(&cset->mg_preload_node);
839	INIT_LIST_HEAD(&cset->mg_node);
840	INIT_HLIST_NODE(&cset->hlist);
841
842	/* Copy the set of subsystem state objects generated in
843	* find_existing_css_set() */
844	memcpy(cset->subsys, template, sizeof(cset->subsys));
845
846	down_write(&css_set_rwsem);
847	/* Add reference counts and links from the new css_set. */
848	list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
849	struct cgroup *c = link->cgrp;
850
851	if (c->root == cgrp->root)
852	c = cgrp;
853	link_css_set(&tmp_links, cset, c);
854	}
855
856	BUG_ON(!list_empty(&tmp_links));
857
858	css_set_count++;
859
860	/* Add @cset to the hash table */
861	key = css_set_hash(cset->subsys);
862	hash_add(css_set_table, &cset->hlist, key);
863
864	for_each_subsys(ss, ssid)
865	list_add_tail(&cset->e_cset_node[ssid],
866	&cset->subsys[ssid]->cgroup->e_csets[ssid]);
867
868	up_write(&css_set_rwsem);
869
870	return cset;
871	}
872
873	static struct cgroup_root cgroup_root_from_kf(struct kernfs_root kf_root)
874	{
875	struct cgroup *root_cgrp = kf_root->kn->priv;
876
877	return root_cgrp->root;
878	}
879
880	static int cgroup_init_root_id(struct cgroup_root *root)
881	{
882	int id;
883
884	lockdep_assert_held(&cgroup_mutex);
885
886	id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
887	if (id < 0)
888	return id;
889
890	root->hierarchy_id = id;
891	return 0;
892	}
893
894	static void cgroup_exit_root_id(struct cgroup_root *root)
895	{
896	lockdep_assert_held(&cgroup_mutex);
897
898	if (root->hierarchy_id) {
899	idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
900	root->hierarchy_id = 0;
901	}
902	}
903
904	static void cgroup_free_root(struct cgroup_root *root)
905	{
906	if (root) {
907	/* hierarchy ID should already have been released */
908	WARN_ON_ONCE(root->hierarchy_id);
909
910	idr_destroy(&root->cgroup_idr);
911	kfree(root);
912	}
913	}
914
915	static void cgroup_destroy_root(struct cgroup_root *root)
916	{
917	struct cgroup *cgrp = &root->cgrp;
918	struct cgrp_cset_link link, tmp_link;
919
920	mutex_lock(&cgroup_mutex);
921
922	BUG_ON(atomic_read(&root->nr_cgrps));
923	BUG_ON(!list_empty(&cgrp->self.children));
924
925	/* Rebind all subsystems back to the default hierarchy */
926	rebind_subsystems(&cgrp_dfl_root, root->subsys_mask);
927
928	/*
929	* Release all the links from cset_links to this hierarchy's
930	* root cgroup
931	*/
932	down_write(&css_set_rwsem);
933
934	list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
935	list_del(&link->cset_link);
936	list_del(&link->cgrp_link);
937	kfree(link);
938	}
939	up_write(&css_set_rwsem);
940
941	if (!list_empty(&root->root_list)) {
942	list_del(&root->root_list);
943	cgroup_root_count--;
944	}
945
946	cgroup_exit_root_id(root);
947
948	mutex_unlock(&cgroup_mutex);
949
950	kernfs_destroy_root(root->kf_root);
951	cgroup_free_root(root);
952	}
953
954	/* look up cgroup associated with given css_set on the specified hierarchy */
955	static struct cgroup cset_cgroup_from_root(struct css_set cset,
956	struct cgroup_root *root)
957	{
958	struct cgroup *res = NULL;
959
960	lockdep_assert_held(&cgroup_mutex);
961	lockdep_assert_held(&css_set_rwsem);
962
963	if (cset == &init_css_set) {
964	res = &root->cgrp;
965	} else {
966	struct cgrp_cset_link *link;
967
968	list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
969	struct cgroup *c = link->cgrp;
970
971	if (c->root == root) {
972	res = c;
973	break;
974	}
975	}
976	}
977
978	BUG_ON(!res);
979	return res;
980	}
981
982	/*
983	* Return the cgroup for "task" from the given hierarchy. Must be
984	* called with cgroup_mutex and css_set_rwsem held.
985	*/
986	static struct cgroup task_cgroup_from_root(struct task_struct task,
987	struct cgroup_root *root)
988	{
989	/*
990	* No need to lock the task - since we hold cgroup_mutex the
991	* task can't change groups, so the only thing that can happen
992	* is that it exits and its css is set back to init_css_set.
993	*/
994	return cset_cgroup_from_root(task_css_set(task), root);
995	}
996
997	/*
998	* A task must hold cgroup_mutex to modify cgroups.
999	*
1000	* Any task can increment and decrement the count field without lock.
1001	* So in general, code holding cgroup_mutex can't rely on the count
1002	* field not changing. However, if the count goes to zero, then only
1003	* cgroup_attach_task() can increment it again. Because a count of zero
1004	* means that no tasks are currently attached, therefore there is no
1005	* way a task attached to that cgroup can fork (the other way to
1006	* increment the count). So code holding cgroup_mutex can safely
1007	* assume that if the count is zero, it will stay zero. Similarly, if
1008	* a task holds cgroup_mutex on a cgroup with zero count, it
1009	* knows that the cgroup won't be removed, as cgroup_rmdir()
1010	* needs that mutex.
1011	*
1012	* A cgroup can only be deleted if both its 'count' of using tasks
1013	* is zero, and its list of 'children' cgroups is empty. Since all
1014	* tasks in the system use _some_ cgroup, and since there is always at
1015	* least one task in the system (init, pid == 1), therefore, root cgroup
1016	* always has either children cgroups and/or using tasks. So we don't
1017	* need a special hack to ensure that root cgroup cannot be deleted.
1018	*
1019	* P.S. One more locking exception. RCU is used to guard the
1020	* update of a tasks cgroup pointer by cgroup_attach_task()
1021	*/
1022
1023	static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask);
1024	static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1025	static const struct file_operations proc_cgroupstats_operations;
1026
1027	static char cgroup_file_name(struct cgroup cgrp, const struct cftype *cft,
1028	char *buf)
1029	{
1030	if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1031	!(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1032	snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1033	cft->ss->name, cft->name);
1034	else
1035	strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1036	return buf;
1037	}
1038
1039	/**
1040	* cgroup_file_mode - deduce file mode of a control file
1041	* @cft: the control file in question
1042	*
1043	* returns cft->mode if ->mode is not 0
1044	* returns S_IRUGO\|S_IWUSR if it has both a read and a write handler
1045	* returns S_IRUGO if it has only a read handler
1046	* returns S_IWUSR if it has only a write hander
1047	*/
1048	static umode_t cgroup_file_mode(const struct cftype *cft)
1049	{
1050	umode_t mode = 0;
1051
1052	if (cft->mode)
1053	return cft->mode;
1054
1055	if (cft->read_u64 \|\| cft->read_s64 \|\| cft->seq_show)
1056	mode \|= S_IRUGO;
1057
1058	if (cft->write_u64 \|\| cft->write_s64 \|\| cft->write)
1059	mode \|= S_IWUSR;
1060
1061	return mode;
1062	}
1063
1064	static void cgroup_get(struct cgroup *cgrp)
1065	{
1066	WARN_ON_ONCE(cgroup_is_dead(cgrp));
1067	css_get(&cgrp->self);
1068	}
1069
1070	static bool cgroup_tryget(struct cgroup *cgrp)
1071	{
1072	return css_tryget(&cgrp->self);
1073	}
1074
1075	static void cgroup_put(struct cgroup *cgrp)
1076	{
1077	css_put(&cgrp->self);
1078	}
1079
1080	/**
1081	* cgroup_calc_child_subsys_mask - calculate child_subsys_mask
1082	* @cgrp: the target cgroup
1083	* @subtree_control: the new subtree_control mask to consider
1084	*
1085	* On the default hierarchy, a subsystem may request other subsystems to be
1086	* enabled together through its ->depends_on mask. In such cases, more
1087	* subsystems than specified in "cgroup.subtree_control" may be enabled.
1088	*
1089	* This function calculates which subsystems need to be enabled if
1090	* @subtree_control is to be applied to @cgrp. The returned mask is always
1091	* a superset of @subtree_control and follows the usual hierarchy rules.
1092	*/
1093	static unsigned long cgroup_calc_child_subsys_mask(struct cgroup *cgrp,
1094	unsigned long subtree_control)
1095	{
1096	struct cgroup *parent = cgroup_parent(cgrp);
1097	unsigned long cur_ss_mask = subtree_control;
1098	struct cgroup_subsys *ss;
1099	int ssid;
1100
1101	lockdep_assert_held(&cgroup_mutex);
1102
1103	if (!cgroup_on_dfl(cgrp))
1104	return cur_ss_mask;
1105
1106	while (true) {
1107	unsigned long new_ss_mask = cur_ss_mask;
1108
1109	for_each_subsys_which(ss, ssid, &cur_ss_mask)
1110	new_ss_mask \|= ss->depends_on;
1111
1112	/*
1113	* Mask out subsystems which aren't available. This can
1114	* happen only if some depended-upon subsystems were bound
1115	* to non-default hierarchies.
1116	*/
1117	if (parent)
1118	new_ss_mask &= parent->child_subsys_mask;
1119	else
1120	new_ss_mask &= cgrp->root->subsys_mask;
1121
1122	if (new_ss_mask == cur_ss_mask)
1123	break;
1124	cur_ss_mask = new_ss_mask;
1125	}
1126
1127	return cur_ss_mask;
1128	}
1129
1130	/**
1131	* cgroup_refresh_child_subsys_mask - update child_subsys_mask
1132	* @cgrp: the target cgroup
1133	*
1134	* Update @cgrp->child_subsys_mask according to the current
1135	* @cgrp->subtree_control using cgroup_calc_child_subsys_mask().
1136	*/
1137	static void cgroup_refresh_child_subsys_mask(struct cgroup *cgrp)
1138	{
1139	cgrp->child_subsys_mask =
1140	cgroup_calc_child_subsys_mask(cgrp, cgrp->subtree_control);
1141	}
1142
1143	/**
1144	* cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1145	* @kn: the kernfs_node being serviced
1146	*
1147	* This helper undoes cgroup_kn_lock_live() and should be invoked before
1148	* the method finishes if locking succeeded. Note that once this function
1149	* returns the cgroup returned by cgroup_kn_lock_live() may become
1150	* inaccessible any time. If the caller intends to continue to access the
1151	* cgroup, it should pin it before invoking this function.
1152	*/
1153	static void cgroup_kn_unlock(struct kernfs_node *kn)
1154	{
1155	struct cgroup *cgrp;
1156
1157	if (kernfs_type(kn) == KERNFS_DIR)
1158	cgrp = kn->priv;
1159	else
1160	cgrp = kn->parent->priv;
1161
1162	mutex_unlock(&cgroup_mutex);
1163
1164	kernfs_unbreak_active_protection(kn);
1165	cgroup_put(cgrp);
1166	}
1167
1168	/**
1169	* cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1170	* @kn: the kernfs_node being serviced
1171	*
1172	* This helper is to be used by a cgroup kernfs method currently servicing
1173	* @kn. It breaks the active protection, performs cgroup locking and
1174	* verifies that the associated cgroup is alive. Returns the cgroup if
1175	* alive; otherwise, %NULL. A successful return should be undone by a
1176	* matching cgroup_kn_unlock() invocation.
1177	*
1178	* Any cgroup kernfs method implementation which requires locking the
1179	* associated cgroup should use this helper. It avoids nesting cgroup
1180	* locking under kernfs active protection and allows all kernfs operations
1181	* including self-removal.
1182	*/
1183	static struct cgroup cgroup_kn_lock_live(struct kernfs_node kn)
1184	{
1185	struct cgroup *cgrp;
1186
1187	if (kernfs_type(kn) == KERNFS_DIR)
1188	cgrp = kn->priv;
1189	else
1190	cgrp = kn->parent->priv;
1191
1192	/*
1193	* We're gonna grab cgroup_mutex which nests outside kernfs
1194	* active_ref. cgroup liveliness check alone provides enough
1195	* protection against removal. Ensure @cgrp stays accessible and
1196	* break the active_ref protection.
1197	*/
1198	if (!cgroup_tryget(cgrp))
1199	return NULL;
1200	kernfs_break_active_protection(kn);
1201
1202	mutex_lock(&cgroup_mutex);
1203
1204	if (!cgroup_is_dead(cgrp))
1205	return cgrp;
1206
1207	cgroup_kn_unlock(kn);
1208	return NULL;
1209	}
1210
1211	static void cgroup_rm_file(struct cgroup cgrp, const struct cftype cft)
1212	{
1213	char name[CGROUP_FILE_NAME_MAX];
1214
1215	lockdep_assert_held(&cgroup_mutex);
1216	kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1217	}
1218
1219	/**
1220	* cgroup_clear_dir - remove subsys files in a cgroup directory
1221	* @cgrp: target cgroup
1222	* @subsys_mask: mask of the subsystem ids whose files should be removed
1223	*/
1224	static void cgroup_clear_dir(struct cgroup *cgrp, unsigned long subsys_mask)
1225	{
1226	struct cgroup_subsys *ss;
1227	int i;
1228
1229	for_each_subsys(ss, i) {
1230	struct cftype *cfts;
1231
1232	if (!(subsys_mask & (1 << i)))
1233	continue;
1234	list_for_each_entry(cfts, &ss->cfts, node)
1235	cgroup_addrm_files(cgrp, cfts, false);
1236	}
1237	}
1238
1239	static int rebind_subsystems(struct cgroup_root *dst_root,
1240	unsigned long ss_mask)
1241	{
1242	struct cgroup_subsys *ss;
1243	unsigned long tmp_ss_mask;
1244	int ssid, i, ret;
1245
1246	lockdep_assert_held(&cgroup_mutex);
1247
1248	for_each_subsys_which(ss, ssid, &ss_mask) {
1249	/* if @ss has non-root csses attached to it, can't move */
1250	if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)))
1251	return -EBUSY;
1252
1253	/* can't move between two non-dummy roots either */
1254	if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1255	return -EBUSY;
1256	}
1257
1258	/* skip creating root files on dfl_root for inhibited subsystems */
1259	tmp_ss_mask = ss_mask;
1260	if (dst_root == &cgrp_dfl_root)
1261	tmp_ss_mask &= ~cgrp_dfl_root_inhibit_ss_mask;
1262
1263	ret = cgroup_populate_dir(&dst_root->cgrp, tmp_ss_mask);
1264	if (ret) {
1265	if (dst_root != &cgrp_dfl_root)
1266	return ret;
1267
1268	/*
1269	* Rebinding back to the default root is not allowed to
1270	* fail. Using both default and non-default roots should
1271	* be rare. Moving subsystems back and forth even more so.
1272	* Just warn about it and continue.
1273	*/
1274	if (cgrp_dfl_root_visible) {
1275	pr_warn("failed to create files (%d) while rebinding 0x%lx to default root\n",
1276	ret, ss_mask);
1277	pr_warn("you may retry by moving them to a different hierarchy and unbinding\n");
1278	}
1279	}
1280
1281	/*
1282	* Nothing can fail from this point on. Remove files for the
1283	* removed subsystems and rebind each subsystem.
1284	*/
1285	for_each_subsys_which(ss, ssid, &ss_mask)
1286	cgroup_clear_dir(&ss->root->cgrp, 1 << ssid);
1287
1288	for_each_subsys_which(ss, ssid, &ss_mask) {
1289	struct cgroup_root *src_root;
1290	struct cgroup_subsys_state *css;
1291	struct css_set *cset;
1292
1293	src_root = ss->root;
1294	css = cgroup_css(&src_root->cgrp, ss);
1295
1296	WARN_ON(!css \|\| cgroup_css(&dst_root->cgrp, ss));
1297
1298	RCU_INIT_POINTER(src_root->cgrp.subsys[ssid], NULL);
1299	rcu_assign_pointer(dst_root->cgrp.subsys[ssid], css);
1300	ss->root = dst_root;
1301	css->cgroup = &dst_root->cgrp;
1302
1303	down_write(&css_set_rwsem);
1304	hash_for_each(css_set_table, i, cset, hlist)
1305	list_move_tail(&cset->e_cset_node[ss->id],
1306	&dst_root->cgrp.e_csets[ss->id]);
1307	up_write(&css_set_rwsem);
1308
1309	src_root->subsys_mask &= ~(1 << ssid);
1310	src_root->cgrp.subtree_control &= ~(1 << ssid);
1311	cgroup_refresh_child_subsys_mask(&src_root->cgrp);
1312
1313	/* default hierarchy doesn't enable controllers by default */
1314	dst_root->subsys_mask \|= 1 << ssid;
1315	if (dst_root != &cgrp_dfl_root) {
1316	dst_root->cgrp.subtree_control \|= 1 << ssid;
1317	cgroup_refresh_child_subsys_mask(&dst_root->cgrp);
1318	}
1319
1320	if (ss->bind)
1321	ss->bind(css);
1322	}
1323
1324	kernfs_activate(dst_root->cgrp.kn);
1325	return 0;
1326	}
1327
1328	static int cgroup_show_options(struct seq_file *seq,
1329	struct kernfs_root *kf_root)
1330	{
1331	struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1332	struct cgroup_subsys *ss;
1333	int ssid;
1334
1335	for_each_subsys(ss, ssid)
1336	if (root->subsys_mask & (1 << ssid))
1337	seq_printf(seq, ",%s", ss->name);
1338	if (root->flags & CGRP_ROOT_NOPREFIX)
1339	seq_puts(seq, ",noprefix");
1340	if (root->flags & CGRP_ROOT_XATTR)
1341	seq_puts(seq, ",xattr");
1342
1343	spin_lock(&release_agent_path_lock);
1344	if (strlen(root->release_agent_path))
1345	seq_printf(seq, ",release_agent=%s", root->release_agent_path);
1346	spin_unlock(&release_agent_path_lock);
1347
1348	if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags))
1349	seq_puts(seq, ",clone_children");
1350	if (strlen(root->name))
1351	seq_printf(seq, ",name=%s", root->name);
1352	return 0;
1353	}
1354
1355	struct cgroup_sb_opts {
1356	unsigned long subsys_mask;
1357	unsigned int flags;
1358	char *release_agent;
1359	bool cpuset_clone_children;
1360	char *name;
1361	/* User explicitly requested empty subsystem */
1362	bool none;
1363	};
1364
1365	static int parse_cgroupfs_options(char data, struct cgroup_sb_opts opts)
1366	{
1367	char token, o = data;
1368	bool all_ss = false, one_ss = false;
1369	unsigned long mask = -1UL;
1370	struct cgroup_subsys *ss;
1371	int nr_opts = 0;
1372	int i;
1373
1374	#ifdef CONFIG_CPUSETS
1375	mask = ~(1U << cpuset_cgrp_id);
1376	#endif
1377
1378	memset(opts, 0, sizeof(*opts));
1379
1380	while ((token = strsep(&o, ",")) != NULL) {
1381	nr_opts++;
1382
1383	if (!*token)
1384	return -EINVAL;
1385	if (!strcmp(token, "none")) {
1386	/* Explicitly have no subsystems */
1387	opts->none = true;
1388	continue;
1389	}
1390	if (!strcmp(token, "all")) {
1391	/* Mutually exclusive option 'all' + subsystem name */
1392	if (one_ss)
1393	return -EINVAL;
1394	all_ss = true;
1395	continue;
1396	}
1397	if (!strcmp(token, "__DEVEL__sane_behavior")) {
1398	opts->flags \|= CGRP_ROOT_SANE_BEHAVIOR;
1399	continue;
1400	}
1401	if (!strcmp(token, "noprefix")) {
1402	opts->flags \|= CGRP_ROOT_NOPREFIX;
1403	continue;
1404	}
1405	if (!strcmp(token, "clone_children")) {
1406	opts->cpuset_clone_children = true;
1407	continue;
1408	}
1409	if (!strcmp(token, "xattr")) {
1410	opts->flags \|= CGRP_ROOT_XATTR;
1411	continue;
1412	}
1413	if (!strncmp(token, "release_agent=", 14)) {
1414	/* Specifying two release agents is forbidden */
1415	if (opts->release_agent)
1416	return -EINVAL;
1417	opts->release_agent =
1418	kstrndup(token + 14, PATH_MAX - 1, GFP_KERNEL);
1419	if (!opts->release_agent)
1420	return -ENOMEM;
1421	continue;
1422	}
1423	if (!strncmp(token, "name=", 5)) {
1424	const char *name = token + 5;
1425	/* Can't specify an empty name */
1426	if (!strlen(name))
1427	return -EINVAL;
1428	/* Must match [\w.-]+ */
1429	for (i = 0; i < strlen(name); i++) {
1430	char c = name[i];
1431	if (isalnum(c))
1432	continue;
1433	if ((c == '.') \|\| (c == '-') \|\| (c == '_'))
1434	continue;
1435	return -EINVAL;
1436	}
1437	/* Specifying two names is forbidden */
1438	if (opts->name)
1439	return -EINVAL;
1440	opts->name = kstrndup(name,
1441	MAX_CGROUP_ROOT_NAMELEN - 1,
1442	GFP_KERNEL);
1443	if (!opts->name)
1444	return -ENOMEM;
1445
1446	continue;
1447	}
1448
1449	for_each_subsys(ss, i) {
1450	if (strcmp(token, ss->name))
1451	continue;
1452	if (ss->disabled)
1453	continue;
1454
1455	/* Mutually exclusive option 'all' + subsystem name */
1456	if (all_ss)
1457	return -EINVAL;
1458	opts->subsys_mask \|= (1 << i);
1459	one_ss = true;
1460
1461	break;
1462	}
1463	if (i == CGROUP_SUBSYS_COUNT)
1464	return -ENOENT;
1465	}
1466
1467	if (opts->flags & CGRP_ROOT_SANE_BEHAVIOR) {
1468	pr_warn("sane_behavior: this is still under development and its behaviors will change, proceed at your own risk\n");
1469	if (nr_opts != 1) {
1470	pr_err("sane_behavior: no other mount options allowed\n");
1471	return -EINVAL;
1472	}
1473	return 0;
1474	}
1475
1476	/*
1477	* If the 'all' option was specified select all the subsystems,
1478	* otherwise if 'none', 'name=' and a subsystem name options were
1479	* not specified, let's default to 'all'
1480	*/
1481	if (all_ss \|\| (!one_ss && !opts->none && !opts->name))
1482	for_each_subsys(ss, i)
1483	if (!ss->disabled)
1484	opts->subsys_mask \|= (1 << i);
1485
1486	/*
1487	* We either have to specify by name or by subsystems. (So all
1488	* empty hierarchies must have a name).
1489	*/
1490	if (!opts->subsys_mask && !opts->name)
1491	return -EINVAL;
1492
1493	/*
1494	* Option noprefix was introduced just for backward compatibility
1495	* with the old cpuset, so we allow noprefix only if mounting just
1496	* the cpuset subsystem.
1497	*/
1498	if ((opts->flags & CGRP_ROOT_NOPREFIX) && (opts->subsys_mask & mask))
1499	return -EINVAL;
1500
1501	/* Can't specify "none" and some subsystems */
1502	if (opts->subsys_mask && opts->none)
1503	return -EINVAL;
1504
1505	return 0;
1506	}
1507
1508	static int cgroup_remount(struct kernfs_root kf_root, int flags, char *data)
1509	{
1510	int ret = 0;
1511	struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1512	struct cgroup_sb_opts opts;
1513	unsigned long added_mask, removed_mask;
1514
1515	if (root == &cgrp_dfl_root) {
1516	pr_err("remount is not allowed\n");
1517	return -EINVAL;
1518	}
1519
1520	mutex_lock(&cgroup_mutex);
1521
1522	/* See what subsystems are wanted */
1523	ret = parse_cgroupfs_options(data, &opts);
1524	if (ret)
1525	goto out_unlock;
1526
1527	if (opts.subsys_mask != root->subsys_mask \|\| opts.release_agent)
1528	pr_warn("option changes via remount are deprecated (pid=%d comm=%s)\n",
1529	task_tgid_nr(current), current->comm);
1530
1531	added_mask = opts.subsys_mask & ~root->subsys_mask;
1532	removed_mask = root->subsys_mask & ~opts.subsys_mask;
1533
1534	/* Don't allow flags or name to change at remount */
1535	if ((opts.flags ^ root->flags) \|\|
1536	(opts.name && strcmp(opts.name, root->name))) {
1537	pr_err("option or name mismatch, new: 0x%x \"%s\", old: 0x%x \"%s\"\n",
1538	opts.flags, opts.name ?: "", root->flags, root->name);
1539	ret = -EINVAL;
1540	goto out_unlock;
1541	}
1542
1543	/* remounting is not allowed for populated hierarchies */
1544	if (!list_empty(&root->cgrp.self.children)) {
1545	ret = -EBUSY;
1546	goto out_unlock;
1547	}
1548
1549	ret = rebind_subsystems(root, added_mask);
1550	if (ret)
1551	goto out_unlock;
1552
1553	rebind_subsystems(&cgrp_dfl_root, removed_mask);
1554
1555	if (opts.release_agent) {
1556	spin_lock(&release_agent_path_lock);
1557	strcpy(root->release_agent_path, opts.release_agent);
1558	spin_unlock(&release_agent_path_lock);
1559	}
1560	out_unlock:
1561	kfree(opts.release_agent);
1562	kfree(opts.name);
1563	mutex_unlock(&cgroup_mutex);
1564	return ret;
1565	}
1566
1567	/*
1568	* To reduce the fork() overhead for systems that are not actually using
1569	* their cgroups capability, we don't maintain the lists running through
1570	* each css_set to its tasks until we see the list actually used - in other
1571	* words after the first mount.
1572	*/
1573	static bool use_task_css_set_links __read_mostly;
1574
1575	static void cgroup_enable_task_cg_lists(void)
1576	{
1577	struct task_struct p, g;
1578
1579	down_write(&css_set_rwsem);
1580
1581	if (use_task_css_set_links)
1582	goto out_unlock;
1583
1584	use_task_css_set_links = true;
1585
1586	/*
1587	* We need tasklist_lock because RCU is not safe against
1588	* while_each_thread(). Besides, a forking task that has passed
1589	* cgroup_post_fork() without seeing use_task_css_set_links = 1
1590	* is not guaranteed to have its child immediately visible in the
1591	* tasklist if we walk through it with RCU.
1592	*/
1593	read_lock(&tasklist_lock);
1594	do_each_thread(g, p) {
1595	WARN_ON_ONCE(!list_empty(&p->cg_list) \|\|
1596	task_css_set(p) != &init_css_set);
1597
1598	/*
1599	* We should check if the process is exiting, otherwise
1600	* it will race with cgroup_exit() in that the list
1601	* entry won't be deleted though the process has exited.
1602	* Do it while holding siglock so that we don't end up
1603	* racing against cgroup_exit().
1604	*/
1605	spin_lock_irq(&p->sighand->siglock);
1606	if (!(p->flags & PF_EXITING)) {
1607	struct css_set *cset = task_css_set(p);
1608
1609	list_add(&p->cg_list, &cset->tasks);
1610	get_css_set(cset);
1611	}
1612	spin_unlock_irq(&p->sighand->siglock);
1613	} while_each_thread(g, p);
1614	read_unlock(&tasklist_lock);
1615	out_unlock:
1616	up_write(&css_set_rwsem);
1617	}
1618
1619	static void init_cgroup_housekeeping(struct cgroup *cgrp)
1620	{
1621	struct cgroup_subsys *ss;
1622	int ssid;
1623
1624	INIT_LIST_HEAD(&cgrp->self.sibling);
1625	INIT_LIST_HEAD(&cgrp->self.children);
1626	INIT_LIST_HEAD(&cgrp->cset_links);
1627	INIT_LIST_HEAD(&cgrp->pidlists);
1628	mutex_init(&cgrp->pidlist_mutex);
1629	cgrp->self.cgroup = cgrp;
1630	cgrp->self.flags \|= CSS_ONLINE;
1631
1632	for_each_subsys(ss, ssid)
1633	INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1634
1635	init_waitqueue_head(&cgrp->offline_waitq);
1636	INIT_WORK(&cgrp->release_agent_work, cgroup_release_agent);
1637	}
1638
1639	static void init_cgroup_root(struct cgroup_root *root,
1640	struct cgroup_sb_opts *opts)
1641	{
1642	struct cgroup *cgrp = &root->cgrp;
1643
1644	INIT_LIST_HEAD(&root->root_list);
1645	atomic_set(&root->nr_cgrps, 1);
1646	cgrp->root = root;
1647	init_cgroup_housekeeping(cgrp);
1648	idr_init(&root->cgroup_idr);
1649
1650	root->flags = opts->flags;
1651	if (opts->release_agent)
1652	strcpy(root->release_agent_path, opts->release_agent);
1653	if (opts->name)
1654	strcpy(root->name, opts->name);
1655	if (opts->cpuset_clone_children)
1656	set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1657	}
1658
1659	static int cgroup_setup_root(struct cgroup_root *root, unsigned long ss_mask)
1660	{
1661	LIST_HEAD(tmp_links);
1662	struct cgroup *root_cgrp = &root->cgrp;
1663	struct cftype *base_files;
1664	struct css_set *cset;
1665	int i, ret;
1666
1667	lockdep_assert_held(&cgroup_mutex);
1668
1669	ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_NOWAIT);
1670	if (ret < 0)
1671	goto out;
1672	root_cgrp->id = ret;
1673
1674	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
1675	GFP_KERNEL);
1676	if (ret)
1677	goto out;
1678
1679	/*
1680	* We're accessing css_set_count without locking css_set_rwsem here,
1681	* but that's OK - it can only be increased by someone holding
1682	* cgroup_lock, and that's us. The worst that can happen is that we
1683	* have some link structures left over
1684	*/
1685	ret = allocate_cgrp_cset_links(css_set_count, &tmp_links);
1686	if (ret)
1687	goto cancel_ref;
1688
1689	ret = cgroup_init_root_id(root);
1690	if (ret)
1691	goto cancel_ref;
1692
1693	root->kf_root = kernfs_create_root(&cgroup_kf_syscall_ops,
1694	KERNFS_ROOT_CREATE_DEACTIVATED,
1695	root_cgrp);
1696	if (IS_ERR(root->kf_root)) {
1697	ret = PTR_ERR(root->kf_root);
1698	goto exit_root_id;
1699	}
1700	root_cgrp->kn = root->kf_root->kn;
1701
1702	if (root == &cgrp_dfl_root)
1703	base_files = cgroup_dfl_base_files;
1704	else
1705	base_files = cgroup_legacy_base_files;
1706
1707	ret = cgroup_addrm_files(root_cgrp, base_files, true);
1708	if (ret)
1709	goto destroy_root;
1710
1711	ret = rebind_subsystems(root, ss_mask);
1712	if (ret)
1713	goto destroy_root;
1714
1715	/*
1716	* There must be no failure case after here, since rebinding takes
1717	* care of subsystems' refcounts, which are explicitly dropped in
1718	* the failure exit path.
1719	*/
1720	list_add(&root->root_list, &cgroup_roots);
1721	cgroup_root_count++;
1722
1723	/*
1724	* Link the root cgroup in this hierarchy into all the css_set
1725	* objects.
1726	*/
1727	down_write(&css_set_rwsem);
1728	hash_for_each(css_set_table, i, cset, hlist)
1729	link_css_set(&tmp_links, cset, root_cgrp);
1730	up_write(&css_set_rwsem);
1731
1732	BUG_ON(!list_empty(&root_cgrp->self.children));
1733	BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1734
1735	kernfs_activate(root_cgrp->kn);
1736	ret = 0;
1737	goto out;
1738
1739	destroy_root:
1740	kernfs_destroy_root(root->kf_root);
1741	root->kf_root = NULL;
1742	exit_root_id:
1743	cgroup_exit_root_id(root);
1744	cancel_ref:
1745	percpu_ref_exit(&root_cgrp->self.refcnt);
1746	out:
1747	free_cgrp_cset_links(&tmp_links);
1748	return ret;
1749	}
1750
1751	static struct dentry cgroup_mount(struct file_system_type fs_type,
1752	int flags, const char *unused_dev_name,
1753	void *data)
1754	{
1755	struct super_block *pinned_sb = NULL;
1756	struct cgroup_subsys *ss;
1757	struct cgroup_root *root;
1758	struct cgroup_sb_opts opts;
1759	struct dentry *dentry;
1760	int ret;
1761	int i;
1762	bool new_sb;
1763
1764	/*
1765	* The first time anyone tries to mount a cgroup, enable the list
1766	* linking each css_set to its tasks and fix up all existing tasks.
1767	*/
1768	if (!use_task_css_set_links)
1769	cgroup_enable_task_cg_lists();
1770
1771	mutex_lock(&cgroup_mutex);
1772
1773	/* First find the desired set of subsystems */
1774	ret = parse_cgroupfs_options(data, &opts);
1775	if (ret)
1776	goto out_unlock;
1777
1778	/* look for a matching existing root */
1779	if (opts.flags & CGRP_ROOT_SANE_BEHAVIOR) {
1780	cgrp_dfl_root_visible = true;
1781	root = &cgrp_dfl_root;
1782	cgroup_get(&root->cgrp);
1783	ret = 0;
1784	goto out_unlock;
1785	}
1786
1787	/*
1788	* Destruction of cgroup root is asynchronous, so subsystems may
1789	* still be dying after the previous unmount. Let's drain the
1790	* dying subsystems. We just need to ensure that the ones
1791	* unmounted previously finish dying and don't care about new ones
1792	* starting. Testing ref liveliness is good enough.
1793	*/
1794	for_each_subsys(ss, i) {
1795	if (!(opts.subsys_mask & (1 << i)) \|\|
1796	ss->root == &cgrp_dfl_root)
1797	continue;
1798
1799	if (!percpu_ref_tryget_live(&ss->root->cgrp.self.refcnt)) {
1800	mutex_unlock(&cgroup_mutex);
1801	msleep(10);
1802	ret = restart_syscall();
1803	goto out_free;
1804	}
1805	cgroup_put(&ss->root->cgrp);
1806	}
1807
1808	for_each_root(root) {
1809	bool name_match = false;
1810
1811	if (root == &cgrp_dfl_root)
1812	continue;
1813
1814	/*
1815	* If we asked for a name then it must match. Also, if
1816	* name matches but sybsys_mask doesn't, we should fail.
1817	* Remember whether name matched.
1818	*/
1819	if (opts.name) {
1820	if (strcmp(opts.name, root->name))
1821	continue;
1822	name_match = true;
1823	}
1824
1825	/*
1826	* If we asked for subsystems (or explicitly for no
1827	* subsystems) then they must match.
1828	*/
1829	if ((opts.subsys_mask \|\| opts.none) &&
1830	(opts.subsys_mask != root->subsys_mask)) {
1831	if (!name_match)
1832	continue;
1833	ret = -EBUSY;
1834	goto out_unlock;
1835	}
1836
1837	if (root->flags ^ opts.flags)
1838	pr_warn("new mount options do not match the existing superblock, will be ignored\n");
1839
1840	/*
1841	* We want to reuse @root whose lifetime is governed by its
1842	* ->cgrp. Let's check whether @root is alive and keep it
1843	* that way. As cgroup_kill_sb() can happen anytime, we
1844	* want to block it by pinning the sb so that @root doesn't
1845	* get killed before mount is complete.
1846	*
1847	* With the sb pinned, tryget_live can reliably indicate
1848	* whether @root can be reused. If it's being killed,
1849	* drain it. We can use wait_queue for the wait but this
1850	* path is super cold. Let's just sleep a bit and retry.
1851	*/
1852	pinned_sb = kernfs_pin_sb(root->kf_root, NULL);
1853	if (IS_ERR(pinned_sb) \|\|
1854	!percpu_ref_tryget_live(&root->cgrp.self.refcnt)) {
1855	mutex_unlock(&cgroup_mutex);
1856	if (!IS_ERR_OR_NULL(pinned_sb))
1857	deactivate_super(pinned_sb);
1858	msleep(10);
1859	ret = restart_syscall();
1860	goto out_free;
1861	}
1862
1863	ret = 0;
1864	goto out_unlock;
1865	}
1866
1867	/*
1868	* No such thing, create a new one. name= matching without subsys
1869	* specification is allowed for already existing hierarchies but we
1870	* can't create new one without subsys specification.
1871	*/
1872	if (!opts.subsys_mask && !opts.none) {
1873	ret = -EINVAL;
1874	goto out_unlock;
1875	}
1876
1877	root = kzalloc(sizeof(*root), GFP_KERNEL);
1878	if (!root) {
1879	ret = -ENOMEM;
1880	goto out_unlock;
1881	}
1882
1883	init_cgroup_root(root, &opts);
1884
1885	ret = cgroup_setup_root(root, opts.subsys_mask);
1886	if (ret)
1887	cgroup_free_root(root);
1888
1889	out_unlock:
1890	mutex_unlock(&cgroup_mutex);
1891	out_free:
1892	kfree(opts.release_agent);
1893	kfree(opts.name);
1894
1895	if (ret)
1896	return ERR_PTR(ret);
1897
1898	dentry = kernfs_mount(fs_type, flags, root->kf_root,
1899	CGROUP_SUPER_MAGIC, &new_sb);
1900	if (IS_ERR(dentry) \|\| !new_sb)
1901	cgroup_put(&root->cgrp);
1902
1903	/*
1904	* If @pinned_sb, we're reusing an existing root and holding an
1905	* extra ref on its sb. Mount is complete. Put the extra ref.
1906	*/
1907	if (pinned_sb) {
1908	WARN_ON(new_sb);
1909	deactivate_super(pinned_sb);
1910	}
1911
1912	return dentry;
1913	}
1914
1915	static void cgroup_kill_sb(struct super_block *sb)
1916	{
1917	struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
1918	struct cgroup_root *root = cgroup_root_from_kf(kf_root);
1919
1920	/*
1921	* If @root doesn't have any mounts or children, start killing it.
1922	* This prevents new mounts by disabling percpu_ref_tryget_live().
1923	* cgroup_mount() may wait for @root's release.
1924	*
1925	* And don't kill the default root.
1926	*/
1927	if (!list_empty(&root->cgrp.self.children) \|\|
1928	root == &cgrp_dfl_root)
1929	cgroup_put(&root->cgrp);
1930	else
1931	percpu_ref_kill(&root->cgrp.self.refcnt);
1932
1933	kernfs_kill_sb(sb);
1934	}
1935
1936	static struct file_system_type cgroup_fs_type = {
1937	.name = "cgroup",
1938	.mount = cgroup_mount,
1939	.kill_sb = cgroup_kill_sb,
1940	};
1941
1942	/**
1943	* task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
1944	* @task: target task
1945	* @buf: the buffer to write the path into
1946	* @buflen: the length of the buffer
1947	*
1948	* Determine @task's cgroup on the first (the one with the lowest non-zero
1949	* hierarchy_id) cgroup hierarchy and copy its path into @buf. This
1950	* function grabs cgroup_mutex and shouldn't be used inside locks used by
1951	* cgroup controller callbacks.
1952	*
1953	* Return value is the same as kernfs_path().
1954	*/
1955	char task_cgroup_path(struct task_struct task, char *buf, size_t buflen)
1956	{
1957	struct cgroup_root *root;
1958	struct cgroup *cgrp;
1959	int hierarchy_id = 1;
1960	char *path = NULL;
1961
1962	mutex_lock(&cgroup_mutex);
1963	down_read(&css_set_rwsem);
1964
1965	root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
1966
1967	if (root) {
1968	cgrp = task_cgroup_from_root(task, root);
1969	path = cgroup_path(cgrp, buf, buflen);
1970	} else {
1971	/* if no hierarchy exists, everyone is in "/" */
1972	if (strlcpy(buf, "/", buflen) < buflen)
1973	path = buf;
1974	}
1975
1976	up_read(&css_set_rwsem);
1977	mutex_unlock(&cgroup_mutex);
1978	return path;
1979	}
1980	EXPORT_SYMBOL_GPL(task_cgroup_path);
1981
1982	/* used to track tasks and other necessary states during migration */
1983	struct cgroup_taskset {
1984	/* the src and dst cset list running through cset->mg_node */
1985	struct list_head src_csets;
1986	struct list_head dst_csets;
1987
1988	/*
1989	* Fields for cgroup_taskset_*() iteration.
1990	*
1991	* Before migration is committed, the target migration tasks are on
1992	* ->mg_tasks of the csets on ->src_csets. After, on ->mg_tasks of
1993	* the csets on ->dst_csets. ->csets point to either ->src_csets
1994	* or ->dst_csets depending on whether migration is committed.
1995	*
1996	* ->cur_csets and ->cur_task point to the current task position
1997	* during iteration.
1998	*/
1999	struct list_head *csets;
2000	struct css_set *cur_cset;
2001	struct task_struct *cur_task;
2002	};
2003
2004	/**
2005	* cgroup_taskset_first - reset taskset and return the first task
2006	* @tset: taskset of interest
2007	*
2008	* @tset iteration is initialized and the first task is returned.
2009	*/
2010	struct task_struct cgroup_taskset_first(struct cgroup_taskset tset)
2011	{
2012	tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2013	tset->cur_task = NULL;
2014
2015	return cgroup_taskset_next(tset);
2016	}
2017
2018	/**
2019	* cgroup_taskset_next - iterate to the next task in taskset
2020	* @tset: taskset of interest
2021	*
2022	* Return the next task in @tset. Iteration must have been initialized
2023	* with cgroup_taskset_first().
2024	*/
2025	struct task_struct cgroup_taskset_next(struct cgroup_taskset tset)
2026	{
2027	struct css_set *cset = tset->cur_cset;
2028	struct task_struct *task = tset->cur_task;
2029
2030	while (&cset->mg_node != tset->csets) {
2031	if (!task)
2032	task = list_first_entry(&cset->mg_tasks,
2033	struct task_struct, cg_list);
2034	else
2035	task = list_next_entry(task, cg_list);
2036
2037	if (&task->cg_list != &cset->mg_tasks) {
2038	tset->cur_cset = cset;
2039	tset->cur_task = task;
2040	return task;
2041	}
2042
2043	cset = list_next_entry(cset, mg_node);
2044	task = NULL;
2045	}
2046
2047	return NULL;
2048	}
2049
2050	/**
2051	* cgroup_task_migrate - move a task from one cgroup to another.
2052	* @old_cgrp: the cgroup @tsk is being migrated from
2053	* @tsk: the task being migrated
2054	* @new_cset: the new css_set @tsk is being attached to
2055	*
2056	* Must be called with cgroup_mutex, threadgroup and css_set_rwsem locked.
2057	*/
2058	static void cgroup_task_migrate(struct cgroup *old_cgrp,
2059	struct task_struct *tsk,
2060	struct css_set *new_cset)
2061	{
2062	struct css_set *old_cset;
2063
2064	lockdep_assert_held(&cgroup_mutex);
2065	lockdep_assert_held(&css_set_rwsem);
2066
2067	/*
2068	* We are synchronized through cgroup_threadgroup_rwsem against
2069	* PF_EXITING setting such that we can't race against cgroup_exit()
2070	* changing the css_set to init_css_set and dropping the old one.
2071	*/
2072	WARN_ON_ONCE(tsk->flags & PF_EXITING);
2073	old_cset = task_css_set(tsk);
2074
2075	get_css_set(new_cset);
2076	rcu_assign_pointer(tsk->cgroups, new_cset);
2077
2078	/*
2079	* Use move_tail so that cgroup_taskset_first() still returns the
2080	* leader after migration. This works because cgroup_migrate()
2081	* ensures that the dst_cset of the leader is the first on the
2082	* tset's dst_csets list.
2083	*/
2084	list_move_tail(&tsk->cg_list, &new_cset->mg_tasks);
2085
2086	/*
2087	* We just gained a reference on old_cset by taking it from the
2088	* task. As trading it for new_cset is protected by cgroup_mutex,
2089	* we're safe to drop it here; it will be freed under RCU.
2090	*/
2091	put_css_set_locked(old_cset);
2092	}
2093
2094	/**
2095	* cgroup_migrate_finish - cleanup after attach
2096	* @preloaded_csets: list of preloaded css_sets
2097	*
2098	* Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See
2099	* those functions for details.
2100	*/
2101	static void cgroup_migrate_finish(struct list_head *preloaded_csets)
2102	{
2103	struct css_set cset, tmp_cset;
2104
2105	lockdep_assert_held(&cgroup_mutex);
2106
2107	down_write(&css_set_rwsem);
2108	list_for_each_entry_safe(cset, tmp_cset, preloaded_csets, mg_preload_node) {
2109	cset->mg_src_cgrp = NULL;
2110	cset->mg_dst_cset = NULL;
2111	list_del_init(&cset->mg_preload_node);
2112	put_css_set_locked(cset);
2113	}
2114	up_write(&css_set_rwsem);
2115	}
2116
2117	/**
2118	* cgroup_migrate_add_src - add a migration source css_set
2119	* @src_cset: the source css_set to add
2120	* @dst_cgrp: the destination cgroup
2121	* @preloaded_csets: list of preloaded css_sets
2122	*
2123	* Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin
2124	* @src_cset and add it to @preloaded_csets, which should later be cleaned
2125	* up by cgroup_migrate_finish().
2126	*
2127	* This function may be called without holding cgroup_threadgroup_rwsem
2128	* even if the target is a process. Threads may be created and destroyed
2129	* but as long as cgroup_mutex is not dropped, no new css_set can be put
2130	* into play and the preloaded css_sets are guaranteed to cover all
2131	* migrations.
2132	*/
2133	static void cgroup_migrate_add_src(struct css_set *src_cset,
2134	struct cgroup *dst_cgrp,
2135	struct list_head *preloaded_csets)
2136	{
2137	struct cgroup *src_cgrp;
2138
2139	lockdep_assert_held(&cgroup_mutex);
2140	lockdep_assert_held(&css_set_rwsem);
2141
2142	src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2143
2144	if (!list_empty(&src_cset->mg_preload_node))
2145	return;
2146
2147	WARN_ON(src_cset->mg_src_cgrp);
2148	WARN_ON(!list_empty(&src_cset->mg_tasks));
2149	WARN_ON(!list_empty(&src_cset->mg_node));
2150
2151	src_cset->mg_src_cgrp = src_cgrp;
2152	get_css_set(src_cset);
2153	list_add(&src_cset->mg_preload_node, preloaded_csets);
2154	}
2155
2156	/**
2157	* cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2158	* @dst_cgrp: the destination cgroup (may be %NULL)
2159	* @preloaded_csets: list of preloaded source css_sets
2160	*
2161	* Tasks are about to be moved to @dst_cgrp and all the source css_sets
2162	* have been preloaded to @preloaded_csets. This function looks up and
2163	* pins all destination css_sets, links each to its source, and append them
2164	* to @preloaded_csets. If @dst_cgrp is %NULL, the destination of each
2165	* source css_set is assumed to be its cgroup on the default hierarchy.
2166	*
2167	* This function must be called after cgroup_migrate_add_src() has been
2168	* called on each migration source css_set. After migration is performed
2169	* using cgroup_migrate(), cgroup_migrate_finish() must be called on
2170	* @preloaded_csets.
2171	*/
2172	static int cgroup_migrate_prepare_dst(struct cgroup *dst_cgrp,
2173	struct list_head *preloaded_csets)
2174	{
2175	LIST_HEAD(csets);
2176	struct css_set src_cset, tmp_cset;
2177
2178	lockdep_assert_held(&cgroup_mutex);
2179
2180	/*
2181	* Except for the root, child_subsys_mask must be zero for a cgroup
2182	* with tasks so that child cgroups don't compete against tasks.
2183	*/
2184	if (dst_cgrp && cgroup_on_dfl(dst_cgrp) && cgroup_parent(dst_cgrp) &&
2185	dst_cgrp->child_subsys_mask)
2186	return -EBUSY;
2187
2188	/* look up the dst cset for each src cset and link it to src */
2189	list_for_each_entry_safe(src_cset, tmp_cset, preloaded_csets, mg_preload_node) {
2190	struct css_set *dst_cset;
2191
2192	dst_cset = find_css_set(src_cset,
2193	dst_cgrp ?: src_cset->dfl_cgrp);
2194	if (!dst_cset)
2195	goto err;
2196
2197	WARN_ON_ONCE(src_cset->mg_dst_cset \|\| dst_cset->mg_dst_cset);
2198
2199	/*
2200	* If src cset equals dst, it's noop. Drop the src.
2201	* cgroup_migrate() will skip the cset too. Note that we
2202	* can't handle src == dst as some nodes are used by both.
2203	*/
2204	if (src_cset == dst_cset) {
2205	src_cset->mg_src_cgrp = NULL;
2206	list_del_init(&src_cset->mg_preload_node);
2207	put_css_set(src_cset);
2208	put_css_set(dst_cset);
2209	continue;
2210	}
2211
2212	src_cset->mg_dst_cset = dst_cset;
2213
2214	if (list_empty(&dst_cset->mg_preload_node))
2215	list_add(&dst_cset->mg_preload_node, &csets);
2216	else
2217	put_css_set(dst_cset);
2218	}
2219
2220	list_splice_tail(&csets, preloaded_csets);
2221	return 0;
2222	err:
2223	cgroup_migrate_finish(&csets);
2224	return -ENOMEM;
2225	}
2226
2227	/**
2228	* cgroup_migrate - migrate a process or task to a cgroup
2229	* @cgrp: the destination cgroup
2230	* @leader: the leader of the process or the task to migrate
2231	* @threadgroup: whether @leader points to the whole process or a single task
2232	*
2233	* Migrate a process or task denoted by @leader to @cgrp. If migrating a
2234	* process, the caller must be holding cgroup_threadgroup_rwsem. The
2235	* caller is also responsible for invoking cgroup_migrate_add_src() and
2236	* cgroup_migrate_prepare_dst() on the targets before invoking this
2237	* function and following up with cgroup_migrate_finish().
2238	*
2239	* As long as a controller's ->can_attach() doesn't fail, this function is
2240	* guaranteed to succeed. This means that, excluding ->can_attach()
2241	* failure, when migrating multiple targets, the success or failure can be
2242	* decided for all targets by invoking group_migrate_prepare_dst() before
2243	* actually starting migrating.
2244	*/
2245	static int cgroup_migrate(struct cgroup cgrp, struct task_struct leader,
2246	bool threadgroup)
2247	{
2248	struct cgroup_taskset tset = {
2249	.src_csets = LIST_HEAD_INIT(tset.src_csets),
2250	.dst_csets = LIST_HEAD_INIT(tset.dst_csets),
2251	.csets = &tset.src_csets,
2252	};
2253	struct cgroup_subsys_state css, failed_css = NULL;
2254	struct css_set cset, tmp_cset;
2255	struct task_struct task, tmp_task;
2256	int i, ret;
2257
2258	/*
2259	* Prevent freeing of tasks while we take a snapshot. Tasks that are
2260	* already PF_EXITING could be freed from underneath us unless we
2261	* take an rcu_read_lock.
2262	*/
2263	down_write(&css_set_rwsem);
2264	rcu_read_lock();
2265	task = leader;
2266	do {
2267	/* @task either already exited or can't exit until the end */
2268	if (task->flags & PF_EXITING)
2269	goto next;
2270
2271	/* leave @task alone if post_fork() hasn't linked it yet */
2272	if (list_empty(&task->cg_list))
2273	goto next;
2274
2275	cset = task_css_set(task);
2276	if (!cset->mg_src_cgrp)
2277	goto next;
2278
2279	/*
2280	* cgroup_taskset_first() must always return the leader.
2281	* Take care to avoid disturbing the ordering.
2282	*/
2283	list_move_tail(&task->cg_list, &cset->mg_tasks);
2284	if (list_empty(&cset->mg_node))
2285	list_add_tail(&cset->mg_node, &tset.src_csets);
2286	if (list_empty(&cset->mg_dst_cset->mg_node))
2287	list_move_tail(&cset->mg_dst_cset->mg_node,
2288	&tset.dst_csets);
2289	next:
2290	if (!threadgroup)
2291	break;
2292	} while_each_thread(leader, task);
2293	rcu_read_unlock();
2294	up_write(&css_set_rwsem);
2295
2296	/* methods shouldn't be called if no task is actually migrating */
2297	if (list_empty(&tset.src_csets))
2298	return 0;
2299
2300	/* check that we can legitimately attach to the cgroup */
2301	for_each_e_css(css, i, cgrp) {
2302	if (css->ss->can_attach) {
2303	ret = css->ss->can_attach(css, &tset);
2304	if (ret) {
2305	failed_css = css;
2306	goto out_cancel_attach;
2307	}
2308	}
2309	}
2310
2311	/*
2312	* Now that we're guaranteed success, proceed to move all tasks to
2313	* the new cgroup. There are no failure cases after here, so this
2314	* is the commit point.
2315	*/
2316	down_write(&css_set_rwsem);
2317	list_for_each_entry(cset, &tset.src_csets, mg_node) {
2318	list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list)
2319	cgroup_task_migrate(cset->mg_src_cgrp, task,
2320	cset->mg_dst_cset);
2321	}
2322	up_write(&css_set_rwsem);
2323
2324	/*
2325	* Migration is committed, all target tasks are now on dst_csets.
2326	* Nothing is sensitive to fork() after this point. Notify
2327	* controllers that migration is complete.
2328	*/
2329	tset.csets = &tset.dst_csets;
2330
2331	for_each_e_css(css, i, cgrp)
2332	if (css->ss->attach)
2333	css->ss->attach(css, &tset);
2334
2335	ret = 0;
2336	goto out_release_tset;
2337
2338	out_cancel_attach:
2339	for_each_e_css(css, i, cgrp) {
2340	if (css == failed_css)
2341	break;
2342	if (css->ss->cancel_attach)
2343	css->ss->cancel_attach(css, &tset);
2344	}
2345	out_release_tset:
2346	down_write(&css_set_rwsem);
2347	list_splice_init(&tset.dst_csets, &tset.src_csets);
2348	list_for_each_entry_safe(cset, tmp_cset, &tset.src_csets, mg_node) {
2349	list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2350	list_del_init(&cset->mg_node);
2351	}
2352	up_write(&css_set_rwsem);
2353	return ret;
2354	}
2355
2356	/**
2357	* cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2358	* @dst_cgrp: the cgroup to attach to
2359	* @leader: the task or the leader of the threadgroup to be attached
2360	* @threadgroup: attach the whole threadgroup?
2361	*
2362	* Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2363	*/
2364	static int cgroup_attach_task(struct cgroup *dst_cgrp,
2365	struct task_struct *leader, bool threadgroup)
2366	{
2367	LIST_HEAD(preloaded_csets);
2368	struct task_struct *task;
2369	int ret;
2370
2371	/* look up all src csets */
2372	down_read(&css_set_rwsem);
2373	rcu_read_lock();
2374	task = leader;
2375	do {
2376	cgroup_migrate_add_src(task_css_set(task), dst_cgrp,
2377	&preloaded_csets);
2378	if (!threadgroup)
2379	break;
2380	} while_each_thread(leader, task);
2381	rcu_read_unlock();
2382	up_read(&css_set_rwsem);
2383
2384	/* prepare dst csets and commit */
2385	ret = cgroup_migrate_prepare_dst(dst_cgrp, &preloaded_csets);
2386	if (!ret)
2387	ret = cgroup_migrate(dst_cgrp, leader, threadgroup);
2388
2389	cgroup_migrate_finish(&preloaded_csets);
2390	return ret;
2391	}
2392
2393	static int cgroup_procs_write_permission(struct task_struct *task,
2394	struct cgroup *dst_cgrp,
2395	struct kernfs_open_file *of)
2396	{
2397	const struct cred *cred = current_cred();
2398	const struct cred *tcred = get_task_cred(task);
2399	int ret = 0;
2400
2401	/*
2402	* even if we're attaching all tasks in the thread group, we only
2403	* need to check permissions on one of them.
2404	*/
2405	if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
2406	!uid_eq(cred->euid, tcred->uid) &&
2407	!uid_eq(cred->euid, tcred->suid))
2408	ret = -EACCES;
2409
2410	if (!ret && cgroup_on_dfl(dst_cgrp)) {
2411	struct super_block *sb = of->file->f_path.dentry->d_sb;
2412	struct cgroup *cgrp;
2413	struct inode *inode;
2414
2415	down_read(&css_set_rwsem);
2416	cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
2417	up_read(&css_set_rwsem);
2418
2419	while (!cgroup_is_descendant(dst_cgrp, cgrp))
2420	cgrp = cgroup_parent(cgrp);
2421
2422	ret = -ENOMEM;
2423	inode = kernfs_get_inode(sb, cgrp->procs_kn);
2424	if (inode) {
2425	ret = inode_permission(inode, MAY_WRITE);
2426	iput(inode);
2427	}
2428	}
2429
2430	put_cred(tcred);
2431	return ret;
2432	}
2433
2434	/*
2435	* Find the task_struct of the task to attach by vpid and pass it along to the
2436	* function to attach either it or all tasks in its threadgroup. Will lock
2437	* cgroup_mutex and threadgroup.
2438	*/
2439	static ssize_t __cgroup_procs_write(struct kernfs_open_file of, char buf,
2440	size_t nbytes, loff_t off, bool threadgroup)
2441	{
2442	struct task_struct *tsk;
2443	struct cgroup *cgrp;
2444	pid_t pid;
2445	int ret;
2446
2447	if (kstrtoint(strstrip(buf), 0, &pid) \|\| pid < 0)
2448	return -EINVAL;
2449
2450	cgrp = cgroup_kn_lock_live(of->kn);
2451	if (!cgrp)
2452	return -ENODEV;
2453
2454	percpu_down_write(&cgroup_threadgroup_rwsem);
2455	rcu_read_lock();
2456	if (pid) {
2457	tsk = find_task_by_vpid(pid);
2458	if (!tsk) {
2459	ret = -ESRCH;
2460	goto out_unlock_rcu;
2461	}
2462	} else {
2463	tsk = current;
2464	}
2465
2466	if (threadgroup)
2467	tsk = tsk->group_leader;
2468
2469	/*
2470	* Workqueue threads may acquire PF_NO_SETAFFINITY and become
2471	* trapped in a cpuset, or RT worker may be born in a cgroup
2472	* with no rt_runtime allocated. Just say no.
2473	*/
2474	if (tsk == kthreadd_task \|\| (tsk->flags & PF_NO_SETAFFINITY)) {
2475	ret = -EINVAL;
2476	goto out_unlock_rcu;
2477	}
2478
2479	get_task_struct(tsk);
2480	rcu_read_unlock();
2481
2482	ret = cgroup_procs_write_permission(tsk, cgrp, of);
2483	if (!ret)
2484	ret = cgroup_attach_task(cgrp, tsk, threadgroup);
2485
2486	put_task_struct(tsk);
2487	goto out_unlock_threadgroup;
2488
2489	out_unlock_rcu:
2490	rcu_read_unlock();
2491	out_unlock_threadgroup:
2492	percpu_up_write(&cgroup_threadgroup_rwsem);
2493	cgroup_kn_unlock(of->kn);
2494	return ret ?: nbytes;
2495	}
2496
2497	/**
2498	* cgroup_attach_task_all - attach task 'tsk' to all cgroups of task 'from'
2499	* @from: attach to all cgroups of a given task
2500	* @tsk: the task to be attached
2501	*/
2502	int cgroup_attach_task_all(struct task_struct from, struct task_struct tsk)
2503	{
2504	struct cgroup_root *root;
2505	int retval = 0;
2506
2507	mutex_lock(&cgroup_mutex);
2508	for_each_root(root) {
2509	struct cgroup *from_cgrp;
2510
2511	if (root == &cgrp_dfl_root)
2512	continue;
2513
2514	down_read(&css_set_rwsem);
2515	from_cgrp = task_cgroup_from_root(from, root);
2516	up_read(&css_set_rwsem);
2517
2518	retval = cgroup_attach_task(from_cgrp, tsk, false);
2519	if (retval)
2520	break;
2521	}
2522	mutex_unlock(&cgroup_mutex);
2523
2524	return retval;
2525	}
2526	EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
2527
2528	static ssize_t cgroup_tasks_write(struct kernfs_open_file *of,
2529	char *buf, size_t nbytes, loff_t off)
2530	{
2531	return __cgroup_procs_write(of, buf, nbytes, off, false);
2532	}
2533
2534	static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
2535	char *buf, size_t nbytes, loff_t off)
2536	{
2537	return __cgroup_procs_write(of, buf, nbytes, off, true);
2538	}
2539
2540	static ssize_t cgroup_release_agent_write(struct kernfs_open_file *of,
2541	char *buf, size_t nbytes, loff_t off)
2542	{
2543	struct cgroup *cgrp;
2544
2545	BUILD_BUG_ON(sizeof(cgrp->root->release_agent_path) < PATH_MAX);
2546
2547	cgrp = cgroup_kn_lock_live(of->kn);
2548	if (!cgrp)
2549	return -ENODEV;
2550	spin_lock(&release_agent_path_lock);
2551	strlcpy(cgrp->root->release_agent_path, strstrip(buf),
2552	sizeof(cgrp->root->release_agent_path));
2553	spin_unlock(&release_agent_path_lock);
2554	cgroup_kn_unlock(of->kn);
2555	return nbytes;
2556	}
2557
2558	static int cgroup_release_agent_show(struct seq_file seq, void v)
2559	{
2560	struct cgroup *cgrp = seq_css(seq)->cgroup;
2561
2562	spin_lock(&release_agent_path_lock);
2563	seq_puts(seq, cgrp->root->release_agent_path);
2564	spin_unlock(&release_agent_path_lock);
2565	seq_putc(seq, '\n');
2566	return 0;
2567	}
2568
2569	static int cgroup_sane_behavior_show(struct seq_file seq, void v)
2570	{
2571	seq_puts(seq, "0\n");
2572	return 0;
2573	}
2574
2575	static void cgroup_print_ss_mask(struct seq_file *seq, unsigned long ss_mask)
2576	{
2577	struct cgroup_subsys *ss;
2578	bool printed = false;
2579	int ssid;
2580
2581	for_each_subsys_which(ss, ssid, &ss_mask) {
2582	if (printed)
2583	seq_putc(seq, ' ');
2584	seq_printf(seq, "%s", ss->name);
2585	printed = true;
2586	}
2587	if (printed)
2588	seq_putc(seq, '\n');
2589	}
2590
2591	/* show controllers which are currently attached to the default hierarchy */
2592	static int cgroup_root_controllers_show(struct seq_file seq, void v)
2593	{
2594	struct cgroup *cgrp = seq_css(seq)->cgroup;
2595
2596	cgroup_print_ss_mask(seq, cgrp->root->subsys_mask &
2597	~cgrp_dfl_root_inhibit_ss_mask);
2598	return 0;
2599	}
2600
2601	/* show controllers which are enabled from the parent */
2602	static int cgroup_controllers_show(struct seq_file seq, void v)
2603	{
2604	struct cgroup *cgrp = seq_css(seq)->cgroup;
2605
2606	cgroup_print_ss_mask(seq, cgroup_parent(cgrp)->subtree_control);
2607	return 0;
2608	}
2609
2610	/* show controllers which are enabled for a given cgroup's children */
2611	static int cgroup_subtree_control_show(struct seq_file seq, void v)
2612	{
2613	struct cgroup *cgrp = seq_css(seq)->cgroup;
2614
2615	cgroup_print_ss_mask(seq, cgrp->subtree_control);
2616	return 0;
2617	}
2618
2619	/**
2620	* cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2621	* @cgrp: root of the subtree to update csses for
2622	*
2623	* @cgrp's child_subsys_mask has changed and its subtree's (self excluded)
2624	* css associations need to be updated accordingly. This function looks up
2625	* all css_sets which are attached to the subtree, creates the matching
2626	* updated css_sets and migrates the tasks to the new ones.
2627	*/
2628	static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2629	{
2630	LIST_HEAD(preloaded_csets);
2631	struct cgroup_subsys_state *css;
2632	struct css_set *src_cset;
2633	int ret;
2634
2635	lockdep_assert_held(&cgroup_mutex);
2636
2637	percpu_down_write(&cgroup_threadgroup_rwsem);
2638
2639	/* look up all csses currently attached to @cgrp's subtree */
2640	down_read(&css_set_rwsem);
2641	css_for_each_descendant_pre(css, cgroup_css(cgrp, NULL)) {
2642	struct cgrp_cset_link *link;
2643
2644	/* self is not affected by child_subsys_mask change */
2645	if (css->cgroup == cgrp)
2646	continue;
2647
2648	list_for_each_entry(link, &css->cgroup->cset_links, cset_link)
2649	cgroup_migrate_add_src(link->cset, cgrp,
2650	&preloaded_csets);
2651	}
2652	up_read(&css_set_rwsem);
2653
2654	/* NULL dst indicates self on default hierarchy */
2655	ret = cgroup_migrate_prepare_dst(NULL, &preloaded_csets);
2656	if (ret)
2657	goto out_finish;
2658
2659	list_for_each_entry(src_cset, &preloaded_csets, mg_preload_node) {
2660	struct task_struct last_task = NULL, task;
2661
2662	/* src_csets precede dst_csets, break on the first dst_cset */
2663	if (!src_cset->mg_src_cgrp)
2664	break;
2665
2666	/*
2667	* All tasks in src_cset need to be migrated to the
2668	* matching dst_cset. Empty it process by process. We
2669	* walk tasks but migrate processes. The leader might even
2670	* belong to a different cset but such src_cset would also
2671	* be among the target src_csets because the default
2672	* hierarchy enforces per-process membership.
2673	*/
2674	while (true) {
2675	down_read(&css_set_rwsem);
2676	task = list_first_entry_or_null(&src_cset->tasks,
2677	struct task_struct, cg_list);
2678	if (task) {
2679	task = task->group_leader;
2680	WARN_ON_ONCE(!task_css_set(task)->mg_src_cgrp);
2681	get_task_struct(task);
2682	}
2683	up_read(&css_set_rwsem);
2684
2685	if (!task)
2686	break;
2687
2688	/* guard against possible infinite loop */
2689	if (WARN(last_task == task,
2690	"cgroup: update_dfl_csses failed to make progress, aborting in inconsistent state\n"))
2691	goto out_finish;
2692	last_task = task;
2693
2694	ret = cgroup_migrate(src_cset->dfl_cgrp, task, true);
2695
2696	put_task_struct(task);
2697
2698	if (WARN(ret, "cgroup: failed to update controllers for the default hierarchy (%d), further operations may crash or hang\n", ret))
2699	goto out_finish;
2700	}
2701	}
2702
2703	out_finish:
2704	cgroup_migrate_finish(&preloaded_csets);
2705	percpu_up_write(&cgroup_threadgroup_rwsem);
2706	return ret;
2707	}
2708
2709	/* change the enabled child controllers for a cgroup in the default hierarchy */
2710	static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
2711	char *buf, size_t nbytes,
2712	loff_t off)
2713	{
2714	unsigned long enable = 0, disable = 0;
2715	unsigned long css_enable, css_disable, old_sc, new_sc, old_ss, new_ss;
2716	struct cgroup cgrp, child;
2717	struct cgroup_subsys *ss;
2718	char *tok;
2719	int ssid, ret;
2720
2721	/*
2722	* Parse input - space separated list of subsystem names prefixed
2723	* with either + or -.
2724	*/
2725	buf = strstrip(buf);
2726	while ((tok = strsep(&buf, " "))) {
2727	unsigned long tmp_ss_mask = ~cgrp_dfl_root_inhibit_ss_mask;
2728
2729	if (tok[0] == '\0')
2730	continue;
2731	for_each_subsys_which(ss, ssid, &tmp_ss_mask) {
2732	if (ss->disabled \|\| strcmp(tok + 1, ss->name))
2733	continue;
2734
2735	if (*tok == '+') {
2736	enable \|= 1 << ssid;
2737	disable &= ~(1 << ssid);
2738	} else if (*tok == '-') {
2739	disable \|= 1 << ssid;
2740	enable &= ~(1 << ssid);
2741	} else {
2742	return -EINVAL;
2743	}
2744	break;
2745	}
2746	if (ssid == CGROUP_SUBSYS_COUNT)
2747	return -EINVAL;
2748	}
2749
2750	cgrp = cgroup_kn_lock_live(of->kn);
2751	if (!cgrp)
2752	return -ENODEV;
2753
2754	for_each_subsys(ss, ssid) {
2755	if (enable & (1 << ssid)) {
2756	if (cgrp->subtree_control & (1 << ssid)) {
2757	enable &= ~(1 << ssid);
2758	continue;
2759	}
2760
2761	/* unavailable or not enabled on the parent? */
2762	if (!(cgrp_dfl_root.subsys_mask & (1 << ssid)) \|\|
2763	(cgroup_parent(cgrp) &&
2764	!(cgroup_parent(cgrp)->subtree_control & (1 << ssid)))) {
2765	ret = -ENOENT;
2766	goto out_unlock;
2767	}
2768	} else if (disable & (1 << ssid)) {
2769	if (!(cgrp->subtree_control & (1 << ssid))) {
2770	disable &= ~(1 << ssid);
2771	continue;
2772	}
2773
2774	/* a child has it enabled? */
2775	cgroup_for_each_live_child(child, cgrp) {
2776	if (child->subtree_control & (1 << ssid)) {
2777	ret = -EBUSY;
2778	goto out_unlock;
2779	}
2780	}
2781	}
2782	}
2783
2784	if (!enable && !disable) {
2785	ret = 0;
2786	goto out_unlock;
2787	}
2788
2789	/*
2790	* Except for the root, subtree_control must be zero for a cgroup
2791	* with tasks so that child cgroups don't compete against tasks.
2792	*/
2793	if (enable && cgroup_parent(cgrp) && !list_empty(&cgrp->cset_links)) {
2794	ret = -EBUSY;
2795	goto out_unlock;
2796	}
2797
2798	/*
2799	* Update subsys masks and calculate what needs to be done. More
2800	* subsystems than specified may need to be enabled or disabled
2801	* depending on subsystem dependencies.
2802	*/
2803	old_sc = cgrp->subtree_control;
2804	old_ss = cgrp->child_subsys_mask;
2805	new_sc = (old_sc \| enable) & ~disable;
2806	new_ss = cgroup_calc_child_subsys_mask(cgrp, new_sc);
2807
2808	css_enable = ~old_ss & new_ss;
2809	css_disable = old_ss & ~new_ss;
2810	enable \|= css_enable;
2811	disable \|= css_disable;
2812
2813	/*
2814	* Because css offlining is asynchronous, userland might try to
2815	* re-enable the same controller while the previous instance is
2816	* still around. In such cases, wait till it's gone using
2817	* offline_waitq.
2818	*/
2819	for_each_subsys_which(ss, ssid, &css_enable) {
2820	cgroup_for_each_live_child(child, cgrp) {
2821	DEFINE_WAIT(wait);
2822
2823	if (!cgroup_css(child, ss))
2824	continue;
2825
2826	cgroup_get(child);
2827	prepare_to_wait(&child->offline_waitq, &wait,
2828	TASK_UNINTERRUPTIBLE);
2829	cgroup_kn_unlock(of->kn);
2830	schedule();
2831	finish_wait(&child->offline_waitq, &wait);
2832	cgroup_put(child);
2833
2834	return restart_syscall();
2835	}
2836	}
2837
2838	cgrp->subtree_control = new_sc;
2839	cgrp->child_subsys_mask = new_ss;
2840
2841	/*
2842	* Create new csses or make the existing ones visible. A css is
2843	* created invisible if it's being implicitly enabled through
2844	* dependency. An invisible css is made visible when the userland
2845	* explicitly enables it.
2846	*/
2847	for_each_subsys(ss, ssid) {
2848	if (!(enable & (1 << ssid)))
2849	continue;
2850
2851	cgroup_for_each_live_child(child, cgrp) {
2852	if (css_enable & (1 << ssid))
2853	ret = create_css(child, ss,
2854	cgrp->subtree_control & (1 << ssid));
2855	else
2856	ret = cgroup_populate_dir(child, 1 << ssid);
2857	if (ret)
2858	goto err_undo_css;
2859	}
2860	}
2861
2862	/*
2863	* At this point, cgroup_e_css() results reflect the new csses
2864	* making the following cgroup_update_dfl_csses() properly update
2865	* css associations of all tasks in the subtree.
2866	*/
2867	ret = cgroup_update_dfl_csses(cgrp);
2868	if (ret)
2869	goto err_undo_css;
2870
2871	/*
2872	* All tasks are migrated out of disabled csses. Kill or hide
2873	* them. A css is hidden when the userland requests it to be
2874	* disabled while other subsystems are still depending on it. The
2875	* css must not actively control resources and be in the vanilla
2876	* state if it's made visible again later. Controllers which may
2877	* be depended upon should provide ->css_reset() for this purpose.
2878	*/
2879	for_each_subsys(ss, ssid) {
2880	if (!(disable & (1 << ssid)))
2881	continue;
2882
2883	cgroup_for_each_live_child(child, cgrp) {
2884	struct cgroup_subsys_state *css = cgroup_css(child, ss);
2885
2886	if (css_disable & (1 << ssid)) {
2887	kill_css(css);
2888	} else {
2889	cgroup_clear_dir(child, 1 << ssid);
2890	if (ss->css_reset)
2891	ss->css_reset(css);
2892	}
2893	}
2894	}
2895
2896	/*
2897	* The effective csses of all the descendants (excluding @cgrp) may
2898	* have changed. Subsystems can optionally subscribe to this event
2899	* by implementing ->css_e_css_changed() which is invoked if any of
2900	* the effective csses seen from the css's cgroup may have changed.
2901	*/
2902	for_each_subsys(ss, ssid) {
2903	struct cgroup_subsys_state *this_css = cgroup_css(cgrp, ss);
2904	struct cgroup_subsys_state *css;
2905
2906	if (!ss->css_e_css_changed \|\| !this_css)
2907	continue;
2908
2909	css_for_each_descendant_pre(css, this_css)
2910	if (css != this_css)
2911	ss->css_e_css_changed(css);
2912	}
2913
2914	kernfs_activate(cgrp->kn);
2915	ret = 0;
2916	out_unlock:
2917	cgroup_kn_unlock(of->kn);
2918	return ret ?: nbytes;
2919
2920	err_undo_css:
2921	cgrp->subtree_control = old_sc;
2922	cgrp->child_subsys_mask = old_ss;
2923
2924	for_each_subsys(ss, ssid) {
2925	if (!(enable & (1 << ssid)))
2926	continue;
2927
2928	cgroup_for_each_live_child(child, cgrp) {
2929	struct cgroup_subsys_state *css = cgroup_css(child, ss);
2930
2931	if (!css)
2932	continue;
2933
2934	if (css_enable & (1 << ssid))
2935	kill_css(css);
2936	else
2937	cgroup_clear_dir(child, 1 << ssid);
2938	}
2939	}
2940	goto out_unlock;
2941	}
2942
2943	static int cgroup_populated_show(struct seq_file seq, void v)
2944	{
2945	seq_printf(seq, "%d\n", (bool)seq_css(seq)->cgroup->populated_cnt);
2946	return 0;
2947	}
2948
2949	static ssize_t cgroup_file_write(struct kernfs_open_file of, char buf,
2950	size_t nbytes, loff_t off)
2951	{
2952	struct cgroup *cgrp = of->kn->parent->priv;
2953	struct cftype *cft = of->kn->priv;
2954	struct cgroup_subsys_state *css;
2955	int ret;
2956
2957	if (cft->write)
2958	return cft->write(of, buf, nbytes, off);
2959
2960	/*
2961	* kernfs guarantees that a file isn't deleted with operations in
2962	* flight, which means that the matching css is and stays alive and
2963	* doesn't need to be pinned. The RCU locking is not necessary
2964	* either. It's just for the convenience of using cgroup_css().
2965	*/
2966	rcu_read_lock();
2967	css = cgroup_css(cgrp, cft->ss);
2968	rcu_read_unlock();
2969
2970	if (cft->write_u64) {
2971	unsigned long long v;
2972	ret = kstrtoull(buf, 0, &v);
2973	if (!ret)
2974	ret = cft->write_u64(css, cft, v);
2975	} else if (cft->write_s64) {
2976	long long v;
2977	ret = kstrtoll(buf, 0, &v);
2978	if (!ret)
2979	ret = cft->write_s64(css, cft, v);
2980	} else {
2981	ret = -EINVAL;
2982	}
2983
2984	return ret ?: nbytes;
2985	}
2986
2987	static void cgroup_seqfile_start(struct seq_file seq, loff_t *ppos)
2988	{
2989	return seq_cft(seq)->seq_start(seq, ppos);
2990	}
2991
2992	static void cgroup_seqfile_next(struct seq_file seq, void v, loff_t ppos)
2993	{
2994	return seq_cft(seq)->seq_next(seq, v, ppos);
2995	}
2996
2997	static void cgroup_seqfile_stop(struct seq_file seq, void v)
2998	{
2999	seq_cft(seq)->seq_stop(seq, v);
3000	}
3001
3002	static int cgroup_seqfile_show(struct seq_file m, void arg)
3003	{
3004	struct cftype *cft = seq_cft(m);
3005	struct cgroup_subsys_state *css = seq_css(m);
3006
3007	if (cft->seq_show)
3008	return cft->seq_show(m, arg);
3009
3010	if (cft->read_u64)
3011	seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3012	else if (cft->read_s64)
3013	seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3014	else
3015	return -EINVAL;
3016	return 0;
3017	}
3018
3019	static struct kernfs_ops cgroup_kf_single_ops = {
3020	.atomic_write_len = PAGE_SIZE,
3021	.write = cgroup_file_write,
3022	.seq_show = cgroup_seqfile_show,
3023	};
3024
3025	static struct kernfs_ops cgroup_kf_ops = {
3026	.atomic_write_len = PAGE_SIZE,
3027	.write = cgroup_file_write,
3028	.seq_start = cgroup_seqfile_start,
3029	.seq_next = cgroup_seqfile_next,
3030	.seq_stop = cgroup_seqfile_stop,
3031	.seq_show = cgroup_seqfile_show,
3032	};
3033
3034	/*
3035	* cgroup_rename - Only allow simple rename of directories in place.
3036	*/
3037	static int cgroup_rename(struct kernfs_node kn, struct kernfs_node new_parent,
3038	const char *new_name_str)
3039	{
3040	struct cgroup *cgrp = kn->priv;
3041	int ret;
3042
3043	if (kernfs_type(kn) != KERNFS_DIR)
3044	return -ENOTDIR;
3045	if (kn->parent != new_parent)
3046	return -EIO;
3047
3048	/*
3049	* This isn't a proper migration and its usefulness is very
3050	* limited. Disallow on the default hierarchy.
3051	*/
3052	if (cgroup_on_dfl(cgrp))
3053	return -EPERM;
3054
3055	/*
3056	* We're gonna grab cgroup_mutex which nests outside kernfs
3057	* active_ref. kernfs_rename() doesn't require active_ref
3058	* protection. Break them before grabbing cgroup_mutex.
3059	*/
3060	kernfs_break_active_protection(new_parent);
3061	kernfs_break_active_protection(kn);
3062
3063	mutex_lock(&cgroup_mutex);
3064
3065	ret = kernfs_rename(kn, new_parent, new_name_str);
3066
3067	mutex_unlock(&cgroup_mutex);
3068
3069	kernfs_unbreak_active_protection(kn);
3070	kernfs_unbreak_active_protection(new_parent);
3071	return ret;
3072	}
3073
3074	/* set uid and gid of cgroup dirs and files to that of the creator */
3075	static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3076	{
3077	struct iattr iattr = { .ia_valid = ATTR_UID \| ATTR_GID,
3078	.ia_uid = current_fsuid(),
3079	.ia_gid = current_fsgid(), };
3080
3081	if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3082	gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3083	return 0;
3084
3085	return kernfs_setattr(kn, &iattr);
3086	}
3087
3088	static int cgroup_add_file(struct cgroup cgrp, struct cftype cft)
3089	{
3090	char name[CGROUP_FILE_NAME_MAX];
3091	struct kernfs_node *kn;
3092	struct lock_class_key *key = NULL;
3093	int ret;
3094
3095	#ifdef CONFIG_DEBUG_LOCK_ALLOC
3096	key = &cft->lockdep_key;
3097	#endif
3098	kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3099	cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3100	NULL, key);
3101	if (IS_ERR(kn))
3102	return PTR_ERR(kn);
3103
3104	ret = cgroup_kn_set_ugid(kn);
3105	if (ret) {
3106	kernfs_remove(kn);
3107	return ret;
3108	}
3109
3110	if (cft->write == cgroup_procs_write)
3111	cgrp->procs_kn = kn;
3112	else if (cft->seq_show == cgroup_populated_show)
3113	cgrp->populated_kn = kn;
3114	return 0;
3115	}
3116
3117	/**
3118	* cgroup_addrm_files - add or remove files to a cgroup directory
3119	* @cgrp: the target cgroup
3120	* @cfts: array of cftypes to be added
3121	* @is_add: whether to add or remove
3122	*
3123	* Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3124	* For removals, this function never fails. If addition fails, this
3125	* function doesn't remove files already added. The caller is responsible
3126	* for cleaning up.
3127	*/
3128	static int cgroup_addrm_files(struct cgroup *cgrp, struct cftype cfts[],
3129	bool is_add)
3130	{
3131	struct cftype *cft;
3132	int ret;
3133
3134	lockdep_assert_held(&cgroup_mutex);
3135
3136	for (cft = cfts; cft->name[0] != '\0'; cft++) {
3137	/* does cft->flags tell us to skip this file on @cgrp? */
3138	if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3139	continue;
3140	if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3141	continue;
3142	if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3143	continue;
3144	if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3145	continue;
3146
3147	if (is_add) {
3148	ret = cgroup_add_file(cgrp, cft);
3149	if (ret) {
3150	pr_warn("%s: failed to add %s, err=%d\n",
3151	__func__, cft->name, ret);
3152	return ret;
3153	}
3154	} else {
3155	cgroup_rm_file(cgrp, cft);
3156	}
3157	}
3158	return 0;
3159	}
3160
3161	static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3162	{
3163	LIST_HEAD(pending);
3164	struct cgroup_subsys *ss = cfts[0].ss;
3165	struct cgroup *root = &ss->root->cgrp;
3166	struct cgroup_subsys_state *css;
3167	int ret = 0;
3168
3169	lockdep_assert_held(&cgroup_mutex);
3170
3171	/* add/rm files for all cgroups created before */
3172	css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3173	struct cgroup *cgrp = css->cgroup;
3174
3175	if (cgroup_is_dead(cgrp))
3176	continue;
3177
3178	ret = cgroup_addrm_files(cgrp, cfts, is_add);
3179	if (ret)
3180	break;
3181	}
3182
3183	if (is_add && !ret)
3184	kernfs_activate(root->kn);
3185	return ret;
3186	}
3187
3188	static void cgroup_exit_cftypes(struct cftype *cfts)
3189	{
3190	struct cftype *cft;
3191
3192	for (cft = cfts; cft->name[0] != '\0'; cft++) {
3193	/* free copy for custom atomic_write_len, see init_cftypes() */
3194	if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3195	kfree(cft->kf_ops);
3196	cft->kf_ops = NULL;
3197	cft->ss = NULL;
3198
3199	/* revert flags set by cgroup core while adding @cfts */
3200	cft->flags &= ~(__CFTYPE_ONLY_ON_DFL \| __CFTYPE_NOT_ON_DFL);
3201	}
3202	}
3203
3204	static int cgroup_init_cftypes(struct cgroup_subsys ss, struct cftype cfts)
3205	{
3206	struct cftype *cft;
3207
3208	for (cft = cfts; cft->name[0] != '\0'; cft++) {
3209	struct kernfs_ops *kf_ops;
3210
3211	WARN_ON(cft->ss \|\| cft->kf_ops);
3212
3213	if (cft->seq_start)
3214	kf_ops = &cgroup_kf_ops;
3215	else
3216	kf_ops = &cgroup_kf_single_ops;
3217
3218	/*
3219	* Ugh... if @cft wants a custom max_write_len, we need to
3220	* make a copy of kf_ops to set its atomic_write_len.
3221	*/
3222	if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3223	kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3224	if (!kf_ops) {
3225	cgroup_exit_cftypes(cfts);
3226	return -ENOMEM;
3227	}
3228	kf_ops->atomic_write_len = cft->max_write_len;
3229	}
3230
3231	cft->kf_ops = kf_ops;
3232	cft->ss = ss;
3233	}
3234
3235	return 0;
3236	}
3237
3238	static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3239	{
3240	lockdep_assert_held(&cgroup_mutex);
3241
3242	if (!cfts \|\| !cfts[0].ss)
3243	return -ENOENT;
3244
3245	list_del(&cfts->node);
3246	cgroup_apply_cftypes(cfts, false);
3247	cgroup_exit_cftypes(cfts);
3248	return 0;
3249	}
3250
3251	/**
3252	* cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3253	* @cfts: zero-length name terminated array of cftypes
3254	*
3255	* Unregister @cfts. Files described by @cfts are removed from all
3256	* existing cgroups and all future cgroups won't have them either. This
3257	* function can be called anytime whether @cfts' subsys is attached or not.
3258	*
3259	* Returns 0 on successful unregistration, -ENOENT if @cfts is not
3260	* registered.
3261	*/
3262	int cgroup_rm_cftypes(struct cftype *cfts)
3263	{
3264	int ret;
3265
3266	mutex_lock(&cgroup_mutex);
3267	ret = cgroup_rm_cftypes_locked(cfts);
3268	mutex_unlock(&cgroup_mutex);
3269	return ret;
3270	}
3271
3272	/**
3273	* cgroup_add_cftypes - add an array of cftypes to a subsystem
3274	* @ss: target cgroup subsystem
3275	* @cfts: zero-length name terminated array of cftypes
3276	*
3277	* Register @cfts to @ss. Files described by @cfts are created for all
3278	* existing cgroups to which @ss is attached and all future cgroups will
3279	* have them too. This function can be called anytime whether @ss is
3280	* attached or not.
3281	*
3282	* Returns 0 on successful registration, -errno on failure. Note that this
3283	* function currently returns 0 as long as @cfts registration is successful
3284	* even if some file creation attempts on existing cgroups fail.
3285	*/
3286	static int cgroup_add_cftypes(struct cgroup_subsys ss, struct cftype cfts)
3287	{
3288	int ret;
3289
3290	if (ss->disabled)
3291	return 0;
3292
3293	if (!cfts \|\| cfts[0].name[0] == '\0')
3294	return 0;
3295
3296	ret = cgroup_init_cftypes(ss, cfts);
3297	if (ret)
3298	return ret;
3299
3300	mutex_lock(&cgroup_mutex);
3301
3302	list_add_tail(&cfts->node, &ss->cfts);
3303	ret = cgroup_apply_cftypes(cfts, true);
3304	if (ret)
3305	cgroup_rm_cftypes_locked(cfts);
3306
3307	mutex_unlock(&cgroup_mutex);
3308	return ret;
3309	}
3310
3311	/**
3312	* cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3313	* @ss: target cgroup subsystem
3314	* @cfts: zero-length name terminated array of cftypes
3315	*
3316	* Similar to cgroup_add_cftypes() but the added files are only used for
3317	* the default hierarchy.
3318	*/
3319	int cgroup_add_dfl_cftypes(struct cgroup_subsys ss, struct cftype cfts)
3320	{
3321	struct cftype *cft;
3322
3323	for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3324	cft->flags \|= __CFTYPE_ONLY_ON_DFL;
3325	return cgroup_add_cftypes(ss, cfts);
3326	}
3327
3328	/**
3329	* cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3330	* @ss: target cgroup subsystem
3331	* @cfts: zero-length name terminated array of cftypes
3332	*
3333	* Similar to cgroup_add_cftypes() but the added files are only used for
3334	* the legacy hierarchies.
3335	*/
3336	int cgroup_add_legacy_cftypes(struct cgroup_subsys ss, struct cftype cfts)
3337	{
3338	struct cftype *cft;
3339
3340	/*
3341	* If legacy_flies_on_dfl, we want to show the legacy files on the
3342	* dfl hierarchy but iff the target subsystem hasn't been updated
3343	* for the dfl hierarchy yet.
3344	*/
3345	if (!cgroup_legacy_files_on_dfl \|\|
3346	ss->dfl_cftypes != ss->legacy_cftypes) {
3347	for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3348	cft->flags \|= __CFTYPE_NOT_ON_DFL;
3349	}
3350
3351	return cgroup_add_cftypes(ss, cfts);
3352	}
3353
3354	/**
3355	* cgroup_task_count - count the number of tasks in a cgroup.
3356	* @cgrp: the cgroup in question
3357	*
3358	* Return the number of tasks in the cgroup.
3359	*/
3360	static int cgroup_task_count(const struct cgroup *cgrp)
3361	{
3362	int count = 0;
3363	struct cgrp_cset_link *link;
3364
3365	down_read(&css_set_rwsem);
3366	list_for_each_entry(link, &cgrp->cset_links, cset_link)
3367	count += atomic_read(&link->cset->refcount);
3368	up_read(&css_set_rwsem);
3369	return count;
3370	}
3371
3372	/**
3373	* css_next_child - find the next child of a given css
3374	* @pos: the current position (%NULL to initiate traversal)
3375	* @parent: css whose children to walk
3376	*
3377	* This function returns the next child of @parent and should be called
3378	* under either cgroup_mutex or RCU read lock. The only requirement is
3379	* that @parent and @pos are accessible. The next sibling is guaranteed to
3380	* be returned regardless of their states.
3381	*
3382	* If a subsystem synchronizes ->css_online() and the start of iteration, a
3383	* css which finished ->css_online() is guaranteed to be visible in the
3384	* future iterations and will stay visible until the last reference is put.
3385	* A css which hasn't finished ->css_online() or already finished
3386	* ->css_offline() may show up during traversal. It's each subsystem's
3387	* responsibility to synchronize against on/offlining.
3388	*/
3389	struct cgroup_subsys_state css_next_child(struct cgroup_subsys_state pos,
3390	struct cgroup_subsys_state *parent)
3391	{
3392	struct cgroup_subsys_state *next;
3393
3394	cgroup_assert_mutex_or_rcu_locked();
3395
3396	/*
3397	* @pos could already have been unlinked from the sibling list.
3398	* Once a cgroup is removed, its ->sibling.next is no longer
3399	* updated when its next sibling changes. CSS_RELEASED is set when
3400	* @pos is taken off list, at which time its next pointer is valid,
3401	* and, as releases are serialized, the one pointed to by the next
3402	* pointer is guaranteed to not have started release yet. This
3403	* implies that if we observe !CSS_RELEASED on @pos in this RCU
3404	* critical section, the one pointed to by its next pointer is
3405	* guaranteed to not have finished its RCU grace period even if we
3406	* have dropped rcu_read_lock() inbetween iterations.
3407	*
3408	* If @pos has CSS_RELEASED set, its next pointer can't be
3409	* dereferenced; however, as each css is given a monotonically
3410	* increasing unique serial number and always appended to the
3411	* sibling list, the next one can be found by walking the parent's
3412	* children until the first css with higher serial number than
3413	* @pos's. While this path can be slower, it happens iff iteration
3414	* races against release and the race window is very small.
3415	*/
3416	if (!pos) {
3417	next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3418	} else if (likely(!(pos->flags & CSS_RELEASED))) {
3419	next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3420	} else {
3421	list_for_each_entry_rcu(next, &parent->children, sibling)
3422	if (next->serial_nr > pos->serial_nr)
3423	break;
3424	}
3425
3426	/*
3427	* @next, if not pointing to the head, can be dereferenced and is
3428	* the next sibling.
3429	*/
3430	if (&next->sibling != &parent->children)
3431	return next;
3432	return NULL;
3433	}
3434
3435	/**
3436	* css_next_descendant_pre - find the next descendant for pre-order walk
3437	* @pos: the current position (%NULL to initiate traversal)
3438	* @root: css whose descendants to walk
3439	*
3440	* To be used by css_for_each_descendant_pre(). Find the next descendant
3441	* to visit for pre-order traversal of @root's descendants. @root is
3442	* included in the iteration and the first node to be visited.
3443	*
3444	* While this function requires cgroup_mutex or RCU read locking, it
3445	* doesn't require the whole traversal to be contained in a single critical
3446	* section. This function will return the correct next descendant as long
3447	* as both @pos and @root are accessible and @pos is a descendant of @root.
3448	*
3449	* If a subsystem synchronizes ->css_online() and the start of iteration, a
3450	* css which finished ->css_online() is guaranteed to be visible in the
3451	* future iterations and will stay visible until the last reference is put.
3452	* A css which hasn't finished ->css_online() or already finished
3453	* ->css_offline() may show up during traversal. It's each subsystem's
3454	* responsibility to synchronize against on/offlining.
3455	*/
3456	struct cgroup_subsys_state *
3457	css_next_descendant_pre(struct cgroup_subsys_state *pos,
3458	struct cgroup_subsys_state *root)
3459	{
3460	struct cgroup_subsys_state *next;
3461
3462	cgroup_assert_mutex_or_rcu_locked();
3463
3464	/* if first iteration, visit @root */
3465	if (!pos)
3466	return root;
3467
3468	/* visit the first child if exists */
3469	next = css_next_child(NULL, pos);
3470	if (next)
3471	return next;
3472
3473	/* no child, visit my or the closest ancestor's next sibling */
3474	while (pos != root) {
3475	next = css_next_child(pos, pos->parent);
3476	if (next)
3477	return next;
3478	pos = pos->parent;
3479	}
3480
3481	return NULL;
3482	}
3483
3484	/**
3485	* css_rightmost_descendant - return the rightmost descendant of a css
3486	* @pos: css of interest
3487	*
3488	* Return the rightmost descendant of @pos. If there's no descendant, @pos
3489	* is returned. This can be used during pre-order traversal to skip
3490	* subtree of @pos.
3491	*
3492	* While this function requires cgroup_mutex or RCU read locking, it
3493	* doesn't require the whole traversal to be contained in a single critical
3494	* section. This function will return the correct rightmost descendant as
3495	* long as @pos is accessible.
3496	*/
3497	struct cgroup_subsys_state *
3498	css_rightmost_descendant(struct cgroup_subsys_state *pos)
3499	{
3500	struct cgroup_subsys_state last, tmp;
3501
3502	cgroup_assert_mutex_or_rcu_locked();
3503
3504	do {
3505	last = pos;
3506	/* ->prev isn't RCU safe, walk ->next till the end */
3507	pos = NULL;
3508	css_for_each_child(tmp, last)
3509	pos = tmp;
3510	} while (pos);
3511
3512	return last;
3513	}
3514
3515	static struct cgroup_subsys_state *
3516	css_leftmost_descendant(struct cgroup_subsys_state *pos)
3517	{
3518	struct cgroup_subsys_state *last;
3519
3520	do {
3521	last = pos;
3522	pos = css_next_child(NULL, pos);
3523	} while (pos);
3524
3525	return last;
3526	}
3527
3528	/**
3529	* css_next_descendant_post - find the next descendant for post-order walk
3530	* @pos: the current position (%NULL to initiate traversal)
3531	* @root: css whose descendants to walk
3532	*
3533	* To be used by css_for_each_descendant_post(). Find the next descendant
3534	* to visit for post-order traversal of @root's descendants. @root is
3535	* included in the iteration and the last node to be visited.
3536	*
3537	* While this function requires cgroup_mutex or RCU read locking, it
3538	* doesn't require the whole traversal to be contained in a single critical
3539	* section. This function will return the correct next descendant as long
3540	* as both @pos and @cgroup are accessible and @pos is a descendant of
3541	* @cgroup.
3542	*
3543	* If a subsystem synchronizes ->css_online() and the start of iteration, a
3544	* css which finished ->css_online() is guaranteed to be visible in the
3545	* future iterations and will stay visible until the last reference is put.
3546	* A css which hasn't finished ->css_online() or already finished
3547	* ->css_offline() may show up during traversal. It's each subsystem's
3548	* responsibility to synchronize against on/offlining.
3549	*/
3550	struct cgroup_subsys_state *
3551	css_next_descendant_post(struct cgroup_subsys_state *pos,
3552	struct cgroup_subsys_state *root)
3553	{
3554	struct cgroup_subsys_state *next;
3555
3556	cgroup_assert_mutex_or_rcu_locked();
3557
3558	/* if first iteration, visit leftmost descendant which may be @root */
3559	if (!pos)
3560	return css_leftmost_descendant(root);
3561
3562	/* if we visited @root, we're done */
3563	if (pos == root)
3564	return NULL;
3565
3566	/* if there's an unvisited sibling, visit its leftmost descendant */
3567	next = css_next_child(pos, pos->parent);
3568	if (next)
3569	return css_leftmost_descendant(next);
3570
3571	/* no sibling left, visit parent */
3572	return pos->parent;
3573	}
3574
3575	/**
3576	* css_has_online_children - does a css have online children
3577	* @css: the target css
3578	*
3579	* Returns %true if @css has any online children; otherwise, %false. This
3580	* function can be called from any context but the caller is responsible
3581	* for synchronizing against on/offlining as necessary.
3582	*/
3583	bool css_has_online_children(struct cgroup_subsys_state *css)
3584	{
3585	struct cgroup_subsys_state *child;
3586	bool ret = false;
3587
3588	rcu_read_lock();
3589	css_for_each_child(child, css) {
3590	if (child->flags & CSS_ONLINE) {
3591	ret = true;
3592	break;
3593	}
3594	}
3595	rcu_read_unlock();
3596	return ret;
3597	}
3598
3599	/**
3600	* css_advance_task_iter - advance a task itererator to the next css_set
3601	* @it: the iterator to advance
3602	*
3603	* Advance @it to the next css_set to walk.
3604	*/
3605	static void css_advance_task_iter(struct css_task_iter *it)
3606	{
3607	struct list_head *l = it->cset_pos;
3608	struct cgrp_cset_link *link;
3609	struct css_set *cset;
3610
3611	/* Advance to the next non-empty css_set */
3612	do {
3613	l = l->next;
3614	if (l == it->cset_head) {
3615	it->cset_pos = NULL;
3616	return;
3617	}
3618
3619	if (it->ss) {
3620	cset = container_of(l, struct css_set,
3621	e_cset_node[it->ss->id]);
3622	} else {
3623	link = list_entry(l, struct cgrp_cset_link, cset_link);
3624	cset = link->cset;
3625	}
3626	} while (list_empty(&cset->tasks) && list_empty(&cset->mg_tasks));
3627
3628	it->cset_pos = l;
3629
3630	if (!list_empty(&cset->tasks))
3631	it->task_pos = cset->tasks.next;
3632	else
3633	it->task_pos = cset->mg_tasks.next;
3634
3635	it->tasks_head = &cset->tasks;
3636	it->mg_tasks_head = &cset->mg_tasks;
3637	}
3638
3639	/**
3640	* css_task_iter_start - initiate task iteration
3641	* @css: the css to walk tasks of
3642	* @it: the task iterator to use
3643	*
3644	* Initiate iteration through the tasks of @css. The caller can call
3645	* css_task_iter_next() to walk through the tasks until the function
3646	* returns NULL. On completion of iteration, css_task_iter_end() must be
3647	* called.
3648	*
3649	* Note that this function acquires a lock which is released when the
3650	* iteration finishes. The caller can't sleep while iteration is in
3651	* progress.
3652	*/
3653	void css_task_iter_start(struct cgroup_subsys_state *css,
3654	struct css_task_iter *it)
3655	__acquires(css_set_rwsem)
3656	{
3657	/* no one should try to iterate before mounting cgroups */
3658	WARN_ON_ONCE(!use_task_css_set_links);
3659
3660	down_read(&css_set_rwsem);
3661
3662	it->ss = css->ss;
3663
3664	if (it->ss)
3665	it->cset_pos = &css->cgroup->e_csets[css->ss->id];
3666	else
3667	it->cset_pos = &css->cgroup->cset_links;
3668
3669	it->cset_head = it->cset_pos;
3670
3671	css_advance_task_iter(it);
3672	}
3673
3674	/**
3675	* css_task_iter_next - return the next task for the iterator
3676	* @it: the task iterator being iterated
3677	*
3678	* The "next" function for task iteration. @it should have been
3679	* initialized via css_task_iter_start(). Returns NULL when the iteration
3680	* reaches the end.
3681	*/
3682	struct task_struct css_task_iter_next(struct css_task_iter it)
3683	{
3684	struct task_struct *res;
3685	struct list_head *l = it->task_pos;
3686
3687	/* If the iterator cg is NULL, we have no tasks */
3688	if (!it->cset_pos)
3689	return NULL;
3690	res = list_entry(l, struct task_struct, cg_list);
3691
3692	/*
3693	* Advance iterator to find next entry. cset->tasks is consumed
3694	* first and then ->mg_tasks. After ->mg_tasks, we move onto the
3695	* next cset.
3696	*/
3697	l = l->next;
3698
3699	if (l == it->tasks_head)
3700	l = it->mg_tasks_head->next;
3701
3702	if (l == it->mg_tasks_head)
3703	css_advance_task_iter(it);
3704	else
3705	it->task_pos = l;
3706
3707	return res;
3708	}
3709
3710	/**
3711	* css_task_iter_end - finish task iteration
3712	* @it: the task iterator to finish
3713	*
3714	* Finish task iteration started by css_task_iter_start().
3715	*/
3716	void css_task_iter_end(struct css_task_iter *it)
3717	__releases(css_set_rwsem)
3718	{
3719	up_read(&css_set_rwsem);
3720	}
3721
3722	/**
3723	* cgroup_trasnsfer_tasks - move tasks from one cgroup to another
3724	* @to: cgroup to which the tasks will be moved
3725	* @from: cgroup in which the tasks currently reside
3726	*
3727	* Locking rules between cgroup_post_fork() and the migration path
3728	* guarantee that, if a task is forking while being migrated, the new child
3729	* is guaranteed to be either visible in the source cgroup after the
3730	* parent's migration is complete or put into the target cgroup. No task
3731	* can slip out of migration through forking.
3732	*/
3733	int cgroup_transfer_tasks(struct cgroup to, struct cgroup from)
3734	{
3735	LIST_HEAD(preloaded_csets);
3736	struct cgrp_cset_link *link;
3737	struct css_task_iter it;
3738	struct task_struct *task;
3739	int ret;
3740
3741	mutex_lock(&cgroup_mutex);
3742
3743	/* all tasks in @from are being moved, all csets are source */
3744	down_read(&css_set_rwsem);
3745	list_for_each_entry(link, &from->cset_links, cset_link)
3746	cgroup_migrate_add_src(link->cset, to, &preloaded_csets);
3747	up_read(&css_set_rwsem);
3748
3749	ret = cgroup_migrate_prepare_dst(to, &preloaded_csets);
3750	if (ret)
3751	goto out_err;
3752
3753	/*
3754	* Migrate tasks one-by-one until @form is empty. This fails iff
3755	* ->can_attach() fails.
3756	*/
3757	do {
3758	css_task_iter_start(&from->self, &it);
3759	task = css_task_iter_next(&it);
3760	if (task)
3761	get_task_struct(task);
3762	css_task_iter_end(&it);
3763
3764	if (task) {
3765	ret = cgroup_migrate(to, task, false);
3766	put_task_struct(task);
3767	}
3768	} while (task && !ret);
3769	out_err:
3770	cgroup_migrate_finish(&preloaded_csets);
3771	mutex_unlock(&cgroup_mutex);
3772	return ret;
3773	}
3774
3775	/*
3776	* Stuff for reading the 'tasks'/'procs' files.
3777	*
3778	* Reading this file can return large amounts of data if a cgroup has
3779	* lots of attached tasks. So it may need several calls to read(),
3780	* but we cannot guarantee that the information we produce is correct
3781	* unless we produce it entirely atomically.
3782	*
3783	*/
3784
3785	/* which pidlist file are we talking about? */
3786	enum cgroup_filetype {
3787	CGROUP_FILE_PROCS,
3788	CGROUP_FILE_TASKS,
3789	};
3790
3791	/*
3792	* A pidlist is a list of pids that virtually represents the contents of one
3793	* of the cgroup files ("procs" or "tasks"). We keep a list of such pidlists,
3794	* a pair (one each for procs, tasks) for each pid namespace that's relevant
3795	* to the cgroup.
3796	*/
3797	struct cgroup_pidlist {
3798	/*
3799	* used to find which pidlist is wanted. doesn't change as long as
3800	* this particular list stays in the list.
3801	*/
3802	struct { enum cgroup_filetype type; struct pid_namespace *ns; } key;
3803	/* array of xids */
3804	pid_t *list;
3805	/* how many elements the above list has */
3806	int length;
3807	/* each of these stored in a list by its cgroup */
3808	struct list_head links;
3809	/* pointer to the cgroup we belong to, for list removal purposes */
3810	struct cgroup *owner;
3811	/* for delayed destruction */
3812	struct delayed_work destroy_dwork;
3813	};
3814
3815	/*
3816	* The following two functions "fix" the issue where there are more pids
3817	* than kmalloc will give memory for; in such cases, we use vmalloc/vfree.
3818	* TODO: replace with a kernel-wide solution to this problem
3819	*/
3820	#define PIDLIST_TOO_LARGE(c) ((c) * sizeof(pid_t) > (PAGE_SIZE * 2))
3821	static void *pidlist_allocate(int count)
3822	{
3823	if (PIDLIST_TOO_LARGE(count))
3824	return vmalloc(count * sizeof(pid_t));
3825	else
3826	return kmalloc(count * sizeof(pid_t), GFP_KERNEL);
3827	}
3828
3829	static void pidlist_free(void *p)
3830	{
3831	kvfree(p);
3832	}
3833
3834	/*
3835	* Used to destroy all pidlists lingering waiting for destroy timer. None
3836	* should be left afterwards.
3837	*/
3838	static void cgroup_pidlist_destroy_all(struct cgroup *cgrp)
3839	{
3840	struct cgroup_pidlist l, tmp_l;
3841
3842	mutex_lock(&cgrp->pidlist_mutex);
3843	list_for_each_entry_safe(l, tmp_l, &cgrp->pidlists, links)
3844	mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork, 0);
3845	mutex_unlock(&cgrp->pidlist_mutex);
3846
3847	flush_workqueue(cgroup_pidlist_destroy_wq);
3848	BUG_ON(!list_empty(&cgrp->pidlists));
3849	}
3850
3851	static void cgroup_pidlist_destroy_work_fn(struct work_struct *work)
3852	{
3853	struct delayed_work *dwork = to_delayed_work(work);
3854	struct cgroup_pidlist *l = container_of(dwork, struct cgroup_pidlist,
3855	destroy_dwork);
3856	struct cgroup_pidlist *tofree = NULL;
3857
3858	mutex_lock(&l->owner->pidlist_mutex);
3859
3860	/*
3861	* Destroy iff we didn't get queued again. The state won't change
3862	* as destroy_dwork can only be queued while locked.
3863	*/
3864	if (!delayed_work_pending(dwork)) {
3865	list_del(&l->links);
3866	pidlist_free(l->list);
3867	put_pid_ns(l->key.ns);
3868	tofree = l;
3869	}
3870
3871	mutex_unlock(&l->owner->pidlist_mutex);
3872	kfree(tofree);
3873	}
3874
3875	/*
3876	* pidlist_uniq - given a kmalloc()ed list, strip out all duplicate entries
3877	* Returns the number of unique elements.
3878	*/
3879	static int pidlist_uniq(pid_t *list, int length)
3880	{
3881	int src, dest = 1;
3882
3883	/*
3884	* we presume the 0th element is unique, so i starts at 1. trivial
3885	* edge cases first; no work needs to be done for either
3886	*/
3887	if (length == 0 \|\| length == 1)
3888	return length;
3889	/* src and dest walk down the list; dest counts unique elements */
3890	for (src = 1; src < length; src++) {
3891	/* find next unique element */
3892	while (list[src] == list[src-1]) {
3893	src++;
3894	if (src == length)
3895	goto after;
3896	}
3897	/* dest always points to where the next unique element goes */
3898	list[dest] = list[src];
3899	dest++;
3900	}
3901	after:
3902	return dest;
3903	}
3904
3905	/*
3906	* The two pid files - task and cgroup.procs - guaranteed that the result
3907	* is sorted, which forced this whole pidlist fiasco. As pid order is
3908	* different per namespace, each namespace needs differently sorted list,
3909	* making it impossible to use, for example, single rbtree of member tasks
3910	* sorted by task pointer. As pidlists can be fairly large, allocating one
3911	* per open file is dangerous, so cgroup had to implement shared pool of
3912	* pidlists keyed by cgroup and namespace.
3913	*
3914	* All this extra complexity was caused by the original implementation
3915	* committing to an entirely unnecessary property. In the long term, we
3916	* want to do away with it. Explicitly scramble sort order if on the
3917	* default hierarchy so that no such expectation exists in the new
3918	* interface.
3919	*
3920	* Scrambling is done by swapping every two consecutive bits, which is
3921	* non-identity one-to-one mapping which disturbs sort order sufficiently.
3922	*/
3923	static pid_t pid_fry(pid_t pid)
3924	{
3925	unsigned a = pid & 0x55555555;
3926	unsigned b = pid & 0xAAAAAAAA;
3927
3928	return (a << 1) \| (b >> 1);
3929	}
3930
3931	static pid_t cgroup_pid_fry(struct cgroup *cgrp, pid_t pid)
3932	{
3933	if (cgroup_on_dfl(cgrp))
3934	return pid_fry(pid);
3935	else
3936	return pid;
3937	}
3938
3939	static int cmppid(const void a, const void b)
3940	{
3941	return (pid_t )a - (pid_t )b;
3942	}
3943
3944	static int fried_cmppid(const void a, const void b)
3945	{
3946	return pid_fry((pid_t )a) - pid_fry((pid_t )b);
3947	}
3948
3949	static struct cgroup_pidlist cgroup_pidlist_find(struct cgroup cgrp,
3950	enum cgroup_filetype type)
3951	{
3952	struct cgroup_pidlist *l;
3953	/* don't need task_nsproxy() if we're looking at ourself */
3954	struct pid_namespace *ns = task_active_pid_ns(current);
3955
3956	lockdep_assert_held(&cgrp->pidlist_mutex);
3957
3958	list_for_each_entry(l, &cgrp->pidlists, links)
3959	if (l->key.type == type && l->key.ns == ns)
3960	return l;
3961	return NULL;
3962	}
3963
3964	/*
3965	* find the appropriate pidlist for our purpose (given procs vs tasks)
3966	* returns with the lock on that pidlist already held, and takes care
3967	* of the use count, or returns NULL with no locks held if we're out of
3968	* memory.
3969	*/
3970	static struct cgroup_pidlist cgroup_pidlist_find_create(struct cgroup cgrp,
3971	enum cgroup_filetype type)
3972	{
3973	struct cgroup_pidlist *l;
3974
3975	lockdep_assert_held(&cgrp->pidlist_mutex);
3976
3977	l = cgroup_pidlist_find(cgrp, type);
3978	if (l)
3979	return l;
3980
3981	/* entry not found; create a new one */
3982	l = kzalloc(sizeof(struct cgroup_pidlist), GFP_KERNEL);
3983	if (!l)
3984	return l;
3985
3986	INIT_DELAYED_WORK(&l->destroy_dwork, cgroup_pidlist_destroy_work_fn);
3987	l->key.type = type;
3988	/* don't need task_nsproxy() if we're looking at ourself */
3989	l->key.ns = get_pid_ns(task_active_pid_ns(current));
3990	l->owner = cgrp;
3991	list_add(&l->links, &cgrp->pidlists);
3992	return l;
3993	}
3994
3995	/*
3996	* Load a cgroup's pidarray with either procs' tgids or tasks' pids
3997	*/
3998	static int pidlist_array_load(struct cgroup *cgrp, enum cgroup_filetype type,
3999	struct cgroup_pidlist **lp)
4000	{
4001	pid_t *array;
4002	int length;
4003	int pid, n = 0; /* used for populating the array */
4004	struct css_task_iter it;
4005	struct task_struct *tsk;
4006	struct cgroup_pidlist *l;
4007
4008	lockdep_assert_held(&cgrp->pidlist_mutex);
4009
4010	/*
4011	* If cgroup gets more users after we read count, we won't have
4012	* enough space - tough. This race is indistinguishable to the
4013	* caller from the case that the additional cgroup users didn't
4014	* show up until sometime later on.
4015	*/
4016	length = cgroup_task_count(cgrp);
4017	array = pidlist_allocate(length);
4018	if (!array)
4019	return -ENOMEM;
4020	/* now, populate the array */
4021	css_task_iter_start(&cgrp->self, &it);
4022	while ((tsk = css_task_iter_next(&it))) {
4023	if (unlikely(n == length))
4024	break;
4025	/* get tgid or pid for procs or tasks file respectively */
4026	if (type == CGROUP_FILE_PROCS)
4027	pid = task_tgid_vnr(tsk);
4028	else
4029	pid = task_pid_vnr(tsk);
4030	if (pid > 0) /* make sure to only use valid results */
4031	array[n++] = pid;
4032	}
4033	css_task_iter_end(&it);
4034	length = n;
4035	/* now sort & (if procs) strip out duplicates */
4036	if (cgroup_on_dfl(cgrp))
4037	sort(array, length, sizeof(pid_t), fried_cmppid, NULL);
4038	else
4039	sort(array, length, sizeof(pid_t), cmppid, NULL);
4040	if (type == CGROUP_FILE_PROCS)
4041	length = pidlist_uniq(array, length);
4042
4043	l = cgroup_pidlist_find_create(cgrp, type);
4044	if (!l) {
4045	pidlist_free(array);
4046	return -ENOMEM;
4047	}
4048
4049	/* store array, freeing old if necessary */
4050	pidlist_free(l->list);
4051	l->list = array;
4052	l->length = length;
4053	*lp = l;
4054	return 0;
4055	}
4056
4057	/**
4058	* cgroupstats_build - build and fill cgroupstats
4059	* @stats: cgroupstats to fill information into
4060	* @dentry: A dentry entry belonging to the cgroup for which stats have
4061	* been requested.
4062	*
4063	* Build and fill cgroupstats so that taskstats can export it to user
4064	* space.
4065	*/
4066	int cgroupstats_build(struct cgroupstats stats, struct dentry dentry)
4067	{
4068	struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
4069	struct cgroup *cgrp;
4070	struct css_task_iter it;
4071	struct task_struct *tsk;
4072
4073	/* it should be kernfs_node belonging to cgroupfs and is a directory */
4074	if (dentry->d_sb->s_type != &cgroup_fs_type \|\| !kn \|\|
4075	kernfs_type(kn) != KERNFS_DIR)
4076	return -EINVAL;
4077
4078	mutex_lock(&cgroup_mutex);
4079
4080	/*
4081	* We aren't being called from kernfs and there's no guarantee on
4082	* @kn->priv's validity. For this and css_tryget_online_from_dir(),
4083	* @kn->priv is RCU safe. Let's do the RCU dancing.
4084	*/
4085	rcu_read_lock();
4086	cgrp = rcu_dereference(kn->priv);
4087	if (!cgrp \|\| cgroup_is_dead(cgrp)) {
4088	rcu_read_unlock();
4089	mutex_unlock(&cgroup_mutex);
4090	return -ENOENT;
4091	}
4092	rcu_read_unlock();
4093
4094	css_task_iter_start(&cgrp->self, &it);
4095	while ((tsk = css_task_iter_next(&it))) {
4096	switch (tsk->state) {
4097	case TASK_RUNNING:
4098	stats->nr_running++;
4099	break;
4100	case TASK_INTERRUPTIBLE:
4101	stats->nr_sleeping++;
4102	break;
4103	case TASK_UNINTERRUPTIBLE:
4104	stats->nr_uninterruptible++;
4105	break;
4106	case TASK_STOPPED:
4107	stats->nr_stopped++;
4108	break;
4109	default:
4110	if (delayacct_is_task_waiting_on_io(tsk))
4111	stats->nr_io_wait++;
4112	break;
4113	}
4114	}
4115	css_task_iter_end(&it);
4116
4117	mutex_unlock(&cgroup_mutex);
4118	return 0;
4119	}
4120
4121
4122	/*
4123	* seq_file methods for the tasks/procs files. The seq_file position is the
4124	* next pid to display; the seq_file iterator is a pointer to the pid
4125	* in the cgroup->l->list array.
4126	*/
4127
4128	static void cgroup_pidlist_start(struct seq_file s, loff_t *pos)
4129	{
4130	/*
4131	* Initially we receive a position value that corresponds to
4132	* one more than the last pid shown (or 0 on the first call or
4133	* after a seek to the start). Use a binary-search to find the
4134	* next pid to display, if any
4135	*/
4136	struct kernfs_open_file *of = s->private;
4137	struct cgroup *cgrp = seq_css(s)->cgroup;
4138	struct cgroup_pidlist *l;
4139	enum cgroup_filetype type = seq_cft(s)->private;
4140	int index = 0, pid = *pos;
4141	int *iter, ret;
4142
4143	mutex_lock(&cgrp->pidlist_mutex);
4144
4145	/*
4146	* !NULL @of->priv indicates that this isn't the first start()
4147	* after open. If the matching pidlist is around, we can use that.
4148	* Look for it. Note that @of->priv can't be used directly. It
4149	* could already have been destroyed.
4150	*/
4151	if (of->priv)
4152	of->priv = cgroup_pidlist_find(cgrp, type);
4153
4154	/*
4155	* Either this is the first start() after open or the matching
4156	* pidlist has been destroyed inbetween. Create a new one.
4157	*/
4158	if (!of->priv) {
4159	ret = pidlist_array_load(cgrp, type,
4160	(struct cgroup_pidlist **)&of->priv);
4161	if (ret)
4162	return ERR_PTR(ret);
4163	}
4164	l = of->priv;
4165
4166	if (pid) {
4167	int end = l->length;
4168
4169	while (index < end) {
4170	int mid = (index + end) / 2;
4171	if (cgroup_pid_fry(cgrp, l->list[mid]) == pid) {
4172	index = mid;
4173	break;
4174	} else if (cgroup_pid_fry(cgrp, l->list[mid]) <= pid)
4175	index = mid + 1;
4176	else
4177	end = mid;
4178	}
4179	}
4180	/* If we're off the end of the array, we're done */
4181	if (index >= l->length)
4182	return NULL;
4183	/* Update the abstract position to be the actual pid that we found */
4184	iter = l->list + index;
4185	pos = cgroup_pid_fry(cgrp, iter);
4186	return iter;
4187	}
4188
4189	static void cgroup_pidlist_stop(struct seq_file s, void v)
4190	{
4191	struct kernfs_open_file *of = s->private;
4192	struct cgroup_pidlist *l = of->priv;
4193
4194	if (l)
4195	mod_delayed_work(cgroup_pidlist_destroy_wq, &l->destroy_dwork,
4196	CGROUP_PIDLIST_DESTROY_DELAY);
4197	mutex_unlock(&seq_css(s)->cgroup->pidlist_mutex);
4198	}
4199
4200	static void cgroup_pidlist_next(struct seq_file s, void v, loff_t pos)
4201	{
4202	struct kernfs_open_file *of = s->private;
4203	struct cgroup_pidlist *l = of->priv;
4204	pid_t *p = v;
4205	pid_t *end = l->list + l->length;
4206	/*
4207	* Advance to the next pid in the array. If this goes off the
4208	* end, we're done
4209	*/
4210	p++;
4211	if (p >= end) {
4212	return NULL;
4213	} else {
4214	pos = cgroup_pid_fry(seq_css(s)->cgroup, p);
4215	return p;
4216	}
4217	}
4218
4219	static int cgroup_pidlist_show(struct seq_file s, void v)
4220	{
4221	seq_printf(s, "%d\n", (int )v);
4222
4223	return 0;
4224	}
4225
4226	static u64 cgroup_read_notify_on_release(struct cgroup_subsys_state *css,
4227	struct cftype *cft)
4228	{
4229	return notify_on_release(css->cgroup);
4230	}
4231
4232	static int cgroup_write_notify_on_release(struct cgroup_subsys_state *css,
4233	struct cftype *cft, u64 val)
4234	{
4235	if (val)
4236	set_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4237	else
4238	clear_bit(CGRP_NOTIFY_ON_RELEASE, &css->cgroup->flags);
4239	return 0;
4240	}
4241
4242	static u64 cgroup_clone_children_read(struct cgroup_subsys_state *css,
4243	struct cftype *cft)
4244	{
4245	return test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4246	}
4247
4248	static int cgroup_clone_children_write(struct cgroup_subsys_state *css,
4249	struct cftype *cft, u64 val)
4250	{
4251	if (val)
4252	set_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4253	else
4254	clear_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags);
4255	return 0;
4256	}
4257
4258	/* cgroup core interface files for the default hierarchy */
4259	static struct cftype cgroup_dfl_base_files[] = {
4260	{
4261	.name = "cgroup.procs",
4262	.seq_start = cgroup_pidlist_start,
4263	.seq_next = cgroup_pidlist_next,
4264	.seq_stop = cgroup_pidlist_stop,
4265	.seq_show = cgroup_pidlist_show,
4266	.private = CGROUP_FILE_PROCS,
4267	.write = cgroup_procs_write,
4268	.mode = S_IRUGO \| S_IWUSR,
4269	},
4270	{
4271	.name = "cgroup.controllers",
4272	.flags = CFTYPE_ONLY_ON_ROOT,
4273	.seq_show = cgroup_root_controllers_show,
4274	},
4275	{
4276	.name = "cgroup.controllers",
4277	.flags = CFTYPE_NOT_ON_ROOT,
4278	.seq_show = cgroup_controllers_show,
4279	},
4280	{
4281	.name = "cgroup.subtree_control",
4282	.seq_show = cgroup_subtree_control_show,
4283	.write = cgroup_subtree_control_write,
4284	},
4285	{
4286	.name = "cgroup.populated",
4287	.flags = CFTYPE_NOT_ON_ROOT,
4288	.seq_show = cgroup_populated_show,
4289	},
4290	{ } /* terminate */
4291	};
4292
4293	/* cgroup core interface files for the legacy hierarchies */
4294	static struct cftype cgroup_legacy_base_files[] = {
4295	{
4296	.name = "cgroup.procs",
4297	.seq_start = cgroup_pidlist_start,
4298	.seq_next = cgroup_pidlist_next,
4299	.seq_stop = cgroup_pidlist_stop,
4300	.seq_show = cgroup_pidlist_show,
4301	.private = CGROUP_FILE_PROCS,
4302	.write = cgroup_procs_write,
4303	.mode = S_IRUGO \| S_IWUSR,
4304	},
4305	{
4306	.name = "cgroup.clone_children",
4307	.read_u64 = cgroup_clone_children_read,
4308	.write_u64 = cgroup_clone_children_write,
4309	},
4310	{
4311	.name = "cgroup.sane_behavior",
4312	.flags = CFTYPE_ONLY_ON_ROOT,
4313	.seq_show = cgroup_sane_behavior_show,
4314	},
4315	{
4316	.name = "tasks",
4317	.seq_start = cgroup_pidlist_start,
4318	.seq_next = cgroup_pidlist_next,
4319	.seq_stop = cgroup_pidlist_stop,
4320	.seq_show = cgroup_pidlist_show,
4321	.private = CGROUP_FILE_TASKS,
4322	.write = cgroup_tasks_write,
4323	.mode = S_IRUGO \| S_IWUSR,
4324	},
4325	{
4326	.name = "notify_on_release",
4327	.read_u64 = cgroup_read_notify_on_release,
4328	.write_u64 = cgroup_write_notify_on_release,
4329	},
4330	{
4331	.name = "release_agent",
4332	.flags = CFTYPE_ONLY_ON_ROOT,
4333	.seq_show = cgroup_release_agent_show,
4334	.write = cgroup_release_agent_write,
4335	.max_write_len = PATH_MAX - 1,
4336	},
4337	{ } /* terminate */
4338	};
4339
4340	/**
4341	* cgroup_populate_dir - create subsys files in a cgroup directory
4342	* @cgrp: target cgroup
4343	* @subsys_mask: mask of the subsystem ids whose files should be added
4344	*
4345	* On failure, no file is added.
4346	*/
4347	static int cgroup_populate_dir(struct cgroup *cgrp, unsigned long subsys_mask)
4348	{
4349	struct cgroup_subsys *ss;
4350	int i, ret = 0;
4351
4352	/* process cftsets of each subsystem */
4353	for_each_subsys(ss, i) {
4354	struct cftype *cfts;
4355
4356	if (!(subsys_mask & (1 << i)))
4357	continue;
4358
4359	list_for_each_entry(cfts, &ss->cfts, node) {
4360	ret = cgroup_addrm_files(cgrp, cfts, true);
4361	if (ret < 0)
4362	goto err;
4363	}
4364	}
4365	return 0;
4366	err:
4367	cgroup_clear_dir(cgrp, subsys_mask);
4368	return ret;
4369	}
4370
4371	/*
4372	* css destruction is four-stage process.
4373	*
4374	* 1. Destruction starts. Killing of the percpu_ref is initiated.
4375	* Implemented in kill_css().
4376	*
4377	* 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4378	* and thus css_tryget_online() is guaranteed to fail, the css can be
4379	* offlined by invoking offline_css(). After offlining, the base ref is
4380	* put. Implemented in css_killed_work_fn().
4381	*
4382	* 3. When the percpu_ref reaches zero, the only possible remaining
4383	* accessors are inside RCU read sections. css_release() schedules the
4384	* RCU callback.
4385	*
4386	* 4. After the grace period, the css can be freed. Implemented in
4387	* css_free_work_fn().
4388	*
4389	* It is actually hairier because both step 2 and 4 require process context
4390	* and thus involve punting to css->destroy_work adding two additional
4391	* steps to the already complex sequence.
4392	*/
4393	static void css_free_work_fn(struct work_struct *work)
4394	{
4395	struct cgroup_subsys_state *css =
4396	container_of(work, struct cgroup_subsys_state, destroy_work);
4397	struct cgroup_subsys *ss = css->ss;
4398	struct cgroup *cgrp = css->cgroup;
4399
4400	percpu_ref_exit(&css->refcnt);
4401
4402	if (ss) {
4403	/* css free path */
4404	int id = css->id;
4405
4406	if (css->parent)
4407	css_put(css->parent);
4408
4409	ss->css_free(css);
4410	cgroup_idr_remove(&ss->css_idr, id);
4411	cgroup_put(cgrp);
4412	} else {
4413	/* cgroup free path */
4414	atomic_dec(&cgrp->root->nr_cgrps);
4415	cgroup_pidlist_destroy_all(cgrp);
4416	cancel_work_sync(&cgrp->release_agent_work);
4417
4418	if (cgroup_parent(cgrp)) {
4419	/*
4420	* We get a ref to the parent, and put the ref when
4421	* this cgroup is being freed, so it's guaranteed
4422	* that the parent won't be destroyed before its
4423	* children.
4424	*/
4425	cgroup_put(cgroup_parent(cgrp));
4426	kernfs_put(cgrp->kn);
4427	kfree(cgrp);
4428	} else {
4429	/*
4430	* This is root cgroup's refcnt reaching zero,
4431	* which indicates that the root should be
4432	* released.
4433	*/
4434	cgroup_destroy_root(cgrp->root);
4435	}
4436	}
4437	}
4438
4439	static void css_free_rcu_fn(struct rcu_head *rcu_head)
4440	{
4441	struct cgroup_subsys_state *css =
4442	container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4443
4444	INIT_WORK(&css->destroy_work, css_free_work_fn);
4445	queue_work(cgroup_destroy_wq, &css->destroy_work);
4446	}
4447
4448	static void css_release_work_fn(struct work_struct *work)
4449	{
4450	struct cgroup_subsys_state *css =
4451	container_of(work, struct cgroup_subsys_state, destroy_work);
4452	struct cgroup_subsys *ss = css->ss;
4453	struct cgroup *cgrp = css->cgroup;
4454
4455	mutex_lock(&cgroup_mutex);
4456
4457	css->flags \|= CSS_RELEASED;
4458	list_del_rcu(&css->sibling);
4459
4460	if (ss) {
4461	/* css release path */
4462	cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4463	if (ss->css_released)
4464	ss->css_released(css);
4465	} else {
4466	/* cgroup release path */
4467	cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4468	cgrp->id = -1;
4469
4470	/*
4471	* There are two control paths which try to determine
4472	* cgroup from dentry without going through kernfs -
4473	* cgroupstats_build() and css_tryget_online_from_dir().
4474	* Those are supported by RCU protecting clearing of
4475	* cgrp->kn->priv backpointer.
4476	*/
4477	RCU_INIT_POINTER((void __rcu __force *)&cgrp->kn->priv, NULL);
4478	}
4479
4480	mutex_unlock(&cgroup_mutex);
4481
4482	call_rcu(&css->rcu_head, css_free_rcu_fn);
4483	}
4484
4485	static void css_release(struct percpu_ref *ref)
4486	{
4487	struct cgroup_subsys_state *css =
4488	container_of(ref, struct cgroup_subsys_state, refcnt);
4489
4490	INIT_WORK(&css->destroy_work, css_release_work_fn);
4491	queue_work(cgroup_destroy_wq, &css->destroy_work);
4492	}
4493
4494	static void init_and_link_css(struct cgroup_subsys_state *css,
4495	struct cgroup_subsys ss, struct cgroup cgrp)
4496	{
4497	lockdep_assert_held(&cgroup_mutex);
4498
4499	cgroup_get(cgrp);
4500
4501	memset(css, 0, sizeof(*css));
4502	css->cgroup = cgrp;
4503	css->ss = ss;
4504	INIT_LIST_HEAD(&css->sibling);
4505	INIT_LIST_HEAD(&css->children);
4506	css->serial_nr = css_serial_nr_next++;
4507
4508	if (cgroup_parent(cgrp)) {
4509	css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4510	css_get(css->parent);
4511	}
4512
4513	BUG_ON(cgroup_css(cgrp, ss));
4514	}
4515
4516	/* invoke ->css_online() on a new CSS and mark it online if successful */
4517	static int online_css(struct cgroup_subsys_state *css)
4518	{
4519	struct cgroup_subsys *ss = css->ss;
4520	int ret = 0;
4521
4522	lockdep_assert_held(&cgroup_mutex);
4523
4524	if (ss->css_online)
4525	ret = ss->css_online(css);
4526	if (!ret) {
4527	css->flags \|= CSS_ONLINE;
4528	rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4529	}
4530	return ret;
4531	}
4532
4533	/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4534	static void offline_css(struct cgroup_subsys_state *css)
4535	{
4536	struct cgroup_subsys *ss = css->ss;
4537
4538	lockdep_assert_held(&cgroup_mutex);
4539
4540	if (!(css->flags & CSS_ONLINE))
4541	return;
4542
4543	if (ss->css_offline)
4544	ss->css_offline(css);
4545
4546	css->flags &= ~CSS_ONLINE;
4547	RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4548
4549	wake_up_all(&css->cgroup->offline_waitq);
4550	}
4551
4552	/**
4553	* create_css - create a cgroup_subsys_state
4554	* @cgrp: the cgroup new css will be associated with
4555	* @ss: the subsys of new css
4556	* @visible: whether to create control knobs for the new css or not
4557	*
4558	* Create a new css associated with @cgrp - @ss pair. On success, the new
4559	* css is online and installed in @cgrp with all interface files created if
4560	* @visible. Returns 0 on success, -errno on failure.
4561	*/
4562	static int create_css(struct cgroup cgrp, struct cgroup_subsys ss,
4563	bool visible)
4564	{
4565	struct cgroup *parent = cgroup_parent(cgrp);
4566	struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4567	struct cgroup_subsys_state *css;
4568	int err;
4569
4570	lockdep_assert_held(&cgroup_mutex);
4571
4572	css = ss->css_alloc(parent_css);
4573	if (IS_ERR(css))
4574	return PTR_ERR(css);
4575
4576	init_and_link_css(css, ss, cgrp);
4577
4578	err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4579	if (err)
4580	goto err_free_css;
4581
4582	err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_NOWAIT);
4583	if (err < 0)
4584	goto err_free_percpu_ref;
4585	css->id = err;
4586
4587	if (visible) {
4588	err = cgroup_populate_dir(cgrp, 1 << ss->id);
4589	if (err)
4590	goto err_free_id;
4591	}
4592
4593	/* @css is ready to be brought online now, make it visible */
4594	list_add_tail_rcu(&css->sibling, &parent_css->children);
4595	cgroup_idr_replace(&ss->css_idr, css, css->id);
4596
4597	err = online_css(css);
4598	if (err)
4599	goto err_list_del;
4600
4601	if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4602	cgroup_parent(parent)) {
4603	pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4604	current->comm, current->pid, ss->name);
4605	if (!strcmp(ss->name, "memory"))
4606	pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4607	ss->warned_broken_hierarchy = true;
4608	}
4609
4610	return 0;
4611
4612	err_list_del:
4613	list_del_rcu(&css->sibling);
4614	cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4615	err_free_id:
4616	cgroup_idr_remove(&ss->css_idr, css->id);
4617	err_free_percpu_ref:
4618	percpu_ref_exit(&css->refcnt);
4619	err_free_css:
4620	call_rcu(&css->rcu_head, css_free_rcu_fn);
4621	return err;
4622	}
4623
4624	static int cgroup_mkdir(struct kernfs_node parent_kn, const char name,
4625	umode_t mode)
4626	{
4627	struct cgroup parent, cgrp;
4628	struct cgroup_root *root;
4629	struct cgroup_subsys *ss;
4630	struct kernfs_node *kn;
4631	struct cftype *base_files;
4632	int ssid, ret;
4633
4634	/* Do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable.
4635	*/
4636	if (strchr(name, '\n'))
4637	return -EINVAL;
4638
4639	parent = cgroup_kn_lock_live(parent_kn);
4640	if (!parent)
4641	return -ENODEV;
4642	root = parent->root;
4643
4644	/* allocate the cgroup and its ID, 0 is reserved for the root */
4645	cgrp = kzalloc(sizeof(*cgrp), GFP_KERNEL);
4646	if (!cgrp) {
4647	ret = -ENOMEM;
4648	goto out_unlock;
4649	}
4650
4651	ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4652	if (ret)
4653	goto out_free_cgrp;
4654
4655	/*
4656	* Temporarily set the pointer to NULL, so idr_find() won't return
4657	* a half-baked cgroup.
4658	*/
4659	cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_NOWAIT);
4660	if (cgrp->id < 0) {
4661	ret = -ENOMEM;
4662	goto out_cancel_ref;
4663	}
4664
4665	init_cgroup_housekeeping(cgrp);
4666
4667	cgrp->self.parent = &parent->self;
4668	cgrp->root = root;
4669
4670	if (notify_on_release(parent))
4671	set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4672
4673	if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4674	set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4675
4676	/* create the directory */
4677	kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4678	if (IS_ERR(kn)) {
4679	ret = PTR_ERR(kn);
4680	goto out_free_id;
4681	}
4682	cgrp->kn = kn;
4683
4684	/*
4685	* This extra ref will be put in cgroup_free_fn() and guarantees
4686	* that @cgrp->kn is always accessible.
4687	*/
4688	kernfs_get(kn);
4689
4690	cgrp->self.serial_nr = css_serial_nr_next++;
4691
4692	/* allocation complete, commit to creation */
4693	list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4694	atomic_inc(&root->nr_cgrps);
4695	cgroup_get(parent);
4696
4697	/*
4698	* @cgrp is now fully operational. If something fails after this
4699	* point, it'll be released via the normal destruction path.
4700	*/
4701	cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4702
4703	ret = cgroup_kn_set_ugid(kn);
4704	if (ret)
4705	goto out_destroy;
4706
4707	if (cgroup_on_dfl(cgrp))
4708	base_files = cgroup_dfl_base_files;
4709	else
4710	base_files = cgroup_legacy_base_files;
4711
4712	ret = cgroup_addrm_files(cgrp, base_files, true);
4713	if (ret)
4714	goto out_destroy;
4715
4716	/* let's create and online css's */
4717	for_each_subsys(ss, ssid) {
4718	if (parent->child_subsys_mask & (1 << ssid)) {
4719	ret = create_css(cgrp, ss,
4720	parent->subtree_control & (1 << ssid));
4721	if (ret)
4722	goto out_destroy;
4723	}
4724	}
4725
4726	/*
4727	* On the default hierarchy, a child doesn't automatically inherit
4728	* subtree_control from the parent. Each is configured manually.
4729	*/
4730	if (!cgroup_on_dfl(cgrp)) {
4731	cgrp->subtree_control = parent->subtree_control;
4732	cgroup_refresh_child_subsys_mask(cgrp);
4733	}
4734
4735	kernfs_activate(kn);
4736
4737	ret = 0;
4738	goto out_unlock;
4739
4740	out_free_id:
4741	cgroup_idr_remove(&root->cgroup_idr, cgrp->id);
4742	out_cancel_ref:
4743	percpu_ref_exit(&cgrp->self.refcnt);
4744	out_free_cgrp:
4745	kfree(cgrp);
4746	out_unlock:
4747	cgroup_kn_unlock(parent_kn);
4748	return ret;
4749
4750	out_destroy:
4751	cgroup_destroy_locked(cgrp);
4752	goto out_unlock;
4753	}
4754
4755	/*
4756	* This is called when the refcnt of a css is confirmed to be killed.
4757	* css_tryget_online() is now guaranteed to fail. Tell the subsystem to
4758	* initate destruction and put the css ref from kill_css().
4759	*/
4760	static void css_killed_work_fn(struct work_struct *work)
4761	{
4762	struct cgroup_subsys_state *css =
4763	container_of(work, struct cgroup_subsys_state, destroy_work);
4764
4765	mutex_lock(&cgroup_mutex);
4766	offline_css(css);
4767	mutex_unlock(&cgroup_mutex);
4768
4769	css_put(css);
4770	}
4771
4772	/* css kill confirmation processing requires process context, bounce */
4773	static void css_killed_ref_fn(struct percpu_ref *ref)
4774	{
4775	struct cgroup_subsys_state *css =
4776	container_of(ref, struct cgroup_subsys_state, refcnt);
4777
4778	INIT_WORK(&css->destroy_work, css_killed_work_fn);
4779	queue_work(cgroup_destroy_wq, &css->destroy_work);
4780	}
4781
4782	/**
4783	* kill_css - destroy a css
4784	* @css: css to destroy
4785	*
4786	* This function initiates destruction of @css by removing cgroup interface
4787	* files and putting its base reference. ->css_offline() will be invoked
4788	* asynchronously once css_tryget_online() is guaranteed to fail and when
4789	* the reference count reaches zero, @css will be released.
4790	*/
4791	static void kill_css(struct cgroup_subsys_state *css)
4792	{
4793	lockdep_assert_held(&cgroup_mutex);
4794
4795	/*
4796	* This must happen before css is disassociated with its cgroup.
4797	* See seq_css() for details.
4798	*/
4799	cgroup_clear_dir(css->cgroup, 1 << css->ss->id);
4800
4801	/*
4802	* Killing would put the base ref, but we need to keep it alive
4803	* until after ->css_offline().
4804	*/
4805	css_get(css);
4806
4807	/*
4808	* cgroup core guarantees that, by the time ->css_offline() is
4809	* invoked, no new css reference will be given out via
4810	* css_tryget_online(). We can't simply call percpu_ref_kill() and
4811	* proceed to offlining css's because percpu_ref_kill() doesn't
4812	* guarantee that the ref is seen as killed on all CPUs on return.
4813	*
4814	* Use percpu_ref_kill_and_confirm() to get notifications as each
4815	* css is confirmed to be seen as killed on all CPUs.
4816	*/
4817	percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4818	}
4819
4820	/**
4821	* cgroup_destroy_locked - the first stage of cgroup destruction
4822	* @cgrp: cgroup to be destroyed
4823	*
4824	* css's make use of percpu refcnts whose killing latency shouldn't be
4825	* exposed to userland and are RCU protected. Also, cgroup core needs to
4826	* guarantee that css_tryget_online() won't succeed by the time
4827	* ->css_offline() is invoked. To satisfy all the requirements,
4828	* destruction is implemented in the following two steps.
4829	*
4830	* s1. Verify @cgrp can be destroyed and mark it dying. Remove all
4831	* userland visible parts and start killing the percpu refcnts of
4832	* css's. Set up so that the next stage will be kicked off once all
4833	* the percpu refcnts are confirmed to be killed.
4834	*
4835	* s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4836	* rest of destruction. Once all cgroup references are gone, the
4837	* cgroup is RCU-freed.
4838	*
4839	* This function implements s1. After this step, @cgrp is gone as far as
4840	* the userland is concerned and a new cgroup with the same name may be
4841	* created. As cgroup doesn't care about the names internally, this
4842	* doesn't cause any problem.
4843	*/
4844	static int cgroup_destroy_locked(struct cgroup *cgrp)
4845	__releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4846	{
4847	struct cgroup_subsys_state *css;
4848	bool empty;
4849	int ssid;
4850
4851	lockdep_assert_held(&cgroup_mutex);
4852
4853	/*
4854	* css_set_rwsem synchronizes access to ->cset_links and prevents
4855	* @cgrp from being removed while put_css_set() is in progress.
4856	*/
4857	down_read(&css_set_rwsem);
4858	empty = list_empty(&cgrp->cset_links);
4859	up_read(&css_set_rwsem);
4860	if (!empty)
4861	return -EBUSY;
4862
4863	/*
4864	* Make sure there's no live children. We can't test emptiness of
4865	* ->self.children as dead children linger on it while being
4866	* drained; otherwise, "rmdir parent/child parent" may fail.
4867	*/
4868	if (css_has_online_children(&cgrp->self))
4869	return -EBUSY;
4870
4871	/*
4872	* Mark @cgrp dead. This prevents further task migration and child
4873	* creation by disabling cgroup_lock_live_group().
4874	*/
4875	cgrp->self.flags &= ~CSS_ONLINE;
4876
4877	/* initiate massacre of all css's */
4878	for_each_css(css, ssid, cgrp)
4879	kill_css(css);
4880
4881	/*
4882	* Remove @cgrp directory along with the base files. @cgrp has an
4883	* extra ref on its kn.
4884	*/
4885	kernfs_remove(cgrp->kn);
4886
4887	check_for_release(cgroup_parent(cgrp));
4888
4889	/* put the base reference */
4890	percpu_ref_kill(&cgrp->self.refcnt);
4891
4892	return 0;
4893	};
4894
4895	static int cgroup_rmdir(struct kernfs_node *kn)
4896	{
4897	struct cgroup *cgrp;
4898	int ret = 0;
4899
4900	cgrp = cgroup_kn_lock_live(kn);
4901	if (!cgrp)
4902	return 0;
4903
4904	ret = cgroup_destroy_locked(cgrp);
4905
4906	cgroup_kn_unlock(kn);
4907	return ret;
4908	}
4909
4910	static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
4911	.remount_fs = cgroup_remount,
4912	.show_options = cgroup_show_options,
4913	.mkdir = cgroup_mkdir,
4914	.rmdir = cgroup_rmdir,
4915	.rename = cgroup_rename,
4916	};
4917
4918	static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
4919	{
4920	struct cgroup_subsys_state *css;
4921
4922	printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
4923
4924	mutex_lock(&cgroup_mutex);
4925
4926	idr_init(&ss->css_idr);
4927	INIT_LIST_HEAD(&ss->cfts);
4928
4929	/* Create the root cgroup state for this subsystem */
4930	ss->root = &cgrp_dfl_root;
4931	css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
4932	/* We don't handle early failures gracefully */
4933	BUG_ON(IS_ERR(css));
4934	init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
4935
4936	/*
4937	* Root csses are never destroyed and we can't initialize
4938	* percpu_ref during early init. Disable refcnting.
4939	*/
4940	css->flags \|= CSS_NO_REF;
4941
4942	if (early) {
4943	/* allocation can't be done safely during early init */
4944	css->id = 1;
4945	} else {
4946	css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
4947	BUG_ON(css->id < 0);
4948	}
4949
4950	/* Update the init_css_set to contain a subsys
4951	* pointer to this state - since the subsystem is
4952	* newly registered, all tasks and hence the
4953	* init_css_set is in the subsystem's root cgroup. */
4954	init_css_set.subsys[ss->id] = css;
4955
4956	have_fork_callback \|= (bool)ss->fork << ss->id;
4957	have_exit_callback \|= (bool)ss->exit << ss->id;
4958
4959	/* At system boot, before all subsystems have been
4960	* registered, no tasks have been forked, so we don't
4961	* need to invoke fork callbacks here. */
4962	BUG_ON(!list_empty(&init_task.tasks));
4963
4964	BUG_ON(online_css(css));
4965
4966	mutex_unlock(&cgroup_mutex);
4967	}
4968
4969	/**
4970	* cgroup_init_early - cgroup initialization at system boot
4971	*
4972	* Initialize cgroups at system boot, and initialize any
4973	* subsystems that request early init.
4974	*/
4975	int __init cgroup_init_early(void)
4976	{
4977	static struct cgroup_sb_opts __initdata opts;
4978	struct cgroup_subsys *ss;
4979	int i;
4980
4981	init_cgroup_root(&cgrp_dfl_root, &opts);
4982	cgrp_dfl_root.cgrp.self.flags \|= CSS_NO_REF;
4983
4984	RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
4985
4986	for_each_subsys(ss, i) {
4987	WARN(!ss->css_alloc \|\| !ss->css_free \|\| ss->name \|\| ss->id,
4988	"invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p name:id=%d:%s\n",
4989	i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
4990	ss->id, ss->name);
4991	WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
4992	"cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
4993
4994	ss->id = i;
4995	ss->name = cgroup_subsys_name[i];
4996
4997	if (ss->early_init)
4998	cgroup_init_subsys(ss, true);
4999	}
5000	return 0;
5001	}
5002
5003	/**
5004	* cgroup_init - cgroup initialization
5005	*
5006	* Register cgroup filesystem and /proc file, and initialize
5007	* any subsystems that didn't request early init.
5008	*/
5009	int __init cgroup_init(void)
5010	{
5011	struct cgroup_subsys *ss;
5012	unsigned long key;
5013	int ssid, err;
5014
5015	BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5016	BUG_ON(cgroup_init_cftypes(NULL, cgroup_dfl_base_files));
5017	BUG_ON(cgroup_init_cftypes(NULL, cgroup_legacy_base_files));
5018
5019	mutex_lock(&cgroup_mutex);
5020
5021	/* Add init_css_set to the hash table */
5022	key = css_set_hash(init_css_set.subsys);
5023	hash_add(css_set_table, &init_css_set.hlist, key);
5024
5025	BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0));
5026
5027	mutex_unlock(&cgroup_mutex);
5028
5029	for_each_subsys(ss, ssid) {
5030	if (ss->early_init) {
5031	struct cgroup_subsys_state *css =
5032	init_css_set.subsys[ss->id];
5033
5034	css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5035	GFP_KERNEL);
5036	BUG_ON(css->id < 0);
5037	} else {
5038	cgroup_init_subsys(ss, false);
5039	}
5040
5041	list_add_tail(&init_css_set.e_cset_node[ssid],
5042	&cgrp_dfl_root.cgrp.e_csets[ssid]);
5043
5044	/*
5045	* Setting dfl_root subsys_mask needs to consider the
5046	* disabled flag and cftype registration needs kmalloc,
5047	* both of which aren't available during early_init.
5048	*/
5049	if (ss->disabled)
5050	continue;
5051
5052	cgrp_dfl_root.subsys_mask \|= 1 << ss->id;
5053
5054	if (cgroup_legacy_files_on_dfl && !ss->dfl_cftypes)
5055	ss->dfl_cftypes = ss->legacy_cftypes;
5056
5057	if (!ss->dfl_cftypes)
5058	cgrp_dfl_root_inhibit_ss_mask \|= 1 << ss->id;
5059
5060	if (ss->dfl_cftypes == ss->legacy_cftypes) {
5061	WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5062	} else {
5063	WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5064	WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5065	}
5066
5067	if (ss->bind)
5068	ss->bind(init_css_set.subsys[ssid]);
5069	}
5070
5071	err = sysfs_create_mount_point(fs_kobj, "cgroup");
5072	if (err)
5073	return err;
5074
5075	err = register_filesystem(&cgroup_fs_type);
5076	if (err < 0) {
5077	sysfs_remove_mount_point(fs_kobj, "cgroup");
5078	return err;
5079	}
5080
5081	proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations);
5082	return 0;
5083	}
5084
5085	static int __init cgroup_wq_init(void)
5086	{
5087	/*
5088	* There isn't much point in executing destruction path in
5089	* parallel. Good chunk is serialized with cgroup_mutex anyway.
5090	* Use 1 for @max_active.
5091	*
5092	* We would prefer to do this in cgroup_init() above, but that
5093	* is called before init_workqueues(): so leave this until after.
5094	*/
5095	cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5096	BUG_ON(!cgroup_destroy_wq);
5097
5098	/*
5099	* Used to destroy pidlists and separate to serve as flush domain.
5100	* Cap @max_active to 1 too.
5101	*/
5102	cgroup_pidlist_destroy_wq = alloc_workqueue("cgroup_pidlist_destroy",
5103	0, 1);
5104	BUG_ON(!cgroup_pidlist_destroy_wq);
5105
5106	return 0;
5107	}
5108	core_initcall(cgroup_wq_init);
5109
5110	/*
5111	* proc_cgroup_show()
5112	* - Print task's cgroup paths into seq_file, one line for each hierarchy
5113	* - Used for /proc/<pid>/cgroup.
5114	*/
5115	int proc_cgroup_show(struct seq_file m, struct pid_namespace ns,
5116	struct pid pid, struct task_struct tsk)
5117	{
5118	char buf, path;
5119	int retval;
5120	struct cgroup_root *root;
5121
5122	retval = -ENOMEM;
5123	buf = kmalloc(PATH_MAX, GFP_KERNEL);
5124	if (!buf)
5125	goto out;
5126
5127	mutex_lock(&cgroup_mutex);
5128	down_read(&css_set_rwsem);
5129
5130	for_each_root(root) {
5131	struct cgroup_subsys *ss;
5132	struct cgroup *cgrp;
5133	int ssid, count = 0;
5134
5135	if (root == &cgrp_dfl_root && !cgrp_dfl_root_visible)
5136	continue;
5137
5138	seq_printf(m, "%d:", root->hierarchy_id);
5139	for_each_subsys(ss, ssid)
5140	if (root->subsys_mask & (1 << ssid))
5141	seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
5142	if (strlen(root->name))
5143	seq_printf(m, "%sname=%s", count ? "," : "",
5144	root->name);
5145	seq_putc(m, ':');
5146	cgrp = task_cgroup_from_root(tsk, root);
5147	path = cgroup_path(cgrp, buf, PATH_MAX);
5148	if (!path) {
5149	retval = -ENAMETOOLONG;
5150	goto out_unlock;
5151	}
5152	seq_puts(m, path);
5153	seq_putc(m, '\n');
5154	}
5155
5156	retval = 0;
5157	out_unlock:
5158	up_read(&css_set_rwsem);
5159	mutex_unlock(&cgroup_mutex);
5160	kfree(buf);
5161	out:
5162	return retval;
5163	}
5164
5165	/* Display information about each subsystem and each hierarchy */
5166	static int proc_cgroupstats_show(struct seq_file m, void v)
5167	{
5168	struct cgroup_subsys *ss;
5169	int i;
5170
5171	seq_puts(m, "#subsys_name\thierarchy\tnum_cgroups\tenabled\n");
5172	/*
5173	* ideally we don't want subsystems moving around while we do this.
5174	* cgroup_mutex is also necessary to guarantee an atomic snapshot of
5175	* subsys/hierarchy state.
5176	*/
5177	mutex_lock(&cgroup_mutex);
5178
5179	for_each_subsys(ss, i)
5180	seq_printf(m, "%s\t%d\t%d\t%d\n",
5181	ss->name, ss->root->hierarchy_id,
5182	atomic_read(&ss->root->nr_cgrps), !ss->disabled);
5183
5184	mutex_unlock(&cgroup_mutex);
5185	return 0;
5186	}
5187
5188	static int cgroupstats_open(struct inode inode, struct file file)
5189	{
5190	return single_open(file, proc_cgroupstats_show, NULL);
5191	}
5192
5193	static const struct file_operations proc_cgroupstats_operations = {
5194	.open = cgroupstats_open,
5195	.read = seq_read,
5196	.llseek = seq_lseek,
5197	.release = single_release,
5198	};
5199
5200	/**
5201	* cgroup_fork - initialize cgroup related fields during copy_process()
5202	* @child: pointer to task_struct of forking parent process.
5203	*
5204	* A task is associated with the init_css_set until cgroup_post_fork()
5205	* attaches it to the parent's css_set. Empty cg_list indicates that
5206	* @child isn't holding reference to its css_set.
5207	*/
5208	void cgroup_fork(struct task_struct *child)
5209	{
5210	RCU_INIT_POINTER(child->cgroups, &init_css_set);
5211	INIT_LIST_HEAD(&child->cg_list);
5212	}
5213
5214	/**
5215	* cgroup_post_fork - called on a new task after adding it to the task list
5216	* @child: the task in question
5217	*
5218	* Adds the task to the list running through its css_set if necessary and
5219	* call the subsystem fork() callbacks. Has to be after the task is
5220	* visible on the task list in case we race with the first call to
5221	* cgroup_task_iter_start() - to guarantee that the new task ends up on its
5222	* list.
5223	*/
5224	void cgroup_post_fork(struct task_struct *child)
5225	{
5226	struct cgroup_subsys *ss;
5227	int i;
5228
5229	/*
5230	* This may race against cgroup_enable_task_cg_lists(). As that
5231	* function sets use_task_css_set_links before grabbing
5232	* tasklist_lock and we just went through tasklist_lock to add
5233	* @child, it's guaranteed that either we see the set
5234	* use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5235	* @child during its iteration.
5236	*
5237	* If we won the race, @child is associated with %current's
5238	* css_set. Grabbing css_set_rwsem guarantees both that the
5239	* association is stable, and, on completion of the parent's
5240	* migration, @child is visible in the source of migration or
5241	* already in the destination cgroup. This guarantee is necessary
5242	* when implementing operations which need to migrate all tasks of
5243	* a cgroup to another.
5244	*
5245	* Note that if we lose to cgroup_enable_task_cg_lists(), @child
5246	* will remain in init_css_set. This is safe because all tasks are
5247	* in the init_css_set before cg_links is enabled and there's no
5248	* operation which transfers all tasks out of init_css_set.
5249	*/
5250	if (use_task_css_set_links) {
5251	struct css_set *cset;
5252
5253	down_write(&css_set_rwsem);
5254	cset = task_css_set(current);
5255	if (list_empty(&child->cg_list)) {
5256	rcu_assign_pointer(child->cgroups, cset);
5257	list_add(&child->cg_list, &cset->tasks);
5258	get_css_set(cset);
5259	}
5260	up_write(&css_set_rwsem);
5261	}
5262
5263	/*
5264	* Call ss->fork(). This must happen after @child is linked on
5265	* css_set; otherwise, @child might change state between ->fork()
5266	* and addition to css_set.
5267	*/
5268	for_each_subsys_which(ss, i, &have_fork_callback)
5269	ss->fork(child);
5270	}
5271
5272	/**
5273	* cgroup_exit - detach cgroup from exiting task
5274	* @tsk: pointer to task_struct of exiting process
5275	*
5276	* Description: Detach cgroup from @tsk and release it.
5277	*
5278	* Note that cgroups marked notify_on_release force every task in
5279	* them to take the global cgroup_mutex mutex when exiting.
5280	* This could impact scaling on very large systems. Be reluctant to
5281	* use notify_on_release cgroups where very high task exit scaling
5282	* is required on large systems.
5283	*
5284	* We set the exiting tasks cgroup to the root cgroup (top_cgroup). We
5285	* call cgroup_exit() while the task is still competent to handle
5286	* notify_on_release(), then leave the task attached to the root cgroup in
5287	* each hierarchy for the remainder of its exit. No need to bother with
5288	* init_css_set refcnting. init_css_set never goes away and we can't race
5289	* with migration path - PF_EXITING is visible to migration path.
5290	*/
5291	void cgroup_exit(struct task_struct *tsk)
5292	{
5293	struct cgroup_subsys *ss;
5294	struct css_set *cset;
5295	bool put_cset = false;
5296	int i;
5297
5298	/*
5299	* Unlink from @tsk from its css_set. As migration path can't race
5300	* with us, we can check cg_list without grabbing css_set_rwsem.
5301	*/
5302	if (!list_empty(&tsk->cg_list)) {
5303	down_write(&css_set_rwsem);
5304	list_del_init(&tsk->cg_list);
5305	up_write(&css_set_rwsem);
5306	put_cset = true;
5307	}
5308
5309	/* Reassign the task to the init_css_set. */
5310	cset = task_css_set(tsk);
5311	RCU_INIT_POINTER(tsk->cgroups, &init_css_set);
5312
5313	/* see cgroup_post_fork() for details */
5314	for_each_subsys_which(ss, i, &have_exit_callback) {
5315	struct cgroup_subsys_state *old_css = cset->subsys[i];
5316	struct cgroup_subsys_state *css = task_css(tsk, i);
5317
5318	ss->exit(css, old_css, tsk);
5319	}
5320
5321	if (put_cset)
5322	put_css_set(cset);
5323	}
5324
5325	static void check_for_release(struct cgroup *cgrp)
5326	{
5327	if (notify_on_release(cgrp) && !cgroup_has_tasks(cgrp) &&
5328	!css_has_online_children(&cgrp->self) && !cgroup_is_dead(cgrp))
5329	schedule_work(&cgrp->release_agent_work);
5330	}
5331
5332	/*
5333	* Notify userspace when a cgroup is released, by running the
5334	* configured release agent with the name of the cgroup (path
5335	* relative to the root of cgroup file system) as the argument.
5336	*
5337	* Most likely, this user command will try to rmdir this cgroup.
5338	*
5339	* This races with the possibility that some other task will be
5340	* attached to this cgroup before it is removed, or that some other
5341	* user task will 'mkdir' a child cgroup of this cgroup. That's ok.
5342	* The presumed 'rmdir' will fail quietly if this cgroup is no longer
5343	* unused, and this cgroup will be reprieved from its death sentence,
5344	* to continue to serve a useful existence. Next time it's released,
5345	* we will get notified again, if it still has 'notify_on_release' set.
5346	*
5347	* The final arg to call_usermodehelper() is UMH_WAIT_EXEC, which
5348	* means only wait until the task is successfully execve()'d. The
5349	* separate release agent task is forked by call_usermodehelper(),
5350	* then control in this thread returns here, without waiting for the
5351	* release agent task. We don't bother to wait because the caller of
5352	* this routine has no use for the exit status of the release agent
5353	* task, so no sense holding our caller up for that.
5354	*/
5355	static void cgroup_release_agent(struct work_struct *work)
5356	{
5357	struct cgroup *cgrp =
5358	container_of(work, struct cgroup, release_agent_work);
5359	char pathbuf = NULL, agentbuf = NULL, *path;
5360	char argv[3], envp[3];
5361
5362	mutex_lock(&cgroup_mutex);
5363
5364	pathbuf = kmalloc(PATH_MAX, GFP_KERNEL);
5365	agentbuf = kstrdup(cgrp->root->release_agent_path, GFP_KERNEL);
5366	if (!pathbuf \|\| !agentbuf)
5367	goto out;
5368
5369	path = cgroup_path(cgrp, pathbuf, PATH_MAX);
5370	if (!path)
5371	goto out;
5372
5373	argv[0] = agentbuf;
5374	argv[1] = path;
5375	argv[2] = NULL;
5376
5377	/* minimal command environment */
5378	envp[0] = "HOME=/";
5379	envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
5380	envp[2] = NULL;
5381
5382	mutex_unlock(&cgroup_mutex);
5383	call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
5384	goto out_free;
5385	out:
5386	mutex_unlock(&cgroup_mutex);
5387	out_free:
5388	kfree(agentbuf);
5389	kfree(pathbuf);
5390	}
5391
5392	static int __init cgroup_disable(char *str)
5393	{
5394	struct cgroup_subsys *ss;
5395	char *token;
5396	int i;
5397
5398	while ((token = strsep(&str, ",")) != NULL) {
5399	if (!*token)
5400	continue;
5401
5402	for_each_subsys(ss, i) {
5403	if (!strcmp(token, ss->name)) {
5404	ss->disabled = 1;
5405	printk(KERN_INFO "Disabling %s control group"
5406	" subsystem\n", ss->name);
5407	break;
5408	}
5409	}
5410	}
5411	return 1;
5412	}
5413	__setup("cgroup_disable=", cgroup_disable);
5414
5415	static int __init cgroup_set_legacy_files_on_dfl(char *str)
5416	{
5417	printk("cgroup: using legacy files on the default hierarchy\n");
5418	cgroup_legacy_files_on_dfl = true;
5419	return 0;
5420	}
5421	__setup("cgroup__DEVEL__legacy_files_on_dfl", cgroup_set_legacy_files_on_dfl);
5422
5423	/**
5424	* css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5425	* @dentry: directory dentry of interest
5426	* @ss: subsystem of interest
5427	*
5428	* If @dentry is a directory for a cgroup which has @ss enabled on it, try
5429	* to get the corresponding css and return it. If such css doesn't exist
5430	* or can't be pinned, an ERR_PTR value is returned.
5431	*/
5432	struct cgroup_subsys_state css_tryget_online_from_dir(struct dentry dentry,
5433	struct cgroup_subsys *ss)
5434	{
5435	struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5436	struct cgroup_subsys_state *css = NULL;
5437	struct cgroup *cgrp;
5438
5439	/* is @dentry a cgroup dir? */
5440	if (dentry->d_sb->s_type != &cgroup_fs_type \|\| !kn \|\|
5441	kernfs_type(kn) != KERNFS_DIR)
5442	return ERR_PTR(-EBADF);
5443
5444	rcu_read_lock();
5445
5446	/*
5447	* This path doesn't originate from kernfs and @kn could already
5448	* have been or be removed at any point. @kn->priv is RCU
5449	* protected for this access. See css_release_work_fn() for details.
5450	*/
5451	cgrp = rcu_dereference(kn->priv);
5452	if (cgrp)
5453	css = cgroup_css(cgrp, ss);
5454
5455	if (!css \|\| !css_tryget_online(css))
5456	css = ERR_PTR(-ENOENT);
5457
5458	rcu_read_unlock();
5459	return css;
5460	}
5461
5462	/**
5463	* css_from_id - lookup css by id
5464	* @id: the cgroup id
5465	* @ss: cgroup subsys to be looked into
5466	*
5467	* Returns the css if there's valid one with @id, otherwise returns NULL.
5468	* Should be called under rcu_read_lock().
5469	*/
5470	struct cgroup_subsys_state css_from_id(int id, struct cgroup_subsys ss)
5471	{
5472	WARN_ON_ONCE(!rcu_read_lock_held());
5473	return id > 0 ? idr_find(&ss->css_idr, id) : NULL;
5474	}
5475
5476	#ifdef CONFIG_CGROUP_DEBUG
5477	static struct cgroup_subsys_state *
5478	debug_css_alloc(struct cgroup_subsys_state *parent_css)
5479	{
5480	struct cgroup_subsys_state css = kzalloc(sizeof(css), GFP_KERNEL);
5481
5482	if (!css)
5483	return ERR_PTR(-ENOMEM);
5484
5485	return css;
5486	}
5487
5488	static void debug_css_free(struct cgroup_subsys_state *css)
5489	{
5490	kfree(css);
5491	}
5492
5493	static u64 debug_taskcount_read(struct cgroup_subsys_state *css,
5494	struct cftype *cft)
5495	{
5496	return cgroup_task_count(css->cgroup);
5497	}
5498
5499	static u64 current_css_set_read(struct cgroup_subsys_state *css,
5500	struct cftype *cft)
5501	{
5502	return (u64)(unsigned long)current->cgroups;
5503	}
5504
5505	static u64 current_css_set_refcount_read(struct cgroup_subsys_state *css,
5506	struct cftype *cft)
5507	{
5508	u64 count;
5509
5510	rcu_read_lock();
5511	count = atomic_read(&task_css_set(current)->refcount);
5512	rcu_read_unlock();
5513	return count;
5514	}
5515
5516	static int current_css_set_cg_links_read(struct seq_file seq, void v)
5517	{
5518	struct cgrp_cset_link *link;
5519	struct css_set *cset;
5520	char *name_buf;
5521
5522	name_buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
5523	if (!name_buf)
5524	return -ENOMEM;
5525
5526	down_read(&css_set_rwsem);
5527	rcu_read_lock();
5528	cset = rcu_dereference(current->cgroups);
5529	list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
5530	struct cgroup *c = link->cgrp;
5531
5532	cgroup_name(c, name_buf, NAME_MAX + 1);
5533	seq_printf(seq, "Root %d group %s\n",
5534	c->root->hierarchy_id, name_buf);
5535	}
5536	rcu_read_unlock();
5537	up_read(&css_set_rwsem);
5538	kfree(name_buf);
5539	return 0;
5540	}
5541
5542	#define MAX_TASKS_SHOWN_PER_CSS 25
5543	static int cgroup_css_links_read(struct seq_file seq, void v)
5544	{
5545	struct cgroup_subsys_state *css = seq_css(seq);
5546	struct cgrp_cset_link *link;
5547
5548	down_read(&css_set_rwsem);
5549	list_for_each_entry(link, &css->cgroup->cset_links, cset_link) {
5550	struct css_set *cset = link->cset;
5551	struct task_struct *task;
5552	int count = 0;
5553
5554	seq_printf(seq, "css_set %p\n", cset);
5555
5556	list_for_each_entry(task, &cset->tasks, cg_list) {
5557	if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5558	goto overflow;
5559	seq_printf(seq, " task %d\n", task_pid_vnr(task));
5560	}
5561
5562	list_for_each_entry(task, &cset->mg_tasks, cg_list) {
5563	if (count++ > MAX_TASKS_SHOWN_PER_CSS)
5564	goto overflow;
5565	seq_printf(seq, " task %d\n", task_pid_vnr(task));
5566	}
5567	continue;
5568	overflow:
5569	seq_puts(seq, " ...\n");
5570	}
5571	up_read(&css_set_rwsem);
5572	return 0;
5573	}
5574
5575	static u64 releasable_read(struct cgroup_subsys_state css, struct cftype cft)
5576	{
5577	return (!cgroup_has_tasks(css->cgroup) &&
5578	!css_has_online_children(&css->cgroup->self));
5579	}
5580
5581	static struct cftype debug_files[] = {
5582	{
5583	.name = "taskcount",
5584	.read_u64 = debug_taskcount_read,
5585	},
5586
5587	{
5588	.name = "current_css_set",
5589	.read_u64 = current_css_set_read,
5590	},
5591
5592	{
5593	.name = "current_css_set_refcount",
5594	.read_u64 = current_css_set_refcount_read,
5595	},
5596
5597	{
5598	.name = "current_css_set_cg_links",
5599	.seq_show = current_css_set_cg_links_read,
5600	},
5601
5602	{
5603	.name = "cgroup_css_links",
5604	.seq_show = cgroup_css_links_read,
5605	},
5606
5607	{
5608	.name = "releasable",
5609	.read_u64 = releasable_read,
5610	},
5611
5612	{ } /* terminate */
5613	};
5614
5615	struct cgroup_subsys debug_cgrp_subsys = {
5616	.css_alloc = debug_css_alloc,
5617	.css_free = debug_css_free,
5618	.legacy_cftypes = debug_files,
5619	};
5620	#endif /* CONFIG_CGROUP_DEBUG */